diff --git a/.github/actions/release-cn-artifacts/action.yaml b/.github/actions/release-cn-artifacts/action.yaml
index 2825d3f5d0..fe78d5a760 100644
--- a/.github/actions/release-cn-artifacts/action.yaml
+++ b/.github/actions/release-cn-artifacts/action.yaml
@@ -37,17 +37,14 @@ inputs:
     description: Whether to push the latest tag of the image
     required: false
     default: 'true'
-  aws-cn-s3-bucket:
-    description: S3 bucket to store released artifacts in CN region
+  proxy-url:
+    description: The url of the S3 proxy server
     required: true
-  aws-cn-access-key-id:
-    description: AWS access key id in CN region
+  proxy-username:
+    description: The username of the S3 proxy
     required: true
-  aws-cn-secret-access-key:
-    description: AWS secret access key in CN region
-    required: true
-  aws-cn-region:
-    description: AWS region in CN
+  proxy-password:
+    description: The password of the S3 proxy
     required: true
   upload-to-s3:
     description: Upload to S3
@@ -77,21 +74,13 @@ runs:
       with:
         path: ${{ inputs.artifacts-dir }}
 
-    - name: Install s5cmd
-      shell: bash
-      run: |
-        wget https://github.com/peak/s5cmd/releases/download/v2.3.0/s5cmd_2.3.0_Linux-64bit.tar.gz
-        tar -xzf s5cmd_2.3.0_Linux-64bit.tar.gz
-        sudo mv s5cmd /usr/local/bin/
-        sudo chmod +x /usr/local/bin/s5cmd
-
     - name: Release artifacts to cn region
       uses: nick-invision/retry@v2
       if: ${{ inputs.upload-to-s3 == 'true' }}
       env:
-        AWS_ACCESS_KEY_ID: ${{ inputs.aws-cn-access-key-id }}
-        AWS_SECRET_ACCESS_KEY: ${{ inputs.aws-cn-secret-access-key }}
-        AWS_REGION: ${{ inputs.aws-cn-region }}
+        PROXY_URL: ${{ inputs.proxy-url }}
+        PROXY_USERNAME: ${{ inputs.proxy-username }}
+        PROXY_PASSWORD: ${{ inputs.proxy-password }}
         UPDATE_VERSION_INFO: ${{ inputs.update-version-info }}
       with:
         max_attempts: ${{ inputs.upload-max-retry-times }}
@@ -99,8 +88,7 @@ runs:
         command: |
           ./.github/scripts/upload-artifacts-to-s3.sh \
             ${{ inputs.artifacts-dir }} \
-            ${{ inputs.version }} \
-            ${{ inputs.aws-cn-s3-bucket }}
+            ${{ inputs.version }}
 
     - name: Push greptimedb image from Dockerhub to ACR
       shell: bash
diff --git a/.github/actions/setup-chaos/action.yml b/.github/actions/setup-chaos/action.yml
index 76eb48cf4a..4b3fb86744 100644
--- a/.github/actions/setup-chaos/action.yml
+++ b/.github/actions/setup-chaos/action.yml
@@ -1,15 +1,16 @@
-name: Setup Kind
-description: Deploy Kind
+name: Setup Chaos Mesh
+description: Install and wait for Chaos Mesh
 runs:
   using: composite
   steps:
   - uses: actions/checkout@v4
-  - name: Create kind cluster
+  - name: Install Chaos Mesh
     shell: bash
     run: |
       helm repo add chaos-mesh https://charts.chaos-mesh.org
+      helm repo update chaos-mesh
       kubectl create ns chaos-mesh
-      helm install chaos-mesh chaos-mesh/chaos-mesh -n=chaos-mesh --version 2.6.3
+      helm install chaos-mesh chaos-mesh/chaos-mesh -n=chaos-mesh --set chaosDaemon.runtime=containerd --set chaosDaemon.socketPath=/run/containerd/containerd.sock --version 2.8.0
   - name: Print Chaos-mesh
     if: always()
     shell: bash
diff --git a/.github/actions/setup-greptimedb-cluster/with-minio-repartition-gc.yaml b/.github/actions/setup-greptimedb-cluster/with-minio-repartition-gc.yaml
index b836606ad6..5b7f43d81b 100644
--- a/.github/actions/setup-greptimedb-cluster/with-minio-repartition-gc.yaml
+++ b/.github/actions/setup-greptimedb-cluster/with-minio-repartition-gc.yaml
@@ -1,3 +1,8 @@
+logging:
+  level: "info"
+  format: "json"
+  filters:
+  - mito2::sst::file=debug
 meta:
   configData: |-
     [runtime]
diff --git a/.github/cargo-blacklist.txt b/.github/cargo-blacklist.txt
index d2f071130e..32e7878a86 100644
--- a/.github/cargo-blacklist.txt
+++ b/.github/cargo-blacklist.txt
@@ -1,3 +1,2 @@
 native-tls
 openssl
-aws-lc-sys
diff --git a/.github/scripts/check-version.sh b/.github/scripts/check-version.sh
index 28c2812ded..1efa3bb4db 100755
--- a/.github/scripts/check-version.sh
+++ b/.github/scripts/check-version.sh
@@ -30,13 +30,72 @@ CLEAN_LATEST=$(echo "$LATEST_VERSION" | sed 's/^v//' | sed 's/-nightly-.*//')
 echo "Current version: $CLEAN_CURRENT"
 echo "Latest release version: $CLEAN_LATEST"
 
-# Use sort -V to compare versions
-HIGHER_VERSION=$(printf "%s\n%s" "$CLEAN_CURRENT" "$CLEAN_LATEST" | sort -V | tail -n1)
+# Function to extract base version (without pre-release suffix)
+get_base_version() {
+  echo "$1" | sed -E 's/-(alpha|beta|rc|pre).*//'
+}
 
-if [ "$HIGHER_VERSION" = "$CLEAN_CURRENT" ]; then
+# Function to check if a version is pre-release
+is_prerelease() {
+  [[ "$1" =~ -(alpha|beta|rc|pre) ]]
+}
+
+# Compare versions properly considering pre-release
+compare_versions() {
+  local current=$1
+  local latest=$2
+
+  # Extract base versions
+  local current_base=$(get_base_version "$current")
+  local latest_base=$(get_base_version "$latest")
+
+  # Compare base versions first
+  HIGHER_BASE=$(printf "%s\n%s" "$current_base" "$latest_base" | sort -V | tail -n1)
+
+  if [ "$HIGHER_BASE" = "$latest_base" ] && [ "$current_base" != "$latest_base" ]; then
+    # Latest has higher base version
+    echo "current_older"
+    return
+  elif [ "$HIGHER_BASE" = "$current_base" ] && [ "$current_base" != "$latest_base" ]; then
+    # Current has higher base version
+    echo "current_newer"
+    return
+  fi
+
+  # Base versions are equal, compare pre-release status
+  if [ "$current_base" = "$latest_base" ]; then
+    # If current is pre-release and latest is not, current is older
+    if is_prerelease "$current" && ! is_prerelease "$latest"; then
+      echo "current_older"
+      return
+    fi
+
+    # If latest is pre-release and current is not, current is newer
+    if ! is_prerelease "$current" && is_prerelease "$latest"; then
+      echo "current_newer"
+      return
+    fi
+  fi
+
+  # Both are same type or different base versions already handled, use sort -V
+  HIGHER_VERSION=$(printf "%s\n%s" "$current" "$latest" | sort -V | tail -n1)
+  if [ "$HIGHER_VERSION" = "$current" ]; then
+    echo "current_newer_or_equal"
+  else
+    echo "current_older"
+  fi
+}
+
+RESULT=$(compare_versions "$CLEAN_CURRENT" "$CLEAN_LATEST")
+
+if [ "$RESULT" = "current_newer" ] || [ "$RESULT" = "current_newer_or_equal" ]; then
   echo "Current version ($CLEAN_CURRENT) is NEWER than or EQUAL to latest ($CLEAN_LATEST)"
-  echo "is-current-version-latest=true" >> $GITHUB_OUTPUT
+  if [ -n "$GITHUB_OUTPUT" ]; then
+    echo "is-current-version-latest=true" >> $GITHUB_OUTPUT
+  fi
 else
   echo "Current version ($CLEAN_CURRENT) is OLDER than latest ($CLEAN_LATEST)"
-  echo "is-current-version-latest=false" >> $GITHUB_OUTPUT
+  if [ -n "$GITHUB_OUTPUT" ]; then
+    echo "is-current-version-latest=false" >> $GITHUB_OUTPUT
+  fi
 fi
diff --git a/.github/scripts/update-dev-builder-version.sh b/.github/scripts/update-dev-builder-version.sh
index 38466760a4..a593f385fd 100755
--- a/.github/scripts/update-dev-builder-version.sh
+++ b/.github/scripts/update-dev-builder-version.sh
@@ -30,8 +30,11 @@ update_dev_builder_version() {
       --body "This PR updates the dev-builder image tag" \
       --base main \
       --head $BRANCH_NAME \
-      --reviewer zyy17 \
-      --reviewer daviderli614
+      --reviewer sunng87 \
+      --reviewer daviderli614 \
+      --reviewer killme2008 \
+      --reviewer evenyag \
+      --reviewer fengjiachun
 }
 
 update_dev_builder_version
diff --git a/.github/scripts/upload-artifacts-to-s3.sh b/.github/scripts/upload-artifacts-to-s3.sh
index 75c8f8d932..9c92859e3b 100755
--- a/.github/scripts/upload-artifacts-to-s3.sh
+++ b/.github/scripts/upload-artifacts-to-s3.sh
@@ -5,16 +5,15 @@ set -o pipefail
 
 ARTIFACTS_DIR=$1
 VERSION=$2
-AWS_S3_BUCKET=$3
 RELEASE_DIRS="releases/greptimedb"
 GREPTIMEDB_REPO="GreptimeTeam/greptimedb"
 
 # Check if necessary variables are set.
 function check_vars() {
-  for var in AWS_S3_BUCKET VERSION ARTIFACTS_DIR; do
+  for var in VERSION ARTIFACTS_DIR; do
     if [ -z "${!var}" ]; then
       echo "$var is not set or empty."
-      echo "Usage: $0 <artifacts-dir> <version> <aws-s3-bucket>"
+      echo "Usage: $0 <artifacts-dir> <version>"
       exit 1
     fi
   done
@@ -33,8 +32,18 @@ function upload_artifacts() {
   #    ├── greptime-darwin-amd64-v0.2.0.sha256sum
   #    └── greptime-darwin-amd64-v0.2.0.tar.gz
   find "$ARTIFACTS_DIR" -type f \( -name "*.tar.gz" -o -name "*.sha256sum" \) | while IFS= read -r file; do
-    s5cmd cp \
-      "$file" "s3://$AWS_S3_BUCKET/$RELEASE_DIRS/$VERSION/$(basename "$file")"
+    filename=$(basename "$file")
+    TARGET_URL="$PROXY_URL/$RELEASE_DIRS/$VERSION"
+
+    curl -X PUT \
+      -u "$PROXY_USERNAME:$PROXY_PASSWORD" \
+      -F "file=@$file" \
+      --max-time 3600 \
+      --connect-timeout 20 \
+      --retry 5 \
+      --retry-delay 10 \
+      --retry-max-time 3000 \
+      "$TARGET_URL"
   done
 }
 
@@ -45,16 +54,34 @@ function update_version_info() {
     if [[ "$VERSION" =~ ^v[0-9]+\.[0-9]+\.[0-9]+$ ]]; then
       echo "Updating latest-version.txt"
       echo "$VERSION" > latest-version.txt
-      s5cmd cp \
-        latest-version.txt "s3://$AWS_S3_BUCKET/$RELEASE_DIRS/latest-version.txt"
+      TARGET_URL="$PROXY_URL/$RELEASE_DIRS"
+
+      curl -X PUT \
+        -u "$PROXY_USERNAME:$PROXY_PASSWORD" \
+        -F "file=@latest-version.txt" \
+        --max-time 3600 \
+        --connect-timeout 20 \
+        --retry 5 \
+        --retry-delay 10 \
+        --retry-max-time 3000 \
+        "$TARGET_URL"
     fi
 
     # If it's the nightly release, update latest-nightly-version.txt.
     if [[ "$VERSION" == *"nightly"* ]]; then
       echo "Updating latest-nightly-version.txt"
       echo "$VERSION" > latest-nightly-version.txt
-      s5cmd cp \
-        latest-nightly-version.txt "s3://$AWS_S3_BUCKET/$RELEASE_DIRS/latest-nightly-version.txt"
+
+      TARGET_URL="$PROXY_URL/$RELEASE_DIRS"
+      curl -X PUT \
+        -u "$PROXY_USERNAME:$PROXY_PASSWORD" \
+        -F "file=@latest-nightly-version.txt" \
+        --max-time 3600 \
+        --connect-timeout 20 \
+        --retry 5 \
+        --retry-delay 10 \
+        --retry-max-time 3000 \
+        "$TARGET_URL"
     fi
   fi
 }
@@ -93,10 +120,10 @@ function main() {
 }
 
 # Usage example:
-#   AWS_ACCESS_KEY_ID=<your_access_key_id> \
-#   AWS_SECRET_ACCESS_KEY=<your_secret_access_key> \
-#   AWS_DEFAULT_REGION=<your_region> \
+#   PROXY_URL=<proxy_url> \
+#   PROXY_USERNAME=<proxy_username> \
+#   PROXY_PASSWORD=<proxy_password> \
 #   UPDATE_VERSION_INFO=true \
 #   DOWNLOAD_ARTIFACTS_FROM_GITHUB=false \
-#     ./upload-artifacts-to-s3.sh <artifacts-dir> <version> <aws-s3-bucket>
+#     ./upload-artifacts-to-s3.sh <artifacts-dir> <version>
 main
diff --git a/.github/workflows/bump-helm-charts-version.yml b/.github/workflows/bump-helm-charts-version.yml
new file mode 100644
index 0000000000..5921ec8a8c
--- /dev/null
+++ b/.github/workflows/bump-helm-charts-version.yml
@@ -0,0 +1,29 @@
+name: Bump helm charts version
+
+on:
+  workflow_dispatch:
+    inputs:
+      version:
+        description: The version to bump (e.g. v1.0.0)
+        required: true
+        type: string
+
+jobs:
+  bump-helm-charts-version:
+    name: Bump helm charts version
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+      pull-requests: write
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Bump helm charts version
+        env:
+          GITHUB_TOKEN: ${{ secrets.HELM_CHARTS_REPO_TOKEN }}
+          VERSION: ${{ inputs.version }}
+        run: |
+          ./.github/scripts/update-helm-charts-version.sh
diff --git a/.github/workflows/bump-homebrew-greptime-version.yml b/.github/workflows/bump-homebrew-greptime-version.yml
new file mode 100644
index 0000000000..af8ca8fc99
--- /dev/null
+++ b/.github/workflows/bump-homebrew-greptime-version.yml
@@ -0,0 +1,29 @@
+name: Bump homebrew greptime version
+
+on:
+  workflow_dispatch:
+    inputs:
+      version:
+        description: The version to bump (e.g. v1.0.0)
+        required: true
+        type: string
+
+jobs:
+  bump-homebrew-greptime-version:
+    name: Bump homebrew greptime version
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+      pull-requests: write
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Bump homebrew greptime version
+        env:
+          GITHUB_TOKEN: ${{ secrets.HOMEBREW_GREPTIME_REPO_TOKEN }}
+          VERSION: ${{ inputs.version }}
+        run: |
+          ./.github/scripts/update-homebrew-greptme-version.sh
diff --git a/.github/workflows/dev-build.yml b/.github/workflows/dev-build.yml
index 021867e4ed..d03fbeff14 100644
--- a/.github/workflows/dev-build.yml
+++ b/.github/workflows/dev-build.yml
@@ -285,10 +285,9 @@ jobs:
           dst-image-registry: ${{ vars.ACR_IMAGE_REGISTRY }}
           dst-image-namespace: ${{ vars.IMAGE_NAMESPACE }}
           version: ${{ needs.allocate-runners.outputs.version }}
-          aws-cn-s3-bucket: ${{ vars.AWS_RELEASE_BUCKET }}
-          aws-cn-access-key-id: ${{ secrets.AWS_CN_ACCESS_KEY_ID }}
-          aws-cn-secret-access-key: ${{ secrets.AWS_CN_SECRET_ACCESS_KEY }}
-          aws-cn-region: ${{ vars.AWS_RELEASE_BUCKET_REGION }}
+          proxy-url: ${{ secrets.PROXY_URL }}
+          proxy-username: ${{ secrets.PROXY_USERNAME }}
+          proxy-password: ${{ secrets.PROXY_PASSWORD }}
           upload-to-s3: ${{ inputs.upload_artifacts_to_s3 }}
           dev-mode: true                     # Only build the standard images(exclude centos images).
           push-latest-tag: false             # Don't push the latest tag to registry.
diff --git a/.github/workflows/develop.yml b/.github/workflows/develop.yml
index 0238e92c8d..d0d2804c6a 100644
--- a/.github/workflows/develop.yml
+++ b/.github/workflows/develop.yml
@@ -319,7 +319,13 @@ jobs:
         include:
           - target: "fuzz_repartition_table"
             mode:
-              name: "Local WAL Repartition GC"
+              name: "Local WAL mito table repartition"
+              minio: true
+              kafka: false
+              values: "with-minio-repartition-gc.yaml"
+          - target: "fuzz_repartition_metric_table"
+            mode:
+              name: "Local WAL metric table repartition"
               minio: true
               kafka: false
               values: "with-minio-repartition-gc.yaml"
@@ -455,6 +461,14 @@ jobs:
           path: /tmp/fuzz-monitor-dumps
           if-no-files-found: warn
           retention-days: 3
+      - name: Upload CSV dumps
+        if: failure()
+        uses: actions/upload-artifact@v4
+        with:
+          name: fuzz-tests-csv-dumps-${{ matrix.mode.name }}-${{ matrix.target }}
+          path: /tmp/greptime-fuzz-dumps
+          if-no-files-found: warn
+          retention-days: 3
       - name: Delete cluster
         if: success()
         shell: bash
@@ -492,6 +506,12 @@ jobs:
               minio: true
               kafka: false
               values: "with-minio.yaml"
+          - target: "fuzz_repartition_table_chaos"
+            mode:
+              name: "Local WAL repartition chaos"
+              minio: true
+              kafka: false
+              values: "with-minio-repartition-gc.yaml"
     steps:
       - name: Remove unused software
         run: |
diff --git a/.github/workflows/nightly-build.yml b/.github/workflows/nightly-build.yml
index 9eaa38c789..14ebb6e715 100644
--- a/.github/workflows/nightly-build.yml
+++ b/.github/workflows/nightly-build.yml
@@ -236,10 +236,9 @@ jobs:
           dst-image-registry: ${{ vars.ACR_IMAGE_REGISTRY }}
           dst-image-namespace: ${{ vars.IMAGE_NAMESPACE }}
           version: ${{ needs.allocate-runners.outputs.version }}
-          aws-cn-s3-bucket: ${{ vars.AWS_RELEASE_BUCKET }}
-          aws-cn-access-key-id: ${{ secrets.AWS_CN_ACCESS_KEY_ID }}
-          aws-cn-secret-access-key: ${{ secrets.AWS_CN_SECRET_ACCESS_KEY }}
-          aws-cn-region: ${{ vars.AWS_RELEASE_BUCKET_REGION }}
+          proxy-url: ${{ secrets.PROXY_URL }}
+          proxy-username: ${{ secrets.PROXY_USERNAME }}
+          proxy-password: ${{ secrets.PROXY_PASSWORD }}
           upload-to-s3: false
           dev-mode: false
           update-version-info: false  # Don't update version info in S3.
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 3b0eb2d68c..9f8f2d9703 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -358,10 +358,9 @@ jobs:
           dst-image-registry: ${{ vars.ACR_IMAGE_REGISTRY }}
           dst-image-namespace: ${{ vars.IMAGE_NAMESPACE }}
           version: ${{ needs.allocate-runners.outputs.version }}
-          aws-cn-s3-bucket: ${{ vars.AWS_RELEASE_BUCKET }}
-          aws-cn-access-key-id: ${{ secrets.AWS_CN_ACCESS_KEY_ID }}
-          aws-cn-secret-access-key: ${{ secrets.AWS_CN_SECRET_ACCESS_KEY }}
-          aws-cn-region: ${{ vars.AWS_RELEASE_BUCKET_REGION }}
+          proxy-url: ${{ secrets.PROXY_URL }}
+          proxy-username: ${{ secrets.PROXY_USERNAME }}
+          proxy-password: ${{ secrets.PROXY_PASSWORD }}
           dev-mode: false
           upload-to-s3: true
           update-version-info: true
diff --git a/.github/workflows/run-multi-lang-tests.yml b/.github/workflows/run-multi-lang-tests.yml
index 4fc405b5b2..52fdca2117 100644
--- a/.github/workflows/run-multi-lang-tests.yml
+++ b/.github/workflows/run-multi-lang-tests.yml
@@ -127,7 +127,7 @@ jobs:
         run: |
           ./bins/greptime standalone start \
             --http-addr 0.0.0.0:${{ inputs.http-port }} \
-            --rpc-addr 0.0.0.0:4001 \
+            --grpc-bind-addr 0.0.0.0:4001 \
             --mysql-addr 0.0.0.0:${{ inputs.mysql-port }} \
             --postgres-addr 0.0.0.0:${{ inputs.postgres-port }} \
             --user-provider=static_user_provider:cmd:${{ inputs.username }}=${{ inputs.password }} > /tmp/greptimedb.log 2>&1 &
diff --git a/.gitignore b/.gitignore
index 862eb8c5b4..0789f3ec10 100644
--- a/.gitignore
+++ b/.gitignore
@@ -65,8 +65,14 @@ greptimedb_data
 # github
 !/.github
 
-# Claude code
+# AI related
 CLAUDE.md
-
-# AGENTS.md
 AGENTS.md
+.codex
+.gemini
+.opencode
+.worktrees/
+
+# local design docs
+docs/specs/
+.vs/ 
diff --git a/Cargo.lock b/Cargo.lock
index 85c2b1ed2d..acfa25df68 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -200,9 +200,9 @@ dependencies = [
 
 [[package]]
 name = "anyhow"
-version = "1.0.98"
+version = "1.0.102"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e16d2d3311acee920a9eb8d33b8cbc1787ce4a264e85f964c2404b969bdcd487"
+checksum = "7f202df86484c868dbad7eaa557ef785d5c66295e41b460ef922eca0723b842c"
 
 [[package]]
 name = "anymap2"
@@ -212,7 +212,7 @@ checksum = "d301b3b94cb4b2f23d7917810addbbaff90738e0ca2be692bd027e70d7e0330c"
 
 [[package]]
 name = "api"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "arrow-schema 57.3.0",
  "common-base",
@@ -565,7 +565,7 @@ dependencies = [
  "arrow-schema 57.3.0",
  "arrow-select 57.3.0",
  "flatbuffers",
- "lz4_flex 0.12.0",
+ "lz4_flex 0.12.1",
  "zstd",
 ]
 
@@ -933,7 +933,7 @@ checksum = "1505bd5d3d116872e7271a6d4e16d81d0c8570876c8de68093a09ac269d8aac0"
 
 [[package]]
 name = "auth"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "async-trait",
@@ -982,6 +982,29 @@ dependencies = [
  "cc",
 ]
 
+[[package]]
+name = "aws-lc-rs"
+version = "1.16.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0ec6fb3fe69024a75fa7e1bfb48aa6cf59706a101658ea01bfd33b2b248a038f"
+dependencies = [
+ "aws-lc-sys",
+ "untrusted 0.7.1",
+ "zeroize",
+]
+
+[[package]]
+name = "aws-lc-sys"
+version = "0.40.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f50037ee5e1e41e7b8f9d161680a725bd1626cb6f8c7e901f91f942850852fe7"
+dependencies = [
+ "cc",
+ "cmake",
+ "dunce",
+ "fs_extra",
+]
+
 [[package]]
 name = "axum"
 version = "0.6.20"
@@ -1129,9 +1152,9 @@ dependencies = [
 
 [[package]]
 name = "backon"
-version = "1.5.1"
+version = "1.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "302eaff5357a264a2c42f127ecb8bac761cf99749fc3dc95677e2743991f99e7"
+checksum = "cffb0e931875b666fc4fcb20fee52e9bbd1ef836fd9e9e04ec21555f9f85f7ef"
 dependencies = [
  "fastrand",
  "gloo-timers",
@@ -1252,7 +1275,7 @@ version = "0.72.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4f72209734318d0b619a5e0f5129918b848c416e122a3c4ce054e03cb87b726f"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "cexpr",
  "clang-sys",
  "itertools 0.13.0",
@@ -1287,11 +1310,11 @@ checksum = "bef38d45163c2f1dde094a7dfd33ccf595c92905c8f8f4fdc18d06fb1037718a"
 
 [[package]]
 name = "bitflags"
-version = "2.9.1"
+version = "2.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1b8e56985ec62d17e9c1001dc89c88ecd7dc08e47eba5ec7c29c7b5eeecde967"
+checksum = "c4512299f36f043ab09a583e57bceb5a5aab7a73db1805848e8fef3c9e8c78b3"
 dependencies = [
- "serde",
+ "serde_core",
 ]
 
 [[package]]
@@ -1523,7 +1546,7 @@ dependencies = [
 
 [[package]]
 name = "cache"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "catalog",
  "common-error",
@@ -1559,7 +1582,7 @@ checksum = "37b2a672a2cb129a2e41c10b1224bb368f9f37a2b16b612598138befd7b37eb5"
 
 [[package]]
 name = "catalog"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "arrow 57.3.0",
@@ -1654,6 +1677,12 @@ version = "0.4.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4f4c707c6a209cbe82d10abd08e1ea8995e9ea937d2550646e02798948992be0"
 
+[[package]]
+name = "cesu8"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6d43a04d8753f35258c91f8ec639f792891f748a1edbd759cf1dcea3382ad83c"
+
 [[package]]
 name = "cexpr"
 version = "0.6.0"
@@ -1894,7 +1923,7 @@ checksum = "b94f61472cee1439c0b966b47e3aca9ae07e45d070759512cd390ea2bebc6675"
 
 [[package]]
 name = "cli"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "async-stream",
  "async-trait",
@@ -1923,6 +1952,7 @@ dependencies = [
  "common-wal",
  "datatypes",
  "etcd-client",
+ "fs2",
  "futures",
  "humantime",
  "meta-client",
@@ -1933,7 +1963,7 @@ dependencies = [
  "paste",
  "query",
  "rand 0.9.1",
- "reqwest",
+ "reqwest 0.13.2",
  "serde",
  "serde_json",
  "servers",
@@ -1946,11 +1976,12 @@ dependencies = [
  "tokio",
  "tracing-appender",
  "url",
+ "uuid",
 ]
 
 [[package]]
 name = "client"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "arc-swap",
@@ -1982,7 +2013,7 @@ dependencies = [
  "serde_json",
  "snafu 0.8.6",
  "store-api",
- "substrait 1.0.0-rc.2",
+ "substrait 1.0.0",
  "tokio",
  "tokio-stream",
  "tonic 0.14.2",
@@ -2022,7 +2053,7 @@ dependencies = [
 
 [[package]]
 name = "cmd"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "async-trait",
@@ -2083,8 +2114,7 @@ dependencies = [
  "query",
  "rand 0.9.1",
  "regex",
- "reqwest",
- "rexpect",
+ "reqwest 0.13.2",
  "serde",
  "serde_json",
  "servers",
@@ -2139,7 +2169,17 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "117725a109d387c937a1533ce01b450cbde6b88abceea8473c4d7a85853cda3c"
 dependencies = [
  "lazy_static",
- "windows-sys 0.59.0",
+ "windows-sys 0.48.0",
+]
+
+[[package]]
+name = "combine"
+version = "4.6.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ba5a308b75df32fe02788e748662718f03fde005016435c444eea572398219fd"
+dependencies = [
+ "bytes",
+ "memchr",
 ]
 
 [[package]]
@@ -2153,32 +2193,21 @@ dependencies = [
  "unicode-width 0.2.1",
 ]
 
-[[package]]
-name = "comma"
-version = "1.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "55b672471b4e9f9e95499ea597ff64941a309b2cdbffcc46f2cc5e2d971fd335"
-
 [[package]]
 name = "common-base"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "ahash 0.8.12",
  "anymap2",
  "async-trait",
  "bitvec",
  "bytes",
- "common-error",
- "common-macro",
  "common-test-util",
  "futures",
  "lazy_static",
- "paste",
  "pin-project",
- "rand 0.9.1",
  "regex",
  "serde",
- "snafu 0.8.6",
  "tokio",
  "toml 0.8.23",
  "zeroize",
@@ -2186,14 +2215,14 @@ dependencies = [
 
 [[package]]
 name = "common-catalog"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "const_format",
 ]
 
 [[package]]
 name = "common-config"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "common-base",
  "common-error",
@@ -2209,7 +2238,6 @@ dependencies = [
  "object-store",
  "serde",
  "serde_json",
- "serde_with",
  "snafu 0.8.6",
  "temp-env",
  "tempfile",
@@ -2218,7 +2246,7 @@ dependencies = [
 
 [[package]]
 name = "common-datasource"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "arrow 57.3.0",
  "arrow-schema 57.3.0",
@@ -2239,7 +2267,6 @@ dependencies = [
  "futures",
  "lazy_static",
  "object-store",
- "object_store_opendal",
  "orc-rust",
  "parquet",
  "paste",
@@ -2254,7 +2281,7 @@ dependencies = [
 
 [[package]]
 name = "common-decimal"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "bigdecimal 0.4.8",
  "common-error",
@@ -2267,7 +2294,7 @@ dependencies = [
 
 [[package]]
 name = "common-error"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "common-macro",
  "http 1.3.1",
@@ -2278,7 +2305,7 @@ dependencies = [
 
 [[package]]
 name = "common-event-recorder"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "async-trait",
@@ -2301,7 +2328,7 @@ dependencies = [
 
 [[package]]
 name = "common-frontend"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "async-trait",
@@ -2322,7 +2349,7 @@ dependencies = [
 
 [[package]]
 name = "common-function"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "ahash 0.8.12",
  "api",
@@ -2360,8 +2387,8 @@ dependencies = [
  "geohash",
  "h3o",
  "hyperloglogplus",
+ "icu_properties",
  "jsonb",
- "jsonpath-rust 0.7.5",
  "memchr",
  "mito-codec",
  "nalgebra",
@@ -2385,7 +2412,7 @@ dependencies = [
 
 [[package]]
 name = "common-greptimedb-telemetry"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "async-trait",
  "common-runtime",
@@ -2393,7 +2420,7 @@ dependencies = [
  "common-test-util",
  "common-version",
  "hyper 0.14.32",
- "reqwest",
+ "reqwest 0.13.2",
  "serde",
  "tempfile",
  "tokio",
@@ -2402,7 +2429,7 @@ dependencies = [
 
 [[package]]
 name = "common-grpc"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "arrow-flight",
@@ -2437,7 +2464,7 @@ dependencies = [
 
 [[package]]
 name = "common-grpc-expr"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "common-base",
@@ -2457,7 +2484,7 @@ dependencies = [
 
 [[package]]
 name = "common-macro"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "greptime-proto",
  "once_cell",
@@ -2468,7 +2495,7 @@ dependencies = [
 
 [[package]]
 name = "common-mem-prof"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "anyhow",
  "common-error",
@@ -2484,11 +2511,10 @@ dependencies = [
 
 [[package]]
 name = "common-memory-manager"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "common-error",
  "common-macro",
- "common-telemetry",
  "humantime",
  "serde",
  "snafu 0.8.6",
@@ -2497,7 +2523,7 @@ dependencies = [
 
 [[package]]
 name = "common-meta"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "anymap2",
  "api",
@@ -2568,7 +2594,7 @@ dependencies = [
 
 [[package]]
 name = "common-options"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "common-grpc",
  "humantime-serde",
@@ -2578,11 +2604,11 @@ dependencies = [
 
 [[package]]
 name = "common-plugins"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 
 [[package]]
 name = "common-pprof"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "common-error",
  "common-macro",
@@ -2593,7 +2619,7 @@ dependencies = [
 
 [[package]]
 name = "common-procedure"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "async-stream",
@@ -2622,7 +2648,7 @@ dependencies = [
 
 [[package]]
 name = "common-procedure-test"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "async-trait",
  "common-procedure",
@@ -2632,7 +2658,7 @@ dependencies = [
 
 [[package]]
 name = "common-query"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "async-trait",
@@ -2658,12 +2684,13 @@ dependencies = [
 
 [[package]]
 name = "common-recordbatch"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "arc-swap",
  "common-base",
  "common-error",
  "common-macro",
+ "common-memory-manager",
  "common-telemetry",
  "common-time",
  "criterion 0.7.0",
@@ -2682,7 +2709,7 @@ dependencies = [
 
 [[package]]
 name = "common-runtime"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "async-trait",
  "clap",
@@ -2711,7 +2738,7 @@ dependencies = [
 
 [[package]]
 name = "common-session"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "serde",
  "strum 0.27.1",
@@ -2719,7 +2746,7 @@ dependencies = [
 
 [[package]]
 name = "common-sql"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "arrow-schema 57.3.0",
  "common-base",
@@ -2739,7 +2766,7 @@ dependencies = [
 
 [[package]]
 name = "common-stat"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "common-base",
  "common-runtime",
@@ -2754,7 +2781,7 @@ dependencies = [
 
 [[package]]
 name = "common-telemetry"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "backtrace",
  "common-base",
@@ -2767,7 +2794,7 @@ dependencies = [
  "once_cell",
  "opentelemetry 0.30.0",
  "opentelemetry-otlp",
- "opentelemetry-semantic-conventions",
+ "opentelemetry-semantic-conventions 0.30.0",
  "opentelemetry_sdk 0.30.0",
  "parking_lot 0.12.4",
  "prometheus 0.14.0",
@@ -2783,7 +2810,7 @@ dependencies = [
 
 [[package]]
 name = "common-test-util"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "client",
  "common-grpc",
@@ -2796,7 +2823,7 @@ dependencies = [
 
 [[package]]
 name = "common-time"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "arrow 57.3.0",
  "chrono",
@@ -2814,7 +2841,7 @@ dependencies = [
 
 [[package]]
 name = "common-version"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "cargo-manifest",
  "const_format",
@@ -2824,7 +2851,7 @@ dependencies = [
 
 [[package]]
 name = "common-wal"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "common-base",
  "common-error",
@@ -2847,7 +2874,7 @@ dependencies = [
 
 [[package]]
 name = "common-workload"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "common-telemetry",
  "serde",
@@ -3330,6 +3357,22 @@ dependencies = [
  "memchr",
 ]
 
+[[package]]
+name = "ctor"
+version = "0.6.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "424e0138278faeb2b401f174ad17e715c829512d74f3d1e81eb43365c2e0590e"
+dependencies = [
+ "ctor-proc-macro",
+ "dtor",
+]
+
+[[package]]
+name = "ctor-proc-macro"
+version = "0.0.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "52560adf09603e58c9a7ee1fe1dcb95a16927b17c127f0ac02d6e768a0e25bc1"
+
 [[package]]
 name = "ctr"
 version = "0.9.2"
@@ -3567,7 +3610,7 @@ dependencies = [
  "itertools 0.14.0",
  "liblzma",
  "log",
- "object_store",
+ "object_store 0.12.5",
  "parking_lot 0.12.4",
  "parquet",
  "rand 0.9.1",
@@ -3599,7 +3642,7 @@ dependencies = [
  "futures",
  "itertools 0.14.0",
  "log",
- "object_store",
+ "object_store 0.12.5",
  "parking_lot 0.12.4",
  "tokio",
 ]
@@ -3623,7 +3666,7 @@ dependencies = [
  "futures",
  "itertools 0.14.0",
  "log",
- "object_store",
+ "object_store 0.12.5",
 ]
 
 [[package]]
@@ -3641,7 +3684,7 @@ dependencies = [
  "itertools 0.14.0",
  "libc",
  "log",
- "object_store",
+ "object_store 0.12.5",
  "parquet",
  "paste",
  "recursive",
@@ -3686,7 +3729,7 @@ dependencies = [
  "itertools 0.14.0",
  "liblzma",
  "log",
- "object_store",
+ "object_store 0.12.5",
  "rand 0.9.1",
  "tokio",
  "tokio-util",
@@ -3713,7 +3756,7 @@ dependencies = [
  "datafusion-session",
  "futures",
  "itertools 0.14.0",
- "object_store",
+ "object_store 0.12.5",
  "tokio",
 ]
 
@@ -3734,7 +3777,7 @@ dependencies = [
  "datafusion-physical-plan",
  "datafusion-session",
  "futures",
- "object_store",
+ "object_store 0.12.5",
  "regex",
  "tokio",
 ]
@@ -3756,7 +3799,7 @@ dependencies = [
  "datafusion-physical-plan",
  "datafusion-session",
  "futures",
- "object_store",
+ "object_store 0.12.5",
  "serde_json",
  "tokio",
  "tokio-stream",
@@ -3785,7 +3828,7 @@ dependencies = [
  "futures",
  "itertools 0.14.0",
  "log",
- "object_store",
+ "object_store 0.12.5",
  "parking_lot 0.12.4",
  "parquet",
  "tokio",
@@ -3811,7 +3854,7 @@ dependencies = [
  "datafusion-physical-expr-common",
  "futures",
  "log",
- "object_store",
+ "object_store 0.12.5",
  "parking_lot 0.12.4",
  "rand 0.9.1",
  "tempfile",
@@ -4020,7 +4063,7 @@ dependencies = [
  "datafusion",
  "futures",
  "futures-util",
- "object_store",
+ "object_store 0.12.5",
  "orc-rust",
  "tokio",
 ]
@@ -4199,7 +4242,7 @@ dependencies = [
  "datafusion",
  "half",
  "itertools 0.14.0",
- "object_store",
+ "object_store 0.12.5",
  "pbjson-types",
  "prost 0.14.1",
  "substrait 0.62.2",
@@ -4209,7 +4252,7 @@ dependencies = [
 
 [[package]]
 name = "datanode"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "arrow-flight",
@@ -4260,7 +4303,7 @@ dependencies = [
  "prometheus 0.14.0",
  "prost 0.14.1",
  "query",
- "reqwest",
+ "reqwest 0.13.2",
  "serde",
  "serde_json",
  "servers",
@@ -4277,7 +4320,7 @@ dependencies = [
 
 [[package]]
 name = "datatypes"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "arrow 57.3.0",
  "arrow-array 57.3.0",
@@ -4672,6 +4715,27 @@ version = "2.0.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ea8a8b81cacc08888170eef4d13b775126db426d0b348bee9d18c2c1eaf123cf"
 
+[[package]]
+name = "dtor"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "404d02eeb088a82cfd873006cb713fe411306c7d182c344905e101fb1167d301"
+dependencies = [
+ "dtor-proc-macro",
+]
+
+[[package]]
+name = "dtor-proc-macro"
+version = "0.0.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f678cf4a922c215c63e0de95eb1ff08a958a81d47e485cf9da1e27bf6305cfa5"
+
+[[package]]
+name = "dunce"
+version = "1.0.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "92773504d58c093f6de2459af4af33faa518c13451eb8f2b5698ed3d36e7c813"
+
 [[package]]
 name = "duration-str"
 version = "0.11.3"
@@ -4955,7 +5019,7 @@ checksum = "37909eebbb50d72f9059c3b6d82c0463f2ff062c9e95845c43a6c9c0355411be"
 
 [[package]]
 name = "file-engine"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "async-trait",
@@ -4975,7 +5039,6 @@ dependencies = [
  "datatypes",
  "futures",
  "object-store",
- "object_store_opendal",
  "serde",
  "serde_json",
  "snafu 0.8.6",
@@ -5026,7 +5089,7 @@ version = "25.2.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1045398c1bfd89168b5fd3f1fc11f6e70b34f6f66300c87d44d3de849463abf1"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "rustc_version",
 ]
 
@@ -5087,7 +5150,7 @@ checksum = "8bf7cc16383c4b8d58b9905a8509f02926ce3058053c056376248d958c9df1e8"
 
 [[package]]
 name = "flow"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "arrow 57.3.0",
@@ -5156,7 +5219,7 @@ dependencies = [
  "sql",
  "store-api",
  "strum 0.27.1",
- "substrait 1.0.0-rc.2",
+ "substrait 1.0.0",
  "table",
  "tokio",
  "tonic 0.14.2",
@@ -5217,7 +5280,7 @@ checksum = "28dd6caf6059519a65843af8fe2a3ae298b14b80179855aeb4adc2c1934ee619"
 
 [[package]]
 name = "frontend"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "arc-swap",
@@ -5261,12 +5324,14 @@ dependencies = [
  "humantime",
  "humantime-serde",
  "hyper-util",
+ "itertools 0.14.0",
  "lazy_static",
  "log-query",
  "meta-client",
  "meta-srv",
  "num_cpus",
  "opentelemetry-proto 0.31.0",
+ "opentelemetry-semantic-conventions 0.31.0",
  "operator",
  "otel-arrow-rust",
  "partition",
@@ -5276,7 +5341,7 @@ dependencies = [
  "prost 0.14.1",
  "query",
  "rand 0.9.1",
- "reqwest",
+ "reqwest 0.13.2",
  "serde",
  "serde_json",
  "servers",
@@ -5372,6 +5437,12 @@ dependencies = [
  "windows-sys 0.52.0",
 ]
 
+[[package]]
+name = "fs_extra"
+version = "1.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "42703706b716c37f96a77aea830392ad231f44c9e9a67872fa5548707e11b11c"
+
 [[package]]
 name = "fsevent-sys"
 version = "4.1.0"
@@ -5664,7 +5735,7 @@ version = "0.20.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7b88256088d75a56f8ecfa070513a775dd9107f6530ef14919dac831af9cfe2b"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "libc",
  "libgit2-sys",
  "log",
@@ -5692,7 +5763,7 @@ dependencies = [
 [[package]]
 name = "greptime-proto"
 version = "0.1.0"
-source = "git+https://github.com/GreptimeTeam/greptime-proto.git?rev=092ba1d01e2da676dca66cca7eebb55009da8ef8#092ba1d01e2da676dca66cca7eebb55009da8ef8"
+source = "git+https://github.com/GreptimeTeam/greptime-proto.git?rev=0de5437582920c8b30d6c34212f161db71d95c50#0de5437582920c8b30d6c34212f161db71d95c50"
 dependencies = [
  "prost 0.14.1",
  "prost-types 0.14.1",
@@ -5702,7 +5773,6 @@ dependencies = [
  "strum_macros 0.25.3",
  "tonic 0.14.2",
  "tonic-prost",
- "tonic-prost-build",
 ]
 
 [[package]]
@@ -6464,7 +6534,7 @@ dependencies = [
 
 [[package]]
 name = "index"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "async-trait",
  "asynchronous-codec",
@@ -6618,7 +6688,7 @@ version = "0.11.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f37dccff2791ab604f9babef0ba14fbe0be30bd368dc541e2b08d07c8aa908f3"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "inotify-sys",
  "libc",
 ]
@@ -6797,6 +6867,93 @@ dependencies = [
  "regex",
 ]
 
+[[package]]
+name = "jiff"
+version = "0.2.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1a3546dc96b6d42c5f24902af9e2538e82e39ad350b0c766eb3fbf2d8f3d8359"
+dependencies = [
+ "jiff-static",
+ "jiff-tzdb-platform",
+ "js-sys",
+ "log",
+ "portable-atomic",
+ "portable-atomic-util",
+ "serde_core",
+ "wasm-bindgen",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "jiff-static"
+version = "0.2.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2a8c8b344124222efd714b73bb41f8b5120b27a7cc1c75593a6ff768d9d05aa4"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
+[[package]]
+name = "jiff-tzdb"
+version = "0.1.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c900ef84826f1338a557697dc8fc601df9ca9af4ac137c7fb61d4c6f2dfd3076"
+
+[[package]]
+name = "jiff-tzdb-platform"
+version = "0.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "875a5a69ac2bab1a891711cf5eccbec1ce0341ea805560dcd90b7a2e925132e8"
+dependencies = [
+ "jiff-tzdb",
+]
+
+[[package]]
+name = "jni"
+version = "0.21.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1a87aa2bb7d2af34197c04845522473242e1aa17c12f4935d5856491a7fb8c97"
+dependencies = [
+ "cesu8",
+ "cfg-if",
+ "combine",
+ "jni-sys 0.3.1",
+ "log",
+ "thiserror 1.0.69",
+ "walkdir",
+ "windows-sys 0.45.0",
+]
+
+[[package]]
+name = "jni-sys"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "41a652e1f9b6e0275df1f15b32661cf0d4b78d4d87ddec5e0c3c20f097433258"
+dependencies = [
+ "jni-sys 0.4.1",
+]
+
+[[package]]
+name = "jni-sys"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c6377a88cb3910bee9b0fa88d4f42e1d2da8e79915598f65fb0c7ee14c878af2"
+dependencies = [
+ "jni-sys-macros",
+]
+
+[[package]]
+name = "jni-sys-macros"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "38c0b942f458fe50cdac086d2f946512305e5631e720728f2a61aabcd47a6264"
+dependencies = [
+ "quote",
+ "syn 2.0.117",
+]
+
 [[package]]
 name = "jobserver"
 version = "0.1.33"
@@ -6809,10 +6966,12 @@ dependencies = [
 
 [[package]]
 name = "js-sys"
-version = "0.3.77"
+version = "0.3.95"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1cfaf33c695fc6e08064efbc1f72ec937429614f25eef83af942d0e227c3a28f"
+checksum = "2964e92d1d9dc3364cae4d718d93f227e3abb088e747d92e0395bfdedf1c12ca"
 dependencies = [
+ "cfg-if",
+ "futures-util",
  "once_cell",
  "wasm-bindgen",
 ]
@@ -6897,16 +7056,18 @@ dependencies = [
 
 [[package]]
 name = "jsonwebtoken"
-version = "9.3.1"
+version = "10.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5a87cc7a48537badeae96744432de36f4be2b4a34a05a5ef32e9dd8a1c169dde"
+checksum = "0529410abe238729a60b108898784df8984c87f6054c9c4fcacc47e4803c1ce1"
 dependencies = [
+ "aws-lc-rs",
  "base64 0.22.1",
+ "getrandom 0.2.16",
  "js-sys",
  "pem",
- "ring",
  "serde",
  "serde_json",
+ "signature",
  "simple_asn1",
 ]
 
@@ -7301,7 +7462,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "07033963ba89ebaf1584d767badaa2e8fcec21aedea6b8c0346d487d49c28667"
 dependencies = [
  "cfg-if",
- "windows-targets 0.52.6",
+ "windows-targets 0.48.5",
 ]
 
 [[package]]
@@ -7336,7 +7497,7 @@ version = "0.1.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1580801010e535496706ba011c15f8532df6b42297d2e471fec38ceadd8c0638"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "libc",
  "redox_syscall 0.5.13",
 ]
@@ -7432,7 +7593,7 @@ checksum = "5e5032e24019045c762d3c0f28f5b6b8bbf38563a65908389bf7978758920897"
 
 [[package]]
 name = "log-query"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "chrono",
  "common-error",
@@ -7444,7 +7605,7 @@ dependencies = [
 
 [[package]]
 name = "log-store"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "async-stream",
  "async-trait",
@@ -7598,18 +7759,18 @@ dependencies = [
 
 [[package]]
 name = "lz4_flex"
-version = "0.11.5"
+version = "0.11.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "08ab2867e3eeeca90e844d1940eab391c9dc5228783db2ed999acbc0a9ed375a"
+checksum = "373f5eceeeab7925e0c1098212f2fbc4d416adec9d35051a6ab251e824c1854a"
 dependencies = [
  "twox-hash",
 ]
 
 [[package]]
 name = "lz4_flex"
-version = "0.12.0"
+version = "0.12.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ab6473172471198271ff72e9379150e9dfd70d8e533e0752a27e515b48dd375e"
+checksum = "98c23545df7ecf1b16c303910a69b079e8e251d60f7dd2cc9b4177f2afaf1746"
 dependencies = [
  "twox-hash",
 ]
@@ -7690,6 +7851,15 @@ version = "0.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ae960838283323069879657ca3de837e9f7bbb4c7bf6ea7f1b290d5e9476d2e0"
 
+[[package]]
+name = "mea"
+version = "0.6.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6747f54621d156e1b47eb6b25f39a941b9fc347f98f67d25d8881ff99e8ed832"
+dependencies = [
+ "slab",
+]
+
 [[package]]
 name = "measure_time"
 version = "0.9.0"
@@ -7724,15 +7894,6 @@ dependencies = [
  "libc",
 ]
 
-[[package]]
-name = "memoffset"
-version = "0.6.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5aa361d4faea93603064a027415f07bd8e1d5c88c9fbf68bf56a285428fd79ce"
-dependencies = [
- "autocfg",
-]
-
 [[package]]
 name = "memoffset"
 version = "0.7.1"
@@ -7744,7 +7905,7 @@ dependencies = [
 
 [[package]]
 name = "meta-client"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "async-trait",
@@ -7755,6 +7916,7 @@ dependencies = [
  "common-meta",
  "common-options",
  "common-telemetry",
+ "common-time",
  "datatypes",
  "futures",
  "futures-util",
@@ -7775,7 +7937,7 @@ dependencies = [
 
 [[package]]
 name = "meta-srv"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "async-trait",
@@ -7846,7 +8008,7 @@ dependencies = [
  "toml 0.8.23",
  "tonic 0.14.2",
  "tower 0.5.2",
- "tower-http 0.6.6",
+ "tower-http 0.6.8",
  "tracing",
  "tracing-subscriber",
  "typetag",
@@ -7875,7 +8037,7 @@ dependencies = [
 
 [[package]]
 name = "metric-engine"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "aquamarine",
@@ -7887,6 +8049,7 @@ dependencies = [
  "common-base",
  "common-error",
  "common-function",
+ "common-grpc",
  "common-macro",
  "common-meta",
  "common-query",
@@ -7975,7 +8138,7 @@ dependencies = [
 
 [[package]]
 name = "mito-codec"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "bytes",
@@ -8000,13 +8163,15 @@ dependencies = [
 
 [[package]]
 name = "mito2"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "aquamarine",
+ "arrow-schema 57.3.0",
  "async-channel 1.9.0",
  "async-stream",
  "async-trait",
+ "base64 0.22.1",
  "bytemuck",
  "bytes",
  "chrono",
@@ -8268,7 +8433,7 @@ dependencies = [
  "base64 0.21.7",
  "bigdecimal 0.4.8",
  "bindgen",
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "bitvec",
  "btoi",
  "byteorder",
@@ -8306,7 +8471,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "34a9141e735d5bb02414a7ac03add09522466d4db65bdd827069f76ae0850e58"
 dependencies = [
  "base64 0.22.1",
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "btoi",
  "byteorder",
  "bytes",
@@ -8399,20 +8564,6 @@ version = "1.0.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "650eef8c711430f1a879fdd01d4745a7deea475becfb90269c06775983bbf086"
 
-[[package]]
-name = "nix"
-version = "0.25.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f346ff70e7dbfd675fe90590b92d59ef2de15a8779ae305ebcbfd3f0caf59be4"
-dependencies = [
- "autocfg",
- "bitflags 1.3.2",
- "cfg-if",
- "libc",
- "memoffset 0.6.5",
- "pin-utils",
-]
-
 [[package]]
 name = "nix"
 version = "0.26.4"
@@ -8422,7 +8573,7 @@ dependencies = [
  "bitflags 1.3.2",
  "cfg-if",
  "libc",
- "memoffset 0.7.1",
+ "memoffset",
  "pin-utils",
 ]
 
@@ -8432,7 +8583,7 @@ version = "0.28.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ab2156c4fce2f8df6c499cc1c763e4394b7482525bf2a9701c9d79d215f519e4"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "cfg-if",
  "cfg_aliases 0.1.1",
  "libc",
@@ -8444,7 +8595,7 @@ version = "0.30.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "74523f3a35e05aba87a1d978330aef40f67b0304ac79c1c00b294c9830543db6"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "cfg-if",
  "cfg_aliases 0.2.1",
  "libc",
@@ -8481,7 +8632,7 @@ version = "8.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2fee8403b3d66ac7b26aee6e40a897d85dc5ce26f44da36b8b73e987cc52e943"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "filetime",
  "fsevent-sys",
  "inotify",
@@ -8738,10 +8889,12 @@ dependencies = [
 
 [[package]]
 name = "object-store"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "anyhow",
+ "async-trait",
  "bytes",
+ "chrono",
  "common-base",
  "common-error",
  "common-macro",
@@ -8751,12 +8904,16 @@ dependencies = [
  "futures",
  "humantime-serde",
  "lazy_static",
- "moka",
+ "object_store 0.12.5",
+ "object_store 0.13.2",
+ "object_store_opendal",
  "opendal",
  "prometheus 0.14.0",
- "reqwest",
+ "rand 0.9.1",
+ "reqwest 0.13.2",
  "serde",
  "snafu 0.8.6",
+ "tempfile",
  "tokio",
  "uuid",
 ]
@@ -8786,15 +8943,43 @@ dependencies = [
 ]
 
 [[package]]
-name = "object_store_opendal"
-version = "0.54.0"
+name = "object_store"
+version = "0.13.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5ce697ee723fdc3eaf6c457abf4059034be15167022b18b619993802cd1443d5"
+checksum = "622acbc9100d3c10e2ee15804b0caa40e55c933d5aa53814cd520805b7958a49"
 dependencies = [
  "async-trait",
  "bytes",
+ "chrono",
+ "futures-channel",
+ "futures-core",
+ "futures-util",
+ "http 1.3.1",
+ "humantime",
+ "itertools 0.14.0",
+ "parking_lot 0.12.4",
+ "percent-encoding",
+ "thiserror 2.0.17",
+ "tokio",
+ "tracing",
+ "url",
+ "walkdir",
+ "wasm-bindgen-futures",
+ "web-time",
+]
+
+[[package]]
+name = "object_store_opendal"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "08298874eee5935c95bcaa393148834f9c53d904461ca15584a041d8a1c907c2"
+dependencies = [
+ "async-trait",
+ "bytes",
+ "chrono",
  "futures",
- "object_store",
+ "mea",
+ "object_store 0.13.2",
  "opendal",
  "pin-project",
  "tokio",
@@ -8844,7 +9029,7 @@ version = "6.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "336b9c63443aceef14bea841b899035ae3abe89b7c486aaf4c5bd8aafedac3f0"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "libc",
  "once_cell",
  "onig_sys",
@@ -8874,33 +9059,245 @@ checksum = "c08d65885ee38876c4f86fa503fb49d7b507c2b62552df7c70b2fce627e06381"
 
 [[package]]
 name = "opendal"
-version = "0.54.1"
+version = "0.56.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "42afda58fa2cf50914402d132cc1caacff116a85d10c72ab2082bb7c50021754"
+checksum = "97b31d3d8e99a85d83b73ec26647f5607b80578ed9375810b6e44ffa3590a236"
+dependencies = [
+ "ctor",
+ "opendal-core",
+ "opendal-layer-concurrent-limit",
+ "opendal-layer-logging",
+ "opendal-layer-prometheus",
+ "opendal-layer-retry",
+ "opendal-layer-timeout",
+ "opendal-layer-tracing",
+ "opendal-service-azblob",
+ "opendal-service-fs",
+ "opendal-service-gcs",
+ "opendal-service-http",
+ "opendal-service-oss",
+ "opendal-service-s3",
+]
+
+[[package]]
+name = "opendal-core"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1849dd2687e173e776d3af5fce1ba3ae47b9dd37a09d1c4deba850ef45fe00ca"
 dependencies = [
  "anyhow",
- "backon",
  "base64 0.22.1",
  "bytes",
- "chrono",
- "crc32c",
  "futures",
- "getrandom 0.2.16",
  "http 1.3.1",
  "http-body 1.0.1",
+ "jiff",
  "log",
  "md-5",
+ "mea",
  "percent-encoding",
- "prometheus 0.14.0",
  "quick-xml 0.38.4",
- "reqsign",
- "reqwest",
+ "reqsign-core",
+ "reqwest 0.13.2",
  "serde",
  "serde_json",
- "sha2",
  "tokio",
- "tracing",
+ "url",
  "uuid",
+ "web-time",
+]
+
+[[package]]
+name = "opendal-layer-concurrent-limit"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "048b1b29c503263bdd80a9afe46a68cd02ea9bd361185b1feab4b151078998e9"
+dependencies = [
+ "futures",
+ "http 1.3.1",
+ "mea",
+ "opendal-core",
+]
+
+[[package]]
+name = "opendal-layer-logging"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d2645adc988b12eda106e2679ae529facfbbaa868ceb706f6f8125c6af15c47b"
+dependencies = [
+ "log",
+ "opendal-core",
+]
+
+[[package]]
+name = "opendal-layer-observe-metrics-common"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9130f0ac11569edc0f70b0e64078b9a12e37a128849d27ea62b0ca7568e8eb97"
+dependencies = [
+ "futures",
+ "http 1.3.1",
+ "opendal-core",
+]
+
+[[package]]
+name = "opendal-layer-prometheus"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "eef98056f8198b5053005e1fbe7d163562f56d9f3be1b73a3792667e7cbaf7be"
+dependencies = [
+ "opendal-core",
+ "opendal-layer-observe-metrics-common",
+ "prometheus 0.14.0",
+]
+
+[[package]]
+name = "opendal-layer-retry"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4eac134ffa4ddda6131a640a84a5315996424b9416c85052f8c64c1a33b70ad4"
+dependencies = [
+ "backon",
+ "log",
+ "opendal-core",
+]
+
+[[package]]
+name = "opendal-layer-timeout"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "619586ab7480c2e3009f6d18eabab18957bc094778fd130bcc38924970a90f4c"
+dependencies = [
+ "opendal-core",
+ "tokio",
+]
+
+[[package]]
+name = "opendal-layer-tracing"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f569b1cfae56851662e0db0e2d8df2f417d40aa281536c13f0c5211e7b56cd6f"
+dependencies = [
+ "futures",
+ "http 1.3.1",
+ "opendal-core",
+ "tracing",
+]
+
+[[package]]
+name = "opendal-service-azblob"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7452bf3ec61cfd81ac9ad9ada17825931e9e371d44a045c6bfab9596c0a2ac3b"
+dependencies = [
+ "base64 0.22.1",
+ "bytes",
+ "http 1.3.1",
+ "log",
+ "opendal-core",
+ "opendal-service-azure-common",
+ "quick-xml 0.38.4",
+ "reqsign-azure-storage",
+ "reqsign-core",
+ "reqsign-file-read-tokio",
+ "serde",
+ "sha2",
+ "uuid",
+]
+
+[[package]]
+name = "opendal-service-azure-common"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ffb0e45d6c8dcf66ce2da20e241bcb80e6e540e109a4ff20f318f6c9b4c54e0c"
+dependencies = [
+ "http 1.3.1",
+ "opendal-core",
+]
+
+[[package]]
+name = "opendal-service-fs"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cf0be0417abeeb0053376d816b90fceb9ca98f20dfb54ebf1f2a282729f83663"
+dependencies = [
+ "bytes",
+ "log",
+ "opendal-core",
+ "serde",
+ "tokio",
+ "xattr",
+]
+
+[[package]]
+name = "opendal-service-gcs"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "70a49477a10163431896d106136117f5670717f9c9e49cf6f710528800c6633a"
+dependencies = [
+ "async-trait",
+ "bytes",
+ "http 1.3.1",
+ "log",
+ "opendal-core",
+ "percent-encoding",
+ "quick-xml 0.38.4",
+ "reqsign-core",
+ "reqsign-file-read-tokio",
+ "reqsign-google",
+ "serde",
+ "serde_json",
+ "tokio",
+]
+
+[[package]]
+name = "opendal-service-http"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6fe73e6978feec293acfb92bfa94bdb9cf1b5be3f7c3f93a4333a25455826005"
+dependencies = [
+ "http 1.3.1",
+ "log",
+ "opendal-core",
+ "serde",
+]
+
+[[package]]
+name = "opendal-service-oss"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "29c8a917829ad06d21b639558532cb0101fe49b040d946d673a73018683fac05"
+dependencies = [
+ "bytes",
+ "http 1.3.1",
+ "log",
+ "opendal-core",
+ "quick-xml 0.38.4",
+ "reqsign-aliyun-oss",
+ "reqsign-core",
+ "reqsign-file-read-tokio",
+ "serde",
+]
+
+[[package]]
+name = "opendal-service-s3"
+version = "0.56.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9dadddeb9bb50b0d30927dd914c298c4ddca47e4c1cfa7674d311f0cf9b051c8"
+dependencies = [
+ "base64 0.22.1",
+ "bytes",
+ "crc32c",
+ "http 1.3.1",
+ "log",
+ "md-5",
+ "opendal-core",
+ "quick-xml 0.38.4",
+ "reqsign-aws-v4",
+ "reqsign-core",
+ "reqsign-file-read-tokio",
+ "serde",
+ "url",
 ]
 
 [[package]]
@@ -8917,7 +9314,7 @@ dependencies = [
 [[package]]
 name = "opensrv-mysql"
 version = "0.8.0"
-source = "git+https://github.com/datafuselabs/opensrv?tag=v0.10.0#074bd8fb81da3c9e6d6a098a482f3380478b9c0b"
+source = "git+https://github.com/GreptimeTeam/opensrv?rev=6c5a451544194b7bb60a8318d155d4f892b49f2c#6c5a451544194b7bb60a8318d155d4f892b49f2c"
 dependencies = [
  "async-trait",
  "byteorder",
@@ -8974,7 +9371,7 @@ dependencies = [
  "bytes",
  "http 1.3.1",
  "opentelemetry 0.30.0",
- "reqwest",
+ "reqwest 0.12.28",
 ]
 
 [[package]]
@@ -8989,7 +9386,7 @@ dependencies = [
  "opentelemetry-proto 0.30.0",
  "opentelemetry_sdk 0.30.0",
  "prost 0.13.5",
- "reqwest",
+ "reqwest 0.12.28",
  "thiserror 2.0.17",
  "tokio",
  "tonic 0.13.1",
@@ -9031,6 +9428,12 @@ version = "0.30.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "83d059a296a47436748557a353c5e6c5705b9470ef6c95cfc52c21a8814ddac2"
 
+[[package]]
+name = "opentelemetry-semantic-conventions"
+version = "0.31.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e62e29dfe041afb8ed2a6c9737ab57db4907285d999ef8ad3a59092a36bdc846"
+
 [[package]]
 name = "opentelemetry_sdk"
 version = "0.30.0"
@@ -9066,7 +9469,7 @@ dependencies = [
 
 [[package]]
 name = "operator"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "ahash 0.8.12",
  "api",
@@ -9112,7 +9515,6 @@ dependencies = [
  "meter-macros",
  "moka",
  "object-store",
- "object_store_opendal",
  "partition",
  "path-slash",
  "prometheus 0.14.0",
@@ -9126,7 +9528,7 @@ dependencies = [
  "sql",
  "sqlparser",
  "store-api",
- "substrait 1.0.0-rc.2",
+ "substrait 1.0.0",
  "table",
  "tokio",
  "tokio-util",
@@ -9150,7 +9552,7 @@ dependencies = [
  "flate2",
  "futures",
  "futures-util",
- "lz4_flex 0.11.5",
+ "lz4_flex 0.11.6",
  "lzokay-native",
  "num",
  "prost 0.13.5",
@@ -9373,11 +9775,11 @@ dependencies = [
  "futures",
  "half",
  "hashbrown 0.16.1",
- "lz4_flex 0.12.0",
+ "lz4_flex 0.12.1",
  "num-bigint",
  "num-integer",
  "num-traits",
- "object_store",
+ "object_store 0.12.5",
  "paste",
  "seq-macro",
  "simdutf8",
@@ -9402,7 +9804,7 @@ checksum = "e3c406c9e2aa74554e662d2c2ee11cd3e73756988800be7e6f5eddb16fed4699"
 
 [[package]]
 name = "partition"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "async-trait",
@@ -9411,6 +9813,7 @@ dependencies = [
  "common-macro",
  "common-meta",
  "common-query",
+ "common-telemetry",
  "criterion 0.7.0",
  "datafusion-common",
  "datafusion-expr",
@@ -9619,11 +10022,12 @@ dependencies = [
 
 [[package]]
 name = "pgwire"
-version = "0.38.0"
+version = "0.38.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "89d5e5a60d3f6e40c91f6a2a7f8d09665e636272bd5611977253559b6651aabb"
+checksum = "24bd4e6b1bfddc5c6420dee6602ec80946700b4c31ddcb64ee190ad6d979c210"
 dependencies = [
  "async-trait",
+ "aws-lc-rs",
  "base64 0.22.1",
  "bytes",
  "chrono",
@@ -9635,7 +10039,6 @@ dependencies = [
  "pg_interval_2",
  "postgres-types",
  "rand 0.10.0",
- "ring",
  "rust_decimal",
  "rustls-pki-types",
  "ryu",
@@ -9757,7 +10160,7 @@ checksum = "8b870d8c151b6f2fb93e84a13146138f05d02ed11c7e7c54f8826aaaf7c9f184"
 
 [[package]]
 name = "pipeline"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "ahash 0.8.12",
  "api",
@@ -9914,7 +10317,7 @@ dependencies = [
 
 [[package]]
 name = "plugins"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "auth",
  "catalog",
@@ -9952,6 +10355,15 @@ version = "1.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f84267b20a16ea918e43c6a88433c2d54fa145c92a811b5b047ccbe153674483"
 
+[[package]]
+name = "portable-atomic-util"
+version = "0.2.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c2a106d1259c23fac8e543272398ae0e3c0b8d33c88ed73d0cc71b0f1d902618"
+dependencies = [
+ "portable-atomic",
+]
+
 [[package]]
 name = "postgres-protocol"
 version = "0.6.8"
@@ -10170,7 +10582,7 @@ version = "0.17.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "cc5b72d8145275d844d4b5f6d4e1eef00c8cd889edb6035c21675d1bb1f45c9f"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "hex",
  "procfs-core",
  "rustix 0.38.44",
@@ -10182,7 +10594,7 @@ version = "0.17.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "239df02d8349b06fc07398a3a1697b06418223b1c7725085e801e7c0fc6a12ec"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "hex",
 ]
 
@@ -10232,7 +10644,7 @@ dependencies = [
 
 [[package]]
 name = "promql"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "ahash 0.8.12",
  "async-trait",
@@ -10241,6 +10653,7 @@ dependencies = [
  "common-macro",
  "common-recordbatch",
  "common-telemetry",
+ "criterion 0.7.0",
  "datafusion",
  "datafusion-common",
  "datafusion-expr",
@@ -10276,7 +10689,7 @@ version = "1.9.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bee689443a2bd0a16ab0348b52ee43e3b2d1b1f931c8aa5c9f8de4c86fbe8c40"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "num-traits",
  "rand 0.9.1",
  "rand_chacha 0.9.0",
@@ -10332,7 +10745,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "22505a5c94da8e3b7c2996394d1c933236c4d743e81a410bcca4e6989fc066a4"
 dependencies = [
  "bytes",
- "heck 0.5.0",
+ "heck 0.4.1",
  "itertools 0.12.1",
  "log",
  "multimap",
@@ -10352,7 +10765,7 @@ version = "0.14.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ac6c3320f9abac597dcbc668774ef006702672474aad53c6d596b62e487b40b1"
 dependencies = [
- "heck 0.5.0",
+ "heck 0.4.1",
  "itertools 0.14.0",
  "log",
  "multimap",
@@ -10583,14 +10996,14 @@ dependencies = [
 
 [[package]]
 name = "puffin"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "async-compression",
  "async-trait",
  "async-walkdir",
  "auto_impl",
  "base64 0.22.1",
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "bytes",
  "common-base",
  "common-error",
@@ -10600,7 +11013,7 @@ dependencies = [
  "common-test-util",
  "derive_builder 0.20.2",
  "futures",
- "lz4_flex 0.11.5",
+ "lz4_flex 0.11.6",
  "moka",
  "pin-project",
  "prometheus 0.14.0",
@@ -10619,7 +11032,7 @@ version = "0.13.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1e8bbe1a966bd2f362681a44f6edce3c2310ac21e4d5067a6e7ec396297a6ea0"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "memchr",
  "unicase",
 ]
@@ -10645,7 +11058,7 @@ dependencies = [
 
 [[package]]
 name = "query"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "ahash 0.8.12",
  "api",
@@ -10675,11 +11088,13 @@ dependencies = [
  "datafusion",
  "datafusion-common",
  "datafusion-expr",
+ "datafusion-expr-common",
  "datafusion-functions",
  "datafusion-optimizer",
  "datafusion-physical-expr",
  "datafusion-sql",
  "datatypes",
+ "either",
  "fastrand",
  "futures",
  "futures-util",
@@ -10712,7 +11127,7 @@ dependencies = [
  "sql",
  "sqlparser",
  "store-api",
- "substrait 1.0.0-rc.2",
+ "substrait 1.0.0",
  "table",
  "tokio",
  "tokio-stream",
@@ -10736,7 +11151,6 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "331e97a1af0bf59823e6eadffe373d7b27f485be8748f71471c662c1f269b7fb"
 dependencies = [
  "memchr",
- "serde",
 ]
 
 [[package]]
@@ -10749,6 +11163,16 @@ dependencies = [
  "serde",
 ]
 
+[[package]]
+name = "quick-xml"
+version = "0.39.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "958f21e8e7ceb5a1aa7fa87fab28e7c75976e0bfe7e23ff069e0a260f894067d"
+dependencies = [
+ "memchr",
+ "serde",
+]
+
 [[package]]
 name = "quinn"
 version = "0.11.8"
@@ -10771,10 +11195,11 @@ dependencies = [
 
 [[package]]
 name = "quinn-proto"
-version = "0.11.12"
+version = "0.11.14"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "49df843a9161c85bb8aae55f101bc0bac8bcafd637a620d9122fd7e0b2f7422e"
+checksum = "434b42fec591c96ef50e21e886936e66d3cc3f737104fdb9b737c40ffb94c098"
 dependencies = [
+ "aws-lc-rs",
  "bytes",
  "getrandom 0.3.3",
  "lru-slab",
@@ -11039,7 +11464,7 @@ version = "0.5.13"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0d04b7d0ee6b4a0207a0a7adb104d23ecb0b47d6beae7152d0fa34b692b29fd6"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
 ]
 
 [[package]]
@@ -11143,42 +11568,123 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "51743d3e274e2b18df81c4dc6caf8a5b8e15dbe799e0dca05c7617380094e884"
 
 [[package]]
-name = "reqsign"
-version = "0.16.5"
+name = "reqsign-aliyun-oss"
+version = "3.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "43451dbf3590a7590684c25fb8d12ecdcc90ed3ac123433e500447c7d77ed701"
+checksum = "57ac2757f3140aa2e213b554148ae0b52733e624fc6723f0cc6bb3d440176c95"
 dependencies = [
  "anyhow",
- "async-trait",
- "base64 0.22.1",
- "chrono",
  "form_urlencoded",
- "getrandom 0.2.16",
- "hex",
- "hmac",
- "home",
  "http 1.3.1",
- "jsonwebtoken",
  "log",
- "once_cell",
  "percent-encoding",
- "quick-xml 0.37.5",
- "rand 0.8.5",
- "reqwest",
- "rsa",
+ "reqsign-core",
  "rust-ini 0.21.1",
  "serde",
  "serde_json",
+]
+
+[[package]]
+name = "reqsign-aws-v4"
+version = "3.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "44eaca382e94505a49f1a4849658d153aebf79d9c1a58e5dd3b10361511e9f43"
+dependencies = [
+ "anyhow",
+ "bytes",
+ "form_urlencoded",
+ "http 1.3.1",
+ "log",
+ "percent-encoding",
+ "quick-xml 0.39.2",
+ "reqsign-core",
+ "rust-ini 0.21.1",
+ "serde",
+ "serde_json",
+ "serde_urlencoded",
  "sha1",
+]
+
+[[package]]
+name = "reqsign-azure-storage"
+version = "3.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7a321980405d596bd34aaf95c4722a3de4128a67fd19e74a81a83aa3fdf082e6"
+dependencies = [
+ "anyhow",
+ "base64 0.22.1",
+ "bytes",
+ "form_urlencoded",
+ "http 1.3.1",
+ "jsonwebtoken",
+ "log",
+ "pem",
+ "percent-encoding",
+ "reqsign-core",
+ "rsa",
+ "serde",
+ "serde_json",
+ "sha1",
+]
+
+[[package]]
+name = "reqsign-core"
+version = "3.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b10302cf0a7d7e7352ba211fc92c3c5bebf1286153e49cc5aa87348078a8e102"
+dependencies = [
+ "anyhow",
+ "base64 0.22.1",
+ "bytes",
+ "form_urlencoded",
+ "futures",
+ "hex",
+ "hmac",
+ "http 1.3.1",
+ "jiff",
+ "log",
+ "percent-encoding",
+ "sha1",
+ "sha2",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "reqsign-file-read-tokio"
+version = "3.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e2d89295b3d17abea31851cc8de55d843d89c52132c864963c38d41920613dc5"
+dependencies = [
+ "anyhow",
+ "reqsign-core",
+ "tokio",
+]
+
+[[package]]
+name = "reqsign-google"
+version = "3.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "35cc609b49c69e76ecaceb775a03f792d1ed3e7755ab3548d4534fd801e3242e"
+dependencies = [
+ "form_urlencoded",
+ "http 1.3.1",
+ "jsonwebtoken",
+ "log",
+ "percent-encoding",
+ "reqsign-aws-v4",
+ "reqsign-core",
+ "rsa",
+ "serde",
+ "serde_json",
  "sha2",
  "tokio",
 ]
 
 [[package]]
 name = "reqwest"
-version = "0.12.24"
+version = "0.12.28"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9d0946410b9f7b082a427e4ef5c8ff541a88b357bc6c637c40db3a68ac70a36f"
+checksum = "eddd3ca559203180a307f12d114c268abf583f59b03cb906fd0b3ff8646c1147"
 dependencies = [
  "base64 0.22.1",
  "bytes",
@@ -11193,13 +11699,54 @@ dependencies = [
  "hyper-util",
  "js-sys",
  "log",
+ "percent-encoding",
+ "pin-project-lite",
+ "quinn",
+ "rustls",
+ "rustls-pki-types",
+ "serde",
+ "serde_json",
+ "serde_urlencoded",
+ "sync_wrapper 1.0.2",
+ "tokio",
+ "tokio-rustls",
+ "tokio-util",
+ "tower 0.5.2",
+ "tower-http 0.6.8",
+ "tower-service",
+ "url",
+ "wasm-bindgen",
+ "wasm-bindgen-futures",
+ "wasm-streams 0.4.2",
+ "web-sys",
+ "webpki-roots 1.0.1",
+]
+
+[[package]]
+name = "reqwest"
+version = "0.13.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ab3f43e3283ab1488b624b44b0e988d0acea0b3214e694730a055cb6b2efa801"
+dependencies = [
+ "base64 0.22.1",
+ "bytes",
+ "futures-core",
+ "futures-util",
+ "http 1.3.1",
+ "http-body 1.0.1",
+ "http-body-util",
+ "hyper 1.6.0",
+ "hyper-rustls",
+ "hyper-util",
+ "js-sys",
+ "log",
  "mime_guess",
  "percent-encoding",
  "pin-project-lite",
  "quinn",
  "rustls",
- "rustls-native-certs 0.8.1",
  "rustls-pki-types",
+ "rustls-platform-verifier",
  "serde",
  "serde_json",
  "serde_urlencoded",
@@ -11208,27 +11755,13 @@ dependencies = [
  "tokio-rustls",
  "tokio-util",
  "tower 0.5.2",
- "tower-http 0.6.6",
+ "tower-http 0.6.8",
  "tower-service",
  "url",
  "wasm-bindgen",
  "wasm-bindgen-futures",
- "wasm-streams",
+ "wasm-streams 0.5.0",
  "web-sys",
- "webpki-roots 1.0.1",
-]
-
-[[package]]
-name = "rexpect"
-version = "0.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "01ff60778f96fb5a48adbe421d21bf6578ed58c0872d712e7e08593c195adff8"
-dependencies = [
- "comma",
- "nix 0.25.1",
- "regex",
- "tempfile",
- "thiserror 1.0.69",
 ]
 
 [[package]]
@@ -11250,7 +11783,7 @@ dependencies = [
  "cfg-if",
  "getrandom 0.2.16",
  "libc",
- "untrusted",
+ "untrusted 0.9.0",
  "windows-sys 0.52.0",
 ]
 
@@ -11345,9 +11878,9 @@ dependencies = [
 
 [[package]]
 name = "rsasl"
-version = "2.2.0"
+version = "2.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c8b534a23662bb559c5c73213be63ecd6524e774d291f3618c2b04b723d184eb"
+checksum = "9f1bcb95b531681a622f3d6972eaab523e17e2aad6d6209f0276628eb1cb5038"
 dependencies = [
  "base64 0.22.1",
  "core2",
@@ -11359,7 +11892,7 @@ dependencies = [
  "serde_json",
  "sha2",
  "stringprep",
- "thiserror 1.0.69",
+ "thiserror 2.0.17",
 ]
 
 [[package]]
@@ -11553,7 +12086,7 @@ version = "0.38.44"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "fdb5bc1ae2baa591800df16c9ca78619bf65c0488b41b96ccec5d11220d8c154"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "errno",
  "libc",
  "linux-raw-sys 0.4.15",
@@ -11566,7 +12099,7 @@ version = "1.0.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c71e83d6afe7ff64890ec6b71d6a69bb8a610ab78ce364b3352876bb4c801266"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "errno",
  "libc",
  "linux-raw-sys 0.9.4",
@@ -11579,6 +12112,7 @@ version = "0.23.28"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7160e3e10bf4535308537f3c4e1641468cd0e485175d6163087c0393c7d46643"
 dependencies = [
+ "aws-lc-rs",
  "log",
  "once_cell",
  "ring",
@@ -11610,7 +12144,7 @@ dependencies = [
  "openssl-probe",
  "rustls-pki-types",
  "schannel",
- "security-framework 3.2.0",
+ "security-framework 3.7.0",
 ]
 
 [[package]]
@@ -11633,14 +12167,42 @@ dependencies = [
 ]
 
 [[package]]
-name = "rustls-webpki"
-version = "0.103.3"
+name = "rustls-platform-verifier"
+version = "0.6.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e4a72fe2bcf7a6ac6fd7d0b9e5cb68aeb7d4c0a0271730218b3e92d43b4eb435"
+checksum = "1d99feebc72bae7ab76ba994bb5e121b8d83d910ca40b36e0921f53becc41784"
 dependencies = [
+ "core-foundation 0.10.1",
+ "core-foundation-sys",
+ "jni",
+ "log",
+ "once_cell",
+ "rustls",
+ "rustls-native-certs 0.8.1",
+ "rustls-platform-verifier-android",
+ "rustls-webpki",
+ "security-framework 3.7.0",
+ "security-framework-sys",
+ "webpki-root-certs",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "rustls-platform-verifier-android"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f87165f0995f63a9fbeea62b64d10b4d9d8e78ec6d7d51fb2125fda7bb36788f"
+
+[[package]]
+name = "rustls-webpki"
+version = "0.103.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "61c429a8649f110dddef65e2a5ad240f747e85f7758a6bccc7e5777bd33f756e"
+dependencies = [
+ "aws-lc-rs",
  "ring",
  "rustls-pki-types",
- "untrusted",
+ "untrusted 0.9.0",
 ]
 
 [[package]]
@@ -11822,7 +12384,7 @@ version = "2.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "897b2245f0b511c87893af39b033e5ca9cce68824c4d7e7630b5a1d339658d02"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "core-foundation 0.9.4",
  "core-foundation-sys",
  "libc",
@@ -11831,11 +12393,11 @@ dependencies = [
 
 [[package]]
 name = "security-framework"
-version = "3.2.0"
+version = "3.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "271720403f46ca04f7ba6f55d438f8bd878d6b8ca0a1046e8228c4145bcbb316"
+checksum = "b7f4bc775c73d9a02cde8bf7b2ec4c9d12743edf609006c7facc23998404cd1d"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "core-foundation 0.10.1",
  "core-foundation-sys",
  "libc",
@@ -11844,9 +12406,9 @@ dependencies = [
 
 [[package]]
 name = "security-framework-sys"
-version = "2.14.0"
+version = "2.17.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "49db231d56a190491cb4aeda9527f1ad45345af50b0851622a7adb8c03b01c32"
+checksum = "6ce2691df843ecc5d231c0b14ece2acc3efb62c0a398c7e1d875f3983ce020e3"
 dependencies = [
  "core-foundation-sys",
  "libc",
@@ -12029,7 +12591,7 @@ dependencies = [
 
 [[package]]
 name = "servers"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "ahash 0.8.12",
  "api",
@@ -12079,6 +12641,7 @@ dependencies = [
  "datafusion-pg-catalog",
  "datatypes",
  "derive_builder 0.20.2",
+ "either",
  "futures",
  "futures-util",
  "headers",
@@ -12096,6 +12659,7 @@ dependencies = [
  "local-ip-address",
  "log-query",
  "loki-proto",
+ "metric-engine",
  "mime_guess",
  "mysql_async",
  "notify",
@@ -12107,6 +12671,7 @@ dependencies = [
  "operator",
  "otel-arrow-rust",
  "parking_lot 0.12.4",
+ "partition",
  "permutation",
  "pg_interval_2",
  "pgwire",
@@ -12121,8 +12686,9 @@ dependencies = [
  "quoted-string",
  "rand 0.9.1",
  "regex",
- "reqwest",
+ "reqwest 0.13.2",
  "rust-embed",
+ "rust_decimal",
  "rustls",
  "rustls-pemfile",
  "rustls-pki-types",
@@ -12131,6 +12697,7 @@ dependencies = [
  "session",
  "simd-json",
  "simdutf8",
+ "smallvec",
  "snafu 0.8.6",
  "snap",
  "socket2 0.5.10",
@@ -12148,8 +12715,9 @@ dependencies = [
  "tokio-util",
  "tonic 0.14.2",
  "tonic-reflection",
+ "tonic-web",
  "tower 0.5.2",
- "tower-http 0.6.6",
+ "tower-http 0.6.8",
  "tracing",
  "tracing-opentelemetry",
  "urlencoding",
@@ -12160,7 +12728,7 @@ dependencies = [
 
 [[package]]
 name = "session"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "ahash 0.8.12",
  "api",
@@ -12346,9 +12914,9 @@ dependencies = [
 
 [[package]]
 name = "slab"
-version = "0.4.10"
+version = "0.4.12"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "04dc19736151f35336d325007ac991178d504a119863a2fcb3758cdb5e52c50d"
+checksum = "0c790de23124f9ab44544d7ac05d60440adc586479ce501c1d6d7da3cd8c9cf5"
 
 [[package]]
 name = "slotmap"
@@ -12415,7 +12983,7 @@ version = "0.8.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1961e2ef424c1424204d3a5d6975f934f56b6d50ff5732382d84ebf460e147f7"
 dependencies = [
- "heck 0.5.0",
+ "heck 0.4.1",
  "proc-macro2",
  "quote",
  "syn 2.0.117",
@@ -12492,7 +13060,7 @@ dependencies = [
 
 [[package]]
 name = "sql"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "arrow-buffer 57.3.0",
@@ -12553,7 +13121,7 @@ dependencies = [
 
 [[package]]
 name = "sqlness-runner"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "async-trait",
  "clap",
@@ -12568,7 +13136,7 @@ dependencies = [
  "local-ip-address",
  "mysql",
  "num_cpus",
- "reqwest",
+ "reqwest 0.12.28",
  "serde",
  "serde_json",
  "sha2",
@@ -12712,7 +13280,7 @@ checksum = "aa003f0038df784eb8fecbbac13affe3da23b45194bd57dba231c8f48199c526"
 dependencies = [
  "atoi",
  "base64 0.22.1",
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "byteorder",
  "bytes",
  "chrono",
@@ -12756,7 +13324,7 @@ checksum = "db58fcd5a53cf07c184b154801ff91347e4c30d17a3562a635ff028ad5deda46"
 dependencies = [
  "atoi",
  "base64 0.22.1",
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "byteorder",
  "chrono",
  "crc",
@@ -12833,7 +13401,7 @@ dependencies = [
 
 [[package]]
 name = "standalone"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "async-trait",
  "catalog",
@@ -12877,7 +13445,7 @@ checksum = "a2eb9349b6444b326872e140eb1cf5e7c522154d69e7a0ffb0fb81c06b37543f"
 
 [[package]]
 name = "store-api"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "aquamarine",
@@ -13069,7 +13637,7 @@ dependencies = [
 
 [[package]]
 name = "substrait"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "async-trait",
  "bytes",
@@ -13191,7 +13759,7 @@ dependencies = [
 
 [[package]]
 name = "table"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "arc-swap",
@@ -13262,7 +13830,7 @@ dependencies = [
  "levenshtein_automata",
  "log",
  "lru",
- "lz4_flex 0.11.5",
+ "lz4_flex 0.11.6",
  "measure_time",
  "memmap2",
  "once_cell",
@@ -13403,9 +13971,9 @@ checksum = "55937e1799185b12863d447f42597ed69d9928686b8d88a1df17376a097d8369"
 
 [[package]]
 name = "tar"
-version = "0.4.44"
+version = "0.4.45"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1d863878d212c87a19c1a610eb53bb01fe12951c0501cf5a0d65f724914a667a"
+checksum = "22692a6476a21fa75fdfc11d452fda482af402c008cdbaf3476414e122040973"
 dependencies = [
  "filetime",
  "libc",
@@ -13461,7 +14029,7 @@ checksum = "8f50febec83f5ee1df3015341d8bd429f2d1cc62bcba7ea2076759d315084683"
 
 [[package]]
 name = "tests-fuzz"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "arbitrary",
  "async-trait",
@@ -13488,7 +14056,8 @@ dependencies = [
  "paste",
  "rand 0.9.1",
  "rand_chacha 0.9.0",
- "reqwest",
+ "reqwest 0.13.2",
+ "rustls",
  "schemars",
  "serde",
  "serde_json",
@@ -13505,7 +14074,7 @@ dependencies = [
 
 [[package]]
 name = "tests-integration"
-version = "1.0.0-rc.2"
+version = "1.0.0"
 dependencies = [
  "api",
  "arrow-flight",
@@ -13513,6 +14082,7 @@ dependencies = [
  "async-trait",
  "auth",
  "axum 0.8.4",
+ "base64 0.22.1",
  "cache",
  "catalog",
  "chrono",
@@ -13528,6 +14098,7 @@ dependencies = [
  "common-grpc",
  "common-memory-manager",
  "common-meta",
+ "common-options",
  "common-procedure",
  "common-procedure-test",
  "common-query",
@@ -13543,6 +14114,7 @@ dependencies = [
  "datanode",
  "datatypes",
  "dotenv",
+ "either",
  "flate2",
  "flow",
  "frontend",
@@ -13558,7 +14130,6 @@ dependencies = [
  "meta-client",
  "meta-srv",
  "mito2",
- "moka",
  "mysql_async",
  "object-store",
  "opentelemetry-proto 0.31.0",
@@ -13567,6 +14138,7 @@ dependencies = [
  "partition",
  "paste",
  "pipeline",
+ "plugins",
  "prost 0.14.1",
  "query",
  "rand 0.9.1",
@@ -13582,7 +14154,7 @@ dependencies = [
  "sqlx",
  "standalone",
  "store-api",
- "substrait 1.0.0-rc.2",
+ "substrait 1.0.0",
  "table",
  "tempfile",
  "time",
@@ -14132,6 +14704,24 @@ dependencies = [
  "tonic-prost",
 ]
 
+[[package]]
+name = "tonic-web"
+version = "0.14.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "29453d84de05f4f1b573db22e6f9f6c95c189a6089a440c9a098aa9dea009299"
+dependencies = [
+ "base64 0.22.1",
+ "bytes",
+ "http 1.3.1",
+ "http-body 1.0.1",
+ "pin-project",
+ "tokio-stream",
+ "tonic 0.14.2",
+ "tower-layer",
+ "tower-service",
+ "tracing",
+]
+
 [[package]]
 name = "tower"
 version = "0.4.13"
@@ -14179,7 +14769,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1e9cd434a998747dd2c4276bc96ee2e0c7a2eadf3cae88e52be55a05fa9053f5"
 dependencies = [
  "base64 0.21.7",
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "bytes",
  "http 1.3.1",
  "http-body 1.0.1",
@@ -14193,13 +14783,13 @@ dependencies = [
 
 [[package]]
 name = "tower-http"
-version = "0.6.6"
+version = "0.6.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "adc82fd73de2a9722ac5da747f12383d2bfdb93591ee6c58486e0097890f05f2"
+checksum = "d4e6559d53cc268e5031cd8429d05415bc4cb4aefc4aa5d6cc35fbf5b924a1f8"
 dependencies = [
  "async-compression",
  "base64 0.22.1",
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "bytes",
  "futures-core",
  "futures-util",
@@ -14553,6 +15143,12 @@ version = "0.2.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "673aac59facbab8a9007c7f6108d11f63b603f7cabff99fabf650fea5c32b861"
 
+[[package]]
+name = "untrusted"
+version = "0.7.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a156c684c91ea7d62626509bce3cb4e1d9ed5c4d978f7b4352658f96a4c26b4a"
+
 [[package]]
 name = "untrusted"
 version = "0.9.0"
@@ -14752,7 +15348,7 @@ dependencies = [
  "itertools 0.14.0",
  "lalrpop",
  "lalrpop-util",
- "lz4_flex 0.11.5",
+ "lz4_flex 0.11.6",
  "md-5",
  "nom 7.1.3",
  "ofb",
@@ -14873,48 +15469,32 @@ checksum = "b8dad83b4f25e74f184f64c43b150b91efe7647395b42289f38e50566d82855b"
 
 [[package]]
 name = "wasm-bindgen"
-version = "0.2.100"
+version = "0.2.118"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1edc8929d7499fc4e8f0be2262a241556cfc54a0bea223790e71446f2aab1ef5"
+checksum = "0bf938a0bacb0469e83c1e148908bd7d5a6010354cf4fb73279b7447422e3a89"
 dependencies = [
  "cfg-if",
  "once_cell",
  "rustversion",
  "wasm-bindgen-macro",
-]
-
-[[package]]
-name = "wasm-bindgen-backend"
-version = "0.2.100"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2f0a0651a5c2bc21487bde11ee802ccaf4c51935d0d3d42a6101f98161700bc6"
-dependencies = [
- "bumpalo",
- "log",
- "proc-macro2",
- "quote",
- "syn 2.0.117",
  "wasm-bindgen-shared",
 ]
 
 [[package]]
 name = "wasm-bindgen-futures"
-version = "0.4.50"
+version = "0.4.68"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "555d470ec0bc3bb57890405e5d4322cc9ea83cebb085523ced7be4144dac1e61"
+checksum = "f371d383f2fb139252e0bfac3b81b265689bf45b6874af544ffa4c975ac1ebf8"
 dependencies = [
- "cfg-if",
  "js-sys",
- "once_cell",
  "wasm-bindgen",
- "web-sys",
 ]
 
 [[package]]
 name = "wasm-bindgen-macro"
-version = "0.2.100"
+version = "0.2.118"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7fe63fc6d09ed3792bd0897b314f53de8e16568c2b3f7982f468c0bf9bd0b407"
+checksum = "eeff24f84126c0ec2db7a449f0c2ec963c6a49efe0698c4242929da037ca28ed"
 dependencies = [
  "quote",
  "wasm-bindgen-macro-support",
@@ -14922,22 +15502,22 @@ dependencies = [
 
 [[package]]
 name = "wasm-bindgen-macro-support"
-version = "0.2.100"
+version = "0.2.118"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8ae87ea40c9f689fc23f209965b6fb8a99ad69aeeb0231408be24920604395de"
+checksum = "9d08065faf983b2b80a79fd87d8254c409281cf7de75fc4b773019824196c904"
 dependencies = [
+ "bumpalo",
  "proc-macro2",
  "quote",
  "syn 2.0.117",
- "wasm-bindgen-backend",
  "wasm-bindgen-shared",
 ]
 
 [[package]]
 name = "wasm-bindgen-shared"
-version = "0.2.100"
+version = "0.2.118"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1a05d73b933a847d6cccdda8f838a22ff101ad9bf93e33684f39c1f5f0eece3d"
+checksum = "5fd04d9e306f1907bd13c6361b5c6bfc7b3b3c095ed3f8a9246390f8dbdee129"
 dependencies = [
  "unicode-ident",
 ]
@@ -14977,13 +15557,26 @@ dependencies = [
  "web-sys",
 ]
 
+[[package]]
+name = "wasm-streams"
+version = "0.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9d1ec4f6517c9e11ae630e200b2b65d193279042e28edd4a2cda233e46670bbb"
+dependencies = [
+ "futures-util",
+ "js-sys",
+ "wasm-bindgen",
+ "wasm-bindgen-futures",
+ "web-sys",
+]
+
 [[package]]
 name = "wasmparser"
 version = "0.244.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "47b807c72e1bac69382b3a6fb3dbe8ea4c0ed87ff5629b8685ae6b9a611028fe"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "hashbrown 0.15.4",
  "indexmap 2.13.0",
  "semver",
@@ -14991,9 +15584,9 @@ dependencies = [
 
 [[package]]
 name = "web-sys"
-version = "0.3.77"
+version = "0.3.95"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "33b6dd2ef9186f1f2072e409e99cd22a975331a6b3591b12c764e0e55c60d5d2"
+checksum = "4f2dfbb17949fa2088e5d39408c48368947b86f7834484e87b73de55bc14d97d"
 dependencies = [
  "js-sys",
  "wasm-bindgen",
@@ -15016,7 +15609,16 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ed63aea5ce73d0ff405984102c42de94fc55a6b75765d621c65262469b3c9b53"
 dependencies = [
  "ring",
- "untrusted",
+ "untrusted 0.9.0",
+]
+
+[[package]]
+name = "webpki-root-certs"
+version = "1.0.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f31141ce3fc3e300ae89b78c0dd67f9708061d1d2eda54b8209346fd6be9a92c"
+dependencies = [
+ "rustls-pki-types",
 ]
 
 [[package]]
@@ -15080,7 +15682,7 @@ version = "0.1.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "cf221c93e13a30d793f7645a0e7762c55d169dbb0a49671918a2319d289b10bb"
 dependencies = [
- "windows-sys 0.59.0",
+ "windows-sys 0.48.0",
 ]
 
 [[package]]
@@ -15250,6 +15852,15 @@ dependencies = [
  "windows-link 0.1.3",
 ]
 
+[[package]]
+name = "windows-sys"
+version = "0.45.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "75283be5efb2831d37ea142365f009c02ec203cd29a3ebecbc093d52315b66d0"
+dependencies = [
+ "windows-targets 0.42.2",
+]
+
 [[package]]
 name = "windows-sys"
 version = "0.48.0"
@@ -15286,6 +15897,21 @@ dependencies = [
  "windows-link 0.2.1",
 ]
 
+[[package]]
+name = "windows-targets"
+version = "0.42.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8e5180c00cd44c9b1c88adb3693291f1cd93605ded80c250a75d472756b4d071"
+dependencies = [
+ "windows_aarch64_gnullvm 0.42.2",
+ "windows_aarch64_msvc 0.42.2",
+ "windows_i686_gnu 0.42.2",
+ "windows_i686_msvc 0.42.2",
+ "windows_x86_64_gnu 0.42.2",
+ "windows_x86_64_gnullvm 0.42.2",
+ "windows_x86_64_msvc 0.42.2",
+]
+
 [[package]]
 name = "windows-targets"
 version = "0.48.5"
@@ -15326,6 +15952,12 @@ dependencies = [
  "windows-link 0.1.3",
 ]
 
+[[package]]
+name = "windows_aarch64_gnullvm"
+version = "0.42.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "597a5118570b68bc08d8d59125332c54f1ba9d9adeedeef5b99b02ba2b0698f8"
+
 [[package]]
 name = "windows_aarch64_gnullvm"
 version = "0.48.5"
@@ -15338,6 +15970,12 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "32a4622180e7a0ec044bb555404c800bc9fd9ec262ec147edd5989ccd0c02cd3"
 
+[[package]]
+name = "windows_aarch64_msvc"
+version = "0.42.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e08e8864a60f06ef0d0ff4ba04124db8b0fb3be5776a5cd47641e942e58c4d43"
+
 [[package]]
 name = "windows_aarch64_msvc"
 version = "0.48.5"
@@ -15350,6 +15988,12 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "09ec2a7bb152e2252b53fa7803150007879548bc709c039df7627cabbd05d469"
 
+[[package]]
+name = "windows_i686_gnu"
+version = "0.42.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c61d927d8da41da96a81f029489353e68739737d3beca43145c8afec9a31a84f"
+
 [[package]]
 name = "windows_i686_gnu"
 version = "0.48.5"
@@ -15368,6 +16012,12 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0eee52d38c090b3caa76c563b86c3a4bd71ef1a819287c19d586d7334ae8ed66"
 
+[[package]]
+name = "windows_i686_msvc"
+version = "0.42.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "44d840b6ec649f480a41c8d80f9c65108b92d89345dd94027bfe06ac444d1060"
+
 [[package]]
 name = "windows_i686_msvc"
 version = "0.48.5"
@@ -15380,6 +16030,12 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "240948bc05c5e7c6dabba28bf89d89ffce3e303022809e73deaefe4f6ec56c66"
 
+[[package]]
+name = "windows_x86_64_gnu"
+version = "0.42.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8de912b8b8feb55c064867cf047dda097f92d51efad5b491dfb98f6bbb70cb36"
+
 [[package]]
 name = "windows_x86_64_gnu"
 version = "0.48.5"
@@ -15392,6 +16048,12 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "147a5c80aabfbf0c7d901cb5895d1de30ef2907eb21fbbab29ca94c5b08b1a78"
 
+[[package]]
+name = "windows_x86_64_gnullvm"
+version = "0.42.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "26d41b46a36d453748aedef1486d5c7a85db22e56aff34643984ea85514e94a3"
+
 [[package]]
 name = "windows_x86_64_gnullvm"
 version = "0.48.5"
@@ -15404,6 +16066,12 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "24d5b23dc417412679681396f2b49f3de8c1473deb516bd34410872eff51ed0d"
 
+[[package]]
+name = "windows_x86_64_msvc"
+version = "0.42.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9aec5da331524158c6d1a4ac0ab1541149c0b9505fde06423b02f5ef0106b9f0"
+
 [[package]]
 name = "windows_x86_64_msvc"
 version = "0.48.5"
@@ -15469,7 +16137,7 @@ version = "0.39.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6f42320e61fe2cfd34354ecb597f86f413484a798ba44a8ca1165c58d42da6c1"
 dependencies = [
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
 ]
 
 [[package]]
@@ -15510,7 +16178,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9d66ea20e9553b30172b5e831994e35fbde2d165325bec84fc43dbf6f4eb9cb2"
 dependencies = [
  "anyhow",
- "bitflags 2.9.1",
+ "bitflags 2.11.1",
  "indexmap 2.13.0",
  "log",
  "serde",
diff --git a/Cargo.toml b/Cargo.toml
index 5041f167c3..c9ddba912a 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -75,7 +75,7 @@ members = [
 resolver = "2"
 
 [workspace.package]
-version = "1.0.0-rc.2"
+version = "1.0.0"
 edition = "2024"
 license = "Apache-2.0"
 
@@ -110,6 +110,7 @@ arrow-schema = { version = "57.3", features = ["serde"] }
 async-stream = "0.3"
 async-trait = "0.1"
 # Remember to update axum-extra, axum-macros when updating axum
+arrow_object_store = { package = "object_store", version = "0.13.2" }
 axum = "0.8"
 axum-extra = "0.10"
 axum-macros = "0.5"
@@ -131,6 +132,7 @@ datafusion = "=52.1"
 datafusion-common = "=52.1"
 datafusion-datasource = "=52.1"
 datafusion-expr = "=52.1"
+datafusion-expr-common = "=52.1"
 datafusion-functions = "=52.1"
 datafusion-functions-aggregate-common = "=52.1"
 datafusion-functions-window-common = "=52.1"
@@ -141,6 +143,7 @@ datafusion-physical-expr = "=52.1"
 datafusion-physical-plan = "=52.1"
 datafusion-sql = "=52.1"
 datafusion-substrait = "=52.1"
+datafusion_object_store = { package = "object_store", version = "0.12.5" }
 deadpool = "0.12"
 deadpool-postgres = "0.14"
 derive_builder = "0.20"
@@ -151,16 +154,18 @@ etcd-client = { version = "0.17", features = [
     "tls",
     "tls-roots",
 ] }
+fs2 = "0.4"
 fst = "0.4.7"
 futures = "0.3"
 futures-util = "0.3"
-greptime-proto = { git = "https://github.com/GreptimeTeam/greptime-proto.git", rev = "092ba1d01e2da676dca66cca7eebb55009da8ef8" }
+greptime-proto = { git = "https://github.com/GreptimeTeam/greptime-proto.git", rev = "0de5437582920c8b30d6c34212f161db71d95c50" }
 hex = "0.4"
 http = "1"
 humantime = "2.1"
 humantime-serde = "1.1"
 hyper = "1.1"
 hyper-util = "0.1"
+icu_properties = "2.0.1"
 itertools = "0.14"
 jsonb = { version = "0.4.4", default-features = false }
 lazy_static = "1.4"
@@ -173,7 +178,7 @@ nalgebra = "0.33"
 nix = { version = "0.30.1", default-features = false, features = ["event", "fs", "process"] }
 notify = "8.0"
 num_cpus = "1.16"
-object_store_opendal = "0.54"
+object_store_opendal = "0.56"
 once_cell = "1.18"
 opentelemetry-proto = { version = "0.31", features = [
     "gen-tonic",
@@ -200,14 +205,17 @@ rand = "0.9"
 ratelimit = "0.10"
 regex = "1.12"
 regex-automata = "0.4"
-reqwest = { version = "0.12", default-features = false, features = [
+reqwest = { version = "0.13", default-features = false, features = [
+    "form",
     "json",
-    "rustls-tls-native-roots",
+    "query",
+    "rustls",
     "stream",
     "multipart",
 ] }
 url = "2.3"
 # Branch: feat/request-timeout
+hostname = "0.4.0"
 rskafka = { git = "https://github.com/GreptimeTeam/rskafka.git", rev = "f5688f83e7da591cda3f2674c2408b4c0ed4ed50", features = [
     "transport-tls",
 ] }
@@ -215,8 +223,6 @@ rstest = "0.25"
 rstest_reuse = "0.7"
 rust_decimal = "1.33"
 rustc-hash = "2.0"
-# It is worth noting that we should try to avoid using aws-lc-rs until it can be compiled on various platforms.
-hostname = "0.4.0"
 rustls = { version = "0.23.25", default-features = false }
 sea-query = "0.32"
 serde = { version = "1.0", features = ["derive"] }
@@ -231,7 +237,8 @@ sqlx = { version = "0.8", default-features = false, features = [
     "any",
     "macros",
     "json",
-    "runtime-tokio-rustls",
+    "runtime-tokio",
+    "tls-rustls-aws-lc-rs",
     "rust_decimal",
 ] }
 strum = { version = "0.27", features = ["derive"] }
@@ -243,7 +250,7 @@ tokio-rustls = { version = "0.26.2", default-features = false }
 tokio-stream = "0.1"
 tokio-util = { version = "0.7", features = ["io-util", "compat"] }
 toml = "0.8.8"
-tonic = { version = "0.14", features = ["tls-ring", "gzip", "zstd"] }
+tonic = { version = "0.14", features = ["tls-aws-lc", "gzip", "zstd"] }
 tower = "0.5"
 tower-http = "0.6"
 tracing = "0.1"
@@ -334,6 +341,7 @@ rev = "5618e779cf2bb4755b499c630fba4c35e91898cb"
 datafusion = { git = "https://github.com/GreptimeTeam/datafusion.git", rev = "02b82535e0160c4545667f36a03e1ff9d1d2e51f" }
 datafusion-common = { git = "https://github.com/GreptimeTeam/datafusion.git", rev = "02b82535e0160c4545667f36a03e1ff9d1d2e51f" }
 datafusion-expr = { git = "https://github.com/GreptimeTeam/datafusion.git", rev = "02b82535e0160c4545667f36a03e1ff9d1d2e51f" }
+datafusion-expr-common = { git = "https://github.com/GreptimeTeam/datafusion.git", rev = "02b82535e0160c4545667f36a03e1ff9d1d2e51f" }
 datafusion-functions = { git = "https://github.com/GreptimeTeam/datafusion.git", rev = "02b82535e0160c4545667f36a03e1ff9d1d2e51f" }
 datafusion-functions-aggregate-common = { git = "https://github.com/GreptimeTeam/datafusion.git", rev = "02b82535e0160c4545667f36a03e1ff9d1d2e51f" }
 datafusion-functions-window-common = { git = "https://github.com/GreptimeTeam/datafusion.git", rev = "02b82535e0160c4545667f36a03e1ff9d1d2e51f" }
diff --git a/Makefile b/Makefile
index 60ea01a3ce..3fd09ad4ea 100644
--- a/Makefile
+++ b/Makefile
@@ -8,7 +8,7 @@ CARGO_BUILD_OPTS := --locked
 IMAGE_REGISTRY ?= docker.io
 IMAGE_NAMESPACE ?= greptime
 IMAGE_TAG ?= latest
-DEV_BUILDER_IMAGE_TAG ?= 2025-10-01-8fe17d43-20251011080129
+DEV_BUILDER_IMAGE_TAG ?= 2026-03-21-9c9d9e9e-20260331090344
 BUILDX_MULTI_PLATFORM_BUILD ?= false
 BUILDX_BUILDER_NAME ?= gtbuilder
 BASE_IMAGE ?= ubuntu
diff --git a/README.md b/README.md
index 290fe0263d..4ed99fa306 100644
--- a/README.md
+++ b/README.md
@@ -131,7 +131,7 @@ docker run -p 127.0.0.1:4000-4003:4000-4003 \
   --name greptime --rm \
   greptime/greptimedb:latest standalone start \
   --http-addr 0.0.0.0:4000 \
-  --rpc-bind-addr 0.0.0.0:4001 \
+  --grpc-bind-addr 0.0.0.0:4001 \
   --mysql-addr 0.0.0.0:4002 \
   --postgres-addr 0.0.0.0:4003
 ```
@@ -175,17 +175,16 @@ cargo run -- standalone start
 
 ## Project Status
 
-> **Status:** RC — marching toward v1.0 GA!
-> **GA (v1.0):** March 2026
+> **Status:** [v1.0 GA](https://github.com/GreptimeTeam/greptimedb/releases/tag/v1.0.0) — generally available and production-ready! 🎉
 
 - Deployed in production handling billions of data points daily
 - Stable APIs, actively maintained, with regular releases ([version info](https://docs.greptime.com/nightly/reference/about-greptimedb-version))
 
-GreptimeDB v1.0 represents a major milestone toward maturity — marking stable APIs, production readiness, and proven performance.
+GreptimeDB v1.0 marks a major milestone — stable APIs, production readiness, and proven performance at scale.
 
-**Roadmap:** [v1.0 highlights and release plan](https://greptime.com/blogs/2025-11-05-greptimedb-v1-highlights) and [2026 roadmap](https://greptime.com/blogs/2026-02-11-greptimedb-roadmap-2026).
+**Learn more:** [v1.0 highlights](https://greptime.com/blogs/2025-11-05-greptimedb-v1-highlights) and [2026 roadmap](https://greptime.com/blogs/2026-02-11-greptimedb-roadmap-2026).
 
-For production use, we recommend using the latest stable release.
+For production use, we recommend v1.0 or later.
 
 If you find this project useful, a ⭐ would mean a lot to us!
 
diff --git a/cliff.toml b/cliff.toml
index 4245203e92..2b35ddab5c 100644
--- a/cliff.toml
+++ b/cliff.toml
@@ -12,7 +12,9 @@ footer = ""
 body = """
 # {{ version }}
 
+{% if timestamp -%}
 Release date: {{ timestamp | date(format="%B %d, %Y") }}
+{% endif -%}
 
 {%- set breakings = commits | filter(attribute="breaking", value=true) -%}
 {%- if breakings | length > 0 %}
@@ -118,7 +120,10 @@ filter_commits = false
 # regex for skipping tags
 # skip_tags = ""
 # regex for ignoring tags
-ignore_tags = ".*-nightly-.*"
+# Ignore nightly tags and build-suffixed release tags such as
+# v1.0.0-rc.2-13cdfa9b5-20260325-1774407105 so their commits are merged into
+# the next visible release section instead of creating extra headings.
+ignore_tags = ".*-nightly-.*|^v[0-9]+\\.[0-9]+\\.[0-9]+(-(alpha|beta|rc)\\.[0-9]+)?-[0-9a-f]{7,}-[0-9]{8}-[0-9]+$"
 # sort the tags topologically
 topo_order = false
 # sort the commits inside sections by oldest/newest order
diff --git a/config/config.md b/config/config.md
index 2ac11dd6e6..dfb4db1cf1 100644
--- a/config/config.md
+++ b/config/config.md
@@ -14,11 +14,12 @@
 | --- | -----| ------- | ----------- |
 | `default_timezone` | String | Unset | The default timezone of the server. |
 | `default_column_prefix` | String | Unset | The default column prefix for auto-created time index and value columns. |
+| `user_provider` | String | Unset | The user provider for authentication.<br/>Examples: "static_user_provider:file:/path/to/users", "static_user_provider:cmd:greptime_user=greptime_pwd" |
 | `max_in_flight_write_bytes` | String | Unset | Maximum total memory for all concurrent write request bodies and messages (HTTP, gRPC, Flight).<br/>Set to 0 to disable the limit. Default: "0" (unlimited) |
 | `write_bytes_exhausted_policy` | String | Unset | Policy when write bytes quota is exhausted.<br/>Options: "wait" (default, 10s timeout), "wait(<duration>)" (e.g., "wait(30s)"), "fail" |
 | `init_regions_in_background` | Bool | `false` | Initialize all regions in the background during the startup.<br/>By default, it provides services after all regions have been initialized. |
 | `init_regions_parallelism` | Integer | `16` | Parallelism of initializing regions. |
-| `max_concurrent_queries` | Integer | `0` | The maximum current queries allowed to be executed. Zero means unlimited.<br/>NOTE: This setting affects scan_memory_limit's privileged tier allocation.<br/>When set, 70% of queries get privileged memory access (full scan_memory_limit).<br/>The remaining 30% get standard tier access (70% of scan_memory_limit). |
+| `max_concurrent_queries` | Integer | `0` | The maximum concurrent queries allowed to be executed. Zero means unlimited. |
 | `enable_telemetry` | Bool | `true` | Enable telemetry to collect anonymous usage data. Enabled by default. |
 | `runtime` | -- | -- | The runtime options. |
 | `runtime.global_rt_size` | Integer | `8` | The number of threads to execute the runtime for global read operations. |
@@ -69,6 +70,11 @@
 | `prom_store` | -- | -- | Prometheus remote storage options |
 | `prom_store.enable` | Bool | `true` | Whether to enable Prometheus remote write and read in HTTP API. |
 | `prom_store.with_metric_engine` | Bool | `true` | Whether to store the data from Prometheus remote write in metric engine. |
+| `prom_store.pending_rows_flush_interval` | String | `0s` | Interval to flush pending rows batcher.<br/>Set to "0s" to disable batching mode in Prometheus Remote Write endpoint |
+| `prom_store.max_batch_rows` | Integer | `100000` | Max rows per pending batch before triggering a flush. |
+| `prom_store.max_concurrent_flushes` | Integer | `256` | Max number of concurrent batch flushes. |
+| `prom_store.worker_channel_capacity` | Integer | `65526` | Capacity of the pending batch worker channel. |
+| `prom_store.max_inflight_requests` | Integer | `3000` | Max inflight write requests before backpressure. |
 | `wal` | -- | -- | The WAL options. |
 | `wal.provider` | String | `raft_engine` | The provider of the WAL.<br/>- `raft_engine`: the wal is stored in the local file system by raft-engine.<br/>- `kafka`: it's remote wal that data is stored in Kafka. |
 | `wal.dir` | String | Unset | The directory to store the WAL files.<br/>**It's only used when the provider is `raft_engine`**. |
@@ -139,7 +145,7 @@
 | `region_engine.mito.max_background_flushes` | Integer | Auto | Max number of running background flush jobs (default: 1/2 of cpu cores). |
 | `region_engine.mito.max_background_compactions` | Integer | Auto | Max number of running background compaction jobs (default: 1/4 of cpu cores). |
 | `region_engine.mito.max_background_purges` | Integer | Auto | Max number of running background purge jobs (default: number of cpu cores). |
-| `region_engine.mito.experimental_compaction_memory_limit` | String | 0 | Memory budget for compaction tasks. Setting it to 0 or "unlimited" disables the limit. |
+| `region_engine.mito.experimental_compaction_memory_limit` | String | 0 | Memory budget for compaction tasks.<br/>Supports absolute size (e.g., "2GiB", "512MB") or percentage of system memory (e.g., "50%").<br/>Setting it to 0 or "unlimited" disables the limit. |
 | `region_engine.mito.experimental_compaction_on_exhausted` | String | wait | Behavior when compaction cannot acquire memory from the budget.<br/>Options: "wait" (default, 10s), "wait(<duration>)", "fail" |
 | `region_engine.mito.auto_flush_interval` | String | `1h` | Interval to auto flush a region if it has not flushed yet. |
 | `region_engine.mito.global_write_buffer_size` | String | Auto | Global write buffer size for all regions. If not set, it's default to 1/8 of OS memory with a max limitation of 1GB. |
@@ -157,12 +163,12 @@
 | `region_engine.mito.enable_refill_cache_on_read` | Bool | `true` | Enable refilling cache on read operations (default: true).<br/>When disabled, cache refilling on read won't happen. |
 | `region_engine.mito.manifest_cache_size` | String | `256MB` | Capacity for manifest cache (default: 256MB). |
 | `region_engine.mito.sst_write_buffer_size` | String | `8MB` | Buffer size for SST writing. |
-| `region_engine.mito.parallel_scan_channel_size` | Integer | `32` | Capacity of the channel to send data from parallel scan tasks to the main task. |
 | `region_engine.mito.max_concurrent_scan_files` | Integer | `384` | Maximum number of SST files to scan concurrently. |
 | `region_engine.mito.allow_stale_entries` | Bool | `false` | Whether to allow stale WAL entries read during replay. |
-| `region_engine.mito.scan_memory_limit` | String | `50%` | Memory limit for table scans across all queries.<br/>Supports absolute size (e.g., "2GB") or percentage of system memory (e.g., "20%").<br/>Setting it to 0 disables the limit.<br/>NOTE: Works with max_concurrent_queries for tiered memory allocation.<br/>- If max_concurrent_queries is set: 70% of queries get full access, 30% get 70% access.<br/>- If max_concurrent_queries is 0 (unlimited): first 20 queries get full access, rest get 70% access. |
+| `region_engine.mito.scan_memory_limit` | String | `50%` | Memory limit for table scans across all queries.<br/>Supports absolute size (e.g., "2GB") or percentage of system memory (e.g., "20%").<br/>Setting it to 0 disables the limit. |
+| `region_engine.mito.scan_memory_on_exhausted` | String | `fail` | Controls what happens when a scan cannot get memory immediately.<br/>"fail" (default) fails fast and is the recommended option for most users.<br/>"wait" / "wait(<duration>)" waits for memory to become available. This is mainly<br/>for advanced tuning in bursty workloads where temporary contention is common and<br/>higher latency is acceptable.<br/>"wait" means "wait(10s)", not unlimited waiting. |
 | `region_engine.mito.min_compaction_interval` | String | `0m` | Minimum time interval between two compactions.<br/>To align with the old behavior, the default value is 0 (no restrictions). |
-| `region_engine.mito.default_experimental_flat_format` | Bool | `false` | Whether to enable experimental flat format as the default format. |
+| `region_engine.mito.default_flat_format` | Bool | `true` | Whether to enable flat format as the default SST format. |
 | `region_engine.mito.index` | -- | -- | The options for index in Mito engine. |
 | `region_engine.mito.index.aux_path` | String | `""` | Auxiliary directory path for the index in filesystem, used to store intermediate files for<br/>creating the index and staging files for searching the index, defaults to `{data_home}/index_intermediate`.<br/>The default name for this directory is `index_intermediate` for backward compatibility.<br/><br/>This path contains two subdirectories:<br/>- `__intm`: for storing intermediate files used during creating index.<br/>- `staging`: for storing staging files used during searching index. |
 | `region_engine.mito.index.staging_size` | String | `2GB` | The max capacity of the staging directory. |
@@ -226,14 +232,12 @@
 | --- | -----| ------- | ----------- |
 | `default_timezone` | String | Unset | The default timezone of the server. |
 | `default_column_prefix` | String | Unset | The default column prefix for auto-created time index and value columns. |
+| `user_provider` | String | Unset | The user provider for authentication.<br/>Examples: "static_user_provider:file:/path/to/users", "static_user_provider:cmd:greptime_user=greptime_pwd" |
 | `max_in_flight_write_bytes` | String | Unset | Maximum total memory for all concurrent write request bodies and messages (HTTP, gRPC, Flight).<br/>Set to 0 to disable the limit. Default: "0" (unlimited) |
 | `write_bytes_exhausted_policy` | String | Unset | Policy when write bytes quota is exhausted.<br/>Options: "wait" (default, 10s timeout), "wait(<duration>)" (e.g., "wait(30s)"), "fail" |
 | `runtime` | -- | -- | The runtime options. |
 | `runtime.global_rt_size` | Integer | `8` | The number of threads to execute the runtime for global read operations. |
 | `runtime.compact_rt_size` | Integer | `4` | The number of threads to execute the runtime for global write operations. |
-| `heartbeat` | -- | -- | The heartbeat options. |
-| `heartbeat.interval` | String | `18s` | Interval for sending heartbeat messages to the metasrv. |
-| `heartbeat.retry_interval` | String | `3s` | Interval for retrying to send heartbeat messages to the metasrv. |
 | `http` | -- | -- | The HTTP server options. |
 | `http.addr` | String | `127.0.0.1:4000` | The address to bind the HTTP server. |
 | `http.timeout` | String | `0s` | HTTP request timeout. Set to 0 to disable timeout. |
@@ -254,7 +258,7 @@
 | `grpc.tls.watch` | Bool | `false` | Watch for Certificate and key file change and auto reload.<br/>For now, gRPC tls config does not support auto reload. |
 | `internal_grpc` | -- | -- | The internal gRPC server options. Internal gRPC port for nodes inside cluster to access frontend. |
 | `internal_grpc.bind_addr` | String | `127.0.0.1:4010` | The address to bind the gRPC server. |
-| `internal_grpc.server_addr` | String | `127.0.0.1:4010` | The address advertised to the metasrv, and used for connections from outside the host.<br/>If left empty or unset, the server will automatically use the IP address of the first network interface<br/>on the host, with the same port number as the one specified in `grpc.bind_addr`. |
+| `internal_grpc.server_addr` | String | `127.0.0.1:4010` | The address advertised to the metasrv, and used for connections from outside the host.<br/>If left empty or unset, the server will automatically use the IP address of the first network interface<br/>on the host, with the same port number as the one specified in `internal_grpc.bind_addr`. |
 | `internal_grpc.runtime_size` | Integer | `8` | The number of server worker threads. |
 | `internal_grpc.flight_compression` | String | `arrow_ipc` | Compression mode for frontend side Arrow IPC service. Available options:<br/>- `none`: disable all compression<br/>- `transport`: only enable gRPC transport compression (zstd)<br/>- `arrow_ipc`: only enable Arrow IPC compression (lz4)<br/>- `all`: enable all compression.<br/>Default to `none` |
 | `internal_grpc.tls` | -- | -- | internal gRPC server TLS options, see `mysql.tls` section. |
@@ -292,6 +296,11 @@
 | `prom_store` | -- | -- | Prometheus remote storage options |
 | `prom_store.enable` | Bool | `true` | Whether to enable Prometheus remote write and read in HTTP API. |
 | `prom_store.with_metric_engine` | Bool | `true` | Whether to store the data from Prometheus remote write in metric engine. |
+| `prom_store.pending_rows_flush_interval` | String | `0s` | Interval to flush pending rows batcher.<br/>Set to "0s" to disable batching mode in Prometheus Remote Write endpoint |
+| `prom_store.max_batch_rows` | Integer | `100000` | Max rows per pending batch before triggering a flush. |
+| `prom_store.max_concurrent_flushes` | Integer | `256` | Max number of concurrent batch flushes. |
+| `prom_store.worker_channel_capacity` | Integer | `65526` | Capacity of the pending batch worker channel. |
+| `prom_store.max_inflight_requests` | Integer | `3000` | Max inflight write requests before backpressure. |
 | `meta_client` | -- | -- | The metasrv client options. |
 | `meta_client.metasrv_addrs` | Array | -- | The addresses of the metasrv. |
 | `meta_client.timeout` | String | `3s` | Operation timeout. |
@@ -352,7 +361,7 @@
 | `region_failure_detector_initialization_delay` | String | `10m` | The delay before starting region failure detection.<br/>This delay helps prevent Metasrv from triggering unnecessary region failovers before all Datanodes are fully started.<br/>Especially useful when the cluster is not deployed with GreptimeDB Operator and maintenance mode is not enabled. |
 | `allow_region_failover_on_local_wal` | Bool | `false` | Whether to allow region failover on local WAL.<br/>**This option is not recommended to be set to true, because it may lead to data loss during failover.** |
 | `node_max_idle_time` | String | `24hours` | Max allowed idle time before removing node info from metasrv memory. |
-| `heartbeat_interval` | String | `3s` | Base heartbeat interval for calculating distributed time constants.<br/>The frontend heartbeat interval is 6 times of the base heartbeat interval.<br/>The flownode/datanode heartbeat interval is 1 times of the base heartbeat interval.<br/>e.g., If the base heartbeat interval is 3s, the frontend heartbeat interval is 18s, the flownode/datanode heartbeat interval is 3s.<br/>If you change this value, you need to change the heartbeat interval of the flownode/frontend/datanode accordingly. |
+| `heartbeat_interval` | String | `3s` | Base heartbeat interval for calculating distributed time constants.<br/>The frontend heartbeat interval is 6 times of the base heartbeat interval.<br/>The flownode/datanode heartbeat interval is 1 times of the base heartbeat interval.<br/>e.g., If the base heartbeat interval is 3s, the frontend heartbeat interval is 18s, the flownode/datanode heartbeat interval is 3s.<br/>Heartbeat intervals are negotiated from metasrv during handshake; local node configs do not override this. |
 | `enable_telemetry` | Bool | `true` | Whether to enable greptimedb telemetry. Enabled by default. |
 | `runtime` | -- | -- | The runtime options. |
 | `runtime.global_rt_size` | Integer | `8` | The number of threads to execute the runtime for global read operations. |
@@ -368,7 +377,7 @@
 | `backend_client.connect_timeout` | String | `3s` | The connect timeout for backend client. |
 | `grpc` | -- | -- | The gRPC server options. |
 | `grpc.bind_addr` | String | `127.0.0.1:3002` | The address to bind the gRPC server. |
-| `grpc.server_addr` | String | `127.0.0.1:3002` | The communication server address for the frontend and datanode to connect to metasrv.<br/>If left empty or unset, the server will automatically use the IP address of the first network interface<br/>on the host, with the same port number as the one specified in `bind_addr`. |
+| `grpc.server_addr` | String | `127.0.0.1:3002` | The communication server address for the frontend and datanode to connect to metasrv.<br/>If left empty or unset, the server will automatically use the IP address of the first network interface<br/>on the host, with the same port number as the one specified in `grpc.bind_addr`. |
 | `grpc.runtime_size` | Integer | `8` | The number of server worker threads. |
 | `grpc.max_recv_message_size` | String | `512MB` | The maximum receive message size for gRPC server. |
 | `grpc.max_send_message_size` | String | `512MB` | The maximum send message size for gRPC server. |
@@ -440,7 +449,7 @@
 | `require_lease_before_startup` | Bool | `false` | Start services after regions have obtained leases.<br/>It will block the datanode start if it can't receive leases in the heartbeat from metasrv. |
 | `init_regions_in_background` | Bool | `false` | Initialize all regions in the background during the startup.<br/>By default, it provides services after all regions have been initialized. |
 | `init_regions_parallelism` | Integer | `16` | Parallelism of initializing regions. |
-| `max_concurrent_queries` | Integer | `0` | The maximum current queries allowed to be executed. Zero means unlimited.<br/>NOTE: This setting affects scan_memory_limit's privileged tier allocation.<br/>When set, 70% of queries get privileged memory access (full scan_memory_limit).<br/>The remaining 30% get standard tier access (70% of scan_memory_limit). |
+| `max_concurrent_queries` | Integer | `0` | The maximum concurrent queries allowed to be executed. Zero means unlimited. |
 | `enable_telemetry` | Bool | `true` | Enable telemetry to collect anonymous usage data. Enabled by default. |
 | `http` | -- | -- | The HTTP server options. |
 | `http.addr` | String | `127.0.0.1:4000` | The address to bind the HTTP server. |
@@ -461,9 +470,6 @@
 | `runtime` | -- | -- | The runtime options. |
 | `runtime.global_rt_size` | Integer | `8` | The number of threads to execute the runtime for global read operations. |
 | `runtime.compact_rt_size` | Integer | `4` | The number of threads to execute the runtime for global write operations. |
-| `heartbeat` | -- | -- | The heartbeat options. |
-| `heartbeat.interval` | String | `3s` | Interval for sending heartbeat messages to the metasrv. |
-| `heartbeat.retry_interval` | String | `3s` | Interval for retrying to send heartbeat messages to the metasrv. |
 | `meta_client` | -- | -- | The metasrv client options. |
 | `meta_client.metasrv_addrs` | Array | -- | The addresses of the metasrv. |
 | `meta_client.timeout` | String | `3s` | Operation timeout. |
@@ -531,7 +537,7 @@
 | `region_engine.mito.max_background_flushes` | Integer | Auto | Max number of running background flush jobs (default: 1/2 of cpu cores). |
 | `region_engine.mito.max_background_compactions` | Integer | Auto | Max number of running background compaction jobs (default: 1/4 of cpu cores). |
 | `region_engine.mito.max_background_purges` | Integer | Auto | Max number of running background purge jobs (default: number of cpu cores). |
-| `region_engine.mito.experimental_compaction_memory_limit` | String | 0 | Memory budget for compaction tasks. Setting it to 0 or "unlimited" disables the limit. |
+| `region_engine.mito.experimental_compaction_memory_limit` | String | 0 | Memory budget for compaction tasks.<br/>Supports absolute size (e.g., "2GiB", "512MB") or percentage of system memory (e.g., "50%").<br/>Setting it to 0 or "unlimited" disables the limit. |
 | `region_engine.mito.experimental_compaction_on_exhausted` | String | wait | Behavior when compaction cannot acquire memory from the budget.<br/>Options: "wait" (default, 10s), "wait(<duration>)", "fail" |
 | `region_engine.mito.auto_flush_interval` | String | `1h` | Interval to auto flush a region if it has not flushed yet. |
 | `region_engine.mito.global_write_buffer_size` | String | Auto | Global write buffer size for all regions. If not set, it's default to 1/8 of OS memory with a max limitation of 1GB. |
@@ -549,12 +555,12 @@
 | `region_engine.mito.enable_refill_cache_on_read` | Bool | `true` | Enable refilling cache on read operations (default: true).<br/>When disabled, cache refilling on read won't happen. |
 | `region_engine.mito.manifest_cache_size` | String | `256MB` | Capacity for manifest cache (default: 256MB). |
 | `region_engine.mito.sst_write_buffer_size` | String | `8MB` | Buffer size for SST writing. |
-| `region_engine.mito.parallel_scan_channel_size` | Integer | `32` | Capacity of the channel to send data from parallel scan tasks to the main task. |
 | `region_engine.mito.max_concurrent_scan_files` | Integer | `384` | Maximum number of SST files to scan concurrently. |
 | `region_engine.mito.allow_stale_entries` | Bool | `false` | Whether to allow stale WAL entries read during replay. |
-| `region_engine.mito.scan_memory_limit` | String | `50%` | Memory limit for table scans across all queries.<br/>Supports absolute size (e.g., "2GB") or percentage of system memory (e.g., "20%").<br/>Setting it to 0 disables the limit.<br/>NOTE: Works with max_concurrent_queries for tiered memory allocation.<br/>- If max_concurrent_queries is set: 70% of queries get full access, 30% get 70% access.<br/>- If max_concurrent_queries is 0 (unlimited): first 20 queries get full access, rest get 70% access. |
+| `region_engine.mito.scan_memory_limit` | String | `50%` | Memory limit for table scans across all queries.<br/>Supports absolute size (e.g., "2GB") or percentage of system memory (e.g., "20%").<br/>Setting it to 0 disables the limit. |
+| `region_engine.mito.scan_memory_on_exhausted` | String | `fail` | Controls what happens when a scan cannot get memory immediately.<br/>"fail" (default) fails fast and is the recommended option for most users.<br/>"wait" / "wait(<duration>)" waits for memory to become available. This is mainly<br/>for advanced tuning in bursty workloads where temporary contention is common and<br/>higher latency is acceptable.<br/>"wait" means "wait(10s)", not unlimited waiting. |
 | `region_engine.mito.min_compaction_interval` | String | `0m` | Minimum time interval between two compactions.<br/>To align with the old behavior, the default value is 0 (no restrictions). |
-| `region_engine.mito.default_experimental_flat_format` | Bool | `false` | Whether to enable experimental flat format as the default format. |
+| `region_engine.mito.default_flat_format` | Bool | `true` | Whether to enable flat format as the default SST format. |
 | `region_engine.mito.index` | -- | -- | The options for index in Mito engine. |
 | `region_engine.mito.index.aux_path` | String | `""` | Auxiliary directory path for the index in filesystem, used to store intermediate files for<br/>creating the index and staging files for searching the index, defaults to `{data_home}/index_intermediate`.<br/>The default name for this directory is `index_intermediate` for backward compatibility.<br/><br/>This path contains two subdirectories:<br/>- `__intm`: for storing intermediate files used during creating index.<br/>- `staging`: for storing staging files used during searching index. |
 | `region_engine.mito.index.staging_size` | String | `2GB` | The max capacity of the staging directory. |
@@ -614,6 +620,7 @@
 | Key | Type | Default | Descriptions |
 | --- | -----| ------- | ----------- |
 | `node_id` | Integer | Unset | The flownode identifier and should be unique in the cluster. |
+| `user_provider` | String | Unset | The user provider for authentication.<br/>Examples: "static_user_provider:file:/path/to/users", "static_user_provider:cmd:greptime_user=greptime_pwd" |
 | `flow` | -- | -- | flow engine options. |
 | `flow.num_workers` | Integer | `0` | The number of flow worker in flownode.<br/>Not setting(or set to 0) this value will use the number of CPU cores divided by 2. |
 | `flow.batching_mode` | -- | -- | -- |
@@ -623,7 +630,6 @@
 | `flow.batching_mode.grpc_conn_timeout` | String | `5s` | The gRPC connection timeout |
 | `flow.batching_mode.experimental_grpc_max_retries` | Integer | `3` | The gRPC max retry number |
 | `flow.batching_mode.experimental_frontend_scan_timeout` | String | `30s` | Flow wait for available frontend timeout,<br/>if failed to find available frontend after frontend_scan_timeout elapsed, return error<br/>which prevent flownode from starting |
-| `flow.batching_mode.experimental_frontend_activity_timeout` | String | `60s` | Frontend activity timeout<br/>if frontend is down(not sending heartbeat) for more than frontend_activity_timeout,<br/>it will be removed from the list that flownode use to connect |
 | `flow.batching_mode.experimental_max_filter_num_per_query` | Integer | `20` | Maximum number of filters allowed in a single query |
 | `flow.batching_mode.experimental_time_window_merge_threshold` | Integer | `3` | Time window merge distance |
 | `flow.batching_mode.read_preference` | String | `Leader` | Read preference of the Frontend client. |
@@ -651,9 +657,6 @@
 | `meta_client.metadata_cache_max_capacity` | Integer | `100000` | The configuration about the cache of the metadata. |
 | `meta_client.metadata_cache_ttl` | String | `10m` | TTL of the metadata cache. |
 | `meta_client.metadata_cache_tti` | String | `5m` | -- |
-| `heartbeat` | -- | -- | The heartbeat options. |
-| `heartbeat.interval` | String | `3s` | Interval for sending heartbeat messages to the metasrv. |
-| `heartbeat.retry_interval` | String | `3s` | Interval for retrying to send heartbeat messages to the metasrv. |
 | `logging` | -- | -- | The logging options. |
 | `logging.dir` | String | `./greptimedb_data/logs` | The directory to store the log files. If set to empty, logs will not be written to files. |
 | `logging.level` | String | Unset | The log level. Can be `info`/`debug`/`warn`/`error`. |
diff --git a/config/datanode.example.toml b/config/datanode.example.toml
index 2631a089e1..fbe205c17a 100644
--- a/config/datanode.example.toml
+++ b/config/datanode.example.toml
@@ -17,10 +17,7 @@ init_regions_in_background = false
 ## Parallelism of initializing regions.
 init_regions_parallelism = 16
 
-## The maximum current queries allowed to be executed. Zero means unlimited.
-## NOTE: This setting affects scan_memory_limit's privileged tier allocation.
-## When set, 70% of queries get privileged memory access (full scan_memory_limit).
-## The remaining 30% get standard tier access (70% of scan_memory_limit).
+## The maximum concurrent queries allowed to be executed. Zero means unlimited.
 max_concurrent_queries = 0
 
 ## Enable telemetry to collect anonymous usage data. Enabled by default.
@@ -83,14 +80,6 @@ watch = false
 ## The number of threads to execute the runtime for global write operations.
 #+ compact_rt_size = 4
 
-## The heartbeat options.
-[heartbeat]
-## Interval for sending heartbeat messages to the metasrv.
-interval = "3s"
-
-## Interval for retrying to send heartbeat messages to the metasrv.
-retry_interval = "3s"
-
 ## The metasrv client options.
 [meta_client]
 ## The addresses of the metasrv.
@@ -449,7 +438,9 @@ compress_manifest = false
 ## @toml2docs:none-default="Auto"
 #+ max_background_purges = 8
 
-## Memory budget for compaction tasks. Setting it to 0 or "unlimited" disables the limit.
+## Memory budget for compaction tasks.
+## Supports absolute size (e.g., "2GiB", "512MB") or percentage of system memory (e.g., "50%").
+## Setting it to 0 or "unlimited" disables the limit.
 ## @toml2docs:none-default="0"
 #+ experimental_compaction_memory_limit = "0"
 
@@ -523,9 +514,6 @@ manifest_cache_size = "256MB"
 ## Buffer size for SST writing.
 sst_write_buffer_size = "8MB"
 
-## Capacity of the channel to send data from parallel scan tasks to the main task.
-parallel_scan_channel_size = 32
-
 ## Maximum number of SST files to scan concurrently.
 max_concurrent_scan_files = 384
 
@@ -535,17 +523,21 @@ allow_stale_entries = false
 ## Memory limit for table scans across all queries.
 ## Supports absolute size (e.g., "2GB") or percentage of system memory (e.g., "20%").
 ## Setting it to 0 disables the limit.
-## NOTE: Works with max_concurrent_queries for tiered memory allocation.
-## - If max_concurrent_queries is set: 70% of queries get full access, 30% get 70% access.
-## - If max_concurrent_queries is 0 (unlimited): first 20 queries get full access, rest get 70% access.
 scan_memory_limit = "50%"
+## Controls what happens when a scan cannot get memory immediately.
+## "fail" (default) fails fast and is the recommended option for most users.
+## "wait" / "wait(<duration>)" waits for memory to become available. This is mainly
+## for advanced tuning in bursty workloads where temporary contention is common and
+## higher latency is acceptable.
+## "wait" means "wait(10s)", not unlimited waiting.
+scan_memory_on_exhausted = "fail"
 
 ## Minimum time interval between two compactions.
 ## To align with the old behavior, the default value is 0 (no restrictions).
 min_compaction_interval = "0m"
 
-## Whether to enable experimental flat format as the default format.
-default_experimental_flat_format = false
+## Whether to enable flat format as the default SST format.
+default_flat_format = true
 
 ## The options for index in Mito engine.
 [region_engine.mito.index]
diff --git a/config/flownode.example.toml b/config/flownode.example.toml
index b13acfc447..2c053e6e8c 100644
--- a/config/flownode.example.toml
+++ b/config/flownode.example.toml
@@ -2,6 +2,11 @@
 ## @toml2docs:none-default
 node_id = 14
 
+## The user provider for authentication.
+## Examples: "static_user_provider:file:/path/to/users", "static_user_provider:cmd:greptime_user=greptime_pwd"
+## @toml2docs:none-default
+#+ user_provider = "static_user_provider:file:/path/to/users"
+
 ## flow engine options.
 [flow]
 ## The number of flow worker in flownode.
@@ -22,10 +27,6 @@ node_id = 14
 ## if failed to find available frontend after frontend_scan_timeout elapsed, return error
 ## which prevent flownode from starting
 #+experimental_frontend_scan_timeout="30s"
-## Frontend activity timeout
-## if frontend is down(not sending heartbeat) for more than frontend_activity_timeout,
-## it will be removed from the list that flownode use to connect
-#+experimental_frontend_activity_timeout="60s"
 ## Maximum number of filters allowed in a single query
 #+experimental_max_filter_num_per_query=20
 ## Time window merge distance
@@ -96,14 +97,6 @@ metadata_cache_ttl = "10m"
 # TTI of the metadata cache.
 metadata_cache_tti = "5m"
 
-## The heartbeat options.
-[heartbeat]
-## Interval for sending heartbeat messages to the metasrv.
-interval = "3s"
-
-## Interval for retrying to send heartbeat messages to the metasrv.
-retry_interval = "3s"
-
 ## The logging options.
 [logging]
 ## The directory to store the log files. If set to empty, logs will not be written to files.
diff --git a/config/frontend.example.toml b/config/frontend.example.toml
index 435504b122..60115e93bb 100644
--- a/config/frontend.example.toml
+++ b/config/frontend.example.toml
@@ -6,6 +6,11 @@ default_timezone = "UTC"
 ## @toml2docs:none-default
 default_column_prefix = "greptime"
 
+## The user provider for authentication.
+## Examples: "static_user_provider:file:/path/to/users", "static_user_provider:cmd:greptime_user=greptime_pwd"
+## @toml2docs:none-default
+#+ user_provider = "static_user_provider:file:/path/to/users"
+
 ## Maximum total memory for all concurrent write request bodies and messages (HTTP, gRPC, Flight).
 ## Set to 0 to disable the limit. Default: "0" (unlimited)
 ## @toml2docs:none-default
@@ -23,14 +28,6 @@ default_column_prefix = "greptime"
 ## The number of threads to execute the runtime for global write operations.
 #+ compact_rt_size = 4
 
-## The heartbeat options.
-[heartbeat]
-## Interval for sending heartbeat messages to the metasrv.
-interval = "18s"
-
-## Interval for retrying to send heartbeat messages to the metasrv.
-retry_interval = "3s"
-
 ## The HTTP server options.
 [http]
 ## The address to bind the HTTP server.
@@ -100,7 +97,7 @@ watch = false
 bind_addr = "127.0.0.1:4010"
 ## The address advertised to the metasrv, and used for connections from outside the host.
 ## If left empty or unset, the server will automatically use the IP address of the first network interface
-## on the host, with the same port number as the one specified in `grpc.bind_addr`.
+## on the host, with the same port number as the one specified in `internal_grpc.bind_addr`.
 server_addr = "127.0.0.1:4010"
 ## The number of server worker threads.
 runtime_size = 8
@@ -214,6 +211,17 @@ enable = true
 enable = true
 ## Whether to store the data from Prometheus remote write in metric engine.
 with_metric_engine = true
+## Interval to flush pending rows batcher.
+## Set to "0s" to disable batching mode in Prometheus Remote Write endpoint
+#+pending_rows_flush_interval = "0s"
+## Max rows per pending batch before triggering a flush.
+#+max_batch_rows = 100000
+## Max number of concurrent batch flushes.
+#+max_concurrent_flushes = 256
+## Capacity of the pending batch worker channel.
+#+worker_channel_capacity = 65526
+## Max inflight write requests before backpressure.
+#+max_inflight_requests = 3000
 
 ## The metasrv client options.
 [meta_client]
diff --git a/config/metasrv.example.toml b/config/metasrv.example.toml
index fed6e3b814..8e27e5be26 100644
--- a/config/metasrv.example.toml
+++ b/config/metasrv.example.toml
@@ -79,7 +79,7 @@ node_max_idle_time = "24hours"
 ## The frontend heartbeat interval is 6 times of the base heartbeat interval.
 ## The flownode/datanode heartbeat interval is 1 times of the base heartbeat interval.
 ## e.g., If the base heartbeat interval is 3s, the frontend heartbeat interval is 18s, the flownode/datanode heartbeat interval is 3s.
-## If you change this value, you need to change the heartbeat interval of the flownode/frontend/datanode accordingly.
+## Heartbeat intervals are negotiated from metasrv during handshake; local node configs do not override this.
 #+ heartbeat_interval = "3s"
 
 ## Whether to enable greptimedb telemetry. Enabled by default.
@@ -136,7 +136,7 @@ ca_cert_path = ""
 bind_addr = "127.0.0.1:3002"
 ## The communication server address for the frontend and datanode to connect to metasrv.
 ## If left empty or unset, the server will automatically use the IP address of the first network interface
-## on the host, with the same port number as the one specified in `bind_addr`.
+## on the host, with the same port number as the one specified in `grpc.bind_addr`.
 server_addr = "127.0.0.1:3002"
 ## The number of server worker threads.
 runtime_size = 8
diff --git a/config/standalone.example.toml b/config/standalone.example.toml
index ef96406316..aa745bb6ba 100644
--- a/config/standalone.example.toml
+++ b/config/standalone.example.toml
@@ -6,6 +6,11 @@ default_timezone = "UTC"
 ## @toml2docs:none-default
 default_column_prefix = "greptime"
 
+## The user provider for authentication.
+## Examples: "static_user_provider:file:/path/to/users", "static_user_provider:cmd:greptime_user=greptime_pwd"
+## @toml2docs:none-default
+#+ user_provider = "static_user_provider:file:/path/to/users"
+
 ## Maximum total memory for all concurrent write request bodies and messages (HTTP, gRPC, Flight).
 ## Set to 0 to disable the limit. Default: "0" (unlimited)
 ## @toml2docs:none-default
@@ -23,10 +28,7 @@ init_regions_in_background = false
 ## Parallelism of initializing regions.
 init_regions_parallelism = 16
 
-## The maximum current queries allowed to be executed. Zero means unlimited.
-## NOTE: This setting affects scan_memory_limit's privileged tier allocation.
-## When set, 70% of queries get privileged memory access (full scan_memory_limit).
-## The remaining 30% get standard tier access (70% of scan_memory_limit).
+## The maximum concurrent queries allowed to be executed. Zero means unlimited.
 max_concurrent_queries = 0
 
 ## Enable telemetry to collect anonymous usage data. Enabled by default.
@@ -176,6 +178,17 @@ enable = true
 enable = true
 ## Whether to store the data from Prometheus remote write in metric engine.
 with_metric_engine = true
+## Interval to flush pending rows batcher.
+## Set to "0s" to disable batching mode in Prometheus Remote Write endpoint
+#+pending_rows_flush_interval = "0s"
+## Max rows per pending batch before triggering a flush.
+#+max_batch_rows = 100000
+## Max number of concurrent batch flushes.
+#+max_concurrent_flushes = 256
+## Capacity of the pending batch worker channel.
+#+worker_channel_capacity = 65526
+## Max inflight write requests before backpressure.
+#+max_inflight_requests = 3000
 
 ## The WAL options.
 [wal]
@@ -541,7 +554,9 @@ compress_manifest = false
 ## @toml2docs:none-default="Auto"
 #+ max_background_purges = 8
 
-## Memory budget for compaction tasks. Setting it to 0 or "unlimited" disables the limit.
+## Memory budget for compaction tasks.
+## Supports absolute size (e.g., "2GiB", "512MB") or percentage of system memory (e.g., "50%").
+## Setting it to 0 or "unlimited" disables the limit.
 ## @toml2docs:none-default="0"
 #+ experimental_compaction_memory_limit = "0"
 
@@ -615,9 +630,6 @@ manifest_cache_size = "256MB"
 ## Buffer size for SST writing.
 sst_write_buffer_size = "8MB"
 
-## Capacity of the channel to send data from parallel scan tasks to the main task.
-parallel_scan_channel_size = 32
-
 ## Maximum number of SST files to scan concurrently.
 max_concurrent_scan_files = 384
 
@@ -627,17 +639,21 @@ allow_stale_entries = false
 ## Memory limit for table scans across all queries.
 ## Supports absolute size (e.g., "2GB") or percentage of system memory (e.g., "20%").
 ## Setting it to 0 disables the limit.
-## NOTE: Works with max_concurrent_queries for tiered memory allocation.
-## - If max_concurrent_queries is set: 70% of queries get full access, 30% get 70% access.
-## - If max_concurrent_queries is 0 (unlimited): first 20 queries get full access, rest get 70% access.
 scan_memory_limit = "50%"
+## Controls what happens when a scan cannot get memory immediately.
+## "fail" (default) fails fast and is the recommended option for most users.
+## "wait" / "wait(<duration>)" waits for memory to become available. This is mainly
+## for advanced tuning in bursty workloads where temporary contention is common and
+## higher latency is acceptable.
+## "wait" means "wait(10s)", not unlimited waiting.
+scan_memory_on_exhausted = "fail"
 
 ## Minimum time interval between two compactions.
 ## To align with the old behavior, the default value is 0 (no restrictions).
 min_compaction_interval = "0m"
 
-## Whether to enable experimental flat format as the default format.
-default_experimental_flat_format = false
+## Whether to enable flat format as the default SST format.
+default_flat_format = true
 
 ## The options for index in Mito engine.
 [region_engine.mito.index]
diff --git a/docker/dev-builder/centos/Dockerfile b/docker/dev-builder/centos/Dockerfile
index 8b29bb3065..344bc1ec2a 100644
--- a/docker/dev-builder/centos/Dockerfile
+++ b/docker/dev-builder/centos/Dockerfile
@@ -7,7 +7,7 @@ RUN sed -i s/mirror.centos.org/vault.centos.org/g /etc/yum.repos.d/*.repo
 RUN sed -i s/^#.*baseurl=http/baseurl=http/g /etc/yum.repos.d/*.repo
 
 # Install dependencies
-RUN ulimit -n 1024000 && yum groupinstall -y 'Development Tools'
+RUN yum groupinstall -y 'Development Tools'
 RUN yum install -y epel-release  \
     openssl \
     openssl-devel  \
diff --git a/docker/docker-compose/cluster-with-etcd.yaml b/docker/docker-compose/cluster-with-etcd.yaml
index 34eee7f4b5..39f421e637 100644
--- a/docker/docker-compose/cluster-with-etcd.yaml
+++ b/docker/docker-compose/cluster-with-etcd.yaml
@@ -85,8 +85,8 @@ services:
     command:
       - metasrv
       - start
-      - --rpc-bind-addr=0.0.0.0:3002
-      - --rpc-server-addr=metasrv:3002
+      - --grpc-bind-addr=0.0.0.0:3002
+      - --grpc-server-addr=metasrv:3002
       - --store-addrs=etcd0:2379
       - --http-addr=0.0.0.0:3000
     healthcheck:
@@ -111,8 +111,8 @@ services:
       - start
       - --node-id=0
       - --data-home=/greptimedb_data
-      - --rpc-bind-addr=0.0.0.0:3001
-      - --rpc-server-addr=datanode0:3001
+      - --grpc-bind-addr=0.0.0.0:3001
+      - --grpc-server-addr=datanode0:3001
       - --metasrv-addrs=metasrv:3002
       - --http-addr=0.0.0.0:5000
     volumes:
@@ -141,7 +141,7 @@ services:
       - start
       - --metasrv-addrs=metasrv:3002
       - --http-addr=0.0.0.0:4000
-      - --rpc-bind-addr=0.0.0.0:4001
+      - --grpc-bind-addr=0.0.0.0:4001
       - --mysql-addr=0.0.0.0:4002
       - --postgres-addr=0.0.0.0:4003
     healthcheck:
@@ -166,8 +166,8 @@ services:
       - start
       - --node-id=0
       - --metasrv-addrs=metasrv:3002
-      - --rpc-bind-addr=0.0.0.0:4004
-      - --rpc-server-addr=flownode0:4004
+      - --grpc-bind-addr=0.0.0.0:4004
+      - --grpc-server-addr=flownode0:4004
       - --http-addr=0.0.0.0:4005
     depends_on:
       frontend0:
diff --git a/docs/rfcs/2025-12-30-export-import-v2.md b/docs/rfcs/2025-12-30-export-import-v2.md
index 197eb7cc9d..6bc8428300 100644
--- a/docs/rfcs/2025-12-30-export-import-v2.md
+++ b/docs/rfcs/2025-12-30-export-import-v2.md
@@ -67,6 +67,7 @@ snapshot-20250101/
 - Self-contained (all information needed for restore)
 - Immutable (content never changes after creation)
 - Verifiable (checksums at file, chunk, and snapshot levels)
+- Schema-only snapshots contain only `manifest.json` and `schema/`; `data/` is absent, `chunks` is empty, and later data append is rejected (use `--force` to recreate)
 
 ### Chunk
 
@@ -116,6 +117,8 @@ greptime export create \
   --schema-only \
   --to s3://my-bucket/snapshots/prod-schema-only
 
+Schema-only snapshots cannot be resumed with data; use `--force` to recreate.
+
 # Export with specific format (default: parquet)
 greptime export create \
   --format csv \
@@ -173,7 +176,9 @@ The manifest is a JSON file containing snapshot metadata and chunk index:
 - `snapshot_id`: Unique identifier (UUID)
 - `catalog`, `schemas`: Catalog and schema list
 - `time_range`: Overall time range covered
+- `schema_only`: Whether the snapshot contains schema only
 - `chunks[]`: Array of chunk metadata
+- `format`: Data format for exported files
 - `checksum`: Snapshot-level SHA256 checksum
 
 **Chunk metadata structure**:
@@ -182,7 +187,7 @@ Each chunk entry in the manifest contains:
 
 - `id`: Chunk identifier (sequential number)
 - `time_range`: Start and end timestamps
-- `status`: Export status (Pending, Completed, Failed)
+- `status`: Export status (Pending, InProgress, Completed, Failed)
 - `files`: List of data files in the chunk directory
 - `checksum`: Chunk-level checksum for integrity verification
 
@@ -292,9 +297,9 @@ Checksums are verified during import before data is written to the database.
 
 **Resume capability**:
 
-- Manifest tracks chunk status (Pending, Completed, Failed)
+- Manifest tracks chunk status (Pending, InProgress, Completed, Failed)
 - Export/import automatically resumes when executed on existing snapshot
-- Skips completed chunks, retries failed chunks, processes pending chunks
+- Skips completed chunks, retries failed/in-progress chunks, processes pending chunks
 - Works across process restarts
 - Use `--force` (export only) to delete existing snapshot and start over
 
diff --git a/docs/rfcs/2026-03-16-flow-inc-query.md b/docs/rfcs/2026-03-16-flow-inc-query.md
new file mode 100644
index 0000000000..8041d37d2b
--- /dev/null
+++ b/docs/rfcs/2026-03-16-flow-inc-query.md
@@ -0,0 +1,190 @@
+---
+Feature Name: Flow Batching Sequence-Based Incremental Query Plan (Lite)
+Tracking Issue: TBD
+Date: 2026-03-16
+Author: @discord9
+---
+
+# Summary
+
+This RFC proposes a correctness-first incremental query mode for Flow batching.
+Flow queries can read only `seq > checkpoint` and advance checkpoints using per-region correctness watermarks.
+When incremental reads are stale or correctness cannot be proven, Flow falls back to full recomputation.
+
+# Motivation
+
+Flow batching still needs to repeatedly compute old data in the same time window, so incremental query can improve Flow performance.
+
+# Goals
+
+1. Add opt-in incremental reads (`seq > given_seq`) for Flow.
+2. Return per-region correctness watermarks for checkpoint advancement.
+3. Keep existing query behavior unchanged unless explicitly enabled.
+4. Define deterministic fallback for stale or unprovable incremental reads.
+
+# Non-Goals
+
+1. No business-schema changes (no synthetic watermark columns in result rows).
+2. No global throughput optimization in v1 (correctness first).
+3. No observational watermark output when correctness is unprovable.
+
+# Proposal
+
+## 1) Query options
+
+Introduce three `QueryContext` extension keys:
+
+- `flow.incremental_after_seqs`
+- `flow.incremental_mode`
+- `flow.return_region_seq`
+
+These options are opt-in and only affect Flow incremental execution paths.
+
+## 2) Scan mapping
+
+When incremental mode is enabled:
+
+- map `after_seq` to `memtable_min_sequence` (exclusive lower bound)
+- keep existing snapshot upper-bound behavior (`memtable_max_sequence`)
+
+Important limitation in v1:
+
+- incremental filtering is correctness-proven only for memtable rows
+- SST files do not preserve detailed row-level sequence metadata; they only expose coarser file-level sequence information
+- therefore `seq > checkpoint` must not assume precise incremental pruning across memtable->SST flush boundaries
+
+If required incremental parameters are missing or invalid, return argument error.
+
+## 3) Stale protection
+
+Add dedicated stale error:
+
+- `IncrementalQueryStale { region_id, given_seq, min_readable_seq }`
+
+Behavior:
+
+- if `given_seq < min_readable_seq`, return stale error
+- if `given_seq == min_readable_seq`, query is valid and reads `seq > given_seq`
+- if `given_seq > min_readable_seq`, query is also valid and reads `seq > given_seq`
+
+`IncrementalQueryStale` also covers the case where rows newer than the checkpoint have crossed a memtable->SST flush boundary and sequence-precise incremental exclusion can no longer be proven.
+In other words, the flush-boundary case is not a separate fallback category in v1; it is one concrete way an incremental cursor becomes stale.
+
+## 4) Watermark return
+
+Extend query metrics with optional per-region watermark map:
+
+- `region_latest_sequences: Vec<(region_id: u64, latest_sequence: u64)>`
+
+Rules:
+
+- only terminal metrics of successful query can advance checkpoints
+- for multi-region query, watermark must be complete map or absent
+- if correctness is unprovable, business rows may return but watermark is absent
+
+## 5) Flow state machine
+
+Checkpoint and watermark state are kept only in flownode memory in v1; they are not persisted as durable flow metadata.
+Cold start or flownode restart therefore always re-enters through a full snapshot read.
+Only after that full query succeeds with a complete correctness watermark may Flow switch back to incremental mode.
+
+Flow starts in full mode, then transitions:
+
+1. Full query succeeds with correctness watermark -> enter incremental mode
+2. Incremental query succeeds with correctness watermark -> advance checkpoint
+3. Incremental stale/failure -> fallback to full mode
+4. Full query without correctness watermark -> remain in full mode
+
+```mermaid
+stateDiagram-v2
+    [*] --> FullSnapshot: Flow starts
+
+    state FullSnapshot {
+        [*] --> RunFull
+        RunFull --> RunFull: Full query succeeds but watermark is unprovable<br/>no region_latest_sequences returned
+    }
+
+    FullSnapshot --> Incremental: Full query succeeds and correctness watermark is returned<br/>(checkpoint updated)
+
+    state Incremental {
+        [*] --> RunInc
+        RunInc --> RunInc: Incremental succeeds<br/>(checkpoint advances)
+    }
+
+    Incremental --> FullSnapshot: IncrementalQueryStale<br/>(cursor too old, fallback required)
+    Incremental --> FullSnapshot: Incremental fails<br/> and fallback policy is triggered
+
+    FullSnapshot --> [*]: Flow stops
+    Incremental --> [*]: Flow stops
+```
+
+### Fallback Policy
+
+Fallback to full mode is deterministic and is triggered by any of the following:
+
+1. `IncrementalQueryStale` is returned.
+2. Incremental query fails with execution errors.
+3. Incremental query succeeds but watermark is absent or incomplete for participating regions.
+
+Policy behavior:
+
+1. Do not advance any checkpoint in the failed/incomplete round.
+2. Switch to full mode for the affected flow/window in the next round.
+3. Return to incremental mode only after a full query succeeds with a complete correctness watermark map.
+
+### Persistence and recovery model
+
+The v1 design is intentionally correctness-first and keeps the progress cursor lightweight:
+
+1. Watermarks/checkpoints live only in flownode memory; v1 does not persist them separately.
+2. On cold start, the flow re-establishes progress by running a successful full-query snapshot read, then resumes incremental mode only after that round returns a complete correctness watermark map.
+3. Sequence-precise incremental correctness is currently limited to rows still visible in memtables.
+4. Once relevant rows have been flushed into SST, the system cannot use `seq > checkpoint` alone to prove precise incremental exclusion, because SST lacks detailed row-level sequence metadata.
+5. In that case the correct behavior is to fall back to full recomputation, not to continue a best-effort incremental scan.
+
+# Distributed and Compatibility Requirements
+
+1. Distributed path must preserve region-level snapshot/read-bound semantics end-to-end.
+2. `snapshot_seqs` transport and `flow.*` options must both be carried correctly.
+   - `snapshot_seqs` means the per-region snapshot upper-bound map: `region_id -> sequence`.
+3. New metrics fields must be backward-compatible (old clients ignore unknown fields).
+
+# Rollout Plan
+
+## Phase 1 (MVP, correctness first)
+
+1. Add extension constants and parsing.
+2. Add incremental scan mapping and stale detection.
+3. Add watermark metrics field and terminal-watermark checkpoint update path.
+4. Complete standalone and distributed passthrough.
+
+## Phase 2 (performance and observability)
+
+1. Improve batching key strategy with sequence/watermark context.
+2. Optimize watermark serialization overhead.
+3. Add metrics: incremental hit rate, fallback rate, fallback window size.
+
+# Testing Plan
+
+1. Unit tests for incremental bounds and stale detection.
+2. Query-path tests for extension mapping and watermark semantics.
+3. Flow integration tests for full->incremental->fallback transitions.
+4. Distributed tests for end-to-end snapshot/watermark propagation.
+5. Compatibility tests for old/new client-server combinations.
+
+# Risks
+
+1. Boundary semantic mismatch (`<` vs `<=`) may cause correctness bugs.
+2. Incomplete distributed propagation can silently invalidate watermark safety.
+3. Frequent fallback can reduce throughput before phase-2 optimizations.
+4. Memtable->SST flushes may force more full recomputation than expected until finer-grained SST sequence tracking exists.
+
+# Alternatives
+
+1. Put watermark into business rows (rejected: schema pollution).
+2. Add new dedicated Flight message type in v1 (deferred to reduce scope).
+
+# Conclusion
+
+This plan enables a practical, correctness-first incremental path for Flow batching.
+It reuses existing sequence scan capability, adds strict stale handling, and advances checkpoints only from correctness-proven per-region watermarks.
diff --git a/docs/rfcs/remote-dyn-filter-rfc.md b/docs/rfcs/remote-dyn-filter-rfc.md
new file mode 100644
index 0000000000..5f2243b15c
--- /dev/null
+++ b/docs/rfcs/remote-dyn-filter-rfc.md
@@ -0,0 +1,136 @@
+---
+Feature Name: Remote Dynamic Filter Propagation
+Tracking Issue: N/A
+Date: 2026-04-08
+Author: @discord9
+---
+
+# RFC: Remote Dynamic Filter Propagation
+
+# Summary
+
+This RFC proposes a remote dynamic filter propagation mechanism for distributed queries. It lets frontend-produced dynamic filters reach remote datanode scans through a lightweight control plane, while preserving one rule: remote dynamic filters are an optimization only, never a correctness dependency.
+
+# Motivation
+
+Today, dynamic filters can improve local execution, but they do not automatically propagate to remote datanode scans in distributed queries. As a result, the frontend may already know that a probe-side scan can be narrowed, while the remote scan still runs with a weaker predicate and loses pruning opportunities.
+
+We want a minimal design that:
+
+- propagates dynamic filter updates to remote scans,
+- keeps filter identity and lifecycle stable across register/update/unregister,
+- and safely degrades when encoding, routing, RPC, or apply logic fails.
+
+# Details
+
+The high-level flow is:
+
+1. A join on the frontend produces an alive dynamic filter.
+2. `MergeScanExec` identifies the remote subscribers, generates a stable `filter_id`, and registers the alive filter into a query-scoped registry.
+3. The initial remote read establishes the corresponding registration on the datanode side.
+4. The frontend registry watches for dynamic filter updates via `wait_update` or generation changes.
+5. Later updates and unregister messages are sent through the existing region unary RPC path.
+6. The datanode applies these updates to query-scoped remote filter state and scan wrappers.
+7. Query finish, cancel, or no-consumer conditions trigger unregister and cleanup.
+
+## Identity
+
+The logical identity of a remote dynamic filter is `query_id + filter_id`.
+
+Region and scan metadata are routing information, not part of filter-state identity. `filter_id` only needs to be stable and unique within a single query.
+
+The current recommendation is to derive `filter_id` from:
+
+- `region_id`
+- `producer-local ordinal`
+- `canonicalized children fingerprint`
+
+The following should not be included:
+
+- `partition`
+- transport metadata
+- memory addresses or temporary runtime object ids
+
+## Transport
+
+This design reuses the existing region unary gRPC path:
+
+- `RegionRequest.body.remote_dyn_filter`
+- `RemoteDynFilterRequest.oneof action`
+  - `update`
+  - `unregister`
+
+The initial remote read is responsible for register and scan setup. The unary RPC path is only for later `update` and `unregister` messages.
+
+## Frontend registry
+
+The frontend uses a query-engine runtime map:
+
+- implementation near `src/query/src/dist_plan/remote_dyn_filter_registry.rs`
+- storage model: `query_id -> Arc<RemoteDynFilterRegistry>`
+
+This registry should not live on a single `MergeScanExec`, and it should not be stored in `QueryContext.mutable_session_data`. It is a query execution runtime object that owns watcher tasks, cleanup tail, and fanout state.
+
+The registry lifecycle has three states:
+
+- `Active`: accepts registrations and sends updates
+- `Closing`: query ended; stop new registrations, send final cleanup messages, drain in-flight RPCs
+- `Closed`: watchers stopped, state removable from the runtime map
+
+The registry may outlive the main query execution briefly for cleanup, but it is not intended to be a long-lived global object.
+
+## Propagation policy
+
+Remote dynamic filters should remain a selective optimization, not an automatic fanout for every filter update.
+
+The frontend may skip remote propagation when the encoded filter becomes too large, fanout cost is too high, or the expected pruning benefit is too small. In those cases, execution should continue with local-only dynamic filtering semantics.
+
+## Responsibilities
+
+On the frontend:
+
+- the join produces alive dynamic filters,
+- `MergeScanExec` bridges producers to remote subscribers,
+- the registry watches updates and fans out RPCs.
+
+On the datanode:
+
+- the unary handler receives `update` and `unregister`,
+- query-scoped remote filter state is keyed by `query_id + filter_id`,
+- remote wrappers apply updates through existing predicate and scan refresh paths.
+
+## Failure semantics
+
+All failures must degrade safely:
+
+- encode failure -> local-only filter
+- RPC failure -> log/metric and degrade
+- early update or missing target -> explicit buffer, drop+metric, or retry policy
+- decode or remap failure -> disable remote optimization only
+
+# Alternatives
+
+## Registry on `MergeScanExec`
+
+Rejected because lifecycle and cleanup would become fragmented across multiple bridge or exec instances in the same query.
+
+## Registry in `QueryContext.mutable_session_data`
+
+Rejected because this is the wrong ownership model. The registry is not session metadata; it is a query runtime object with watcher tasks and cleanup behavior.
+
+## Long-lived global manager
+
+Rejected for now because it is heavier than necessary. A query-engine runtime map is sufficient for the current design.
+
+# Drawbacks
+
+- The design introduces extra query runtime state and cleanup logic on both frontend and datanode.
+- The initial version only covers the current minimal filter form and leaves larger membership propagation to later work.
+- A clear policy is still required for updates that arrive before scan registration.
+
+# Unresolved questions
+
+1. Should children fingerprint canonicalization become a shared helper?
+2. What is the strict timing relationship between `is_complete` and final unregister?
+3. Does the runtime map need a background sweep task, or is explicit reap enough?
+4. How should large build-side membership evolve beyond `IN` in later work?
diff --git a/flake.lock b/flake.lock
index bec6e18d9a..3c3ee4bd67 100644
--- a/flake.lock
+++ b/flake.lock
@@ -8,11 +8,11 @@
         "rust-analyzer-src": "rust-analyzer-src"
       },
       "locked": {
-        "lastModified": 1770794449,
-        "narHash": "sha256-1nFkhcZx9+Sdw5OXwJqp5TxvGncqRqLeK781v0XV3WI=",
+        "lastModified": 1774250935,
+        "narHash": "sha256-mWID0WFgTnd9hbEeaPNX+YYWF70JN3r7zBouEqERJOE=",
         "owner": "nix-community",
         "repo": "fenix",
-        "rev": "b19d93fdf9761e6101f8cb5765d638bacebd9a1b",
+        "rev": "64d7705e8c37d650cfb1aa99c24a8ce46597f29e",
         "type": "github"
       },
       "original": {
@@ -41,11 +41,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1770617025,
-        "narHash": "sha256-1jZvgZoAagZZB6NwGRv2T2ezPy+X6EFDsJm+YSlsvEs=",
+        "lastModified": 1774244481,
+        "narHash": "sha256-4XfMXU0DjN83o6HWZoKG9PegCvKvIhNUnRUI19vzTcQ=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "2db38e08fdadcc0ce3232f7279bab59a15b94482",
+        "rev": "4590696c8693fea477850fe379a01544293ca4e2",
         "type": "github"
       },
       "original": {
@@ -65,11 +65,11 @@
     "rust-analyzer-src": {
       "flake": false,
       "locked": {
-        "lastModified": 1770702974,
-        "narHash": "sha256-CbvWu72rpGHK5QynoXwuOnVzxX7njF2LYgk8wRSiAQ0=",
+        "lastModified": 1774221325,
+        "narHash": "sha256-aEIdkqB8gtQZtEbogdUb5iyfcZpKIlD3FkG8ANu73/I=",
         "owner": "rust-lang",
         "repo": "rust-analyzer",
-        "rev": "07a594815f7c1d6e7e39f21ddeeedb75b21795f4",
+        "rev": "b42b63f390a4dab14e6efa34a70e67f5b087cc62",
         "type": "github"
       },
       "original": {
diff --git a/flake.nix b/flake.nix
index 6a02f4f05f..8dc84136a0 100644
--- a/flake.nix
+++ b/flake.nix
@@ -20,7 +20,7 @@
         lib = nixpkgs.lib;
         rustToolchain = fenix.packages.${system}.fromToolchainName {
           name = (lib.importTOML ./rust-toolchain.toml).toolchain.channel;
-          sha256 = "sha256-GCGEXGZeJySLND0KU5TdtTrqFV76TF3UdvAHSUegSsk=";
+          sha256 = "sha256-rboGKQLH4eDuiY01SINOqmXUFUNr9F4awoFZGzib17o=";
         };
       in
       {
diff --git a/grafana/dashboards/metrics/cluster/dashboard.json b/grafana/dashboards/metrics/cluster/dashboard.json
index 918eb3c4e8..d56df06a3b 100644
--- a/grafana/dashboards/metrics/cluster/dashboard.json
+++ b/grafana/dashboards/metrics/cluster/dashboard.json
@@ -25,7 +25,7 @@
   "editable": true,
   "fiscalYearStartMonth": 0,
   "graphTooltip": 1,
-  "id": 34,
+  "id": 33,
   "links": [],
   "panels": [
     {
@@ -927,7 +927,7 @@
       "type": "stat"
     },
     {
-      "collapsed": true,
+      "collapsed": false,
       "gridPos": {
         "h": 1,
         "w": 24,
@@ -935,226 +935,226 @@
         "y": 9
       },
       "id": 275,
-      "panels": [
+      "panels": [],
+      "title": "Ingestion",
+      "type": "row"
+    },
+    {
+      "datasource": {
+        "type": "prometheus",
+        "uid": "${metrics}"
+      },
+      "description": "Total ingestion rate.\n\nHere we listed 3 primary protocols:\n\n- Prometheus remote write\n- Greptime's gRPC API (when using our ingest SDK)\n- Log ingestion http API\n",
+      "fieldConfig": {
+        "defaults": {
+          "color": {
+            "mode": "palette-classic"
+          },
+          "custom": {
+            "axisBorderShow": false,
+            "axisCenteredZero": false,
+            "axisColorMode": "text",
+            "axisLabel": "",
+            "axisPlacement": "auto",
+            "barAlignment": 0,
+            "barWidthFactor": 0.6,
+            "drawStyle": "line",
+            "fillOpacity": 0,
+            "gradientMode": "none",
+            "hideFrom": {
+              "legend": false,
+              "tooltip": false,
+              "viz": false
+            },
+            "insertNulls": false,
+            "lineInterpolation": "linear",
+            "lineWidth": 1,
+            "pointSize": 5,
+            "scaleDistribution": {
+              "type": "linear"
+            },
+            "showPoints": "auto",
+            "spanNulls": false,
+            "stacking": {
+              "group": "A",
+              "mode": "none"
+            },
+            "thresholdsStyle": {
+              "mode": "off"
+            }
+          },
+          "mappings": [],
+          "thresholds": {
+            "mode": "absolute",
+            "steps": [
+              {
+                "color": "green"
+              },
+              {
+                "color": "red",
+                "value": 80
+              }
+            ]
+          },
+          "unit": "rowsps"
+        },
+        "overrides": []
+      },
+      "gridPos": {
+        "h": 6,
+        "w": 24,
+        "x": 0,
+        "y": 10
+      },
+      "id": 193,
+      "options": {
+        "legend": {
+          "calcs": [
+            "mean"
+          ],
+          "displayMode": "table",
+          "placement": "bottom",
+          "showLegend": true
+        },
+        "tooltip": {
+          "hideZeros": false,
+          "mode": "single",
+          "sort": "none"
+        }
+      },
+      "pluginVersion": "11.6.0",
+      "targets": [
         {
           "datasource": {
             "type": "prometheus",
             "uid": "${metrics}"
           },
-          "description": "Total ingestion rate.\n\nHere we listed 3 primary protocols:\n\n- Prometheus remote write\n- Greptime's gRPC API (when using our ingest SDK)\n- Log ingestion http API\n",
-          "fieldConfig": {
-            "defaults": {
-              "color": {
-                "mode": "palette-classic"
-              },
-              "custom": {
-                "axisBorderShow": false,
-                "axisCenteredZero": false,
-                "axisColorMode": "text",
-                "axisLabel": "",
-                "axisPlacement": "auto",
-                "barAlignment": 0,
-                "drawStyle": "line",
-                "fillOpacity": 0,
-                "gradientMode": "none",
-                "hideFrom": {
-                  "legend": false,
-                  "tooltip": false,
-                  "viz": false
-                },
-                "insertNulls": false,
-                "lineInterpolation": "linear",
-                "lineWidth": 1,
-                "pointSize": 5,
-                "scaleDistribution": {
-                  "type": "linear"
-                },
-                "showPoints": "auto",
-                "spanNulls": false,
-                "stacking": {
-                  "group": "A",
-                  "mode": "none"
-                },
-                "thresholdsStyle": {
-                  "mode": "off"
-                }
-              },
-              "mappings": [],
-              "thresholds": {
-                "mode": "absolute",
-                "steps": [
-                  {
-                    "color": "green"
-                  },
-                  {
-                    "color": "red",
-                    "value": 80
-                  }
-                ]
-              },
-              "unit": "rowsps"
-            },
-            "overrides": []
+          "editorMode": "code",
+          "expr": "sum(rate(greptime_table_operator_ingest_rows{instance=~\"$frontend\"}[$__rate_interval]))",
+          "hide": false,
+          "instant": false,
+          "legendFormat": "ingestion",
+          "range": true,
+          "refId": "C"
+        }
+      ],
+      "title": "Total Ingestion Rate",
+      "type": "timeseries"
+    },
+    {
+      "datasource": {
+        "type": "prometheus",
+        "uid": "${metrics}"
+      },
+      "description": "Total ingestion rate.\n\nHere we listed 3 primary protocols:\n\n- Prometheus remote write\n- Greptime's gRPC API (when using our ingest SDK)\n- Log ingestion http API\n",
+      "fieldConfig": {
+        "defaults": {
+          "color": {
+            "mode": "palette-classic"
           },
-          "gridPos": {
-            "h": 6,
-            "w": 24,
-            "x": 0,
-            "y": 18
-          },
-          "id": 193,
-          "options": {
-            "legend": {
-              "calcs": [
-                "mean"
-              ],
-              "displayMode": "table",
-              "placement": "bottom",
-              "showLegend": true
+          "custom": {
+            "axisBorderShow": false,
+            "axisCenteredZero": false,
+            "axisColorMode": "text",
+            "axisLabel": "",
+            "axisPlacement": "auto",
+            "barAlignment": 0,
+            "barWidthFactor": 0.6,
+            "drawStyle": "line",
+            "fillOpacity": 0,
+            "gradientMode": "none",
+            "hideFrom": {
+              "legend": false,
+              "tooltip": false,
+              "viz": false
             },
-            "tooltip": {
-              "hideZeros": false,
-              "mode": "single",
-              "sort": "none"
+            "insertNulls": false,
+            "lineInterpolation": "linear",
+            "lineWidth": 1,
+            "pointSize": 5,
+            "scaleDistribution": {
+              "type": "linear"
+            },
+            "showPoints": "auto",
+            "spanNulls": false,
+            "stacking": {
+              "group": "A",
+              "mode": "none"
+            },
+            "thresholdsStyle": {
+              "mode": "off"
             }
           },
-          "pluginVersion": "12.0.0",
-          "targets": [
-            {
-              "datasource": {
-                "type": "prometheus",
-                "uid": "${metrics}"
+          "mappings": [],
+          "thresholds": {
+            "mode": "absolute",
+            "steps": [
+              {
+                "color": "green"
               },
-              "editorMode": "code",
-              "expr": "sum(rate(greptime_table_operator_ingest_rows{instance=~\"$frontend\"}[$__rate_interval]))",
-              "hide": false,
-              "instant": false,
-              "legendFormat": "ingestion",
-              "range": true,
-              "refId": "C"
-            }
+              {
+                "color": "red",
+                "value": 80
+              }
+            ]
+          },
+          "unit": "rowsps"
+        },
+        "overrides": []
+      },
+      "gridPos": {
+        "h": 6,
+        "w": 24,
+        "x": 0,
+        "y": 16
+      },
+      "id": 277,
+      "options": {
+        "legend": {
+          "calcs": [
+            "mean"
           ],
-          "title": "Total Ingestion Rate",
-          "type": "timeseries"
+          "displayMode": "table",
+          "placement": "bottom",
+          "showLegend": true
+        },
+        "tooltip": {
+          "hideZeros": false,
+          "mode": "single",
+          "sort": "none"
+        }
+      },
+      "pluginVersion": "11.6.0",
+      "targets": [
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "editorMode": "code",
+          "expr": "sum(rate(greptime_servers_http_logs_ingestion_counter[$__rate_interval]))",
+          "hide": false,
+          "instant": false,
+          "legendFormat": "http-logs",
+          "range": true,
+          "refId": "http_logs"
         },
         {
           "datasource": {
             "type": "prometheus",
             "uid": "${metrics}"
           },
-          "description": "Total ingestion rate.\n\nHere we listed 3 primary protocols:\n\n- Prometheus remote write\n- Greptime's gRPC API (when using our ingest SDK)\n- Log ingestion http API\n",
-          "fieldConfig": {
-            "defaults": {
-              "color": {
-                "mode": "palette-classic"
-              },
-              "custom": {
-                "axisBorderShow": false,
-                "axisCenteredZero": false,
-                "axisColorMode": "text",
-                "axisLabel": "",
-                "axisPlacement": "auto",
-                "barAlignment": 0,
-                "barWidthFactor": 0.6,
-                "drawStyle": "line",
-                "fillOpacity": 0,
-                "gradientMode": "none",
-                "hideFrom": {
-                  "legend": false,
-                  "tooltip": false,
-                  "viz": false
-                },
-                "insertNulls": false,
-                "lineInterpolation": "linear",
-                "lineWidth": 1,
-                "pointSize": 5,
-                "scaleDistribution": {
-                  "type": "linear"
-                },
-                "showPoints": "auto",
-                "spanNulls": false,
-                "stacking": {
-                  "group": "A",
-                  "mode": "none"
-                },
-                "thresholdsStyle": {
-                  "mode": "off"
-                }
-              },
-              "mappings": [],
-              "thresholds": {
-                "mode": "absolute",
-                "steps": [
-                  {
-                    "color": "green"
-                  },
-                  {
-                    "color": "red",
-                    "value": 80
-                  }
-                ]
-              },
-              "unit": "rowsps"
-            },
-            "overrides": []
-          },
-          "gridPos": {
-            "h": 6,
-            "w": 24,
-            "x": 0,
-            "y": 70
-          },
-          "id": 277,
-          "options": {
-            "legend": {
-              "calcs": [
-                "mean"
-              ],
-              "displayMode": "table",
-              "placement": "bottom",
-              "showLegend": true
-            },
-            "tooltip": {
-              "hideZeros": false,
-              "mode": "single",
-              "sort": "none"
-            }
-          },
-          "pluginVersion": "12.0.0",
-          "targets": [
-            {
-              "datasource": {
-                "type": "prometheus",
-                "uid": "${metrics}"
-              },
-              "editorMode": "code",
-              "expr": "sum(rate(greptime_servers_http_logs_ingestion_counter[$__rate_interval]))",
-              "hide": false,
-              "instant": false,
-              "legendFormat": "http-logs",
-              "range": true,
-              "refId": "http_logs"
-            },
-            {
-              "datasource": {
-                "type": "prometheus",
-                "uid": "${metrics}"
-              },
-              "editorMode": "code",
-              "expr": "sum(rate(greptime_servers_prometheus_remote_write_samples[$__rate_interval]))",
-              "hide": false,
-              "instant": false,
-              "legendFormat": "prometheus-remote-write",
-              "range": true,
-              "refId": "prometheus-remote-write"
-            }
-          ],
-          "title": "Ingestion Rate by Type",
-          "type": "timeseries"
+          "editorMode": "code",
+          "expr": "sum(rate(greptime_servers_prometheus_remote_write_samples[$__rate_interval]))",
+          "hide": false,
+          "instant": false,
+          "legendFormat": "prometheus-remote-write",
+          "range": true,
+          "refId": "prometheus-remote-write"
         }
       ],
-      "title": "Ingestion",
-      "type": "row"
+      "title": "Ingestion Rate by Type",
+      "type": "timeseries"
     },
     {
       "collapsed": true,
@@ -1162,7 +1162,7 @@
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 10
+        "y": 22
       },
       "id": 276,
       "panels": [
@@ -1231,7 +1231,7 @@
             "h": 6,
             "w": 24,
             "x": 0,
-            "y": 19
+            "y": 246
           },
           "id": 255,
           "options": {
@@ -1303,7 +1303,7 @@
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 11
+        "y": 23
       },
       "id": 274,
       "panels": [
@@ -1371,7 +1371,7 @@
             "h": 10,
             "w": 12,
             "x": 0,
-            "y": 20
+            "y": 247
           },
           "id": 256,
           "options": {
@@ -1486,7 +1486,7 @@
             "h": 10,
             "w": 12,
             "x": 12,
-            "y": 20
+            "y": 247
           },
           "id": 262,
           "options": {
@@ -1603,7 +1603,7 @@
             "h": 10,
             "w": 12,
             "x": 0,
-            "y": 30
+            "y": 257
           },
           "id": 266,
           "options": {
@@ -1718,7 +1718,7 @@
             "h": 10,
             "w": 12,
             "x": 12,
-            "y": 30
+            "y": 257
           },
           "id": 268,
           "options": {
@@ -1835,7 +1835,7 @@
             "h": 10,
             "w": 12,
             "x": 0,
-            "y": 40
+            "y": 267
           },
           "id": 269,
           "options": {
@@ -1950,7 +1950,7 @@
             "h": 10,
             "w": 12,
             "x": 12,
-            "y": 40
+            "y": 267
           },
           "id": 271,
           "options": {
@@ -2067,7 +2067,7 @@
             "h": 10,
             "w": 12,
             "x": 0,
-            "y": 50
+            "y": 277
           },
           "id": 272,
           "options": {
@@ -2182,7 +2182,7 @@
             "h": 10,
             "w": 12,
             "x": 12,
-            "y": 50
+            "y": 277
           },
           "id": 273,
           "options": {
@@ -2245,7 +2245,7 @@
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 12
+        "y": 24
       },
       "id": 280,
       "panels": [
@@ -2314,7 +2314,7 @@
             "h": 8,
             "w": 12,
             "x": 0,
-            "y": 61
+            "y": 288
           },
           "id": 281,
           "options": {
@@ -2415,7 +2415,7 @@
             "h": 8,
             "w": 12,
             "x": 12,
-            "y": 61
+            "y": 288
           },
           "id": 282,
           "options": {
@@ -2516,7 +2516,7 @@
             "h": 8,
             "w": 12,
             "x": 0,
-            "y": 69
+            "y": 296
           },
           "id": 283,
           "options": {
@@ -2617,7 +2617,7 @@
             "h": 8,
             "w": 12,
             "x": 12,
-            "y": 69
+            "y": 296
           },
           "id": 284,
           "options": {
@@ -2716,7 +2716,7 @@
             "h": 8,
             "w": 12,
             "x": 0,
-            "y": 77
+            "y": 304
           },
           "id": 285,
           "options": {
@@ -2817,7 +2817,7 @@
             "h": 8,
             "w": 12,
             "x": 12,
-            "y": 77
+            "y": 304
           },
           "id": 286,
           "options": {
@@ -2919,7 +2919,7 @@
             "h": 8,
             "w": 12,
             "x": 0,
-            "y": 85
+            "y": 312
           },
           "id": 287,
           "options": {
@@ -3020,7 +3020,7 @@
             "h": 8,
             "w": 12,
             "x": 12,
-            "y": 85
+            "y": 312
           },
           "id": 288,
           "options": {
@@ -3064,7 +3064,7 @@
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 13
+        "y": 25
       },
       "id": 289,
       "panels": [
@@ -3133,7 +3133,7 @@
             "h": 6,
             "w": 24,
             "x": 0,
-            "y": 62
+            "y": 289
           },
           "id": 292,
           "options": {
@@ -3234,7 +3234,7 @@
             "h": 8,
             "w": 12,
             "x": 0,
-            "y": 68
+            "y": 295
           },
           "id": 290,
           "options": {
@@ -3335,7 +3335,7 @@
             "h": 8,
             "w": 12,
             "x": 12,
-            "y": 68
+            "y": 295
           },
           "id": 291,
           "options": {
@@ -3464,7 +3464,7 @@
             "h": 8,
             "w": 12,
             "x": 0,
-            "y": 76
+            "y": 303
           },
           "id": 336,
           "options": {
@@ -3518,2519 +3518,2520 @@
       "type": "row"
     },
     {
-      "collapsed": false,
+      "collapsed": true,
       "gridPos": {
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 14
+        "y": 26
       },
       "id": 293,
-      "panels": [],
+      "panels": [
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Request QPS per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "ops"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 193
+          },
+          "id": 294,
+          "options": {
+            "legend": {
+              "calcs": [
+                "mean"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod, type) (rate(greptime_mito_handle_request_elapsed_count{instance=~\"$datanode\"}[$__rate_interval]))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Request OPS per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Request P99 per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 193
+          },
+          "id": 295,
+          "options": {
+            "legend": {
+              "calcs": [
+                "mean"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(instance, pod, le, type) (rate(greptime_mito_handle_request_elapsed_bucket{instance=~\"$datanode\"}[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Request P99 per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Write Buffer per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "decbytes"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 225
+          },
+          "id": 296,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_mito_write_buffer_bytes{instance=~\"$datanode\"}",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Write Buffer per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Ingestion size by row counts.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "rowsps"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 225
+          },
+          "id": 297,
+          "options": {
+            "legend": {
+              "calcs": [
+                "mean"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by (instance, pod) (rate(greptime_mito_write_rows_total{instance=~\"$datanode\"}[$__rate_interval]))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Write Rows per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Flush QPS per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "ops"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 233
+          },
+          "id": 298,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod, reason) (rate(greptime_mito_flush_requests_total{instance=~\"$datanode\"}[$__rate_interval]))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{reason}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Flush OPS per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Write Stall per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              }
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 233
+          },
+          "id": 299,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod) (greptime_mito_write_stall_total{instance=~\"$datanode\"})",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Write Stall per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Read Stage OPS per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "ops"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 241
+          },
+          "id": 300,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod) (rate(greptime_mito_read_stage_elapsed_count{instance=~\"$datanode\", stage=\"total\"}[$__rate_interval]))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Read Stage OPS per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Read Stage P99 per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": [
+              {
+                "__systemRef": "hideSeriesFrom",
+                "matcher": {
+                  "id": "byNames",
+                  "options": {
+                    "mode": "exclude",
+                    "names": [
+                      "[10.39.145.87:4000]-[mycluster-datanode-1]-[cache_miss_read]"
+                    ],
+                    "prefix": "All except:",
+                    "readOnly": true
+                  }
+                },
+                "properties": [
+                  {
+                    "id": "custom.hideFrom",
+                    "value": {
+                      "legend": false,
+                      "tooltip": false,
+                      "viz": true
+                    }
+                  }
+                ]
+              }
+            ]
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 241
+          },
+          "id": 301,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(instance, pod, le, stage) (rate(greptime_mito_read_stage_elapsed_bucket{instance=~\"$datanode\"}[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Read Stage P99 per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Write Stage P99 per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 249
+          },
+          "id": 302,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true,
+              "sortBy": "Last *",
+              "sortDesc": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(instance, pod, le, stage) (rate(greptime_mito_write_stage_elapsed_bucket{instance=~\"$datanode\"}[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Write Stage P99 per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Compaction OPS per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "ops"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 249
+          },
+          "id": 303,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod) (rate(greptime_mito_compaction_total_elapsed_count{instance=~\"$datanode\"}[$__rate_interval]))",
+              "instant": false,
+              "legendFormat": "[{{ instance }}]-[{{pod}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Compaction OPS per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Compaction latency by stage",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 257
+          },
+          "id": 304,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(instance, pod, le, stage) (rate(greptime_mito_compaction_stage_elapsed_bucket{instance=~\"$datanode\"}[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-p99",
+              "range": true,
+              "refId": "A"
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod, stage) (rate(greptime_mito_compaction_stage_elapsed_sum{instance=~\"$datanode\"}[$__rate_interval]))/sum by(instance, pod, stage) (rate(greptime_mito_compaction_stage_elapsed_count{instance=~\"$datanode\"}[$__rate_interval]))",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-avg",
+              "range": true,
+              "refId": "B"
+            }
+          ],
+          "title": "Compaction Elapsed Time per Instance by Stage",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Compaction P99 per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 257
+          },
+          "id": 305,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(instance, pod, le,stage) (rate(greptime_mito_compaction_total_elapsed_bucket{instance=~\"$datanode\"}[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-compaction",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Compaction P99 per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Write-ahead logs write size as bytes. This chart includes stats of p95 and p99 size by instance, total WAL write rate.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "bytes"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 265
+          },
+          "id": 306,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.95, sum by(le,instance, pod) (rate(raft_engine_write_size_bucket[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-req-size-p95",
+              "range": true,
+              "refId": "A"
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(le,instance,pod) (rate(raft_engine_write_size_bucket[$__rate_interval])))",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-req-size-p99",
+              "range": true,
+              "refId": "B"
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by (instance, pod)(rate(raft_engine_write_size_sum[$__rate_interval]))",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-throughput",
+              "range": true,
+              "refId": "C"
+            }
+          ],
+          "title": "WAL write size",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Cached Bytes per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "decbytes"
+            },
+            "overrides": [
+              {
+                "__systemRef": "hideSeriesFrom",
+                "matcher": {
+                  "id": "byNames",
+                  "options": {
+                    "mode": "exclude",
+                    "names": [
+                      "[10.39.145.79:4000]-[mycluster-datanode-0]-[file]"
+                    ],
+                    "prefix": "All except:",
+                    "readOnly": true
+                  }
+                },
+                "properties": [
+                  {
+                    "id": "custom.hideFrom",
+                    "value": {
+                      "legend": false,
+                      "tooltip": false,
+                      "viz": true
+                    }
+                  }
+                ]
+              }
+            ]
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 265
+          },
+          "id": 307,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_mito_cache_bytes{instance=~\"$datanode\"}",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Cached Bytes per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Ongoing compaction task count",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 273
+          },
+          "id": 308,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_mito_inflight_compaction_count",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Inflight Compaction",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Raft engine (local disk) log store sync latency, p99",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 273
+          },
+          "id": 310,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(le, type, node, instance, pod) (rate(raft_engine_sync_log_duration_seconds_bucket[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-p99",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "WAL sync duration seconds",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Write-ahead log operations latency at p99",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 281
+          },
+          "id": 311,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(le,logstore,optype,instance, pod) (rate(greptime_logstore_op_elapsed_bucket[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{logstore}}]-[{{optype}}]-p99",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Log Store op duration seconds",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Ongoing flush task count",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 281
+          },
+          "id": 312,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_mito_inflight_flush_count",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Inflight Flush",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Compaction oinput output bytes",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "bytes"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 289
+          },
+          "id": 335,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod) (greptime_mito_compaction_input_bytes)",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-input",
+              "range": true,
+              "refId": "A"
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod) (greptime_mito_compaction_output_bytes)",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-output",
+              "range": true,
+              "refId": "B"
+            }
+          ],
+          "title": "Compaction Input/Output Bytes",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Per-stage elapsed time for region worker to handle bulk insert region requests.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": [
+              {
+                "__systemRef": "hideSeriesFrom",
+                "matcher": {
+                  "id": "byNames",
+                  "options": {
+                    "mode": "exclude",
+                    "names": [
+                      "[192.168.50.8:4000]-[]-[prepare_bulk_request]-AVG",
+                      "[minipc-1:4000]-[]-[bulk_extend]-AVG",
+                      "[minipc-1:4000]-[]-[prepare_bulk_request]-AVG",
+                      "[minipc-1:4000]-[]-[process_bulk_req]-AVG",
+                      "[minipc-1:4000]-[]-[write_bulk]-AVG"
+                    ],
+                    "prefix": "All except:",
+                    "readOnly": true
+                  }
+                },
+                "properties": [
+                  {
+                    "id": "custom.hideFrom",
+                    "value": {
+                      "legend": false,
+                      "tooltip": false,
+                      "viz": true
+                    }
+                  }
+                ]
+              }
+            ]
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 289
+          },
+          "id": 337,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.95, sum by(le,instance, stage, pod) (rate(greptime_region_worker_handle_write_bucket[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-P95",
+              "range": true,
+              "refId": "A"
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, stage, pod) (rate(greptime_region_worker_handle_write_sum[$__rate_interval]))/sum by(instance, stage, pod) (rate(greptime_region_worker_handle_write_count[$__rate_interval]))",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-AVG",
+              "range": true,
+              "refId": "B"
+            }
+          ],
+          "title": "Region Worker Handle Bulk Insert Requests",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Compaction oinput output bytes",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 297
+          },
+          "id": 348,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod) (greptime_mito_memtable_active_series_count)",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-series",
+              "range": true,
+              "refId": "A"
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod) (greptime_mito_memtable_field_builder_count)",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-field_builders",
+              "range": true,
+              "refId": "B"
+            }
+          ],
+          "title": "Active Series and Field Builders Count",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Per-stage elapsed time for region worker to decode requests.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 297
+          },
+          "id": 338,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "builder",
+              "expr": "histogram_quantile(0.95, sum by(le, instance, stage, pod) (rate(greptime_datanode_convert_region_request_bucket[$__rate_interval])))",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-P95",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(le,instance, stage, pod) (rate(greptime_datanode_convert_region_request_sum[$__rate_interval]))/sum by(le,instance, stage, pod) (rate(greptime_datanode_convert_region_request_count[$__rate_interval]))",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-AVG",
+              "range": true,
+              "refId": "B"
+            }
+          ],
+          "title": "Region Worker Convert Requests",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "The local cache miss of the datanode.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              }
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 305
+          },
+          "id": 349,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by (instance,pod, type) (rate(greptime_mito_cache_miss{instance=~\"$datanode\"}[$__rate_interval]))",
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Cache Miss",
+          "type": "timeseries"
+        }
+      ],
       "title": "Mito Engine",
       "type": "row"
     },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Request QPS per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "ops"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 15
-      },
-      "id": 294,
-      "options": {
-        "legend": {
-          "calcs": [
-            "mean"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod, type) (rate(greptime_mito_handle_request_elapsed_count{instance=~\"$datanode\"}[$__rate_interval]))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Request OPS per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Request P99 per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 15
-      },
-      "id": 295,
-      "options": {
-        "legend": {
-          "calcs": [
-            "mean"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(instance, pod, le, type) (rate(greptime_mito_handle_request_elapsed_bucket{instance=~\"$datanode\"}[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Request P99 per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Write Buffer per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "decbytes"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 23
-      },
-      "id": 296,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_mito_write_buffer_bytes{instance=~\"$datanode\"}",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Write Buffer per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Ingestion size by row counts.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "rowsps"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 23
-      },
-      "id": 297,
-      "options": {
-        "legend": {
-          "calcs": [
-            "mean"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by (instance, pod) (rate(greptime_mito_write_rows_total{instance=~\"$datanode\"}[$__rate_interval]))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Write Rows per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Flush QPS per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "ops"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 31
-      },
-      "id": 298,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod, reason) (rate(greptime_mito_flush_requests_total{instance=~\"$datanode\"}[$__rate_interval]))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{reason}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Flush OPS per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Write Stall per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          }
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 31
-      },
-      "id": 299,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod) (greptime_mito_write_stall_total{instance=~\"$datanode\"})",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Write Stall per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Read Stage OPS per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "ops"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 39
-      },
-      "id": 300,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod) (rate(greptime_mito_read_stage_elapsed_count{instance=~\"$datanode\", stage=\"total\"}[$__rate_interval]))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Read Stage OPS per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Read Stage P99 per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": [
-          {
-            "__systemRef": "hideSeriesFrom",
-            "matcher": {
-              "id": "byNames",
-              "options": {
-                "mode": "exclude",
-                "names": [
-                  "[10.39.145.87:4000]-[mycluster-datanode-1]-[cache_miss_read]"
-                ],
-                "prefix": "All except:",
-                "readOnly": true
-              }
-            },
-            "properties": [
-              {
-                "id": "custom.hideFrom",
-                "value": {
-                  "legend": false,
-                  "tooltip": false,
-                  "viz": true
-                }
-              }
-            ]
-          }
-        ]
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 39
-      },
-      "id": 301,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(instance, pod, le, stage) (rate(greptime_mito_read_stage_elapsed_bucket{instance=~\"$datanode\"}[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Read Stage P99 per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Write Stage P99 per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 47
-      },
-      "id": 302,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true,
-          "sortBy": "Last *",
-          "sortDesc": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(instance, pod, le, stage) (rate(greptime_mito_write_stage_elapsed_bucket{instance=~\"$datanode\"}[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Write Stage P99 per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Compaction OPS per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "ops"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 47
-      },
-      "id": 303,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod) (rate(greptime_mito_compaction_total_elapsed_count{instance=~\"$datanode\"}[$__rate_interval]))",
-          "instant": false,
-          "legendFormat": "[{{ instance }}]-[{{pod}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Compaction OPS per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Compaction latency by stage",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 55
-      },
-      "id": 304,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(instance, pod, le, stage) (rate(greptime_mito_compaction_stage_elapsed_bucket{instance=~\"$datanode\"}[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-p99",
-          "range": true,
-          "refId": "A"
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod, stage) (rate(greptime_mito_compaction_stage_elapsed_sum{instance=~\"$datanode\"}[$__rate_interval]))/sum by(instance, pod, stage) (rate(greptime_mito_compaction_stage_elapsed_count{instance=~\"$datanode\"}[$__rate_interval]))",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-avg",
-          "range": true,
-          "refId": "B"
-        }
-      ],
-      "title": "Compaction Elapsed Time per Instance by Stage",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Compaction P99 per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 55
-      },
-      "id": 305,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(instance, pod, le,stage) (rate(greptime_mito_compaction_total_elapsed_bucket{instance=~\"$datanode\"}[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-compaction",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Compaction P99 per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Write-ahead logs write size as bytes. This chart includes stats of p95 and p99 size by instance, total WAL write rate.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "bytes"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 63
-      },
-      "id": 306,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.95, sum by(le,instance, pod) (rate(raft_engine_write_size_bucket[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-req-size-p95",
-          "range": true,
-          "refId": "A"
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(le,instance,pod) (rate(raft_engine_write_size_bucket[$__rate_interval])))",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-req-size-p99",
-          "range": true,
-          "refId": "B"
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by (instance, pod)(rate(raft_engine_write_size_sum[$__rate_interval]))",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-throughput",
-          "range": true,
-          "refId": "C"
-        }
-      ],
-      "title": "WAL write size",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Cached Bytes per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "decbytes"
-        },
-        "overrides": [
-          {
-            "__systemRef": "hideSeriesFrom",
-            "matcher": {
-              "id": "byNames",
-              "options": {
-                "mode": "exclude",
-                "names": [
-                  "[10.39.145.79:4000]-[mycluster-datanode-0]-[file]"
-                ],
-                "prefix": "All except:",
-                "readOnly": true
-              }
-            },
-            "properties": [
-              {
-                "id": "custom.hideFrom",
-                "value": {
-                  "legend": false,
-                  "tooltip": false,
-                  "viz": true
-                }
-              }
-            ]
-          }
-        ]
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 63
-      },
-      "id": 307,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_mito_cache_bytes{instance=~\"$datanode\"}",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Cached Bytes per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Ongoing compaction task count",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 71
-      },
-      "id": 308,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_mito_inflight_compaction_count",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Inflight Compaction",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Raft engine (local disk) log store sync latency, p99",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 71
-      },
-      "id": 310,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(le, type, node, instance, pod) (rate(raft_engine_sync_log_duration_seconds_bucket[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-p99",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "WAL sync duration seconds",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Write-ahead log operations latency at p99",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 79
-      },
-      "id": 311,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(le,logstore,optype,instance, pod) (rate(greptime_logstore_op_elapsed_bucket[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{logstore}}]-[{{optype}}]-p99",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Log Store op duration seconds",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Ongoing flush task count",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 79
-      },
-      "id": 312,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_mito_inflight_flush_count",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Inflight Flush",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Compaction oinput output bytes",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "bytes"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 87
-      },
-      "id": 335,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod) (greptime_mito_compaction_input_bytes)",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-input",
-          "range": true,
-          "refId": "A"
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod) (greptime_mito_compaction_output_bytes)",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-output",
-          "range": true,
-          "refId": "B"
-        }
-      ],
-      "title": "Compaction Input/Output Bytes",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Per-stage elapsed time for region worker to handle bulk insert region requests.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": [
-          {
-            "__systemRef": "hideSeriesFrom",
-            "matcher": {
-              "id": "byNames",
-              "options": {
-                "mode": "exclude",
-                "names": [
-                  "[192.168.50.8:4000]-[]-[prepare_bulk_request]-AVG",
-                  "[minipc-1:4000]-[]-[bulk_extend]-AVG",
-                  "[minipc-1:4000]-[]-[prepare_bulk_request]-AVG",
-                  "[minipc-1:4000]-[]-[process_bulk_req]-AVG",
-                  "[minipc-1:4000]-[]-[write_bulk]-AVG"
-                ],
-                "prefix": "All except:",
-                "readOnly": true
-              }
-            },
-            "properties": [
-              {
-                "id": "custom.hideFrom",
-                "value": {
-                  "legend": false,
-                  "tooltip": false,
-                  "viz": true
-                }
-              }
-            ]
-          }
-        ]
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 87
-      },
-      "id": 337,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.95, sum by(le,instance, stage, pod) (rate(greptime_region_worker_handle_write_bucket[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-P95",
-          "range": true,
-          "refId": "A"
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, stage, pod) (rate(greptime_region_worker_handle_write_sum[$__rate_interval]))/sum by(instance, stage, pod) (rate(greptime_region_worker_handle_write_count[$__rate_interval]))",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-AVG",
-          "range": true,
-          "refId": "B"
-        }
-      ],
-      "title": "Region Worker Handle Bulk Insert Requests",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Compaction oinput output bytes",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 95
-      },
-      "id": 348,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod) (greptime_mito_memtable_active_series_count)",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-series",
-          "range": true,
-          "refId": "A"
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod) (greptime_mito_memtable_field_builder_count)",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-field_builders",
-          "range": true,
-          "refId": "B"
-        }
-      ],
-      "title": "Active Series and Field Builders Count",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Per-stage elapsed time for region worker to decode requests.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 95
-      },
-      "id": 338,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "builder",
-          "expr": "histogram_quantile(0.95, sum by(le, instance, stage, pod) (rate(greptime_datanode_convert_region_request_bucket[$__rate_interval])))",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-P95",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(le,instance, stage, pod) (rate(greptime_datanode_convert_region_request_sum[$__rate_interval]))/sum by(le,instance, stage, pod) (rate(greptime_datanode_convert_region_request_count[$__rate_interval]))",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-AVG",
-          "range": true,
-          "refId": "B"
-        }
-      ],
-      "title": "Region Worker Convert Requests",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "The local cache miss of the datanode.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          }
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 103
-      },
-      "id": 349,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by (instance,pod, type) (rate(greptime_mito_cache_miss{instance=~\"$datanode\"}[$__rate_interval]))",
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Cache Miss",
-      "type": "timeseries"
-    },
     {
       "collapsed": true,
       "gridPos": {
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 111
+        "y": 27
       },
       "id": 313,
       "panels": [
@@ -6098,7 +6099,7 @@
             "h": 10,
             "w": 24,
             "x": 0,
-            "y": 232
+            "y": 459
           },
           "id": 314,
           "options": {
@@ -6198,7 +6199,7 @@
             "h": 7,
             "w": 12,
             "x": 0,
-            "y": 242
+            "y": 469
           },
           "id": 315,
           "options": {
@@ -6298,7 +6299,7 @@
             "h": 7,
             "w": 12,
             "x": 12,
-            "y": 242
+            "y": 469
           },
           "id": 316,
           "options": {
@@ -6398,7 +6399,7 @@
             "h": 7,
             "w": 12,
             "x": 0,
-            "y": 249
+            "y": 476
           },
           "id": 317,
           "options": {
@@ -6498,7 +6499,7 @@
             "h": 7,
             "w": 12,
             "x": 12,
-            "y": 249
+            "y": 476
           },
           "id": 318,
           "options": {
@@ -6598,7 +6599,7 @@
             "h": 7,
             "w": 12,
             "x": 0,
-            "y": 256
+            "y": 483
           },
           "id": 319,
           "options": {
@@ -6699,7 +6700,7 @@
             "h": 7,
             "w": 12,
             "x": 12,
-            "y": 256
+            "y": 483
           },
           "id": 320,
           "options": {
@@ -6800,7 +6801,7 @@
             "h": 7,
             "w": 12,
             "x": 0,
-            "y": 263
+            "y": 490
           },
           "id": 321,
           "options": {
@@ -6900,7 +6901,7 @@
             "h": 7,
             "w": 12,
             "x": 12,
-            "y": 263
+            "y": 490
           },
           "id": 322,
           "options": {
@@ -7000,7 +7001,7 @@
             "h": 7,
             "w": 12,
             "x": 0,
-            "y": 270
+            "y": 497
           },
           "id": 323,
           "options": {
@@ -7103,7 +7104,7 @@
             "h": 7,
             "w": 12,
             "x": 12,
-            "y": 270
+            "y": 497
           },
           "id": 334,
           "options": {
@@ -7144,1827 +7145,1829 @@
       "type": "row"
     },
     {
-      "collapsed": false,
+      "collapsed": true,
       "gridPos": {
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 112
+        "y": 28
       },
       "id": 351,
-      "panels": [],
+      "panels": [
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Triggered region flush total",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 195
+          },
+          "id": 350,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "meta_triggered_region_flush_total",
+              "instant": false,
+              "legendFormat": "{{pod}}-{{topic_name}}",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Triggered region flush total",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Triggered region checkpoint total",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 195
+          },
+          "id": 352,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "meta_triggered_region_checkpoint_total",
+              "instant": false,
+              "legendFormat": "{{pod}}-{{topic_name}}",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Triggered region checkpoint total",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Topic estimated max replay size",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "bytes"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 227
+          },
+          "id": 353,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "meta_topic_estimated_replay_size",
+              "instant": false,
+              "legendFormat": "{{pod}}-{{topic_name}}",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Topic estimated replay size",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Kafka logstore's bytes traffic",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "bytes"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 227
+          },
+          "id": 354,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "rate(greptime_logstore_kafka_client_bytes_total[$__rate_interval])",
+              "instant": false,
+              "legendFormat": "{{pod}}-{{logstore}}",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Kafka logstore's bytes traffic",
+          "type": "timeseries"
+        }
+      ],
       "title": "Remote WAL",
       "type": "row"
     },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Triggered region flush total",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 113
-      },
-      "id": 350,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "meta_triggered_region_flush_total",
-          "instant": false,
-          "legendFormat": "{{pod}}-{{topic_name}}",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Triggered region flush total",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Triggered region checkpoint total",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 113
-      },
-      "id": 352,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "meta_triggered_region_checkpoint_total",
-          "instant": false,
-          "legendFormat": "{{pod}}-{{topic_name}}",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Triggered region checkpoint total",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Topic estimated max replay size",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "bytes"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 121
-      },
-      "id": 353,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "meta_topic_estimated_replay_size",
-          "instant": false,
-          "legendFormat": "{{pod}}-{{topic_name}}",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Topic estimated replay size",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Kafka logstore's bytes traffic",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "bytes"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 121
-      },
-      "id": 354,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "rate(greptime_logstore_kafka_client_bytes_total[$__rate_interval])",
-          "instant": false,
-          "legendFormat": "{{pod}}-{{logstore}}",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Kafka logstore's bytes traffic",
-      "type": "timeseries"
-    },
-    {
-      "collapsed": false,
-      "gridPos": {
-        "h": 1,
-        "w": 24,
-        "x": 0,
-        "y": 129
-      },
-      "id": 324,
-      "panels": [],
-      "title": "Metasrv",
-      "type": "row"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Counter of region migration by source and destination",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "thresholds"
-          },
-          "custom": {
-            "axisPlacement": "auto",
-            "fillOpacity": 70,
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "lineWidth": 1
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          }
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 130
-      },
-      "id": 325,
-      "options": {
-        "colWidth": 0.9,
-        "legend": {
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": false
-        },
-        "rowHeight": 0.9,
-        "showValue": "auto",
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_meta_region_migration_stat{datanode_type=\"src\"}",
-          "instant": false,
-          "legendFormat": "from-datanode-{{datanode_id}}",
-          "range": true,
-          "refId": "A"
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_meta_region_migration_stat{datanode_type=\"desc\"}",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "to-datanode-{{datanode_id}}",
-          "range": true,
-          "refId": "B"
-        }
-      ],
-      "title": "Region migration datanode",
-      "type": "status-history"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Counter of region migration error",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 130
-      },
-      "id": 326,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_meta_region_migration_error",
-          "instant": false,
-          "legendFormat": "{{pod}}-{{state}}-{{error_type}}",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Region migration error",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "binBps"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 138
-      },
-      "id": 327,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_datanode_load",
-          "instant": false,
-          "legendFormat": "Datanode-{{datanode_id}}-writeload",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Datanode load",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Displays the rate of SQL executions processed by the Meta service using the RDS backend.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 138
-      },
-      "id": 339,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "rate(greptime_meta_rds_pg_sql_execute_elapsed_ms_count[$__rate_interval])",
-          "instant": false,
-          "legendFormat": "{{pod}} {{op}} {{type}} {{result}} ",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Rate of SQL Executions (RDS)",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Measures the response time of SQL executions via the RDS backend. ",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "ms"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 146
-      },
-      "id": 340,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.90, sum by(pod, op, type, result, le) (rate(greptime_meta_rds_pg_sql_execute_elapsed_ms_bucket[$__rate_interval])))",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "{{pod}} {{op}} {{type}} {{result}} p90",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        }
-      ],
-      "title": "SQL Execution Latency (RDS)",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Shows latency of Meta handlers by pod and handler name, useful for monitoring handler performance and detecting latency spikes.\n",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 146
-      },
-      "id": 341,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.90, sum by(pod, le, name) (\n  rate(greptime_meta_handler_execute_bucket[$__rate_interval])\n))",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "{{pod}} {{name}} p90",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        }
-      ],
-      "title": "Handler Execution Latency",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Shows p90 heartbeat message sizes, helping track network usage and identify anomalies in heartbeat payload.\n",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "bytes"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 154
-      },
-      "id": 342,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, sum by(pod, le) (greptime_meta_heartbeat_stat_memory_size_bucket))",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "{{pod}}",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        }
-      ],
-      "title": "Heartbeat Packet Size",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 154
-      },
-      "id": 345,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "rate(greptime_meta_heartbeat_rate[$__rate_interval])",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "{{pod}}",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        }
-      ],
-      "title": "Meta Heartbeat Receive Rate",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 162
-      },
-      "id": 343,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(pod, le, op, target) (greptime_meta_kv_request_elapsed_bucket))",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "{{pod}}-{{op}} p99",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        }
-      ],
-      "title": "Meta KV Ops Latency",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 162
-      },
-      "id": 346,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "rate(greptime_meta_kv_request_elapsed_count[$__rate_interval])",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "{{pod}}-{{op}} p99",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        }
-      ],
-      "title": "Rate of meta KV Ops",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 170
-      },
-      "id": 347,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_tables_bucket))",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "CreateLogicalTables-{{step}} p90",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_table))",
-          "fullMetaSearch": false,
-          "hide": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "CreateTable-{{step}} p90",
-          "range": true,
-          "refId": "B",
-          "useBackend": false
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_view))",
-          "fullMetaSearch": false,
-          "hide": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "CreateView-{{step}} p90",
-          "range": true,
-          "refId": "C",
-          "useBackend": false
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_flow))",
-          "fullMetaSearch": false,
-          "hide": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "CreateFlow-{{step}} p90",
-          "range": true,
-          "refId": "D",
-          "useBackend": false
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_drop_table))",
-          "fullMetaSearch": false,
-          "hide": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "DropTable-{{step}} p90",
-          "range": true,
-          "refId": "E",
-          "useBackend": false
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_alter_table))",
-          "fullMetaSearch": false,
-          "hide": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "AlterTable-{{step}} p90",
-          "range": true,
-          "refId": "F",
-          "useBackend": false
-        }
-      ],
-      "title": "DDL Latency",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Reconciliation stats",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 170
-      },
-      "id": 355,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_meta_reconciliation_stats",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "{{pod}}-{{table_type}}-{{type}}",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Reconciliation stats",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Elapsed of Reconciliation steps",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 178
-      },
-      "id": 356,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, greptime_meta_reconciliation_procedure_bucket)",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "{{procedure_name}}-{{step}}-P90",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Reconciliation steps",
-      "type": "timeseries"
-    },
     {
       "collapsed": true,
       "gridPos": {
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 186
+        "y": 29
+      },
+      "id": 324,
+      "panels": [
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Counter of region migration by source and destination",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "thresholds"
+              },
+              "custom": {
+                "axisPlacement": "auto",
+                "fillOpacity": 70,
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "lineWidth": 1
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              }
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 196
+          },
+          "id": 325,
+          "options": {
+            "colWidth": 0.9,
+            "legend": {
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": false
+            },
+            "rowHeight": 0.9,
+            "showValue": "auto",
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_meta_region_migration_stat{datanode_type=\"src\"}",
+              "instant": false,
+              "legendFormat": "from-datanode-{{datanode_id}}",
+              "range": true,
+              "refId": "A"
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_meta_region_migration_stat{datanode_type=\"desc\"}",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "to-datanode-{{datanode_id}}",
+              "range": true,
+              "refId": "B"
+            }
+          ],
+          "title": "Region migration datanode",
+          "type": "status-history"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Counter of region migration error",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 196
+          },
+          "id": 326,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_meta_region_migration_error",
+              "instant": false,
+              "legendFormat": "{{pod}}-{{state}}-{{error_type}}",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Region migration error",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "binBps"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 228
+          },
+          "id": 327,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_datanode_load",
+              "instant": false,
+              "legendFormat": "Datanode-{{datanode_id}}-writeload",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Datanode load",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Displays the rate of SQL executions processed by the Meta service using the RDS backend.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 228
+          },
+          "id": 339,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "rate(greptime_meta_rds_pg_sql_execute_elapsed_ms_count[$__rate_interval])",
+              "instant": false,
+              "legendFormat": "{{pod}} {{op}} {{type}} {{result}} ",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Rate of SQL Executions (RDS)",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Measures the response time of SQL executions via the RDS backend. ",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "ms"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 236
+          },
+          "id": 340,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.90, sum by(pod, op, type, result, le) (rate(greptime_meta_rds_pg_sql_execute_elapsed_ms_bucket[$__rate_interval])))",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "{{pod}} {{op}} {{type}} {{result}} p90",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            }
+          ],
+          "title": "SQL Execution Latency (RDS)",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Shows latency of Meta handlers by pod and handler name, useful for monitoring handler performance and detecting latency spikes.\n",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 236
+          },
+          "id": 341,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.90, sum by(pod, le, name) (\n  rate(greptime_meta_handler_execute_bucket[$__rate_interval])\n))",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "{{pod}} {{name}} p90",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            }
+          ],
+          "title": "Handler Execution Latency",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Shows p90 heartbeat message sizes, helping track network usage and identify anomalies in heartbeat payload.\n",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "bytes"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 244
+          },
+          "id": 342,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, sum by(pod, le) (greptime_meta_heartbeat_stat_memory_size_bucket))",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "{{pod}}",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            }
+          ],
+          "title": "Heartbeat Packet Size",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 244
+          },
+          "id": 345,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "rate(greptime_meta_heartbeat_rate[$__rate_interval])",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "{{pod}}",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            }
+          ],
+          "title": "Meta Heartbeat Receive Rate",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 252
+          },
+          "id": 343,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(pod, le, op, target) (greptime_meta_kv_request_elapsed_bucket))",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "{{pod}}-{{op}} p99",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            }
+          ],
+          "title": "Meta KV Ops Latency",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 252
+          },
+          "id": 346,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "rate(greptime_meta_kv_request_elapsed_count[$__rate_interval])",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "{{pod}}-{{op}} p99",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            }
+          ],
+          "title": "Rate of meta KV Ops",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 260
+          },
+          "id": 347,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_tables_bucket))",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "CreateLogicalTables-{{step}} p90",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_table))",
+              "fullMetaSearch": false,
+              "hide": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "CreateTable-{{step}} p90",
+              "range": true,
+              "refId": "B",
+              "useBackend": false
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_view))",
+              "fullMetaSearch": false,
+              "hide": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "CreateView-{{step}} p90",
+              "range": true,
+              "refId": "C",
+              "useBackend": false
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_flow))",
+              "fullMetaSearch": false,
+              "hide": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "CreateFlow-{{step}} p90",
+              "range": true,
+              "refId": "D",
+              "useBackend": false
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_drop_table))",
+              "fullMetaSearch": false,
+              "hide": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "DropTable-{{step}} p90",
+              "range": true,
+              "refId": "E",
+              "useBackend": false
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_alter_table))",
+              "fullMetaSearch": false,
+              "hide": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "AlterTable-{{step}} p90",
+              "range": true,
+              "refId": "F",
+              "useBackend": false
+            }
+          ],
+          "title": "DDL Latency",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Reconciliation stats",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 260
+          },
+          "id": 355,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_meta_reconciliation_stats",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "{{pod}}-{{table_type}}-{{type}}",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Reconciliation stats",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Elapsed of Reconciliation steps",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 268
+          },
+          "id": 356,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, greptime_meta_reconciliation_procedure_bucket)",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "{{procedure_name}}-{{step}}-P90",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Reconciliation steps",
+          "type": "timeseries"
+        }
+      ],
+      "title": "Metasrv",
+      "type": "row"
+    },
+    {
+      "collapsed": true,
+      "gridPos": {
+        "h": 1,
+        "w": 24,
+        "x": 0,
+        "y": 30
       },
       "id": 328,
       "panels": [
@@ -9031,7 +9034,7 @@
             "h": 8,
             "w": 12,
             "x": 0,
-            "y": 3055
+            "y": 3282
           },
           "id": 329,
           "options": {
@@ -9126,7 +9129,7 @@
             "h": 8,
             "w": 12,
             "x": 12,
-            "y": 3055
+            "y": 3282
           },
           "id": 330,
           "options": {
@@ -9234,7 +9237,7 @@
             "h": 8,
             "w": 9,
             "x": 0,
-            "y": 3063
+            "y": 3290
           },
           "id": 331,
           "options": {
@@ -9342,7 +9345,7 @@
             "h": 8,
             "w": 9,
             "x": 9,
-            "y": 3063
+            "y": 3290
           },
           "id": 332,
           "options": {
@@ -9437,7 +9440,7 @@
             "h": 8,
             "w": 6,
             "x": 18,
-            "y": 3063
+            "y": 3290
           },
           "id": 333,
           "options": {
@@ -9474,12 +9477,12 @@
       "type": "row"
     },
     {
-      "collapsed": true,
+      "collapsed": false,
       "gridPos": {
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 187
+        "y": 31
       },
       "id": 357,
       "panels": [],
@@ -9550,7 +9553,7 @@
         "h": 8,
         "w": 24,
         "x": 0,
-        "y": 188
+        "y": 32
       },
       "id": 358,
       "options": {
@@ -9648,7 +9651,7 @@
         "h": 8,
         "w": 12,
         "x": 0,
-        "y": 196
+        "y": 40
       },
       "id": 359,
       "options": {
@@ -9761,7 +9764,7 @@
         "h": 8,
         "w": 12,
         "x": 12,
-        "y": 196
+        "y": 40
       },
       "id": 360,
       "options": {
@@ -9861,7 +9864,7 @@
         "h": 8,
         "w": 12,
         "x": 0,
-        "y": 204
+        "y": 48
       },
       "id": 361,
       "options": {
@@ -9974,7 +9977,7 @@
         "h": 8,
         "w": 12,
         "x": 12,
-        "y": 204
+        "y": 48
       },
       "id": 364,
       "options": {
@@ -10074,7 +10077,7 @@
         "h": 8,
         "w": 12,
         "x": 0,
-        "y": 212
+        "y": 56
       },
       "id": 363,
       "options": {
@@ -10187,7 +10190,7 @@
         "h": 8,
         "w": 12,
         "x": 12,
-        "y": 212
+        "y": 56
       },
       "id": 362,
       "options": {
@@ -10221,6 +10224,477 @@
       ],
       "title": "Save Alert Failure Rate",
       "type": "timeseries"
+    },
+    {
+      "collapsed": false,
+      "gridPos": {
+        "h": 1,
+        "w": 24,
+        "x": 0,
+        "y": 64
+      },
+      "id": 366,
+      "panels": [],
+      "title": "Hotspot",
+      "type": "row"
+    },
+    {
+      "datasource": {
+        "type": "mysql",
+        "uid": "${information_schema}"
+      },
+      "fieldConfig": {
+        "defaults": {
+          "color": {
+            "mode": "thresholds"
+          },
+          "custom": {
+            "align": "auto",
+            "cellOptions": {
+              "type": "auto",
+              "wrapText": false
+            },
+            "inspect": false
+          },
+          "fieldMinMax": false,
+          "mappings": [],
+          "thresholds": {
+            "mode": "absolute",
+            "steps": [
+              {
+                "color": "green"
+              },
+              {
+                "color": "red",
+                "value": 80
+              }
+            ]
+          }
+        },
+        "overrides": [
+          {
+            "matcher": {
+              "id": "byName",
+              "options": "disk_size"
+            },
+            "properties": [
+              {
+                "id": "unit",
+                "value": "bytes"
+              }
+            ]
+          }
+        ]
+      },
+      "gridPos": {
+        "h": 18,
+        "w": 24,
+        "x": 0,
+        "y": 65
+      },
+      "id": 365,
+      "options": {
+        "cellHeight": "sm",
+        "footer": {
+          "countRows": false,
+          "enablePagination": true,
+          "fields": [
+            "region_rows",
+            "disk_size"
+          ],
+          "reducer": [
+            "sum"
+          ],
+          "show": true
+        },
+        "showHeader": true,
+        "sortBy": [
+          {
+            "desc": true,
+            "displayName": "table_schema"
+          }
+        ]
+      },
+      "pluginVersion": "11.6.0",
+      "targets": [
+        {
+          "dataset": "information_schema",
+          "datasource": {
+            "type": "mysql",
+            "uid": "${information_schema}"
+          },
+          "editorMode": "code",
+          "format": "table",
+          "rawQuery": true,
+          "rawSql": "WITH table_stats AS (\n  SELECT\n    table_id,\n    COUNT(*) AS region_count,\n    SUM(disk_size) AS total_disk_size,\n    SUM(region_rows) as total_region_rows\n  FROM information_schema.region_statistics\n  WHERE region_role = 'Leader'\n  GROUP BY table_id\n  HAVING COUNT(*) > 1\n)\n\nSELECT\n  t.table_schema,\n  t.table_name,\n\n  r.region_id,\n  t.table_id,\n  r.region_number,\n\n  p.partition_description,\n\n\n  ROUND(\n    r.disk_size * 100.0\n      / NULLIF(ts.total_disk_size, 0),\n    2\n  ) AS disk_size_share_percent,\n\n  r.disk_size,\n  \n  ROUND(\n    r.region_rows * 100.0\n      / NULLIF(ts.total_region_rows, 0),\n    2\n  ) AS region_rows_share_percent,\n  r.region_rows\n\nFROM information_schema.region_statistics r\n\nJOIN table_stats ts\n  ON r.table_id = ts.table_id\n\nJOIN information_schema.tables t\n  ON r.table_id = t.table_id\n\nLEFT JOIN information_schema.partitions p\n  ON p.table_schema = t.table_schema\n AND p.table_name = t.table_name\n AND p.greptime_partition_id = r.region_id\n\nWHERE r.region_role = 'Leader'\n\nORDER BY region_rows_share_percent DESC\nLIMIT 100;",
+          "refId": "A",
+          "sql": {
+            "columns": [
+              {
+                "parameters": [],
+                "type": "function"
+              }
+            ],
+            "groupBy": [
+              {
+                "property": {
+                  "type": "string"
+                },
+                "type": "groupBy"
+              }
+            ],
+            "limit": 50
+          }
+        }
+      ],
+      "title": "Hotspot Regions",
+      "type": "table"
+    },
+    {
+      "datasource": {
+        "type": "prometheus",
+        "uid": "${metrics}"
+      },
+      "description": "Write load of each datanode over time.",
+      "fieldConfig": {
+        "defaults": {
+          "color": {
+            "mode": "palette-classic"
+          },
+          "custom": {
+            "axisBorderShow": false,
+            "axisCenteredZero": false,
+            "axisColorMode": "text",
+            "axisLabel": "",
+            "axisPlacement": "auto",
+            "barAlignment": 0,
+            "barWidthFactor": 0.6,
+            "drawStyle": "line",
+            "fillOpacity": 0,
+            "gradientMode": "none",
+            "hideFrom": {
+              "legend": false,
+              "tooltip": false,
+              "viz": false
+            },
+            "insertNulls": false,
+            "lineInterpolation": "linear",
+            "lineWidth": 1,
+            "pointSize": 5,
+            "scaleDistribution": {
+              "type": "linear"
+            },
+            "showPoints": "auto",
+            "spanNulls": false,
+            "stacking": {
+              "group": "A",
+              "mode": "none"
+            },
+            "thresholdsStyle": {
+              "mode": "off"
+            }
+          },
+          "fieldMinMax": false,
+          "mappings": [],
+          "thresholds": {
+            "mode": "absolute",
+            "steps": [
+              {
+                "color": "green"
+              },
+              {
+                "color": "red",
+                "value": 80
+              }
+            ]
+          },
+          "unit": "binBps"
+        },
+        "overrides": [
+          {
+            "__systemRef": "hideSeriesFrom",
+            "matcher": {
+              "id": "byNames",
+              "options": {
+                "mode": "exclude",
+                "names": [
+                  "total_region_rows",
+                  "total_written_bytes_since_open",
+                  "total_disk_size",
+                  "total_sst_size",
+                  "total_sst_num",
+                  "total_memtable_size",
+                  "total_manifest_size",
+                  "total_index_size",
+                  "cluster_disk_ratio_percent",
+                  "cluster_sst_ratio_percent",
+                  "cluster_rows_ratio_percent",
+                  "max_region_disk_size",
+                  "max_region_rows",
+                  "leader_region_count",
+                  "disk_ratio_percent",
+                  "data_size"
+                ],
+                "prefix": "All except:",
+                "readOnly": true
+              }
+            },
+            "properties": []
+          }
+        ]
+      },
+      "gridPos": {
+        "h": 11,
+        "w": 12,
+        "x": 0,
+        "y": 83
+      },
+      "id": 368,
+      "options": {
+        "legend": {
+          "calcs": [],
+          "displayMode": "list",
+          "placement": "bottom",
+          "showLegend": true
+        },
+        "tooltip": {
+          "hideZeros": false,
+          "mode": "single",
+          "sort": "none"
+        }
+      },
+      "pluginVersion": "11.6.0",
+      "targets": [
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "disableTextWrap": false,
+          "editorMode": "builder",
+          "expr": "greptime_datanode_history_load",
+          "fullMetaSearch": false,
+          "includeNullMetadata": true,
+          "instant": false,
+          "legendFormat": "datanode-{{datanode_id}}({{instance}})",
+          "range": true,
+          "refId": "A",
+          "useBackend": false
+        }
+      ],
+      "title": "Datanode Load(Write)",
+      "type": "timeseries"
+    },
+    {
+      "datasource": {
+        "type": "prometheus",
+        "uid": "${metrics}"
+      },
+      "description": "Distribution of write load across datanodes.",
+      "fieldConfig": {
+        "defaults": {
+          "color": {
+            "mode": "palette-classic"
+          },
+          "custom": {
+            "hideFrom": {
+              "legend": false,
+              "tooltip": false,
+              "viz": false
+            }
+          },
+          "fieldMinMax": false,
+          "mappings": [],
+          "min": 0,
+          "unit": "binBps"
+        },
+        "overrides": []
+      },
+      "gridPos": {
+        "h": 11,
+        "w": 12,
+        "x": 12,
+        "y": 83
+      },
+      "id": 369,
+      "options": {
+        "displayLabels": [
+          "percent"
+        ],
+        "legend": {
+          "displayMode": "list",
+          "placement": "right",
+          "showLegend": true
+        },
+        "pieType": "donut",
+        "reduceOptions": {
+          "calcs": [
+            "lastNotNull"
+          ],
+          "fields": "",
+          "values": false
+        },
+        "tooltip": {
+          "hideZeros": false,
+          "mode": "single",
+          "sort": "none"
+        }
+      },
+      "pluginVersion": "11.6.0",
+      "targets": [
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "disableTextWrap": false,
+          "editorMode": "builder",
+          "exemplar": false,
+          "expr": "greptime_datanode_history_load",
+          "format": "time_series",
+          "fullMetaSearch": false,
+          "includeNullMetadata": true,
+          "instant": true,
+          "legendFormat": "datanode-{{datanode_id}}({{instance}})",
+          "range": false,
+          "refId": "A",
+          "useBackend": false
+        }
+      ],
+      "title": "Datanode Load(Write) Distribution",
+      "type": "piechart"
+    },
+    {
+      "datasource": {
+        "type": "mysql",
+        "uid": "${information_schema}"
+      },
+      "description": "Distribution of leader regions and data size across datanodes.",
+      "fieldConfig": {
+        "defaults": {
+          "color": {
+            "mode": "palette-classic"
+          },
+          "custom": {
+            "hideFrom": {
+              "legend": false,
+              "tooltip": false,
+              "viz": false
+            }
+          },
+          "fieldMinMax": false,
+          "mappings": [],
+          "unit": "bytes"
+        },
+        "overrides": [
+          {
+            "__systemRef": "hideSeriesFrom",
+            "matcher": {
+              "id": "byNames",
+              "options": {
+                "mode": "exclude",
+                "names": [
+                  "total_region_rows",
+                  "total_written_bytes_since_open",
+                  "total_disk_size",
+                  "total_sst_size",
+                  "total_sst_num",
+                  "total_memtable_size",
+                  "total_manifest_size",
+                  "total_index_size",
+                  "cluster_disk_ratio_percent",
+                  "cluster_sst_ratio_percent",
+                  "cluster_rows_ratio_percent",
+                  "max_region_disk_size",
+                  "max_region_rows",
+                  "leader_region_count",
+                  "disk_ratio_percent",
+                  "data_size"
+                ],
+                "prefix": "All except:",
+                "readOnly": true
+              }
+            },
+            "properties": [
+              {
+                "id": "custom.hideFrom",
+                "value": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": true
+                }
+              }
+            ]
+          }
+        ]
+      },
+      "gridPos": {
+        "h": 11,
+        "w": 12,
+        "x": 0,
+        "y": 94
+      },
+      "id": 367,
+      "options": {
+        "displayLabels": [
+          "value"
+        ],
+        "legend": {
+          "displayMode": "list",
+          "placement": "right",
+          "showLegend": true,
+          "values": [
+            "percent"
+          ]
+        },
+        "pieType": "pie",
+        "reduceOptions": {
+          "calcs": [
+            "lastNotNull"
+          ],
+          "fields": "/^data_size$/",
+          "values": true
+        },
+        "tooltip": {
+          "hideZeros": false,
+          "mode": "single",
+          "sort": "none"
+        }
+      },
+      "pluginVersion": "11.6.0",
+      "targets": [
+        {
+          "dataset": "information_schema",
+          "datasource": {
+            "type": "mysql",
+            "uid": "${information_schema}"
+          },
+          "editorMode": "code",
+          "format": "table",
+          "rawQuery": true,
+          "rawSql": "WITH leader_regions AS (\n  SELECT\n    CONCAT(\n      'datanode-',\n      p.peer_id,\n      ' (',\n      p.peer_addr,\n      ')'\n    ) AS datanode,\n    r.disk_size\n  FROM information_schema.region_statistics r\n  JOIN information_schema.region_peers p\n    ON r.region_id = p.region_id\n  WHERE r.region_role = 'Leader'\n    AND p.is_leader = 'Yes'\n)\n\nSELECT\n  datanode,\n  COUNT(*) AS leader_region_count,\n  SUM(disk_size) AS data_size\nFROM leader_regions\nGROUP BY datanode\nORDER BY data_size DESC;",
+          "refId": "A",
+          "sql": {
+            "columns": [
+              {
+                "parameters": [],
+                "type": "function"
+              }
+            ],
+            "groupBy": [
+              {
+                "property": {
+                  "type": "string"
+                },
+                "type": "groupBy"
+              }
+            ],
+            "limit": 50
+          }
+        }
+      ],
+      "title": "Datanode Data Distribution",
+      "type": "piechart"
     }
   ],
   "preload": false,
@@ -10354,12 +10828,12 @@
     ]
   },
   "time": {
-    "from": "now-1h",
+    "from": "now-2h",
     "to": "now"
   },
   "timepicker": {},
   "timezone": "",
-  "title": "GreptimeDB",
-  "uid": "dejf3k5e7g2kgb",
-  "version": 15
-}
+  "title": "GreptimeDB Copy",
+  "uid": "dflfbxbwvvchsa",
+  "version": 37
+}
\ No newline at end of file
diff --git a/grafana/dashboards/metrics/cluster/dashboard.md b/grafana/dashboards/metrics/cluster/dashboard.md
index c0bc49770b..c398ea7f90 100644
--- a/grafana/dashboards/metrics/cluster/dashboard.md
+++ b/grafana/dashboards/metrics/cluster/dashboard.md
@@ -142,3 +142,87 @@
   rate(greptime_trigger_save_alert_record_elapsed_bucket[$__rate_interval])
 )` | `timeseries` | Elapsed time to persist trigger alert records. | `prometheus` | `s` | `[{{instance}}]-[{{pod}}]-[{{storage_type}}]-p99` |
 | Save Alert Failure Rate | `rate(greptime_trigger_save_alert_record_failure_count[$__rate_interval])` | `timeseries` | Rate of failures when persisting trigger alert records. | `prometheus` | `none` | `__auto` |
+# Hotspot
+| Title | Query | Type | Description | Datasource | Unit | Legend Format |
+| --- | --- | --- | --- | --- | --- | --- |
+| Hotspot Regions | `WITH table_stats AS (
+  SELECT
+    table_id,
+    COUNT(*) AS region_count,
+    SUM(disk_size) AS total_disk_size,
+    SUM(region_rows) as total_region_rows
+  FROM information_schema.region_statistics
+  WHERE region_role = 'Leader'
+  GROUP BY table_id
+  HAVING COUNT(*) > 1
+)
+
+SELECT
+  t.table_schema,
+  t.table_name,
+
+  r.region_id,
+  t.table_id,
+  r.region_number,
+
+  p.partition_description,
+
+
+  ROUND(
+    r.disk_size * 100.0
+      / NULLIF(ts.total_disk_size, 0),
+    2
+  ) AS disk_size_share_percent,
+
+  r.disk_size,
+  
+  ROUND(
+    r.region_rows * 100.0
+      / NULLIF(ts.total_region_rows, 0),
+    2
+  ) AS region_rows_share_percent,
+  r.region_rows
+
+FROM information_schema.region_statistics r
+
+JOIN table_stats ts
+  ON r.table_id = ts.table_id
+
+JOIN information_schema.tables t
+  ON r.table_id = t.table_id
+
+LEFT JOIN information_schema.partitions p
+  ON p.table_schema = t.table_schema
+ AND p.table_name = t.table_name
+ AND p.greptime_partition_id = r.region_id
+
+WHERE r.region_role = 'Leader'
+
+ORDER BY region_rows_share_percent DESC
+LIMIT 100;` | `table` |  | `mysql` | -- | -- |
+| Datanode Load(Write) | `greptime_datanode_history_load` | `timeseries` | Write load of each datanode over time. | `prometheus` | `binBps` | `datanode-{{datanode_id}}({{instance}})` |
+| Datanode Load(Write) Distribution | `greptime_datanode_history_load` | `piechart` | Distribution of write load across datanodes. | `prometheus` | `binBps` | `datanode-{{datanode_id}}({{instance}})` |
+| Datanode Data Distribution | `WITH leader_regions AS (
+  SELECT
+    CONCAT(
+      'datanode-',
+      p.peer_id,
+      ' (',
+      p.peer_addr,
+      ')'
+    ) AS datanode,
+    r.disk_size
+  FROM information_schema.region_statistics r
+  JOIN information_schema.region_peers p
+    ON r.region_id = p.region_id
+  WHERE r.region_role = 'Leader'
+    AND p.is_leader = 'Yes'
+)
+
+SELECT
+  datanode,
+  COUNT(*) AS leader_region_count,
+  SUM(disk_size) AS data_size
+FROM leader_regions
+GROUP BY datanode
+ORDER BY data_size DESC;` | `piechart` | Distribution of leader regions and data size across datanodes. | `mysql` | `bytes` | -- |
diff --git a/grafana/dashboards/metrics/cluster/dashboard.yaml b/grafana/dashboards/metrics/cluster/dashboard.yaml
index 934cf58c0c..2617016083 100644
--- a/grafana/dashboards/metrics/cluster/dashboard.yaml
+++ b/grafana/dashboards/metrics/cluster/dashboard.yaml
@@ -1153,3 +1153,65 @@ groups:
                 type: prometheus
                 uid: ${metrics}
               legendFormat: __auto
+    - title: Hotspot
+      panels:
+        - title: Hotspot Regions
+          type: table
+          queries:
+            - expr: "WITH table_stats AS (\n  SELECT\n    table_id,\n    COUNT(*) AS region_count,\n    SUM(disk_size) AS total_disk_size,\n    SUM(region_rows) as total_region_rows\n  FROM information_schema.region_statistics\n  WHERE region_role = 'Leader'\n  GROUP BY table_id\n  HAVING COUNT(*) > 1\n)\n\nSELECT\n  t.table_schema,\n  t.table_name,\n\n  r.region_id,\n  t.table_id,\n  r.region_number,\n\n  p.partition_description,\n\n\n  ROUND(\n    r.disk_size * 100.0\n      / NULLIF(ts.total_disk_size, 0),\n    2\n  ) AS disk_size_share_percent,\n\n  r.disk_size,\n  \n  ROUND(\n    r.region_rows * 100.0\n      / NULLIF(ts.total_region_rows, 0),\n    2\n  ) AS region_rows_share_percent,\n  r.region_rows\n\nFROM information_schema.region_statistics r\n\nJOIN table_stats ts\n  ON r.table_id = ts.table_id\n\nJOIN information_schema.tables t\n  ON r.table_id = t.table_id\n\nLEFT JOIN information_schema.partitions p\n  ON p.table_schema = t.table_schema\n AND p.table_name = t.table_name\n AND p.greptime_partition_id = r.region_id\n\nWHERE r.region_role = 'Leader'\n\nORDER BY region_rows_share_percent DESC\nLIMIT 100;"
+              datasource:
+                type: mysql
+                uid: ${information_schema}
+        - title: Datanode Load(Write)
+          type: timeseries
+          description: Write load of each datanode over time.
+          unit: binBps
+          queries:
+            - expr: greptime_datanode_history_load
+              datasource:
+                type: prometheus
+                uid: ${metrics}
+              legendFormat: datanode-{{datanode_id}}({{instance}})
+        - title: Datanode Load(Write) Distribution
+          type: piechart
+          description: Distribution of write load across datanodes.
+          unit: binBps
+          queries:
+            - expr: greptime_datanode_history_load
+              datasource:
+                type: prometheus
+                uid: ${metrics}
+              legendFormat: datanode-{{datanode_id}}({{instance}})
+        - title: Datanode Data Distribution
+          type: piechart
+          description: Distribution of leader regions and data size across datanodes.
+          unit: bytes
+          queries:
+            - expr: |-
+                WITH leader_regions AS (
+                  SELECT
+                    CONCAT(
+                      'datanode-',
+                      p.peer_id,
+                      ' (',
+                      p.peer_addr,
+                      ')'
+                    ) AS datanode,
+                    r.disk_size
+                  FROM information_schema.region_statistics r
+                  JOIN information_schema.region_peers p
+                    ON r.region_id = p.region_id
+                  WHERE r.region_role = 'Leader'
+                    AND p.is_leader = 'Yes'
+                )
+
+                SELECT
+                  datanode,
+                  COUNT(*) AS leader_region_count,
+                  SUM(disk_size) AS data_size
+                FROM leader_regions
+                GROUP BY datanode
+                ORDER BY data_size DESC;
+              datasource:
+                type: mysql
+                uid: ${information_schema}
diff --git a/grafana/dashboards/metrics/standalone/dashboard.json b/grafana/dashboards/metrics/standalone/dashboard.json
index 28f38f9dd7..229bdd30ea 100644
--- a/grafana/dashboards/metrics/standalone/dashboard.json
+++ b/grafana/dashboards/metrics/standalone/dashboard.json
@@ -25,7 +25,7 @@
   "editable": true,
   "fiscalYearStartMonth": 0,
   "graphTooltip": 1,
-  "id": 34,
+  "id": 33,
   "links": [],
   "panels": [
     {
@@ -927,7 +927,7 @@
       "type": "stat"
     },
     {
-      "collapsed": true,
+      "collapsed": false,
       "gridPos": {
         "h": 1,
         "w": 24,
@@ -935,226 +935,226 @@
         "y": 9
       },
       "id": 275,
-      "panels": [
+      "panels": [],
+      "title": "Ingestion",
+      "type": "row"
+    },
+    {
+      "datasource": {
+        "type": "prometheus",
+        "uid": "${metrics}"
+      },
+      "description": "Total ingestion rate.\n\nHere we listed 3 primary protocols:\n\n- Prometheus remote write\n- Greptime's gRPC API (when using our ingest SDK)\n- Log ingestion http API\n",
+      "fieldConfig": {
+        "defaults": {
+          "color": {
+            "mode": "palette-classic"
+          },
+          "custom": {
+            "axisBorderShow": false,
+            "axisCenteredZero": false,
+            "axisColorMode": "text",
+            "axisLabel": "",
+            "axisPlacement": "auto",
+            "barAlignment": 0,
+            "barWidthFactor": 0.6,
+            "drawStyle": "line",
+            "fillOpacity": 0,
+            "gradientMode": "none",
+            "hideFrom": {
+              "legend": false,
+              "tooltip": false,
+              "viz": false
+            },
+            "insertNulls": false,
+            "lineInterpolation": "linear",
+            "lineWidth": 1,
+            "pointSize": 5,
+            "scaleDistribution": {
+              "type": "linear"
+            },
+            "showPoints": "auto",
+            "spanNulls": false,
+            "stacking": {
+              "group": "A",
+              "mode": "none"
+            },
+            "thresholdsStyle": {
+              "mode": "off"
+            }
+          },
+          "mappings": [],
+          "thresholds": {
+            "mode": "absolute",
+            "steps": [
+              {
+                "color": "green"
+              },
+              {
+                "color": "red",
+                "value": 80
+              }
+            ]
+          },
+          "unit": "rowsps"
+        },
+        "overrides": []
+      },
+      "gridPos": {
+        "h": 6,
+        "w": 24,
+        "x": 0,
+        "y": 10
+      },
+      "id": 193,
+      "options": {
+        "legend": {
+          "calcs": [
+            "mean"
+          ],
+          "displayMode": "table",
+          "placement": "bottom",
+          "showLegend": true
+        },
+        "tooltip": {
+          "hideZeros": false,
+          "mode": "single",
+          "sort": "none"
+        }
+      },
+      "pluginVersion": "11.6.0",
+      "targets": [
         {
           "datasource": {
             "type": "prometheus",
             "uid": "${metrics}"
           },
-          "description": "Total ingestion rate.\n\nHere we listed 3 primary protocols:\n\n- Prometheus remote write\n- Greptime's gRPC API (when using our ingest SDK)\n- Log ingestion http API\n",
-          "fieldConfig": {
-            "defaults": {
-              "color": {
-                "mode": "palette-classic"
-              },
-              "custom": {
-                "axisBorderShow": false,
-                "axisCenteredZero": false,
-                "axisColorMode": "text",
-                "axisLabel": "",
-                "axisPlacement": "auto",
-                "barAlignment": 0,
-                "drawStyle": "line",
-                "fillOpacity": 0,
-                "gradientMode": "none",
-                "hideFrom": {
-                  "legend": false,
-                  "tooltip": false,
-                  "viz": false
-                },
-                "insertNulls": false,
-                "lineInterpolation": "linear",
-                "lineWidth": 1,
-                "pointSize": 5,
-                "scaleDistribution": {
-                  "type": "linear"
-                },
-                "showPoints": "auto",
-                "spanNulls": false,
-                "stacking": {
-                  "group": "A",
-                  "mode": "none"
-                },
-                "thresholdsStyle": {
-                  "mode": "off"
-                }
-              },
-              "mappings": [],
-              "thresholds": {
-                "mode": "absolute",
-                "steps": [
-                  {
-                    "color": "green"
-                  },
-                  {
-                    "color": "red",
-                    "value": 80
-                  }
-                ]
-              },
-              "unit": "rowsps"
-            },
-            "overrides": []
+          "editorMode": "code",
+          "expr": "sum(rate(greptime_table_operator_ingest_rows{}[$__rate_interval]))",
+          "hide": false,
+          "instant": false,
+          "legendFormat": "ingestion",
+          "range": true,
+          "refId": "C"
+        }
+      ],
+      "title": "Total Ingestion Rate",
+      "type": "timeseries"
+    },
+    {
+      "datasource": {
+        "type": "prometheus",
+        "uid": "${metrics}"
+      },
+      "description": "Total ingestion rate.\n\nHere we listed 3 primary protocols:\n\n- Prometheus remote write\n- Greptime's gRPC API (when using our ingest SDK)\n- Log ingestion http API\n",
+      "fieldConfig": {
+        "defaults": {
+          "color": {
+            "mode": "palette-classic"
           },
-          "gridPos": {
-            "h": 6,
-            "w": 24,
-            "x": 0,
-            "y": 18
-          },
-          "id": 193,
-          "options": {
-            "legend": {
-              "calcs": [
-                "mean"
-              ],
-              "displayMode": "table",
-              "placement": "bottom",
-              "showLegend": true
+          "custom": {
+            "axisBorderShow": false,
+            "axisCenteredZero": false,
+            "axisColorMode": "text",
+            "axisLabel": "",
+            "axisPlacement": "auto",
+            "barAlignment": 0,
+            "barWidthFactor": 0.6,
+            "drawStyle": "line",
+            "fillOpacity": 0,
+            "gradientMode": "none",
+            "hideFrom": {
+              "legend": false,
+              "tooltip": false,
+              "viz": false
             },
-            "tooltip": {
-              "hideZeros": false,
-              "mode": "single",
-              "sort": "none"
+            "insertNulls": false,
+            "lineInterpolation": "linear",
+            "lineWidth": 1,
+            "pointSize": 5,
+            "scaleDistribution": {
+              "type": "linear"
+            },
+            "showPoints": "auto",
+            "spanNulls": false,
+            "stacking": {
+              "group": "A",
+              "mode": "none"
+            },
+            "thresholdsStyle": {
+              "mode": "off"
             }
           },
-          "pluginVersion": "12.0.0",
-          "targets": [
-            {
-              "datasource": {
-                "type": "prometheus",
-                "uid": "${metrics}"
+          "mappings": [],
+          "thresholds": {
+            "mode": "absolute",
+            "steps": [
+              {
+                "color": "green"
               },
-              "editorMode": "code",
-              "expr": "sum(rate(greptime_table_operator_ingest_rows{}[$__rate_interval]))",
-              "hide": false,
-              "instant": false,
-              "legendFormat": "ingestion",
-              "range": true,
-              "refId": "C"
-            }
+              {
+                "color": "red",
+                "value": 80
+              }
+            ]
+          },
+          "unit": "rowsps"
+        },
+        "overrides": []
+      },
+      "gridPos": {
+        "h": 6,
+        "w": 24,
+        "x": 0,
+        "y": 16
+      },
+      "id": 277,
+      "options": {
+        "legend": {
+          "calcs": [
+            "mean"
           ],
-          "title": "Total Ingestion Rate",
-          "type": "timeseries"
+          "displayMode": "table",
+          "placement": "bottom",
+          "showLegend": true
+        },
+        "tooltip": {
+          "hideZeros": false,
+          "mode": "single",
+          "sort": "none"
+        }
+      },
+      "pluginVersion": "11.6.0",
+      "targets": [
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "editorMode": "code",
+          "expr": "sum(rate(greptime_servers_http_logs_ingestion_counter[$__rate_interval]))",
+          "hide": false,
+          "instant": false,
+          "legendFormat": "http-logs",
+          "range": true,
+          "refId": "http_logs"
         },
         {
           "datasource": {
             "type": "prometheus",
             "uid": "${metrics}"
           },
-          "description": "Total ingestion rate.\n\nHere we listed 3 primary protocols:\n\n- Prometheus remote write\n- Greptime's gRPC API (when using our ingest SDK)\n- Log ingestion http API\n",
-          "fieldConfig": {
-            "defaults": {
-              "color": {
-                "mode": "palette-classic"
-              },
-              "custom": {
-                "axisBorderShow": false,
-                "axisCenteredZero": false,
-                "axisColorMode": "text",
-                "axisLabel": "",
-                "axisPlacement": "auto",
-                "barAlignment": 0,
-                "barWidthFactor": 0.6,
-                "drawStyle": "line",
-                "fillOpacity": 0,
-                "gradientMode": "none",
-                "hideFrom": {
-                  "legend": false,
-                  "tooltip": false,
-                  "viz": false
-                },
-                "insertNulls": false,
-                "lineInterpolation": "linear",
-                "lineWidth": 1,
-                "pointSize": 5,
-                "scaleDistribution": {
-                  "type": "linear"
-                },
-                "showPoints": "auto",
-                "spanNulls": false,
-                "stacking": {
-                  "group": "A",
-                  "mode": "none"
-                },
-                "thresholdsStyle": {
-                  "mode": "off"
-                }
-              },
-              "mappings": [],
-              "thresholds": {
-                "mode": "absolute",
-                "steps": [
-                  {
-                    "color": "green"
-                  },
-                  {
-                    "color": "red",
-                    "value": 80
-                  }
-                ]
-              },
-              "unit": "rowsps"
-            },
-            "overrides": []
-          },
-          "gridPos": {
-            "h": 6,
-            "w": 24,
-            "x": 0,
-            "y": 70
-          },
-          "id": 277,
-          "options": {
-            "legend": {
-              "calcs": [
-                "mean"
-              ],
-              "displayMode": "table",
-              "placement": "bottom",
-              "showLegend": true
-            },
-            "tooltip": {
-              "hideZeros": false,
-              "mode": "single",
-              "sort": "none"
-            }
-          },
-          "pluginVersion": "12.0.0",
-          "targets": [
-            {
-              "datasource": {
-                "type": "prometheus",
-                "uid": "${metrics}"
-              },
-              "editorMode": "code",
-              "expr": "sum(rate(greptime_servers_http_logs_ingestion_counter[$__rate_interval]))",
-              "hide": false,
-              "instant": false,
-              "legendFormat": "http-logs",
-              "range": true,
-              "refId": "http_logs"
-            },
-            {
-              "datasource": {
-                "type": "prometheus",
-                "uid": "${metrics}"
-              },
-              "editorMode": "code",
-              "expr": "sum(rate(greptime_servers_prometheus_remote_write_samples[$__rate_interval]))",
-              "hide": false,
-              "instant": false,
-              "legendFormat": "prometheus-remote-write",
-              "range": true,
-              "refId": "prometheus-remote-write"
-            }
-          ],
-          "title": "Ingestion Rate by Type",
-          "type": "timeseries"
+          "editorMode": "code",
+          "expr": "sum(rate(greptime_servers_prometheus_remote_write_samples[$__rate_interval]))",
+          "hide": false,
+          "instant": false,
+          "legendFormat": "prometheus-remote-write",
+          "range": true,
+          "refId": "prometheus-remote-write"
         }
       ],
-      "title": "Ingestion",
-      "type": "row"
+      "title": "Ingestion Rate by Type",
+      "type": "timeseries"
     },
     {
       "collapsed": true,
@@ -1162,7 +1162,7 @@
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 10
+        "y": 22
       },
       "id": 276,
       "panels": [
@@ -1231,7 +1231,7 @@
             "h": 6,
             "w": 24,
             "x": 0,
-            "y": 19
+            "y": 246
           },
           "id": 255,
           "options": {
@@ -1303,7 +1303,7 @@
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 11
+        "y": 23
       },
       "id": 274,
       "panels": [
@@ -1371,7 +1371,7 @@
             "h": 10,
             "w": 12,
             "x": 0,
-            "y": 20
+            "y": 247
           },
           "id": 256,
           "options": {
@@ -1486,7 +1486,7 @@
             "h": 10,
             "w": 12,
             "x": 12,
-            "y": 20
+            "y": 247
           },
           "id": 262,
           "options": {
@@ -1603,7 +1603,7 @@
             "h": 10,
             "w": 12,
             "x": 0,
-            "y": 30
+            "y": 257
           },
           "id": 266,
           "options": {
@@ -1718,7 +1718,7 @@
             "h": 10,
             "w": 12,
             "x": 12,
-            "y": 30
+            "y": 257
           },
           "id": 268,
           "options": {
@@ -1835,7 +1835,7 @@
             "h": 10,
             "w": 12,
             "x": 0,
-            "y": 40
+            "y": 267
           },
           "id": 269,
           "options": {
@@ -1950,7 +1950,7 @@
             "h": 10,
             "w": 12,
             "x": 12,
-            "y": 40
+            "y": 267
           },
           "id": 271,
           "options": {
@@ -2067,7 +2067,7 @@
             "h": 10,
             "w": 12,
             "x": 0,
-            "y": 50
+            "y": 277
           },
           "id": 272,
           "options": {
@@ -2182,7 +2182,7 @@
             "h": 10,
             "w": 12,
             "x": 12,
-            "y": 50
+            "y": 277
           },
           "id": 273,
           "options": {
@@ -2245,7 +2245,7 @@
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 12
+        "y": 24
       },
       "id": 280,
       "panels": [
@@ -2314,7 +2314,7 @@
             "h": 8,
             "w": 12,
             "x": 0,
-            "y": 61
+            "y": 288
           },
           "id": 281,
           "options": {
@@ -2415,7 +2415,7 @@
             "h": 8,
             "w": 12,
             "x": 12,
-            "y": 61
+            "y": 288
           },
           "id": 282,
           "options": {
@@ -2516,7 +2516,7 @@
             "h": 8,
             "w": 12,
             "x": 0,
-            "y": 69
+            "y": 296
           },
           "id": 283,
           "options": {
@@ -2617,7 +2617,7 @@
             "h": 8,
             "w": 12,
             "x": 12,
-            "y": 69
+            "y": 296
           },
           "id": 284,
           "options": {
@@ -2716,7 +2716,7 @@
             "h": 8,
             "w": 12,
             "x": 0,
-            "y": 77
+            "y": 304
           },
           "id": 285,
           "options": {
@@ -2817,7 +2817,7 @@
             "h": 8,
             "w": 12,
             "x": 12,
-            "y": 77
+            "y": 304
           },
           "id": 286,
           "options": {
@@ -2919,7 +2919,7 @@
             "h": 8,
             "w": 12,
             "x": 0,
-            "y": 85
+            "y": 312
           },
           "id": 287,
           "options": {
@@ -3020,7 +3020,7 @@
             "h": 8,
             "w": 12,
             "x": 12,
-            "y": 85
+            "y": 312
           },
           "id": 288,
           "options": {
@@ -3064,7 +3064,7 @@
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 13
+        "y": 25
       },
       "id": 289,
       "panels": [
@@ -3133,7 +3133,7 @@
             "h": 6,
             "w": 24,
             "x": 0,
-            "y": 62
+            "y": 289
           },
           "id": 292,
           "options": {
@@ -3234,7 +3234,7 @@
             "h": 8,
             "w": 12,
             "x": 0,
-            "y": 68
+            "y": 295
           },
           "id": 290,
           "options": {
@@ -3335,7 +3335,7 @@
             "h": 8,
             "w": 12,
             "x": 12,
-            "y": 68
+            "y": 295
           },
           "id": 291,
           "options": {
@@ -3464,7 +3464,7 @@
             "h": 8,
             "w": 12,
             "x": 0,
-            "y": 76
+            "y": 303
           },
           "id": 336,
           "options": {
@@ -3518,2519 +3518,2520 @@
       "type": "row"
     },
     {
-      "collapsed": false,
+      "collapsed": true,
       "gridPos": {
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 14
+        "y": 26
       },
       "id": 293,
-      "panels": [],
+      "panels": [
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Request QPS per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "ops"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 193
+          },
+          "id": 294,
+          "options": {
+            "legend": {
+              "calcs": [
+                "mean"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod, type) (rate(greptime_mito_handle_request_elapsed_count{}[$__rate_interval]))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Request OPS per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Request P99 per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 193
+          },
+          "id": 295,
+          "options": {
+            "legend": {
+              "calcs": [
+                "mean"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(instance, pod, le, type) (rate(greptime_mito_handle_request_elapsed_bucket{}[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Request P99 per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Write Buffer per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "decbytes"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 225
+          },
+          "id": 296,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_mito_write_buffer_bytes{}",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Write Buffer per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Ingestion size by row counts.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "rowsps"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 225
+          },
+          "id": 297,
+          "options": {
+            "legend": {
+              "calcs": [
+                "mean"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by (instance, pod) (rate(greptime_mito_write_rows_total{}[$__rate_interval]))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Write Rows per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Flush QPS per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "ops"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 233
+          },
+          "id": 298,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod, reason) (rate(greptime_mito_flush_requests_total{}[$__rate_interval]))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{reason}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Flush OPS per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Write Stall per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              }
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 233
+          },
+          "id": 299,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod) (greptime_mito_write_stall_total{})",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Write Stall per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Read Stage OPS per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "ops"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 241
+          },
+          "id": 300,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod) (rate(greptime_mito_read_stage_elapsed_count{ stage=\"total\"}[$__rate_interval]))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Read Stage OPS per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Read Stage P99 per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": [
+              {
+                "__systemRef": "hideSeriesFrom",
+                "matcher": {
+                  "id": "byNames",
+                  "options": {
+                    "mode": "exclude",
+                    "names": [
+                      "[10.39.145.87:4000]-[mycluster-datanode-1]-[cache_miss_read]"
+                    ],
+                    "prefix": "All except:",
+                    "readOnly": true
+                  }
+                },
+                "properties": [
+                  {
+                    "id": "custom.hideFrom",
+                    "value": {
+                      "legend": false,
+                      "tooltip": false,
+                      "viz": true
+                    }
+                  }
+                ]
+              }
+            ]
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 241
+          },
+          "id": 301,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(instance, pod, le, stage) (rate(greptime_mito_read_stage_elapsed_bucket{}[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Read Stage P99 per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Write Stage P99 per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 249
+          },
+          "id": 302,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true,
+              "sortBy": "Last *",
+              "sortDesc": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(instance, pod, le, stage) (rate(greptime_mito_write_stage_elapsed_bucket{}[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Write Stage P99 per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Compaction OPS per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "ops"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 249
+          },
+          "id": 303,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod) (rate(greptime_mito_compaction_total_elapsed_count{}[$__rate_interval]))",
+              "instant": false,
+              "legendFormat": "[{{ instance }}]-[{{pod}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Compaction OPS per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Compaction latency by stage",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 257
+          },
+          "id": 304,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(instance, pod, le, stage) (rate(greptime_mito_compaction_stage_elapsed_bucket{}[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-p99",
+              "range": true,
+              "refId": "A"
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod, stage) (rate(greptime_mito_compaction_stage_elapsed_sum{}[$__rate_interval]))/sum by(instance, pod, stage) (rate(greptime_mito_compaction_stage_elapsed_count{}[$__rate_interval]))",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-avg",
+              "range": true,
+              "refId": "B"
+            }
+          ],
+          "title": "Compaction Elapsed Time per Instance by Stage",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Compaction P99 per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 257
+          },
+          "id": 305,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(instance, pod, le,stage) (rate(greptime_mito_compaction_total_elapsed_bucket{}[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-compaction",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Compaction P99 per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Write-ahead logs write size as bytes. This chart includes stats of p95 and p99 size by instance, total WAL write rate.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "bytes"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 265
+          },
+          "id": 306,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.95, sum by(le,instance, pod) (rate(raft_engine_write_size_bucket[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-req-size-p95",
+              "range": true,
+              "refId": "A"
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(le,instance,pod) (rate(raft_engine_write_size_bucket[$__rate_interval])))",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-req-size-p99",
+              "range": true,
+              "refId": "B"
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by (instance, pod)(rate(raft_engine_write_size_sum[$__rate_interval]))",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-throughput",
+              "range": true,
+              "refId": "C"
+            }
+          ],
+          "title": "WAL write size",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Cached Bytes per Instance.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "decbytes"
+            },
+            "overrides": [
+              {
+                "__systemRef": "hideSeriesFrom",
+                "matcher": {
+                  "id": "byNames",
+                  "options": {
+                    "mode": "exclude",
+                    "names": [
+                      "[10.39.145.79:4000]-[mycluster-datanode-0]-[file]"
+                    ],
+                    "prefix": "All except:",
+                    "readOnly": true
+                  }
+                },
+                "properties": [
+                  {
+                    "id": "custom.hideFrom",
+                    "value": {
+                      "legend": false,
+                      "tooltip": false,
+                      "viz": true
+                    }
+                  }
+                ]
+              }
+            ]
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 265
+          },
+          "id": 307,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "multi",
+              "sort": "desc"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_mito_cache_bytes{}",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Cached Bytes per Instance",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Ongoing compaction task count",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 273
+          },
+          "id": 308,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_mito_inflight_compaction_count",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Inflight Compaction",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Raft engine (local disk) log store sync latency, p99",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 273
+          },
+          "id": 310,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(le, type, node, instance, pod) (rate(raft_engine_sync_log_duration_seconds_bucket[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-p99",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "WAL sync duration seconds",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Write-ahead log operations latency at p99",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 281
+          },
+          "id": 311,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(le,logstore,optype,instance, pod) (rate(greptime_logstore_op_elapsed_bucket[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{logstore}}]-[{{optype}}]-p99",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Log Store op duration seconds",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Ongoing flush task count",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 281
+          },
+          "id": 312,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_mito_inflight_flush_count",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Inflight Flush",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Compaction oinput output bytes",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "bytes"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 289
+          },
+          "id": 335,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod) (greptime_mito_compaction_input_bytes)",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-input",
+              "range": true,
+              "refId": "A"
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod) (greptime_mito_compaction_output_bytes)",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-output",
+              "range": true,
+              "refId": "B"
+            }
+          ],
+          "title": "Compaction Input/Output Bytes",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Per-stage elapsed time for region worker to handle bulk insert region requests.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": [
+              {
+                "__systemRef": "hideSeriesFrom",
+                "matcher": {
+                  "id": "byNames",
+                  "options": {
+                    "mode": "exclude",
+                    "names": [
+                      "[192.168.50.8:4000]-[]-[prepare_bulk_request]-AVG",
+                      "[minipc-1:4000]-[]-[bulk_extend]-AVG",
+                      "[minipc-1:4000]-[]-[prepare_bulk_request]-AVG",
+                      "[minipc-1:4000]-[]-[process_bulk_req]-AVG",
+                      "[minipc-1:4000]-[]-[write_bulk]-AVG"
+                    ],
+                    "prefix": "All except:",
+                    "readOnly": true
+                  }
+                },
+                "properties": [
+                  {
+                    "id": "custom.hideFrom",
+                    "value": {
+                      "legend": false,
+                      "tooltip": false,
+                      "viz": true
+                    }
+                  }
+                ]
+              }
+            ]
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 289
+          },
+          "id": 337,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.95, sum by(le,instance, stage, pod) (rate(greptime_region_worker_handle_write_bucket[$__rate_interval])))",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-P95",
+              "range": true,
+              "refId": "A"
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, stage, pod) (rate(greptime_region_worker_handle_write_sum[$__rate_interval]))/sum by(instance, stage, pod) (rate(greptime_region_worker_handle_write_count[$__rate_interval]))",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-AVG",
+              "range": true,
+              "refId": "B"
+            }
+          ],
+          "title": "Region Worker Handle Bulk Insert Requests",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Compaction oinput output bytes",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 297
+          },
+          "id": 348,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod) (greptime_mito_memtable_active_series_count)",
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-series",
+              "range": true,
+              "refId": "A"
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(instance, pod) (greptime_mito_memtable_field_builder_count)",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-field_builders",
+              "range": true,
+              "refId": "B"
+            }
+          ],
+          "title": "Active Series and Field Builders Count",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Per-stage elapsed time for region worker to decode requests.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "points",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 297
+          },
+          "id": 338,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "builder",
+              "expr": "histogram_quantile(0.95, sum by(le, instance, stage, pod) (rate(greptime_datanode_convert_region_request_bucket[$__rate_interval])))",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-P95",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by(le,instance, stage, pod) (rate(greptime_datanode_convert_region_request_sum[$__rate_interval]))/sum by(le,instance, stage, pod) (rate(greptime_datanode_convert_region_request_count[$__rate_interval]))",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-AVG",
+              "range": true,
+              "refId": "B"
+            }
+          ],
+          "title": "Region Worker Convert Requests",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "The local cache miss of the datanode.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              }
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 305
+          },
+          "id": 349,
+          "options": {
+            "legend": {
+              "calcs": [
+                "lastNotNull"
+              ],
+              "displayMode": "table",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "sum by (instance,pod, type) (rate(greptime_mito_cache_miss{}[$__rate_interval]))",
+              "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Cache Miss",
+          "type": "timeseries"
+        }
+      ],
       "title": "Mito Engine",
       "type": "row"
     },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Request QPS per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "ops"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 15
-      },
-      "id": 294,
-      "options": {
-        "legend": {
-          "calcs": [
-            "mean"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod, type) (rate(greptime_mito_handle_request_elapsed_count{}[$__rate_interval]))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Request OPS per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Request P99 per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 15
-      },
-      "id": 295,
-      "options": {
-        "legend": {
-          "calcs": [
-            "mean"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(instance, pod, le, type) (rate(greptime_mito_handle_request_elapsed_bucket{}[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Request P99 per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Write Buffer per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "decbytes"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 23
-      },
-      "id": 296,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_mito_write_buffer_bytes{}",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Write Buffer per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Ingestion size by row counts.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "rowsps"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 23
-      },
-      "id": 297,
-      "options": {
-        "legend": {
-          "calcs": [
-            "mean"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by (instance, pod) (rate(greptime_mito_write_rows_total{}[$__rate_interval]))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Write Rows per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Flush QPS per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "ops"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 31
-      },
-      "id": 298,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod, reason) (rate(greptime_mito_flush_requests_total{}[$__rate_interval]))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{reason}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Flush OPS per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Write Stall per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          }
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 31
-      },
-      "id": 299,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod) (greptime_mito_write_stall_total{})",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Write Stall per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Read Stage OPS per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "ops"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 39
-      },
-      "id": 300,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod) (rate(greptime_mito_read_stage_elapsed_count{ stage=\"total\"}[$__rate_interval]))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Read Stage OPS per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Read Stage P99 per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": [
-          {
-            "__systemRef": "hideSeriesFrom",
-            "matcher": {
-              "id": "byNames",
-              "options": {
-                "mode": "exclude",
-                "names": [
-                  "[10.39.145.87:4000]-[mycluster-datanode-1]-[cache_miss_read]"
-                ],
-                "prefix": "All except:",
-                "readOnly": true
-              }
-            },
-            "properties": [
-              {
-                "id": "custom.hideFrom",
-                "value": {
-                  "legend": false,
-                  "tooltip": false,
-                  "viz": true
-                }
-              }
-            ]
-          }
-        ]
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 39
-      },
-      "id": 301,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(instance, pod, le, stage) (rate(greptime_mito_read_stage_elapsed_bucket{}[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Read Stage P99 per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Write Stage P99 per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 47
-      },
-      "id": 302,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true,
-          "sortBy": "Last *",
-          "sortDesc": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(instance, pod, le, stage) (rate(greptime_mito_write_stage_elapsed_bucket{}[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Write Stage P99 per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Compaction OPS per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "ops"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 47
-      },
-      "id": 303,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod) (rate(greptime_mito_compaction_total_elapsed_count{}[$__rate_interval]))",
-          "instant": false,
-          "legendFormat": "[{{ instance }}]-[{{pod}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Compaction OPS per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Compaction latency by stage",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 55
-      },
-      "id": 304,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(instance, pod, le, stage) (rate(greptime_mito_compaction_stage_elapsed_bucket{}[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-p99",
-          "range": true,
-          "refId": "A"
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod, stage) (rate(greptime_mito_compaction_stage_elapsed_sum{}[$__rate_interval]))/sum by(instance, pod, stage) (rate(greptime_mito_compaction_stage_elapsed_count{}[$__rate_interval]))",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-avg",
-          "range": true,
-          "refId": "B"
-        }
-      ],
-      "title": "Compaction Elapsed Time per Instance by Stage",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Compaction P99 per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 55
-      },
-      "id": 305,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(instance, pod, le,stage) (rate(greptime_mito_compaction_total_elapsed_bucket{}[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-compaction",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Compaction P99 per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Write-ahead logs write size as bytes. This chart includes stats of p95 and p99 size by instance, total WAL write rate.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "bytes"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 63
-      },
-      "id": 306,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.95, sum by(le,instance, pod) (rate(raft_engine_write_size_bucket[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-req-size-p95",
-          "range": true,
-          "refId": "A"
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(le,instance,pod) (rate(raft_engine_write_size_bucket[$__rate_interval])))",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-req-size-p99",
-          "range": true,
-          "refId": "B"
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by (instance, pod)(rate(raft_engine_write_size_sum[$__rate_interval]))",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-throughput",
-          "range": true,
-          "refId": "C"
-        }
-      ],
-      "title": "WAL write size",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Cached Bytes per Instance.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "decbytes"
-        },
-        "overrides": [
-          {
-            "__systemRef": "hideSeriesFrom",
-            "matcher": {
-              "id": "byNames",
-              "options": {
-                "mode": "exclude",
-                "names": [
-                  "[10.39.145.79:4000]-[mycluster-datanode-0]-[file]"
-                ],
-                "prefix": "All except:",
-                "readOnly": true
-              }
-            },
-            "properties": [
-              {
-                "id": "custom.hideFrom",
-                "value": {
-                  "legend": false,
-                  "tooltip": false,
-                  "viz": true
-                }
-              }
-            ]
-          }
-        ]
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 63
-      },
-      "id": 307,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "multi",
-          "sort": "desc"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_mito_cache_bytes{}",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Cached Bytes per Instance",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Ongoing compaction task count",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 71
-      },
-      "id": 308,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_mito_inflight_compaction_count",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Inflight Compaction",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Raft engine (local disk) log store sync latency, p99",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 71
-      },
-      "id": 310,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(le, type, node, instance, pod) (rate(raft_engine_sync_log_duration_seconds_bucket[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-p99",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "WAL sync duration seconds",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Write-ahead log operations latency at p99",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 79
-      },
-      "id": 311,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(le,logstore,optype,instance, pod) (rate(greptime_logstore_op_elapsed_bucket[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{logstore}}]-[{{optype}}]-p99",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Log Store op duration seconds",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Ongoing flush task count",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 79
-      },
-      "id": 312,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_mito_inflight_flush_count",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Inflight Flush",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Compaction oinput output bytes",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "bytes"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 87
-      },
-      "id": 335,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod) (greptime_mito_compaction_input_bytes)",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-input",
-          "range": true,
-          "refId": "A"
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod) (greptime_mito_compaction_output_bytes)",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-output",
-          "range": true,
-          "refId": "B"
-        }
-      ],
-      "title": "Compaction Input/Output Bytes",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Per-stage elapsed time for region worker to handle bulk insert region requests.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": [
-          {
-            "__systemRef": "hideSeriesFrom",
-            "matcher": {
-              "id": "byNames",
-              "options": {
-                "mode": "exclude",
-                "names": [
-                  "[192.168.50.8:4000]-[]-[prepare_bulk_request]-AVG",
-                  "[minipc-1:4000]-[]-[bulk_extend]-AVG",
-                  "[minipc-1:4000]-[]-[prepare_bulk_request]-AVG",
-                  "[minipc-1:4000]-[]-[process_bulk_req]-AVG",
-                  "[minipc-1:4000]-[]-[write_bulk]-AVG"
-                ],
-                "prefix": "All except:",
-                "readOnly": true
-              }
-            },
-            "properties": [
-              {
-                "id": "custom.hideFrom",
-                "value": {
-                  "legend": false,
-                  "tooltip": false,
-                  "viz": true
-                }
-              }
-            ]
-          }
-        ]
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 87
-      },
-      "id": 337,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.95, sum by(le,instance, stage, pod) (rate(greptime_region_worker_handle_write_bucket[$__rate_interval])))",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-P95",
-          "range": true,
-          "refId": "A"
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, stage, pod) (rate(greptime_region_worker_handle_write_sum[$__rate_interval]))/sum by(instance, stage, pod) (rate(greptime_region_worker_handle_write_count[$__rate_interval]))",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-AVG",
-          "range": true,
-          "refId": "B"
-        }
-      ],
-      "title": "Region Worker Handle Bulk Insert Requests",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Compaction oinput output bytes",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 95
-      },
-      "id": 348,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod) (greptime_mito_memtable_active_series_count)",
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-series",
-          "range": true,
-          "refId": "A"
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(instance, pod) (greptime_mito_memtable_field_builder_count)",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-field_builders",
-          "range": true,
-          "refId": "B"
-        }
-      ],
-      "title": "Active Series and Field Builders Count",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Per-stage elapsed time for region worker to decode requests.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "points",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 95
-      },
-      "id": 338,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "builder",
-          "expr": "histogram_quantile(0.95, sum by(le, instance, stage, pod) (rate(greptime_datanode_convert_region_request_bucket[$__rate_interval])))",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-P95",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by(le,instance, stage, pod) (rate(greptime_datanode_convert_region_request_sum[$__rate_interval]))/sum by(le,instance, stage, pod) (rate(greptime_datanode_convert_region_request_count[$__rate_interval]))",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{stage}}]-AVG",
-          "range": true,
-          "refId": "B"
-        }
-      ],
-      "title": "Region Worker Convert Requests",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "The local cache miss of the datanode.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          }
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 103
-      },
-      "id": 349,
-      "options": {
-        "legend": {
-          "calcs": [
-            "lastNotNull"
-          ],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "sum by (instance,pod, type) (rate(greptime_mito_cache_miss{}[$__rate_interval]))",
-          "legendFormat": "[{{instance}}]-[{{pod}}]-[{{type}}]",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Cache Miss",
-      "type": "timeseries"
-    },
     {
       "collapsed": true,
       "gridPos": {
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 111
+        "y": 27
       },
       "id": 313,
       "panels": [
@@ -6098,7 +6099,7 @@
             "h": 10,
             "w": 24,
             "x": 0,
-            "y": 232
+            "y": 459
           },
           "id": 314,
           "options": {
@@ -6198,7 +6199,7 @@
             "h": 7,
             "w": 12,
             "x": 0,
-            "y": 242
+            "y": 469
           },
           "id": 315,
           "options": {
@@ -6298,7 +6299,7 @@
             "h": 7,
             "w": 12,
             "x": 12,
-            "y": 242
+            "y": 469
           },
           "id": 316,
           "options": {
@@ -6398,7 +6399,7 @@
             "h": 7,
             "w": 12,
             "x": 0,
-            "y": 249
+            "y": 476
           },
           "id": 317,
           "options": {
@@ -6498,7 +6499,7 @@
             "h": 7,
             "w": 12,
             "x": 12,
-            "y": 249
+            "y": 476
           },
           "id": 318,
           "options": {
@@ -6598,7 +6599,7 @@
             "h": 7,
             "w": 12,
             "x": 0,
-            "y": 256
+            "y": 483
           },
           "id": 319,
           "options": {
@@ -6699,7 +6700,7 @@
             "h": 7,
             "w": 12,
             "x": 12,
-            "y": 256
+            "y": 483
           },
           "id": 320,
           "options": {
@@ -6800,7 +6801,7 @@
             "h": 7,
             "w": 12,
             "x": 0,
-            "y": 263
+            "y": 490
           },
           "id": 321,
           "options": {
@@ -6900,7 +6901,7 @@
             "h": 7,
             "w": 12,
             "x": 12,
-            "y": 263
+            "y": 490
           },
           "id": 322,
           "options": {
@@ -7000,7 +7001,7 @@
             "h": 7,
             "w": 12,
             "x": 0,
-            "y": 270
+            "y": 497
           },
           "id": 323,
           "options": {
@@ -7103,7 +7104,7 @@
             "h": 7,
             "w": 12,
             "x": 12,
-            "y": 270
+            "y": 497
           },
           "id": 334,
           "options": {
@@ -7144,1827 +7145,1829 @@
       "type": "row"
     },
     {
-      "collapsed": false,
+      "collapsed": true,
       "gridPos": {
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 112
+        "y": 28
       },
       "id": 351,
-      "panels": [],
+      "panels": [
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Triggered region flush total",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 195
+          },
+          "id": 350,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "meta_triggered_region_flush_total",
+              "instant": false,
+              "legendFormat": "{{pod}}-{{topic_name}}",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Triggered region flush total",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Triggered region checkpoint total",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 195
+          },
+          "id": 352,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "meta_triggered_region_checkpoint_total",
+              "instant": false,
+              "legendFormat": "{{pod}}-{{topic_name}}",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Triggered region checkpoint total",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Topic estimated max replay size",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "bytes"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 227
+          },
+          "id": 353,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "meta_topic_estimated_replay_size",
+              "instant": false,
+              "legendFormat": "{{pod}}-{{topic_name}}",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Topic estimated replay size",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Kafka logstore's bytes traffic",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "bytes"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 227
+          },
+          "id": 354,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "rate(greptime_logstore_kafka_client_bytes_total[$__rate_interval])",
+              "instant": false,
+              "legendFormat": "{{pod}}-{{logstore}}",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Kafka logstore's bytes traffic",
+          "type": "timeseries"
+        }
+      ],
       "title": "Remote WAL",
       "type": "row"
     },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Triggered region flush total",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 113
-      },
-      "id": 350,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "meta_triggered_region_flush_total",
-          "instant": false,
-          "legendFormat": "{{pod}}-{{topic_name}}",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Triggered region flush total",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Triggered region checkpoint total",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 113
-      },
-      "id": 352,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "meta_triggered_region_checkpoint_total",
-          "instant": false,
-          "legendFormat": "{{pod}}-{{topic_name}}",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Triggered region checkpoint total",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Topic estimated max replay size",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "bytes"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 121
-      },
-      "id": 353,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "meta_topic_estimated_replay_size",
-          "instant": false,
-          "legendFormat": "{{pod}}-{{topic_name}}",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Topic estimated replay size",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Kafka logstore's bytes traffic",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "bytes"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 121
-      },
-      "id": 354,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "rate(greptime_logstore_kafka_client_bytes_total[$__rate_interval])",
-          "instant": false,
-          "legendFormat": "{{pod}}-{{logstore}}",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Kafka logstore's bytes traffic",
-      "type": "timeseries"
-    },
-    {
-      "collapsed": false,
-      "gridPos": {
-        "h": 1,
-        "w": 24,
-        "x": 0,
-        "y": 129
-      },
-      "id": 324,
-      "panels": [],
-      "title": "Metasrv",
-      "type": "row"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Counter of region migration by source and destination",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "thresholds"
-          },
-          "custom": {
-            "axisPlacement": "auto",
-            "fillOpacity": 70,
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "lineWidth": 1
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          }
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 130
-      },
-      "id": 325,
-      "options": {
-        "colWidth": 0.9,
-        "legend": {
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": false
-        },
-        "rowHeight": 0.9,
-        "showValue": "auto",
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_meta_region_migration_stat{datanode_type=\"src\"}",
-          "instant": false,
-          "legendFormat": "from-datanode-{{datanode_id}}",
-          "range": true,
-          "refId": "A"
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_meta_region_migration_stat{datanode_type=\"desc\"}",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "to-datanode-{{datanode_id}}",
-          "range": true,
-          "refId": "B"
-        }
-      ],
-      "title": "Region migration datanode",
-      "type": "status-history"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Counter of region migration error",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 130
-      },
-      "id": 326,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_meta_region_migration_error",
-          "instant": false,
-          "legendFormat": "{{pod}}-{{state}}-{{error_type}}",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Region migration error",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "binBps"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 138
-      },
-      "id": 327,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_datanode_load",
-          "instant": false,
-          "legendFormat": "Datanode-{{datanode_id}}-writeload",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Datanode load",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Displays the rate of SQL executions processed by the Meta service using the RDS backend.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 138
-      },
-      "id": 339,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "rate(greptime_meta_rds_pg_sql_execute_elapsed_ms_count[$__rate_interval])",
-          "instant": false,
-          "legendFormat": "{{pod}} {{op}} {{type}} {{result}} ",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Rate of SQL Executions (RDS)",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Measures the response time of SQL executions via the RDS backend. ",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "ms"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 146
-      },
-      "id": 340,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.90, sum by(pod, op, type, result, le) (rate(greptime_meta_rds_pg_sql_execute_elapsed_ms_bucket[$__rate_interval])))",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "{{pod}} {{op}} {{type}} {{result}} p90",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        }
-      ],
-      "title": "SQL Execution Latency (RDS)",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Shows latency of Meta handlers by pod and handler name, useful for monitoring handler performance and detecting latency spikes.\n",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 146
-      },
-      "id": 341,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.90, sum by(pod, le, name) (\n  rate(greptime_meta_handler_execute_bucket[$__rate_interval])\n))",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "{{pod}} {{name}} p90",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        }
-      ],
-      "title": "Handler Execution Latency",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Shows p90 heartbeat message sizes, helping track network usage and identify anomalies in heartbeat payload.\n",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "bytes"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 154
-      },
-      "id": 342,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, sum by(pod, le) (greptime_meta_heartbeat_stat_memory_size_bucket))",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "{{pod}}",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        }
-      ],
-      "title": "Heartbeat Packet Size",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 154
-      },
-      "id": 345,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "rate(greptime_meta_heartbeat_rate[$__rate_interval])",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "{{pod}}",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        }
-      ],
-      "title": "Meta Heartbeat Receive Rate",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 162
-      },
-      "id": 343,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.99, sum by(pod, le, op, target) (greptime_meta_kv_request_elapsed_bucket))",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "{{pod}}-{{op}} p99",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        }
-      ],
-      "title": "Meta KV Ops Latency",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "none"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 162
-      },
-      "id": 346,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "rate(greptime_meta_kv_request_elapsed_count[$__rate_interval])",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "{{pod}}-{{op}} p99",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        }
-      ],
-      "title": "Rate of meta KV Ops",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 170
-      },
-      "id": 347,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_tables_bucket))",
-          "fullMetaSearch": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "CreateLogicalTables-{{step}} p90",
-          "range": true,
-          "refId": "A",
-          "useBackend": false
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_table))",
-          "fullMetaSearch": false,
-          "hide": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "CreateTable-{{step}} p90",
-          "range": true,
-          "refId": "B",
-          "useBackend": false
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_view))",
-          "fullMetaSearch": false,
-          "hide": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "CreateView-{{step}} p90",
-          "range": true,
-          "refId": "C",
-          "useBackend": false
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_flow))",
-          "fullMetaSearch": false,
-          "hide": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "CreateFlow-{{step}} p90",
-          "range": true,
-          "refId": "D",
-          "useBackend": false
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_drop_table))",
-          "fullMetaSearch": false,
-          "hide": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "DropTable-{{step}} p90",
-          "range": true,
-          "refId": "E",
-          "useBackend": false
-        },
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "disableTextWrap": false,
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_alter_table))",
-          "fullMetaSearch": false,
-          "hide": false,
-          "includeNullMetadata": true,
-          "instant": false,
-          "legendFormat": "AlterTable-{{step}} p90",
-          "range": true,
-          "refId": "F",
-          "useBackend": false
-        }
-      ],
-      "title": "DDL Latency",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Reconciliation stats",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 12,
-        "y": 170
-      },
-      "id": 355,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "greptime_meta_reconciliation_stats",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "{{pod}}-{{table_type}}-{{type}}",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Reconciliation stats",
-      "type": "timeseries"
-    },
-    {
-      "datasource": {
-        "type": "prometheus",
-        "uid": "${metrics}"
-      },
-      "description": "Elapsed of Reconciliation steps",
-      "fieldConfig": {
-        "defaults": {
-          "color": {
-            "mode": "palette-classic"
-          },
-          "custom": {
-            "axisBorderShow": false,
-            "axisCenteredZero": false,
-            "axisColorMode": "text",
-            "axisLabel": "",
-            "axisPlacement": "auto",
-            "barAlignment": 0,
-            "barWidthFactor": 0.6,
-            "drawStyle": "line",
-            "fillOpacity": 0,
-            "gradientMode": "none",
-            "hideFrom": {
-              "legend": false,
-              "tooltip": false,
-              "viz": false
-            },
-            "insertNulls": false,
-            "lineInterpolation": "linear",
-            "lineWidth": 1,
-            "pointSize": 5,
-            "scaleDistribution": {
-              "type": "linear"
-            },
-            "showPoints": "auto",
-            "spanNulls": false,
-            "stacking": {
-              "group": "A",
-              "mode": "none"
-            },
-            "thresholdsStyle": {
-              "mode": "off"
-            }
-          },
-          "mappings": [],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              {
-                "color": "green"
-              },
-              {
-                "color": "red",
-                "value": 80
-              }
-            ]
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": {
-        "h": 8,
-        "w": 12,
-        "x": 0,
-        "y": 178
-      },
-      "id": 356,
-      "options": {
-        "legend": {
-          "calcs": [],
-          "displayMode": "list",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": {
-          "hideZeros": false,
-          "mode": "single",
-          "sort": "none"
-        }
-      },
-      "pluginVersion": "11.6.0",
-      "targets": [
-        {
-          "datasource": {
-            "type": "prometheus",
-            "uid": "${metrics}"
-          },
-          "editorMode": "code",
-          "expr": "histogram_quantile(0.9, greptime_meta_reconciliation_procedure_bucket)",
-          "hide": false,
-          "instant": false,
-          "legendFormat": "{{procedure_name}}-{{step}}-P90",
-          "range": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Reconciliation steps",
-      "type": "timeseries"
-    },
     {
       "collapsed": true,
       "gridPos": {
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 186
+        "y": 29
+      },
+      "id": 324,
+      "panels": [
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Counter of region migration by source and destination",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "thresholds"
+              },
+              "custom": {
+                "axisPlacement": "auto",
+                "fillOpacity": 70,
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "lineWidth": 1
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              }
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 196
+          },
+          "id": 325,
+          "options": {
+            "colWidth": 0.9,
+            "legend": {
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": false
+            },
+            "rowHeight": 0.9,
+            "showValue": "auto",
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_meta_region_migration_stat{datanode_type=\"src\"}",
+              "instant": false,
+              "legendFormat": "from-datanode-{{datanode_id}}",
+              "range": true,
+              "refId": "A"
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_meta_region_migration_stat{datanode_type=\"desc\"}",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "to-datanode-{{datanode_id}}",
+              "range": true,
+              "refId": "B"
+            }
+          ],
+          "title": "Region migration datanode",
+          "type": "status-history"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Counter of region migration error",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 196
+          },
+          "id": 326,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_meta_region_migration_error",
+              "instant": false,
+              "legendFormat": "{{pod}}-{{state}}-{{error_type}}",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Region migration error",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "binBps"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 228
+          },
+          "id": 327,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_datanode_load",
+              "instant": false,
+              "legendFormat": "Datanode-{{datanode_id}}-writeload",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Datanode load",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Displays the rate of SQL executions processed by the Meta service using the RDS backend.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 228
+          },
+          "id": 339,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "rate(greptime_meta_rds_pg_sql_execute_elapsed_ms_count[$__rate_interval])",
+              "instant": false,
+              "legendFormat": "{{pod}} {{op}} {{type}} {{result}} ",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Rate of SQL Executions (RDS)",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Measures the response time of SQL executions via the RDS backend. ",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "ms"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 236
+          },
+          "id": 340,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.90, sum by(pod, op, type, result, le) (rate(greptime_meta_rds_pg_sql_execute_elapsed_ms_bucket[$__rate_interval])))",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "{{pod}} {{op}} {{type}} {{result}} p90",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            }
+          ],
+          "title": "SQL Execution Latency (RDS)",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Shows latency of Meta handlers by pod and handler name, useful for monitoring handler performance and detecting latency spikes.\n",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 236
+          },
+          "id": 341,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.90, sum by(pod, le, name) (\n  rate(greptime_meta_handler_execute_bucket[$__rate_interval])\n))",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "{{pod}} {{name}} p90",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            }
+          ],
+          "title": "Handler Execution Latency",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Shows p90 heartbeat message sizes, helping track network usage and identify anomalies in heartbeat payload.\n",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "bytes"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 244
+          },
+          "id": 342,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, sum by(pod, le) (greptime_meta_heartbeat_stat_memory_size_bucket))",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "{{pod}}",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            }
+          ],
+          "title": "Heartbeat Packet Size",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 244
+          },
+          "id": 345,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "rate(greptime_meta_heartbeat_rate[$__rate_interval])",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "{{pod}}",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            }
+          ],
+          "title": "Meta Heartbeat Receive Rate",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 252
+          },
+          "id": 343,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.99, sum by(pod, le, op, target) (greptime_meta_kv_request_elapsed_bucket))",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "{{pod}}-{{op}} p99",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            }
+          ],
+          "title": "Meta KV Ops Latency",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "none"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 252
+          },
+          "id": 346,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "rate(greptime_meta_kv_request_elapsed_count[$__rate_interval])",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "{{pod}}-{{op}} p99",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            }
+          ],
+          "title": "Rate of meta KV Ops",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Gauge of load information of each datanode, collected via heartbeat between datanode and metasrv. This information is for metasrv to schedule workloads.",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 260
+          },
+          "id": 347,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_tables_bucket))",
+              "fullMetaSearch": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "CreateLogicalTables-{{step}} p90",
+              "range": true,
+              "refId": "A",
+              "useBackend": false
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_table))",
+              "fullMetaSearch": false,
+              "hide": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "CreateTable-{{step}} p90",
+              "range": true,
+              "refId": "B",
+              "useBackend": false
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_view))",
+              "fullMetaSearch": false,
+              "hide": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "CreateView-{{step}} p90",
+              "range": true,
+              "refId": "C",
+              "useBackend": false
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_create_flow))",
+              "fullMetaSearch": false,
+              "hide": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "CreateFlow-{{step}} p90",
+              "range": true,
+              "refId": "D",
+              "useBackend": false
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_drop_table))",
+              "fullMetaSearch": false,
+              "hide": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "DropTable-{{step}} p90",
+              "range": true,
+              "refId": "E",
+              "useBackend": false
+            },
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "disableTextWrap": false,
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, sum by(le, pod, step) (greptime_meta_procedure_alter_table))",
+              "fullMetaSearch": false,
+              "hide": false,
+              "includeNullMetadata": true,
+              "instant": false,
+              "legendFormat": "AlterTable-{{step}} p90",
+              "range": true,
+              "refId": "F",
+              "useBackend": false
+            }
+          ],
+          "title": "DDL Latency",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Reconciliation stats",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 260
+          },
+          "id": 355,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "greptime_meta_reconciliation_stats",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "{{pod}}-{{table_type}}-{{type}}",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Reconciliation stats",
+          "type": "timeseries"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "description": "Elapsed of Reconciliation steps",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisBorderShow": false,
+                "axisCenteredZero": false,
+                "axisColorMode": "text",
+                "axisLabel": "",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "barWidthFactor": 0.6,
+                "drawStyle": "line",
+                "fillOpacity": 0,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": false
+                },
+                "insertNulls": false,
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green"
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              },
+              "unit": "s"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 268
+          },
+          "id": 356,
+          "options": {
+            "legend": {
+              "calcs": [],
+              "displayMode": "list",
+              "placement": "bottom",
+              "showLegend": true
+            },
+            "tooltip": {
+              "hideZeros": false,
+              "mode": "single",
+              "sort": "none"
+            }
+          },
+          "pluginVersion": "11.6.0",
+          "targets": [
+            {
+              "datasource": {
+                "type": "prometheus",
+                "uid": "${metrics}"
+              },
+              "editorMode": "code",
+              "expr": "histogram_quantile(0.9, greptime_meta_reconciliation_procedure_bucket)",
+              "hide": false,
+              "instant": false,
+              "legendFormat": "{{procedure_name}}-{{step}}-P90",
+              "range": true,
+              "refId": "A"
+            }
+          ],
+          "title": "Reconciliation steps",
+          "type": "timeseries"
+        }
+      ],
+      "title": "Metasrv",
+      "type": "row"
+    },
+    {
+      "collapsed": true,
+      "gridPos": {
+        "h": 1,
+        "w": 24,
+        "x": 0,
+        "y": 30
       },
       "id": 328,
       "panels": [
@@ -9031,7 +9034,7 @@
             "h": 8,
             "w": 12,
             "x": 0,
-            "y": 3055
+            "y": 3282
           },
           "id": 329,
           "options": {
@@ -9126,7 +9129,7 @@
             "h": 8,
             "w": 12,
             "x": 12,
-            "y": 3055
+            "y": 3282
           },
           "id": 330,
           "options": {
@@ -9234,7 +9237,7 @@
             "h": 8,
             "w": 9,
             "x": 0,
-            "y": 3063
+            "y": 3290
           },
           "id": 331,
           "options": {
@@ -9342,7 +9345,7 @@
             "h": 8,
             "w": 9,
             "x": 9,
-            "y": 3063
+            "y": 3290
           },
           "id": 332,
           "options": {
@@ -9437,7 +9440,7 @@
             "h": 8,
             "w": 6,
             "x": 18,
-            "y": 3063
+            "y": 3290
           },
           "id": 333,
           "options": {
@@ -9474,12 +9477,12 @@
       "type": "row"
     },
     {
-      "collapsed": true,
+      "collapsed": false,
       "gridPos": {
         "h": 1,
         "w": 24,
         "x": 0,
-        "y": 187
+        "y": 31
       },
       "id": 357,
       "panels": [],
@@ -9550,7 +9553,7 @@
         "h": 8,
         "w": 24,
         "x": 0,
-        "y": 188
+        "y": 32
       },
       "id": 358,
       "options": {
@@ -9648,7 +9651,7 @@
         "h": 8,
         "w": 12,
         "x": 0,
-        "y": 196
+        "y": 40
       },
       "id": 359,
       "options": {
@@ -9761,7 +9764,7 @@
         "h": 8,
         "w": 12,
         "x": 12,
-        "y": 196
+        "y": 40
       },
       "id": 360,
       "options": {
@@ -9861,7 +9864,7 @@
         "h": 8,
         "w": 12,
         "x": 0,
-        "y": 204
+        "y": 48
       },
       "id": 361,
       "options": {
@@ -9974,7 +9977,7 @@
         "h": 8,
         "w": 12,
         "x": 12,
-        "y": 204
+        "y": 48
       },
       "id": 364,
       "options": {
@@ -10074,7 +10077,7 @@
         "h": 8,
         "w": 12,
         "x": 0,
-        "y": 212
+        "y": 56
       },
       "id": 363,
       "options": {
@@ -10187,7 +10190,7 @@
         "h": 8,
         "w": 12,
         "x": 12,
-        "y": 212
+        "y": 56
       },
       "id": 362,
       "options": {
@@ -10221,6 +10224,477 @@
       ],
       "title": "Save Alert Failure Rate",
       "type": "timeseries"
+    },
+    {
+      "collapsed": false,
+      "gridPos": {
+        "h": 1,
+        "w": 24,
+        "x": 0,
+        "y": 64
+      },
+      "id": 366,
+      "panels": [],
+      "title": "Hotspot",
+      "type": "row"
+    },
+    {
+      "datasource": {
+        "type": "mysql",
+        "uid": "${information_schema}"
+      },
+      "fieldConfig": {
+        "defaults": {
+          "color": {
+            "mode": "thresholds"
+          },
+          "custom": {
+            "align": "auto",
+            "cellOptions": {
+              "type": "auto",
+              "wrapText": false
+            },
+            "inspect": false
+          },
+          "fieldMinMax": false,
+          "mappings": [],
+          "thresholds": {
+            "mode": "absolute",
+            "steps": [
+              {
+                "color": "green"
+              },
+              {
+                "color": "red",
+                "value": 80
+              }
+            ]
+          }
+        },
+        "overrides": [
+          {
+            "matcher": {
+              "id": "byName",
+              "options": "disk_size"
+            },
+            "properties": [
+              {
+                "id": "unit",
+                "value": "bytes"
+              }
+            ]
+          }
+        ]
+      },
+      "gridPos": {
+        "h": 18,
+        "w": 24,
+        "x": 0,
+        "y": 65
+      },
+      "id": 365,
+      "options": {
+        "cellHeight": "sm",
+        "footer": {
+          "countRows": false,
+          "enablePagination": true,
+          "fields": [
+            "region_rows",
+            "disk_size"
+          ],
+          "reducer": [
+            "sum"
+          ],
+          "show": true
+        },
+        "showHeader": true,
+        "sortBy": [
+          {
+            "desc": true,
+            "displayName": "table_schema"
+          }
+        ]
+      },
+      "pluginVersion": "11.6.0",
+      "targets": [
+        {
+          "dataset": "information_schema",
+          "datasource": {
+            "type": "mysql",
+            "uid": "${information_schema}"
+          },
+          "editorMode": "code",
+          "format": "table",
+          "rawQuery": true,
+          "rawSql": "WITH table_stats AS (\n  SELECT\n    table_id,\n    COUNT(*) AS region_count,\n    SUM(disk_size) AS total_disk_size,\n    SUM(region_rows) as total_region_rows\n  FROM information_schema.region_statistics\n  WHERE region_role = 'Leader'\n  GROUP BY table_id\n  HAVING COUNT(*) > 1\n)\n\nSELECT\n  t.table_schema,\n  t.table_name,\n\n  r.region_id,\n  t.table_id,\n  r.region_number,\n\n  p.partition_description,\n\n\n  ROUND(\n    r.disk_size * 100.0\n      / NULLIF(ts.total_disk_size, 0),\n    2\n  ) AS disk_size_share_percent,\n\n  r.disk_size,\n  \n  ROUND(\n    r.region_rows * 100.0\n      / NULLIF(ts.total_region_rows, 0),\n    2\n  ) AS region_rows_share_percent,\n  r.region_rows\n\nFROM information_schema.region_statistics r\n\nJOIN table_stats ts\n  ON r.table_id = ts.table_id\n\nJOIN information_schema.tables t\n  ON r.table_id = t.table_id\n\nLEFT JOIN information_schema.partitions p\n  ON p.table_schema = t.table_schema\n AND p.table_name = t.table_name\n AND p.greptime_partition_id = r.region_id\n\nWHERE r.region_role = 'Leader'\n\nORDER BY region_rows_share_percent DESC\nLIMIT 100;",
+          "refId": "A",
+          "sql": {
+            "columns": [
+              {
+                "parameters": [],
+                "type": "function"
+              }
+            ],
+            "groupBy": [
+              {
+                "property": {
+                  "type": "string"
+                },
+                "type": "groupBy"
+              }
+            ],
+            "limit": 50
+          }
+        }
+      ],
+      "title": "Hotspot Regions",
+      "type": "table"
+    },
+    {
+      "datasource": {
+        "type": "prometheus",
+        "uid": "${metrics}"
+      },
+      "description": "Write load of each datanode over time.",
+      "fieldConfig": {
+        "defaults": {
+          "color": {
+            "mode": "palette-classic"
+          },
+          "custom": {
+            "axisBorderShow": false,
+            "axisCenteredZero": false,
+            "axisColorMode": "text",
+            "axisLabel": "",
+            "axisPlacement": "auto",
+            "barAlignment": 0,
+            "barWidthFactor": 0.6,
+            "drawStyle": "line",
+            "fillOpacity": 0,
+            "gradientMode": "none",
+            "hideFrom": {
+              "legend": false,
+              "tooltip": false,
+              "viz": false
+            },
+            "insertNulls": false,
+            "lineInterpolation": "linear",
+            "lineWidth": 1,
+            "pointSize": 5,
+            "scaleDistribution": {
+              "type": "linear"
+            },
+            "showPoints": "auto",
+            "spanNulls": false,
+            "stacking": {
+              "group": "A",
+              "mode": "none"
+            },
+            "thresholdsStyle": {
+              "mode": "off"
+            }
+          },
+          "fieldMinMax": false,
+          "mappings": [],
+          "thresholds": {
+            "mode": "absolute",
+            "steps": [
+              {
+                "color": "green"
+              },
+              {
+                "color": "red",
+                "value": 80
+              }
+            ]
+          },
+          "unit": "binBps"
+        },
+        "overrides": [
+          {
+            "__systemRef": "hideSeriesFrom",
+            "matcher": {
+              "id": "byNames",
+              "options": {
+                "mode": "exclude",
+                "names": [
+                  "total_region_rows",
+                  "total_written_bytes_since_open",
+                  "total_disk_size",
+                  "total_sst_size",
+                  "total_sst_num",
+                  "total_memtable_size",
+                  "total_manifest_size",
+                  "total_index_size",
+                  "cluster_disk_ratio_percent",
+                  "cluster_sst_ratio_percent",
+                  "cluster_rows_ratio_percent",
+                  "max_region_disk_size",
+                  "max_region_rows",
+                  "leader_region_count",
+                  "disk_ratio_percent",
+                  "data_size"
+                ],
+                "prefix": "All except:",
+                "readOnly": true
+              }
+            },
+            "properties": []
+          }
+        ]
+      },
+      "gridPos": {
+        "h": 11,
+        "w": 12,
+        "x": 0,
+        "y": 83
+      },
+      "id": 368,
+      "options": {
+        "legend": {
+          "calcs": [],
+          "displayMode": "list",
+          "placement": "bottom",
+          "showLegend": true
+        },
+        "tooltip": {
+          "hideZeros": false,
+          "mode": "single",
+          "sort": "none"
+        }
+      },
+      "pluginVersion": "11.6.0",
+      "targets": [
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "disableTextWrap": false,
+          "editorMode": "builder",
+          "expr": "greptime_datanode_history_load",
+          "fullMetaSearch": false,
+          "includeNullMetadata": true,
+          "instant": false,
+          "legendFormat": "datanode-{{datanode_id}}({{instance}})",
+          "range": true,
+          "refId": "A",
+          "useBackend": false
+        }
+      ],
+      "title": "Datanode Load(Write)",
+      "type": "timeseries"
+    },
+    {
+      "datasource": {
+        "type": "prometheus",
+        "uid": "${metrics}"
+      },
+      "description": "Distribution of write load across datanodes.",
+      "fieldConfig": {
+        "defaults": {
+          "color": {
+            "mode": "palette-classic"
+          },
+          "custom": {
+            "hideFrom": {
+              "legend": false,
+              "tooltip": false,
+              "viz": false
+            }
+          },
+          "fieldMinMax": false,
+          "mappings": [],
+          "min": 0,
+          "unit": "binBps"
+        },
+        "overrides": []
+      },
+      "gridPos": {
+        "h": 11,
+        "w": 12,
+        "x": 12,
+        "y": 83
+      },
+      "id": 369,
+      "options": {
+        "displayLabels": [
+          "percent"
+        ],
+        "legend": {
+          "displayMode": "list",
+          "placement": "right",
+          "showLegend": true
+        },
+        "pieType": "donut",
+        "reduceOptions": {
+          "calcs": [
+            "lastNotNull"
+          ],
+          "fields": "",
+          "values": false
+        },
+        "tooltip": {
+          "hideZeros": false,
+          "mode": "single",
+          "sort": "none"
+        }
+      },
+      "pluginVersion": "11.6.0",
+      "targets": [
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${metrics}"
+          },
+          "disableTextWrap": false,
+          "editorMode": "builder",
+          "exemplar": false,
+          "expr": "greptime_datanode_history_load",
+          "format": "time_series",
+          "fullMetaSearch": false,
+          "includeNullMetadata": true,
+          "instant": true,
+          "legendFormat": "datanode-{{datanode_id}}({{instance}})",
+          "range": false,
+          "refId": "A",
+          "useBackend": false
+        }
+      ],
+      "title": "Datanode Load(Write) Distribution",
+      "type": "piechart"
+    },
+    {
+      "datasource": {
+        "type": "mysql",
+        "uid": "${information_schema}"
+      },
+      "description": "Distribution of leader regions and data size across datanodes.",
+      "fieldConfig": {
+        "defaults": {
+          "color": {
+            "mode": "palette-classic"
+          },
+          "custom": {
+            "hideFrom": {
+              "legend": false,
+              "tooltip": false,
+              "viz": false
+            }
+          },
+          "fieldMinMax": false,
+          "mappings": [],
+          "unit": "bytes"
+        },
+        "overrides": [
+          {
+            "__systemRef": "hideSeriesFrom",
+            "matcher": {
+              "id": "byNames",
+              "options": {
+                "mode": "exclude",
+                "names": [
+                  "total_region_rows",
+                  "total_written_bytes_since_open",
+                  "total_disk_size",
+                  "total_sst_size",
+                  "total_sst_num",
+                  "total_memtable_size",
+                  "total_manifest_size",
+                  "total_index_size",
+                  "cluster_disk_ratio_percent",
+                  "cluster_sst_ratio_percent",
+                  "cluster_rows_ratio_percent",
+                  "max_region_disk_size",
+                  "max_region_rows",
+                  "leader_region_count",
+                  "disk_ratio_percent",
+                  "data_size"
+                ],
+                "prefix": "All except:",
+                "readOnly": true
+              }
+            },
+            "properties": [
+              {
+                "id": "custom.hideFrom",
+                "value": {
+                  "legend": false,
+                  "tooltip": false,
+                  "viz": true
+                }
+              }
+            ]
+          }
+        ]
+      },
+      "gridPos": {
+        "h": 11,
+        "w": 12,
+        "x": 0,
+        "y": 94
+      },
+      "id": 367,
+      "options": {
+        "displayLabels": [
+          "value"
+        ],
+        "legend": {
+          "displayMode": "list",
+          "placement": "right",
+          "showLegend": true,
+          "values": [
+            "percent"
+          ]
+        },
+        "pieType": "pie",
+        "reduceOptions": {
+          "calcs": [
+            "lastNotNull"
+          ],
+          "fields": "/^data_size$/",
+          "values": true
+        },
+        "tooltip": {
+          "hideZeros": false,
+          "mode": "single",
+          "sort": "none"
+        }
+      },
+      "pluginVersion": "11.6.0",
+      "targets": [
+        {
+          "dataset": "information_schema",
+          "datasource": {
+            "type": "mysql",
+            "uid": "${information_schema}"
+          },
+          "editorMode": "code",
+          "format": "table",
+          "rawQuery": true,
+          "rawSql": "WITH leader_regions AS (\n  SELECT\n    CONCAT(\n      'datanode-',\n      p.peer_id,\n      ' (',\n      p.peer_addr,\n      ')'\n    ) AS datanode,\n    r.disk_size\n  FROM information_schema.region_statistics r\n  JOIN information_schema.region_peers p\n    ON r.region_id = p.region_id\n  WHERE r.region_role = 'Leader'\n    AND p.is_leader = 'Yes'\n)\n\nSELECT\n  datanode,\n  COUNT(*) AS leader_region_count,\n  SUM(disk_size) AS data_size\nFROM leader_regions\nGROUP BY datanode\nORDER BY data_size DESC;",
+          "refId": "A",
+          "sql": {
+            "columns": [
+              {
+                "parameters": [],
+                "type": "function"
+              }
+            ],
+            "groupBy": [
+              {
+                "property": {
+                  "type": "string"
+                },
+                "type": "groupBy"
+              }
+            ],
+            "limit": 50
+          }
+        }
+      ],
+      "title": "Datanode Data Distribution",
+      "type": "piechart"
     }
   ],
   "preload": false,
@@ -10354,12 +10828,12 @@
     ]
   },
   "time": {
-    "from": "now-1h",
+    "from": "now-2h",
     "to": "now"
   },
   "timepicker": {},
   "timezone": "",
-  "title": "GreptimeDB",
-  "uid": "dejf3k5e7g2kgb",
-  "version": 15
-}
+  "title": "GreptimeDB Copy",
+  "uid": "dflfbxbwvvchsa",
+  "version": 37
+}
\ No newline at end of file
diff --git a/grafana/dashboards/metrics/standalone/dashboard.md b/grafana/dashboards/metrics/standalone/dashboard.md
index 63b45208c5..d2a4ddba0a 100644
--- a/grafana/dashboards/metrics/standalone/dashboard.md
+++ b/grafana/dashboards/metrics/standalone/dashboard.md
@@ -142,3 +142,87 @@
   rate(greptime_trigger_save_alert_record_elapsed_bucket[$__rate_interval])
 )` | `timeseries` | Elapsed time to persist trigger alert records. | `prometheus` | `s` | `[{{instance}}]-[{{pod}}]-[{{storage_type}}]-p99` |
 | Save Alert Failure Rate | `rate(greptime_trigger_save_alert_record_failure_count[$__rate_interval])` | `timeseries` | Rate of failures when persisting trigger alert records. | `prometheus` | `none` | `__auto` |
+# Hotspot
+| Title | Query | Type | Description | Datasource | Unit | Legend Format |
+| --- | --- | --- | --- | --- | --- | --- |
+| Hotspot Regions | `WITH table_stats AS (
+  SELECT
+    table_id,
+    COUNT(*) AS region_count,
+    SUM(disk_size) AS total_disk_size,
+    SUM(region_rows) as total_region_rows
+  FROM information_schema.region_statistics
+  WHERE region_role = 'Leader'
+  GROUP BY table_id
+  HAVING COUNT(*) > 1
+)
+
+SELECT
+  t.table_schema,
+  t.table_name,
+
+  r.region_id,
+  t.table_id,
+  r.region_number,
+
+  p.partition_description,
+
+
+  ROUND(
+    r.disk_size * 100.0
+      / NULLIF(ts.total_disk_size, 0),
+    2
+  ) AS disk_size_share_percent,
+
+  r.disk_size,
+  
+  ROUND(
+    r.region_rows * 100.0
+      / NULLIF(ts.total_region_rows, 0),
+    2
+  ) AS region_rows_share_percent,
+  r.region_rows
+
+FROM information_schema.region_statistics r
+
+JOIN table_stats ts
+  ON r.table_id = ts.table_id
+
+JOIN information_schema.tables t
+  ON r.table_id = t.table_id
+
+LEFT JOIN information_schema.partitions p
+  ON p.table_schema = t.table_schema
+ AND p.table_name = t.table_name
+ AND p.greptime_partition_id = r.region_id
+
+WHERE r.region_role = 'Leader'
+
+ORDER BY region_rows_share_percent DESC
+LIMIT 100;` | `table` |  | `mysql` | -- | -- |
+| Datanode Load(Write) | `greptime_datanode_history_load` | `timeseries` | Write load of each datanode over time. | `prometheus` | `binBps` | `datanode-{{datanode_id}}({{instance}})` |
+| Datanode Load(Write) Distribution | `greptime_datanode_history_load` | `piechart` | Distribution of write load across datanodes. | `prometheus` | `binBps` | `datanode-{{datanode_id}}({{instance}})` |
+| Datanode Data Distribution | `WITH leader_regions AS (
+  SELECT
+    CONCAT(
+      'datanode-',
+      p.peer_id,
+      ' (',
+      p.peer_addr,
+      ')'
+    ) AS datanode,
+    r.disk_size
+  FROM information_schema.region_statistics r
+  JOIN information_schema.region_peers p
+    ON r.region_id = p.region_id
+  WHERE r.region_role = 'Leader'
+    AND p.is_leader = 'Yes'
+)
+
+SELECT
+  datanode,
+  COUNT(*) AS leader_region_count,
+  SUM(disk_size) AS data_size
+FROM leader_regions
+GROUP BY datanode
+ORDER BY data_size DESC;` | `piechart` | Distribution of leader regions and data size across datanodes. | `mysql` | `bytes` | -- |
diff --git a/grafana/dashboards/metrics/standalone/dashboard.yaml b/grafana/dashboards/metrics/standalone/dashboard.yaml
index 32f771f66f..db1d2b6a7a 100644
--- a/grafana/dashboards/metrics/standalone/dashboard.yaml
+++ b/grafana/dashboards/metrics/standalone/dashboard.yaml
@@ -1153,3 +1153,65 @@ groups:
                 type: prometheus
                 uid: ${metrics}
               legendFormat: __auto
+    - title: Hotspot
+      panels:
+        - title: Hotspot Regions
+          type: table
+          queries:
+            - expr: "WITH table_stats AS (\n  SELECT\n    table_id,\n    COUNT(*) AS region_count,\n    SUM(disk_size) AS total_disk_size,\n    SUM(region_rows) as total_region_rows\n  FROM information_schema.region_statistics\n  WHERE region_role = 'Leader'\n  GROUP BY table_id\n  HAVING COUNT(*) > 1\n)\n\nSELECT\n  t.table_schema,\n  t.table_name,\n\n  r.region_id,\n  t.table_id,\n  r.region_number,\n\n  p.partition_description,\n\n\n  ROUND(\n    r.disk_size * 100.0\n      / NULLIF(ts.total_disk_size, 0),\n    2\n  ) AS disk_size_share_percent,\n\n  r.disk_size,\n  \n  ROUND(\n    r.region_rows * 100.0\n      / NULLIF(ts.total_region_rows, 0),\n    2\n  ) AS region_rows_share_percent,\n  r.region_rows\n\nFROM information_schema.region_statistics r\n\nJOIN table_stats ts\n  ON r.table_id = ts.table_id\n\nJOIN information_schema.tables t\n  ON r.table_id = t.table_id\n\nLEFT JOIN information_schema.partitions p\n  ON p.table_schema = t.table_schema\n AND p.table_name = t.table_name\n AND p.greptime_partition_id = r.region_id\n\nWHERE r.region_role = 'Leader'\n\nORDER BY region_rows_share_percent DESC\nLIMIT 100;"
+              datasource:
+                type: mysql
+                uid: ${information_schema}
+        - title: Datanode Load(Write)
+          type: timeseries
+          description: Write load of each datanode over time.
+          unit: binBps
+          queries:
+            - expr: greptime_datanode_history_load
+              datasource:
+                type: prometheus
+                uid: ${metrics}
+              legendFormat: datanode-{{datanode_id}}({{instance}})
+        - title: Datanode Load(Write) Distribution
+          type: piechart
+          description: Distribution of write load across datanodes.
+          unit: binBps
+          queries:
+            - expr: greptime_datanode_history_load
+              datasource:
+                type: prometheus
+                uid: ${metrics}
+              legendFormat: datanode-{{datanode_id}}({{instance}})
+        - title: Datanode Data Distribution
+          type: piechart
+          description: Distribution of leader regions and data size across datanodes.
+          unit: bytes
+          queries:
+            - expr: |-
+                WITH leader_regions AS (
+                  SELECT
+                    CONCAT(
+                      'datanode-',
+                      p.peer_id,
+                      ' (',
+                      p.peer_addr,
+                      ')'
+                    ) AS datanode,
+                    r.disk_size
+                  FROM information_schema.region_statistics r
+                  JOIN information_schema.region_peers p
+                    ON r.region_id = p.region_id
+                  WHERE r.region_role = 'Leader'
+                    AND p.is_leader = 'Yes'
+                )
+
+                SELECT
+                  datanode,
+                  COUNT(*) AS leader_region_count,
+                  SUM(disk_size) AS data_size
+                FROM leader_regions
+                GROUP BY datanode
+                ORDER BY data_size DESC;
+              datasource:
+                type: mysql
+                uid: ${information_schema}
diff --git a/rust-toolchain.toml b/rust-toolchain.toml
index 58b88a3894..d16edecca8 100644
--- a/rust-toolchain.toml
+++ b/rust-toolchain.toml
@@ -1,2 +1,2 @@
 [toolchain]
-channel = "nightly-2025-10-01"
+channel = "nightly-2026-03-21"
diff --git a/src/api/src/helper.rs b/src/api/src/helper.rs
index 4664c0434b..3b305d90c5 100644
--- a/src/api/src/helper.rs
+++ b/src/api/src/helper.rs
@@ -444,7 +444,7 @@ impl TryFrom<ConcreteDataType> for ColumnDataTypeWrapper {
                         JsonFormat::Jsonb => Some(ColumnDataTypeExtension {
                             type_ext: Some(TypeExt::JsonType(JsonTypeExtension::JsonBinary.into())),
                         }),
-                        JsonFormat::Native(native_type) => {
+                        JsonFormat::Json2(native_type) => {
                             if native_type.is_null() {
                                 None
                             } else {
@@ -919,6 +919,7 @@ pub fn encode_json_value(value: JsonValue) -> v1::JsonValue {
                     .collect::<Vec<_>>();
                 Some(json_value::Value::Object(JsonObject { entries }))
             }
+            JsonVariant::Variant(x) => Some(json_value::Value::Variant(x)),
         };
         v1::JsonValue { value }
     }
@@ -952,6 +953,7 @@ fn decode_json_value(value: &v1::JsonValue) -> JsonValueRef<'_> {
             })
             .collect::<BTreeMap<_, _>>()
             .into(),
+        json_value::Value::Variant(x) => x.as_slice().into(),
     }
 }
 
diff --git a/src/auth/tests/mod.rs b/src/auth/tests/mod.rs
index 65db96a13f..4abbf89e5c 100644
--- a/src/auth/tests/mod.rs
+++ b/src/auth/tests/mod.rs
@@ -12,8 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 use std::sync::Arc;
 
 use api::v1::greptime_request::Request;
diff --git a/src/cache/Cargo.toml b/src/cache/Cargo.toml
index 3c128fe1ad..814eb9e484 100644
--- a/src/cache/Cargo.toml
+++ b/src/cache/Cargo.toml
@@ -9,6 +9,6 @@ catalog.workspace = true
 common-error.workspace = true
 common-macro.workspace = true
 common-meta.workspace = true
-moka.workspace = true
+moka = { workspace = true, features = ["future"] }
 partition.workspace = true
 snafu.workspace = true
diff --git a/src/catalog/src/kvbackend.rs b/src/catalog/src/kvbackend.rs
index 334acc999c..77eb3f2048 100644
--- a/src/catalog/src/kvbackend.rs
+++ b/src/catalog/src/kvbackend.rs
@@ -20,6 +20,6 @@ mod table_cache;
 pub use builder::{
     CatalogManagerConfigurator, CatalogManagerConfiguratorRef, KvBackendCatalogManagerBuilder,
 };
-pub use client::{CachedKvBackend, CachedKvBackendBuilder, MetaKvBackend};
+pub use client::{CachedKvBackend, CachedKvBackendBuilder, new_read_only_meta_kv_backend};
 pub use manager::KvBackendCatalogManager;
 pub use table_cache::{TableCache, TableCacheRef, new_table_cache};
diff --git a/src/catalog/src/kvbackend/client.rs b/src/catalog/src/kvbackend/client.rs
index f74509217f..69b31d6232 100644
--- a/src/catalog/src/kvbackend/client.rs
+++ b/src/catalog/src/kvbackend/client.rs
@@ -21,7 +21,10 @@ use std::time::Duration;
 use common_error::ext::BoxedError;
 use common_meta::cache_invalidator::KvCacheInvalidator;
 use common_meta::error::Error::CacheNotGet;
-use common_meta::error::{CacheNotGetSnafu, Error, ExternalSnafu, GetKvCacheSnafu, Result};
+use common_meta::error::{
+    CacheNotGetSnafu, Error, ExternalSnafu, GetKvCacheSnafu, Result, UnsupportedSnafu,
+};
+use common_meta::kv_backend::read_only::ReadOnlyKvBackend;
 use common_meta::kv_backend::txn::{Txn, TxnResponse};
 use common_meta::kv_backend::{KvBackend, KvBackendRef, TxnService};
 use common_meta::rpc::KeyValue;
@@ -357,19 +360,35 @@ impl CachedKvBackend {
 }
 
 #[derive(Debug)]
-pub struct MetaKvBackend {
-    pub client: Arc<MetaClient>,
+pub(crate) struct MetaKvBackend {
+    client: Arc<MetaClient>,
 }
 
 impl MetaKvBackend {
     /// Constructs a [MetaKvBackend].
-    pub fn new(client: Arc<MetaClient>) -> MetaKvBackend {
+    fn new(client: Arc<MetaClient>) -> MetaKvBackend {
         MetaKvBackend { client }
     }
 }
 
+pub fn new_read_only_meta_kv_backend(client: Arc<MetaClient>) -> KvBackendRef {
+    Arc::new(ReadOnlyKvBackend::new(Arc::new(MetaKvBackend::new(client))))
+}
+
+#[async_trait::async_trait]
 impl TxnService for MetaKvBackend {
     type Error = Error;
+
+    async fn txn(&self, _txn: Txn) -> Result<TxnResponse> {
+        UnsupportedSnafu {
+            operation: "MetaKvBackend txn",
+        }
+        .fail()
+    }
+
+    fn max_txn_ops(&self) -> usize {
+        usize::MAX
+    }
 }
 
 /// Implement `KvBackend` trait for `MetaKvBackend` instead of opendal's `Accessor` since
@@ -465,6 +484,9 @@ mod tests {
     use std::sync::atomic::{AtomicU32, Ordering};
 
     use async_trait::async_trait;
+    use common_meta::kv_backend::memory::MemoryKvBackend;
+    use common_meta::kv_backend::read_only::ReadOnlyKvBackend;
+    use common_meta::kv_backend::txn::{Txn, TxnOp};
     use common_meta::kv_backend::{KvBackend, TxnService};
     use common_meta::rpc::KeyValue;
     use common_meta::rpc::store::{
@@ -473,8 +495,9 @@ mod tests {
         PutResponse, RangeRequest, RangeResponse,
     };
     use dashmap::DashMap;
+    use meta_client::client::MetaClientBuilder;
 
-    use super::CachedKvBackend;
+    use super::{CachedKvBackend, new_read_only_meta_kv_backend};
 
     #[derive(Default)]
     pub struct SimpleKvBackend {
@@ -579,6 +602,62 @@ mod tests {
         }
     }
 
+    #[tokio::test]
+    async fn test_cached_kv_backend_rejects_writes_with_read_only_inner() {
+        let inner = Arc::new(MemoryKvBackend::<common_meta::error::Error>::new());
+        let cached_kv = CachedKvBackend::wrap(Arc::new(ReadOnlyKvBackend::new(inner)));
+
+        let err = cached_kv
+            .put(PutRequest {
+                key: b"k1".to_vec(),
+                value: b"v1".to_vec(),
+                prev_kv: false,
+            })
+            .await
+            .unwrap_err();
+
+        assert!(matches!(
+            err,
+            common_meta::error::Error::ReadOnlyKvBackend { .. }
+        ));
+    }
+
+    #[tokio::test]
+    async fn test_read_only_meta_kv_backend_rejects_writes() {
+        let meta_client = Arc::new(MetaClientBuilder::frontend_default_options().build());
+        let backend = new_read_only_meta_kv_backend(meta_client);
+
+        let err = backend
+            .put(PutRequest {
+                key: b"k1".to_vec(),
+                value: b"v1".to_vec(),
+                prev_kv: false,
+            })
+            .await
+            .unwrap_err();
+
+        assert!(matches!(
+            err,
+            common_meta::error::Error::ReadOnlyKvBackend { .. }
+        ));
+    }
+
+    #[tokio::test]
+    async fn test_read_only_meta_kv_backend_does_not_emulate_txn() {
+        let meta_client = Arc::new(MetaClientBuilder::frontend_default_options().build());
+        let backend = new_read_only_meta_kv_backend(meta_client);
+
+        let result = backend
+            .txn(Txn::new().and_then(vec![TxnOp::Get(b"k1".to_vec())]))
+            .await;
+        let err = match result {
+            Ok(_) => panic!("expected unsupported txn error"),
+            Err(err) => err,
+        };
+
+        assert!(matches!(err, common_meta::error::Error::Unsupported { .. }));
+    }
+
     async fn add_some_vals(kv_backend: &impl KvBackend) {
         kv_backend
             .put(PutRequest {
diff --git a/src/catalog/src/kvbackend/table_cache.rs b/src/catalog/src/kvbackend/table_cache.rs
index ea328c3e17..42b3fbc74b 100644
--- a/src/catalog/src/kvbackend/table_cache.rs
+++ b/src/catalog/src/kvbackend/table_cache.rs
@@ -65,11 +65,13 @@ fn init_factory(
 
 fn invalidator<'a>(
     cache: &'a Cache<TableName, TableRef>,
-    ident: &'a CacheIdent,
+    idents: &'a [&CacheIdent],
 ) -> BoxFuture<'a, MetaResult<()>> {
     Box::pin(async move {
-        if let CacheIdent::TableName(table_name) = ident {
-            cache.invalidate(table_name).await
+        for ident in idents {
+            if let CacheIdent::TableName(table_name) = ident {
+                cache.invalidate(table_name).await
+            }
         }
         Ok(())
     })
diff --git a/src/catalog/src/lib.rs b/src/catalog/src/lib.rs
index 9c31e809fd..a701473551 100644
--- a/src/catalog/src/lib.rs
+++ b/src/catalog/src/lib.rs
@@ -12,9 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
-#![feature(try_blocks)]
-
 use std::any::Any;
 use std::fmt::{Debug, Formatter};
 use std::sync::Arc;
diff --git a/src/catalog/src/memory/manager.rs b/src/catalog/src/memory/manager.rs
index 571cd06468..6e747f62ed 100644
--- a/src/catalog/src/memory/manager.rs
+++ b/src/catalog/src/memory/manager.rs
@@ -132,15 +132,13 @@ impl CatalogManager for MemoryCatalogManager {
         table_name: &str,
         _query_ctx: Option<&QueryContext>,
     ) -> Result<Option<TableRef>> {
-        let result = try {
-            self.catalogs
-                .read()
-                .unwrap()
-                .get(catalog)?
-                .get(schema)?
-                .get(table_name)
-                .cloned()?
-        };
+        let catalogs = self.catalogs.read().unwrap();
+        let result = catalogs
+            .get(catalog)
+            .and_then(|c| c.get(schema))
+            .and_then(|s| s.get(table_name))
+            .cloned();
+
         Ok(result)
     }
 
@@ -149,8 +147,8 @@ impl CatalogManager for MemoryCatalogManager {
             .catalogs
             .read()
             .unwrap()
-            .iter()
-            .flat_map(|(_, schema_entries)| schema_entries.values())
+            .values()
+            .flat_map(|schema_entries| schema_entries.values())
             .flat_map(|tables| tables.values())
             .find(|t| t.table_info().ident.table_id == table_id)
             .map(|t| t.table_info()))
diff --git a/src/catalog/src/process_manager.rs b/src/catalog/src/process_manager.rs
index 1eebfec627..49be14933c 100644
--- a/src/catalog/src/process_manager.rs
+++ b/src/catalog/src/process_manager.rs
@@ -58,6 +58,8 @@ pub enum QueryStatement {
     Sql(Statement),
     // The optional string is the alias of the PromQL query.
     Promql(EvalStmt, Option<String>),
+    /// Logical plan with original query string
+    Plan(String),
 }
 
 impl Display for QueryStatement {
@@ -71,6 +73,7 @@ impl Display for QueryStatement {
                     write!(f, "{}", eval_stmt)
                 }
             }
+            QueryStatement::Plan(query) => write!(f, "{}", query),
         }
     }
 }
@@ -170,7 +173,7 @@ impl ProcessManager {
         let mut processes = vec![];
         if let Some(remote_frontend_selector) = self.frontend_selector.as_ref() {
             let frontends = remote_frontend_selector
-                .select(|node| node.peer.addr != self.server_addr)
+                .select(|peer| peer.addr != self.server_addr)
                 .await
                 .context(error::InvokeFrontendSnafu)?;
             for mut f in frontends {
@@ -208,7 +211,7 @@ impl ProcessManager {
                 .frontend_selector
                 .as_ref()
                 .context(error::MetaClientMissingSnafu)?
-                .select(|node| node.peer.addr == server_addr)
+                .select(|peer| peer.addr == server_addr)
                 .await
                 .context(error::InvokeFrontendSnafu)?;
             ensure!(
@@ -369,6 +372,9 @@ impl SlowQueryTimer {
             QueryStatement::Sql(stmt) => {
                 slow_query_event.query = stmt.to_string();
             }
+            QueryStatement::Plan(query) => {
+                slow_query_event.query = query.clone();
+            }
         }
 
         match self.record_type {
@@ -395,7 +401,7 @@ impl SlowQueryTimer {
 
 impl Drop for SlowQueryTimer {
     fn drop(&mut self) {
-        // Calculate the elaspsed duration since the timer is created.
+        // Calculate the elapsed duration since the timer is created.
         let elapsed = self.start.elapsed();
         if elapsed > self.threshold {
             // Only capture a portion of slow queries based on sample_ratio.
diff --git a/src/catalog/src/system_schema.rs b/src/catalog/src/system_schema.rs
index 00b96e5292..6c6a5fc202 100644
--- a/src/catalog/src/system_schema.rs
+++ b/src/catalog/src/system_schema.rs
@@ -139,12 +139,16 @@ impl DataSource for SystemTableDataSource {
         &self,
         request: ScanRequest,
     ) -> std::result::Result<SendableRecordBatchStream, BoxedError> {
-        let projected_schema = match &request.projection {
+        let projection = request
+            .projection_input
+            .as_ref()
+            .map(|input| input.projection.clone());
+
+        let projected_schema = match projection.as_ref() {
             Some(projection) => self.try_project(projection)?,
             None => self.table.schema(),
         };
 
-        let projection = request.projection.clone();
         let stream = self
             .table
             .to_stream(request)
diff --git a/src/catalog/src/system_schema/information_schema/flows.rs b/src/catalog/src/system_schema/information_schema/flows.rs
index a751716620..8002b17888 100644
--- a/src/catalog/src/system_schema/information_schema/flows.rs
+++ b/src/catalog/src/system_schema/information_schema/flows.rs
@@ -16,6 +16,7 @@ use std::sync::{Arc, Weak};
 
 use common_catalog::consts::INFORMATION_SCHEMA_FLOW_TABLE_ID;
 use common_error::ext::BoxedError;
+use common_meta::ddl::create_flow::FlowType;
 use common_meta::key::FlowId;
 use common_meta::key::flow::FlowMetadataManager;
 use common_meta::key::flow::flow_info::FlowInfoValue;
@@ -71,6 +72,7 @@ pub const CREATED_TIME: &str = "created_time";
 pub const UPDATED_TIME: &str = "updated_time";
 pub const LAST_EXECUTION_TIME: &str = "last_execution_time";
 pub const SOURCE_TABLE_NAMES: &str = "source_table_names";
+pub const FLOWNODE_ADDRS: &str = "flownode_addrs";
 
 /// The `information_schema.flows` to provides information about flows in databases.
 #[derive(Debug)]
@@ -95,7 +97,8 @@ impl InformationSchemaFlows {
         }
     }
 
-    /// for complex fields(including [`SOURCE_TABLE_IDS`], [`FLOWNODE_IDS`] and [`OPTIONS`]), it will be serialized to json string for now
+    /// for complex fields(including [`SOURCE_TABLE_IDS`], [`FLOWNODE_IDS`], [`OPTIONS`] and
+    /// [`FLOWNODE_ADDRS`]), it will be serialized to json string for now
     /// TODO(discord9): use a better way to store complex fields like json type
     pub(crate) fn schema() -> SchemaRef {
         Arc::new(Schema::new(
@@ -119,6 +122,7 @@ impl InformationSchemaFlows {
                     true,
                 ),
                 (SOURCE_TABLE_NAMES, CDT::string_datatype(), true),
+                (FLOWNODE_ADDRS, CDT::string_datatype(), true),
             ]
             .into_iter()
             .map(|(name, ty, nullable)| ColumnSchema::new(name, ty, nullable))
@@ -165,6 +169,10 @@ impl InformationSchemaFlows {
             expire_after: flow_info.expire_after(),
             eval_interval: flow_info.eval_interval(),
             comment,
+            flow_options: sql::statements::OptionMap::from_filtered_string_map(
+                flow_info.options(),
+                &[FlowType::FLOW_TYPE_KEY],
+            ),
             query,
         };
 
@@ -239,6 +247,7 @@ struct InformationSchemaFlowsBuilder {
     updated_time: TimestampMillisecondVectorBuilder,
     last_execution_time: TimestampMillisecondVectorBuilder,
     source_table_names: StringVectorBuilder,
+    flownode_addr_groups: StringVectorBuilder,
 }
 
 impl InformationSchemaFlowsBuilder {
@@ -269,6 +278,7 @@ impl InformationSchemaFlowsBuilder {
             updated_time: TimestampMillisecondVectorBuilder::with_capacity(INIT_CAPACITY),
             last_execution_time: TimestampMillisecondVectorBuilder::with_capacity(INIT_CAPACITY),
             source_table_names: StringVectorBuilder::with_capacity(INIT_CAPACITY),
+            flownode_addr_groups: StringVectorBuilder::with_capacity(INIT_CAPACITY),
         }
     }
 
@@ -378,6 +388,21 @@ impl InformationSchemaFlowsBuilder {
                     .get(&flow_id)
                     .map(|v| TimestampMillisecond::new(*v))
             }));
+        let flownode_addrs = self
+            .flow_metadata_manager
+            .flownode_addrs(flow_id)
+            .await
+            .map_err(BoxedError::new)
+            .context(InternalSnafu)?;
+        if flownode_addrs.is_empty() {
+            self.flownode_addr_groups.push(None);
+        } else {
+            let flownode_addrs_json =
+                serde_json::to_string(&flownode_addrs).with_context(|_| JsonSnafu {
+                    input: format!("{:?}", flownode_addrs),
+                })?;
+            self.flownode_addr_groups.push(Some(&flownode_addrs_json));
+        }
 
         let mut source_table_names = vec![];
         let catalog_manager = self
@@ -413,6 +438,7 @@ impl InformationSchemaFlowsBuilder {
             Arc::new(self.updated_time.finish()),
             Arc::new(self.last_execution_time.finish()),
             Arc::new(self.source_table_names.finish()),
+            Arc::new(self.flownode_addr_groups.finish()),
         ];
         RecordBatch::new(self.schema.clone(), columns).context(CreateRecordBatchSnafu)
     }
diff --git a/src/catalog/src/system_schema/information_schema/region_peers.rs b/src/catalog/src/system_schema/information_schema/region_peers.rs
index 5bc91d207e..9eddb061f8 100644
--- a/src/catalog/src/system_schema/information_schema/region_peers.rs
+++ b/src/catalog/src/system_schema/information_schema/region_peers.rs
@@ -267,7 +267,7 @@ impl InformationSchemaRegionPeersBuilder {
             ];
 
             if !predicates.eval(&row) {
-                return;
+                continue;
             }
 
             self.table_catalogs.push(Some(table_catalog));
@@ -331,3 +331,87 @@ impl DfPartitionStream for InformationSchemaRegionPeers {
         ))
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use api::v1::meta::Peer;
+    use arrow::array::AsArray;
+    use common_meta::rpc::router::{Region, RegionRoute};
+    use datafusion::common::ScalarValue;
+    use datafusion::logical_expr::{BinaryExpr, Expr, Operator, col};
+    use store_api::storage::{RegionId, ScanRequest};
+
+    use super::*;
+
+    fn new_region_route(table_id: u32, region_number: u32, peer_id: u64) -> RegionRoute {
+        RegionRoute {
+            region: Region {
+                id: RegionId::new(table_id, region_number),
+                ..Default::default()
+            },
+            leader_peer: Some(Peer {
+                id: peer_id,
+                addr: format!("127.0.0.1:{}", 3000 + peer_id),
+            }),
+            follower_peers: vec![],
+            leader_state: None,
+            leader_down_since: None,
+            write_route_policy: None,
+        }
+    }
+
+    #[test]
+    fn test_add_region_peers_predicate_filters_correctly() {
+        let schema = InformationSchemaRegionPeers::schema();
+        let mut builder = InformationSchemaRegionPeersBuilder::new(
+            schema,
+            "greptime".to_string(),
+            Weak::<KvBackendCatalogManager>::new(),
+        );
+
+        let table_id = 1;
+        // 3 regions: region_number 0, 1, 2
+        let routes = vec![
+            new_region_route(table_id, 0, 1),
+            new_region_route(table_id, 1, 2),
+            new_region_route(table_id, 2, 3),
+        ];
+
+        // Build a predicate that matches only the last region (region_number=2).
+        // With the old `return` bug, encountering the first non-matching region
+        // (region_number=0) would exit add_region_peers entirely, so region_number=2
+        // would never be found.
+        let target_region_id = RegionId::new(table_id, 2).as_u64();
+        let filter = Expr::BinaryExpr(BinaryExpr::new(
+            Box::new(col(REGION_ID)),
+            Operator::Eq,
+            Box::new(Expr::Literal(
+                ScalarValue::UInt64(Some(target_region_id)),
+                None,
+            )),
+        ));
+        let request = ScanRequest {
+            filters: vec![filter],
+            ..Default::default()
+        };
+        let predicates = Predicates::from_scan_request(&Some(request));
+
+        builder.add_region_peers(
+            "greptime",
+            "public",
+            "test_table",
+            &predicates,
+            table_id,
+            &routes,
+        );
+
+        let batch = builder.finish().unwrap();
+        // Should have exactly 1 row for the matching region
+        assert_eq!(batch.num_rows(), 1);
+        // Verify it's the correct region
+        let region_id_col = batch
+            .column(3)
+            .as_primitive::<arrow::datatypes::UInt64Type>();
+        assert_eq!(region_id_col.value(0), target_region_id);
+    }
+}
diff --git a/src/catalog/src/system_schema/information_schema/ssts.rs b/src/catalog/src/system_schema/information_schema/ssts.rs
index 1c0d507a29..520fc6d485 100644
--- a/src/catalog/src/system_schema/information_schema/ssts.rs
+++ b/src/catalog/src/system_schema/information_schema/ssts.rs
@@ -63,7 +63,7 @@ impl InformationTable for InformationSchemaSstsManifest {
     }
 
     fn to_stream(&self, request: ScanRequest) -> Result<SendableRecordBatchStream> {
-        let schema = if let Some(p) = &request.projection {
+        let schema = if let Some(p) = request.projection_indices() {
             Arc::new(self.schema.try_project(p).context(ProjectSchemaSnafu)?)
         } else {
             self.schema.clone()
@@ -117,7 +117,7 @@ impl InformationTable for InformationSchemaSstsStorage {
     }
 
     fn to_stream(&self, request: ScanRequest) -> Result<SendableRecordBatchStream> {
-        let schema = if let Some(p) = &request.projection {
+        let schema = if let Some(p) = request.projection_indices() {
             Arc::new(self.schema.try_project(p).context(ProjectSchemaSnafu)?)
         } else {
             self.schema.clone()
@@ -172,7 +172,7 @@ impl InformationTable for InformationSchemaSstsIndexMeta {
     }
 
     fn to_stream(&self, request: ScanRequest) -> Result<SendableRecordBatchStream> {
-        let schema = if let Some(p) = &request.projection {
+        let schema = if let Some(p) = request.projection_indices() {
             Arc::new(self.schema.try_project(p).context(ProjectSchemaSnafu)?)
         } else {
             self.schema.clone()
diff --git a/src/catalog/src/system_schema/information_schema/tables.rs b/src/catalog/src/system_schema/information_schema/tables.rs
index 248fb243dd..6175c17d39 100644
--- a/src/catalog/src/system_schema/information_schema/tables.rs
+++ b/src/catalog/src/system_schema/information_schema/tables.rs
@@ -372,22 +372,16 @@ impl InformationSchemaTablesBuilder {
         self.table_types.push(Some(table_type_text));
         self.table_ids.push(Some(table_id));
 
-        let data_length = region_stats.iter().map(|stat| stat.sst_size).sum();
-        let table_rows = region_stats.iter().map(|stat| stat.num_rows).sum();
-        let index_length = region_stats.iter().map(|stat| stat.index_size).sum();
+        let data_length: u64 = region_stats.iter().map(|stat| stat.sst_size).sum();
+        let table_rows: u64 = region_stats.iter().map(|stat| stat.num_rows).sum();
+        let index_length: u64 = region_stats.iter().map(|stat| stat.index_size).sum();
 
-        // It's not precise, but it is acceptable for long-term data storage.
-        let avg_row_length = if table_rows > 0 {
-            let total_data_length = data_length
-                + region_stats
-                    .iter()
-                    .map(|stat| stat.memtable_size)
-                    .sum::<u64>();
-
-            total_data_length / table_rows
-        } else {
-            0
-        };
+        let total_data_length: u64 = data_length
+            + region_stats
+                .iter()
+                .map(|stat| stat.memtable_size)
+                .sum::<u64>();
+        let avg_row_length = total_data_length.checked_div(table_rows).unwrap_or(0);
 
         self.data_length.push(Some(data_length));
         self.index_length.push(Some(index_length));
diff --git a/src/catalog/src/system_schema/pg_catalog.rs b/src/catalog/src/system_schema/pg_catalog.rs
index 08aad2d6dd..feec46ff90 100644
--- a/src/catalog/src/system_schema/pg_catalog.rs
+++ b/src/catalog/src/system_schema/pg_catalog.rs
@@ -74,12 +74,10 @@ impl PGCatalogProvider {
         )
         .expect("Failed to initialize PgCatalogSchemaProvider");
 
-        let mut table_ids = HashMap::new();
-        let mut table_id = PG_CATALOG_TABLE_ID_START;
-        for name in PG_CATALOG_TABLES {
-            table_ids.insert(*name, table_id);
-            table_id += 1;
-        }
+        let table_ids: HashMap<_, _> = (PG_CATALOG_TABLE_ID_START..)
+            .zip(PG_CATALOG_TABLES.iter())
+            .map(|(id, name)| (*name, id))
+            .collect();
 
         let mut provider = Self {
             catalog_name,
diff --git a/src/catalog/src/table_source.rs b/src/catalog/src/table_source.rs
index 132e02fe14..fd78cc2573 100644
--- a/src/catalog/src/table_source.rs
+++ b/src/catalog/src/table_source.rs
@@ -15,7 +15,6 @@
 use std::collections::HashMap;
 use std::sync::Arc;
 
-use bytes::Bytes;
 use common_catalog::format_full_table_name;
 use common_query::logical_plan::{SubstraitPlanDecoderRef, rename_logical_plan_columns};
 use datafusion::common::{ResolvedTableReference, TableReference};
@@ -151,7 +150,7 @@ impl DfTableSourceProvider {
         let catalog_list = Arc::new(DummyCatalogList::new(self.catalog_manager.clone()));
         let logical_plan = self
             .plan_decoder
-            .decode(Bytes::from(view_info.view_info.clone()), catalog_list, true)
+            .decode(view_info.view_info.clone().into(), catalog_list, false)
             .await
             .context(DecodePlanSnafu {
                 name: &table.table_info().name,
@@ -191,7 +190,7 @@ impl DfTableSourceProvider {
                 plan_columns
                     .iter()
                     .map(|c| c.as_str())
-                    .zip(columns.into_iter())
+                    .zip(columns)
                     .collect(),
             )
             .context(ProjectViewColumnsSnafu)?
diff --git a/src/cli/Cargo.toml b/src/cli/Cargo.toml
index 46e79efd00..8c4150cd46 100644
--- a/src/cli/Cargo.toml
+++ b/src/cli/Cargo.toml
@@ -44,6 +44,7 @@ common-version.workspace = true
 common-wal.workspace = true
 datatypes.workspace = true
 etcd-client.workspace = true
+fs2.workspace = true
 futures.workspace = true
 humantime.workspace = true
 meta-client.workspace = true
@@ -65,6 +66,8 @@ store-api.workspace = true
 table.workspace = true
 tokio.workspace = true
 tracing-appender.workspace = true
+url.workspace = true
+uuid.workspace = true
 
 [dev-dependencies]
 common-meta = { workspace = true, features = ["testing"] }
@@ -72,4 +75,3 @@ common-test-util.workspace = true
 common-version.workspace = true
 serde.workspace = true
 tempfile.workspace = true
-url.workspace = true
diff --git a/src/cli/src/common/object_store.rs b/src/cli/src/common/object_store.rs
index 0b8372e509..129fb29cb7 100644
--- a/src/cli/src/common/object_store.rs
+++ b/src/cli/src/common/object_store.rs
@@ -220,18 +220,8 @@ impl PrefixedAzblobConnection {
             name: "AzBlob",
             required: [
                 (&self.azblob_container, "container"),
-                (&self.azblob_root, "root"),
-                (&self.azblob_account_name, "account name"),
                 (&self.azblob_endpoint, "endpoint"),
-            ],
-            custom_validator: |missing: &mut Vec<&str>| {
-                // account_key is only required if sas_token is not provided
-                if self.azblob_sas_token.is_none()
-                    && self.azblob_account_key.is_empty()
-                {
-                    missing.push("account key (when sas_token is not provided)");
-                }
-            }
+            ]
         )
     }
 }
diff --git a/src/cli/src/common/store.rs b/src/cli/src/common/store.rs
index 373c96a37a..34baaba4ff 100644
--- a/src/cli/src/common/store.rs
+++ b/src/cli/src/common/store.rs
@@ -153,17 +153,11 @@ impl StoreConfig {
                 BackendImpl::PostgresStore => {
                     let table_name = &self.meta_table_name;
                     let tls_config = self.tls_config();
-                    let pool = meta_srv::utils::postgres::create_postgres_pool(
+                    Ok(meta_srv::utils::postgres::build_postgres_kv_backend(
                         store_addrs,
                         None,
                         tls_config,
-                    )
-                    .await
-                    .map_err(BoxedError::new)?;
-                    let schema_name = self.meta_schema_name.as_deref();
-                    Ok(common_meta::kv_backend::rds::PgStore::with_pg_pool(
-                        pool,
-                        schema_name,
+                        self.meta_schema_name.as_deref(),
                         table_name,
                         max_txn_ops,
                         self.auto_create_schema,
@@ -175,12 +169,9 @@ impl StoreConfig {
                 BackendImpl::MysqlStore => {
                     let table_name = &self.meta_table_name;
                     let tls_config = self.tls_config();
-                    let pool =
-                        meta_srv::utils::mysql::create_mysql_pool(store_addrs, tls_config.as_ref())
-                            .await
-                            .map_err(BoxedError::new)?;
-                    Ok(common_meta::kv_backend::rds::MySqlStore::with_mysql_pool(
-                        pool,
+                    Ok(meta_srv::utils::mysql::build_mysql_kv_backend(
+                        store_addrs,
+                        tls_config.as_ref(),
                         table_name,
                         max_txn_ops,
                     )
diff --git a/src/cli/src/data.rs b/src/cli/src/data.rs
index 5966040a3b..114886542e 100644
--- a/src/cli/src/data.rs
+++ b/src/cli/src/data.rs
@@ -13,7 +13,12 @@
 // limitations under the License.
 
 mod export;
+pub mod export_v2;
 mod import;
+pub mod import_v2;
+pub(crate) mod path;
+pub mod snapshot_storage;
+pub(crate) mod sql;
 mod storage_export;
 
 use clap::Subcommand;
@@ -22,15 +27,24 @@ use common_error::ext::BoxedError;
 
 use crate::Tool;
 use crate::data::export::ExportCommand;
+use crate::data::export_v2::ExportV2Command;
 use crate::data::import::ImportCommand;
+use crate::data::import_v2::ImportV2Command;
 
 pub(crate) const COPY_PATH_PLACEHOLDER: &str = "<PATH/TO/FILES>";
 
 /// Command for data operations including exporting data from and importing data into GreptimeDB.
 #[derive(Subcommand)]
 pub enum DataCommand {
+    /// Export data (V1 - legacy).
     Export(ExportCommand),
+    /// Import data (V1 - legacy).
     Import(ImportCommand),
+    /// Export V2 - JSON-based schema export with manifest support.
+    #[clap(subcommand)]
+    ExportV2(ExportV2Command),
+    /// Import V2 - Import from V2 snapshot.
+    ImportV2(ImportV2Command),
 }
 
 impl DataCommand {
@@ -38,6 +52,8 @@ impl DataCommand {
         match self {
             DataCommand::Export(cmd) => cmd.build().await,
             DataCommand::Import(cmd) => cmd.build().await,
+            DataCommand::ExportV2(cmd) => cmd.build().await,
+            DataCommand::ImportV2(cmd) => cmd.build().await,
         }
     }
 }
diff --git a/src/cli/src/data/export.rs b/src/cli/src/data/export.rs
index 1cdb159336..148c27316e 100644
--- a/src/cli/src/data/export.rs
+++ b/src/cli/src/data/export.rs
@@ -107,13 +107,16 @@ pub struct ExportCommand {
     #[clap(long, value_parser = humantime::parse_duration)]
     timeout: Option<Duration>,
 
-    /// The proxy server address to connect, if set, will override the system proxy.
+    /// The proxy server address to connect.
     ///
-    /// The default behavior will use the system proxy if neither `proxy` nor `no_proxy` is set.
+    /// If set, it overrides the system proxy unless `--no-proxy` is specified.
+    /// If neither `--proxy` nor `--no-proxy` is set, system proxy (env) may be used.
     #[clap(long)]
     proxy: Option<String>,
 
-    /// Disable proxy server, if set, will not use any proxy.
+    /// Disable all proxy usage (ignores `--proxy` and system proxy).
+    ///
+    /// When set and `--proxy` is not provided, this explicitly disables system proxy.
     #[clap(long)]
     no_proxy: bool,
 
@@ -173,6 +176,7 @@ impl ExportCommand {
             // Treats `None` as `0s` to disable server-side default timeout.
             self.timeout.unwrap_or_default(),
             proxy,
+            self.no_proxy,
         );
 
         Ok(Box::new(Export {
@@ -454,8 +458,10 @@ impl Export {
 
     /// build operator with preference for file system
     async fn build_prefer_fs_operator(&self) -> Result<ObjectStore> {
-        if self.storage_type.is_remote_storage() && self.ddl_local_dir.is_some() {
-            let root = self.ddl_local_dir.as_ref().unwrap().clone();
+        if self.storage_type.is_remote_storage()
+            && let Some(ddl_local_dir) = &self.ddl_local_dir
+        {
+            let root = ddl_local_dir.clone();
             let op = new_fs_object_store(&root).map_err(|e| Error::Other {
                 source: e,
                 location: snafu::location!(),
@@ -1078,7 +1084,7 @@ mod tests {
 
     #[tokio::test]
     async fn test_export_command_build_with_azblob_empty_account_name() {
-        // Test Azure Blob with empty account_name
+        // account_name is optional for Azure Blob validation
         let cmd = ExportCommand::parse_from([
             "export",
             "--addr",
@@ -1086,30 +1092,19 @@ mod tests {
             "--azblob",
             "--azblob-container",
             "test-container",
-            "--azblob-root",
-            "test-root",
             "--azblob-account-name",
             "", // Empty account name
-            "--azblob-account-key",
-            MOCK_AZBLOB_ACCOUNT_KEY_B64,
             "--azblob-endpoint",
             "https://account.blob.core.windows.net",
         ]);
 
         let result = cmd.build().await;
-        assert!(result.is_err());
-        if let Err(err) = result {
-            assert!(
-                err.to_string().contains("AzBlob account name must be set"),
-                "Actual error: {}",
-                err
-            );
-        }
+        assert!(result.is_ok(), "Empty account_name should succeed");
     }
 
     #[tokio::test]
     async fn test_export_command_build_with_azblob_missing_account_key() {
-        // Missing account key
+        // account_key is optional for Azure Blob validation
         let cmd = ExportCommand::parse_from([
             "export",
             "--addr",
@@ -1117,24 +1112,12 @@ mod tests {
             "--azblob",
             "--azblob-container",
             "test-container",
-            "--azblob-root",
-            "test-root",
-            "--azblob-account-name",
-            "test-account",
             "--azblob-endpoint",
             "https://account.blob.core.windows.net",
         ]);
 
         let result = cmd.build().await;
-        assert!(result.is_err());
-        if let Err(err) = result {
-            assert!(
-                err.to_string()
-                    .contains("AzBlob account key (when sas_token is not provided) must be set"),
-                "Actual error: {}",
-                err
-            );
-        }
+        assert!(result.is_ok(), "Missing account_key should succeed");
     }
 
     // ==================== Gap 3: Boundary cases ====================
@@ -1232,21 +1215,58 @@ mod tests {
             "--azblob",
             "--azblob-container",
             "test-container",
-            "--azblob-root",
-            "test-root",
-            "--azblob-account-name",
-            "test-account",
-            "--azblob-account-key",
-            MOCK_AZBLOB_ACCOUNT_KEY_B64,
             "--azblob-endpoint",
             "https://account.blob.core.windows.net",
-            // No sas_token
         ]);
 
         let result = cmd.build().await;
         assert!(result.is_ok(), "Minimal AzBlob config should succeed");
     }
 
+    #[tokio::test]
+    async fn test_export_command_build_with_azblob_missing_endpoint() {
+        let cmd = ExportCommand::parse_from([
+            "export",
+            "--addr",
+            "127.0.0.1:4000",
+            "--azblob",
+            "--azblob-container",
+            "test-container",
+        ]);
+
+        let result = cmd.build().await;
+        assert!(result.is_err());
+        if let Err(err) = result {
+            assert!(
+                err.to_string().contains("AzBlob endpoint must be set"),
+                "Actual error: {}",
+                err
+            );
+        }
+    }
+
+    #[tokio::test]
+    async fn test_export_command_build_with_azblob_missing_container() {
+        let cmd = ExportCommand::parse_from([
+            "export",
+            "--addr",
+            "127.0.0.1:4000",
+            "--azblob",
+            "--azblob-endpoint",
+            "https://account.blob.core.windows.net",
+        ]);
+
+        let result = cmd.build().await;
+        assert!(result.is_err());
+        if let Err(err) = result {
+            assert!(
+                err.to_string().contains("AzBlob container must be set"),
+                "Actual error: {}",
+                err
+            );
+        }
+    }
+
     #[tokio::test]
     async fn test_export_command_build_with_local_and_s3() {
         // Both output-dir and S3 - S3 should take precedence
@@ -1281,7 +1301,7 @@ mod tests {
 
     #[tokio::test]
     async fn test_export_command_build_with_azblob_only_sas_token() {
-        // Azure Blob with sas_token but no account_key - should succeed
+        // Azure Blob with sas_token but no credentials - should still succeed
         let cmd = ExportCommand::parse_from([
             "export",
             "--addr",
@@ -1289,15 +1309,10 @@ mod tests {
             "--azblob",
             "--azblob-container",
             "test-container",
-            "--azblob-root",
-            "test-root",
-            "--azblob-account-name",
-            "test-account",
             "--azblob-endpoint",
             "https://account.blob.core.windows.net",
             "--azblob-sas-token",
             "test-sas-token",
-            // No account_key
         ]);
 
         let result = cmd.build().await;
@@ -1318,10 +1333,6 @@ mod tests {
             "--azblob",
             "--azblob-container",
             "test-container",
-            "--azblob-root",
-            "test-root",
-            "--azblob-account-name",
-            "test-account",
             "--azblob-account-key",
             "", // Empty account_key is OK if sas_token is provided
             "--azblob-endpoint",
diff --git a/src/cli/src/data/export_v2.rs b/src/cli/src/data/export_v2.rs
new file mode 100644
index 0000000000..9bdf3c63f5
--- /dev/null
+++ b/src/cli/src/data/export_v2.rs
@@ -0,0 +1,52 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Export V2 module.
+//!
+//! This module provides the V2 implementation of database export functionality,
+//! featuring:
+//! - JSON-based schema export (version-agnostic)
+//! - Manifest-based snapshot management
+//! - Support for multiple storage backends (S3, OSS, GCS, Azure Blob, local FS)
+//! - Resume capability for interrupted exports
+//!
+//! # Example
+//!
+//! ```bash
+//! # Export schema only
+//! greptime cli data export-v2 create \
+//!   --addr 127.0.0.1:4000 \
+//!   --to file:///tmp/snapshot \
+//!   --schema-only
+//!
+//! # Export with time range
+//! greptime cli data export-v2 create \
+//!   --addr 127.0.0.1:4000 \
+//!   --to s3://bucket/snapshots/prod-20250101 \
+//!   --start-time 2025-01-01T00:00:00Z \
+//!   --end-time 2025-01-31T23:59:59Z
+//! ```
+
+mod chunker;
+mod command;
+mod coordinator;
+pub(crate) mod data;
+pub mod error;
+pub mod extractor;
+pub mod manifest;
+pub mod schema;
+pub use command::ExportV2Command;
+
+#[cfg(test)]
+mod tests;
diff --git a/src/cli/src/data/export_v2/chunker.rs b/src/cli/src/data/export_v2/chunker.rs
new file mode 100644
index 0000000000..260d95fae9
--- /dev/null
+++ b/src/cli/src/data/export_v2/chunker.rs
@@ -0,0 +1,103 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::time::Duration;
+
+use chrono::Duration as ChronoDuration;
+
+use crate::data::export_v2::manifest::{ChunkMeta, TimeRange};
+
+pub fn generate_chunks(time_range: &TimeRange, window: Duration) -> Vec<ChunkMeta> {
+    let (Some(start), Some(end)) = (time_range.start, time_range.end) else {
+        return vec![ChunkMeta::new(1, time_range.clone())];
+    };
+
+    if start == end {
+        return vec![ChunkMeta::skipped(1, time_range.clone())];
+    }
+
+    if start > end {
+        return Vec::new();
+    }
+
+    let window = match ChronoDuration::from_std(window) {
+        Ok(window) if window > ChronoDuration::zero() => window,
+        _ => return vec![ChunkMeta::new(1, time_range.clone())],
+    };
+
+    let mut chunks = Vec::new();
+    let mut cursor = start;
+    let mut id = 1;
+
+    while cursor < end {
+        let next = cursor
+            .checked_add_signed(window)
+            .map_or(end, |timestamp| timestamp.min(end));
+        chunks.push(ChunkMeta::new(id, TimeRange::new(Some(cursor), Some(next))));
+        id += 1;
+        cursor = next;
+    }
+
+    chunks
+}
+
+#[cfg(test)]
+mod tests {
+    use chrono::{TimeZone, Utc};
+
+    use super::*;
+    use crate::data::export_v2::manifest::ChunkStatus;
+
+    #[test]
+    fn test_generate_chunks_unbounded() {
+        let range = TimeRange::unbounded();
+        let chunks = generate_chunks(&range, Duration::from_secs(3600));
+        assert_eq!(chunks.len(), 1);
+        assert_eq!(chunks[0].time_range, range);
+    }
+
+    #[test]
+    fn test_generate_chunks_split() {
+        let start = Utc.with_ymd_and_hms(2025, 1, 1, 0, 0, 0).unwrap();
+        let end = Utc.with_ymd_and_hms(2025, 1, 1, 3, 0, 0).unwrap();
+        let range = TimeRange::new(Some(start), Some(end));
+
+        let chunks = generate_chunks(&range, Duration::from_secs(3600));
+        assert_eq!(chunks.len(), 3);
+        assert_eq!(chunks[0].time_range.start, Some(start));
+        assert_eq!(
+            chunks[2].time_range.end,
+            Some(Utc.with_ymd_and_hms(2025, 1, 1, 3, 0, 0).unwrap())
+        );
+    }
+
+    #[test]
+    fn test_generate_chunks_empty_range() {
+        let start = Utc.with_ymd_and_hms(2025, 1, 1, 0, 0, 0).unwrap();
+        let range = TimeRange::new(Some(start), Some(start));
+        let chunks = generate_chunks(&range, Duration::from_secs(3600));
+        assert_eq!(chunks.len(), 1);
+        assert_eq!(chunks[0].status, ChunkStatus::Skipped);
+        assert_eq!(chunks[0].time_range, range);
+    }
+
+    #[test]
+    fn test_generate_chunks_invalid_range_is_empty() {
+        let start = Utc.with_ymd_and_hms(2025, 1, 1, 1, 0, 0).unwrap();
+        let end = Utc.with_ymd_and_hms(2025, 1, 1, 0, 0, 0).unwrap();
+        let range = TimeRange::new(Some(start), Some(end));
+        let chunks = generate_chunks(&range, Duration::from_secs(3600));
+        assert!(chunks.is_empty());
+    }
+}
diff --git a/src/cli/src/data/export_v2/command.rs b/src/cli/src/data/export_v2/command.rs
new file mode 100644
index 0000000000..a7da6d3862
--- /dev/null
+++ b/src/cli/src/data/export_v2/command.rs
@@ -0,0 +1,1194 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Export V2 CLI commands.
+
+use std::collections::HashSet;
+use std::time::Duration;
+
+use async_trait::async_trait;
+use clap::{Parser, Subcommand};
+use common_error::ext::BoxedError;
+use common_telemetry::info;
+use serde_json::Value;
+use snafu::{OptionExt, ResultExt};
+
+use crate::Tool;
+use crate::common::ObjectStoreConfig;
+use crate::data::export_v2::coordinator::export_data;
+use crate::data::export_v2::error::{
+    ChunkTimeWindowRequiresBoundsSnafu, DatabaseSnafu, EmptyResultSnafu,
+    ManifestVersionMismatchSnafu, Result, ResumeConfigMismatchSnafu, SchemaOnlyArgsNotAllowedSnafu,
+    SchemaOnlyModeMismatchSnafu, UnexpectedValueTypeSnafu,
+};
+use crate::data::export_v2::extractor::SchemaExtractor;
+use crate::data::export_v2::manifest::{
+    ChunkMeta, DataFormat, MANIFEST_FILE, MANIFEST_VERSION, Manifest, TimeRange,
+};
+use crate::data::path::ddl_path_for_schema;
+use crate::data::snapshot_storage::{OpenDalStorage, SnapshotStorage, validate_uri};
+use crate::data::sql::{escape_sql_identifier, escape_sql_literal};
+use crate::database::{DatabaseClient, parse_proxy_opts};
+
+/// Export V2 commands.
+#[derive(Debug, Subcommand)]
+pub enum ExportV2Command {
+    /// Create a new snapshot.
+    Create(ExportCreateCommand),
+    /// List snapshots under a parent location.
+    List(ExportListCommand),
+}
+
+impl ExportV2Command {
+    pub async fn build(&self) -> std::result::Result<Box<dyn Tool>, BoxedError> {
+        match self {
+            ExportV2Command::Create(cmd) => cmd.build().await,
+            ExportV2Command::List(cmd) => cmd.build().await,
+        }
+    }
+}
+
+/// List snapshots under a parent location.
+#[derive(Debug, Parser)]
+pub struct ExportListCommand {
+    /// Parent storage location whose direct subdirectories are snapshots.
+    #[clap(long)]
+    location: String,
+
+    /// Object store configuration for remote storage backends.
+    #[clap(flatten)]
+    storage: ObjectStoreConfig,
+}
+
+impl ExportListCommand {
+    pub async fn build(&self) -> std::result::Result<Box<dyn Tool>, BoxedError> {
+        validate_uri(&self.location).map_err(BoxedError::new)?;
+        let storage = OpenDalStorage::from_parent_uri(&self.location, &self.storage)
+            .map_err(BoxedError::new)?;
+
+        Ok(Box::new(ExportList {
+            location: self.location.clone(),
+            storage,
+        }))
+    }
+}
+
+/// Export list tool implementation.
+pub struct ExportList {
+    location: String,
+    storage: OpenDalStorage,
+}
+
+#[async_trait]
+impl Tool for ExportList {
+    async fn do_work(&self) -> std::result::Result<(), BoxedError> {
+        self.run().await.map_err(BoxedError::new)
+    }
+}
+
+impl ExportList {
+    async fn run(&self) -> Result<()> {
+        let result = scan_snapshots(&self.storage).await?;
+
+        println!("Scanning: {}", self.location);
+        if result.snapshots.is_empty() {
+            println!("No snapshots found.");
+        } else {
+            print_snapshot_list(&result.snapshots, result.unreadable.len());
+        }
+        print_unreadable_warnings(&result.unreadable);
+
+        Ok(())
+    }
+}
+
+/// Create a new snapshot.
+#[derive(Debug, Parser)]
+pub struct ExportCreateCommand {
+    /// Server address to connect (e.g., 127.0.0.1:4000).
+    #[clap(long)]
+    addr: String,
+
+    /// Target storage location (e.g., s3://bucket/path, file:///tmp/backup).
+    #[clap(long)]
+    to: String,
+
+    /// Catalog name.
+    #[clap(long, default_value = "greptime")]
+    catalog: String,
+
+    /// Schema list to export (default: all non-system schemas).
+    /// Can be specified multiple times or comma-separated.
+    #[clap(long, value_delimiter = ',')]
+    schemas: Vec<String>,
+
+    /// Export schema only, no data.
+    #[clap(long)]
+    schema_only: bool,
+
+    /// Time range start (ISO 8601 format, e.g., 2024-01-01T00:00:00Z).
+    #[clap(long)]
+    start_time: Option<String>,
+
+    /// Time range end (ISO 8601 format, e.g., 2024-12-31T23:59:59Z).
+    #[clap(long)]
+    end_time: Option<String>,
+
+    /// Chunk time window (e.g., 1h, 6h, 1d, 7d).
+    /// Requires both --start-time and --end-time when specified.
+    #[clap(long, value_parser = humantime::parse_duration)]
+    chunk_time_window: Option<Duration>,
+
+    /// Data format: parquet, csv, json.
+    #[clap(long, value_enum, default_value = "parquet")]
+    format: DataFormat,
+
+    /// Delete existing snapshot and recreate.
+    #[clap(long)]
+    force: bool,
+
+    /// Parallelism for COPY DATABASE execution (server-side, per schema per chunk).
+    #[clap(long, default_value = "1")]
+    parallelism: usize,
+
+    /// Basic authentication (user:password).
+    #[clap(long)]
+    auth_basic: Option<String>,
+
+    /// Request timeout.
+    #[clap(long, value_parser = humantime::parse_duration)]
+    timeout: Option<Duration>,
+
+    /// Proxy server address.
+    ///
+    /// If set, it overrides the system proxy unless `--no-proxy` is specified.
+    /// If neither `--proxy` nor `--no-proxy` is set, system proxy (env) may be used.
+    #[clap(long)]
+    proxy: Option<String>,
+
+    /// Disable all proxy usage (ignores `--proxy` and system proxy).
+    ///
+    /// When set and `--proxy` is not provided, this explicitly disables system proxy.
+    #[clap(long)]
+    no_proxy: bool,
+
+    /// Object store configuration for remote storage backends.
+    #[clap(flatten)]
+    storage: ObjectStoreConfig,
+}
+
+impl ExportCreateCommand {
+    pub async fn build(&self) -> std::result::Result<Box<dyn Tool>, BoxedError> {
+        // Validate URI format
+        validate_uri(&self.to).map_err(BoxedError::new)?;
+
+        let time_range = TimeRange::parse(self.start_time.as_deref(), self.end_time.as_deref())
+            .map_err(BoxedError::new)?;
+        if self.chunk_time_window.is_some() && !time_range.is_bounded() {
+            return ChunkTimeWindowRequiresBoundsSnafu
+                .fail()
+                .map_err(BoxedError::new);
+        }
+        if self.schema_only {
+            let mut invalid_args = Vec::new();
+            if self.start_time.is_some() {
+                invalid_args.push("--start-time");
+            }
+            if self.end_time.is_some() {
+                invalid_args.push("--end-time");
+            }
+            if self.chunk_time_window.is_some() {
+                invalid_args.push("--chunk-time-window");
+            }
+            if self.format != DataFormat::Parquet {
+                invalid_args.push("--format");
+            }
+            if self.parallelism != 1 {
+                invalid_args.push("--parallelism");
+            }
+            if !invalid_args.is_empty() {
+                return SchemaOnlyArgsNotAllowedSnafu {
+                    args: invalid_args.join(", "),
+                }
+                .fail()
+                .map_err(BoxedError::new);
+            }
+        }
+
+        // Parse schemas (empty vec means all schemas)
+        let schemas = if self.schemas.is_empty() {
+            None
+        } else {
+            Some(self.schemas.clone())
+        };
+
+        // Build storage
+        let storage = OpenDalStorage::from_uri(&self.to, &self.storage).map_err(BoxedError::new)?;
+
+        // Build database client
+        let proxy = parse_proxy_opts(self.proxy.clone(), self.no_proxy)?;
+        let database_client = DatabaseClient::new(
+            self.addr.clone(),
+            self.catalog.clone(),
+            self.auth_basic.clone(),
+            self.timeout.unwrap_or(Duration::from_secs(60)),
+            proxy,
+            self.no_proxy,
+        );
+
+        Ok(Box::new(ExportCreate {
+            config: ExportConfig {
+                catalog: self.catalog.clone(),
+                schemas,
+                schema_only: self.schema_only,
+                format: self.format,
+                force: self.force,
+                time_range,
+                chunk_time_window: self.chunk_time_window,
+                parallelism: self.parallelism,
+                snapshot_uri: self.to.clone(),
+                storage_config: self.storage.clone(),
+            },
+            storage: Box::new(storage),
+            database_client,
+        }))
+    }
+}
+
+/// Export tool implementation.
+pub struct ExportCreate {
+    config: ExportConfig,
+    storage: Box<dyn SnapshotStorage>,
+    database_client: DatabaseClient,
+}
+
+struct ExportConfig {
+    catalog: String,
+    schemas: Option<Vec<String>>,
+    schema_only: bool,
+    format: DataFormat,
+    force: bool,
+    time_range: TimeRange,
+    chunk_time_window: Option<Duration>,
+    parallelism: usize,
+    snapshot_uri: String,
+    storage_config: ObjectStoreConfig,
+}
+
+#[async_trait]
+impl Tool for ExportCreate {
+    async fn do_work(&self) -> std::result::Result<(), BoxedError> {
+        self.run().await.map_err(BoxedError::new)
+    }
+}
+
+impl ExportCreate {
+    async fn run(&self) -> Result<()> {
+        // 1. Check if snapshot exists
+        let exists = self.storage.exists().await?;
+
+        if exists {
+            if self.config.force {
+                info!("Deleting existing snapshot (--force)");
+                self.storage.delete_snapshot().await?;
+            } else {
+                // Resume mode - read existing manifest
+                let mut manifest = self.storage.read_manifest().await?;
+
+                // Check version compatibility
+                if manifest.version != MANIFEST_VERSION {
+                    return ManifestVersionMismatchSnafu {
+                        expected: MANIFEST_VERSION,
+                        found: manifest.version,
+                    }
+                    .fail();
+                }
+
+                validate_resume_config(&manifest, &self.config)?;
+
+                info!(
+                    "Resuming existing snapshot: {} (completed: {}/{} chunks)",
+                    manifest.snapshot_id,
+                    manifest.completed_count(),
+                    manifest.chunks.len()
+                );
+
+                if manifest.is_complete() {
+                    info!("Snapshot is already complete");
+                    return Ok(());
+                }
+
+                if manifest.schema_only {
+                    return Ok(());
+                }
+
+                export_data(
+                    self.storage.as_ref(),
+                    &self.database_client,
+                    &self.config.snapshot_uri,
+                    &self.config.storage_config,
+                    &mut manifest,
+                    self.config.parallelism,
+                )
+                .await?;
+                return Ok(());
+            }
+        }
+
+        // 2. Get schema list
+        let extractor = SchemaExtractor::new(&self.database_client, &self.config.catalog);
+        let schema_snapshot = extractor.extract(self.config.schemas.as_deref()).await?;
+
+        let schema_names: Vec<String> = schema_snapshot
+            .schemas
+            .iter()
+            .map(|s| s.name.clone())
+            .collect();
+        info!("Exporting schemas: {:?}", schema_names);
+
+        // 3. Create manifest
+        let mut manifest = Manifest::new_for_export(
+            self.config.catalog.clone(),
+            schema_names.clone(),
+            self.config.schema_only,
+            self.config.time_range.clone(),
+            self.config.format,
+            self.config.chunk_time_window,
+        )?;
+
+        // 4. Write schema files
+        self.storage.write_schema(&schema_snapshot).await?;
+        info!("Exported {} schemas", schema_snapshot.schemas.len());
+
+        // 5. Export DDL files for import recovery.
+        let ddl_by_schema = self.build_ddl_by_schema(&schema_names).await?;
+        for (schema, ddl) in ddl_by_schema {
+            let ddl_path = ddl_path_for_schema(&schema);
+            self.storage.write_text(&ddl_path, &ddl).await?;
+            info!("Exported DDL for schema {} to {}", schema, ddl_path);
+        }
+
+        // 6. Write manifest after schema artifacts and before any data export.
+        //
+        // The manifest is the snapshot commit point: only write it after the schema
+        // index and all DDL files are durable, so a crash cannot leave a "valid"
+        // snapshot that is missing required schema artifacts. For full exports we
+        // still need the manifest before data copy starts, because chunk resume is
+        // tracked by updating this manifest in place.
+        self.storage.write_manifest(&manifest).await?;
+        info!("Snapshot created: {}", manifest.snapshot_id);
+
+        if !self.config.schema_only {
+            export_data(
+                self.storage.as_ref(),
+                &self.database_client,
+                &self.config.snapshot_uri,
+                &self.config.storage_config,
+                &mut manifest,
+                self.config.parallelism,
+            )
+            .await?;
+        }
+
+        Ok(())
+    }
+
+    async fn build_ddl_by_schema(&self, schema_names: &[String]) -> Result<Vec<(String, String)>> {
+        let mut schemas = schema_names.to_vec();
+        schemas.sort();
+
+        let mut ddl_by_schema = Vec::with_capacity(schemas.len());
+        for schema in schemas {
+            let create_database = self.show_create("DATABASE", &schema, None).await?;
+
+            let (mut physical_tables, mut tables, mut views) =
+                self.get_schema_objects(&schema).await?;
+            physical_tables.sort();
+            let mut physical_ddls = Vec::with_capacity(physical_tables.len());
+            for table in physical_tables {
+                physical_ddls.push(self.show_create("TABLE", &schema, Some(&table)).await?);
+            }
+
+            tables.sort();
+            let mut table_ddls = Vec::with_capacity(tables.len());
+            for table in tables {
+                table_ddls.push(self.show_create("TABLE", &schema, Some(&table)).await?);
+            }
+
+            views.sort();
+            let mut view_ddls = Vec::with_capacity(views.len());
+            for view in views {
+                view_ddls.push(self.show_create("VIEW", &schema, Some(&view)).await?);
+            }
+
+            let ddl = build_schema_ddl(
+                &schema,
+                create_database,
+                physical_ddls,
+                table_ddls,
+                view_ddls,
+            );
+            ddl_by_schema.push((schema, ddl));
+        }
+
+        Ok(ddl_by_schema)
+    }
+
+    async fn get_schema_objects(
+        &self,
+        schema: &str,
+    ) -> Result<(Vec<String>, Vec<String>, Vec<String>)> {
+        let physical_tables = self.get_metric_physical_tables(schema).await?;
+        let physical_set: HashSet<&str> = physical_tables.iter().map(String::as_str).collect();
+        let sql = format!(
+            "SELECT table_name, table_type FROM information_schema.tables \
+             WHERE table_catalog = '{}' AND table_schema = '{}' \
+             AND (table_type = 'BASE TABLE' OR table_type = 'VIEW')",
+            escape_sql_literal(&self.config.catalog),
+            escape_sql_literal(schema)
+        );
+        let records: Option<Vec<Vec<Value>>> = self
+            .database_client
+            .sql_in_public(&sql)
+            .await
+            .context(DatabaseSnafu)?;
+
+        let mut tables = Vec::new();
+        let mut views = Vec::new();
+        if let Some(rows) = records {
+            for row in rows {
+                let name = match row.first() {
+                    Some(Value::String(name)) => name.clone(),
+                    _ => return UnexpectedValueTypeSnafu.fail(),
+                };
+                let table_type = match row.get(1) {
+                    Some(Value::String(table_type)) => table_type.as_str(),
+                    _ => return UnexpectedValueTypeSnafu.fail(),
+                };
+                if !physical_set.contains(name.as_str()) {
+                    if table_type == "VIEW" {
+                        views.push(name);
+                    } else {
+                        tables.push(name);
+                    }
+                }
+            }
+        }
+
+        Ok((physical_tables, tables, views))
+    }
+
+    async fn get_metric_physical_tables(&self, schema: &str) -> Result<Vec<String>> {
+        let sql = format!(
+            "SELECT DISTINCT table_name FROM information_schema.columns \
+             WHERE table_catalog = '{}' AND table_schema = '{}' AND column_name = '__tsid'",
+            escape_sql_literal(&self.config.catalog),
+            escape_sql_literal(schema)
+        );
+        let records: Option<Vec<Vec<Value>>> = self
+            .database_client
+            .sql_in_public(&sql)
+            .await
+            .context(DatabaseSnafu)?;
+
+        let mut tables = HashSet::new();
+        if let Some(rows) = records {
+            for row in rows {
+                let name = match row.first() {
+                    Some(Value::String(name)) => name.clone(),
+                    _ => return UnexpectedValueTypeSnafu.fail(),
+                };
+                tables.insert(name);
+            }
+        }
+
+        Ok(tables.into_iter().collect())
+    }
+
+    async fn show_create(
+        &self,
+        show_type: &str,
+        schema: &str,
+        table: Option<&str>,
+    ) -> Result<String> {
+        let sql = match table {
+            Some(table) => format!(
+                r#"SHOW CREATE {} "{}"."{}"."{}""#,
+                show_type,
+                escape_sql_identifier(&self.config.catalog),
+                escape_sql_identifier(schema),
+                escape_sql_identifier(table)
+            ),
+            None => format!(
+                r#"SHOW CREATE {} "{}"."{}""#,
+                show_type,
+                escape_sql_identifier(&self.config.catalog),
+                escape_sql_identifier(schema)
+            ),
+        };
+
+        let records: Option<Vec<Vec<Value>>> = self
+            .database_client
+            .sql_in_public(&sql)
+            .await
+            .context(DatabaseSnafu)?;
+        let rows = records.context(EmptyResultSnafu)?;
+        let row = rows.first().context(EmptyResultSnafu)?;
+        let Some(Value::String(create)) = row.get(1) else {
+            return UnexpectedValueTypeSnafu.fail();
+        };
+
+        Ok(format!("{};\n", create))
+    }
+}
+
+fn build_schema_ddl(
+    schema: &str,
+    create_database: String,
+    physical_tables: Vec<String>,
+    tables: Vec<String>,
+    views: Vec<String>,
+) -> String {
+    let mut ddl = String::new();
+    ddl.push_str(&format!("-- Schema: {}\n", schema));
+    ddl.push_str(&create_database);
+    for stmt in physical_tables {
+        ddl.push_str(&stmt);
+    }
+    for stmt in tables {
+        ddl.push_str(&stmt);
+    }
+    for stmt in views {
+        ddl.push_str(&stmt);
+    }
+    ddl.push('\n');
+    ddl
+}
+
+fn validate_resume_config(manifest: &Manifest, config: &ExportConfig) -> Result<()> {
+    if manifest.schema_only != config.schema_only {
+        return SchemaOnlyModeMismatchSnafu {
+            existing_schema_only: manifest.schema_only,
+            requested_schema_only: config.schema_only,
+        }
+        .fail();
+    }
+
+    if manifest.catalog != config.catalog {
+        return ResumeConfigMismatchSnafu {
+            field: "catalog",
+            existing: manifest.catalog.clone(),
+            requested: config.catalog.clone(),
+        }
+        .fail();
+    }
+
+    // If no schema filter is provided on resume, inherit the existing snapshot
+    // selection instead of reinterpreting the request as "all schemas".
+    if let Some(requested_schemas) = &config.schemas
+        && !schema_selection_matches(&manifest.schemas, requested_schemas)
+    {
+        return ResumeConfigMismatchSnafu {
+            field: "schemas",
+            existing: format_schema_selection(&manifest.schemas),
+            requested: format_schema_selection(requested_schemas),
+        }
+        .fail();
+    }
+
+    if manifest.time_range != config.time_range {
+        return ResumeConfigMismatchSnafu {
+            field: "time_range",
+            existing: format!("{:?}", manifest.time_range),
+            requested: format!("{:?}", config.time_range),
+        }
+        .fail();
+    }
+
+    if manifest.format != config.format {
+        return ResumeConfigMismatchSnafu {
+            field: "format",
+            existing: manifest.format.to_string(),
+            requested: config.format.to_string(),
+        }
+        .fail();
+    }
+
+    let expected_plan = Manifest::new_for_export(
+        manifest.catalog.clone(),
+        manifest.schemas.clone(),
+        config.schema_only,
+        config.time_range.clone(),
+        config.format,
+        config.chunk_time_window,
+    )?;
+    if !chunk_plan_matches(manifest, &expected_plan) {
+        return ResumeConfigMismatchSnafu {
+            field: "chunk plan",
+            existing: format_chunk_plan(&manifest.chunks),
+            requested: format_chunk_plan(&expected_plan.chunks),
+        }
+        .fail();
+    }
+
+    Ok(())
+}
+
+fn schema_selection_matches(existing: &[String], requested: &[String]) -> bool {
+    canonical_schema_selection(existing) == canonical_schema_selection(requested)
+}
+
+fn canonical_schema_selection(schemas: &[String]) -> Vec<String> {
+    let mut canonicalized = Vec::new();
+    let mut seen = HashSet::new();
+
+    for schema in schemas {
+        let normalized = schema.to_ascii_lowercase();
+        if seen.insert(normalized.clone()) {
+            canonicalized.push(normalized);
+        }
+    }
+
+    canonicalized.sort();
+    canonicalized
+}
+
+fn format_schema_selection(schemas: &[String]) -> String {
+    format!("[{}]", schemas.join(", "))
+}
+
+fn chunk_plan_matches(existing: &Manifest, expected: &Manifest) -> bool {
+    existing.chunks.len() == expected.chunks.len()
+        && existing
+            .chunks
+            .iter()
+            .zip(&expected.chunks)
+            .all(|(left, right)| left.id == right.id && left.time_range == right.time_range)
+}
+
+fn format_chunk_plan(chunks: &[ChunkMeta]) -> String {
+    let items = chunks
+        .iter()
+        .map(|chunk| format!("#{}:{:?}", chunk.id, chunk.time_range))
+        .collect::<Vec<_>>();
+    format!("[{}]", items.join(", "))
+}
+
+#[derive(Debug)]
+struct SnapshotListEntry {
+    path: String,
+    manifest: Manifest,
+}
+
+#[derive(Debug, Default)]
+struct SnapshotScanResult {
+    snapshots: Vec<SnapshotListEntry>,
+    unreadable: Vec<String>,
+}
+
+async fn scan_snapshots(storage: &OpenDalStorage) -> Result<SnapshotScanResult> {
+    let mut result = SnapshotScanResult::default();
+    for dir in storage.list_direct_child_dirs().await? {
+        let manifest_path = format!("{}/{}", dir.trim_matches('/'), MANIFEST_FILE);
+        let Some(data) = storage.read_file_if_exists(&manifest_path).await? else {
+            continue;
+        };
+
+        match serde_json::from_slice::<Manifest>(&data) {
+            Ok(manifest) => result.snapshots.push(SnapshotListEntry {
+                path: format!("{}/", dir.trim_matches('/')),
+                manifest,
+            }),
+            Err(_) => result
+                .unreadable
+                .push(format!("{}/", dir.trim_matches('/'))),
+        }
+    }
+
+    result
+        .snapshots
+        .sort_by_key(|entry| std::cmp::Reverse(entry.manifest.created_at));
+    result.unreadable.sort();
+    Ok(result)
+}
+
+fn print_snapshot_list(snapshots: &[SnapshotListEntry], unreadable_count: usize) {
+    if unreadable_count == 0 {
+        println!("Found {} snapshots:", snapshots.len());
+    } else {
+        println!(
+            "Found {} snapshots ({} {} skipped: unreadable manifest):",
+            snapshots.len(),
+            unreadable_count,
+            directory_word(unreadable_count)
+        );
+    }
+    println!();
+    println!(
+        "  {:<24}  {:<36}  {:<19}  {:<9}  {:<7}  {:<6}  Status",
+        "Path", "ID", "Created", "Catalog", "Schemas", "Chunks"
+    );
+    println!(
+        "  {:<24}  {:<36}  {:<19}  {:<9}  {:<7}  {:<6}  {:<10}",
+        "-".repeat(24),
+        "-".repeat(36),
+        "-".repeat(19),
+        "-".repeat(9),
+        "-".repeat(7),
+        "-".repeat(6),
+        "-".repeat(10)
+    );
+    for entry in snapshots {
+        let manifest = &entry.manifest;
+        println!(
+            "  {:<24}  {:<36}  {:<19}  {:<9}  {:<7}  {:<6}  {}",
+            entry.path,
+            manifest.snapshot_id,
+            manifest.created_at.format("%Y-%m-%d %H:%M:%S"),
+            manifest.catalog,
+            manifest.schemas.len(),
+            format_list_chunks(manifest),
+            snapshot_status(manifest)
+        );
+    }
+}
+
+fn print_unreadable_warnings(unreadable: &[String]) {
+    if unreadable.is_empty() {
+        return;
+    }
+
+    println!();
+    println!(
+        "Warning: {} {} had corrupt/unreadable manifest.json:",
+        unreadable.len(),
+        directory_word(unreadable.len())
+    );
+    for path in unreadable {
+        println!("  - {}", path);
+    }
+}
+
+fn directory_word(count: usize) -> &'static str {
+    if count == 1 {
+        "directory"
+    } else {
+        "directories"
+    }
+}
+
+fn snapshot_status(manifest: &Manifest) -> &'static str {
+    if manifest.schema_only {
+        "schema-only"
+    } else if manifest.is_complete() {
+        "complete"
+    } else {
+        "incomplete"
+    }
+}
+
+fn format_list_chunks(manifest: &Manifest) -> String {
+    let total = manifest.chunks.len();
+    if total == 0 {
+        return "0".to_string();
+    }
+
+    format!(
+        "{}/{}",
+        manifest.completed_count() + manifest.skipped_count(),
+        total
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use chrono::TimeZone;
+    use clap::Parser;
+    use tempfile::tempdir;
+    use url::Url;
+
+    use super::*;
+    use crate::data::path::ddl_path_for_schema;
+
+    #[test]
+    fn test_ddl_path_for_schema() {
+        assert_eq!(ddl_path_for_schema("public"), "schema/ddl/public.sql");
+        assert_eq!(
+            ddl_path_for_schema("../evil"),
+            "schema/ddl/%2E%2E%2Fevil.sql"
+        );
+    }
+
+    #[test]
+    fn test_build_schema_ddl_order() {
+        let ddl = build_schema_ddl(
+            "public",
+            "CREATE DATABASE public;\n".to_string(),
+            vec!["PHYSICAL;\n".to_string()],
+            vec!["TABLE;\n".to_string()],
+            vec!["VIEW;\n".to_string()],
+        );
+
+        let db_pos = ddl.find("CREATE DATABASE").unwrap();
+        let physical_pos = ddl.find("PHYSICAL;").unwrap();
+        let table_pos = ddl.find("TABLE;").unwrap();
+        let view_pos = ddl.find("VIEW;").unwrap();
+        assert!(db_pos < physical_pos);
+        assert!(physical_pos < table_pos);
+        assert!(table_pos < view_pos);
+    }
+
+    #[tokio::test]
+    async fn test_build_rejects_chunk_window_without_bounds() {
+        let cmd = ExportCreateCommand::parse_from([
+            "export-v2-create",
+            "--addr",
+            "127.0.0.1:4000",
+            "--to",
+            "file:///tmp/export-v2-test",
+            "--chunk-time-window",
+            "1h",
+        ]);
+
+        let result = cmd.build().await;
+        assert!(result.is_err());
+        let error = result.err().unwrap().to_string();
+
+        assert!(error.contains("chunk_time_window requires both --start-time and --end-time"));
+    }
+
+    #[tokio::test]
+    async fn test_build_rejects_data_export_args_in_schema_only_mode() {
+        let cmd = ExportCreateCommand::parse_from([
+            "export-v2-create",
+            "--addr",
+            "127.0.0.1:4000",
+            "--to",
+            "file:///tmp/export-v2-test",
+            "--schema-only",
+            "--start-time",
+            "2024-01-01T00:00:00Z",
+            "--end-time",
+            "2024-01-02T00:00:00Z",
+            "--chunk-time-window",
+            "1h",
+            "--format",
+            "csv",
+            "--parallelism",
+            "2",
+        ]);
+
+        let error = cmd.build().await.err().unwrap().to_string();
+
+        assert!(error.contains("--schema-only cannot be used with data export arguments"));
+        assert!(error.contains("--start-time"));
+        assert!(error.contains("--end-time"));
+        assert!(error.contains("--chunk-time-window"));
+        assert!(error.contains("--format"));
+        assert!(error.contains("--parallelism"));
+    }
+
+    #[test]
+    fn test_schema_only_mode_mismatch_error_message() {
+        let error = crate::data::export_v2::error::SchemaOnlyModeMismatchSnafu {
+            existing_schema_only: false,
+            requested_schema_only: true,
+        }
+        .build()
+        .to_string();
+
+        assert!(error.contains("existing: false"));
+        assert!(error.contains("requested: true"));
+    }
+
+    #[test]
+    fn test_validate_resume_config_rejects_catalog_mismatch() {
+        let manifest = Manifest::new_for_export(
+            "greptime".to_string(),
+            vec!["public".to_string()],
+            false,
+            TimeRange::unbounded(),
+            DataFormat::Parquet,
+            None,
+        )
+        .unwrap();
+        let config = ExportConfig {
+            catalog: "other".to_string(),
+            schemas: None,
+            schema_only: false,
+            format: DataFormat::Parquet,
+            force: false,
+            time_range: TimeRange::unbounded(),
+            chunk_time_window: None,
+            parallelism: 1,
+            snapshot_uri: "file:///tmp/snapshot".to_string(),
+            storage_config: ObjectStoreConfig::default(),
+        };
+
+        let error = validate_resume_config(&manifest, &config)
+            .err()
+            .unwrap()
+            .to_string();
+        assert!(error.contains("catalog"));
+    }
+
+    #[test]
+    fn test_validate_resume_config_accepts_schema_selection_with_different_case_and_order() {
+        let manifest = Manifest::new_for_export(
+            "greptime".to_string(),
+            vec!["public".to_string(), "analytics".to_string()],
+            false,
+            TimeRange::unbounded(),
+            DataFormat::Parquet,
+            None,
+        )
+        .unwrap();
+        let config = ExportConfig {
+            catalog: "greptime".to_string(),
+            schemas: Some(vec![
+                "ANALYTICS".to_string(),
+                "PUBLIC".to_string(),
+                "public".to_string(),
+            ]),
+            schema_only: false,
+            format: DataFormat::Parquet,
+            force: false,
+            time_range: TimeRange::unbounded(),
+            chunk_time_window: None,
+            parallelism: 1,
+            snapshot_uri: "file:///tmp/snapshot".to_string(),
+            storage_config: ObjectStoreConfig::default(),
+        };
+
+        assert!(validate_resume_config(&manifest, &config).is_ok());
+    }
+
+    #[test]
+    fn test_validate_resume_config_rejects_chunk_plan_mismatch() {
+        let start = chrono::Utc.with_ymd_and_hms(2025, 1, 1, 0, 0, 0).unwrap();
+        let end = chrono::Utc.with_ymd_and_hms(2025, 1, 1, 2, 0, 0).unwrap();
+        let time_range = TimeRange::new(Some(start), Some(end));
+        let manifest = Manifest::new_for_export(
+            "greptime".to_string(),
+            vec!["public".to_string()],
+            false,
+            time_range.clone(),
+            DataFormat::Parquet,
+            None,
+        )
+        .unwrap();
+        let config = ExportConfig {
+            catalog: "greptime".to_string(),
+            schemas: None,
+            schema_only: false,
+            format: DataFormat::Parquet,
+            force: false,
+            time_range,
+            chunk_time_window: Some(Duration::from_secs(3600)),
+            parallelism: 1,
+            snapshot_uri: "file:///tmp/snapshot".to_string(),
+            storage_config: ObjectStoreConfig::default(),
+        };
+
+        let error = validate_resume_config(&manifest, &config)
+            .err()
+            .unwrap()
+            .to_string();
+        assert!(error.contains("chunk plan"));
+    }
+
+    #[test]
+    fn test_validate_resume_config_rejects_format_mismatch() {
+        let manifest = Manifest::new_for_export(
+            "greptime".to_string(),
+            vec!["public".to_string()],
+            false,
+            TimeRange::unbounded(),
+            DataFormat::Parquet,
+            None,
+        )
+        .unwrap();
+        let config = ExportConfig {
+            catalog: "greptime".to_string(),
+            schemas: None,
+            schema_only: false,
+            format: DataFormat::Csv,
+            force: false,
+            time_range: TimeRange::unbounded(),
+            chunk_time_window: None,
+            parallelism: 1,
+            snapshot_uri: "file:///tmp/snapshot".to_string(),
+            storage_config: ObjectStoreConfig::default(),
+        };
+
+        let error = validate_resume_config(&manifest, &config)
+            .err()
+            .unwrap()
+            .to_string();
+        assert!(error.contains("format"));
+    }
+
+    #[test]
+    fn test_validate_resume_config_rejects_time_range_mismatch() {
+        let start = chrono::Utc.with_ymd_and_hms(2025, 1, 1, 0, 0, 0).unwrap();
+        let end = chrono::Utc.with_ymd_and_hms(2025, 1, 1, 1, 0, 0).unwrap();
+        let manifest = Manifest::new_for_export(
+            "greptime".to_string(),
+            vec!["public".to_string()],
+            false,
+            TimeRange::new(Some(start), Some(end)),
+            DataFormat::Parquet,
+            None,
+        )
+        .unwrap();
+        let config = ExportConfig {
+            catalog: "greptime".to_string(),
+            schemas: None,
+            schema_only: false,
+            format: DataFormat::Parquet,
+            force: false,
+            time_range: TimeRange::new(Some(start), Some(start)),
+            chunk_time_window: None,
+            parallelism: 1,
+            snapshot_uri: "file:///tmp/snapshot".to_string(),
+            storage_config: ObjectStoreConfig::default(),
+        };
+
+        let error = validate_resume_config(&manifest, &config)
+            .err()
+            .unwrap()
+            .to_string();
+        assert!(error.contains("time_range"));
+    }
+
+    #[tokio::test]
+    async fn test_scan_snapshots_sorts_and_tracks_unreadable_manifests() {
+        let dir = tempdir().unwrap();
+        write_test_manifest(
+            dir.path(),
+            "older",
+            test_manifest(
+                chrono::Utc.with_ymd_and_hms(2026, 1, 1, 0, 0, 0).unwrap(),
+                false,
+                true,
+            ),
+        );
+        write_test_manifest(
+            dir.path(),
+            "newer",
+            test_manifest(
+                chrono::Utc.with_ymd_and_hms(2026, 2, 1, 0, 0, 0).unwrap(),
+                false,
+                true,
+            ),
+        );
+
+        std::fs::create_dir_all(dir.path().join("empty-dir")).unwrap();
+        std::fs::create_dir_all(dir.path().join("not-snapshot")).unwrap();
+        std::fs::write(dir.path().join("not-snapshot").join("data.txt"), "x").unwrap();
+        std::fs::create_dir_all(dir.path().join("broken")).unwrap();
+        std::fs::write(dir.path().join("broken").join(MANIFEST_FILE), "{not-json").unwrap();
+
+        let uri = Url::from_directory_path(dir.path()).unwrap().to_string();
+        let storage = OpenDalStorage::from_file_uri(&uri).unwrap();
+        let result = scan_snapshots(&storage).await.unwrap();
+
+        assert_eq!(result.snapshots.len(), 2);
+        assert_eq!(
+            result.snapshots[0].manifest.created_at,
+            chrono::Utc.with_ymd_and_hms(2026, 2, 1, 0, 0, 0).unwrap()
+        );
+        assert_eq!(
+            result.snapshots[1].manifest.created_at,
+            chrono::Utc.with_ymd_and_hms(2026, 1, 1, 0, 0, 0).unwrap()
+        );
+        assert_eq!(result.unreadable, vec!["broken/".to_string()]);
+        assert_eq!(result.snapshots[0].path, "newer/");
+        assert_eq!(result.snapshots[1].path, "older/");
+    }
+
+    #[test]
+    fn test_snapshot_list_status_and_chunk_summary() {
+        let schema_only = test_manifest(
+            chrono::Utc.with_ymd_and_hms(2026, 1, 1, 0, 0, 0).unwrap(),
+            true,
+            true,
+        );
+        assert_eq!(snapshot_status(&schema_only), "schema-only");
+        assert_eq!(format_list_chunks(&schema_only), "0");
+
+        let complete = test_manifest(
+            chrono::Utc.with_ymd_and_hms(2026, 1, 1, 0, 0, 0).unwrap(),
+            false,
+            true,
+        );
+        assert_eq!(snapshot_status(&complete), "complete");
+        assert_eq!(format_list_chunks(&complete), "2/2");
+
+        let incomplete = test_manifest(
+            chrono::Utc.with_ymd_and_hms(2026, 1, 1, 0, 0, 0).unwrap(),
+            false,
+            false,
+        );
+        assert_eq!(snapshot_status(&incomplete), "incomplete");
+        assert_eq!(format_list_chunks(&incomplete), "1/2");
+    }
+
+    fn write_test_manifest(root: &std::path::Path, dir: &str, manifest: Manifest) {
+        let snapshot_dir = root.join(dir);
+        std::fs::create_dir_all(&snapshot_dir).unwrap();
+        std::fs::write(
+            snapshot_dir.join(MANIFEST_FILE),
+            serde_json::to_vec_pretty(&manifest).unwrap(),
+        )
+        .unwrap();
+    }
+
+    fn test_manifest(
+        created_at: chrono::DateTime<chrono::Utc>,
+        schema_only: bool,
+        complete: bool,
+    ) -> Manifest {
+        let mut manifest = Manifest::new_for_export(
+            "greptime".to_string(),
+            vec!["public".to_string(), "analytics".to_string()],
+            schema_only,
+            TimeRange::unbounded(),
+            DataFormat::Parquet,
+            None,
+        )
+        .unwrap();
+        manifest.created_at = created_at;
+        manifest.updated_at = created_at;
+
+        if !schema_only {
+            manifest.chunks.clear();
+            let mut first = ChunkMeta::new(1, TimeRange::unbounded());
+            first.mark_completed(vec!["data/public/chunk_1/file.parquet".to_string()], None);
+            manifest.chunks.push(first);
+
+            if complete {
+                manifest
+                    .chunks
+                    .push(ChunkMeta::skipped(2, TimeRange::unbounded()));
+            } else {
+                manifest
+                    .chunks
+                    .push(ChunkMeta::new(2, TimeRange::unbounded()));
+            }
+        }
+
+        manifest
+    }
+}
diff --git a/src/cli/src/data/export_v2/coordinator.rs b/src/cli/src/data/export_v2/coordinator.rs
new file mode 100644
index 0000000000..d96c01d693
--- /dev/null
+++ b/src/cli/src/data/export_v2/coordinator.rs
@@ -0,0 +1,166 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use common_telemetry::info;
+
+use crate::common::ObjectStoreConfig;
+use crate::data::export_v2::data::{CopyOptions, build_copy_target, execute_copy_database};
+use crate::data::export_v2::error::Result;
+use crate::data::export_v2::manifest::{ChunkStatus, DataFormat, Manifest, TimeRange};
+use crate::data::path::data_dir_for_schema_chunk;
+use crate::data::snapshot_storage::{SnapshotStorage, StorageScheme};
+use crate::database::DatabaseClient;
+
+struct ExportContext<'a> {
+    storage: &'a dyn SnapshotStorage,
+    database_client: &'a DatabaseClient,
+    snapshot_uri: &'a str,
+    storage_config: &'a ObjectStoreConfig,
+    catalog: &'a str,
+    schemas: &'a [String],
+    format: DataFormat,
+    parallelism: usize,
+}
+
+pub async fn export_data(
+    storage: &dyn SnapshotStorage,
+    database_client: &DatabaseClient,
+    snapshot_uri: &str,
+    storage_config: &ObjectStoreConfig,
+    manifest: &mut Manifest,
+    parallelism: usize,
+) -> Result<()> {
+    if manifest.chunks.is_empty() {
+        return Ok(());
+    }
+
+    for idx in 0..manifest.chunks.len() {
+        if matches!(
+            manifest.chunks[idx].status,
+            ChunkStatus::Completed | ChunkStatus::Skipped
+        ) {
+            continue;
+        }
+
+        let (chunk_id, time_range) = mark_chunk_in_progress(manifest, idx);
+        manifest.touch();
+        storage.write_manifest(manifest).await?;
+
+        let context = ExportContext {
+            storage,
+            database_client,
+            snapshot_uri,
+            storage_config,
+            catalog: &manifest.catalog,
+            schemas: &manifest.schemas,
+            format: manifest.format,
+            parallelism,
+        };
+        let export_result = export_chunk(&context, chunk_id, time_range).await;
+
+        let result = match export_result {
+            Ok(files) => {
+                mark_chunk_completed(manifest, idx, files);
+                Ok(())
+            }
+            Err(err) => {
+                mark_chunk_failed(manifest, idx, err.to_string());
+                Err(err)
+            }
+        };
+
+        manifest.touch();
+        storage.write_manifest(manifest).await?;
+
+        result?;
+    }
+
+    Ok(())
+}
+
+fn mark_chunk_in_progress(manifest: &mut Manifest, idx: usize) -> (u32, TimeRange) {
+    let chunk = &mut manifest.chunks[idx];
+    chunk.mark_in_progress();
+    (chunk.id, chunk.time_range.clone())
+}
+
+fn mark_chunk_completed(manifest: &mut Manifest, idx: usize, files: Vec<String>) {
+    let chunk = &mut manifest.chunks[idx];
+    if files.is_empty() {
+        chunk.mark_skipped();
+    } else {
+        chunk.mark_completed(files, None);
+    }
+}
+
+fn mark_chunk_failed(manifest: &mut Manifest, idx: usize, error: String) {
+    let chunk = &mut manifest.chunks[idx];
+    chunk.mark_failed(error);
+}
+
+async fn export_chunk(
+    context: &ExportContext<'_>,
+    chunk_id: u32,
+    time_range: TimeRange,
+) -> Result<Vec<String>> {
+    let scheme = StorageScheme::from_uri(context.snapshot_uri)?;
+    let needs_dir = matches!(scheme, StorageScheme::File);
+    let copy_options = CopyOptions {
+        format: context.format,
+        time_range,
+        parallelism: context.parallelism,
+    };
+
+    for schema in context.schemas {
+        let prefix = data_dir_for_schema_chunk(schema, chunk_id);
+        if needs_dir {
+            context.storage.create_dir_all(&prefix).await?;
+        }
+
+        let target = build_copy_target(
+            context.snapshot_uri,
+            context.storage_config,
+            schema,
+            chunk_id,
+        )?;
+        execute_copy_database(
+            context.database_client,
+            context.catalog,
+            schema,
+            &target,
+            &copy_options,
+        )
+        .await?;
+    }
+
+    let files = list_chunk_files(context.storage, context.schemas, chunk_id).await?;
+    info!("Collected {} files for chunk {}", files.len(), chunk_id);
+    Ok(files)
+}
+
+async fn list_chunk_files(
+    storage: &dyn SnapshotStorage,
+    schemas: &[String],
+    chunk_id: u32,
+) -> Result<Vec<String>> {
+    let mut files = Vec::new();
+
+    for schema in schemas {
+        let prefix = data_dir_for_schema_chunk(schema, chunk_id);
+        files.extend(storage.list_files_recursive(&prefix).await?);
+    }
+
+    files.sort();
+    Ok(files)
+}
diff --git a/src/cli/src/data/export_v2/data.rs b/src/cli/src/data/export_v2/data.rs
new file mode 100644
index 0000000000..dce924a717
--- /dev/null
+++ b/src/cli/src/data/export_v2/data.rs
@@ -0,0 +1,538 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use common_base::secrets::{ExposeSecret, SecretString};
+use common_telemetry::info;
+use object_store::util::{join_path, normalize_path};
+use snafu::ResultExt;
+use url::Url;
+
+use crate::common::ObjectStoreConfig;
+use crate::data::export_v2::error::{DatabaseSnafu, InvalidUriSnafu, Result, UrlParseSnafu};
+use crate::data::export_v2::manifest::{DataFormat, TimeRange};
+use crate::data::path::data_dir_for_schema_chunk;
+use crate::data::snapshot_storage::StorageScheme;
+use crate::data::sql::{escape_sql_identifier, escape_sql_literal};
+use crate::database::DatabaseClient;
+
+pub(super) struct CopyOptions {
+    pub(super) format: DataFormat,
+    pub(super) time_range: TimeRange,
+    pub(super) parallelism: usize,
+}
+
+pub(super) struct CopyTarget {
+    pub(super) location: String,
+    pub(super) connection: String,
+    secrets: Vec<Option<String>>,
+}
+
+pub(crate) struct CopySource {
+    pub(crate) location: String,
+    pub(crate) connection: String,
+    secrets: Vec<Option<String>>,
+}
+
+impl CopyTarget {
+    fn mask_sql(&self, sql: &str) -> String {
+        mask_secrets(sql, &self.secrets)
+    }
+}
+
+impl CopySource {
+    fn mask_sql(&self, sql: &str) -> String {
+        mask_secrets(sql, &self.secrets)
+    }
+}
+
+pub(super) fn build_copy_target(
+    snapshot_uri: &str,
+    storage: &ObjectStoreConfig,
+    schema: &str,
+    chunk_id: u32,
+) -> Result<CopyTarget> {
+    let location = build_copy_location(snapshot_uri, storage, schema, chunk_id)?;
+    Ok(CopyTarget {
+        location: location.location,
+        connection: location.connection,
+        secrets: location.secrets,
+    })
+}
+
+pub(crate) fn build_copy_source(
+    snapshot_uri: &str,
+    storage: &ObjectStoreConfig,
+    schema: &str,
+    chunk_id: u32,
+) -> Result<CopySource> {
+    let location = build_copy_location(snapshot_uri, storage, schema, chunk_id)?;
+    Ok(CopySource {
+        location: location.location,
+        connection: location.connection,
+        secrets: location.secrets,
+    })
+}
+
+struct CopyLocation {
+    location: String,
+    connection: String,
+    secrets: Vec<Option<String>>,
+}
+
+fn build_copy_location(
+    snapshot_uri: &str,
+    storage: &ObjectStoreConfig,
+    schema: &str,
+    chunk_id: u32,
+) -> Result<CopyLocation> {
+    let url = Url::parse(snapshot_uri).context(UrlParseSnafu)?;
+    let scheme = StorageScheme::from_uri(snapshot_uri)?;
+    let suffix = data_dir_for_schema_chunk(schema, chunk_id);
+
+    match scheme {
+        StorageScheme::File => {
+            let root = url.to_file_path().map_err(|_| {
+                InvalidUriSnafu {
+                    uri: snapshot_uri,
+                    reason: "file:// URI must use an absolute path like file:///tmp/backup",
+                }
+                .build()
+            })?;
+            let location = normalize_path(&format!("{}/{}", root.to_string_lossy(), suffix));
+            Ok(CopyLocation {
+                location,
+                connection: String::new(),
+                secrets: Vec::new(),
+            })
+        }
+        StorageScheme::S3 => {
+            let (bucket, root) = extract_bucket_root(&url, snapshot_uri)?;
+            let location = format!("s3://{}/{}", bucket, join_root(&root, &suffix));
+            let (connection, secrets) = build_s3_connection(storage);
+            Ok(CopyLocation {
+                location,
+                connection,
+                secrets,
+            })
+        }
+        StorageScheme::Oss => {
+            let (bucket, root) = extract_bucket_root(&url, snapshot_uri)?;
+            let location = format!("oss://{}/{}", bucket, join_root(&root, &suffix));
+            let (connection, secrets) = build_oss_connection(storage);
+            Ok(CopyLocation {
+                location,
+                connection,
+                secrets,
+            })
+        }
+        StorageScheme::Gcs => {
+            let (bucket, root) = extract_bucket_root(&url, snapshot_uri)?;
+            let location = format!("gcs://{}/{}", bucket, join_root(&root, &suffix));
+            let (connection, secrets) = build_gcs_connection(storage, snapshot_uri)?;
+            Ok(CopyLocation {
+                location,
+                connection,
+                secrets,
+            })
+        }
+        StorageScheme::Azblob => {
+            let (bucket, root) = extract_bucket_root(&url, snapshot_uri)?;
+            let location = format!("azblob://{}/{}", bucket, join_root(&root, &suffix));
+            let (connection, secrets) = build_azblob_connection(storage);
+            Ok(CopyLocation {
+                location,
+                connection,
+                secrets,
+            })
+        }
+    }
+}
+
+pub(super) async fn execute_copy_database(
+    database_client: &DatabaseClient,
+    catalog: &str,
+    schema: &str,
+    target: &CopyTarget,
+    options: &CopyOptions,
+) -> Result<()> {
+    let with_options = build_with_options(options);
+    let sql = format!(
+        r#"COPY DATABASE "{}"."{}" TO '{}' WITH ({}){};"#,
+        escape_sql_identifier(catalog),
+        escape_sql_identifier(schema),
+        escape_sql_literal(&target.location),
+        with_options,
+        target.connection
+    );
+    let safe_sql = target.mask_sql(&sql);
+    info!("Executing sql: {}", safe_sql);
+    database_client
+        .sql_in_public(&sql)
+        .await
+        .context(DatabaseSnafu)?;
+    Ok(())
+}
+
+pub(crate) async fn execute_copy_database_from(
+    database_client: &DatabaseClient,
+    catalog: &str,
+    schema: &str,
+    source: &CopySource,
+    format: DataFormat,
+) -> Result<()> {
+    let sql = format!(
+        r#"COPY DATABASE "{}"."{}" FROM '{}' WITH (FORMAT='{}'){};"#,
+        escape_sql_identifier(catalog),
+        escape_sql_identifier(schema),
+        escape_sql_literal(&source.location),
+        format,
+        source.connection
+    );
+    let safe_sql = source.mask_sql(&sql);
+    info!("Executing sql: {}", safe_sql);
+    database_client
+        .sql_in_public(&sql)
+        .await
+        .context(DatabaseSnafu)?;
+    Ok(())
+}
+
+fn build_with_options(options: &CopyOptions) -> String {
+    let mut parts = vec![format!("FORMAT='{}'", options.format)];
+    if let Some(start) = options.time_range.start {
+        parts.push(format!(
+            "START_TIME='{}'",
+            escape_sql_literal(&start.to_rfc3339())
+        ));
+    }
+    if let Some(end) = options.time_range.end {
+        parts.push(format!(
+            "END_TIME='{}'",
+            escape_sql_literal(&end.to_rfc3339())
+        ));
+    }
+    parts.push(format!("PARALLELISM={}", options.parallelism));
+    parts.join(", ")
+}
+
+fn extract_bucket_root(url: &Url, snapshot_uri: &str) -> Result<(String, String)> {
+    let bucket = url.host_str().unwrap_or("").to_string();
+    if bucket.is_empty() {
+        return InvalidUriSnafu {
+            uri: snapshot_uri,
+            reason: "URI must include bucket/container in host",
+        }
+        .fail();
+    }
+    let root = url
+        .path()
+        .trim_start_matches('/')
+        .trim_end_matches('/')
+        .to_string();
+    Ok((bucket, root))
+}
+
+fn join_root(root: &str, suffix: &str) -> String {
+    join_path(root, suffix).trim_start_matches('/').to_string()
+}
+
+fn build_s3_connection(storage: &ObjectStoreConfig) -> (String, Vec<Option<String>>) {
+    let access_key_id = expose_optional_secret(&storage.s3.s3_access_key_id);
+    let secret_access_key = expose_optional_secret(&storage.s3.s3_secret_access_key);
+
+    let mut options = Vec::new();
+    if let Some(access_key_id) = &access_key_id {
+        options.push(format!(
+            "ACCESS_KEY_ID='{}'",
+            escape_sql_literal(access_key_id)
+        ));
+    }
+    if let Some(secret_access_key) = &secret_access_key {
+        options.push(format!(
+            "SECRET_ACCESS_KEY='{}'",
+            escape_sql_literal(secret_access_key)
+        ));
+    }
+    if let Some(region) = &storage.s3.s3_region {
+        options.push(format!("REGION='{}'", escape_sql_literal(region)));
+    }
+    if let Some(endpoint) = &storage.s3.s3_endpoint {
+        options.push(format!("ENDPOINT='{}'", escape_sql_literal(endpoint)));
+    }
+
+    let secrets = vec![access_key_id, secret_access_key];
+    let connection = if options.is_empty() {
+        String::new()
+    } else {
+        format!(" CONNECTION ({})", options.join(", "))
+    };
+    (connection, secrets)
+}
+
+fn build_oss_connection(storage: &ObjectStoreConfig) -> (String, Vec<Option<String>>) {
+    let access_key_id = expose_optional_secret(&storage.oss.oss_access_key_id);
+    let access_key_secret = expose_optional_secret(&storage.oss.oss_access_key_secret);
+
+    let mut options = Vec::new();
+    if let Some(access_key_id) = &access_key_id {
+        options.push(format!(
+            "ACCESS_KEY_ID='{}'",
+            escape_sql_literal(access_key_id)
+        ));
+    }
+    if let Some(access_key_secret) = &access_key_secret {
+        options.push(format!(
+            "ACCESS_KEY_SECRET='{}'",
+            escape_sql_literal(access_key_secret)
+        ));
+    }
+    if !storage.oss.oss_endpoint.is_empty() {
+        options.push(format!(
+            "ENDPOINT='{}'",
+            escape_sql_literal(&storage.oss.oss_endpoint)
+        ));
+    }
+
+    let secrets = vec![access_key_id, access_key_secret];
+    let connection = if options.is_empty() {
+        String::new()
+    } else {
+        format!(" CONNECTION ({})", options.join(", "))
+    };
+    (connection, secrets)
+}
+
+fn build_gcs_connection(
+    storage: &ObjectStoreConfig,
+    snapshot_uri: &str,
+) -> Result<(String, Vec<Option<String>>)> {
+    let credential_path = expose_optional_secret(&storage.gcs.gcs_credential_path);
+    let credential = expose_optional_secret(&storage.gcs.gcs_credential);
+
+    if credential.is_none() && credential_path.is_some() {
+        return InvalidUriSnafu {
+            uri: snapshot_uri,
+            reason: "gcs_credential_path is not supported for server-side COPY; provide gcs_credential or rely on server-side ADC",
+        }
+        .fail();
+    }
+
+    let mut options = Vec::new();
+    if let Some(credential) = &credential {
+        options.push(format!("CREDENTIAL='{}'", escape_sql_literal(credential)));
+    }
+    if !storage.gcs.gcs_scope.is_empty() {
+        options.push(format!(
+            "SCOPE='{}'",
+            escape_sql_literal(&storage.gcs.gcs_scope)
+        ));
+    }
+    if !storage.gcs.gcs_endpoint.is_empty() {
+        options.push(format!(
+            "ENDPOINT='{}'",
+            escape_sql_literal(&storage.gcs.gcs_endpoint)
+        ));
+    }
+
+    let connection = if options.is_empty() {
+        String::new()
+    } else {
+        format!(" CONNECTION ({})", options.join(", "))
+    };
+    let secrets = vec![credential_path, credential];
+    Ok((connection, secrets))
+}
+
+fn build_azblob_connection(storage: &ObjectStoreConfig) -> (String, Vec<Option<String>>) {
+    let account_name = expose_optional_secret(&storage.azblob.azblob_account_name);
+    let account_key = expose_optional_secret(&storage.azblob.azblob_account_key);
+    let sas_token = storage.azblob.azblob_sas_token.clone();
+
+    let mut options = Vec::new();
+    if let Some(account_name) = &account_name {
+        options.push(format!(
+            "ACCOUNT_NAME='{}'",
+            escape_sql_literal(account_name)
+        ));
+    }
+    if let Some(account_key) = &account_key {
+        options.push(format!("ACCOUNT_KEY='{}'", escape_sql_literal(account_key)));
+    }
+    if let Some(sas_token) = &sas_token {
+        options.push(format!("SAS_TOKEN='{}'", escape_sql_literal(sas_token)));
+    }
+    if !storage.azblob.azblob_endpoint.is_empty() {
+        options.push(format!(
+            "ENDPOINT='{}'",
+            escape_sql_literal(&storage.azblob.azblob_endpoint)
+        ));
+    }
+
+    let secrets = vec![account_name, account_key, sas_token];
+    let connection = if options.is_empty() {
+        String::new()
+    } else {
+        format!(" CONNECTION ({})", options.join(", "))
+    };
+    (connection, secrets)
+}
+
+fn expose_optional_secret(secret: &Option<SecretString>) -> Option<String> {
+    secret.as_ref().map(|s| s.expose_secret().to_owned())
+}
+
+fn mask_secrets(sql: &str, secrets: &[Option<String>]) -> String {
+    let mut masked = sql.to_string();
+    for secret in secrets {
+        if let Some(secret) = secret
+            && !secret.is_empty()
+        {
+            let escaped = escape_sql_literal(secret);
+            if escaped != *secret {
+                masked = masked.replace(&escaped, "[REDACTED]");
+            }
+            masked = masked.replace(secret, "[REDACTED]");
+        }
+    }
+    masked
+}
+
+#[cfg(test)]
+mod tests {
+    use common_base::secrets::SecretString;
+    use common_test_util::temp_dir::create_temp_dir;
+
+    use super::*;
+    use crate::common::{PrefixedAzblobConnection, PrefixedGcsConnection, PrefixedOssConnection};
+
+    #[test]
+    fn test_build_oss_connection_includes_endpoint() {
+        let storage = ObjectStoreConfig {
+            oss: PrefixedOssConnection {
+                oss_endpoint: "https://oss.example.com".to_string(),
+                oss_access_key_id: Some(SecretString::from("key_id".to_string())),
+                oss_access_key_secret: Some(SecretString::from("key_secret".to_string())),
+                ..Default::default()
+            },
+            ..Default::default()
+        };
+
+        let (connection, _) = build_oss_connection(&storage);
+        assert!(connection.contains("ENDPOINT='https://oss.example.com'"));
+    }
+
+    #[test]
+    fn test_build_gcs_connection_uses_scope_and_inline_credential() {
+        let storage = ObjectStoreConfig {
+            gcs: PrefixedGcsConnection {
+                gcs_scope: "scope-a".to_string(),
+                gcs_endpoint: "https://storage.googleapis.com".to_string(),
+                gcs_credential: Some(SecretString::from("credential-json".to_string())),
+                ..Default::default()
+            },
+            ..Default::default()
+        };
+
+        let (connection, _) = build_gcs_connection(&storage, "gcs://bucket/root").unwrap();
+        assert!(connection.contains("CREDENTIAL='credential-json'"));
+        assert!(connection.contains("SCOPE='scope-a'"));
+        assert!(connection.contains("ENDPOINT='https://storage.googleapis.com'"));
+        assert!(!connection.contains("CREDENTIAL_PATH"));
+    }
+
+    #[test]
+    fn test_build_gcs_connection_rejects_credential_path_only() {
+        let storage = ObjectStoreConfig {
+            gcs: PrefixedGcsConnection {
+                gcs_scope: "scope-a".to_string(),
+                gcs_credential_path: Some(SecretString::from("/tmp/creds.json".to_string())),
+                ..Default::default()
+            },
+            ..Default::default()
+        };
+
+        let error = build_gcs_connection(&storage, "gcs://bucket/root")
+            .expect_err("credential_path-only should be rejected")
+            .to_string();
+        assert!(error.contains("gcs_credential_path is not supported"));
+    }
+
+    #[test]
+    fn test_build_azblob_connection_includes_endpoint() {
+        let storage = ObjectStoreConfig {
+            azblob: PrefixedAzblobConnection {
+                azblob_account_name: Some(SecretString::from("account".to_string())),
+                azblob_account_key: Some(SecretString::from("key".to_string())),
+                azblob_endpoint: "https://blob.example.com".to_string(),
+                ..Default::default()
+            },
+            ..Default::default()
+        };
+
+        let (connection, _) = build_azblob_connection(&storage);
+        assert!(connection.contains("ENDPOINT='https://blob.example.com'"));
+    }
+
+    #[test]
+    fn test_build_azblob_connection_redacts_sas_token() {
+        let storage = ObjectStoreConfig {
+            azblob: PrefixedAzblobConnection {
+                azblob_account_name: Some(SecretString::from("account".to_string())),
+                azblob_account_key: Some(SecretString::from("key".to_string())),
+                azblob_sas_token: Some("sig=secret-token".to_string()),
+                ..Default::default()
+            },
+            ..Default::default()
+        };
+
+        let (connection, secrets) = build_azblob_connection(&storage);
+        let masked = mask_secrets(&connection, &secrets);
+
+        assert!(connection.contains("SAS_TOKEN='sig=secret-token'"));
+        assert!(masked.contains("SAS_TOKEN='[REDACTED]'"));
+        assert!(!masked.contains("sig=secret-token"));
+    }
+
+    #[test]
+    fn test_mask_secrets_redacts_sql_escaped_literals() {
+        let sql =
+            "COPY DATABASE \"greptime\".\"public\" TO 's3://bucket' CONNECTION (SECRET='ab''cd');";
+        let masked = mask_secrets(sql, &[Some("ab'cd".to_string())]);
+
+        assert!(!masked.contains("ab'cd"));
+        assert!(!masked.contains("ab''cd"));
+        assert!(masked.contains("SECRET='[REDACTED]'"));
+    }
+
+    #[test]
+    fn test_build_copy_target_decodes_file_uri_path() {
+        let storage = ObjectStoreConfig::default();
+        let snapshot_root = create_temp_dir("my backup");
+        let snapshot_uri = Url::from_file_path(snapshot_root.path())
+            .expect("absolute platform path should convert to file:// URI")
+            .to_string();
+        let expected = normalize_path(&format!(
+            "{}/{}",
+            snapshot_root.path().to_string_lossy(),
+            data_dir_for_schema_chunk("public", 7)
+        ));
+        let target = build_copy_target(&snapshot_uri, &storage, "public", 7)
+            .expect("file:// copy target should be built");
+
+        assert!(snapshot_uri.contains("%20"));
+        assert!(!target.location.contains("%20"));
+        assert!(target.location.contains("my backup"));
+        assert_eq!(target.location, expected);
+    }
+}
diff --git a/src/cli/src/data/export_v2/error.rs b/src/cli/src/data/export_v2/error.rs
new file mode 100644
index 0000000000..ec860fecfa
--- /dev/null
+++ b/src/cli/src/data/export_v2/error.rs
@@ -0,0 +1,223 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::any::Any;
+
+use common_error::ext::ErrorExt;
+use common_error::status_code::StatusCode;
+use common_macro::stack_trace_debug;
+use snafu::{Location, Snafu};
+
+#[derive(Snafu)]
+#[snafu(visibility(pub))]
+#[stack_trace_debug]
+pub enum Error {
+    #[snafu(display("Invalid URI '{}': {}", uri, reason))]
+    InvalidUri {
+        uri: String,
+        reason: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Unsupported storage scheme: {}", scheme))]
+    UnsupportedScheme {
+        scheme: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Storage operation '{}' failed", operation))]
+    StorageOperation {
+        operation: String,
+        #[snafu(source)]
+        error: object_store::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Failed to parse manifest"))]
+    ManifestParse {
+        #[snafu(source)]
+        error: serde_json::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Failed to serialize manifest"))]
+    ManifestSerialize {
+        #[snafu(source)]
+        error: serde_json::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Failed to decode text file as UTF-8"))]
+    TextDecode {
+        #[snafu(source)]
+        error: std::string::FromUtf8Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display(
+        "Cannot resume snapshot with a different schema_only mode (existing: {}, requested: {}). Use --force to recreate.",
+        existing_schema_only,
+        requested_schema_only
+    ))]
+    SchemaOnlyModeMismatch {
+        existing_schema_only: bool,
+        requested_schema_only: bool,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display(
+        "Cannot resume snapshot with different {} (existing: {}, requested: {}). Use --force to recreate.",
+        field,
+        existing,
+        requested
+    ))]
+    ResumeConfigMismatch {
+        field: String,
+        existing: String,
+        requested: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Failed to parse time: invalid format: {}", input))]
+    TimeParseInvalidFormat {
+        input: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Failed to parse time: end_time is before start_time"))]
+    TimeParseEndBeforeStart {
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display(
+        "chunk_time_window requires both --start-time and --end-time to be specified"
+    ))]
+    ChunkTimeWindowRequiresBounds {
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("--schema-only cannot be used with data export arguments: {}", args))]
+    SchemaOnlyArgsNotAllowed {
+        args: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Empty result from query"))]
+    EmptyResult {
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Unexpected value type in query result"))]
+    UnexpectedValueType {
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Database error"))]
+    Database {
+        #[snafu(source)]
+        error: crate::error::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Snapshot not found at '{}'", uri))]
+    SnapshotNotFound {
+        uri: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Schema '{}' not found in catalog '{}'", schema, catalog))]
+    SchemaNotFound {
+        catalog: String,
+        schema: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Failed to parse URL"))]
+    UrlParse {
+        #[snafu(source)]
+        error: url::ParseError,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Failed to build object store"))]
+    BuildObjectStore {
+        #[snafu(source)]
+        error: object_store::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Manifest version mismatch: expected {}, found {}", expected, found))]
+    ManifestVersionMismatch {
+        expected: u32,
+        found: u32,
+        #[snafu(implicit)]
+        location: Location,
+    },
+}
+
+pub type Result<T> = std::result::Result<T, Error>;
+
+impl ErrorExt for Error {
+    fn status_code(&self) -> StatusCode {
+        match self {
+            Error::InvalidUri { .. }
+            | Error::UnsupportedScheme { .. }
+            | Error::SchemaOnlyModeMismatch { .. }
+            | Error::ResumeConfigMismatch { .. }
+            | Error::ManifestVersionMismatch { .. }
+            | Error::SchemaOnlyArgsNotAllowed { .. } => StatusCode::InvalidArguments,
+            Error::TimeParseInvalidFormat { .. }
+            | Error::TimeParseEndBeforeStart { .. }
+            | Error::ChunkTimeWindowRequiresBounds { .. } => StatusCode::InvalidArguments,
+
+            Error::StorageOperation { .. }
+            | Error::ManifestParse { .. }
+            | Error::ManifestSerialize { .. }
+            | Error::TextDecode { .. }
+            | Error::BuildObjectStore { .. } => StatusCode::StorageUnavailable,
+
+            Error::EmptyResult { .. }
+            | Error::UnexpectedValueType { .. }
+            | Error::UrlParse { .. } => StatusCode::Internal,
+
+            Error::Database { error, .. } => error.status_code(),
+
+            Error::SnapshotNotFound { .. } => StatusCode::InvalidArguments,
+            Error::SchemaNotFound { .. } => StatusCode::DatabaseNotFound,
+        }
+    }
+
+    fn as_any(&self) -> &dyn Any {
+        self
+    }
+}
diff --git a/src/cli/src/data/export_v2/extractor.rs b/src/cli/src/data/export_v2/extractor.rs
new file mode 100644
index 0000000000..ae15b199af
--- /dev/null
+++ b/src/cli/src/data/export_v2/extractor.rs
@@ -0,0 +1,254 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Schema extraction from information_schema.
+//!
+//! For V2 DDL-only snapshots, extractor only persists the schema index.
+
+use std::collections::{HashMap, HashSet};
+
+use serde_json::Value;
+use snafu::ResultExt;
+
+use crate::data::export_v2::error::{
+    DatabaseSnafu, EmptyResultSnafu, Result, SchemaNotFoundSnafu, UnexpectedValueTypeSnafu,
+};
+use crate::data::export_v2::schema::{SchemaDefinition, SchemaSnapshot};
+use crate::data::sql::escape_sql_literal;
+use crate::database::DatabaseClient;
+
+/// System schemas that should be excluded from export.
+const SYSTEM_SCHEMAS: &[&str] = &["information_schema", "pg_catalog"];
+
+/// Extracts schema definitions from information_schema.
+pub struct SchemaExtractor<'a> {
+    client: &'a DatabaseClient,
+    catalog: &'a str,
+}
+
+impl<'a> SchemaExtractor<'a> {
+    /// Creates a new schema extractor.
+    pub fn new(client: &'a DatabaseClient, catalog: &'a str) -> Self {
+        Self { client, catalog }
+    }
+
+    /// Extracts the schema index for the given schemas.
+    ///
+    /// If `schemas` is None, extracts all non-system schemas.
+    pub async fn extract(&self, schemas: Option<&[String]>) -> Result<SchemaSnapshot> {
+        let mut snapshot = SchemaSnapshot::new();
+
+        let schema_names = match schemas {
+            Some(names) => self.validate_schemas(names).await?,
+            None => self.get_all_schemas().await?,
+        };
+
+        for schema_name in &schema_names {
+            let schema_def = self.extract_schema_definition(schema_name).await?;
+            snapshot.add_schema(schema_def);
+        }
+
+        Ok(snapshot)
+    }
+
+    /// Gets all non-system schemas in the catalog.
+    async fn get_all_schemas(&self) -> Result<Vec<String>> {
+        let sql = format!(
+            "SELECT schema_name FROM information_schema.schemata \
+             WHERE catalog_name = '{}'",
+            escape_sql_literal(self.catalog)
+        );
+
+        let records = self.query(&sql).await?;
+        let mut schemas = Vec::new();
+
+        for row in records {
+            let name = extract_string(&row, 0)?;
+            if !SYSTEM_SCHEMAS.contains(&name.as_str()) {
+                schemas.push(name);
+            }
+        }
+
+        Ok(schemas)
+    }
+
+    /// Validates that all specified schemas exist.
+    async fn validate_schemas(&self, schemas: &[String]) -> Result<Vec<String>> {
+        let all_schemas = self.get_all_schemas().await?;
+        dedupe_canonicalized_schemas(schemas, &all_schemas, self.catalog)
+    }
+
+    /// Extracts schema (database) definition.
+    async fn extract_schema_definition(&self, schema: &str) -> Result<SchemaDefinition> {
+        let sql = format!(
+            "SELECT schema_name, options FROM information_schema.schemata \
+             WHERE catalog_name = '{}' AND schema_name = '{}'",
+            escape_sql_literal(self.catalog),
+            escape_sql_literal(schema)
+        );
+
+        let records = self.query(&sql).await?;
+        if records.is_empty() {
+            return SchemaNotFoundSnafu {
+                catalog: self.catalog,
+                schema,
+            }
+            .fail();
+        }
+
+        let name = extract_string(&records[0], 0)?;
+        let options = extract_optional_string(&records[0], 1)
+            .map(|opts| parse_options(&opts))
+            .unwrap_or_default();
+
+        Ok(SchemaDefinition {
+            catalog: self.catalog.to_string(),
+            name,
+            options,
+        })
+    }
+
+    /// Executes a SQL query and returns the results.
+    async fn query(&self, sql: &str) -> Result<Vec<Vec<Value>>> {
+        self.client
+            .sql_in_public(sql)
+            .await
+            .context(DatabaseSnafu)?
+            .ok_or_else(|| EmptyResultSnafu.build())
+    }
+}
+
+/// Extracts a string value from a row.
+fn extract_string(row: &[Value], index: usize) -> Result<String> {
+    match row.get(index) {
+        Some(Value::String(s)) => Ok(s.clone()),
+        Some(Value::Null) => UnexpectedValueTypeSnafu.fail(),
+        _ => UnexpectedValueTypeSnafu.fail(),
+    }
+}
+
+/// Extracts an optional string value from a row.
+fn extract_optional_string(row: &[Value], index: usize) -> Option<String> {
+    match row.get(index) {
+        Some(Value::String(s)) if !s.is_empty() => Some(s.clone()),
+        _ => None,
+    }
+}
+
+/// Parses options string into a HashMap.
+fn parse_options(options_str: &str) -> HashMap<String, String> {
+    if let Ok(map) = serde_json::from_str::<HashMap<String, String>>(options_str) {
+        return map;
+    }
+
+    let mut options = HashMap::new();
+    for line in options_str.lines() {
+        let trimmed = line.trim();
+        if trimmed.is_empty() {
+            continue;
+        }
+
+        if let Some((key, value)) = parse_quoted_option_line(trimmed) {
+            options.insert(key, value);
+            continue;
+        }
+
+        for part in trimmed.split_whitespace() {
+            if let Some((key, value)) = part.split_once('=') {
+                options.insert(key.to_string(), value.to_string());
+            }
+        }
+    }
+    options
+}
+
+fn parse_quoted_option_line(line: &str) -> Option<(String, String)> {
+    let key = line.strip_prefix('\'')?;
+    let (key, rest) = key.split_once("'='")?;
+    let value = rest.strip_suffix('\'')?;
+    Some((key.to_string(), value.to_string()))
+}
+
+fn dedupe_canonicalized_schemas(
+    requested: &[String],
+    available: &[String],
+    catalog: &str,
+) -> Result<Vec<String>> {
+    let mut canonicalized = Vec::new();
+    let mut seen = HashSet::new();
+
+    for schema in requested {
+        let Some(canonical) = available.iter().find(|s| s.eq_ignore_ascii_case(schema)) else {
+            return SchemaNotFoundSnafu { catalog, schema }.fail();
+        };
+
+        if seen.insert(canonical.to_ascii_lowercase()) {
+            canonicalized.push(canonical.clone());
+        }
+    }
+
+    Ok(canonicalized)
+}
+
+#[cfg(test)]
+mod tests {
+    use serde_json::Value;
+
+    use super::*;
+
+    #[test]
+    fn test_parse_options_json() {
+        let opts = r#"{"ttl": "30d", "custom": "value"}"#;
+        let parsed = parse_options(opts);
+        assert_eq!(parsed.get("ttl"), Some(&"30d".to_string()));
+        assert_eq!(parsed.get("custom"), Some(&"value".to_string()));
+    }
+
+    #[test]
+    fn test_parse_options_key_value() {
+        let opts = "ttl=30d custom=value";
+        let parsed = parse_options(opts);
+        assert_eq!(parsed.get("ttl"), Some(&"30d".to_string()));
+        assert_eq!(parsed.get("custom"), Some(&"value".to_string()));
+    }
+
+    #[test]
+    fn test_parse_options_schema_display_format() {
+        let opts = "'ttl'='30d'\n'custom'='value with spaces'\n";
+        let parsed = parse_options(opts);
+        assert_eq!(parsed.get("ttl"), Some(&"30d".to_string()));
+        assert_eq!(parsed.get("custom"), Some(&"value with spaces".to_string()));
+    }
+
+    #[test]
+    fn test_extract_string_rejects_null() {
+        let row = vec![Value::Null];
+        assert!(extract_string(&row, 0).is_err());
+    }
+
+    #[test]
+    fn test_dedupe_canonicalized_schemas() {
+        let available = vec!["public".to_string(), "test_db".to_string()];
+        let requested = vec![
+            "PUBLIC".to_string(),
+            "public".to_string(),
+            "Test_Db".to_string(),
+        ];
+
+        let canonicalized = dedupe_canonicalized_schemas(&requested, &available, "greptime")
+            .expect("schemas should be canonicalized");
+
+        assert_eq!(canonicalized, vec!["public", "test_db"]);
+    }
+}
diff --git a/src/cli/src/data/export_v2/manifest.rs b/src/cli/src/data/export_v2/manifest.rs
new file mode 100644
index 0000000000..918288bb51
--- /dev/null
+++ b/src/cli/src/data/export_v2/manifest.rs
@@ -0,0 +1,569 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Manifest data structures for Export/Import V2.
+
+use std::time::Duration;
+use std::{fmt, str};
+
+use chrono::{DateTime, Utc};
+use serde::{Deserialize, Serialize};
+use uuid::Uuid;
+
+use crate::data::export_v2::chunker::generate_chunks;
+use crate::data::export_v2::error::{
+    ChunkTimeWindowRequiresBoundsSnafu, Result as ExportResult, TimeParseEndBeforeStartSnafu,
+    TimeParseInvalidFormatSnafu,
+};
+
+/// Current manifest format version.
+pub const MANIFEST_VERSION: u32 = 1;
+
+/// Manifest file name within snapshot directory.
+pub const MANIFEST_FILE: &str = "manifest.json";
+
+/// Time range for data export (half-open interval: [start, end)).
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
+pub struct TimeRange {
+    /// Start time (inclusive). None means earliest available data.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub start: Option<DateTime<Utc>>,
+    /// End time (exclusive). None means current time.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub end: Option<DateTime<Utc>>,
+}
+
+impl TimeRange {
+    /// Creates a new time range with specified bounds.
+    pub fn new(start: Option<DateTime<Utc>>, end: Option<DateTime<Utc>>) -> Self {
+        Self { start, end }
+    }
+
+    /// Creates an unbounded time range (all data).
+    pub fn unbounded() -> Self {
+        Self {
+            start: None,
+            end: None,
+        }
+    }
+
+    /// Returns true if this time range is unbounded.
+    pub fn is_unbounded(&self) -> bool {
+        self.start.is_none() && self.end.is_none()
+    }
+
+    /// Returns true if both bounds are specified.
+    pub fn is_bounded(&self) -> bool {
+        self.start.is_some() && self.end.is_some()
+    }
+
+    /// Parses a time range from optional RFC3339 strings.
+    pub fn parse(start: Option<&str>, end: Option<&str>) -> ExportResult<Self> {
+        let start = start.map(parse_time).transpose()?;
+        let end = end.map(parse_time).transpose()?;
+
+        if let (Some(start), Some(end)) = (start, end)
+            && end < start
+        {
+            return TimeParseEndBeforeStartSnafu.fail();
+        }
+
+        Ok(Self::new(start, end))
+    }
+}
+
+fn parse_time(input: &str) -> ExportResult<DateTime<Utc>> {
+    DateTime::parse_from_rfc3339(input)
+        .map(|dt| dt.with_timezone(&Utc))
+        .map_err(|_| TimeParseInvalidFormatSnafu { input }.build())
+}
+
+impl Default for TimeRange {
+    fn default() -> Self {
+        Self::unbounded()
+    }
+}
+
+/// Status of a chunk during export/import.
+#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq, Default)]
+#[serde(rename_all = "snake_case")]
+pub enum ChunkStatus {
+    /// Chunk is pending export.
+    #[default]
+    Pending,
+    /// Chunk export is in progress.
+    InProgress,
+    /// Chunk export completed successfully.
+    Completed,
+    /// Chunk had no data to export.
+    Skipped,
+    /// Chunk export failed.
+    Failed,
+}
+
+/// Metadata for a single chunk of exported data.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ChunkMeta {
+    /// Chunk identifier (sequential number starting from 1).
+    pub id: u32,
+    /// Time range covered by this chunk.
+    pub time_range: TimeRange,
+    /// Export status.
+    pub status: ChunkStatus,
+    /// List of data files in this chunk (relative paths from snapshot root).
+    #[serde(default)]
+    pub files: Vec<String>,
+    /// SHA256 checksum of all files in this chunk (aggregated).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub checksum: Option<String>,
+    /// Error message if status is Failed.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub error: Option<String>,
+}
+
+impl ChunkMeta {
+    /// Creates a new pending chunk with the given id and time range.
+    pub fn new(id: u32, time_range: TimeRange) -> Self {
+        Self {
+            id,
+            time_range,
+            status: ChunkStatus::Pending,
+            files: vec![],
+            checksum: None,
+            error: None,
+        }
+    }
+
+    /// Creates a skipped chunk with the given id and time range.
+    pub fn skipped(id: u32, time_range: TimeRange) -> Self {
+        let mut chunk = Self::new(id, time_range);
+        chunk.mark_skipped();
+        chunk
+    }
+
+    /// Marks this chunk as in progress.
+    pub fn mark_in_progress(&mut self) {
+        self.status = ChunkStatus::InProgress;
+        self.error = None;
+    }
+
+    /// Marks this chunk as completed with the given files and checksum.
+    pub fn mark_completed(&mut self, files: Vec<String>, checksum: Option<String>) {
+        self.status = ChunkStatus::Completed;
+        self.files = files;
+        self.checksum = checksum;
+        self.error = None;
+    }
+
+    /// Marks this chunk as skipped because no data files were produced.
+    pub fn mark_skipped(&mut self) {
+        self.status = ChunkStatus::Skipped;
+        self.files.clear();
+        self.checksum = None;
+        self.error = None;
+    }
+
+    /// Marks this chunk as failed with the given error message.
+    pub fn mark_failed(&mut self, error: String) {
+        self.status = ChunkStatus::Failed;
+        self.error = Some(error);
+    }
+}
+
+/// Supported data formats for export.
+#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq, Default, clap::ValueEnum)]
+#[serde(rename_all = "lowercase")]
+#[value(rename_all = "lowercase")]
+pub enum DataFormat {
+    /// Apache Parquet format (default, recommended for production).
+    #[default]
+    Parquet,
+    /// CSV format (human-readable).
+    Csv,
+    /// JSON format (structured text).
+    Json,
+}
+
+impl fmt::Display for DataFormat {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match self {
+            DataFormat::Parquet => write!(f, "parquet"),
+            DataFormat::Csv => write!(f, "csv"),
+            DataFormat::Json => write!(f, "json"),
+        }
+    }
+}
+
+impl str::FromStr for DataFormat {
+    type Err = String;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s.to_lowercase().as_str() {
+            "parquet" => Ok(DataFormat::Parquet),
+            "csv" => Ok(DataFormat::Csv),
+            "json" => Ok(DataFormat::Json),
+            _ => Err(format!(
+                "invalid format '{}': expected one of parquet, csv, json",
+                s
+            )),
+        }
+    }
+}
+
+/// Snapshot manifest containing all metadata.
+///
+/// The manifest is stored as `manifest.json` in the snapshot root directory.
+/// It contains:
+/// - Snapshot identification (UUID, timestamps)
+/// - Scope (catalog, schemas, time range)
+/// - Export configuration (format, schema_only)
+/// - Chunk metadata for resume support
+/// - Integrity checksums
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct Manifest {
+    /// Manifest format version for compatibility checking.
+    pub version: u32,
+    /// Unique snapshot identifier.
+    pub snapshot_id: Uuid,
+    /// Catalog name.
+    pub catalog: String,
+    /// List of schemas included in this snapshot.
+    pub schemas: Vec<String>,
+    /// Overall time range covered by this snapshot.
+    pub time_range: TimeRange,
+    /// Whether this is a schema-only snapshot (no data).
+    pub schema_only: bool,
+    /// Data format used for export.
+    pub format: DataFormat,
+    /// Chunk metadata (empty for schema-only snapshots).
+    #[serde(default)]
+    pub chunks: Vec<ChunkMeta>,
+    /// Snapshot-level SHA256 checksum (aggregated from all chunks).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub checksum: Option<String>,
+    /// Creation timestamp.
+    pub created_at: DateTime<Utc>,
+    /// Last updated timestamp.
+    pub updated_at: DateTime<Utc>,
+}
+
+impl Manifest {
+    pub fn new_for_export(
+        catalog: String,
+        schemas: Vec<String>,
+        schema_only: bool,
+        time_range: TimeRange,
+        format: DataFormat,
+        chunk_time_window: Option<Duration>,
+    ) -> ExportResult<Self> {
+        if chunk_time_window.is_some() && !time_range.is_bounded() {
+            return ChunkTimeWindowRequiresBoundsSnafu.fail();
+        }
+
+        let mut manifest = if schema_only {
+            Self::new_schema_only(catalog, schemas)
+        } else {
+            Self::new_full(catalog, schemas, time_range, format)
+        };
+
+        if !schema_only {
+            manifest.chunks = match chunk_time_window {
+                Some(window) => generate_chunks(&manifest.time_range, window),
+                None => generate_single_chunk(&manifest.time_range),
+            };
+            manifest.touch();
+        }
+
+        Ok(manifest)
+    }
+
+    /// Creates a new manifest for schema-only export.
+    pub fn new_schema_only(catalog: String, schemas: Vec<String>) -> Self {
+        let now = Utc::now();
+        Self {
+            version: MANIFEST_VERSION,
+            snapshot_id: Uuid::new_v4(),
+            catalog,
+            schemas,
+            time_range: TimeRange::unbounded(),
+            schema_only: true,
+            format: DataFormat::Parquet,
+            chunks: vec![],
+            checksum: None,
+            created_at: now,
+            updated_at: now,
+        }
+    }
+
+    /// Creates a new manifest for full export with time range and format.
+    pub fn new_full(
+        catalog: String,
+        schemas: Vec<String>,
+        time_range: TimeRange,
+        format: DataFormat,
+    ) -> Self {
+        let now = Utc::now();
+        Self {
+            version: MANIFEST_VERSION,
+            snapshot_id: Uuid::new_v4(),
+            catalog,
+            schemas,
+            time_range,
+            schema_only: false,
+            format,
+            chunks: vec![],
+            checksum: None,
+            created_at: now,
+            updated_at: now,
+        }
+    }
+
+    /// Returns true if all chunks are completed (or if schema-only).
+    pub fn is_complete(&self) -> bool {
+        self.schema_only
+            || (!self.chunks.is_empty()
+                && self
+                    .chunks
+                    .iter()
+                    .all(|c| matches!(c.status, ChunkStatus::Completed | ChunkStatus::Skipped)))
+    }
+
+    /// Returns the number of pending chunks.
+    pub fn pending_count(&self) -> usize {
+        self.chunks
+            .iter()
+            .filter(|c| c.status == ChunkStatus::Pending)
+            .count()
+    }
+
+    /// Returns the number of in-progress chunks.
+    pub fn in_progress_count(&self) -> usize {
+        self.chunks
+            .iter()
+            .filter(|c| c.status == ChunkStatus::InProgress)
+            .count()
+    }
+
+    /// Returns the number of completed chunks.
+    pub fn completed_count(&self) -> usize {
+        self.chunks
+            .iter()
+            .filter(|c| c.status == ChunkStatus::Completed)
+            .count()
+    }
+
+    /// Returns the number of skipped chunks.
+    pub fn skipped_count(&self) -> usize {
+        self.chunks
+            .iter()
+            .filter(|c| c.status == ChunkStatus::Skipped)
+            .count()
+    }
+
+    /// Returns the number of failed chunks.
+    pub fn failed_count(&self) -> usize {
+        self.chunks
+            .iter()
+            .filter(|c| c.status == ChunkStatus::Failed)
+            .count()
+    }
+
+    /// Updates the `updated_at` timestamp to now.
+    pub fn touch(&mut self) {
+        self.updated_at = Utc::now();
+    }
+
+    /// Adds a chunk to the manifest.
+    pub fn add_chunk(&mut self, chunk: ChunkMeta) {
+        self.chunks.push(chunk);
+        self.touch();
+    }
+
+    /// Updates a chunk by id.
+    pub fn update_chunk(&mut self, id: u32, updater: impl FnOnce(&mut ChunkMeta)) {
+        if let Some(chunk) = self.chunks.iter_mut().find(|c| c.id == id) {
+            updater(chunk);
+            self.touch();
+        }
+    }
+}
+
+fn generate_single_chunk(time_range: &TimeRange) -> Vec<ChunkMeta> {
+    if let (Some(start), Some(end)) = (time_range.start, time_range.end) {
+        if start == end {
+            return vec![ChunkMeta::skipped(1, time_range.clone())];
+        }
+        if start > end {
+            return Vec::new();
+        }
+    }
+    vec![ChunkMeta::new(1, time_range.clone())]
+}
+
+#[cfg(test)]
+mod tests {
+    use std::time::Duration;
+
+    use chrono::{TimeZone, Utc};
+
+    use super::*;
+
+    #[test]
+    fn test_time_range_serialization() {
+        let range = TimeRange::unbounded();
+        let json = serde_json::to_string(&range).unwrap();
+        assert_eq!(json, "{}");
+
+        let range: TimeRange = serde_json::from_str("{}").unwrap();
+        assert!(range.is_unbounded());
+    }
+
+    #[test]
+    fn test_manifest_schema_only() {
+        let manifest =
+            Manifest::new_schema_only("greptime".to_string(), vec!["public".to_string()]);
+
+        assert_eq!(manifest.version, MANIFEST_VERSION);
+        assert!(manifest.schema_only);
+        assert!(manifest.chunks.is_empty());
+        assert!(manifest.is_complete());
+    }
+
+    #[test]
+    fn test_generate_single_chunk_zero_width_range_is_skipped() {
+        let ts = Utc.with_ymd_and_hms(2025, 1, 1, 0, 0, 0).unwrap();
+        let chunks = generate_single_chunk(&TimeRange::new(Some(ts), Some(ts)));
+
+        assert_eq!(chunks.len(), 1);
+        assert_eq!(chunks[0].status, ChunkStatus::Skipped);
+        assert_eq!(chunks[0].time_range.start, Some(ts));
+        assert_eq!(chunks[0].time_range.end, Some(ts));
+    }
+
+    #[test]
+    fn test_generate_single_chunk_invalid_range_is_empty() {
+        let start = Utc.with_ymd_and_hms(2025, 1, 1, 1, 0, 0).unwrap();
+        let end = Utc.with_ymd_and_hms(2025, 1, 1, 0, 0, 0).unwrap();
+        let chunks = generate_single_chunk(&TimeRange::new(Some(start), Some(end)));
+
+        assert!(chunks.is_empty());
+    }
+
+    #[test]
+    fn test_manifest_full() {
+        let manifest = Manifest::new_full(
+            "greptime".to_string(),
+            vec!["public".to_string()],
+            TimeRange::unbounded(),
+            DataFormat::Parquet,
+        );
+
+        assert!(!manifest.schema_only);
+        assert!(manifest.chunks.is_empty());
+        assert!(!manifest.is_complete());
+    }
+
+    #[test]
+    fn test_data_format_parsing() {
+        assert_eq!(
+            "parquet".parse::<DataFormat>().unwrap(),
+            DataFormat::Parquet
+        );
+        assert_eq!("CSV".parse::<DataFormat>().unwrap(), DataFormat::Csv);
+        assert_eq!("JSON".parse::<DataFormat>().unwrap(), DataFormat::Json);
+        assert!("invalid".parse::<DataFormat>().is_err());
+    }
+
+    #[test]
+    fn test_chunk_status_transitions() {
+        let mut chunk = ChunkMeta::new(1, TimeRange::unbounded());
+        assert_eq!(chunk.status, ChunkStatus::Pending);
+
+        chunk.mark_in_progress();
+        assert_eq!(chunk.status, ChunkStatus::InProgress);
+
+        chunk.mark_completed(
+            vec!["file1.parquet".to_string()],
+            Some("abc123".to_string()),
+        );
+        assert_eq!(chunk.status, ChunkStatus::Completed);
+        assert_eq!(chunk.files.len(), 1);
+
+        chunk.mark_skipped();
+        assert_eq!(chunk.status, ChunkStatus::Skipped);
+        assert!(chunk.files.is_empty());
+    }
+
+    #[test]
+    fn test_manifest_is_complete_when_chunks_are_completed_or_skipped() {
+        let mut manifest = Manifest::new_full(
+            "greptime".to_string(),
+            vec!["public".to_string()],
+            TimeRange::unbounded(),
+            DataFormat::Parquet,
+        );
+        manifest.add_chunk(ChunkMeta::new(1, TimeRange::unbounded()));
+        manifest.add_chunk(ChunkMeta::new(2, TimeRange::unbounded()));
+
+        manifest.update_chunk(1, |chunk| {
+            chunk.mark_completed(vec!["a.parquet".to_string()], None)
+        });
+        manifest.update_chunk(2, |chunk| chunk.mark_skipped());
+
+        assert!(manifest.is_complete());
+        assert_eq!(manifest.completed_count(), 1);
+        assert_eq!(manifest.skipped_count(), 1);
+    }
+
+    #[test]
+    fn test_manifest_chunk_time_window_none_single_chunk() {
+        let start = Utc.with_ymd_and_hms(2025, 1, 1, 0, 0, 0).unwrap();
+        let end = Utc.with_ymd_and_hms(2025, 1, 2, 0, 0, 0).unwrap();
+        let range = TimeRange::new(Some(start), Some(end));
+        let manifest = Manifest::new_for_export(
+            "greptime".to_string(),
+            vec!["public".to_string()],
+            false,
+            range.clone(),
+            DataFormat::Parquet,
+            None,
+        )
+        .unwrap();
+
+        assert_eq!(manifest.chunks.len(), 1);
+        assert_eq!(manifest.chunks[0].time_range, range);
+    }
+
+    #[test]
+    fn test_time_range_parse_requires_order() {
+        let result = TimeRange::parse(Some("2025-01-02T00:00:00Z"), Some("2025-01-01T00:00:00Z"));
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_new_for_export_with_chunk_window_requires_bounded_range() {
+        let result = Manifest::new_for_export(
+            "greptime".to_string(),
+            vec!["public".to_string()],
+            false,
+            TimeRange::new(
+                None,
+                Some(Utc.with_ymd_and_hms(2025, 1, 2, 0, 0, 0).unwrap()),
+            ),
+            DataFormat::Parquet,
+            Some(Duration::from_secs(3600)),
+        );
+        assert!(result.is_err());
+    }
+}
diff --git a/src/cli/src/data/export_v2/schema.rs b/src/cli/src/data/export_v2/schema.rs
new file mode 100644
index 0000000000..1aab6ac900
--- /dev/null
+++ b/src/cli/src/data/export_v2/schema.rs
@@ -0,0 +1,98 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Minimal schema index structures for Export/Import V2.
+//!
+//! The canonical schema representation is the per-schema DDL file under
+//! `schema/ddl/`. `schemas.json` only records which schemas exist in a snapshot.
+
+use std::collections::HashMap;
+
+use serde::{Deserialize, Serialize};
+
+/// Schema directory name within snapshot.
+pub const SCHEMA_DIR: &str = "schema";
+
+/// DDL directory name within schema directory.
+pub const DDL_DIR: &str = "ddl";
+
+/// Schema definition file name.
+pub const SCHEMAS_FILE: &str = "schemas.json";
+
+/// Schema (database) definition.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
+pub struct SchemaDefinition {
+    /// Catalog name.
+    pub catalog: String,
+    /// Schema (database) name.
+    pub name: String,
+    /// Schema options (if any).
+    #[serde(default, skip_serializing_if = "HashMap::is_empty")]
+    pub options: HashMap<String, String>,
+}
+
+/// Minimal schema index stored in a snapshot.
+#[derive(Debug, Clone, Default, Serialize, Deserialize, PartialEq, Eq)]
+pub struct SchemaSnapshot {
+    /// Schema (database) definitions.
+    pub schemas: Vec<SchemaDefinition>,
+}
+
+impl SchemaSnapshot {
+    /// Creates an empty schema snapshot.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Adds a schema definition.
+    pub fn add_schema(&mut self, schema: SchemaDefinition) {
+        self.schemas.push(schema);
+    }
+
+    /// Filters the snapshot to only include specified schemas.
+    pub fn filter_schemas(&self, schemas: &[String]) -> Self {
+        Self {
+            schemas: self
+                .schemas
+                .iter()
+                .filter(|s| schemas.contains(&s.name))
+                .cloned()
+                .collect(),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_schema_snapshot_filter() {
+        let mut snapshot = SchemaSnapshot::new();
+        snapshot.add_schema(SchemaDefinition {
+            catalog: "greptime".to_string(),
+            name: "public".to_string(),
+            options: HashMap::new(),
+        });
+        snapshot.add_schema(SchemaDefinition {
+            catalog: "greptime".to_string(),
+            name: "private".to_string(),
+            options: HashMap::new(),
+        });
+
+        let filtered = snapshot.filter_schemas(&["public".to_string()]);
+        assert_eq!(filtered.schemas.len(), 1);
+        assert_eq!(filtered.schemas[0].name, "public");
+    }
+}
diff --git a/src/cli/src/data/export_v2/tests.rs b/src/cli/src/data/export_v2/tests.rs
new file mode 100644
index 0000000000..92f45c6aec
--- /dev/null
+++ b/src/cli/src/data/export_v2/tests.rs
@@ -0,0 +1,885 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::env;
+use std::time::Duration;
+
+use clap::Parser;
+use common_error::ext::BoxedError;
+use serde_json::Value;
+use snafu::ResultExt;
+use tempfile::tempdir;
+use url::Url;
+
+use super::command::ExportCreateCommand;
+use crate::common::ObjectStoreConfig;
+use crate::data::export_v2::manifest::ChunkStatus;
+use crate::data::import_v2::ImportV2Command;
+use crate::data::snapshot_storage::{OpenDalStorage, SnapshotStorage};
+use crate::data::sql::escape_sql_identifier;
+use crate::database::DatabaseClient;
+use crate::error::{FileIoSnafu, InvalidArgumentsSnafu, OtherSnafu, Result};
+
+async fn query_count(database_client: &DatabaseClient, schema: &str, table: &str) -> Result<u64> {
+    let sql = format!("SELECT COUNT(*) FROM {}", escape_sql_identifier(table));
+    let rows = database_client.sql(&sql, schema).await?;
+    let first_row = rows.as_ref().and_then(|rows| rows.first()).ok_or_else(|| {
+        InvalidArgumentsSnafu {
+            msg: format!("empty result for query: {sql}"),
+        }
+        .build()
+    })?;
+    let first_value = first_row.first().ok_or_else(|| {
+        InvalidArgumentsSnafu {
+            msg: format!("no first column for query: {sql}"),
+        }
+        .build()
+    })?;
+    match first_value {
+        Value::Number(n) => n.as_u64().ok_or_else(|| {
+            InvalidArgumentsSnafu {
+                msg: format!("count is not u64 for query: {sql}"),
+            }
+            .build()
+        }),
+        _ => InvalidArgumentsSnafu {
+            msg: format!("unexpected count type for query: {sql}"),
+        }
+        .fail(),
+    }
+}
+
+#[tokio::test]
+#[ignore]
+async fn export_import_v2_schema_parity_e2e() -> Result<()> {
+    let addr = env::var("GREPTIME_ADDR").unwrap_or_else(|_| "127.0.0.1:4000".to_string());
+    let catalog = env::var("GREPTIME_CATALOG").unwrap_or_else(|_| "greptime".to_string());
+    let auth_basic = env::var("GREPTIME_AUTH_BASIC").ok();
+    let schema = "test_db_schema_parity";
+
+    let database_client = DatabaseClient::new(
+        addr.clone(),
+        catalog.clone(),
+        auth_basic.clone(),
+        Duration::from_secs(60),
+        None,
+        false,
+    );
+
+    database_client
+        .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+        .await?;
+    database_client
+        .sql_in_public(&format!("CREATE DATABASE {schema}"))
+        .await?;
+    database_client
+        .sql(
+            "CREATE TABLE metrics (\
+                ts TIMESTAMP TIME INDEX, \
+                host STRING PRIMARY KEY, \
+                cpu DOUBLE DEFAULT 0.0, \
+                region_name STRING \
+            ) ENGINE = mito WITH (ttl='7d', 'compaction.type'='twcs')",
+            schema,
+        )
+        .await?;
+    database_client
+        .sql(
+            "CREATE TABLE logs (\
+                ts TIMESTAMP TIME INDEX, \
+                app STRING PRIMARY KEY, \
+                msg STRING NOT NULL COMMENT 'log message' \
+            ) ENGINE = mito",
+            schema,
+        )
+        .await?;
+    database_client
+        .sql(
+            "CREATE TABLE metrics_physical (\
+                ts TIMESTAMP TIME INDEX, \
+                host STRING, \
+                region_name STRING, \
+                cpu DOUBLE DEFAULT 0.0, \
+                PRIMARY KEY (host, region_name) \
+            ) ENGINE = metric WITH (physical_metric_table='true')",
+            schema,
+        )
+        .await?;
+    database_client
+        .sql(
+            "CREATE TABLE metrics_logical (\
+                ts TIMESTAMP TIME INDEX, \
+                host STRING, \
+                region_name STRING, \
+                cpu DOUBLE DEFAULT 0.0, \
+                PRIMARY KEY (host, region_name) \
+            ) ENGINE = metric WITH (on_physical_table='metrics_physical')",
+            schema,
+        )
+        .await?;
+    database_client
+        .sql(
+            "CREATE VIEW metrics_view AS SELECT * FROM metrics WHERE cpu > 0.5",
+            schema,
+        )
+        .await?;
+
+    let src_dir = tempdir().context(FileIoSnafu)?;
+    let src_uri = Url::from_directory_path(src_dir.path())
+        .map_err(|_| {
+            InvalidArgumentsSnafu {
+                msg: "invalid temp dir path".to_string(),
+            }
+            .build()
+        })?
+        .to_string();
+
+    let mut export_args = vec![
+        "export-v2-create",
+        "--addr",
+        &addr,
+        "--to",
+        &src_uri,
+        "--catalog",
+        &catalog,
+        "--schemas",
+        schema,
+        "--schema-only",
+    ];
+    if let Some(auth) = &auth_basic {
+        export_args.push("--auth-basic");
+        export_args.push(auth);
+    }
+    let export_cmd = ExportCreateCommand::parse_from(export_args);
+    export_cmd
+        .build()
+        .await
+        .context(OtherSnafu)?
+        .do_work()
+        .await
+        .context(OtherSnafu)?;
+
+    database_client
+        .sql_in_public(&format!("DROP DATABASE {schema}"))
+        .await?;
+
+    let mut import_args = vec![
+        "import-v2",
+        "--addr",
+        &addr,
+        "--from",
+        &src_uri,
+        "--catalog",
+        &catalog,
+        "--schemas",
+        schema,
+    ];
+    if let Some(auth) = &auth_basic {
+        import_args.push("--auth-basic");
+        import_args.push(auth);
+    }
+    let import_cmd = ImportV2Command::parse_from(import_args);
+    import_cmd
+        .build()
+        .await
+        .context(OtherSnafu)?
+        .do_work()
+        .await
+        .context(OtherSnafu)?;
+
+    let dst_dir = tempdir().context(FileIoSnafu)?;
+    let dst_uri = Url::from_directory_path(dst_dir.path())
+        .map_err(|_| {
+            InvalidArgumentsSnafu {
+                msg: "invalid temp dir path".to_string(),
+            }
+            .build()
+        })?
+        .to_string();
+
+    let mut export_args = vec![
+        "export-v2-create",
+        "--addr",
+        &addr,
+        "--to",
+        &dst_uri,
+        "--catalog",
+        &catalog,
+        "--schemas",
+        schema,
+        "--schema-only",
+    ];
+    if let Some(auth) = &auth_basic {
+        export_args.push("--auth-basic");
+        export_args.push(auth);
+    }
+    let export_cmd = ExportCreateCommand::parse_from(export_args);
+    export_cmd
+        .build()
+        .await
+        .context(OtherSnafu)?
+        .do_work()
+        .await
+        .context(OtherSnafu)?;
+
+    let storage_config = ObjectStoreConfig::default();
+    let src_storage = OpenDalStorage::from_uri(&src_uri, &storage_config)
+        .map_err(BoxedError::new)
+        .context(OtherSnafu)?;
+    let dst_storage = OpenDalStorage::from_uri(&dst_uri, &storage_config)
+        .map_err(BoxedError::new)
+        .context(OtherSnafu)?;
+
+    let src_schema_snapshot = src_storage
+        .read_schema()
+        .await
+        .map_err(BoxedError::new)
+        .context(OtherSnafu)?;
+    let dst_schema_snapshot = dst_storage
+        .read_schema()
+        .await
+        .map_err(BoxedError::new)
+        .context(OtherSnafu)?;
+    assert_eq!(src_schema_snapshot, dst_schema_snapshot);
+
+    database_client
+        .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+        .await?;
+
+    Ok(())
+}
+
+#[tokio::test]
+#[ignore]
+async fn import_v2_ddl_dry_run_e2e() -> Result<()> {
+    let addr = env::var("GREPTIME_ADDR").unwrap_or_else(|_| "127.0.0.1:4000".to_string());
+    let catalog = env::var("GREPTIME_CATALOG").unwrap_or_else(|_| "greptime".to_string());
+    let auth_basic = env::var("GREPTIME_AUTH_BASIC").ok();
+    let schema = "test_db_ddl_dry_run";
+
+    let database_client = DatabaseClient::new(
+        addr.clone(),
+        catalog.clone(),
+        auth_basic.clone(),
+        Duration::from_secs(60),
+        None,
+        false,
+    );
+
+    database_client
+        .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+        .await?;
+    database_client
+        .sql_in_public(&format!("CREATE DATABASE {schema}"))
+        .await?;
+    database_client
+        .sql(
+            "CREATE TABLE metrics (\
+                ts TIMESTAMP TIME INDEX, \
+                host STRING PRIMARY KEY, \
+                cpu DOUBLE DEFAULT 0.0, \
+                region_name STRING \
+            ) ENGINE = mito WITH (ttl='7d', 'compaction.type'='twcs')",
+            schema,
+        )
+        .await?;
+    database_client
+        .sql(
+            "CREATE TABLE logs (\
+                ts TIMESTAMP TIME INDEX, \
+                app STRING PRIMARY KEY, \
+                msg STRING NOT NULL COMMENT 'log message' \
+            ) ENGINE = mito",
+            schema,
+        )
+        .await?;
+
+    let src_dir = tempdir().context(FileIoSnafu)?;
+    let src_uri = Url::from_directory_path(src_dir.path())
+        .map_err(|_| {
+            InvalidArgumentsSnafu {
+                msg: "invalid temp dir path".to_string(),
+            }
+            .build()
+        })?
+        .to_string();
+
+    let mut export_args = vec![
+        "export-v2-create",
+        "--addr",
+        &addr,
+        "--to",
+        &src_uri,
+        "--catalog",
+        &catalog,
+        "--schemas",
+        schema,
+        "--schema-only",
+    ];
+    if let Some(auth) = &auth_basic {
+        export_args.push("--auth-basic");
+        export_args.push(auth);
+    }
+    let export_cmd = ExportCreateCommand::parse_from(export_args);
+    export_cmd
+        .build()
+        .await
+        .context(OtherSnafu)?
+        .do_work()
+        .await
+        .context(OtherSnafu)?;
+
+    let mut import_args = vec![
+        "import-v2",
+        "--addr",
+        &addr,
+        "--from",
+        &src_uri,
+        "--catalog",
+        &catalog,
+        "--schemas",
+        schema,
+        "--dry-run",
+    ];
+    if let Some(auth) = &auth_basic {
+        import_args.push("--auth-basic");
+        import_args.push(auth);
+    }
+    let import_cmd = ImportV2Command::parse_from(import_args);
+    import_cmd
+        .build()
+        .await
+        .context(OtherSnafu)?
+        .do_work()
+        .await
+        .context(OtherSnafu)?;
+
+    database_client
+        .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+        .await?;
+
+    Ok(())
+}
+
+#[tokio::test]
+#[ignore]
+async fn export_import_v2_data_roundtrip_e2e() -> Result<()> {
+    let addr = env::var("GREPTIME_ADDR").unwrap_or_else(|_| "127.0.0.1:4000".to_string());
+    let catalog = env::var("GREPTIME_CATALOG").unwrap_or_else(|_| "greptime".to_string());
+    let auth_basic = env::var("GREPTIME_AUTH_BASIC").ok();
+    let schema = "test_db_data_roundtrip";
+
+    let database_client = DatabaseClient::new(
+        addr.clone(),
+        catalog.clone(),
+        auth_basic.clone(),
+        Duration::from_secs(60),
+        None,
+        false,
+    );
+
+    database_client
+        .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+        .await?;
+    database_client
+        .sql_in_public(&format!("CREATE DATABASE {schema}"))
+        .await?;
+    database_client
+        .sql(
+            "CREATE TABLE metrics (\
+                ts TIMESTAMP TIME INDEX, \
+                host STRING PRIMARY KEY, \
+                cpu DOUBLE \
+            ) ENGINE=mito",
+            schema,
+        )
+        .await?;
+    database_client
+        .sql(
+            "INSERT INTO metrics (ts, host, cpu) VALUES \
+             ('2025-01-01T00:00:00Z', 'h1', 1.0), \
+             ('2025-01-01T01:00:00Z', 'h2', 2.0)",
+            schema,
+        )
+        .await?;
+
+    let expected_rows = query_count(&database_client, schema, "metrics").await?;
+
+    let src_dir = tempdir().context(FileIoSnafu)?;
+    let src_uri = Url::from_directory_path(src_dir.path())
+        .map_err(|_| {
+            InvalidArgumentsSnafu {
+                msg: "invalid temp dir path".to_string(),
+            }
+            .build()
+        })?
+        .to_string();
+
+    let mut export_args = vec![
+        "export-v2-create",
+        "--addr",
+        &addr,
+        "--to",
+        &src_uri,
+        "--catalog",
+        &catalog,
+        "--schemas",
+        schema,
+    ];
+    if let Some(auth) = &auth_basic {
+        export_args.push("--auth-basic");
+        export_args.push(auth);
+    }
+    let export_cmd = ExportCreateCommand::parse_from(export_args);
+    export_cmd
+        .build()
+        .await
+        .context(OtherSnafu)?
+        .do_work()
+        .await
+        .context(OtherSnafu)?;
+
+    database_client
+        .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+        .await?;
+
+    let mut import_args = vec![
+        "import-v2",
+        "--addr",
+        &addr,
+        "--from",
+        &src_uri,
+        "--catalog",
+        &catalog,
+        "--schemas",
+        schema,
+    ];
+    if let Some(auth) = &auth_basic {
+        import_args.push("--auth-basic");
+        import_args.push(auth);
+    }
+    let import_cmd = ImportV2Command::parse_from(import_args);
+    import_cmd
+        .build()
+        .await
+        .context(OtherSnafu)?
+        .do_work()
+        .await
+        .context(OtherSnafu)?;
+
+    let actual_rows = query_count(&database_client, schema, "metrics").await?;
+    assert_eq!(actual_rows, expected_rows);
+
+    database_client
+        .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+        .await?;
+
+    Ok(())
+}
+
+#[tokio::test]
+#[ignore]
+async fn import_v2_fails_on_incomplete_snapshot_e2e() -> Result<()> {
+    let addr = env::var("GREPTIME_ADDR").unwrap_or_else(|_| "127.0.0.1:4000".to_string());
+    let catalog = env::var("GREPTIME_CATALOG").unwrap_or_else(|_| "greptime".to_string());
+    let auth_basic = env::var("GREPTIME_AUTH_BASIC").ok();
+    let schema = "test_db_incomplete_snapshot";
+
+    let database_client = DatabaseClient::new(
+        addr.clone(),
+        catalog.clone(),
+        auth_basic.clone(),
+        Duration::from_secs(60),
+        None,
+        false,
+    );
+
+    database_client
+        .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+        .await?;
+    database_client
+        .sql_in_public(&format!("CREATE DATABASE {schema}"))
+        .await?;
+    database_client
+        .sql(
+            "CREATE TABLE metrics (\
+                ts TIMESTAMP TIME INDEX, \
+                host STRING PRIMARY KEY, \
+                cpu DOUBLE \
+            ) ENGINE=mito",
+            schema,
+        )
+        .await?;
+    database_client
+        .sql(
+            "INSERT INTO metrics (ts, host, cpu) VALUES ('2025-01-01T00:00:00Z', 'h1', 1.0)",
+            schema,
+        )
+        .await?;
+
+    let src_dir = tempdir().context(FileIoSnafu)?;
+    let src_uri = Url::from_directory_path(src_dir.path())
+        .map_err(|_| {
+            InvalidArgumentsSnafu {
+                msg: "invalid temp dir path".to_string(),
+            }
+            .build()
+        })?
+        .to_string();
+
+    let mut export_args = vec![
+        "export-v2-create",
+        "--addr",
+        &addr,
+        "--to",
+        &src_uri,
+        "--catalog",
+        &catalog,
+        "--schemas",
+        schema,
+    ];
+    if let Some(auth) = &auth_basic {
+        export_args.push("--auth-basic");
+        export_args.push(auth);
+    }
+    let export_cmd = ExportCreateCommand::parse_from(export_args);
+    export_cmd
+        .build()
+        .await
+        .context(OtherSnafu)?
+        .do_work()
+        .await
+        .context(OtherSnafu)?;
+
+    let storage_config = ObjectStoreConfig::default();
+    let storage = OpenDalStorage::from_uri(&src_uri, &storage_config)
+        .map_err(BoxedError::new)
+        .context(OtherSnafu)?;
+    let mut manifest = storage
+        .read_manifest()
+        .await
+        .map_err(BoxedError::new)
+        .context(OtherSnafu)?;
+    if let Some(first_chunk) = manifest.chunks.first_mut() {
+        first_chunk.status = ChunkStatus::Failed;
+    }
+    storage
+        .write_manifest(&manifest)
+        .await
+        .map_err(BoxedError::new)
+        .context(OtherSnafu)?;
+
+    database_client
+        .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+        .await?;
+
+    let mut import_args = vec![
+        "import-v2",
+        "--addr",
+        &addr,
+        "--from",
+        &src_uri,
+        "--catalog",
+        &catalog,
+        "--schemas",
+        schema,
+    ];
+    if let Some(auth) = &auth_basic {
+        import_args.push("--auth-basic");
+        import_args.push(auth);
+    }
+    let import_cmd = ImportV2Command::parse_from(import_args);
+    let err = import_cmd
+        .build()
+        .await
+        .context(OtherSnafu)?
+        .do_work()
+        .await
+        .expect_err("import should fail on incomplete snapshot");
+    assert!(err.to_string().contains("Incomplete snapshot"));
+
+    Ok(())
+}
+
+#[tokio::test]
+#[ignore]
+async fn import_v2_schema_filter_data_e2e() -> Result<()> {
+    let addr = env::var("GREPTIME_ADDR").unwrap_or_else(|_| "127.0.0.1:4000".to_string());
+    let catalog = env::var("GREPTIME_CATALOG").unwrap_or_else(|_| "greptime".to_string());
+    let auth_basic = env::var("GREPTIME_AUTH_BASIC").ok();
+    let schema_a = "test_db_filter_a";
+    let schema_b = "test_db_filter_b";
+
+    let database_client = DatabaseClient::new(
+        addr.clone(),
+        catalog.clone(),
+        auth_basic.clone(),
+        Duration::from_secs(60),
+        None,
+        false,
+    );
+
+    for schema in [schema_a, schema_b] {
+        database_client
+            .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+            .await?;
+        database_client
+            .sql_in_public(&format!("CREATE DATABASE {schema}"))
+            .await?;
+        database_client
+            .sql(
+                "CREATE TABLE metrics (\
+                    ts TIMESTAMP TIME INDEX, \
+                    host STRING PRIMARY KEY, \
+                    cpu DOUBLE \
+                ) ENGINE=mito",
+                schema,
+            )
+            .await?;
+    }
+    database_client
+        .sql(
+            "INSERT INTO metrics (ts, host, cpu) VALUES ('2025-01-01T00:00:00Z', 'a1', 1.0)",
+            schema_a,
+        )
+        .await?;
+    database_client
+        .sql(
+            "INSERT INTO metrics (ts, host, cpu) VALUES ('2025-01-01T00:00:00Z', 'b1', 2.0)",
+            schema_b,
+        )
+        .await?;
+
+    let expected_rows_a = query_count(&database_client, schema_a, "metrics").await?;
+
+    let src_dir = tempdir().context(FileIoSnafu)?;
+    let src_uri = Url::from_directory_path(src_dir.path())
+        .map_err(|_| {
+            InvalidArgumentsSnafu {
+                msg: "invalid temp dir path".to_string(),
+            }
+            .build()
+        })?
+        .to_string();
+
+    let mut export_args = vec![
+        "export-v2-create",
+        "--addr",
+        &addr,
+        "--to",
+        &src_uri,
+        "--catalog",
+        &catalog,
+        "--schemas",
+        schema_a,
+        "--schemas",
+        schema_b,
+    ];
+    if let Some(auth) = &auth_basic {
+        export_args.push("--auth-basic");
+        export_args.push(auth);
+    }
+    let export_cmd = ExportCreateCommand::parse_from(export_args);
+    export_cmd
+        .build()
+        .await
+        .context(OtherSnafu)?
+        .do_work()
+        .await
+        .context(OtherSnafu)?;
+
+    for schema in [schema_a, schema_b] {
+        database_client
+            .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+            .await?;
+    }
+
+    let mut import_args = vec![
+        "import-v2",
+        "--addr",
+        &addr,
+        "--from",
+        &src_uri,
+        "--catalog",
+        &catalog,
+        "--schemas",
+        schema_a,
+    ];
+    if let Some(auth) = &auth_basic {
+        import_args.push("--auth-basic");
+        import_args.push(auth);
+    }
+    let import_cmd = ImportV2Command::parse_from(import_args);
+    import_cmd
+        .build()
+        .await
+        .context(OtherSnafu)?
+        .do_work()
+        .await
+        .context(OtherSnafu)?;
+
+    let actual_rows_a = query_count(&database_client, schema_a, "metrics").await?;
+    assert_eq!(actual_rows_a, expected_rows_a);
+
+    let schema_b_query = database_client
+        .sql("SELECT COUNT(*) FROM metrics", schema_b)
+        .await;
+    assert!(schema_b_query.is_err(), "schema_b should not be imported");
+
+    for schema in [schema_a, schema_b] {
+        database_client
+            .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+            .await?;
+    }
+
+    Ok(())
+}
+
+#[tokio::test]
+#[ignore]
+async fn export_import_v2_skipped_chunk_e2e() -> Result<()> {
+    let addr = env::var("GREPTIME_ADDR").unwrap_or_else(|_| "127.0.0.1:4000".to_string());
+    let catalog = env::var("GREPTIME_CATALOG").unwrap_or_else(|_| "greptime".to_string());
+    let auth_basic = env::var("GREPTIME_AUTH_BASIC").ok();
+    let schema = "test_db_skipped_chunk";
+
+    let database_client = DatabaseClient::new(
+        addr.clone(),
+        catalog.clone(),
+        auth_basic.clone(),
+        Duration::from_secs(60),
+        None,
+        false,
+    );
+
+    database_client
+        .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+        .await?;
+    database_client
+        .sql_in_public(&format!("CREATE DATABASE {schema}"))
+        .await?;
+    database_client
+        .sql(
+            "CREATE TABLE metrics (\
+                ts TIMESTAMP TIME INDEX, \
+                host STRING PRIMARY KEY, \
+                cpu DOUBLE \
+            ) ENGINE=mito",
+            schema,
+        )
+        .await?;
+    database_client
+        .sql(
+            "INSERT INTO metrics (ts, host, cpu) VALUES \
+             ('2025-01-01T00:00:00Z', 'h1', 1.0), \
+             ('2025-01-01T01:00:00Z', 'h2', 2.0)",
+            schema,
+        )
+        .await?;
+
+    let src_dir = tempdir().context(FileIoSnafu)?;
+    let src_uri = Url::from_directory_path(src_dir.path())
+        .map_err(|_| {
+            InvalidArgumentsSnafu {
+                msg: "invalid temp dir path".to_string(),
+            }
+            .build()
+        })?
+        .to_string();
+
+    let mut export_args = vec![
+        "export-v2-create",
+        "--addr",
+        &addr,
+        "--to",
+        &src_uri,
+        "--catalog",
+        &catalog,
+        "--schemas",
+        schema,
+        "--start-time",
+        "2025-01-01T00:00:00Z",
+        "--end-time",
+        "2025-01-01T02:00:00Z",
+        "--chunk-time-window",
+        "1h",
+    ];
+    if let Some(auth) = &auth_basic {
+        export_args.push("--auth-basic");
+        export_args.push(auth);
+    }
+    let export_cmd = ExportCreateCommand::parse_from(export_args);
+    export_cmd
+        .build()
+        .await
+        .context(OtherSnafu)?
+        .do_work()
+        .await
+        .context(OtherSnafu)?;
+
+    let storage_config = ObjectStoreConfig::default();
+    let storage = OpenDalStorage::from_uri(&src_uri, &storage_config)
+        .map_err(BoxedError::new)
+        .context(OtherSnafu)?;
+    let mut manifest = storage
+        .read_manifest()
+        .await
+        .map_err(BoxedError::new)
+        .context(OtherSnafu)?;
+    assert_eq!(manifest.chunks.len(), 2);
+    manifest.chunks[0].status = ChunkStatus::Skipped;
+    manifest.chunks[0].files.clear();
+    storage
+        .write_manifest(&manifest)
+        .await
+        .map_err(BoxedError::new)
+        .context(OtherSnafu)?;
+
+    database_client
+        .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+        .await?;
+
+    let mut import_args = vec![
+        "import-v2",
+        "--addr",
+        &addr,
+        "--from",
+        &src_uri,
+        "--catalog",
+        &catalog,
+        "--schemas",
+        schema,
+    ];
+    if let Some(auth) = &auth_basic {
+        import_args.push("--auth-basic");
+        import_args.push(auth);
+    }
+    let import_cmd = ImportV2Command::parse_from(import_args);
+    import_cmd
+        .build()
+        .await
+        .context(OtherSnafu)?
+        .do_work()
+        .await
+        .context(OtherSnafu)?;
+
+    let actual_rows = query_count(&database_client, schema, "metrics").await?;
+    assert_eq!(actual_rows, 1);
+
+    database_client
+        .sql_in_public(&format!("DROP DATABASE IF EXISTS {schema}"))
+        .await?;
+
+    Ok(())
+}
diff --git a/src/cli/src/data/import.rs b/src/cli/src/data/import.rs
index ffe8b62c7e..f5c234f1a7 100644
--- a/src/cli/src/data/import.rs
+++ b/src/cli/src/data/import.rs
@@ -81,13 +81,16 @@ pub struct ImportCommand {
     #[clap(long, value_parser = humantime::parse_duration)]
     timeout: Option<Duration>,
 
-    /// The proxy server address to connect, if set, will override the system proxy.
+    /// The proxy server address to connect.
     ///
-    /// The default behavior will use the system proxy if neither `proxy` nor `no_proxy` is set.
+    /// If set, it overrides the system proxy unless `--no-proxy` is specified.
+    /// If neither `--proxy` nor `--no-proxy` is set, system proxy (env) may be used.
     #[clap(long)]
     proxy: Option<String>,
 
-    /// Disable proxy server, if set, will not use any proxy.
+    /// Disable all proxy usage (ignores `--proxy` and system proxy).
+    ///
+    /// When set and `--proxy` is not provided, this explicitly disables system proxy.
     #[clap(long, default_value = "false")]
     no_proxy: bool,
 }
@@ -104,6 +107,7 @@ impl ImportCommand {
             // Treats `None` as `0s` to disable server-side default timeout.
             self.timeout.unwrap_or_default(),
             proxy,
+            self.no_proxy,
         );
 
         Ok(Box::new(Import {
@@ -314,6 +318,7 @@ mod tests {
                 None,
                 Duration::from_secs(0),
                 None,
+                false,
             ),
             input_dir: input_dir.to_string(),
             parallelism: 1,
diff --git a/src/cli/src/data/import_v2.rs b/src/cli/src/data/import_v2.rs
new file mode 100644
index 0000000000..b197f59b3b
--- /dev/null
+++ b/src/cli/src/data/import_v2.rs
@@ -0,0 +1,43 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Import V2 module.
+//!
+//! This module provides the V2 implementation of database import functionality,
+//! featuring:
+//! - DDL-based schema import
+//! - Dry-run mode for verification
+//!
+//! # Example
+//!
+//! ```bash
+//! # Dry-run import (verify without executing)
+//! greptime cli data import-v2 \
+//!   --addr 127.0.0.1:4000 \
+//!   --from file:///tmp/snapshot \
+//!   --dry-run
+//!
+//! # Actual import
+//! greptime cli data import-v2 \
+//!   --addr 127.0.0.1:4000 \
+//!   --from s3://bucket/snapshots/prod-20250101
+//! ```
+
+mod command;
+pub(crate) mod coordinator;
+pub mod error;
+pub mod executor;
+pub(crate) mod state;
+
+pub use command::ImportV2Command;
diff --git a/src/cli/src/data/import_v2/command.rs b/src/cli/src/data/import_v2/command.rs
new file mode 100644
index 0000000000..9ee60252a0
--- /dev/null
+++ b/src/cli/src/data/import_v2/command.rs
@@ -0,0 +1,974 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Import V2 CLI command.
+
+use std::collections::HashSet;
+use std::time::Duration;
+
+use async_trait::async_trait;
+use clap::Parser;
+use common_error::ext::BoxedError;
+use common_telemetry::info;
+use snafu::{OptionExt, ResultExt};
+
+use crate::Tool;
+use crate::common::ObjectStoreConfig;
+use crate::data::export_v2::data::{build_copy_source, execute_copy_database_from};
+use crate::data::export_v2::manifest::{ChunkMeta, ChunkStatus, DataFormat, MANIFEST_VERSION};
+use crate::data::import_v2::coordinator::{
+    ImportResumeConfig, ImportTaskExecutor, build_import_tasks, chunk_has_schema_files,
+    import_with_resume_session, prepare_import_resume,
+};
+use crate::data::import_v2::error::{
+    ChunkImportFailedSnafu, EmptyChunkManifestSnafu, ImportStatePathUnavailableSnafu,
+    IncompleteSnapshotSnafu, ManifestVersionMismatchSnafu, MissingChunkDataSnafu, Result,
+    SchemaNotInSnapshotSnafu, SnapshotStorageSnafu,
+};
+use crate::data::import_v2::executor::{DdlExecutor, DdlStatement};
+use crate::data::import_v2::state::{ImportTaskKey, default_state_path};
+use crate::data::path::{data_dir_for_schema_chunk, ddl_path_for_schema};
+use crate::data::snapshot_storage::{OpenDalStorage, SnapshotStorage, validate_uri};
+use crate::database::{DatabaseClient, parse_proxy_opts};
+
+/// Import from a snapshot.
+#[derive(Debug, Parser)]
+pub struct ImportV2Command {
+    /// Server address to connect (e.g., 127.0.0.1:4000).
+    #[clap(long)]
+    addr: String,
+
+    /// Source snapshot location (e.g., s3://bucket/path, file:///tmp/backup).
+    #[clap(long)]
+    from: String,
+
+    /// Target catalog name.
+    #[clap(long, default_value = "greptime")]
+    catalog: String,
+
+    /// Schema list to import (default: all in snapshot).
+    /// Can be specified multiple times or comma-separated.
+    #[clap(long, value_delimiter = ',')]
+    schemas: Vec<String>,
+
+    /// Verify without importing (dry-run).
+    #[clap(long)]
+    dry_run: bool,
+
+    /// Basic authentication (user:password).
+    #[clap(long)]
+    auth_basic: Option<String>,
+
+    /// Request timeout.
+    #[clap(long, value_parser = humantime::parse_duration)]
+    timeout: Option<Duration>,
+
+    /// Proxy server address.
+    ///
+    /// If set, it overrides the system proxy unless `--no-proxy` is specified.
+    /// If neither `--proxy` nor `--no-proxy` is set, system proxy (env) may be used.
+    #[clap(long)]
+    proxy: Option<String>,
+
+    /// Disable all proxy usage (ignores `--proxy` and system proxy).
+    ///
+    /// When set and `--proxy` is not provided, this explicitly disables system proxy.
+    #[clap(long)]
+    no_proxy: bool,
+
+    /// Object store configuration for remote storage backends.
+    #[clap(flatten)]
+    storage: ObjectStoreConfig,
+}
+
+impl ImportV2Command {
+    pub async fn build(&self) -> std::result::Result<Box<dyn Tool>, BoxedError> {
+        // Validate URI format
+        validate_uri(&self.from)
+            .context(SnapshotStorageSnafu)
+            .map_err(BoxedError::new)?;
+
+        // Parse schemas (empty vec means all schemas)
+        let schemas = if self.schemas.is_empty() {
+            None
+        } else {
+            Some(self.schemas.clone())
+        };
+
+        // Build storage
+        let storage = OpenDalStorage::from_uri(&self.from, &self.storage)
+            .context(SnapshotStorageSnafu)
+            .map_err(BoxedError::new)?;
+
+        // Build database client
+        let proxy = parse_proxy_opts(self.proxy.clone(), self.no_proxy)?;
+        let database_client = DatabaseClient::new(
+            self.addr.clone(),
+            self.catalog.clone(),
+            self.auth_basic.clone(),
+            self.timeout.unwrap_or(Duration::from_secs(60)),
+            proxy,
+            self.no_proxy,
+        );
+
+        Ok(Box::new(Import {
+            catalog: self.catalog.clone(),
+            schemas,
+            dry_run: self.dry_run,
+            snapshot_uri: self.from.clone(),
+            storage_config: self.storage.clone(),
+            storage: Box::new(storage),
+            database_client,
+        }))
+    }
+}
+
+/// Import tool implementation.
+pub struct Import {
+    catalog: String,
+    schemas: Option<Vec<String>>,
+    dry_run: bool,
+    snapshot_uri: String,
+    storage_config: ObjectStoreConfig,
+    storage: Box<dyn SnapshotStorage>,
+    database_client: DatabaseClient,
+}
+
+#[async_trait]
+impl Tool for Import {
+    async fn do_work(&self) -> std::result::Result<(), BoxedError> {
+        self.run().await.map_err(BoxedError::new)
+    }
+}
+
+impl Import {
+    async fn run(&self) -> Result<()> {
+        // 1. Read manifest
+        let manifest = self
+            .storage
+            .read_manifest()
+            .await
+            .context(SnapshotStorageSnafu)?;
+
+        info!(
+            "Loading snapshot: {} (version: {}, schema_only: {})",
+            manifest.snapshot_id, manifest.version, manifest.schema_only
+        );
+
+        // Check version compatibility
+        if manifest.version != MANIFEST_VERSION {
+            return ManifestVersionMismatchSnafu {
+                expected: MANIFEST_VERSION,
+                found: manifest.version,
+            }
+            .fail();
+        }
+
+        info!("Snapshot contains {} schema(s)", manifest.schemas.len());
+
+        // 2. Determine schemas to import
+        let schemas_to_import = match &self.schemas {
+            Some(filter) => canonicalize_schema_filter(filter, &manifest.schemas)?,
+            None => manifest.schemas.clone(),
+        };
+
+        info!("Importing schemas: {:?}", schemas_to_import);
+
+        // 3. Read DDL statements
+        let ddl_statements = self.read_ddl_statements(&schemas_to_import).await?;
+
+        info!("Generated {} DDL statements", ddl_statements.len());
+
+        let data_tasks = if !manifest.schema_only && !manifest.chunks.is_empty() {
+            validate_data_snapshot(self.storage.as_ref(), &manifest.chunks, &schemas_to_import)
+                .await?;
+            build_import_tasks(&manifest.chunks, &schemas_to_import)
+        } else {
+            Vec::new()
+        };
+
+        // 4. Dry-run mode: print DDL and exit
+        if self.dry_run {
+            info!("Dry-run mode - DDL statements to execute:");
+            println!();
+            for (i, stmt) in ddl_statements.iter().enumerate() {
+                println!("-- Statement {}", i + 1);
+                println!("{};", stmt.sql);
+                println!();
+            }
+            if !manifest.schema_only && !manifest.chunks.is_empty() {
+                for line in format_data_import_plan(&manifest.chunks, &schemas_to_import) {
+                    println!("{line}");
+                }
+                println!();
+            }
+            return Ok(());
+        }
+
+        let mut resume_session = if !data_tasks.is_empty() {
+            let state_path = default_state_path(
+                &manifest.snapshot_id.to_string(),
+                self.database_client.addr(),
+                &self.catalog,
+                &schemas_to_import,
+            )
+            .context(ImportStatePathUnavailableSnafu {
+                snapshot_id: manifest.snapshot_id.to_string(),
+            })?;
+            Some(
+                prepare_import_resume(ImportResumeConfig {
+                    snapshot_id: manifest.snapshot_id.to_string(),
+                    target_addr: self.database_client.addr().to_string(),
+                    catalog: self.catalog.clone(),
+                    schemas: schemas_to_import.clone(),
+                    state_path,
+                    tasks: data_tasks,
+                })
+                .await?,
+            )
+        } else {
+            None
+        };
+
+        let skip_ddl = resume_session
+            .as_ref()
+            .map(|session| session.should_skip_ddl())
+            .unwrap_or(false);
+
+        // 5. Execute DDL unless a previous run already completed it.
+        let ddl_executed = if skip_ddl {
+            info!(
+                "Existing import state has DDL marked completed; skipping DDL execution and resuming data import"
+            );
+            false
+        } else {
+            let executor = DdlExecutor::new(&self.database_client);
+            executor.execute_strict(&ddl_statements).await?;
+            if let Some(session) = resume_session.as_mut() {
+                session.mark_ddl_completed().await?;
+            }
+            true
+        };
+
+        if let Some(resume_session) = resume_session {
+            let executor = CopyDatabaseImportTaskExecutor {
+                import: self,
+                format: manifest.format,
+            };
+            import_with_resume_session(resume_session, &executor).await?;
+        }
+
+        if ddl_executed {
+            info!(
+                "Import completed: {} DDL statements executed",
+                ddl_statements.len()
+            );
+        } else {
+            info!("Import completed: DDL execution skipped");
+        }
+
+        Ok(())
+    }
+
+    async fn read_ddl_statements(&self, schemas: &[String]) -> Result<Vec<DdlStatement>> {
+        let mut statements = Vec::new();
+        for schema in schemas {
+            let path = ddl_path_for_schema(schema);
+            let content = self
+                .storage
+                .read_text(&path)
+                .await
+                .context(SnapshotStorageSnafu)?;
+            statements.extend(
+                parse_ddl_statements(&content)
+                    .into_iter()
+                    .map(|sql| ddl_statement_for_schema(schema, sql)),
+            );
+        }
+
+        Ok(statements)
+    }
+}
+
+struct CopyDatabaseImportTaskExecutor<'a> {
+    import: &'a Import,
+    format: DataFormat,
+}
+
+#[async_trait]
+impl ImportTaskExecutor for CopyDatabaseImportTaskExecutor<'_> {
+    async fn import_task(&self, task: &ImportTaskKey) -> Result<()> {
+        let source = build_copy_source(
+            &self.import.snapshot_uri,
+            &self.import.storage_config,
+            &task.schema,
+            task.chunk_id,
+        )
+        .context(ChunkImportFailedSnafu {
+            chunk_id: task.chunk_id,
+            schema: task.schema.clone(),
+        })?;
+
+        execute_copy_database_from(
+            &self.import.database_client,
+            &self.import.catalog,
+            &task.schema,
+            &source,
+            self.format,
+        )
+        .await
+        .context(ChunkImportFailedSnafu {
+            chunk_id: task.chunk_id,
+            schema: task.schema.clone(),
+        })
+    }
+}
+
+fn parse_ddl_statements(content: &str) -> Vec<String> {
+    let mut statements = Vec::new();
+    let mut current = String::new();
+    let mut chars = content.chars().peekable();
+    let mut in_single_quote = false;
+    let mut in_double_quote = false;
+    let mut in_line_comment = false;
+    let mut in_block_comment = false;
+
+    while let Some(ch) = chars.next() {
+        if in_line_comment {
+            if ch == '\n' {
+                in_line_comment = false;
+                current.push('\n');
+            }
+            continue;
+        }
+
+        if in_block_comment {
+            if ch == '*' && chars.peek() == Some(&'/') {
+                chars.next();
+                in_block_comment = false;
+            }
+            continue;
+        }
+
+        if in_single_quote {
+            current.push(ch);
+            if ch == '\'' {
+                if chars.peek() == Some(&'\'') {
+                    current.push(chars.next().expect("peeked quote must exist"));
+                } else {
+                    in_single_quote = false;
+                }
+            }
+            continue;
+        }
+
+        if in_double_quote {
+            current.push(ch);
+            if ch == '"' {
+                if chars.peek() == Some(&'"') {
+                    current.push(chars.next().expect("peeked quote must exist"));
+                } else {
+                    in_double_quote = false;
+                }
+            }
+            continue;
+        }
+
+        match ch {
+            '-' if chars.peek() == Some(&'-') => {
+                chars.next();
+                in_line_comment = true;
+            }
+            '/' if chars.peek() == Some(&'*') => {
+                chars.next();
+                in_block_comment = true;
+            }
+            '\'' => {
+                in_single_quote = true;
+                current.push(ch);
+            }
+            '"' => {
+                in_double_quote = true;
+                current.push(ch);
+            }
+            ';' => {
+                let statement = current.trim();
+                if !statement.is_empty() {
+                    statements.push(statement.to_string());
+                }
+                current.clear();
+            }
+            _ => current.push(ch),
+        }
+    }
+
+    let statement = current.trim();
+    if !statement.is_empty() {
+        statements.push(statement.to_string());
+    }
+
+    statements
+}
+
+fn ddl_statement_for_schema(schema: &str, sql: String) -> DdlStatement {
+    if is_schema_scoped_statement(&sql) {
+        DdlStatement::with_execution_schema(sql, schema.to_string())
+    } else {
+        DdlStatement::new(sql)
+    }
+}
+
+fn is_schema_scoped_statement(sql: &str) -> bool {
+    let trimmed = sql.trim_start();
+    if !starts_with_keyword(trimmed, "CREATE") {
+        return false;
+    }
+
+    let Some(rest) = trimmed.get("CREATE".len()..) else {
+        return false;
+    };
+    let mut rest = rest.trim_start();
+    if starts_with_keyword(rest, "OR") {
+        let Some(next) = rest.get("OR".len()..) else {
+            return false;
+        };
+        rest = next.trim_start();
+        if !starts_with_keyword(rest, "REPLACE") {
+            return false;
+        }
+        let Some(next) = rest.get("REPLACE".len()..) else {
+            return false;
+        };
+        rest = next.trim_start();
+    }
+
+    if starts_with_keyword(rest, "EXTERNAL") {
+        let Some(next) = rest.get("EXTERNAL".len()..) else {
+            return false;
+        };
+        rest = next.trim_start();
+    }
+
+    starts_with_keyword(rest, "TABLE") || starts_with_keyword(rest, "VIEW")
+}
+
+fn starts_with_keyword(input: &str, keyword: &str) -> bool {
+    input
+        .get(0..keyword.len())
+        .map(|s| s.eq_ignore_ascii_case(keyword))
+        .unwrap_or(false)
+        && input
+            .as_bytes()
+            .get(keyword.len())
+            .map(|b| !b.is_ascii_alphanumeric() && *b != b'_')
+            .unwrap_or(true)
+}
+
+fn canonicalize_schema_filter(
+    filter: &[String],
+    manifest_schemas: &[String],
+) -> Result<Vec<String>> {
+    let mut canonicalized = Vec::new();
+    let mut seen = HashSet::new();
+
+    for schema in filter {
+        let canonical = manifest_schemas
+            .iter()
+            .find(|candidate| candidate.eq_ignore_ascii_case(schema))
+            .cloned()
+            .ok_or_else(|| {
+                SchemaNotInSnapshotSnafu {
+                    schema: schema.clone(),
+                }
+                .build()
+            })?;
+
+        if seen.insert(canonical.to_ascii_lowercase()) {
+            canonicalized.push(canonical);
+        }
+    }
+
+    Ok(canonicalized)
+}
+
+fn validate_chunk_statuses(chunks: &[ChunkMeta]) -> Result<()> {
+    let invalid_chunk = chunks
+        .iter()
+        .find(|chunk| !matches!(chunk.status, ChunkStatus::Completed | ChunkStatus::Skipped));
+
+    if let Some(chunk) = invalid_chunk {
+        return IncompleteSnapshotSnafu {
+            chunk_id: chunk.id,
+            status: chunk.status,
+        }
+        .fail();
+    }
+
+    Ok(())
+}
+
+fn format_data_import_plan(chunks: &[ChunkMeta], schemas: &[String]) -> Vec<String> {
+    let mut lines = vec!["-- Data import plan:".to_string()];
+    for chunk in chunks {
+        lines.push(format!("-- Chunk {}: {:?}", chunk.id, chunk.status));
+        for schema in schemas {
+            if chunk_has_schema_files(chunk, schema) {
+                lines.push(format!("--   {} -> COPY DATABASE FROM", schema));
+            }
+        }
+    }
+    lines
+}
+
+async fn validate_data_snapshot(
+    storage: &dyn SnapshotStorage,
+    chunks: &[ChunkMeta],
+    schemas: &[String],
+) -> Result<()> {
+    validate_chunk_statuses(chunks)?;
+    let actual_prefixes = collect_chunk_data_prefixes(storage).await?;
+
+    for chunk in chunks {
+        if chunk.status == ChunkStatus::Skipped {
+            continue;
+        }
+        if chunk.files.is_empty() {
+            return EmptyChunkManifestSnafu { chunk_id: chunk.id }.fail();
+        }
+        for schema in schemas {
+            validate_chunk_schema_files(chunk, schema, &actual_prefixes)?;
+        }
+    }
+
+    Ok(())
+}
+
+async fn collect_chunk_data_prefixes(storage: &dyn SnapshotStorage) -> Result<HashSet<String>> {
+    let files = storage
+        .list_files_recursive("data/")
+        .await
+        .context(SnapshotStorageSnafu)?;
+    let mut prefixes = HashSet::new();
+
+    for path in files {
+        let normalized = path.trim_start_matches('/');
+        let mut parts = normalized.splitn(4, '/');
+        let Some(root) = parts.next() else {
+            continue;
+        };
+        let Some(schema) = parts.next() else {
+            continue;
+        };
+        let Some(chunk_id) = parts.next() else {
+            continue;
+        };
+        if root != "data" {
+            continue;
+        }
+        prefixes.insert(format!("data/{schema}/{chunk_id}/"));
+    }
+
+    Ok(prefixes)
+}
+
+fn validate_chunk_schema_files(
+    chunk: &ChunkMeta,
+    schema: &str,
+    actual_prefixes: &HashSet<String>,
+) -> Result<bool> {
+    if !chunk_has_schema_files(chunk, schema) {
+        return Ok(false);
+    }
+
+    let prefix = data_dir_for_schema_chunk(schema, chunk.id);
+    if !actual_prefixes.contains(&prefix) {
+        return MissingChunkDataSnafu {
+            chunk_id: chunk.id,
+            schema: schema.to_string(),
+            path: prefix,
+        }
+        .fail();
+    }
+
+    Ok(true)
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::{HashMap, HashSet};
+
+    use async_trait::async_trait;
+
+    use super::*;
+    use crate::data::export_v2::manifest::{ChunkMeta, ChunkStatus, Manifest, TimeRange};
+    use crate::data::export_v2::schema::SchemaSnapshot;
+    use crate::data::snapshot_storage::SnapshotStorage;
+
+    struct StubStorage {
+        manifest: Manifest,
+        files_by_prefix: HashMap<String, Vec<String>>,
+    }
+
+    #[async_trait]
+    impl SnapshotStorage for StubStorage {
+        async fn exists(&self) -> crate::data::export_v2::error::Result<bool> {
+            Ok(true)
+        }
+
+        async fn read_manifest(&self) -> crate::data::export_v2::error::Result<Manifest> {
+            Ok(self.manifest.clone())
+        }
+
+        async fn write_manifest(
+            &self,
+            _manifest: &Manifest,
+        ) -> crate::data::export_v2::error::Result<()> {
+            unimplemented!("not needed in import_v2::command tests")
+        }
+
+        async fn read_text(&self, _path: &str) -> crate::data::export_v2::error::Result<String> {
+            unimplemented!("not needed in import_v2::command tests")
+        }
+
+        async fn write_text(
+            &self,
+            _path: &str,
+            _content: &str,
+        ) -> crate::data::export_v2::error::Result<()> {
+            unimplemented!("not needed in import_v2::command tests")
+        }
+
+        async fn write_schema(
+            &self,
+            _snapshot: &SchemaSnapshot,
+        ) -> crate::data::export_v2::error::Result<()> {
+            unimplemented!("not needed in import_v2::command tests")
+        }
+
+        async fn create_dir_all(&self, _path: &str) -> crate::data::export_v2::error::Result<()> {
+            unimplemented!("not needed in import_v2::command tests")
+        }
+
+        async fn list_files_recursive(
+            &self,
+            prefix: &str,
+        ) -> crate::data::export_v2::error::Result<Vec<String>> {
+            Ok(self
+                .files_by_prefix
+                .iter()
+                .filter(|(candidate, _)| candidate.starts_with(prefix))
+                .flat_map(|(_, files)| files.clone())
+                .collect())
+        }
+
+        async fn delete_snapshot(&self) -> crate::data::export_v2::error::Result<()> {
+            unimplemented!("not needed in import_v2::command tests")
+        }
+    }
+
+    #[test]
+    fn test_parse_ddl_statements() {
+        let content = r#"
+-- Schema: public
+CREATE DATABASE public;
+CREATE TABLE t (ts TIMESTAMP TIME INDEX, host STRING, PRIMARY KEY (host)) ENGINE=mito;
+
+-- comment
+CREATE VIEW v AS SELECT * FROM t;
+"#;
+        let statements = parse_ddl_statements(content);
+        assert_eq!(statements.len(), 3);
+        assert!(statements[0].starts_with("CREATE DATABASE public"));
+        assert!(statements[1].starts_with("CREATE TABLE t"));
+        assert!(statements[2].starts_with("CREATE VIEW v"));
+    }
+
+    #[test]
+    fn test_parse_ddl_statements_preserves_semicolons_in_string_literals() {
+        let content = r#"
+CREATE TABLE t (
+    host STRING DEFAULT 'a;b'
+);
+CREATE VIEW v AS SELECT ';' AS marker;
+"#;
+
+        let statements = parse_ddl_statements(content);
+
+        assert_eq!(statements.len(), 2);
+        assert!(statements[0].contains("'a;b'"));
+        assert!(statements[1].contains("';' AS marker"));
+    }
+
+    #[test]
+    fn test_parse_ddl_statements_handles_comments_without_splitting() {
+        let content = r#"
+-- leading comment
+CREATE TABLE t (ts TIMESTAMP TIME INDEX); /* block; comment */
+CREATE VIEW v AS SELECT 1;
+"#;
+
+        let statements = parse_ddl_statements(content);
+
+        assert_eq!(statements.len(), 2);
+        assert!(statements[0].starts_with("CREATE TABLE t"));
+        assert!(statements[1].starts_with("CREATE VIEW v"));
+    }
+
+    #[test]
+    fn test_canonicalize_schema_filter_uses_manifest_casing() {
+        let filter = vec!["TEST_DB".to_string(), "PUBLIC".to_string()];
+        let manifest_schemas = vec!["test_db".to_string(), "public".to_string()];
+
+        let canonicalized = canonicalize_schema_filter(&filter, &manifest_schemas).unwrap();
+
+        assert_eq!(canonicalized, vec!["test_db", "public"]);
+    }
+
+    #[test]
+    fn test_canonicalize_schema_filter_dedupes_case_insensitive_matches() {
+        let filter = vec![
+            "TEST_DB".to_string(),
+            "test_db".to_string(),
+            "PUBLIC".to_string(),
+            "public".to_string(),
+        ];
+        let manifest_schemas = vec!["test_db".to_string(), "public".to_string()];
+
+        let canonicalized = canonicalize_schema_filter(&filter, &manifest_schemas).unwrap();
+
+        assert_eq!(canonicalized, vec!["test_db", "public"]);
+    }
+
+    #[test]
+    fn test_canonicalize_schema_filter_rejects_missing_schema() {
+        let filter = vec!["missing".to_string()];
+        let manifest_schemas = vec!["test_db".to_string()];
+
+        let error = canonicalize_schema_filter(&filter, &manifest_schemas)
+            .expect_err("missing schema should fail")
+            .to_string();
+
+        assert!(error.contains("missing"));
+    }
+
+    #[test]
+    fn test_ddl_statement_for_schema_create_table_uses_execution_schema() {
+        let stmt = ddl_statement_for_schema(
+            "test_db",
+            "CREATE TABLE metrics (ts TIMESTAMP TIME INDEX) ENGINE=mito".to_string(),
+        );
+        assert_eq!(stmt.execution_schema.as_deref(), Some("test_db"));
+    }
+
+    #[test]
+    fn test_ddl_statement_for_schema_create_view_uses_execution_schema() {
+        let stmt = ddl_statement_for_schema(
+            "test_db",
+            "CREATE VIEW metrics_view AS SELECT * FROM metrics".to_string(),
+        );
+        assert_eq!(stmt.execution_schema.as_deref(), Some("test_db"));
+    }
+
+    #[test]
+    fn test_ddl_statement_for_schema_create_or_replace_view_uses_execution_schema() {
+        let stmt = ddl_statement_for_schema(
+            "test_db",
+            "CREATE OR REPLACE VIEW metrics_view AS SELECT * FROM metrics".to_string(),
+        );
+        assert_eq!(stmt.execution_schema.as_deref(), Some("test_db"));
+    }
+
+    #[test]
+    fn test_ddl_statement_for_schema_create_external_table_uses_execution_schema() {
+        let stmt = ddl_statement_for_schema(
+            "test_db",
+            "CREATE EXTERNAL TABLE IF NOT EXISTS ext_metrics (ts TIMESTAMP TIME INDEX) ENGINE=file"
+                .to_string(),
+        );
+        assert_eq!(stmt.execution_schema.as_deref(), Some("test_db"));
+    }
+
+    #[test]
+    fn test_ddl_statement_for_schema_create_database_uses_public_context() {
+        let stmt = ddl_statement_for_schema("test_db", "CREATE DATABASE test_db".to_string());
+        assert_eq!(stmt.execution_schema, None);
+    }
+
+    #[test]
+    fn test_starts_with_keyword_requires_word_boundary() {
+        assert!(starts_with_keyword("CREATE TABLE t", "CREATE"));
+        assert!(!starts_with_keyword("CREATED TABLE t", "CREATE"));
+        assert!(!starts_with_keyword("TABLESPACE foo", "TABLE"));
+    }
+
+    #[test]
+    fn test_validate_chunk_statuses_rejects_failed_chunk() {
+        let mut failed = ChunkMeta::new(3, TimeRange::unbounded());
+        failed.status = ChunkStatus::Failed;
+
+        let error = validate_chunk_statuses(&[failed]).expect_err("failed chunk should error");
+        assert!(error.to_string().contains("Incomplete snapshot"));
+    }
+
+    #[test]
+    fn test_validate_chunk_statuses_accepts_completed_and_skipped_chunks() {
+        let mut completed = ChunkMeta::new(1, TimeRange::unbounded());
+        completed.status = ChunkStatus::Completed;
+        let skipped = ChunkMeta::skipped(2, TimeRange::unbounded());
+
+        assert!(validate_chunk_statuses(&[completed, skipped]).is_ok());
+    }
+
+    #[test]
+    fn test_chunk_has_schema_files_matches_encoded_schema_prefix() {
+        let mut chunk = ChunkMeta::new(7, TimeRange::unbounded());
+        chunk.files = vec![
+            "data/public/7/a.parquet".to_string(),
+            "data/%E6%B5%8B%E8%AF%95/7/b.parquet".to_string(),
+        ];
+
+        assert!(chunk_has_schema_files(&chunk, "public"));
+        assert!(chunk_has_schema_files(&chunk, "测试"));
+        assert!(!chunk_has_schema_files(&chunk, "metrics"));
+    }
+
+    #[test]
+    fn test_format_data_import_plan_includes_matching_schemas_only() {
+        let mut completed = ChunkMeta::new(1, TimeRange::unbounded());
+        completed.status = ChunkStatus::Completed;
+        completed.files = vec![
+            "data/public/1/a.parquet".to_string(),
+            "data/%E6%B5%8B%E8%AF%95/1/b.parquet".to_string(),
+        ];
+        let skipped = ChunkMeta::skipped(2, TimeRange::unbounded());
+
+        let lines = format_data_import_plan(
+            &[completed, skipped],
+            &[
+                "public".to_string(),
+                "测试".to_string(),
+                "metrics".to_string(),
+            ],
+        );
+
+        assert_eq!(lines[0], "-- Data import plan:");
+        assert!(lines.contains(&"-- Chunk 1: Completed".to_string()));
+        assert!(lines.contains(&"--   public -> COPY DATABASE FROM".to_string()));
+        assert!(lines.contains(&"--   测试 -> COPY DATABASE FROM".to_string()));
+        assert!(!lines.contains(&"--   metrics -> COPY DATABASE FROM".to_string()));
+        assert!(lines.contains(&"-- Chunk 2: Skipped".to_string()));
+    }
+
+    #[tokio::test]
+    async fn test_collect_chunk_data_prefixes_indexes_present_prefixes() {
+        let storage = StubStorage {
+            manifest: Manifest::new_schema_only("greptime".to_string(), vec!["public".to_string()]),
+            files_by_prefix: HashMap::from([
+                (
+                    "data/public/7/".to_string(),
+                    vec!["data/public/7/a.parquet".to_string()],
+                ),
+                (
+                    "data/%E6%B5%8B%E8%AF%95/9/".to_string(),
+                    vec!["data/%E6%B5%8B%E8%AF%95/9/b.parquet".to_string()],
+                ),
+            ]),
+        };
+
+        let prefixes = collect_chunk_data_prefixes(&storage).await.unwrap();
+
+        assert!(prefixes.contains("data/public/7/"));
+        assert!(prefixes.contains("data/%E6%B5%8B%E8%AF%95/9/"));
+    }
+
+    #[test]
+    fn test_validate_chunk_schema_files_accepts_present_prefix() {
+        let mut chunk = ChunkMeta::new(7, TimeRange::unbounded());
+        chunk.files = vec!["data/public/7/a.parquet".to_string()];
+        let actual_prefixes = HashSet::from(["data/public/7/".to_string()]);
+
+        assert!(validate_chunk_schema_files(&chunk, "public", &actual_prefixes).unwrap());
+    }
+
+    #[test]
+    fn test_validate_chunk_schema_files_rejects_missing_prefix() {
+        let mut chunk = ChunkMeta::new(7, TimeRange::unbounded());
+        chunk.files = vec!["data/public/7/a.parquet".to_string()];
+
+        let error = validate_chunk_schema_files(&chunk, "public", &HashSet::new())
+            .expect_err("missing chunk prefix should fail")
+            .to_string();
+        assert!(error.contains("marked completed but no files were found"));
+    }
+
+    #[test]
+    fn test_validate_chunk_schema_files_skips_absent_schema() {
+        let mut chunk = ChunkMeta::new(7, TimeRange::unbounded());
+        chunk.files = vec!["data/public/7/a.parquet".to_string()];
+
+        assert!(!validate_chunk_schema_files(&chunk, "metrics", &HashSet::new()).unwrap());
+    }
+
+    #[tokio::test]
+    async fn test_validate_data_snapshot_rejects_failed_chunk_before_dry_run() {
+        let mut failed = ChunkMeta::new(3, TimeRange::unbounded());
+        failed.status = ChunkStatus::Failed;
+
+        let storage = StubStorage {
+            manifest: Manifest::new_schema_only("greptime".to_string(), vec!["public".to_string()]),
+            files_by_prefix: HashMap::new(),
+        };
+
+        let error = validate_data_snapshot(&storage, &[failed], &["public".to_string()])
+            .await
+            .expect_err("failed chunk should reject dry-run validation")
+            .to_string();
+        assert!(error.contains("Incomplete snapshot"));
+    }
+
+    #[tokio::test]
+    async fn test_validate_data_snapshot_rejects_missing_chunk_prefix_before_dry_run() {
+        let mut completed = ChunkMeta::new(7, TimeRange::unbounded());
+        completed.status = ChunkStatus::Completed;
+        completed.files = vec!["data/public/7/a.parquet".to_string()];
+
+        let storage = StubStorage {
+            manifest: Manifest::new_schema_only("greptime".to_string(), vec!["public".to_string()]),
+            files_by_prefix: HashMap::new(),
+        };
+
+        let error = validate_data_snapshot(&storage, &[completed], &["public".to_string()])
+            .await
+            .expect_err("missing chunk prefix should reject dry-run validation")
+            .to_string();
+        assert!(error.contains("marked completed but no files were found"));
+    }
+
+    #[tokio::test]
+    async fn test_validate_data_snapshot_rejects_completed_chunk_with_empty_manifest() {
+        let mut completed = ChunkMeta::new(7, TimeRange::unbounded());
+        completed.status = ChunkStatus::Completed;
+
+        let storage = StubStorage {
+            manifest: Manifest::new_schema_only("greptime".to_string(), vec!["public".to_string()]),
+            files_by_prefix: HashMap::new(),
+        };
+
+        let error = validate_data_snapshot(&storage, &[completed], &["public".to_string()])
+            .await
+            .expect_err("empty completed chunk should reject validation")
+            .to_string();
+        assert!(error.contains("file manifest is empty"));
+    }
+}
diff --git a/src/cli/src/data/import_v2/coordinator.rs b/src/cli/src/data/import_v2/coordinator.rs
new file mode 100644
index 0000000000..fe2d016464
--- /dev/null
+++ b/src/cli/src/data/import_v2/coordinator.rs
@@ -0,0 +1,695 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::BTreeSet;
+use std::path::{Path, PathBuf};
+use std::time::Instant;
+
+use async_trait::async_trait;
+use common_telemetry::{info, warn};
+
+use crate::data::export_v2::manifest::{ChunkMeta, ChunkStatus};
+use crate::data::import_v2::error::{
+    ImportStateDdlIncompleteSnafu, ImportStateMismatchSnafu, Result,
+};
+use crate::data::import_v2::state::{
+    ImportState, ImportStateLockGuard, ImportTaskKey, ImportTaskStatus, canonical_schema_selection,
+    delete_import_state, load_import_state, save_import_state, try_acquire_import_state_lock,
+};
+use crate::data::path::data_dir_for_schema_chunk;
+
+#[async_trait]
+pub(crate) trait ImportTaskExecutor {
+    async fn import_task(&self, task: &ImportTaskKey) -> Result<()>;
+}
+
+pub(crate) struct ImportResumeConfig {
+    pub(crate) snapshot_id: String,
+    pub(crate) target_addr: String,
+    pub(crate) catalog: String,
+    pub(crate) schemas: Vec<String>,
+    pub(crate) state_path: PathBuf,
+    pub(crate) tasks: Vec<ImportTaskKey>,
+}
+
+pub(crate) struct ImportResumeSession {
+    config: ImportResumeConfig,
+    state: ImportState,
+    lock: ImportStateLockGuard,
+}
+
+impl ImportResumeSession {
+    pub(crate) fn should_skip_ddl(&self) -> bool {
+        self.state.ddl_completed
+    }
+
+    /// Marks DDL as completed and persists the state. Must be called after a
+    /// successful DDL run on a fresh session, so that crashes after this point
+    /// resume into the data-import phase instead of replaying DDL.
+    pub(crate) async fn mark_ddl_completed(&mut self) -> Result<()> {
+        self.state.mark_ddl_completed();
+        save_import_state(&self.config.state_path, &self.state).await
+    }
+}
+
+pub(crate) fn chunk_has_schema_files(chunk: &ChunkMeta, schema: &str) -> bool {
+    let prefix = data_dir_for_schema_chunk(schema, chunk.id);
+    chunk.files.iter().any(|path| {
+        let normalized = path.trim_start_matches('/');
+        normalized.starts_with(&prefix)
+    })
+}
+
+pub(crate) fn build_import_tasks(chunks: &[ChunkMeta], schemas: &[String]) -> Vec<ImportTaskKey> {
+    let mut tasks = Vec::new();
+    for chunk in chunks {
+        if chunk.status == ChunkStatus::Skipped {
+            continue;
+        }
+        // TODO: build a per-chunk schema index if chunk file manifests become large.
+        for schema in schemas {
+            if chunk_has_schema_files(chunk, schema) {
+                tasks.push(ImportTaskKey::new(chunk.id, schema.clone()));
+            }
+        }
+    }
+    tasks
+}
+
+pub(crate) async fn prepare_import_resume(
+    config: ImportResumeConfig,
+) -> Result<ImportResumeSession> {
+    // Validate the request before touching the state file or acquiring the
+    // lock. Duplicate task keys would corrupt the resume bookkeeping because
+    // status lookups use linear `find()` and only ever see the first match.
+    validate_config_tasks(&config)?;
+
+    let lock = try_acquire_import_state_lock(&config.state_path)?;
+    let state = match load_import_state(&config.state_path).await? {
+        Some(loaded) => {
+            validate_state_matches(&loaded, &config)?;
+            loaded
+        }
+        None => {
+            // Persist a fresh state immediately so that any crash after this
+            // point is recoverable as a resume. `ddl_completed=false` on a
+            // loaded state therefore means a previous run reached this point
+            // but did not confirm DDL completion - DDL must be (re-)run before
+            // data import is allowed.
+            let fresh = ImportState::new(
+                &config.snapshot_id,
+                &config.target_addr,
+                &config.catalog,
+                &config.schemas,
+                config.tasks.clone(),
+            );
+            save_import_state(&config.state_path, &fresh).await?;
+            fresh
+        }
+    };
+
+    Ok(ImportResumeSession {
+        config,
+        state,
+        lock,
+    })
+}
+
+pub(crate) async fn import_with_resume_session<E>(
+    session: ImportResumeSession,
+    executor: &E,
+) -> Result<()>
+where
+    E: ImportTaskExecutor + Sync,
+{
+    let ImportResumeSession {
+        config,
+        mut state,
+        lock,
+    } = session;
+
+    // The state machine requires DDL to be explicitly marked completed before
+    // data import; otherwise a caller could import data and leave a state that
+    // replays DDL on the next resume. Surface the misuse instead of silently
+    // importing.
+    if !state.ddl_completed {
+        return ImportStateDdlIncompleteSnafu {
+            path: config.state_path.display().to_string(),
+        }
+        .fail();
+    }
+
+    let completed = state
+        .tasks
+        .iter()
+        .filter(|task| task.status == ImportTaskStatus::Completed)
+        .count();
+    info!(
+        "Import resume state: {} completed, {} pending, path: {}",
+        completed,
+        state.tasks.len().saturating_sub(completed),
+        config.state_path.display()
+    );
+
+    let import_start = Instant::now();
+    for (idx, task) in config.tasks.iter().enumerate() {
+        if state.task_status(task.chunk_id, &task.schema) == Some(ImportTaskStatus::Completed) {
+            info!(
+                "[{}/{}] Chunk {} schema {}: already completed, skipped",
+                idx + 1,
+                config.tasks.len(),
+                task.chunk_id,
+                task.schema
+            );
+            continue;
+        }
+
+        info!(
+            "[{}/{}] Chunk {} schema {}: importing...",
+            idx + 1,
+            config.tasks.len(),
+            task.chunk_id,
+            task.schema
+        );
+        state.set_task_status(
+            task.chunk_id,
+            &task.schema,
+            ImportTaskStatus::InProgress,
+            None,
+        )?;
+        save_import_state(&config.state_path, &state).await?;
+
+        let task_start = Instant::now();
+        let result = executor.import_task(task).await;
+
+        match result {
+            Ok(()) => {
+                // The task itself succeeded. If we cannot persist the
+                // Completed marker, the next resume will replay it (potentially
+                // duplicating data depending on engine semantics), but we must
+                // not pretend the import as a whole failed - return the persist
+                // error so the operator notices, after logging the success.
+                update_status_and_save(
+                    &config,
+                    &mut state,
+                    task,
+                    ImportTaskStatus::Completed,
+                    None,
+                )
+                .await?;
+                info!(
+                    "[{}/{}] Chunk {} schema {}: done in {:?}",
+                    idx + 1,
+                    config.tasks.len(),
+                    task.chunk_id,
+                    task.schema,
+                    task_start.elapsed()
+                );
+            }
+            Err(task_error) => {
+                // Persist Failed best-effort, but always surface the original
+                // task error to the caller. State persistence problems are
+                // logged so they are not silently lost.
+                if let Err(persist_error) = update_status_and_save(
+                    &config,
+                    &mut state,
+                    task,
+                    ImportTaskStatus::Failed,
+                    Some(task_error.to_string()),
+                )
+                .await
+                {
+                    warn!(
+                        "Failed to persist Failed status for chunk {} schema {} after task error ({}); state file may be out of date: {}",
+                        task.chunk_id, task.schema, task_error, persist_error
+                    );
+                }
+                return Err(task_error);
+            }
+        }
+    }
+
+    delete_import_state(&config.state_path).await?;
+    info!("Data import finished in {:?}", import_start.elapsed());
+    drop(lock);
+    Ok(())
+}
+
+async fn update_status_and_save(
+    config: &ImportResumeConfig,
+    state: &mut ImportState,
+    task: &ImportTaskKey,
+    status: ImportTaskStatus,
+    error_message: Option<String>,
+) -> Result<()> {
+    // set_task_status only fails if the task isn't in the state; that would
+    // indicate a logic bug since `task` came from the same config. Surface it
+    // instead of swallowing.
+    state.set_task_status(task.chunk_id, &task.schema, status, error_message)?;
+    save_import_state(&config.state_path, state).await
+}
+
+fn validate_state_matches(state: &ImportState, config: &ImportResumeConfig) -> Result<()> {
+    if state.snapshot_id != config.snapshot_id {
+        return state_mismatch(
+            config,
+            format!(
+                "snapshot_id differs (state: {}, requested: {})",
+                state.snapshot_id, config.snapshot_id
+            ),
+        );
+    }
+    // Target addresses are compared literally; hostname normalization is left to the caller.
+    if state.target_addr != config.target_addr {
+        return state_mismatch(
+            config,
+            format!(
+                "target_addr differs (state: {}, requested: {})",
+                state.target_addr, config.target_addr
+            ),
+        );
+    }
+    if state.catalog != config.catalog {
+        return state_mismatch(
+            config,
+            format!(
+                "catalog differs (state: {}, requested: {})",
+                state.catalog, config.catalog
+            ),
+        );
+    }
+
+    let requested_schemas = canonical_schema_selection(&config.schemas);
+    if state.schemas != requested_schemas {
+        return state_mismatch(
+            config,
+            format!(
+                "schemas differ (state: {:?}, requested: {:?})",
+                state.schemas, requested_schemas
+            ),
+        );
+    }
+
+    if task_set_from_state(state, &config.state_path)? != task_set_from_config(config)? {
+        return state_mismatch(config, "task set differs".to_string());
+    }
+
+    Ok(())
+}
+
+fn state_mismatch(config: &ImportResumeConfig, reason: String) -> Result<()> {
+    ImportStateMismatchSnafu {
+        path: config.state_path.display().to_string(),
+        reason,
+    }
+    .fail()
+}
+
+fn task_set_from_state<'a>(
+    state: &'a ImportState,
+    state_path: &Path,
+) -> Result<BTreeSet<(u32, &'a str)>> {
+    let mut tasks = BTreeSet::new();
+    for task in &state.tasks {
+        if !tasks.insert((task.chunk_id, task.schema.as_str())) {
+            return ImportStateMismatchSnafu {
+                path: state_path.display().to_string(),
+                reason: format!(
+                    "duplicate task key in state (chunk_id: {}, schema: {})",
+                    task.chunk_id, task.schema
+                ),
+            }
+            .fail();
+        }
+    }
+    Ok(tasks)
+}
+
+fn task_set_from_config(config: &ImportResumeConfig) -> Result<BTreeSet<(u32, &str)>> {
+    let mut tasks = BTreeSet::new();
+    for task in &config.tasks {
+        if !tasks.insert((task.chunk_id, task.schema.as_str())) {
+            return ImportStateMismatchSnafu {
+                path: config.state_path.display().to_string(),
+                reason: format!(
+                    "duplicate task key in request (chunk_id: {}, schema: {})",
+                    task.chunk_id, task.schema
+                ),
+            }
+            .fail();
+        }
+    }
+    Ok(tasks)
+}
+
+fn validate_config_tasks(config: &ImportResumeConfig) -> Result<()> {
+    task_set_from_config(config).map(|_| ())
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::atomic::{AtomicUsize, Ordering};
+    use std::sync::{Arc, Mutex};
+
+    use super::*;
+    use crate::data::export_v2::manifest::{ChunkMeta, TimeRange};
+    use crate::data::import_v2::error::TestTaskFailedSnafu;
+
+    #[derive(Debug, Clone, Copy)]
+    enum FailureMode {
+        Fatal,
+        RetryableThenSuccess { failures: usize },
+    }
+
+    struct RecordingExecutor {
+        imported: Arc<Mutex<Vec<ImportTaskKey>>>,
+        fail_task: Option<ImportTaskKey>,
+        failure_mode: Option<FailureMode>,
+        attempts: Arc<AtomicUsize>,
+    }
+
+    #[async_trait]
+    impl ImportTaskExecutor for RecordingExecutor {
+        async fn import_task(&self, task: &ImportTaskKey) -> Result<()> {
+            let attempt = self.attempts.fetch_add(1, Ordering::SeqCst);
+            if self.fail_task.as_ref() == Some(task) {
+                match self.failure_mode {
+                    Some(FailureMode::Fatal) => {
+                        return TestTaskFailedSnafu {
+                            message: "fatal failure".to_string(),
+                            retryable: false,
+                        }
+                        .fail();
+                    }
+                    Some(FailureMode::RetryableThenSuccess { failures }) if attempt < failures => {
+                        return TestTaskFailedSnafu {
+                            message: "retryable failure".to_string(),
+                            retryable: true,
+                        }
+                        .fail();
+                    }
+                    _ => {}
+                }
+            }
+            self.imported.lock().unwrap().push(task.clone());
+            Ok(())
+        }
+    }
+
+    fn recording_executor(imported: Arc<Mutex<Vec<ImportTaskKey>>>) -> RecordingExecutor {
+        RecordingExecutor {
+            imported,
+            fail_task: None,
+            failure_mode: None,
+            attempts: Arc::new(AtomicUsize::new(0)),
+        }
+    }
+
+    fn config(path: PathBuf, tasks: Vec<ImportTaskKey>) -> ImportResumeConfig {
+        ImportResumeConfig {
+            snapshot_id: "snapshot-1".to_string(),
+            target_addr: "127.0.0.1:4000".to_string(),
+            catalog: "greptime".to_string(),
+            schemas: vec!["public".to_string(), "analytics".to_string()],
+            state_path: path,
+            tasks,
+        }
+    }
+
+    async fn run_import_with_resume<E>(config: ImportResumeConfig, executor: &E) -> Result<()>
+    where
+        E: ImportTaskExecutor + Sync,
+    {
+        // Mirror the production caller: mark DDL completed for fresh sessions
+        // so the data-import guard is satisfied. Tests that want to exercise
+        // the unsafe path drive prepare/import directly.
+        let mut session = prepare_import_resume(config).await?;
+        if !session.should_skip_ddl() {
+            session.mark_ddl_completed().await?;
+        }
+        import_with_resume_session(session, executor).await
+    }
+
+    #[test]
+    fn test_build_import_tasks_skips_skipped_chunks_and_missing_schema_files() {
+        let mut completed = ChunkMeta::new(1, TimeRange::unbounded());
+        completed.status = ChunkStatus::Completed;
+        completed.files = vec!["data/public/1/file.parquet".to_string()];
+        let mut skipped = ChunkMeta::new(2, TimeRange::unbounded());
+        skipped.status = ChunkStatus::Skipped;
+        skipped.files = vec!["data/public/2/file.parquet".to_string()];
+
+        let tasks = build_import_tasks(
+            &[completed, skipped],
+            &["public".to_string(), "analytics".to_string()],
+        );
+
+        assert_eq!(tasks, vec![ImportTaskKey::new(1, "public")]);
+    }
+
+    #[tokio::test]
+    async fn test_import_with_resume_skips_completed_tasks() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("import_state.json");
+        let tasks = vec![
+            ImportTaskKey::new(1, "public"),
+            ImportTaskKey::new(2, "analytics"),
+        ];
+        let mut state = ImportState::new(
+            "snapshot-1",
+            "127.0.0.1:4000",
+            "greptime",
+            &["public".to_string(), "analytics".to_string()],
+            tasks.clone(),
+        );
+        state.mark_ddl_completed();
+        state
+            .set_task_status(1, "public", ImportTaskStatus::Completed, None)
+            .unwrap();
+        save_import_state(&path, &state).await.unwrap();
+
+        let imported = Arc::new(Mutex::new(Vec::new()));
+        let executor = recording_executor(imported.clone());
+
+        run_import_with_resume(config(path.clone(), tasks), &executor)
+            .await
+            .unwrap();
+
+        assert_eq!(
+            imported.lock().unwrap().clone(),
+            vec![ImportTaskKey::new(2, "analytics")]
+        );
+        assert!(load_import_state(&path).await.unwrap().is_none());
+    }
+
+    #[tokio::test]
+    async fn test_import_with_resume_persists_failed_task() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("import_state.json");
+        let failed_task = ImportTaskKey::new(1, "public");
+        let tasks = vec![failed_task.clone()];
+        let imported = Arc::new(Mutex::new(Vec::new()));
+        let executor = RecordingExecutor {
+            imported,
+            fail_task: Some(failed_task.clone()),
+            failure_mode: Some(FailureMode::Fatal),
+            attempts: Arc::new(AtomicUsize::new(0)),
+        };
+
+        let error = run_import_with_resume(config(path.clone(), tasks), &executor)
+            .await
+            .unwrap_err();
+        assert!(matches!(
+            error,
+            crate::data::import_v2::error::Error::TestTaskFailed {
+                retryable: false,
+                ..
+            }
+        ));
+
+        let state = load_import_state(&path).await.unwrap().unwrap();
+        assert_eq!(
+            state.task_status(failed_task.chunk_id, &failed_task.schema),
+            Some(ImportTaskStatus::Failed)
+        );
+    }
+
+    #[tokio::test]
+    async fn test_import_with_resume_rejects_mismatched_state_identity() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("import_state.json");
+        let tasks = vec![ImportTaskKey::new(1, "public")];
+        let state = ImportState::new(
+            "snapshot-1",
+            "127.0.0.1:4001",
+            "greptime",
+            &["public".to_string(), "analytics".to_string()],
+            tasks.clone(),
+        );
+        save_import_state(&path, &state).await.unwrap();
+
+        let imported = Arc::new(Mutex::new(Vec::new()));
+        let executor = recording_executor(imported);
+
+        let error = run_import_with_resume(config(path, tasks), &executor)
+            .await
+            .unwrap_err();
+
+        assert!(matches!(
+            error,
+            crate::data::import_v2::error::Error::ImportStateMismatch { .. }
+        ));
+    }
+
+    #[tokio::test]
+    async fn test_prepare_import_resume_reports_existing_state_before_ddl() {
+        let dir = tempfile::tempdir().unwrap();
+        let tasks = vec![ImportTaskKey::new(1, "public")];
+
+        let fresh_session =
+            prepare_import_resume(config(dir.path().join("fresh_state.json"), tasks.clone()))
+                .await
+                .unwrap();
+        assert!(!fresh_session.should_skip_ddl());
+        drop(fresh_session);
+
+        let existing_path = dir.path().join("existing_state.json");
+        let mut state = ImportState::new(
+            "snapshot-1",
+            "127.0.0.1:4000",
+            "greptime",
+            &["public".to_string(), "analytics".to_string()],
+            tasks.clone(),
+        );
+        state.mark_ddl_completed();
+        save_import_state(&existing_path, &state).await.unwrap();
+
+        let resume_session = prepare_import_resume(config(existing_path, tasks))
+            .await
+            .unwrap();
+        assert!(resume_session.should_skip_ddl());
+    }
+
+    #[tokio::test]
+    async fn test_import_with_resume_rejects_duplicate_state_tasks() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("import_state.json");
+        let tasks = vec![ImportTaskKey::new(1, "public")];
+        let mut state = ImportState::new(
+            "snapshot-1",
+            "127.0.0.1:4000",
+            "greptime",
+            &["public".to_string(), "analytics".to_string()],
+            tasks.clone(),
+        );
+        state.tasks.push(state.tasks[0].clone());
+        save_import_state(&path, &state).await.unwrap();
+
+        let imported = Arc::new(Mutex::new(Vec::new()));
+        let executor = recording_executor(imported);
+
+        let error = run_import_with_resume(config(path, tasks), &executor)
+            .await
+            .unwrap_err();
+
+        assert!(matches!(
+            error,
+            crate::data::import_v2::error::Error::ImportStateMismatch { .. }
+        ));
+    }
+
+    #[tokio::test]
+    async fn test_import_with_resume_rejects_data_import_when_ddl_incomplete() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("import_state.json");
+        let tasks = vec![ImportTaskKey::new(1, "public")];
+
+        // prepare creates fresh state with ddl_completed=false; calling
+        // import_with_resume_session directly (without mark_ddl_completed)
+        // must be rejected.
+        let session = prepare_import_resume(config(path, tasks)).await.unwrap();
+        let imported = Arc::new(Mutex::new(Vec::new()));
+        let executor = recording_executor(imported.clone());
+
+        let error = import_with_resume_session(session, &executor)
+            .await
+            .unwrap_err();
+
+        assert!(matches!(
+            error,
+            crate::data::import_v2::error::Error::ImportStateDdlIncomplete { .. }
+        ));
+        assert!(imported.lock().unwrap().is_empty());
+    }
+
+    #[tokio::test]
+    async fn test_prepare_import_resume_rejects_duplicate_request_tasks_on_fresh_state() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("import_state.json");
+        let task = ImportTaskKey::new(1, "public");
+        // No state file yet - duplicate detection must run before the fresh
+        // state is persisted, otherwise corrupted bookkeeping would be
+        // written to disk and observed only on a later resume.
+        let error =
+            match prepare_import_resume(config(path.clone(), vec![task.clone(), task])).await {
+                Ok(_) => panic!("duplicate request tasks should be rejected"),
+                Err(error) => error,
+            };
+
+        assert!(matches!(
+            error,
+            crate::data::import_v2::error::Error::ImportStateMismatch { .. }
+        ));
+        assert!(load_import_state(&path).await.unwrap().is_none());
+    }
+
+    #[tokio::test]
+    async fn test_import_with_resume_does_not_retry_retryable_task_error() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("import_state.json");
+        let failed_task = ImportTaskKey::new(1, "public");
+        let tasks = vec![failed_task.clone()];
+        let imported = Arc::new(Mutex::new(Vec::new()));
+        let attempts = Arc::new(AtomicUsize::new(0));
+        let executor = RecordingExecutor {
+            imported: imported.clone(),
+            fail_task: Some(failed_task.clone()),
+            // If task import were retried, the second attempt would succeed.
+            // COPY DATABASE FROM failures are ambiguous, so retryable errors
+            // must still stop immediately to avoid duplicate rows.
+            failure_mode: Some(FailureMode::RetryableThenSuccess { failures: 1 }),
+            attempts: attempts.clone(),
+        };
+
+        let error = run_import_with_resume(config(path.clone(), tasks), &executor)
+            .await
+            .unwrap_err();
+
+        assert!(matches!(
+            error,
+            crate::data::import_v2::error::Error::TestTaskFailed {
+                retryable: true,
+                ..
+            }
+        ));
+        assert_eq!(attempts.load(Ordering::SeqCst), 1);
+        assert!(imported.lock().unwrap().is_empty());
+
+        let state = load_import_state(&path).await.unwrap().unwrap();
+        assert_eq!(
+            state.task_status(failed_task.chunk_id, &failed_task.schema),
+            Some(ImportTaskStatus::Failed)
+        );
+    }
+}
diff --git a/src/cli/src/data/import_v2/error.rs b/src/cli/src/data/import_v2/error.rs
new file mode 100644
index 0000000000..165f1c0118
--- /dev/null
+++ b/src/cli/src/data/import_v2/error.rs
@@ -0,0 +1,222 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::any::Any;
+
+use common_error::ext::ErrorExt;
+use common_error::status_code::StatusCode;
+use common_macro::stack_trace_debug;
+use snafu::{Location, Snafu};
+
+use crate::data::export_v2::manifest::ChunkStatus;
+
+#[derive(Snafu)]
+#[snafu(visibility(pub))]
+#[stack_trace_debug]
+pub enum Error {
+    #[snafu(display("Snapshot not found at '{}'", uri))]
+    SnapshotNotFound {
+        uri: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Manifest version mismatch: expected {}, found {}", expected, found))]
+    ManifestVersionMismatch {
+        expected: u32,
+        found: u32,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Schema '{}' not found in snapshot", schema))]
+    SchemaNotInSnapshot {
+        schema: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Incomplete snapshot: chunk {} has status {:?}", chunk_id, status))]
+    IncompleteSnapshot {
+        chunk_id: u32,
+        status: ChunkStatus,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display(
+        "Snapshot is inconsistent: chunk {} is marked completed but its file manifest is empty",
+        chunk_id
+    ))]
+    EmptyChunkManifest {
+        chunk_id: u32,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display(
+        "Snapshot is inconsistent: chunk {} for schema '{}' is marked completed but no files were found under '{}'",
+        chunk_id,
+        schema,
+        path
+    ))]
+    MissingChunkData {
+        chunk_id: u32,
+        schema: String,
+        path: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Chunk {} import failed for schema '{}'", chunk_id, schema))]
+    ChunkImportFailed {
+        chunk_id: u32,
+        schema: String,
+        #[snafu(source)]
+        error: crate::data::export_v2::error::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Snapshot storage error"))]
+    SnapshotStorage {
+        #[snafu(source)]
+        error: crate::data::export_v2::error::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Database error"))]
+    Database {
+        #[snafu(source)]
+        error: crate::error::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Failed to parse import state file"))]
+    ImportStateParse {
+        #[snafu(source)]
+        error: serde_json::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Import state I/O failed at '{}': {}", path, error))]
+    ImportStateIo {
+        path: String,
+        #[snafu(source)]
+        error: std::io::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Import state is already locked at '{}'", path))]
+    ImportStateLocked {
+        path: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display(
+        "Failed to determine import state path for snapshot '{}'. Set HOME, USERPROFILE, or run from a valid current directory.",
+        snapshot_id
+    ))]
+    ImportStatePathUnavailable {
+        snapshot_id: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display(
+        "Import state at '{}' does not match current import: {}. Either rerun with matching import arguments, or delete the state file to start over (DDL will be re-executed).",
+        path,
+        reason
+    ))]
+    ImportStateMismatch {
+        path: String,
+        reason: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[cfg(test)]
+    #[snafu(display("Test task failed: {}", message))]
+    TestTaskFailed {
+        message: String,
+        retryable: bool,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display(
+        "Import state references unknown task: chunk {}, schema '{}'",
+        chunk_id,
+        schema
+    ))]
+    ImportStateUnknownTask {
+        chunk_id: u32,
+        schema: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display(
+        "Import state at '{}' is not ready for data import: DDL has not been marked completed",
+        path
+    ))]
+    ImportStateDdlIncomplete {
+        path: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+}
+
+pub type Result<T> = std::result::Result<T, Error>;
+
+impl ErrorExt for Error {
+    fn status_code(&self) -> StatusCode {
+        match self {
+            Error::SnapshotNotFound { .. }
+            | Error::SchemaNotInSnapshot { .. }
+            | Error::ManifestVersionMismatch { .. }
+            | Error::IncompleteSnapshot { .. }
+            | Error::EmptyChunkManifest { .. }
+            | Error::MissingChunkData { .. } => StatusCode::InvalidArguments,
+            Error::ImportStatePathUnavailable { .. }
+            | Error::ImportStateUnknownTask { .. }
+            | Error::ImportStateDdlIncomplete { .. } => StatusCode::Unexpected,
+            Error::ImportStateMismatch { .. } => StatusCode::InvalidArguments,
+            #[cfg(test)]
+            Error::TestTaskFailed { retryable, .. } => {
+                if *retryable {
+                    StatusCode::StorageUnavailable
+                } else {
+                    StatusCode::InvalidArguments
+                }
+            }
+            Error::Database { error, .. } => error.status_code(),
+            Error::SnapshotStorage { error, .. } | Error::ChunkImportFailed { error, .. } => {
+                error.status_code()
+            }
+            Error::ImportStateParse { .. } => StatusCode::Internal,
+            Error::ImportStateIo { .. } => StatusCode::StorageUnavailable,
+            Error::ImportStateLocked { .. } => StatusCode::IllegalState,
+        }
+    }
+
+    fn as_any(&self) -> &dyn Any {
+        self
+    }
+}
diff --git a/src/cli/src/data/import_v2/executor.rs b/src/cli/src/data/import_v2/executor.rs
new file mode 100644
index 0000000000..3f2bf66ae6
--- /dev/null
+++ b/src/cli/src/data/import_v2/executor.rs
@@ -0,0 +1,122 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! DDL execution for import.
+
+use common_telemetry::info;
+use snafu::ResultExt;
+
+use crate::data::import_v2::error::{DatabaseSnafu, Result};
+use crate::database::DatabaseClient;
+
+/// A DDL statement with an explicit execution schema context.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct DdlStatement {
+    pub sql: String,
+    pub execution_schema: Option<String>,
+}
+
+impl DdlStatement {
+    pub fn new(sql: String) -> Self {
+        Self {
+            sql,
+            execution_schema: None,
+        }
+    }
+
+    pub fn with_execution_schema(sql: String, schema: String) -> Self {
+        Self {
+            sql,
+            execution_schema: Some(schema),
+        }
+    }
+}
+
+/// Executes DDL statements against the database.
+pub struct DdlExecutor<'a> {
+    client: &'a DatabaseClient,
+}
+
+impl<'a> DdlExecutor<'a> {
+    /// Creates a new DDL executor.
+    pub fn new(client: &'a DatabaseClient) -> Self {
+        Self { client }
+    }
+
+    /// Executes a list of DDL statements, stopping on first error.
+    pub async fn execute_strict(&self, statements: &[DdlStatement]) -> Result<()> {
+        let total = statements.len();
+
+        for (i, stmt) in statements.iter().enumerate() {
+            let preview = preview_sql(&stmt.sql);
+
+            info!("Executing DDL ({}/{}): {}", i + 1, total, preview);
+
+            if let Some(schema) = stmt.execution_schema.as_deref() {
+                self.client
+                    .sql(&stmt.sql, schema)
+                    .await
+                    .context(DatabaseSnafu)?;
+            } else {
+                self.client
+                    .sql_in_public(&stmt.sql)
+                    .await
+                    .context(DatabaseSnafu)?;
+            }
+        }
+
+        Ok(())
+    }
+}
+
+fn preview_sql(sql: &str) -> String {
+    let mut chars = sql.chars();
+    let preview: String = chars.by_ref().take(80).collect();
+    if chars.next().is_some() {
+        format!("{preview}...")
+    } else {
+        preview
+    }
+}
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_statement_without_execution_schema_uses_public() {
+        let stmt = DdlStatement::new("CREATE DATABASE IF NOT EXISTS test_db".to_string());
+        assert_eq!(stmt.execution_schema, None);
+    }
+
+    #[test]
+    fn test_statement_with_execution_schema_preserves_context() {
+        let stmt = DdlStatement::with_execution_schema(
+            r#"CREATE TABLE IF NOT EXISTS "my""schema"."metrics" (ts TIMESTAMP TIME INDEX)"#
+                .to_string(),
+            r#"my"schema"#.to_string(),
+        );
+        assert_eq!(stmt.execution_schema.as_deref(), Some(r#"my"schema"#));
+    }
+
+    #[test]
+    fn test_preview_sql_truncates_at_char_boundary() {
+        let sql = format!(
+            "CREATE TABLE {} (ts TIMESTAMP TIME INDEX)",
+            "测".repeat(100)
+        );
+        let preview = preview_sql(&sql);
+        assert!(preview.ends_with("..."));
+        assert!(preview.is_char_boundary(preview.len()));
+    }
+}
diff --git a/src/cli/src/data/import_v2/state.rs b/src/cli/src/data/import_v2/state.rs
new file mode 100644
index 0000000000..ba431ac62f
--- /dev/null
+++ b/src/cli/src/data/import_v2/state.rs
@@ -0,0 +1,804 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::path::{Path, PathBuf};
+use std::sync::atomic::{AtomicU64, Ordering};
+
+use chrono::{DateTime, Utc};
+use fs2::FileExt;
+use serde::{Deserialize, Serialize};
+use snafu::{IntoError, OptionExt, ResultExt};
+use tokio::io::AsyncWriteExt;
+
+use crate::data::import_v2::error::{
+    ImportStateIoSnafu, ImportStateLockedSnafu, ImportStateParseSnafu, ImportStateUnknownTaskSnafu,
+    Result,
+};
+use crate::data::path::encode_path_segment;
+
+const IMPORT_STATE_ROOT: &str = ".greptime";
+const IMPORT_STATE_DIR: &str = "import_state";
+static IMPORT_STATE_TMP_ID: AtomicU64 = AtomicU64::new(0);
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub(crate) enum ImportTaskStatus {
+    Pending,
+    InProgress,
+    Completed,
+    Failed,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub(crate) struct ImportTaskKey {
+    pub(crate) chunk_id: u32,
+    pub(crate) schema: String,
+}
+
+impl ImportTaskKey {
+    pub(crate) fn new(chunk_id: u32, schema: impl Into<String>) -> Self {
+        Self {
+            chunk_id,
+            schema: schema.into(),
+        }
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub(crate) struct ImportTaskState {
+    pub(crate) chunk_id: u32,
+    pub(crate) schema: String,
+    pub(crate) status: ImportTaskStatus,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub(crate) error: Option<String>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub(crate) struct ImportState {
+    pub(crate) snapshot_id: String,
+    pub(crate) target_addr: String,
+    pub(crate) catalog: String,
+    pub(crate) schemas: Vec<String>,
+    #[serde(default)]
+    pub(crate) ddl_completed: bool,
+    pub(crate) updated_at: DateTime<Utc>,
+    // Tasks are (chunk-schema) tuples and can reach the tens of thousands;
+    // linear scans here are accepted because per-task work is dominated by
+    // network I/O and an fsync, but if the bound grows further this should be
+    // backed by a HashMap<(chunk_id, schema), index> rebuilt after load.
+    pub(crate) tasks: Vec<ImportTaskState>,
+}
+
+impl ImportState {
+    pub(crate) fn new<I>(
+        snapshot_id: impl Into<String>,
+        target_addr: impl Into<String>,
+        catalog: impl Into<String>,
+        schemas: &[String],
+        tasks: I,
+    ) -> Self
+    where
+        I: IntoIterator<Item = ImportTaskKey>,
+    {
+        Self {
+            snapshot_id: snapshot_id.into(),
+            target_addr: target_addr.into(),
+            catalog: catalog.into(),
+            schemas: canonical_schema_selection(schemas),
+            ddl_completed: false,
+            updated_at: Utc::now(),
+            tasks: tasks
+                .into_iter()
+                .map(|task| ImportTaskState {
+                    chunk_id: task.chunk_id,
+                    schema: task.schema,
+                    status: ImportTaskStatus::Pending,
+                    error: None,
+                })
+                .collect(),
+        }
+    }
+
+    pub(crate) fn mark_ddl_completed(&mut self) {
+        self.ddl_completed = true;
+        self.updated_at = Utc::now();
+    }
+
+    pub(crate) fn task_status(&self, chunk_id: u32, schema: &str) -> Option<ImportTaskStatus> {
+        self.tasks
+            .iter()
+            .find(|task| task.chunk_id == chunk_id && task.schema == schema)
+            .map(|task| task.status)
+    }
+
+    pub(crate) fn set_task_status(
+        &mut self,
+        chunk_id: u32,
+        schema: &str,
+        status: ImportTaskStatus,
+        error: Option<String>,
+    ) -> Result<()> {
+        let task = self
+            .tasks
+            .iter_mut()
+            .find(|task| task.chunk_id == chunk_id && task.schema == schema)
+            .context(ImportStateUnknownTaskSnafu {
+                chunk_id,
+                schema: schema.to_string(),
+            })?;
+        task.status = status;
+        task.error = error;
+        self.updated_at = Utc::now();
+        Ok(())
+    }
+}
+
+#[derive(Debug)]
+pub(crate) struct ImportStateLockGuard {
+    file: std::fs::File,
+}
+
+impl Drop for ImportStateLockGuard {
+    fn drop(&mut self) {
+        let _ = self.file.unlock();
+    }
+}
+
+pub(crate) fn default_state_path(
+    snapshot_id: &str,
+    target_addr: &str,
+    catalog: &str,
+    schemas: &[String],
+) -> Option<PathBuf> {
+    let home = default_home_dir_with(|key| std::env::var_os(key));
+    let cwd = std::env::current_dir().ok();
+    default_state_path_with(
+        home.as_deref(),
+        cwd.as_deref(),
+        snapshot_id,
+        target_addr,
+        catalog,
+        schemas,
+    )
+}
+
+fn default_home_dir_with<F>(get: F) -> Option<PathBuf>
+where
+    F: Fn(&str) -> Option<std::ffi::OsString>,
+{
+    get("HOME")
+        .or_else(|| get("USERPROFILE"))
+        .map(PathBuf::from)
+        .or_else(|| {
+            let drive = get("HOMEDRIVE")?;
+            let path = get("HOMEPATH")?;
+            Some(PathBuf::from(drive).join(path))
+        })
+}
+
+fn default_state_path_with(
+    home: Option<&Path>,
+    cwd: Option<&Path>,
+    snapshot_id: &str,
+    target_addr: &str,
+    catalog: &str,
+    schemas: &[String],
+) -> Option<PathBuf> {
+    let file_name = import_state_file_name(snapshot_id, target_addr, catalog, schemas);
+    match (home, cwd) {
+        (Some(home), _) => Some(
+            home.join(IMPORT_STATE_ROOT)
+                .join(IMPORT_STATE_DIR)
+                .join(file_name),
+        ),
+        (None, Some(cwd)) => Some(cwd.join(file_name)),
+        (None, None) => None,
+    }
+}
+
+fn import_state_file_name(
+    snapshot_id: &str,
+    target_addr: &str,
+    catalog: &str,
+    schemas: &[String],
+) -> String {
+    format!(
+        ".import_state_{}_{}_{}.json",
+        encode_path_segment(snapshot_id),
+        encode_path_segment(target_addr),
+        import_identity_hash(catalog, schemas)
+    )
+}
+
+pub(crate) fn canonical_schema_selection(schemas: &[String]) -> Vec<String> {
+    let mut canonicalized = schemas
+        .iter()
+        .map(|schema| schema.to_ascii_lowercase())
+        .collect::<Vec<_>>();
+    canonicalized.sort();
+    canonicalized.dedup();
+    canonicalized
+}
+
+/// FNV-1a over `(catalog, schemas)`. The output is part of the persisted state
+/// filename, so we cannot use `std::collections::hash_map::DefaultHasher` -
+/// Rust does not guarantee its algorithm across releases, which would make a
+/// state file written by one toolchain undiscoverable by another.
+fn import_identity_hash(catalog: &str, schemas: &[String]) -> String {
+    const FNV_OFFSET: u64 = 0xcbf29ce484222325;
+    const FNV_PRIME: u64 = 0x100000001b3;
+
+    fn hash_bytes(mut hash: u64, bytes: &[u8]) -> u64 {
+        for byte in bytes {
+            hash ^= u64::from(*byte);
+            hash = hash.wrapping_mul(FNV_PRIME);
+        }
+        hash
+    }
+
+    let mut hash = FNV_OFFSET;
+    hash = hash_bytes(hash, catalog.as_bytes());
+    // 0xff cannot appear in valid UTF-8, so it works as an unambiguous
+    // field separator between adjacent identifiers.
+    hash = hash_bytes(hash, &[0xff]);
+    for schema in canonical_schema_selection(schemas) {
+        hash = hash_bytes(hash, schema.as_bytes());
+        hash = hash_bytes(hash, &[0xff]);
+    }
+    format!("{hash:016x}")
+}
+
+pub(crate) async fn load_import_state(path: &Path) -> Result<Option<ImportState>> {
+    match tokio::fs::read(path).await {
+        Ok(bytes) => {
+            let mut state: ImportState =
+                serde_json::from_slice(&bytes).context(ImportStateParseSnafu)?;
+            normalize_import_state_for_resume(&mut state);
+            Ok(Some(state))
+        }
+        Err(error) if error.kind() == std::io::ErrorKind::NotFound => Ok(None),
+        Err(source) => Err(source).context(ImportStateIoSnafu {
+            path: path.display().to_string(),
+        }),
+    }
+}
+
+/// Caller must hold the lock acquired via `try_acquire_import_state_lock`.
+pub(crate) async fn save_import_state(path: &Path, state: &ImportState) -> Result<()> {
+    if let Some(parent) = path.parent() {
+        tokio::fs::create_dir_all(parent)
+            .await
+            .context(ImportStateIoSnafu {
+                path: parent.display().to_string(),
+            })?;
+    }
+
+    let bytes =
+        serde_json::to_vec_pretty(state).expect("ImportState should always be serializable");
+    let tmp_path = unique_tmp_path(path);
+    let mut file = tokio::fs::File::create(&tmp_path)
+        .await
+        .context(ImportStateIoSnafu {
+            path: tmp_path.display().to_string(),
+        })?;
+    file.write_all(&bytes).await.context(ImportStateIoSnafu {
+        path: tmp_path.display().to_string(),
+    })?;
+    file.sync_all().await.context(ImportStateIoSnafu {
+        path: tmp_path.display().to_string(),
+    })?;
+    // Close before rename; Windows forbids renaming an open file.
+    drop(file);
+
+    tokio::fs::rename(&tmp_path, path)
+        .await
+        .context(ImportStateIoSnafu {
+            path: path.display().to_string(),
+        })?;
+    sync_parent_dir(path).await?;
+    Ok(())
+}
+
+pub(crate) fn try_acquire_import_state_lock(path: &Path) -> Result<ImportStateLockGuard> {
+    if let Some(parent) = path.parent() {
+        std::fs::create_dir_all(parent).context(ImportStateIoSnafu {
+            path: parent.display().to_string(),
+        })?;
+    }
+
+    let lock_path = import_state_lock_path(path);
+    let file = std::fs::OpenOptions::new()
+        .create(true)
+        .read(true)
+        .write(true)
+        .truncate(false)
+        .open(&lock_path)
+        .context(ImportStateIoSnafu {
+            path: lock_path.display().to_string(),
+        })?;
+    file.try_lock_exclusive().map_err(|error| {
+        if is_lock_contention(&error) {
+            ImportStateLockedSnafu {
+                path: lock_path.display().to_string(),
+            }
+            .build()
+        } else {
+            ImportStateIoSnafu {
+                path: lock_path.display().to_string(),
+            }
+            .into_error(error)
+        }
+    })?;
+
+    Ok(ImportStateLockGuard { file })
+}
+
+fn is_lock_contention(error: &std::io::Error) -> bool {
+    error.kind() == std::io::ErrorKind::WouldBlock
+        || error.raw_os_error() == fs2::lock_contended_error().raw_os_error()
+}
+
+fn unique_tmp_path(path: &Path) -> PathBuf {
+    let pid = std::process::id();
+    let seq = IMPORT_STATE_TMP_ID.fetch_add(1, Ordering::Relaxed);
+    let file_name = path.file_name().unwrap_or_default().to_string_lossy();
+    path.with_file_name(format!("{file_name}.{pid}.{seq}.tmp"))
+}
+
+fn import_state_lock_path(path: &Path) -> PathBuf {
+    let file_name = path.file_name().unwrap_or_default().to_string_lossy();
+    path.with_file_name(format!("{file_name}.lock"))
+}
+
+fn normalize_import_state_for_resume(state: &mut ImportState) {
+    for task in &mut state.tasks {
+        if task.status == ImportTaskStatus::InProgress {
+            task.status = ImportTaskStatus::Pending;
+            task.error = None;
+        }
+    }
+}
+
+pub(crate) async fn delete_import_state(path: &Path) -> Result<()> {
+    match tokio::fs::remove_file(path).await {
+        Ok(()) => {
+            sync_parent_dir(path).await?;
+            Ok(())
+        }
+        Err(error) if error.kind() == std::io::ErrorKind::NotFound => Ok(()),
+        Err(source) => Err(source).context(ImportStateIoSnafu {
+            path: path.display().to_string(),
+        }),
+    }
+}
+
+#[cfg(unix)]
+async fn sync_parent_dir(path: &Path) -> Result<()> {
+    let Some(parent) = path.parent() else {
+        return Ok(());
+    };
+
+    let dir = tokio::fs::File::open(parent)
+        .await
+        .context(ImportStateIoSnafu {
+            path: parent.display().to_string(),
+        })?;
+    dir.sync_all().await.context(ImportStateIoSnafu {
+        path: parent.display().to_string(),
+    })?;
+    Ok(())
+}
+
+#[cfg(not(unix))]
+async fn sync_parent_dir(_path: &Path) -> Result<()> {
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use std::process::Command;
+
+    use chrono::Utc;
+    use tempfile::tempdir;
+
+    use super::*;
+
+    const CHILD_LOCK_PATH_ENV: &str = "GREPTIME_IMPORT_STATE_LOCK_PATH";
+    const CHILD_LOCK_TEST: &str =
+        "data::import_v2::state::tests::test_try_acquire_import_state_lock_child_process";
+
+    fn schemas() -> Vec<String> {
+        vec!["public".to_string(), "analytics".to_string()]
+    }
+
+    fn tasks() -> Vec<ImportTaskKey> {
+        vec![
+            ImportTaskKey::new(1, "public"),
+            ImportTaskKey::new(2, "analytics"),
+        ]
+    }
+
+    #[test]
+    fn test_import_state_new_initializes_pending_tasks() {
+        let state = ImportState::new(
+            "snapshot-1",
+            "127.0.0.1:4000",
+            "greptime",
+            &schemas(),
+            tasks(),
+        );
+
+        assert_eq!(state.snapshot_id, "snapshot-1");
+        assert_eq!(state.target_addr, "127.0.0.1:4000");
+        assert_eq!(state.catalog, "greptime");
+        assert_eq!(state.schemas, vec!["analytics", "public"]);
+        assert_eq!(state.tasks.len(), 2);
+        assert_eq!(state.tasks[0].status, ImportTaskStatus::Pending);
+        assert_eq!(state.tasks[1].status, ImportTaskStatus::Pending);
+    }
+
+    #[test]
+    fn test_set_task_status_updates_timestamp_and_error() {
+        let mut state = ImportState::new(
+            "snapshot-1",
+            "127.0.0.1:4000",
+            "greptime",
+            &schemas(),
+            tasks(),
+        );
+        let before = state.updated_at;
+        state.updated_at = Utc::now() - chrono::Duration::seconds(10);
+
+        state
+            .set_task_status(
+                1,
+                "public",
+                ImportTaskStatus::Failed,
+                Some("timeout".to_string()),
+            )
+            .unwrap();
+        assert_eq!(
+            state.task_status(1, "public"),
+            Some(ImportTaskStatus::Failed)
+        );
+        assert_eq!(state.tasks[0].error.as_deref(), Some("timeout"));
+        assert!(state.updated_at > before);
+    }
+
+    #[test]
+    fn test_set_task_status_rejects_unknown_task() {
+        let mut state = ImportState::new(
+            "snapshot-1",
+            "127.0.0.1:4000",
+            "greptime",
+            &schemas(),
+            tasks(),
+        );
+
+        let error = state
+            .set_task_status(99, "public", ImportTaskStatus::Completed, None)
+            .unwrap_err();
+
+        assert!(matches!(
+            error,
+            crate::data::import_v2::error::Error::ImportStateUnknownTask { chunk_id, schema, .. }
+                if chunk_id == 99 && schema == "public"
+        ));
+    }
+
+    #[tokio::test]
+    async fn test_save_and_load_import_state_round_trip() {
+        let dir = tempdir().unwrap();
+        let path = dir.path().join("import_state.json");
+        let mut state = ImportState::new(
+            "snapshot-1",
+            "127.0.0.1:4000",
+            "greptime",
+            &schemas(),
+            tasks(),
+        );
+        state
+            .set_task_status(2, "analytics", ImportTaskStatus::Completed, None)
+            .unwrap();
+
+        save_import_state(&path, &state).await.unwrap();
+        let loaded = load_import_state(&path).await.unwrap().unwrap();
+
+        assert_eq!(loaded.snapshot_id, state.snapshot_id);
+        assert_eq!(loaded.target_addr, state.target_addr);
+        assert_eq!(loaded.catalog, state.catalog);
+        assert_eq!(loaded.schemas, state.schemas);
+        assert_eq!(loaded.tasks, state.tasks);
+    }
+
+    #[tokio::test]
+    async fn test_save_import_state_overwrites_existing_file() {
+        let dir = tempdir().unwrap();
+        let path = dir.path().join("import_state.json");
+        let mut state = ImportState::new(
+            "snapshot-1",
+            "127.0.0.1:4000",
+            "greptime",
+            &schemas(),
+            tasks(),
+        );
+        save_import_state(&path, &state).await.unwrap();
+
+        state
+            .set_task_status(1, "public", ImportTaskStatus::Completed, None)
+            .unwrap();
+        save_import_state(&path, &state).await.unwrap();
+
+        let loaded = load_import_state(&path).await.unwrap().unwrap();
+        assert_eq!(
+            loaded.task_status(1, "public"),
+            Some(ImportTaskStatus::Completed)
+        );
+    }
+
+    #[test]
+    fn test_load_import_state_resets_in_progress_to_pending() {
+        let mut state = ImportState::new(
+            "snapshot-1",
+            "127.0.0.1:4000",
+            "greptime",
+            &schemas(),
+            tasks(),
+        );
+        state
+            .set_task_status(
+                2,
+                "analytics",
+                ImportTaskStatus::InProgress,
+                Some("running".to_string()),
+            )
+            .unwrap();
+
+        normalize_import_state_for_resume(&mut state);
+
+        assert_eq!(
+            state.task_status(1, "public"),
+            Some(ImportTaskStatus::Pending)
+        );
+        assert_eq!(
+            state.task_status(2, "analytics"),
+            Some(ImportTaskStatus::Pending)
+        );
+        assert_eq!(state.tasks[1].error, None);
+    }
+
+    #[test]
+    fn test_unique_tmp_path_generates_distinct_paths() {
+        let dir = tempdir().unwrap();
+        let path = dir.path().join("import_state.json");
+
+        let first = unique_tmp_path(&path);
+        let second = unique_tmp_path(&path);
+
+        assert_ne!(first, second);
+        assert!(first.starts_with(dir.path()));
+        assert!(second.starts_with(dir.path()));
+        assert!(
+            first
+                .file_name()
+                .unwrap()
+                .to_string_lossy()
+                .ends_with(".tmp")
+        );
+        assert!(
+            second
+                .file_name()
+                .unwrap()
+                .to_string_lossy()
+                .ends_with(".tmp")
+        );
+    }
+
+    #[test]
+    fn test_lock_contention_detection_accepts_platform_error() {
+        let error = fs2::lock_contended_error();
+
+        assert!(is_lock_contention(&error));
+    }
+
+    #[test]
+    fn test_try_acquire_import_state_lock_rejects_second_holder() {
+        let dir = tempdir().unwrap();
+        let path = dir.path().join("import_state.json");
+
+        let _first = try_acquire_import_state_lock(&path).unwrap();
+        // Import state locking guards concurrent CLI processes, so validate cross-process exclusion.
+        let output = Command::new(std::env::current_exe().unwrap())
+            .arg(CHILD_LOCK_TEST)
+            .arg("--ignored")
+            .arg("--exact")
+            .env(CHILD_LOCK_PATH_ENV, &path)
+            .output()
+            .unwrap();
+
+        assert!(
+            output.status.success(),
+            "child lock test failed\nstdout:\n{}\nstderr:\n{}",
+            String::from_utf8_lossy(&output.stdout),
+            String::from_utf8_lossy(&output.stderr)
+        );
+        let stdout = String::from_utf8_lossy(&output.stdout);
+        assert!(
+            stdout.contains("1 passed"),
+            "child lock test did not run the expected ignored test\nstdout:\n{stdout}"
+        );
+    }
+
+    #[test]
+    #[ignore = "spawned by test_try_acquire_import_state_lock_rejects_second_holder"]
+    fn test_try_acquire_import_state_lock_child_process() {
+        let path = std::env::var_os(CHILD_LOCK_PATH_ENV)
+            .expect("child lock path must be set by the parent test");
+        let path = PathBuf::from(path);
+        let error = try_acquire_import_state_lock(&path).unwrap_err();
+
+        assert!(matches!(
+            error,
+            crate::data::import_v2::error::Error::ImportStateLocked { .. }
+        ));
+    }
+
+    #[tokio::test]
+    async fn test_delete_import_state_ignores_missing_file() {
+        let dir = tempdir().unwrap();
+        let path = dir.path().join("missing.json");
+
+        delete_import_state(&path).await.unwrap();
+    }
+
+    #[test]
+    fn test_default_state_path_prefers_home_and_encodes_snapshot_id() {
+        let home = tempdir().unwrap();
+        let cwd = tempdir().unwrap();
+
+        let path = default_state_path_with(
+            Some(home.path()),
+            Some(cwd.path()),
+            "../snapshot",
+            "127.0.0.1:4000",
+            "greptime",
+            &schemas(),
+        )
+        .unwrap();
+
+        assert_eq!(
+            path.parent().unwrap(),
+            home.path().join(IMPORT_STATE_ROOT).join(IMPORT_STATE_DIR)
+        );
+        let file_name = path.file_name().unwrap().to_string_lossy();
+        assert!(file_name.starts_with(".import_state_%2E%2E%2Fsnapshot_127%2E0%2E0%2E1%3A4000_"));
+        assert!(file_name.ends_with(".json"));
+    }
+
+    #[test]
+    fn test_default_state_path_falls_back_to_cwd_when_home_missing() {
+        let cwd = tempdir().unwrap();
+
+        let path = default_state_path_with(
+            None,
+            Some(cwd.path()),
+            "snapshot-1",
+            "target-a",
+            "greptime",
+            &schemas(),
+        )
+        .unwrap();
+
+        assert_eq!(path.parent().unwrap(), cwd.path());
+        let file_name = path.file_name().unwrap().to_string_lossy();
+        assert!(file_name.starts_with(".import_state_snapshot-1_target-a_"));
+        assert!(file_name.ends_with(".json"));
+    }
+
+    #[test]
+    fn test_default_state_path_isolated_by_target_addr() {
+        let cwd = tempdir().unwrap();
+
+        let first = default_state_path_with(
+            None,
+            Some(cwd.path()),
+            "snapshot-1",
+            "127.0.0.1:4000",
+            "greptime",
+            &schemas(),
+        )
+        .unwrap();
+        let second = default_state_path_with(
+            None,
+            Some(cwd.path()),
+            "snapshot-1",
+            "127.0.0.1:4001",
+            "greptime",
+            &schemas(),
+        )
+        .unwrap();
+
+        assert_ne!(first, second);
+    }
+
+    #[test]
+    fn test_default_state_path_isolated_by_catalog_and_schemas() {
+        let cwd = tempdir().unwrap();
+        let public_only = vec!["public".to_string()];
+        let analytics_only = vec!["analytics".to_string()];
+
+        let first = default_state_path_with(
+            None,
+            Some(cwd.path()),
+            "snapshot-1",
+            "127.0.0.1:4000",
+            "greptime",
+            &public_only,
+        )
+        .unwrap();
+        let second = default_state_path_with(
+            None,
+            Some(cwd.path()),
+            "snapshot-1",
+            "127.0.0.1:4000",
+            "other",
+            &public_only,
+        )
+        .unwrap();
+        let third = default_state_path_with(
+            None,
+            Some(cwd.path()),
+            "snapshot-1",
+            "127.0.0.1:4000",
+            "greptime",
+            &analytics_only,
+        )
+        .unwrap();
+
+        assert_ne!(first, second);
+        assert_ne!(first, third);
+    }
+
+    #[test]
+    fn test_default_home_dir_prefers_home() {
+        let detected = default_home_dir_with(|key| match key {
+            "HOME" => Some(std::ffi::OsString::from("/tmp/home")),
+            "USERPROFILE" => Some(std::ffi::OsString::from("/tmp/userprofile")),
+            _ => None,
+        });
+
+        assert_eq!(detected, Some(PathBuf::from("/tmp/home")));
+    }
+
+    #[test]
+    fn test_default_home_dir_falls_back_to_userprofile() {
+        let detected = default_home_dir_with(|key| match key {
+            "USERPROFILE" => Some(std::ffi::OsString::from("/tmp/userprofile")),
+            _ => None,
+        });
+
+        assert_eq!(detected, Some(PathBuf::from("/tmp/userprofile")));
+    }
+
+    #[test]
+    fn test_default_home_dir_falls_back_to_home_drive_and_path() {
+        let detected = default_home_dir_with(|key| match key {
+            "HOMEDRIVE" => Some(std::ffi::OsString::from("/tmp")),
+            "HOMEPATH" => Some(std::ffi::OsString::from("windows-home")),
+            _ => None,
+        });
+
+        assert_eq!(detected, Some(PathBuf::from("/tmp").join("windows-home")));
+    }
+}
diff --git a/src/cli/src/data/path.rs b/src/cli/src/data/path.rs
new file mode 100644
index 0000000000..2df81f62c8
--- /dev/null
+++ b/src/cli/src/data/path.rs
@@ -0,0 +1,89 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Shared path helpers for export/import data files.
+
+use crate::data::export_v2::schema::{DDL_DIR, SCHEMA_DIR};
+
+pub(crate) fn ddl_path_for_schema(schema: &str) -> String {
+    format!(
+        "{}/{}/{}.sql",
+        SCHEMA_DIR,
+        DDL_DIR,
+        encode_path_segment(schema)
+    )
+}
+
+pub(crate) fn data_dir_for_schema_chunk(schema: &str, chunk_id: u32) -> String {
+    format!("data/{}/{}/", encode_path_segment(schema), chunk_id)
+}
+
+pub(crate) fn encode_path_segment(value: &str) -> String {
+    let mut encoded = String::with_capacity(value.len());
+    for byte in value.bytes() {
+        match byte {
+            b'A'..=b'Z' | b'a'..=b'z' | b'0'..=b'9' | b'-' | b'_' => {
+                encoded.push(byte as char);
+            }
+            _ => {
+                encoded.push('%');
+                encoded.push(hex_char(byte >> 4));
+                encoded.push(hex_char(byte & 0x0F));
+            }
+        }
+    }
+    encoded
+}
+
+fn hex_char(nibble: u8) -> char {
+    match nibble {
+        0..=9 => (b'0' + nibble) as char,
+        10..=15 => (b'A' + (nibble - 10)) as char,
+        _ => unreachable!("nibble must be in 0..=15"),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_encode_path_segment_preserves_safe_ascii() {
+        assert_eq!(encode_path_segment("test_db"), "test_db");
+    }
+
+    #[test]
+    fn test_encode_path_segment_escapes_path_traversal_chars() {
+        assert_eq!(encode_path_segment("../evil"), "%2E%2E%2Fevil");
+        assert_eq!(encode_path_segment(r"..\\evil"), "%2E%2E%5C%5Cevil");
+    }
+
+    #[test]
+    fn test_ddl_path_for_schema_encodes_schema_segment() {
+        assert_eq!(ddl_path_for_schema("public"), "schema/ddl/public.sql");
+        assert_eq!(
+            ddl_path_for_schema("../evil"),
+            "schema/ddl/%2E%2E%2Fevil.sql"
+        );
+    }
+
+    #[test]
+    fn test_data_dir_for_schema_chunk_encodes_schema_segment() {
+        assert_eq!(data_dir_for_schema_chunk("public", 1), "data/public/1/");
+        assert_eq!(
+            data_dir_for_schema_chunk("../evil", 7),
+            "data/%2E%2E%2Fevil/7/"
+        );
+    }
+}
diff --git a/src/cli/src/data/snapshot_storage.rs b/src/cli/src/data/snapshot_storage.rs
new file mode 100644
index 0000000000..2ee07b5586
--- /dev/null
+++ b/src/cli/src/data/snapshot_storage.rs
@@ -0,0 +1,863 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Storage abstraction for Export/Import V2.
+//!
+//! This module provides a unified interface for reading and writing snapshot data
+//! to various storage backends (S3, OSS, GCS, Azure Blob, local filesystem).
+
+use std::collections::BTreeSet;
+
+use async_trait::async_trait;
+use futures::TryStreamExt;
+use object_store::services::{Azblob, Fs, Gcs, Oss, S3};
+use object_store::util::{with_instrument_layers, with_retry_layers};
+use object_store::{
+    AzblobConnection, ErrorKind, GcsConnection, ObjectStore, OssConnection, S3Connection,
+};
+use snafu::ResultExt;
+use url::Url;
+
+use crate::common::ObjectStoreConfig;
+use crate::data::export_v2::error::{
+    BuildObjectStoreSnafu, InvalidUriSnafu, ManifestParseSnafu, ManifestSerializeSnafu, Result,
+    SnapshotNotFoundSnafu, StorageOperationSnafu, TextDecodeSnafu, UnsupportedSchemeSnafu,
+    UrlParseSnafu,
+};
+use crate::data::export_v2::manifest::{MANIFEST_FILE, Manifest};
+#[cfg(test)]
+use crate::data::export_v2::schema::SchemaDefinition;
+use crate::data::export_v2::schema::{SCHEMA_DIR, SCHEMAS_FILE, SchemaSnapshot};
+
+struct RemoteLocation {
+    bucket_or_container: String,
+    root: String,
+}
+
+/// URI schemes supported for snapshot storage.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum StorageScheme {
+    /// Amazon S3.
+    S3,
+    /// Alibaba Cloud OSS.
+    Oss,
+    /// Google Cloud Storage.
+    Gcs,
+    /// Azure Blob Storage.
+    Azblob,
+    /// Local filesystem (file://).
+    File,
+}
+
+impl StorageScheme {
+    /// Parses storage scheme from URI.
+    pub fn from_uri(uri: &str) -> Result<Self> {
+        let url = Url::parse(uri).context(UrlParseSnafu)?;
+
+        match url.scheme() {
+            "s3" => Ok(Self::S3),
+            "oss" => Ok(Self::Oss),
+            "gs" | "gcs" => Ok(Self::Gcs),
+            "azblob" => Ok(Self::Azblob),
+            "file" => Ok(Self::File),
+            scheme => UnsupportedSchemeSnafu { scheme }.fail(),
+        }
+    }
+}
+
+/// Extracts bucket/container and root path from a URI.
+fn extract_remote_location_with_root_policy(
+    uri: &str,
+    allow_empty_root: bool,
+) -> Result<RemoteLocation> {
+    let url = Url::parse(uri).context(UrlParseSnafu)?;
+    let bucket_or_container = url.host_str().unwrap_or("").to_string();
+    if bucket_or_container.is_empty() {
+        return InvalidUriSnafu {
+            uri,
+            reason: "URI must include bucket/container in host",
+        }
+        .fail();
+    }
+
+    let root = url.path().trim_start_matches('/').to_string();
+    if root.is_empty() && !allow_empty_root {
+        return InvalidUriSnafu {
+            uri,
+            reason: "snapshot URI must include a non-empty path after the bucket/container",
+        }
+        .fail();
+    }
+
+    Ok(RemoteLocation {
+        bucket_or_container,
+        root,
+    })
+}
+
+/// Validates that a URI has a proper scheme.
+///
+/// Rejects bare paths (e.g., `/tmp/backup`, `./backup`) because:
+/// - Schema export (CLI) and data export (server) run in different processes
+/// - Using bare paths would split the snapshot across machines
+///
+/// Supported URI schemes:
+/// - `s3://bucket/path` - Amazon S3
+/// - `oss://bucket/path` - Alibaba Cloud OSS
+/// - `gs://bucket/path` - Google Cloud Storage
+/// - `azblob://container/path` - Azure Blob Storage
+/// - `file:///absolute/path` - Local filesystem
+pub fn validate_uri(uri: &str) -> Result<StorageScheme> {
+    // Must have a scheme
+    if !uri.contains("://") {
+        return InvalidUriSnafu {
+            uri,
+            reason: "URI must have a scheme (e.g., s3://, file://). Bare paths are not supported.",
+        }
+        .fail();
+    }
+
+    StorageScheme::from_uri(uri)
+}
+
+fn schema_index_path() -> String {
+    format!("{}/{}", SCHEMA_DIR, SCHEMAS_FILE)
+}
+
+/// Extracts the absolute filesystem path from a file:// URI.
+fn extract_file_path_from_uri(uri: &str) -> Result<String> {
+    let url = Url::parse(uri).context(UrlParseSnafu)?;
+
+    match url.host_str() {
+        Some(host) if !host.is_empty() && host != "localhost" => InvalidUriSnafu {
+            uri,
+            reason: "file:// URI must use an absolute path like file:///tmp/backup",
+        }
+        .fail(),
+        _ => url
+            .to_file_path()
+            .map_err(|_| {
+                InvalidUriSnafu {
+                    uri,
+                    reason: "file:// URI must use an absolute path like file:///tmp/backup",
+                }
+                .build()
+            })
+            .map(|path| path.to_string_lossy().into_owned()),
+    }
+}
+
+async fn ensure_snapshot_exists(storage: &OpenDalStorage) -> Result<()> {
+    if storage.exists().await? {
+        Ok(())
+    } else {
+        SnapshotNotFoundSnafu {
+            uri: storage.target_uri.as_str(),
+        }
+        .fail()
+    }
+}
+
+/// Snapshot storage abstraction.
+///
+/// Provides operations for reading and writing snapshot data to various storage backends.
+#[async_trait]
+pub trait SnapshotStorage: Send + Sync {
+    /// Checks if a snapshot exists at this location (manifest.json exists).
+    async fn exists(&self) -> Result<bool>;
+
+    /// Reads the manifest file.
+    async fn read_manifest(&self) -> Result<Manifest>;
+
+    /// Writes the manifest file.
+    async fn write_manifest(&self, manifest: &Manifest) -> Result<()>;
+
+    /// Writes the schema index to schema/schemas.json.
+    async fn write_schema(&self, schema: &SchemaSnapshot) -> Result<()>;
+
+    /// Writes a text file to a relative path under the snapshot root.
+    async fn write_text(&self, path: &str, content: &str) -> Result<()>;
+
+    /// Reads a text file from a relative path under the snapshot root.
+    async fn read_text(&self, path: &str) -> Result<String>;
+
+    /// Creates a directory-like prefix under the snapshot root when needed by the backend.
+    async fn create_dir_all(&self, path: &str) -> Result<()>;
+
+    /// Lists files recursively under a relative prefix.
+    async fn list_files_recursive(&self, prefix: &str) -> Result<Vec<String>>;
+
+    /// Deletes the entire snapshot (for --force).
+    async fn delete_snapshot(&self) -> Result<()>;
+}
+
+/// OpenDAL-based implementation of SnapshotStorage.
+pub struct OpenDalStorage {
+    object_store: ObjectStore,
+    target_uri: String,
+}
+
+impl OpenDalStorage {
+    fn new_operator_rooted(object_store: ObjectStore, target_uri: &str) -> Self {
+        Self {
+            object_store,
+            target_uri: target_uri.to_string(),
+        }
+    }
+
+    fn finish_local_store(object_store: ObjectStore) -> ObjectStore {
+        with_instrument_layers(object_store, false)
+    }
+
+    fn finish_remote_store(object_store: ObjectStore) -> ObjectStore {
+        with_instrument_layers(with_retry_layers(object_store), false)
+    }
+
+    fn ensure_backend_enabled(uri: &str, enabled: bool, reason: &'static str) -> Result<()> {
+        if enabled {
+            Ok(())
+        } else {
+            InvalidUriSnafu { uri, reason }.fail()
+        }
+    }
+
+    fn validate_remote_config<E: std::fmt::Display>(
+        uri: &str,
+        backend: &str,
+        result: std::result::Result<(), E>,
+    ) -> Result<()> {
+        result.map_err(|error| {
+            InvalidUriSnafu {
+                uri,
+                reason: format!("invalid {} config: {}", backend, error),
+            }
+            .build()
+        })
+    }
+
+    /// Creates a new storage from a file:// URI.
+    pub fn from_file_uri(uri: &str) -> Result<Self> {
+        let path = extract_file_path_from_uri(uri)?;
+
+        let builder = Fs::default().root(&path);
+        let object_store = ObjectStore::new(builder)
+            .context(BuildObjectStoreSnafu)?
+            .finish();
+        Ok(Self::new_operator_rooted(
+            Self::finish_local_store(object_store),
+            uri,
+        ))
+    }
+
+    fn from_file_uri_with_config(uri: &str, storage: &ObjectStoreConfig) -> Result<Self> {
+        if storage.enable_s3 || storage.enable_oss || storage.enable_gcs || storage.enable_azblob {
+            return InvalidUriSnafu {
+                uri,
+                reason: "file:// cannot be used with remote storage flags",
+            }
+            .fail();
+        }
+
+        Self::from_file_uri(uri)
+    }
+
+    fn from_s3_uri(uri: &str, storage: &ObjectStoreConfig) -> Result<Self> {
+        Self::from_s3_uri_with_root_policy(uri, storage, false)
+    }
+
+    fn from_s3_uri_with_root_policy(
+        uri: &str,
+        storage: &ObjectStoreConfig,
+        allow_empty_root: bool,
+    ) -> Result<Self> {
+        Self::ensure_backend_enabled(
+            uri,
+            storage.enable_s3,
+            "s3:// requires --s3 and related options",
+        )?;
+
+        let location = extract_remote_location_with_root_policy(uri, allow_empty_root)?;
+        let mut config = storage.s3.clone();
+        config.s3_bucket = location.bucket_or_container;
+        config.s3_root = location.root;
+        Self::validate_remote_config(uri, "s3", config.validate())?;
+
+        let conn: S3Connection = config.into();
+        let object_store = ObjectStore::new(S3::from(&conn))
+            .context(BuildObjectStoreSnafu)?
+            .finish();
+        Ok(Self::new_operator_rooted(
+            Self::finish_remote_store(object_store),
+            uri,
+        ))
+    }
+
+    fn from_oss_uri(uri: &str, storage: &ObjectStoreConfig) -> Result<Self> {
+        Self::from_oss_uri_with_root_policy(uri, storage, false)
+    }
+
+    fn from_oss_uri_with_root_policy(
+        uri: &str,
+        storage: &ObjectStoreConfig,
+        allow_empty_root: bool,
+    ) -> Result<Self> {
+        Self::ensure_backend_enabled(
+            uri,
+            storage.enable_oss,
+            "oss:// requires --oss and related options",
+        )?;
+
+        let location = extract_remote_location_with_root_policy(uri, allow_empty_root)?;
+        let mut config = storage.oss.clone();
+        config.oss_bucket = location.bucket_or_container;
+        config.oss_root = location.root;
+        Self::validate_remote_config(uri, "oss", config.validate())?;
+
+        let conn: OssConnection = config.into();
+        let object_store = ObjectStore::new(Oss::from(&conn))
+            .context(BuildObjectStoreSnafu)?
+            .finish();
+        Ok(Self::new_operator_rooted(
+            Self::finish_remote_store(object_store),
+            uri,
+        ))
+    }
+
+    fn from_gcs_uri(uri: &str, storage: &ObjectStoreConfig) -> Result<Self> {
+        Self::from_gcs_uri_with_root_policy(uri, storage, false)
+    }
+
+    fn from_gcs_uri_with_root_policy(
+        uri: &str,
+        storage: &ObjectStoreConfig,
+        allow_empty_root: bool,
+    ) -> Result<Self> {
+        Self::ensure_backend_enabled(
+            uri,
+            storage.enable_gcs,
+            "gs:// or gcs:// requires --gcs and related options",
+        )?;
+
+        let location = extract_remote_location_with_root_policy(uri, allow_empty_root)?;
+        let mut config = storage.gcs.clone();
+        config.gcs_bucket = location.bucket_or_container;
+        config.gcs_root = location.root;
+        // GCS validate() rejects empty root, unlike S3/OSS/Azblob.
+        if allow_empty_root && config.gcs_root.is_empty() {
+            Self::validate_gcs_parent_config(uri, &config)?;
+        } else {
+            Self::validate_remote_config(uri, "gcs", config.validate())?;
+        }
+
+        let conn: GcsConnection = config.into();
+        let object_store = ObjectStore::new(Gcs::from(&conn))
+            .context(BuildObjectStoreSnafu)?
+            .finish();
+        Ok(Self::new_operator_rooted(
+            Self::finish_remote_store(object_store),
+            uri,
+        ))
+    }
+
+    fn validate_gcs_parent_config(
+        uri: &str,
+        config: &crate::common::PrefixedGcsConnection,
+    ) -> Result<()> {
+        if config.gcs_bucket.is_empty() {
+            return InvalidUriSnafu {
+                uri,
+                reason: "invalid gcs config: GCS bucket must be set when --gcs is enabled.",
+            }
+            .fail();
+        }
+        if config.gcs_scope.is_empty() {
+            return InvalidUriSnafu {
+                uri,
+                reason: "invalid gcs config: GCS scope must be set when --gcs is enabled.",
+            }
+            .fail();
+        }
+        Ok(())
+    }
+
+    fn from_azblob_uri(uri: &str, storage: &ObjectStoreConfig) -> Result<Self> {
+        Self::from_azblob_uri_with_root_policy(uri, storage, false)
+    }
+
+    fn from_azblob_uri_with_root_policy(
+        uri: &str,
+        storage: &ObjectStoreConfig,
+        allow_empty_root: bool,
+    ) -> Result<Self> {
+        Self::ensure_backend_enabled(
+            uri,
+            storage.enable_azblob,
+            "azblob:// requires --azblob and related options",
+        )?;
+
+        let location = extract_remote_location_with_root_policy(uri, allow_empty_root)?;
+        let mut config = storage.azblob.clone();
+        config.azblob_container = location.bucket_or_container;
+        config.azblob_root = location.root;
+        Self::validate_remote_config(uri, "azblob", config.validate())?;
+
+        let conn: AzblobConnection = config.into();
+        let object_store = ObjectStore::new(Azblob::from(&conn))
+            .context(BuildObjectStoreSnafu)?
+            .finish();
+        Ok(Self::new_operator_rooted(
+            Self::finish_remote_store(object_store),
+            uri,
+        ))
+    }
+
+    /// Creates a new storage from a URI and object store config.
+    pub fn from_uri(uri: &str, storage: &ObjectStoreConfig) -> Result<Self> {
+        match StorageScheme::from_uri(uri)? {
+            StorageScheme::File => Self::from_file_uri_with_config(uri, storage),
+            StorageScheme::S3 => Self::from_s3_uri(uri, storage),
+            StorageScheme::Oss => Self::from_oss_uri(uri, storage),
+            StorageScheme::Gcs => Self::from_gcs_uri(uri, storage),
+            StorageScheme::Azblob => Self::from_azblob_uri(uri, storage),
+        }
+    }
+
+    /// Creates storage rooted at a snapshot parent URI.
+    ///
+    /// Parent-oriented commands such as `export-v2 list` may scan bucket/container
+    /// roots. Snapshot-oriented commands must keep using `from_uri`, which rejects
+    /// empty remote roots to avoid unsafe snapshot operations at bucket scope.
+    pub fn from_parent_uri(uri: &str, storage: &ObjectStoreConfig) -> Result<Self> {
+        match StorageScheme::from_uri(uri)? {
+            StorageScheme::File => Self::from_file_uri_with_config(uri, storage),
+            StorageScheme::S3 => Self::from_s3_uri_with_root_policy(uri, storage, true),
+            StorageScheme::Oss => Self::from_oss_uri_with_root_policy(uri, storage, true),
+            StorageScheme::Gcs => Self::from_gcs_uri_with_root_policy(uri, storage, true),
+            StorageScheme::Azblob => Self::from_azblob_uri_with_root_policy(uri, storage, true),
+        }
+    }
+
+    /// Reads a file as bytes.
+    async fn read_file(&self, path: &str) -> Result<Vec<u8>> {
+        let data = self
+            .object_store
+            .read(path)
+            .await
+            .context(StorageOperationSnafu {
+                operation: format!("read {}", path),
+            })?;
+        Ok(data.to_vec())
+    }
+
+    /// Reads a file as bytes if it exists.
+    pub(crate) async fn read_file_if_exists(&self, path: &str) -> Result<Option<Vec<u8>>> {
+        match self.object_store.read(path).await {
+            Ok(data) => Ok(Some(data.to_vec())),
+            Err(error) if error.kind() == ErrorKind::NotFound => Ok(None),
+            Err(error) => Err(error).context(StorageOperationSnafu {
+                operation: format!("read {}", path),
+            }),
+        }
+    }
+
+    /// Writes bytes to a file.
+    async fn write_file(&self, path: &str, data: Vec<u8>) -> Result<()> {
+        self.object_store
+            .write(path, data)
+            .await
+            .map(|_| ())
+            .context(StorageOperationSnafu {
+                operation: format!("write {}", path),
+            })
+    }
+
+    /// Checks if a file exists using stat.
+    async fn file_exists(&self, path: &str) -> Result<bool> {
+        match self.object_store.stat(path).await {
+            Ok(_) => Ok(true),
+            Err(e) if e.kind() == object_store::ErrorKind::NotFound => Ok(false),
+            Err(e) => Err(e).context(StorageOperationSnafu {
+                operation: format!("check exists {}", path),
+            }),
+        }
+    }
+
+    /// Lists direct child directory names under the storage root.
+    pub(crate) async fn list_direct_child_dirs(&self) -> Result<Vec<String>> {
+        let mut lister = match self.object_store.lister_with("/").recursive(false).await {
+            Ok(lister) => lister,
+            Err(error) if error.kind() == ErrorKind::NotFound => return Ok(Vec::new()),
+            Err(error) => {
+                return Err(error).context(StorageOperationSnafu {
+                    operation: "list /",
+                });
+            }
+        };
+
+        let mut dirs = BTreeSet::new();
+        while let Some(entry) = lister.try_next().await.context(StorageOperationSnafu {
+            operation: "list /",
+        })? {
+            let path = entry.path().trim_matches('/');
+            if path.is_empty() {
+                continue;
+            }
+
+            if entry.metadata().is_dir()
+                && let Some(name) = path.split('/').next()
+            {
+                dirs.insert(name.to_string());
+            }
+        }
+
+        Ok(dirs.into_iter().collect())
+    }
+
+    #[cfg(test)]
+    pub async fn read_schema(&self) -> Result<SchemaSnapshot> {
+        let schemas_path = schema_index_path();
+        let schemas: Vec<SchemaDefinition> = if self.file_exists(&schemas_path).await? {
+            let data = self.read_file(&schemas_path).await?;
+            serde_json::from_slice(&data).context(ManifestParseSnafu)?
+        } else {
+            vec![]
+        };
+
+        Ok(SchemaSnapshot { schemas })
+    }
+}
+
+#[async_trait]
+impl SnapshotStorage for OpenDalStorage {
+    async fn exists(&self) -> Result<bool> {
+        self.file_exists(MANIFEST_FILE).await
+    }
+
+    async fn read_manifest(&self) -> Result<Manifest> {
+        ensure_snapshot_exists(self).await?;
+
+        let data = self.read_file(MANIFEST_FILE).await?;
+        serde_json::from_slice(&data).context(ManifestParseSnafu)
+    }
+
+    async fn write_manifest(&self, manifest: &Manifest) -> Result<()> {
+        let data = serde_json::to_vec_pretty(manifest).context(ManifestSerializeSnafu)?;
+        self.write_file(MANIFEST_FILE, data).await
+    }
+
+    async fn write_schema(&self, schema: &SchemaSnapshot) -> Result<()> {
+        let schemas_path = schema_index_path();
+        let schemas_data =
+            serde_json::to_vec_pretty(&schema.schemas).context(ManifestSerializeSnafu)?;
+        self.write_file(&schemas_path, schemas_data).await
+    }
+
+    async fn write_text(&self, path: &str, content: &str) -> Result<()> {
+        self.write_file(path, content.as_bytes().to_vec()).await
+    }
+
+    async fn read_text(&self, path: &str) -> Result<String> {
+        let data = self.read_file(path).await?;
+        String::from_utf8(data).context(TextDecodeSnafu)
+    }
+
+    async fn create_dir_all(&self, path: &str) -> Result<()> {
+        self.object_store
+            .create_dir(path)
+            .await
+            .context(StorageOperationSnafu {
+                operation: format!("create dir {}", path),
+            })
+    }
+
+    async fn list_files_recursive(&self, prefix: &str) -> Result<Vec<String>> {
+        let mut lister = match self.object_store.lister_with(prefix).recursive(true).await {
+            Ok(lister) => lister,
+            Err(error) if error.kind() == ErrorKind::NotFound => return Ok(Vec::new()),
+            Err(error) => {
+                return Err(error).context(StorageOperationSnafu {
+                    operation: format!("list {}", prefix),
+                });
+            }
+        };
+
+        let mut files = Vec::new();
+        while let Some(entry) = lister.try_next().await.context(StorageOperationSnafu {
+            operation: format!("list {}", prefix),
+        })? {
+            if entry.metadata().is_dir() {
+                continue;
+            }
+            files.push(entry.path().to_string());
+        }
+        Ok(files)
+    }
+
+    async fn delete_snapshot(&self) -> Result<()> {
+        self.object_store
+            .delete_with("/")
+            .recursive(true)
+            .await
+            .context(StorageOperationSnafu {
+                operation: "delete snapshot",
+            })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+    use std::path::Path;
+
+    use object_store::ObjectStore;
+    use object_store::services::Fs;
+    use tempfile::tempdir;
+    use url::Url;
+
+    use super::*;
+    use crate::data::export_v2::manifest::{DataFormat, TimeRange};
+    use crate::data::export_v2::schema::SchemaDefinition;
+
+    fn make_storage_with_rooted_fs(dir: &std::path::Path) -> OpenDalStorage {
+        let object_store = ObjectStore::new(Fs::default().root(dir.to_str().unwrap()))
+            .unwrap()
+            .finish();
+        OpenDalStorage::new_operator_rooted(
+            OpenDalStorage::finish_local_store(object_store),
+            Url::from_directory_path(dir).unwrap().as_ref(),
+        )
+    }
+
+    #[test]
+    fn test_validate_uri_valid() {
+        assert_eq!(validate_uri("s3://bucket/path").unwrap(), StorageScheme::S3);
+        assert_eq!(
+            validate_uri("oss://bucket/path").unwrap(),
+            StorageScheme::Oss
+        );
+        assert_eq!(
+            validate_uri("gs://bucket/path").unwrap(),
+            StorageScheme::Gcs
+        );
+        assert_eq!(
+            validate_uri("gcs://bucket/path").unwrap(),
+            StorageScheme::Gcs
+        );
+        assert_eq!(
+            validate_uri("azblob://container/path").unwrap(),
+            StorageScheme::Azblob
+        );
+        assert_eq!(
+            validate_uri("file:///tmp/backup").unwrap(),
+            StorageScheme::File
+        );
+    }
+
+    #[test]
+    fn test_validate_uri_invalid() {
+        // Bare paths should be rejected
+        assert!(validate_uri("/tmp/backup").is_err());
+        assert!(validate_uri("./backup").is_err());
+        assert!(validate_uri("backup").is_err());
+
+        // Unknown schemes
+        assert!(validate_uri("ftp://server/path").is_err());
+    }
+
+    #[test]
+    fn test_extract_remote_location_requires_non_empty_root() {
+        assert!(extract_remote_location_with_root_policy("s3://bucket", false).is_err());
+        assert!(extract_remote_location_with_root_policy("s3://bucket/", false).is_err());
+        assert!(extract_remote_location_with_root_policy("oss://bucket", false).is_err());
+        assert!(extract_remote_location_with_root_policy("gs://bucket", false).is_err());
+        assert!(extract_remote_location_with_root_policy("azblob://container", false).is_err());
+    }
+
+    #[test]
+    fn test_extract_remote_location_allows_empty_root_when_permitted() {
+        let location = extract_remote_location_with_root_policy("s3://bucket", true).unwrap();
+        assert_eq!(location.bucket_or_container, "bucket");
+        assert_eq!(location.root, "");
+
+        let location =
+            extract_remote_location_with_root_policy("azblob://container/", true).unwrap();
+        assert_eq!(location.bucket_or_container, "container");
+        assert_eq!(location.root, "");
+    }
+
+    #[test]
+    fn test_parent_storage_allows_s3_bucket_root() {
+        let mut storage = ObjectStoreConfig {
+            enable_s3: true,
+            ..Default::default()
+        };
+        storage.s3.s3_region = Some("us-east-1".to_string());
+
+        assert!(OpenDalStorage::from_uri("s3://bucket", &storage).is_err());
+        assert!(OpenDalStorage::from_parent_uri("s3://bucket", &storage).is_ok());
+    }
+
+    #[cfg(not(windows))]
+    #[test]
+    fn test_extract_path_from_uri_unix_examples() {
+        assert_eq!(
+            extract_file_path_from_uri("file:///tmp/backup").unwrap(),
+            "/tmp/backup"
+        );
+        assert_eq!(
+            extract_file_path_from_uri("file://localhost/tmp/backup").unwrap(),
+            "/tmp/backup"
+        );
+        assert_eq!(
+            extract_file_path_from_uri("file:///tmp/my%20backup").unwrap(),
+            "/tmp/my backup"
+        );
+        assert_eq!(
+            extract_file_path_from_uri("file://localhost/tmp/my%20backup").unwrap(),
+            "/tmp/my backup"
+        );
+    }
+
+    #[test]
+    fn test_extract_file_path_from_uri_rejects_file_host() {
+        assert!(extract_file_path_from_uri("file://tmp/backup").is_err());
+    }
+
+    #[test]
+    fn test_extract_file_path_from_uri_round_trips_directory_url() {
+        let dir = tempdir().unwrap();
+        let uri = Url::from_directory_path(dir.path()).unwrap().to_string();
+        let path = extract_file_path_from_uri(&uri).unwrap();
+
+        assert_eq!(Path::new(&path), dir.path());
+    }
+
+    #[tokio::test]
+    async fn test_read_manifest_reports_requested_uri() {
+        let dir = tempdir().unwrap();
+        let uri = Url::from_directory_path(dir.path()).unwrap().to_string();
+        let storage = OpenDalStorage::from_file_uri(&uri).unwrap();
+
+        let error = storage.read_manifest().await.unwrap_err().to_string();
+
+        assert!(error.contains(uri.as_str()));
+    }
+
+    #[tokio::test]
+    async fn test_manifest_round_trip() {
+        let dir = tempdir().unwrap();
+        let storage = make_storage_with_rooted_fs(dir.path());
+
+        let manifest = Manifest::new_full(
+            "greptime".to_string(),
+            vec!["public".to_string()],
+            TimeRange::unbounded(),
+            DataFormat::Parquet,
+        );
+
+        storage.write_manifest(&manifest).await.unwrap();
+        let loaded = storage.read_manifest().await.unwrap();
+
+        assert_eq!(loaded.catalog, manifest.catalog);
+        assert_eq!(loaded.schemas, manifest.schemas);
+        assert_eq!(loaded.schema_only, manifest.schema_only);
+        assert_eq!(loaded.format, manifest.format);
+        assert_eq!(loaded.snapshot_id, manifest.snapshot_id);
+    }
+
+    #[tokio::test]
+    async fn test_schema_round_trip() {
+        let dir = tempdir().unwrap();
+        let storage = make_storage_with_rooted_fs(dir.path());
+
+        let mut snapshot = SchemaSnapshot::new();
+        snapshot.add_schema(SchemaDefinition {
+            catalog: "greptime".to_string(),
+            name: "test_db".to_string(),
+            options: HashMap::from([("ttl".to_string(), "7d".to_string())]),
+        });
+
+        storage.write_schema(&snapshot).await.unwrap();
+        let loaded = storage.read_schema().await.unwrap();
+
+        assert_eq!(loaded, snapshot);
+    }
+
+    #[tokio::test]
+    async fn test_text_round_trip() {
+        let dir = tempdir().unwrap();
+        let storage = make_storage_with_rooted_fs(dir.path());
+        let content = "CREATE TABLE metrics (ts TIMESTAMP TIME INDEX);";
+
+        storage
+            .write_text("schema/ddl/public.sql", content)
+            .await
+            .unwrap();
+        let loaded = storage.read_text("schema/ddl/public.sql").await.unwrap();
+
+        assert_eq!(loaded, content);
+    }
+
+    #[tokio::test]
+    async fn test_read_text_rejects_invalid_utf8() {
+        let dir = tempdir().unwrap();
+        let storage = make_storage_with_rooted_fs(dir.path());
+
+        storage
+            .write_file("schema/ddl/public.sql", vec![0xff, 0xfe, 0xfd])
+            .await
+            .unwrap();
+
+        let error = storage
+            .read_text("schema/ddl/public.sql")
+            .await
+            .unwrap_err();
+        assert!(error.to_string().contains("UTF-8"));
+    }
+
+    #[tokio::test]
+    async fn test_exists_follows_manifest_presence() {
+        let dir = tempdir().unwrap();
+        let storage = make_storage_with_rooted_fs(dir.path());
+
+        assert!(!storage.exists().await.unwrap());
+
+        storage
+            .write_manifest(&Manifest::new_schema_only(
+                "greptime".to_string(),
+                vec!["public".to_string()],
+            ))
+            .await
+            .unwrap();
+
+        assert!(storage.exists().await.unwrap());
+    }
+
+    #[tokio::test]
+    async fn test_delete_snapshot_only_removes_rooted_contents() {
+        let parent = tempdir().unwrap();
+        let snapshot_root = parent.path().join("snapshot");
+        let sibling = parent.path().join("sibling");
+        std::fs::create_dir_all(&snapshot_root).unwrap();
+        std::fs::create_dir_all(&sibling).unwrap();
+        std::fs::write(snapshot_root.join("manifest.json"), b"{}").unwrap();
+        std::fs::write(sibling.join("keep.txt"), b"keep").unwrap();
+
+        let storage = make_storage_with_rooted_fs(&snapshot_root);
+        storage.delete_snapshot().await.unwrap();
+
+        assert!(!snapshot_root.join("manifest.json").exists());
+        assert!(sibling.join("keep.txt").exists());
+    }
+}
diff --git a/src/cli/src/data/sql.rs b/src/cli/src/data/sql.rs
new file mode 100644
index 0000000000..7de4206b26
--- /dev/null
+++ b/src/cli/src/data/sql.rs
@@ -0,0 +1,40 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Shared SQL escaping helpers for CLI-generated statements.
+
+pub(crate) fn escape_sql_literal(value: &str) -> String {
+    value.replace('\'', "''")
+}
+
+pub(crate) fn escape_sql_identifier(value: &str) -> String {
+    value.replace('"', "\"\"")
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_escape_sql_literal_escapes_single_quotes() {
+        assert_eq!(escape_sql_literal("test_db"), "test_db");
+        assert_eq!(escape_sql_literal("te'st"), "te''st");
+    }
+
+    #[test]
+    fn test_escape_sql_identifier_escapes_double_quotes() {
+        assert_eq!(escape_sql_identifier("test_db"), "test_db");
+        assert_eq!(escape_sql_identifier(r#"te"st"#), r#"te""st"#);
+    }
+}
diff --git a/src/cli/src/database.rs b/src/cli/src/database.rs
index db98c38e38..cba27129dd 100644
--- a/src/cli/src/database.rs
+++ b/src/cli/src/database.rs
@@ -36,6 +36,7 @@ pub struct DatabaseClient {
     auth_header: Option<String>,
     timeout: Duration,
     proxy: Option<reqwest::Proxy>,
+    no_proxy: bool,
 }
 
 pub fn parse_proxy_opts(
@@ -61,6 +62,7 @@ impl DatabaseClient {
         auth_basic: Option<String>,
         timeout: Duration,
         proxy: Option<reqwest::Proxy>,
+        no_proxy: bool,
     ) -> Self {
         let auth_header = if let Some(basic) = auth_basic {
             let encoded = general_purpose::STANDARD.encode(basic);
@@ -69,7 +71,9 @@ impl DatabaseClient {
             None
         };
 
-        if let Some(ref proxy) = proxy {
+        if no_proxy {
+            common_telemetry::info!("Proxy disabled");
+        } else if let Some(ref proxy) = proxy {
             common_telemetry::info!("Using proxy: {:?}", proxy);
         } else {
             common_telemetry::info!("Using system proxy(if any)");
@@ -81,9 +85,14 @@ impl DatabaseClient {
             auth_header,
             timeout,
             proxy,
+            no_proxy,
         }
     }
 
+    pub fn addr(&self) -> &str {
+        &self.addr
+    }
+
     pub async fn sql_in_public(&self, sql: &str) -> Result<Option<Vec<Vec<Value>>>> {
         self.sql(sql, DEFAULT_SCHEMA_NAME).await
     }
@@ -95,12 +104,14 @@ impl DatabaseClient {
             ("db", format!("{}-{}", self.catalog, schema)),
             ("sql", sql.to_string()),
         ];
-        let client = self
-            .proxy
-            .clone()
-            .map(|proxy| reqwest::Client::builder().proxy(proxy).build())
-            .unwrap_or_else(|| Ok(reqwest::Client::new()))
-            .context(BuildClientSnafu)?;
+        let mut builder = reqwest::Client::builder();
+        if let Some(proxy) = self.proxy.clone() {
+            builder = builder.proxy(proxy);
+        }
+        if self.no_proxy {
+            builder = builder.no_proxy();
+        }
+        let client = builder.build().context(BuildClientSnafu)?;
         let mut request = client
             .post(&url)
             .form(&params)
diff --git a/src/cli/src/lib.rs b/src/cli/src/lib.rs
index acf5df4086..4305da9c8f 100644
--- a/src/cli/src/lib.rs
+++ b/src/cli/src/lib.rs
@@ -29,7 +29,7 @@ pub use database::DatabaseClient;
 use error::Result;
 
 pub use crate::bench::BenchTableMetadataCommand;
-pub use crate::data::DataCommand;
+pub use crate::data::{DataCommand, export_v2, import_v2};
 pub use crate::metadata::MetadataCommand;
 
 #[async_trait]
diff --git a/src/cli/src/metadata.rs b/src/cli/src/metadata.rs
index 5800293cb6..beb62a2deb 100644
--- a/src/cli/src/metadata.rs
+++ b/src/cli/src/metadata.rs
@@ -21,7 +21,7 @@ use clap::Subcommand;
 use common_error::ext::BoxedError;
 
 use crate::Tool;
-use crate::metadata::control::{DelCommand, GetCommand};
+use crate::metadata::control::{DelCommand, GetCommand, PutCommand};
 use crate::metadata::repair::RepairCommand;
 use crate::metadata::snapshot::SnapshotCommand;
 
@@ -37,6 +37,8 @@ pub enum MetadataCommand {
     #[clap(subcommand)]
     Del(DelCommand),
     #[clap(subcommand)]
+    Put(PutCommand),
+    #[clap(subcommand)]
     Repair(RepairCommand),
 }
 
@@ -47,6 +49,7 @@ impl MetadataCommand {
             MetadataCommand::Repair(cmd) => cmd.build().await,
             MetadataCommand::Get(cmd) => cmd.build().await,
             MetadataCommand::Del(cmd) => cmd.build().await,
+            MetadataCommand::Put(cmd) => cmd.build().await,
         }
     }
 }
diff --git a/src/cli/src/metadata/control.rs b/src/cli/src/metadata/control.rs
index 3fba582ce5..eaf5b167b5 100644
--- a/src/cli/src/metadata/control.rs
+++ b/src/cli/src/metadata/control.rs
@@ -14,9 +14,12 @@
 
 mod del;
 mod get;
+mod put;
+mod selector;
 #[cfg(test)]
 mod test_utils;
 mod utils;
 
 pub(crate) use del::DelCommand;
 pub(crate) use get::GetCommand;
+pub(crate) use put::PutCommand;
diff --git a/src/cli/src/metadata/control/del/table.rs b/src/cli/src/metadata/control/del/table.rs
index d1346928fe..c92e573224 100644
--- a/src/cli/src/metadata/control/del/table.rs
+++ b/src/cli/src/metadata/control/del/table.rs
@@ -14,111 +14,59 @@
 
 use async_trait::async_trait;
 use clap::Parser;
-use client::{DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME};
-use common_catalog::format_full_table_name;
 use common_error::ext::BoxedError;
 use common_meta::ddl::utils::get_region_wal_options;
 use common_meta::key::TableMetadataManager;
-use common_meta::key::table_name::TableNameManager;
 use common_meta::kv_backend::KvBackendRef;
 use store_api::storage::TableId;
 
 use crate::Tool;
 use crate::common::StoreConfig;
-use crate::error::{InvalidArgumentsSnafu, TableNotFoundSnafu};
+use crate::error::TableNotFoundSnafu;
 use crate::metadata::control::del::CLI_TOMBSTONE_PREFIX;
-use crate::metadata::control::utils::get_table_id_by_name;
+use crate::metadata::control::selector::TableSelector;
 
 /// Delete table metadata logically from the metadata store.
 #[derive(Debug, Default, Parser)]
 pub struct DelTableCommand {
-    /// The table id to delete from the metadata store.
-    #[clap(long)]
-    table_id: Option<u32>,
-
-    /// The table name to delete from the metadata store.
-    #[clap(long)]
-    table_name: Option<String>,
-
-    /// The schema name of the table.
-    #[clap(long, default_value = DEFAULT_SCHEMA_NAME)]
-    schema_name: String,
-
-    /// The catalog name of the table.
-    #[clap(long, default_value = DEFAULT_CATALOG_NAME)]
-    catalog_name: String,
+    #[clap(flatten)]
+    selector: TableSelector,
 
     /// The store config.
     #[clap(flatten)]
     store: StoreConfig,
 }
 
-impl DelTableCommand {
-    fn validate(&self) -> Result<(), BoxedError> {
-        if matches!(
-            (&self.table_id, &self.table_name),
-            (Some(_), Some(_)) | (None, None)
-        ) {
-            return Err(BoxedError::new(
-                InvalidArgumentsSnafu {
-                    msg: "You must specify either --table-id or --table-name.",
-                }
-                .build(),
-            ));
-        }
-        Ok(())
-    }
-}
-
 impl DelTableCommand {
     pub async fn build(&self) -> Result<Box<dyn Tool>, BoxedError> {
-        self.validate()?;
+        self.selector.validate()?;
         let kv_backend = self.store.build().await?;
         Ok(Box::new(DelTableTool {
-            table_id: self.table_id,
-            table_name: self.table_name.clone(),
-            schema_name: self.schema_name.clone(),
-            catalog_name: self.catalog_name.clone(),
-            table_name_manager: TableNameManager::new(kv_backend.clone()),
+            selector: self.selector.clone(),
             table_metadata_deleter: TableMetadataDeleter::new(kv_backend),
         }))
     }
 }
 
 struct DelTableTool {
-    table_id: Option<u32>,
-    table_name: Option<String>,
-    schema_name: String,
-    catalog_name: String,
-    table_name_manager: TableNameManager,
+    selector: TableSelector,
     table_metadata_deleter: TableMetadataDeleter,
 }
 
 #[async_trait]
 impl Tool for DelTableTool {
     async fn do_work(&self) -> Result<(), BoxedError> {
-        let table_id = if let Some(table_name) = &self.table_name {
-            let catalog_name = &self.catalog_name;
-            let schema_name = &self.schema_name;
-
-            let Some(table_id) = get_table_id_by_name(
-                &self.table_name_manager,
-                catalog_name,
-                schema_name,
-                table_name,
+        let Some(table_id) = self
+            .selector
+            .resolve_table_id(
+                self.table_metadata_deleter
+                    .table_metadata_manager
+                    .table_name_manager(),
             )
             .await?
-            else {
-                println!(
-                    "Table({}) not found",
-                    format_full_table_name(catalog_name, schema_name, table_name)
-                );
-                return Ok(());
-            };
-            table_id
-        } else {
-            // Safety: we have validated that table_id or table_name is not None
-            self.table_id.unwrap()
+        else {
+            println!("Table({}) not found", self.selector.formatted_table_name());
+            return Ok(());
         };
         self.table_metadata_deleter.delete(table_id).await?;
         println!("Table({}) deleted", table_id);
@@ -182,6 +130,7 @@ mod tests {
     use std::collections::HashMap;
     use std::sync::Arc;
 
+    use clap::Parser;
     use common_error::ext::ErrorExt;
     use common_error::status_code::StatusCode;
     use common_meta::key::TableMetadataManager;
@@ -192,9 +141,83 @@ mod tests {
     use common_meta::rpc::store::RangeRequest;
 
     use crate::metadata::control::del::CLI_TOMBSTONE_PREFIX;
-    use crate::metadata::control::del::table::TableMetadataDeleter;
+    use crate::metadata::control::del::table::{DelTableCommand, TableMetadataDeleter};
     use crate::metadata::control::test_utils::prepare_physical_table_metadata;
 
+    #[tokio::test]
+    async fn test_del_table_selector_requires_single_target() {
+        let command = DelTableCommand::parse_from([
+            "table",
+            "--backend",
+            "memory-store",
+            "--store-addrs",
+            "memory://",
+        ]);
+
+        let err = match command.build().await {
+            Ok(_) => panic!("expected validation failure"),
+            Err(err) => err,
+        };
+        assert!(
+            err.output_msg()
+                .contains("You must specify either --table-id or --table-name.")
+        );
+    }
+
+    #[tokio::test]
+    async fn test_del_table_selector_rejects_both_targets() {
+        let command = DelTableCommand::parse_from([
+            "table",
+            "--table-id",
+            "1024",
+            "--table-name",
+            "my_table",
+            "--backend",
+            "memory-store",
+            "--store-addrs",
+            "memory://",
+        ]);
+
+        let err = match command.build().await {
+            Ok(_) => panic!("expected validation failure"),
+            Err(err) => err,
+        };
+        assert!(
+            err.output_msg()
+                .contains("You must specify either --table-id or --table-name.")
+        );
+    }
+
+    #[tokio::test]
+    async fn test_del_table_command_builds_tool_with_table_id() {
+        let command = DelTableCommand::parse_from([
+            "table",
+            "--table-id",
+            "1024",
+            "--backend",
+            "memory-store",
+            "--store-addrs",
+            "memory://",
+        ]);
+
+        let _tool = command.build().await.unwrap();
+    }
+
+    #[tokio::test]
+    async fn test_del_table_command_builds_tool_with_table_name() {
+        let command = DelTableCommand::parse_from([
+            "table",
+            "--table-name",
+            "my_table",
+            "--backend",
+            "memory-store",
+            "--store-addrs",
+            "memory://",
+        ]);
+
+        let _tool = command.build().await.unwrap();
+    }
+
     #[tokio::test]
     async fn test_delete_table_not_found() {
         let kv_backend = Arc::new(MemoryKvBackend::new()) as KvBackendRef;
diff --git a/src/cli/src/metadata/control/get.rs b/src/cli/src/metadata/control/get.rs
index 1de2bb370e..8a1de07acf 100644
--- a/src/cli/src/metadata/control/get.rs
+++ b/src/cli/src/metadata/control/get.rs
@@ -16,8 +16,6 @@ use std::cmp::min;
 
 use async_trait::async_trait;
 use clap::{Parser, Subcommand};
-use client::{DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME};
-use common_catalog::format_full_table_name;
 use common_error::ext::BoxedError;
 use common_meta::key::TableMetadataManager;
 use common_meta::key::table_info::TableInfoKey;
@@ -29,8 +27,8 @@ use futures::TryStreamExt;
 
 use crate::Tool;
 use crate::common::StoreConfig;
-use crate::error::InvalidArgumentsSnafu;
-use crate::metadata::control::utils::{decode_key_value, get_table_id_by_name, json_formatter};
+use crate::metadata::control::selector::TableSelector;
+use crate::metadata::control::utils::{decode_key_value, json_formatter};
 
 /// Getting metadata from metadata store.
 #[derive(Subcommand)]
@@ -120,21 +118,8 @@ impl Tool for GetKeyTool {
 /// Get table metadata from the metadata store via table id.
 #[derive(Debug, Default, Parser)]
 pub struct GetTableCommand {
-    /// Get table metadata by table id.
-    #[clap(long)]
-    table_id: Option<u32>,
-
-    /// Get table metadata by table name.
-    #[clap(long)]
-    table_name: Option<String>,
-
-    /// The schema name of the table.
-    #[clap(long, default_value = DEFAULT_SCHEMA_NAME)]
-    schema_name: String,
-
-    /// The catalog name of the table.
-    #[clap(long, default_value = DEFAULT_CATALOG_NAME)]
-    catalog_name: String,
+    #[clap(flatten)]
+    selector: TableSelector,
 
     /// Pretty print the output.
     #[clap(long, default_value = "false")]
@@ -144,29 +129,9 @@ pub struct GetTableCommand {
     store: StoreConfig,
 }
 
-impl GetTableCommand {
-    pub fn validate(&self) -> Result<(), BoxedError> {
-        if matches!(
-            (&self.table_id, &self.table_name),
-            (Some(_), Some(_)) | (None, None)
-        ) {
-            return Err(BoxedError::new(
-                InvalidArgumentsSnafu {
-                    msg: "You must specify either --table-id or --table-name.",
-                }
-                .build(),
-            ));
-        }
-        Ok(())
-    }
-}
-
 struct GetTableTool {
     kvbackend: KvBackendRef,
-    table_id: Option<u32>,
-    table_name: Option<String>,
-    schema_name: String,
-    catalog_name: String,
+    selector: TableSelector,
     pretty: bool,
 }
 
@@ -178,24 +143,9 @@ impl Tool for GetTableTool {
         let table_info_manager = table_metadata_manager.table_info_manager();
         let table_route_manager = table_metadata_manager.table_route_manager();
 
-        let table_id = if let Some(table_name) = &self.table_name {
-            let catalog_name = &self.catalog_name;
-            let schema_name = &self.schema_name;
-
-            let Some(table_id) =
-                get_table_id_by_name(table_name_manager, catalog_name, schema_name, table_name)
-                    .await?
-            else {
-                println!(
-                    "Table({}) not found",
-                    format_full_table_name(catalog_name, schema_name, table_name)
-                );
-                return Ok(());
-            };
-            table_id
-        } else {
-            // Safety: we have validated that table_id or table_name is not None
-            self.table_id.unwrap()
+        let Some(table_id) = self.selector.resolve_table_id(table_name_manager).await? else {
+            println!("Table({}) not found", self.selector.formatted_table_name());
+            return Ok(());
         };
 
         let table_info = table_info_manager
@@ -233,15 +183,94 @@ impl Tool for GetTableTool {
 
 impl GetTableCommand {
     pub async fn build(&self) -> Result<Box<dyn Tool>, BoxedError> {
-        self.validate()?;
+        self.selector.validate()?;
         let kvbackend = self.store.build().await?;
         Ok(Box::new(GetTableTool {
             kvbackend,
-            table_id: self.table_id,
-            table_name: self.table_name.clone(),
-            schema_name: self.schema_name.clone(),
-            catalog_name: self.catalog_name.clone(),
+            selector: self.selector.clone(),
             pretty: self.pretty,
         }))
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use clap::Parser;
+    use common_error::ext::ErrorExt;
+
+    use super::GetTableCommand;
+
+    #[tokio::test]
+    async fn test_get_table_selector_requires_single_target() {
+        let command = GetTableCommand::parse_from([
+            "table",
+            "--backend",
+            "memory-store",
+            "--store-addrs",
+            "memory://",
+        ]);
+
+        let err = match command.build().await {
+            Ok(_) => panic!("expected validation failure"),
+            Err(err) => err,
+        };
+        assert!(
+            err.output_msg()
+                .contains("You must specify either --table-id or --table-name.")
+        );
+    }
+
+    #[tokio::test]
+    async fn test_get_table_selector_rejects_both_targets() {
+        let command = GetTableCommand::parse_from([
+            "table",
+            "--table-id",
+            "1024",
+            "--table-name",
+            "my_table",
+            "--backend",
+            "memory-store",
+            "--store-addrs",
+            "memory://",
+        ]);
+
+        let err = match command.build().await {
+            Ok(_) => panic!("expected validation failure"),
+            Err(err) => err,
+        };
+        assert!(
+            err.output_msg()
+                .contains("You must specify either --table-id or --table-name.")
+        );
+    }
+
+    #[tokio::test]
+    async fn test_get_table_command_builds_tool_with_table_id() {
+        let command = GetTableCommand::parse_from([
+            "table",
+            "--table-id",
+            "1024",
+            "--backend",
+            "memory-store",
+            "--store-addrs",
+            "memory://",
+        ]);
+
+        let _tool = command.build().await.unwrap();
+    }
+
+    #[tokio::test]
+    async fn test_get_table_command_builds_tool_with_table_name() {
+        let command = GetTableCommand::parse_from([
+            "table",
+            "--table-name",
+            "my_table",
+            "--backend",
+            "memory-store",
+            "--store-addrs",
+            "memory://",
+        ]);
+
+        let _tool = command.build().await.unwrap();
+    }
+}
diff --git a/src/cli/src/metadata/control/put.rs b/src/cli/src/metadata/control/put.rs
new file mode 100644
index 0000000000..c2c535c57a
--- /dev/null
+++ b/src/cli/src/metadata/control/put.rs
@@ -0,0 +1,56 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+mod key;
+mod table;
+
+use clap::Subcommand;
+use common_error::ext::BoxedError;
+use snafu::ResultExt;
+use tokio::io::{AsyncRead, AsyncReadExt};
+
+use crate::Tool;
+use crate::error::FileIoSnafu;
+use crate::metadata::control::put::key::PutKeyCommand;
+use crate::metadata::control::put::table::PutTableCommand;
+
+pub(crate) async fn read_value<R>(mut reader: R) -> Result<Vec<u8>, BoxedError>
+where
+    R: AsyncRead + Unpin,
+{
+    let mut value = Vec::new();
+    reader
+        .read_to_end(&mut value)
+        .await
+        .context(FileIoSnafu)
+        .map_err(BoxedError::new)?;
+    Ok(value)
+}
+
+/// Subcommand for putting metadata into the metadata store.
+#[derive(Subcommand)]
+pub enum PutCommand {
+    Key(PutKeyCommand),
+    #[clap(subcommand)]
+    Table(PutTableCommand),
+}
+
+impl PutCommand {
+    pub async fn build(&self) -> Result<Box<dyn Tool>, BoxedError> {
+        match self {
+            PutCommand::Key(cmd) => cmd.build().await,
+            PutCommand::Table(cmd) => cmd.build().await,
+        }
+    }
+}
diff --git a/src/cli/src/metadata/control/put/key.rs b/src/cli/src/metadata/control/put/key.rs
new file mode 100644
index 0000000000..7becfd72dc
--- /dev/null
+++ b/src/cli/src/metadata/control/put/key.rs
@@ -0,0 +1,444 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use async_trait::async_trait;
+use clap::Parser;
+use common_error::ext::BoxedError;
+use common_meta::key::catalog_name::{CatalogNameKey, CatalogNameValue};
+use common_meta::key::flow::flow_state::FlowStateValue;
+use common_meta::key::flow::{
+    flow_info_key_prefix, flow_name_key_prefix, flow_route_key_prefix, flow_state_full_key,
+    flownode_flow_key_prefix, table_flow_key_prefix,
+};
+use common_meta::key::node_address::{NodeAddressKey, NodeAddressValue};
+use common_meta::key::schema_name::{SchemaNameKey, SchemaNameValue};
+use common_meta::key::table_repart::{TableRepartKey, TableRepartValue};
+use common_meta::key::topic_name::{TopicNameKey, TopicNameValue};
+use common_meta::key::topic_region::{TopicRegionKey, TopicRegionValue};
+use common_meta::key::view_info::{ViewInfoKey, ViewInfoValue};
+use common_meta::key::{
+    CATALOG_NAME_KEY_PREFIX, DATANODE_TABLE_KEY_PREFIX, KAFKA_TOPIC_KEY_PREFIX, MetadataKey,
+    MetadataValue, NODE_ADDRESS_PREFIX, SCHEMA_NAME_KEY_PREFIX, TABLE_INFO_KEY_PREFIX,
+    TABLE_NAME_KEY_PREFIX, TABLE_REPART_PREFIX, TABLE_ROUTE_PREFIX, TOPIC_REGION_PREFIX,
+    VIEW_INFO_KEY_PREFIX,
+};
+use common_meta::kv_backend::KvBackendRef;
+use common_meta::rpc::store::PutRequest;
+
+use crate::Tool;
+use crate::common::StoreConfig;
+use crate::error::InvalidArgumentsSnafu;
+use crate::metadata::control::put::read_value;
+
+/// Put a key-value pair into the metadata store.
+#[derive(Debug, Default, Parser)]
+pub struct PutKeyCommand {
+    /// The key to put into the metadata store.
+    key: String,
+
+    /// Read the value to put into the metadata store from standard input.
+    #[clap(long, required = true)]
+    value_stdin: bool,
+
+    /// Skip metadata validation before writing.
+    #[clap(long)]
+    no_validate: bool,
+
+    #[clap(flatten)]
+    store: StoreConfig,
+}
+
+impl PutKeyCommand {
+    pub async fn build(&self) -> Result<Box<dyn Tool>, BoxedError> {
+        let kv_backend = self.store.build().await?;
+        self.build_tool(tokio::io::stdin(), kv_backend).await
+    }
+
+    async fn build_tool<R>(
+        &self,
+        reader: R,
+        kv_backend: KvBackendRef,
+    ) -> Result<Box<dyn Tool>, BoxedError>
+    where
+        R: tokio::io::AsyncRead + Unpin,
+    {
+        Ok(Box::new(PutKeyTool {
+            kv_backend,
+            key: self.key.clone(),
+            value: read_value(reader).await?,
+            no_validate: self.no_validate,
+        }))
+    }
+}
+
+struct PutKeyTool {
+    kv_backend: KvBackendRef,
+    key: String,
+    value: Vec<u8>,
+    no_validate: bool,
+}
+
+#[async_trait]
+impl Tool for PutKeyTool {
+    async fn do_work(&self) -> Result<(), BoxedError> {
+        if !self.no_validate {
+            validate_metadata_value(&self.key, &self.value)?;
+        }
+
+        let request = PutRequest::new()
+            .with_key(self.key.as_bytes())
+            .with_value(self.value.clone());
+        self.kv_backend
+            .put(request)
+            .await
+            .map_err(BoxedError::new)?;
+
+        println!("Key({}) updated", self.key);
+        Ok(())
+    }
+}
+
+fn validate_metadata_value(key: &str, value: &[u8]) -> Result<(), BoxedError> {
+    if let Some(reason) = unsupported_direct_put_reason(key) {
+        return Err(BoxedError::new(
+            InvalidArgumentsSnafu {
+                msg: format!("{reason}, use --no-validate to bypass"),
+            }
+            .build(),
+        ));
+    }
+
+    if key == flow_state_full_key() {
+        validate_value(key, value, FlowStateValue::try_from_raw_value)?;
+        return Ok(());
+    } else if matches_key_prefix(key, VIEW_INFO_KEY_PREFIX) {
+        validate_key(ViewInfoKey::from_bytes(key.as_bytes()), key)?;
+        validate_value(key, value, ViewInfoValue::try_from_raw_value)?;
+        return Ok(());
+    } else if matches_key_prefix(key, CATALOG_NAME_KEY_PREFIX) {
+        validate_key(CatalogNameKey::from_bytes(key.as_bytes()), key)?;
+        validate_value(key, value, CatalogNameValue::try_from_raw_value)?;
+        return Ok(());
+    } else if matches_key_prefix(key, SCHEMA_NAME_KEY_PREFIX) {
+        validate_key(SchemaNameKey::from_bytes(key.as_bytes()), key)?;
+        validate_value(key, value, SchemaNameValue::try_from_raw_value)?;
+        return Ok(());
+    } else if matches_key_prefix(key, TABLE_REPART_PREFIX) {
+        validate_key(TableRepartKey::from_bytes(key.as_bytes()), key)?;
+        validate_value(key, value, TableRepartValue::try_from_raw_value)?;
+        return Ok(());
+    } else if matches_key_prefix(key, NODE_ADDRESS_PREFIX) {
+        validate_key(NodeAddressKey::from_bytes(key.as_bytes()), key)?;
+        validate_value(key, value, NodeAddressValue::try_from_raw_value)?;
+        return Ok(());
+    } else if matches_key_prefix(key, KAFKA_TOPIC_KEY_PREFIX) {
+        validate_key(TopicNameKey::from_bytes(key.as_bytes()), key)?;
+        validate_value(key, value, TopicNameValue::try_from_raw_value)?;
+        return Ok(());
+    } else if matches_key_prefix(key, TOPIC_REGION_PREFIX) {
+        validate_key(TopicRegionKey::from_bytes(key.as_bytes()), key)?;
+        validate_value(key, value, TopicRegionValue::try_from_raw_value)?;
+        return Ok(());
+    }
+
+    Err(BoxedError::new(
+        InvalidArgumentsSnafu {
+            msg: format!(
+                "Unsupported metadata key for validation: {key}, use --no-validate to bypass"
+            ),
+        }
+        .build(),
+    ))
+}
+
+/// Returns the rejection reason for keys that should not be updated by `put key`.
+///
+/// These keys may be decodable, but they are not safe to update via raw KV writes.
+/// `__table_route/*` is the canonical example.
+fn unsupported_direct_put_reason(key: &str) -> Option<String> {
+    let flow_info_prefix = flow_info_key_prefix();
+    let flow_name_prefix = flow_name_key_prefix();
+    let flow_route_prefix = flow_route_key_prefix();
+    let table_flow_prefix = table_flow_key_prefix();
+    let flownode_flow_prefix = flownode_flow_key_prefix();
+
+    let (prefix, target) = [
+        (TABLE_ROUTE_PREFIX, "table route metadata"),
+        (TABLE_INFO_KEY_PREFIX, "table info metadata"),
+        (TABLE_NAME_KEY_PREFIX, "table name metadata"),
+        (DATANODE_TABLE_KEY_PREFIX, "datanode table metadata"),
+        (&flow_info_prefix, "flow info metadata"),
+        (&flow_name_prefix, "flow name metadata"),
+        (&flow_route_prefix, "flow route metadata"),
+        (&table_flow_prefix, "flow source table metadata"),
+        (&flownode_flow_prefix, "flownode flow metadata"),
+    ]
+    .into_iter()
+    .find(|(prefix, _)| matches_key_prefix(key, prefix))?;
+
+    Some(format!(
+        "Direct put is not supported for {target} ({prefix}*); use a dedicated metadata update interface instead"
+    ))
+}
+
+fn matches_key_prefix(key: &str, prefix: &str) -> bool {
+    key == prefix
+        || key
+            .strip_prefix(prefix)
+            .is_some_and(|rest| rest.starts_with('/'))
+}
+
+fn validate_value<T, F>(key: &str, value: &[u8], parser: F) -> Result<(), BoxedError>
+where
+    F: FnOnce(&[u8]) -> common_meta::error::Result<T>,
+{
+    parser(value).map_err(|e| {
+        BoxedError::new(
+            InvalidArgumentsSnafu {
+                msg: format!("Invalid metadata value for key: {key}: {e}"),
+            }
+            .build(),
+        )
+    })?;
+    Ok(())
+}
+
+fn validate_key<T>(result: common_meta::error::Result<T>, key: &str) -> Result<(), BoxedError> {
+    result.map_err(|e| {
+        BoxedError::new(
+            InvalidArgumentsSnafu {
+                msg: format!("Invalid metadata key: {key}: {e}"),
+            }
+            .build(),
+        )
+    })?;
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::BTreeMap;
+    use std::sync::Arc;
+
+    use clap::Parser;
+    use common_error::ext::{BoxedError, ErrorExt};
+    use common_meta::key::flow::flow_state::FlowStateValue;
+    use common_meta::key::flow::flow_state_full_key;
+    use common_meta::key::schema_name::SchemaNameValue;
+    use common_meta::key::topic_name::TopicNameValue;
+    use common_meta::key::{KAFKA_TOPIC_KEY_PREFIX, MetadataValue, SCHEMA_NAME_KEY_PREFIX};
+    use common_meta::kv_backend::KvBackendRef;
+    use common_meta::kv_backend::memory::MemoryKvBackend;
+    use tokio::io::BufReader;
+
+    use super::{
+        PutKeyCommand, PutKeyTool, TABLE_ROUTE_PREFIX, matches_key_prefix,
+        unsupported_direct_put_reason, validate_metadata_value,
+    };
+    use crate::Tool;
+
+    impl PutKeyCommand {
+        async fn build_for_test<R>(
+            &self,
+            reader: R,
+            kv_backend: KvBackendRef,
+        ) -> Result<Box<dyn Tool>, BoxedError>
+        where
+            R: tokio::io::AsyncRead + Unpin,
+        {
+            self.build_tool(reader, kv_backend).await
+        }
+    }
+
+    #[test]
+    fn test_validate_supported_key_success() {
+        let value = SchemaNameValue::default().try_as_raw_value().unwrap();
+
+        validate_metadata_value(&format!("{SCHEMA_NAME_KEY_PREFIX}/greptime/public"), &value)
+            .unwrap();
+    }
+
+    #[test]
+    fn test_validate_supported_key_invalid_value() {
+        let err = validate_metadata_value(
+            &format!("{KAFKA_TOPIC_KEY_PREFIX}/test-topic"),
+            b"not-a-valid-json-value",
+        )
+        .unwrap_err();
+
+        assert!(err.output_msg().contains("Invalid metadata value for key"));
+    }
+
+    #[test]
+    fn test_validate_complex_key_fails() {
+        let value = serde_json::to_vec(&BTreeMap::<u32, u32>::new()).unwrap();
+        let err =
+            validate_metadata_value(&format!("{TABLE_ROUTE_PREFIX}/1024"), &value).unwrap_err();
+
+        assert!(
+            err.output_msg()
+                .contains("Direct put is not supported for table route metadata")
+        );
+    }
+
+    #[test]
+    fn test_validate_unknown_key_fails() {
+        let err = validate_metadata_value("__unknown/foo", b"{}").unwrap_err();
+
+        assert!(
+            err.output_msg()
+                .contains("Unsupported metadata key for validation")
+        );
+    }
+
+    #[test]
+    fn test_validate_invalid_supported_key_fails() {
+        let value = SchemaNameValue::default().try_as_raw_value().unwrap();
+        let err = validate_metadata_value("__schema_name/greptime", &value).unwrap_err();
+
+        assert!(
+            err.output_msg()
+                .contains("Invalid metadata key: __schema_name/greptime")
+        );
+    }
+
+    #[test]
+    fn test_unsupported_direct_put_reason_covers_complex_keys() {
+        let cases = [
+            "__table_route/1024",
+            "__table_info/1024",
+            "__table_name/greptime/public/demo",
+            "__dn_table/1/1024",
+            "__flow/route/1/1",
+        ];
+
+        for key in cases {
+            assert!(unsupported_direct_put_reason(key).is_some(), "key: {key}");
+        }
+    }
+
+    #[test]
+    fn test_matches_key_prefix() {
+        assert!(matches_key_prefix("__table_route", "__table_route"));
+        assert!(matches_key_prefix("__table_route/1024", "__table_route"));
+        assert!(!matches_key_prefix(
+            "__table_route_extra/1024",
+            "__table_route"
+        ));
+        assert!(!matches_key_prefix("__table_routex", "__table_route"));
+        assert!(!matches_key_prefix(
+            "__topic_name/kafka_backup/foo",
+            "__topic_name/kafka"
+        ));
+    }
+
+    #[test]
+    fn test_validate_exact_flow_state_key() {
+        let value = FlowStateValue::new(BTreeMap::new(), BTreeMap::new())
+            .try_as_raw_value()
+            .unwrap();
+
+        validate_metadata_value(&flow_state_full_key(), &value).unwrap();
+    }
+
+    #[tokio::test]
+    async fn test_put_key_tool_writes_supported_key() {
+        let kv_backend = Arc::new(MemoryKvBackend::new()) as KvBackendRef;
+        let value = TopicNameValue::new(42).try_as_raw_value().unwrap();
+        let key = format!("{KAFKA_TOPIC_KEY_PREFIX}/test-topic");
+        let tool = PutKeyTool {
+            kv_backend: kv_backend.clone(),
+            key: key.clone(),
+            value: value.clone(),
+            no_validate: false,
+        };
+
+        tool.do_work().await.unwrap();
+
+        let stored = kv_backend.get(key.as_bytes()).await.unwrap().unwrap();
+        assert_eq!(stored.value, value);
+    }
+
+    #[tokio::test]
+    async fn test_put_key_tool_bypasses_validation() {
+        let kv_backend = Arc::new(MemoryKvBackend::new()) as KvBackendRef;
+        let key = format!("{TABLE_ROUTE_PREFIX}/1024");
+        let value = b"not-json".to_vec();
+        let tool = PutKeyTool {
+            kv_backend: kv_backend.clone(),
+            key: key.clone(),
+            value: value.clone(),
+            no_validate: true,
+        };
+
+        tool.do_work().await.unwrap();
+
+        let stored = kv_backend.get(key.as_bytes()).await.unwrap().unwrap();
+        assert_eq!(stored.value, value);
+    }
+
+    #[test]
+    fn test_put_key_command_requires_value_stdin() {
+        let err = PutKeyCommand::try_parse_from([
+            "key",
+            "__topic_name/kafka/test-cli-topic",
+            "--backend",
+            "memory-store",
+            "--store-addrs",
+            "memory://",
+        ])
+        .unwrap_err();
+
+        assert_eq!(err.kind(), clap::error::ErrorKind::MissingRequiredArgument);
+    }
+
+    #[tokio::test]
+    async fn test_put_key_command_builds_tool_with_stdin() {
+        let value = TopicNameValue::new(7).try_as_raw_value().unwrap();
+        let command = PutKeyCommand::parse_from([
+            "key",
+            "__topic_name/kafka/test-cli-topic",
+            "--value-stdin",
+            "--backend",
+            "memory-store",
+            "--store-addrs",
+            "memory://",
+        ]);
+
+        let tool = command
+            .build_for_test(
+                BufReader::new(value.as_slice()),
+                Arc::new(MemoryKvBackend::new()) as KvBackendRef,
+            )
+            .await
+            .unwrap();
+        tool.do_work().await.unwrap();
+    }
+
+    #[tokio::test]
+    async fn test_put_key_command_validate_failure() {
+        let tool = PutKeyTool {
+            kv_backend: Arc::new(MemoryKvBackend::new()) as KvBackendRef,
+            key: "__table_route/1024".to_string(),
+            value: b"{}".to_vec(),
+            no_validate: false,
+        };
+        let err = tool.do_work().await.unwrap_err();
+
+        assert!(
+            err.output_msg()
+                .contains("Direct put is not supported for table route metadata")
+        );
+    }
+}
diff --git a/src/cli/src/metadata/control/put/table.rs b/src/cli/src/metadata/control/put/table.rs
new file mode 100644
index 0000000000..29e5c546d5
--- /dev/null
+++ b/src/cli/src/metadata/control/put/table.rs
@@ -0,0 +1,687 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::HashSet;
+
+use async_trait::async_trait;
+use clap::{Parser, Subcommand};
+use common_error::ext::BoxedError;
+use common_meta::key::datanode_table::{DatanodeTableKey, RegionInfo};
+use common_meta::key::table_info::TableInfoValue;
+use common_meta::key::table_route::TableRouteValue;
+use common_meta::key::{
+    DeserializedValueWithBytes, MetadataValue, RegionDistribution, TableMetadataManager,
+};
+use common_meta::kv_backend::KvBackendRef;
+use common_meta::rpc::router::{RegionRoute, region_distribution};
+use snafu::{OptionExt, ensure};
+use store_api::storage::TableId;
+use table::metadata::TableInfo;
+
+use crate::Tool;
+use crate::common::StoreConfig;
+use crate::error::{Error, InvalidArgumentsSnafu, TableNotFoundSnafu, UnexpectedSnafu};
+use crate::metadata::control::put::read_value;
+use crate::metadata::control::selector::TableSelector;
+
+/// Put table metadata into the metadata store.
+#[derive(Subcommand)]
+pub enum PutTableCommand {
+    Info(PutTableInfoCommand),
+    Route(PutTableRouteCommand),
+}
+
+impl PutTableCommand {
+    pub async fn build(&self) -> Result<Box<dyn Tool>, BoxedError> {
+        match self {
+            PutTableCommand::Info(cmd) => cmd.build().await,
+            PutTableCommand::Route(cmd) => cmd.build().await,
+        }
+    }
+}
+
+/// Put table info into the metadata store.
+#[derive(Debug, Parser)]
+pub struct PutTableInfoCommand {
+    #[clap(flatten)]
+    selector: TableSelector,
+
+    /// Read the JSON-encoded [`TableInfoValue`] from standard input.
+    #[clap(long, required = true)]
+    value_stdin: bool,
+
+    #[clap(flatten)]
+    store: StoreConfig,
+}
+
+impl PutTableInfoCommand {
+    pub async fn build(&self) -> Result<Box<dyn Tool>, BoxedError> {
+        let kv_backend = self.store.build().await?;
+        self.build_tool(tokio::io::stdin(), kv_backend).await
+    }
+
+    async fn build_tool<R>(
+        &self,
+        reader: R,
+        kv_backend: KvBackendRef,
+    ) -> Result<Box<dyn Tool>, BoxedError>
+    where
+        R: tokio::io::AsyncRead + Unpin,
+    {
+        self.selector.validate()?;
+        Ok(Box::new(PutTableInfoTool {
+            kv_backend,
+            selector: self.selector.clone(),
+            value: read_value(reader).await?,
+        }))
+    }
+}
+
+struct PutTableInfoTool {
+    kv_backend: KvBackendRef,
+    selector: TableSelector,
+    value: Vec<u8>,
+}
+
+#[async_trait]
+impl Tool for PutTableInfoTool {
+    async fn do_work(&self) -> Result<(), BoxedError> {
+        let table_metadata_manager = TableMetadataManager::new(self.kv_backend.clone());
+        let Some(table_id) = self
+            .selector
+            .resolve_table_id(table_metadata_manager.table_name_manager())
+            .await?
+        else {
+            return Err(BoxedError::new(
+                UnexpectedSnafu {
+                    msg: format!("Table({}) not found", self.selector.formatted_table_name()),
+                }
+                .build(),
+            ));
+        };
+
+        let (current_table_info, current_table_route) =
+            load_table_metadata(&table_metadata_manager, table_id).await?;
+        let new_table_info = TableInfoValue::try_from_raw_value(&self.value)
+            .map_err(|e| {
+                BoxedError::new(
+                    InvalidArgumentsSnafu {
+                        msg: format!("Invalid table info JSON: {e}"),
+                    }
+                    .build(),
+                )
+            })?
+            .table_info;
+        validate_table_info(table_id, &current_table_info.table_info, &new_table_info)
+            .map_err(BoxedError::new)?;
+
+        let region_distribution =
+            physical_region_distribution(current_table_route.get_inner_ref())?;
+
+        if current_table_info.table_info != new_table_info {
+            table_metadata_manager
+                .update_table_info(&current_table_info, region_distribution, new_table_info)
+                .await
+                .map_err(BoxedError::new)?;
+            println!("Table({table_id}) info updated");
+        }
+
+        Ok(())
+    }
+}
+
+/// Put table route into the metadata store.
+#[derive(Debug, Parser)]
+pub struct PutTableRouteCommand {
+    #[clap(flatten)]
+    selector: TableSelector,
+
+    /// Read the JSON-encoded [`TableRouteValue`] from standard input.
+    #[clap(long, required = true)]
+    value_stdin: bool,
+
+    #[clap(flatten)]
+    store: StoreConfig,
+}
+
+impl PutTableRouteCommand {
+    pub async fn build(&self) -> Result<Box<dyn Tool>, BoxedError> {
+        let kv_backend = self.store.build().await?;
+        self.build_tool(tokio::io::stdin(), kv_backend).await
+    }
+
+    async fn build_tool<R>(
+        &self,
+        reader: R,
+        kv_backend: KvBackendRef,
+    ) -> Result<Box<dyn Tool>, BoxedError>
+    where
+        R: tokio::io::AsyncRead + Unpin,
+    {
+        self.selector.validate()?;
+        Ok(Box::new(PutTableRouteTool {
+            kv_backend,
+            selector: self.selector.clone(),
+            value: read_value(reader).await?,
+        }))
+    }
+}
+
+struct PutTableRouteTool {
+    kv_backend: KvBackendRef,
+    selector: TableSelector,
+    value: Vec<u8>,
+}
+
+#[async_trait]
+impl Tool for PutTableRouteTool {
+    async fn do_work(&self) -> Result<(), BoxedError> {
+        let table_metadata_manager = TableMetadataManager::new(self.kv_backend.clone());
+        let Some(table_id) = self
+            .selector
+            .resolve_table_id(table_metadata_manager.table_name_manager())
+            .await?
+        else {
+            return Err(BoxedError::new(
+                UnexpectedSnafu {
+                    msg: format!("Table({}) not found", self.selector.formatted_table_name()),
+                }
+                .build(),
+            ));
+        };
+
+        let (current_table_info, current_table_route) =
+            load_table_metadata(&table_metadata_manager, table_id).await?;
+        let current_region_routes = current_table_route
+            .region_routes()
+            .map_err(BoxedError::new)?;
+        let new_table_route = TableRouteValue::try_from_raw_value(&self.value).map_err(|e| {
+            BoxedError::new(
+                InvalidArgumentsSnafu {
+                    msg: format!("Invalid table route JSON: {e}"),
+                }
+                .build(),
+            )
+        })?;
+        let new_region_routes = new_table_route.region_routes().map_err(BoxedError::new)?;
+        validate_table_route(table_id, new_region_routes, current_region_routes)
+            .map_err(BoxedError::new)?;
+        let region_info =
+            load_region_info(&table_metadata_manager, table_id, current_region_routes).await?;
+        let new_region_options = current_table_info.table_info.to_region_options();
+        let new_region_wal_options = region_info.region_wal_options.clone();
+
+        if current_table_route.get_inner_ref() != &new_table_route {
+            table_metadata_manager
+                .update_table_route(
+                    table_id,
+                    region_info,
+                    &current_table_route,
+                    new_region_routes.clone(),
+                    &new_region_options,
+                    &new_region_wal_options,
+                )
+                .await
+                .map_err(BoxedError::new)?;
+            println!("Table({table_id}) route updated");
+        }
+
+        Ok(())
+    }
+}
+
+fn validate_table_route(
+    table_id: TableId,
+    new_region_routes: &[RegionRoute],
+    current_region_route: &[RegionRoute],
+) -> Result<(), Error> {
+    let current_region_ids = current_region_route
+        .iter()
+        .map(|r| r.region.id)
+        .collect::<HashSet<_>>();
+    for route in new_region_routes {
+        ensure!(
+            route.region.id.table_id() == table_id,
+            InvalidArgumentsSnafu {
+                msg: format!(
+                    "Invalid table route: all region routes must have table id {table_id}, but got {}",
+                    route.region.id.table_id()
+                ),
+            }
+        );
+        // Ensure the region in new route exists in current route
+        current_region_ids
+            .contains(&route.region.id)
+            .then_some(())
+            .context(InvalidArgumentsSnafu {
+                msg: format!(
+                    "Invalid table route: region {} does not exist in current routes",
+                    route.region.id
+                ),
+            })?;
+    }
+
+    Ok(())
+}
+
+fn validate_table_info(
+    table_id: TableId,
+    current_table_info: &TableInfo,
+    new_table_info: &TableInfo,
+) -> Result<(), Error> {
+    ensure!(
+        new_table_info.ident.table_id == table_id,
+        InvalidArgumentsSnafu {
+            msg: format!(
+                "Invalid table info: expected table id {table_id}, got {}",
+                new_table_info.ident.table_id
+            ),
+        }
+    );
+
+    ensure!(
+        current_table_info.catalog_name == new_table_info.catalog_name,
+        InvalidArgumentsSnafu {
+            msg: format!(
+                "Invalid table info: catalog name is immutable, expected {}, got {}",
+                current_table_info.catalog_name, new_table_info.catalog_name
+            ),
+        }
+    );
+
+    ensure!(
+        current_table_info.schema_name == new_table_info.schema_name,
+        InvalidArgumentsSnafu {
+            msg: format!(
+                "Invalid table info: schema name is immutable, expected {}, got {}",
+                current_table_info.schema_name, new_table_info.schema_name
+            ),
+        }
+    );
+
+    ensure!(
+        current_table_info.name == new_table_info.name,
+        InvalidArgumentsSnafu {
+            msg: format!(
+                "Invalid table info: table name is immutable, expected {}, got {}",
+                current_table_info.name, new_table_info.name
+            ),
+        }
+    );
+
+    Ok(())
+}
+
+async fn load_region_info(
+    table_metadata_manager: &TableMetadataManager,
+    table_id: TableId,
+    region_routes: &[RegionRoute],
+) -> Result<RegionInfo, BoxedError> {
+    let datanode_id = region_distribution(region_routes)
+        .into_keys()
+        .next()
+        .ok_or_else(|| {
+            BoxedError::new(
+                UnexpectedSnafu {
+                    msg: format!(
+                        "Missing datanode assignment for physical table route: {table_id}"
+                    ),
+                }
+                .build(),
+            )
+        })?;
+
+    table_metadata_manager
+        .datanode_table_manager()
+        .get(&DatanodeTableKey::new(datanode_id, table_id))
+        .await
+        .map_err(BoxedError::new)?
+        .map(|value| value.region_info)
+        .ok_or_else(|| {
+            BoxedError::new(
+                UnexpectedSnafu {
+                    msg: format!(
+                        "Missing datanode table metadata for physical table route: {table_id}"
+                    ),
+                }
+                .build(),
+            )
+        })
+}
+
+async fn load_table_metadata(
+    table_metadata_manager: &TableMetadataManager,
+    table_id: TableId,
+) -> Result<
+    (
+        DeserializedValueWithBytes<TableInfoValue>,
+        DeserializedValueWithBytes<TableRouteValue>,
+    ),
+    BoxedError,
+> {
+    let (table_info, table_route) = table_metadata_manager
+        .get_full_table_info(table_id)
+        .await
+        .map_err(BoxedError::new)?;
+    let table_info =
+        table_info.ok_or_else(|| BoxedError::new(TableNotFoundSnafu { table_id }.build()))?;
+    let table_route =
+        table_route.ok_or_else(|| BoxedError::new(TableNotFoundSnafu { table_id }.build()))?;
+    Ok((table_info, table_route))
+}
+
+fn physical_region_distribution(
+    table_route: &TableRouteValue,
+) -> Result<Option<RegionDistribution>, BoxedError> {
+    if !table_route.is_physical() {
+        return Ok(None);
+    }
+
+    table_route
+        .region_routes()
+        .map(|routes| Some(region_distribution(routes)))
+        .map_err(BoxedError::new)
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+    use std::sync::Arc;
+
+    use clap::Parser;
+    use common_error::ext::{BoxedError, ErrorExt};
+    use common_meta::key::TableMetadataManager;
+    use common_meta::key::datanode_table::{DatanodeTableKey, DatanodeTableManager};
+    use common_meta::key::table_info::TableInfoValue;
+    use common_meta::key::table_route::TableRouteValue;
+    use common_meta::kv_backend::KvBackendRef;
+    use common_meta::kv_backend::memory::MemoryKvBackend;
+    use common_meta::peer::Peer;
+    use common_meta::rpc::router::RegionRoute;
+    use store_api::storage::RegionId;
+    use tokio::io::BufReader;
+
+    use super::{
+        PutTableInfoCommand, PutTableInfoTool, PutTableRouteCommand, PutTableRouteTool,
+        validate_table_route,
+    };
+    use crate::Tool;
+    use crate::metadata::control::selector::TableSelector;
+    use crate::metadata::control::test_utils::prepare_physical_table_metadata;
+
+    impl PutTableInfoCommand {
+        async fn build_for_test<R>(
+            &self,
+            reader: R,
+            kv_backend: KvBackendRef,
+        ) -> Result<Box<dyn Tool>, BoxedError>
+        where
+            R: tokio::io::AsyncRead + Unpin,
+        {
+            self.build_tool(reader, kv_backend).await
+        }
+    }
+
+    impl PutTableRouteCommand {
+        async fn build_for_test<R>(
+            &self,
+            reader: R,
+            kv_backend: KvBackendRef,
+        ) -> Result<Box<dyn Tool>, BoxedError>
+        where
+            R: tokio::io::AsyncRead + Unpin,
+        {
+            self.build_tool(reader, kv_backend).await
+        }
+    }
+
+    #[tokio::test]
+    async fn test_put_table_selector_validation() {
+        let command = PutTableInfoCommand::parse_from([
+            "info",
+            "--value-stdin",
+            "--backend",
+            "memory-store",
+            "--store-addrs",
+            "memory://",
+        ]);
+
+        let err = match command.build().await {
+            Ok(_) => panic!("expected validation failure"),
+            Err(err) => err,
+        };
+        assert!(
+            err.output_msg()
+                .contains("You must specify either --table-id or --table-name.")
+        );
+    }
+
+    #[tokio::test]
+    async fn test_put_table_command_builds_tool_with_table_name() {
+        let command = PutTableInfoCommand::parse_from([
+            "info",
+            "--table-name",
+            "my_table",
+            "--value-stdin",
+            "--backend",
+            "memory-store",
+            "--store-addrs",
+            "memory://",
+        ]);
+
+        let _tool = command
+            .build_for_test(
+                BufReader::new(&b"{}"[..]),
+                Arc::new(MemoryKvBackend::new()) as KvBackendRef,
+            )
+            .await
+            .unwrap();
+    }
+
+    #[tokio::test]
+    async fn test_put_table_info_rejects_table_name_change() {
+        let kv_backend = Arc::new(MemoryKvBackend::new()) as KvBackendRef;
+        let table_metadata_manager = TableMetadataManager::new(kv_backend.clone());
+        let table_id = 1024;
+        let (table_info, table_route) =
+            prepare_physical_table_metadata("old_table", table_id).await;
+        table_metadata_manager
+            .create_table_metadata(
+                table_info.clone(),
+                TableRouteValue::Physical(table_route),
+                HashMap::new(),
+            )
+            .await
+            .unwrap();
+
+        let mut new_table_info = table_info;
+        new_table_info.name = "new_table".to_string();
+        let tool = PutTableInfoTool {
+            kv_backend: kv_backend.clone(),
+            selector: TableSelector::with_table_id(table_id),
+            value: serde_json::to_vec(&TableInfoValue::new(new_table_info)).unwrap(),
+        };
+
+        let err = tool.do_work().await.unwrap_err();
+        assert!(
+            err.output_msg()
+                .contains("Invalid table info: table name is immutable")
+        );
+    }
+
+    #[tokio::test]
+    async fn test_put_table_info_rejects_schema_change() {
+        let kv_backend = Arc::new(MemoryKvBackend::new()) as KvBackendRef;
+        let table_metadata_manager = TableMetadataManager::new(kv_backend.clone());
+        let table_id = 1024;
+        let (table_info, table_route) =
+            prepare_physical_table_metadata("old_table", table_id).await;
+        table_metadata_manager
+            .create_table_metadata(
+                table_info.clone(),
+                TableRouteValue::Physical(table_route),
+                HashMap::new(),
+            )
+            .await
+            .unwrap();
+
+        let mut new_table_info = table_info;
+        new_table_info.schema_name = "another_schema".to_string();
+        let tool = PutTableInfoTool {
+            kv_backend,
+            selector: TableSelector::with_table_id(table_id),
+            value: serde_json::to_vec(&TableInfoValue::new(new_table_info)).unwrap(),
+        };
+
+        let err = tool.do_work().await.unwrap_err();
+        assert!(
+            err.output_msg()
+                .contains("Invalid table info: schema name is immutable")
+        );
+    }
+
+    #[tokio::test]
+    async fn test_put_table_route_updates_route_and_datanode_table() {
+        let kv_backend = Arc::new(MemoryKvBackend::new()) as KvBackendRef;
+        let table_metadata_manager = TableMetadataManager::new(kv_backend.clone());
+        let table_id = 1024;
+        let (table_info, table_route) = prepare_physical_table_metadata("my_table", table_id).await;
+        table_metadata_manager
+            .create_table_metadata(
+                table_info,
+                TableRouteValue::Physical(table_route.clone()),
+                HashMap::new(),
+            )
+            .await
+            .unwrap();
+
+        let mut region_routes = table_route.region_routes.clone();
+        region_routes[0].leader_peer = Some(Peer::empty(2));
+        let new_table_route = TableRouteValue::physical(region_routes);
+        let tool = PutTableRouteTool {
+            kv_backend: kv_backend.clone(),
+            selector: TableSelector::with_table_id(table_id),
+            value: serde_json::to_vec(&new_table_route).unwrap(),
+        };
+
+        tool.do_work().await.unwrap();
+
+        let (_, current_route) = table_metadata_manager
+            .get_full_table_info(table_id)
+            .await
+            .unwrap();
+        let current_route = current_route.unwrap().into_inner();
+        assert_eq!(
+            current_route.region_routes().unwrap(),
+            new_table_route.region_routes().unwrap()
+        );
+
+        let datanode_table_manager = DatanodeTableManager::new(kv_backend);
+        let updated = datanode_table_manager
+            .get(&DatanodeTableKey::new(2, table_id))
+            .await
+            .unwrap();
+        assert!(updated.is_some());
+    }
+
+    #[tokio::test]
+    async fn test_put_table_route_rejects_logical_route() {
+        let kv_backend = Arc::new(MemoryKvBackend::new()) as KvBackendRef;
+        let table_metadata_manager = TableMetadataManager::new(kv_backend.clone());
+        let table_id = 1024;
+        let (table_info, table_route) = prepare_physical_table_metadata("my_table", table_id).await;
+        table_metadata_manager
+            .create_table_metadata(
+                table_info,
+                TableRouteValue::Physical(table_route),
+                HashMap::new(),
+            )
+            .await
+            .unwrap();
+
+        let tool = PutTableRouteTool {
+            kv_backend,
+            selector: TableSelector::with_table_id(table_id),
+            value: serde_json::to_vec(&TableRouteValue::logical(table_id + 1)).unwrap(),
+        };
+
+        let err = tool.do_work().await.unwrap_err();
+        assert!(err.output_msg().contains("non-physical TableRouteValue."));
+    }
+
+    #[test]
+    fn test_validate_table_route_rejects_new_region_not_in_current_route() {
+        let table_id = 1024;
+        let current_region_routes = vec![
+            RegionRoute {
+                region: common_meta::rpc::router::Region {
+                    id: RegionId::new(table_id, 1),
+                    ..Default::default()
+                },
+                ..Default::default()
+            },
+            RegionRoute {
+                region: common_meta::rpc::router::Region {
+                    id: RegionId::new(table_id, 2),
+                    ..Default::default()
+                },
+                ..Default::default()
+            },
+        ];
+        let new_region_routes = vec![
+            RegionRoute {
+                region: common_meta::rpc::router::Region {
+                    id: RegionId::new(table_id, 1),
+                    ..Default::default()
+                },
+                ..Default::default()
+            },
+            RegionRoute {
+                region: common_meta::rpc::router::Region {
+                    id: RegionId::new(table_id, 3),
+                    ..Default::default()
+                },
+                ..Default::default()
+            },
+        ];
+
+        let err =
+            validate_table_route(table_id, &current_region_routes, &new_region_routes).unwrap_err();
+
+        assert!(err.to_string().contains("does not exist in current routes"));
+    }
+
+    #[tokio::test]
+    async fn test_put_table_command_builds_tool() {
+        let value = serde_json::to_vec(&TableRouteValue::logical(1025)).unwrap();
+        let command = PutTableRouteCommand::parse_from([
+            "route",
+            "--table-id",
+            "1024",
+            "--value-stdin",
+            "--backend",
+            "memory-store",
+            "--store-addrs",
+            "memory://",
+        ]);
+
+        let _tool = command
+            .build_for_test(
+                BufReader::new(value.as_slice()),
+                Arc::new(MemoryKvBackend::new()) as KvBackendRef,
+            )
+            .await
+            .unwrap();
+    }
+}
diff --git a/src/cli/src/metadata/control/selector.rs b/src/cli/src/metadata/control/selector.rs
new file mode 100644
index 0000000000..6d3c47fa9c
--- /dev/null
+++ b/src/cli/src/metadata/control/selector.rs
@@ -0,0 +1,100 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use clap::Args;
+use client::{DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME};
+use common_catalog::format_full_table_name;
+use common_error::ext::BoxedError;
+use common_meta::key::table_name::TableNameManager;
+use store_api::storage::TableId;
+
+use crate::error::InvalidArgumentsSnafu;
+use crate::metadata::control::utils::get_table_id_by_name;
+
+/// Selects a table by id or by fully qualified name.
+#[derive(Debug, Clone, Default, Args)]
+pub(crate) struct TableSelector {
+    /// The table id to select from the metadata store.
+    #[clap(long)]
+    table_id: Option<u32>,
+
+    /// The table name to select from the metadata store.
+    #[clap(long)]
+    table_name: Option<String>,
+
+    /// The schema name of the table.
+    #[clap(long, default_value = DEFAULT_SCHEMA_NAME)]
+    schema_name: String,
+
+    /// The catalog name of the table.
+    #[clap(long, default_value = DEFAULT_CATALOG_NAME)]
+    catalog_name: String,
+}
+
+impl TableSelector {
+    pub(crate) fn validate(&self) -> Result<(), BoxedError> {
+        if matches!(
+            (&self.table_id, &self.table_name),
+            (Some(_), Some(_)) | (None, None)
+        ) {
+            return Err(BoxedError::new(
+                InvalidArgumentsSnafu {
+                    msg: "You must specify either --table-id or --table-name.",
+                }
+                .build(),
+            ));
+        }
+
+        Ok(())
+    }
+
+    pub(crate) async fn resolve_table_id(
+        &self,
+        table_name_manager: &TableNameManager,
+    ) -> Result<Option<TableId>, BoxedError> {
+        if let Some(table_id) = self.table_id {
+            return Ok(Some(table_id));
+        }
+
+        get_table_id_by_name(
+            table_name_manager,
+            &self.catalog_name,
+            &self.schema_name,
+            self.table_name
+                .as_deref()
+                .expect("validated table selector"),
+        )
+        .await
+    }
+
+    pub(crate) fn formatted_table_name(&self) -> String {
+        format_full_table_name(
+            &self.catalog_name,
+            &self.schema_name,
+            self.table_name.as_deref().unwrap_or_default(),
+        )
+    }
+}
+
+#[cfg(test)]
+impl TableSelector {
+    pub(crate) fn with_table_id(table_id: u32) -> Self {
+        Self {
+            table_id: Some(table_id),
+            table_name: None,
+            schema_name: DEFAULT_SCHEMA_NAME.to_string(),
+            catalog_name: DEFAULT_CATALOG_NAME.to_string(),
+        }
+    }
+}
diff --git a/src/cli/src/metadata/snapshot.rs b/src/cli/src/metadata/snapshot.rs
index 648d3a687d..59b2599ad9 100644
--- a/src/cli/src/metadata/snapshot.rs
+++ b/src/cli/src/metadata/snapshot.rs
@@ -16,7 +16,7 @@ use async_trait::async_trait;
 use clap::{Parser, Subcommand};
 use common_error::ext::BoxedError;
 use common_meta::snapshot::MetadataSnapshotManager;
-use object_store::{ObjectStore, Scheme};
+use object_store::{ObjectStore, services};
 
 use crate::Tool;
 use crate::common::{ObjectStoreConfig, StoreConfig, new_fs_object_store};
@@ -276,7 +276,7 @@ fn build_object_store_and_resolve_file_path(
         None => new_fs_object_store(fs_root)?,
     };
 
-    let file_path = if matches!(object_store.info().scheme(), Scheme::Fs) {
+    let file_path = if object_store.info().scheme() == services::FS_SCHEME {
         resolve_relative_path_with_current_dir(file_path).map_err(BoxedError::new)?
     } else {
         file_path.to_string()
diff --git a/src/client/src/database.rs b/src/client/src/database.rs
index 6a7ac62fc3..13ff9fc651 100644
--- a/src/client/src/database.rs
+++ b/src/client/src/database.rs
@@ -14,7 +14,9 @@
 
 use std::pin::Pin;
 use std::str::FromStr;
-use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::{Arc, RwLock};
+use std::task::{Context, Poll};
 
 use api::v1::auth_header::AuthScheme;
 use api::v1::ddl_request::Expr as DdlExpr;
@@ -25,6 +27,7 @@ use api::v1::{
     AlterTableExpr, AuthHeader, Basic, CreateTableExpr, DdlRequest, GreptimeRequest,
     InsertRequests, QueryRequest, RequestHeader, RowInsertRequests,
 };
+use arc_swap::ArcSwapOption;
 use arrow_flight::{FlightData, Ticket};
 use async_stream::stream;
 use base64::Engine;
@@ -33,17 +36,18 @@ use common_catalog::build_db_string;
 use common_catalog::consts::{DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME};
 use common_error::ext::BoxedError;
 use common_grpc::flight::do_put::DoPutResponse;
-use common_grpc::flight::{FlightDecoder, FlightMessage};
+use common_grpc::flight::{FLOW_EXTENSIONS_METADATA_KEY, FlightDecoder, FlightMessage};
 use common_query::Output;
+use common_recordbatch::adapter::RecordBatchMetrics;
 use common_recordbatch::error::ExternalSnafu;
-use common_recordbatch::{RecordBatch, RecordBatchStreamWrapper};
+use common_recordbatch::{OrderOption, RecordBatch, RecordBatchStream, RecordBatchStreamWrapper};
 use common_telemetry::tracing::Span;
 use common_telemetry::tracing_context::W3cTrace;
 use common_telemetry::{error, warn};
 use futures::future;
 use futures_util::{Stream, StreamExt, TryStreamExt};
 use prost::Message;
-use snafu::{OptionExt, ResultExt, ensure};
+use snafu::{OptionExt, ResultExt};
 use tonic::metadata::{AsciiMetadataKey, AsciiMetadataValue, MetadataMap, MetadataValue};
 use tonic::transport::Channel;
 
@@ -57,6 +61,313 @@ type FlightDataStream = Pin<Box<dyn Stream<Item = FlightData> + Send>>;
 
 type DoPutResponseStream = Pin<Box<dyn Stream<Item = Result<DoPutResponse>>>>;
 
+/// Terminal metrics associated with a query output.
+///
+/// For streaming outputs, metrics are only final after the stream is fully
+/// drained and [`Self::is_ready`] returns `true`.
+#[derive(Debug, Clone, Default)]
+pub struct OutputMetrics {
+    inner: Arc<OutputMetricsInner>,
+}
+
+#[derive(Debug, Default)]
+struct OutputMetricsInner {
+    metrics: RwLock<Option<RecordBatchMetrics>>,
+    ready: AtomicBool,
+}
+
+impl OutputMetrics {
+    fn new() -> Self {
+        Self::default()
+    }
+
+    /// Replaces the current terminal metrics snapshot.
+    pub fn update(&self, metrics: Option<RecordBatchMetrics>) {
+        *self.inner.metrics.write().unwrap() = metrics;
+    }
+
+    /// Marks the terminal metrics as final for this output.
+    pub fn mark_ready(&self) {
+        let _ = self
+            .inner
+            .ready
+            .compare_exchange(false, true, Ordering::AcqRel, Ordering::Acquire);
+    }
+
+    /// Returns whether terminal metrics are final.
+    ///
+    /// Streaming outputs become ready only after the stream reaches EOF.
+    pub fn is_ready(&self) -> bool {
+        self.inner.ready.load(Ordering::Acquire)
+    }
+
+    /// Returns the latest terminal metrics snapshot, if any.
+    pub fn get(&self) -> Option<RecordBatchMetrics> {
+        self.inner.metrics.read().unwrap().clone()
+    }
+
+    /// Returns proved per-region watermarks.
+    ///
+    /// Entries whose watermark is `None` are intentionally omitted because they
+    /// represent participating regions whose terminal sequence bound was not
+    /// provable.
+    pub fn region_watermark_map(&self) -> Option<std::collections::HashMap<u64, u64>> {
+        Some(
+            self.get()?
+                .region_watermarks
+                .into_iter()
+                .filter_map(|entry| entry.watermark.map(|seq| (entry.region_id, seq)))
+                .collect::<std::collections::HashMap<_, _>>(),
+        )
+    }
+
+    /// Returns all regions that participated in terminal metric collection,
+    /// including entries whose watermark is `None`.
+    pub fn participating_regions(&self) -> Option<std::collections::BTreeSet<u64>> {
+        Some(
+            self.get()?
+                .region_watermarks
+                .into_iter()
+                .map(|entry| entry.region_id)
+                .collect::<std::collections::BTreeSet<_>>(),
+        )
+    }
+}
+
+/// Query output together with a handle for its terminal metrics.
+///
+/// The contained [`OutputMetrics`] lets callers read stream terminal metrics
+/// after consuming `output`. For non-stream outputs, metrics are ready
+/// immediately.
+#[derive(Debug)]
+pub struct OutputWithMetrics {
+    pub output: Output,
+    pub metrics: OutputMetrics,
+}
+
+impl OutputWithMetrics {
+    /// Wraps an output with a terminal metrics handle.
+    ///
+    /// Stream outputs update the handle as the stream is consumed. Non-stream
+    /// outputs are marked ready immediately.
+    pub fn from_output(output: Output) -> Self {
+        let terminal_metrics = OutputMetrics::new();
+        let output = attach_terminal_metrics(output, &terminal_metrics);
+        Self {
+            output,
+            metrics: terminal_metrics,
+        }
+    }
+
+    /// Returns proved per-region watermarks from the terminal metrics.
+    pub fn region_watermark_map(&self) -> Option<std::collections::HashMap<u64, u64>> {
+        self.metrics.region_watermark_map()
+    }
+
+    /// Returns all regions participating in terminal metric collection.
+    pub fn participating_regions(&self) -> Option<std::collections::BTreeSet<u64>> {
+        self.metrics.participating_regions()
+    }
+
+    /// Drops the terminal metrics handle and returns the original output.
+    pub fn into_output(self) -> Output {
+        self.output
+    }
+}
+
+fn parse_terminal_metrics(metrics_json: &str) -> Result<RecordBatchMetrics> {
+    serde_json::from_str(metrics_json).map_err(|e| {
+        IllegalFlightMessagesSnafu {
+            reason: format!("Invalid terminal metrics message: {e}"),
+        }
+        .build()
+    })
+}
+
+struct StreamWithMetrics {
+    stream: common_recordbatch::SendableRecordBatchStream,
+    metrics: OutputMetrics,
+}
+
+impl StreamWithMetrics {
+    fn new(stream: common_recordbatch::SendableRecordBatchStream, metrics: OutputMetrics) -> Self {
+        Self { stream, metrics }
+    }
+
+    fn sync_terminal_metrics(&self) {
+        self.metrics.update(self.stream.metrics());
+    }
+}
+
+impl RecordBatchStream for StreamWithMetrics {
+    fn name(&self) -> &str {
+        self.stream.name()
+    }
+
+    fn schema(&self) -> datatypes::schema::SchemaRef {
+        self.stream.schema()
+    }
+
+    fn output_ordering(&self) -> Option<&[OrderOption]> {
+        self.stream.output_ordering()
+    }
+
+    fn metrics(&self) -> Option<RecordBatchMetrics> {
+        self.sync_terminal_metrics();
+        self.metrics.get()
+    }
+}
+
+impl Stream for StreamWithMetrics {
+    type Item = common_recordbatch::error::Result<RecordBatch>;
+
+    fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        let polled = Pin::new(&mut self.stream).poll_next(cx);
+        if let Poll::Ready(None) = &polled {
+            self.sync_terminal_metrics();
+            self.metrics.mark_ready();
+        }
+        polled
+    }
+
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        self.stream.size_hint()
+    }
+}
+
+fn attach_terminal_metrics(output: Output, terminal_metrics: &OutputMetrics) -> Output {
+    let Output { data, meta } = output;
+    let data = match data {
+        common_query::OutputData::Stream(stream) => {
+            terminal_metrics.update(stream.metrics());
+            common_query::OutputData::Stream(Box::pin(StreamWithMetrics::new(
+                stream,
+                terminal_metrics.clone(),
+            )))
+        }
+        other => {
+            terminal_metrics.mark_ready();
+            other
+        }
+    };
+    Output::new(data, meta)
+}
+
+async fn output_from_flight_message_stream<S>(
+    mut flight_message_stream: S,
+) -> Result<OutputWithMetrics>
+where
+    S: Stream<Item = Result<FlightMessage>> + Send + Unpin + 'static,
+{
+    let Some(first_flight_message) = flight_message_stream.next().await else {
+        return IllegalFlightMessagesSnafu {
+            reason: "Expect the response not to be empty",
+        }
+        .fail();
+    };
+
+    let first_flight_message = first_flight_message?;
+
+    match first_flight_message {
+        FlightMessage::AffectedRows { rows, metrics } => {
+            let terminal_metrics = OutputMetrics::new();
+            if let Some(metrics) = metrics {
+                terminal_metrics.update(Some(parse_terminal_metrics(&metrics)?));
+            }
+            let next_message = flight_message_stream.next().await.transpose()?;
+            match next_message {
+                None => terminal_metrics.mark_ready(),
+                Some(FlightMessage::Metrics(s)) if terminal_metrics.get().is_none() => {
+                    terminal_metrics.update(Some(parse_terminal_metrics(&s)?));
+                    terminal_metrics.mark_ready();
+                }
+                Some(FlightMessage::Metrics(_)) => {
+                    return IllegalFlightMessagesSnafu {
+                        reason: "'AffectedRows' Flight metadata already carries Metrics and cannot be followed by another Metrics message",
+                    }
+                    .fail();
+                }
+                Some(other) => {
+                    return IllegalFlightMessagesSnafu {
+                        reason: format!(
+                            "'AffectedRows' Flight message can only be followed by a Metrics message, got {other:?}"
+                        ),
+                    }
+                    .fail();
+                }
+            }
+            Ok(OutputWithMetrics {
+                output: Output::new_with_affected_rows(rows),
+                metrics: terminal_metrics,
+            })
+        }
+        FlightMessage::RecordBatch(_) | FlightMessage::Metrics(_) => IllegalFlightMessagesSnafu {
+            reason: "The first flight message cannot be a RecordBatch or Metrics message",
+        }
+        .fail(),
+        FlightMessage::Schema(schema) => {
+            let metrics = Arc::new(ArcSwapOption::from(None));
+            let metrics_ref = metrics.clone();
+            let schema = Arc::new(
+                datatypes::schema::Schema::try_from(schema).context(error::ConvertSchemaSnafu)?,
+            );
+            let schema_cloned = schema.clone();
+            let stream = Box::pin(stream!({
+                while let Some(flight_message_item) = flight_message_stream.next().await {
+                    let flight_message = match flight_message_item {
+                        Ok(message) => message,
+                        Err(e) => {
+                            yield Err(BoxedError::new(e)).context(ExternalSnafu);
+                            break;
+                        }
+                    };
+
+                    match flight_message {
+                        FlightMessage::RecordBatch(arrow_batch) => {
+                            yield Ok(RecordBatch::from_df_record_batch(
+                                schema_cloned.clone(),
+                                arrow_batch,
+                            ))
+                        }
+                        FlightMessage::Metrics(s) => {
+                            match parse_terminal_metrics(&s) {
+                                Ok(m) => {
+                                    metrics_ref.swap(Some(Arc::new(m)));
+                                }
+                                Err(e) => {
+                                    yield Err(BoxedError::new(e)).context(ExternalSnafu);
+                                }
+                            };
+                        }
+                        FlightMessage::AffectedRows { .. } | FlightMessage::Schema(_) => {
+                            yield IllegalFlightMessagesSnafu {
+                                reason: format!(
+                                    "A Schema message must be succeeded exclusively by a set of RecordBatch messages, flight_message: {:?}",
+                                    flight_message
+                                )
+                            }
+                            .fail()
+                            .map_err(BoxedError::new)
+                            .context(ExternalSnafu);
+                            break;
+                        }
+                    }
+                }
+            }));
+            let record_batch_stream = RecordBatchStreamWrapper {
+                schema,
+                stream,
+                output_ordering: None,
+                metrics,
+                span: Span::current(),
+            };
+            Ok(OutputWithMetrics::from_output(Output::new_with_stream(
+                Box::pin(record_batch_stream),
+            )))
+        }
+    }
+}
+
 #[derive(Clone, Debug, Default)]
 pub struct Database {
     // The "catalog" and "schema" to be used in processing the requests at the server side.
@@ -238,6 +549,22 @@ impl Database {
         Ok(())
     }
 
+    fn put_flow_extensions(
+        metadata: &mut MetadataMap,
+        flow_extensions: &[(&str, &str)],
+    ) -> Result<()> {
+        if flow_extensions.is_empty() {
+            return Ok(());
+        }
+
+        let value = serde_json::to_string(&flow_extensions.to_vec())
+            .expect("flow extension pairs should serialize");
+        let key = AsciiMetadataKey::from_static(FLOW_EXTENSIONS_METADATA_KEY);
+        let value = AsciiMetadataValue::from_str(&value).context(InvalidTonicMetadataValueSnafu)?;
+        metadata.insert(key, value);
+        Ok(())
+    }
+
     /// Make a request to the database.
     pub async fn handle(&self, request: Request) -> Result<u32> {
         let mut client = make_database_client(&self.client)?;
@@ -333,15 +660,58 @@ impl Database {
         let request = Request::Query(QueryRequest {
             query: Some(Query::Sql(sql.as_ref().to_string())),
         });
-        self.do_get(request, hints).await
+        self.do_get(request, hints, &[])
+            .await
+            .map(OutputWithMetrics::into_output)
+    }
+
+    /// Executes a SQL query and returns the output with terminal metrics.
+    ///
+    /// For stream outputs, callers must consume the stream before reading final
+    /// terminal metrics from [`OutputWithMetrics::metrics`].
+    pub async fn sql_with_terminal_metrics<S>(
+        &self,
+        sql: S,
+        hints: &[(&str, &str)],
+    ) -> Result<OutputWithMetrics>
+    where
+        S: AsRef<str>,
+    {
+        self.query_with_terminal_metrics_and_flow_extensions(
+            QueryRequest {
+                query: Some(Query::Sql(sql.as_ref().to_string())),
+            },
+            hints,
+            &[],
+        )
+        .await
     }
 
     /// Executes a logical plan directly without SQL parsing.
     pub async fn logical_plan(&self, logical_plan: Vec<u8>) -> Result<Output> {
-        let request = Request::Query(QueryRequest {
-            query: Some(Query::LogicalPlan(logical_plan)),
-        });
-        self.do_get(request, &[]).await
+        self.query_with_terminal_metrics_and_flow_extensions(
+            QueryRequest {
+                query: Some(Query::LogicalPlan(logical_plan)),
+            },
+            &[],
+            &[],
+        )
+        .await
+        .map(OutputWithMetrics::into_output)
+    }
+
+    /// Executes a query and carries flow extensions through Flight metadata.
+    ///
+    /// This is the lower-level terminal-metrics API for Flow callers that need
+    /// to pass JSON-bearing flow extensions without going through hint metadata.
+    pub async fn query_with_terminal_metrics_and_flow_extensions(
+        &self,
+        request: QueryRequest,
+        hints: &[(&str, &str)],
+        flow_extensions: &[(&str, &str)],
+    ) -> Result<OutputWithMetrics> {
+        self.do_get(Request::Query(request), hints, flow_extensions)
+            .await
     }
 
     /// Creates a new table using the provided table expression.
@@ -349,7 +719,9 @@ impl Database {
         let request = Request::Ddl(DdlRequest {
             expr: Some(DdlExpr::CreateTable(expr)),
         });
-        self.do_get(request, &[]).await
+        self.do_get(request, &[], &[])
+            .await
+            .map(OutputWithMetrics::into_output)
     }
 
     /// Alters an existing table using the provided alter expression.
@@ -357,17 +729,26 @@ impl Database {
         let request = Request::Ddl(DdlRequest {
             expr: Some(DdlExpr::AlterTable(expr)),
         });
-        self.do_get(request, &[]).await
+        self.do_get(request, &[], &[])
+            .await
+            .map(OutputWithMetrics::into_output)
     }
 
-    async fn do_get(&self, request: Request, hints: &[(&str, &str)]) -> Result<Output> {
+    async fn do_get(
+        &self,
+        request: Request,
+        hints: &[(&str, &str)],
+        flow_extensions: &[(&str, &str)],
+    ) -> Result<OutputWithMetrics> {
         let request = self.to_rpc_request(request);
         let request = Ticket {
             ticket: request.encode_to_vec().into(),
         };
 
         let mut request = tonic::Request::new(request);
-        Self::put_hints(request.metadata_mut(), hints)?;
+        let metadata = request.metadata_mut();
+        Self::put_hints(metadata, hints)?;
+        Self::put_flow_extensions(metadata, flow_extensions)?;
 
         let mut client = self.client.make_flight_client(false, false)?;
 
@@ -389,7 +770,7 @@ impl Database {
         let flight_data_stream = response.into_inner();
         let mut decoder = FlightDecoder::default();
 
-        let mut flight_message_stream = flight_data_stream.map(move |flight_data| {
+        let flight_message_stream = flight_data_stream.map(move |flight_data| {
             flight_data
                 .map_err(Error::from)
                 .and_then(|data| decoder.try_decode(&data).context(ConvertFlightDataSnafu))?
@@ -398,70 +779,7 @@ impl Database {
                 })
         });
 
-        let Some(first_flight_message) = flight_message_stream.next().await else {
-            return IllegalFlightMessagesSnafu {
-                reason: "Expect the response not to be empty",
-            }
-            .fail();
-        };
-
-        let first_flight_message = first_flight_message?;
-
-        match first_flight_message {
-            FlightMessage::AffectedRows(rows) => {
-                ensure!(
-                    flight_message_stream.next().await.is_none(),
-                    IllegalFlightMessagesSnafu {
-                        reason: "Expect 'AffectedRows' Flight messages to be the one and the only!"
-                    }
-                );
-                Ok(Output::new_with_affected_rows(rows))
-            }
-            FlightMessage::RecordBatch(_) | FlightMessage::Metrics(_) => {
-                IllegalFlightMessagesSnafu {
-                    reason: "The first flight message cannot be a RecordBatch or Metrics message",
-                }
-                .fail()
-            }
-            FlightMessage::Schema(schema) => {
-                let schema = Arc::new(
-                    datatypes::schema::Schema::try_from(schema)
-                        .context(error::ConvertSchemaSnafu)?,
-                );
-                let schema_cloned = schema.clone();
-                let stream = Box::pin(stream!({
-                    while let Some(flight_message) = flight_message_stream.next().await {
-                        let flight_message = flight_message
-                            .map_err(BoxedError::new)
-                            .context(ExternalSnafu)?;
-                        match flight_message {
-                            FlightMessage::RecordBatch(arrow_batch) => {
-                                yield Ok(RecordBatch::from_df_record_batch(
-                                    schema_cloned.clone(),
-                                    arrow_batch,
-                                ))
-                            }
-                            FlightMessage::Metrics(_) => {}
-                            FlightMessage::AffectedRows(_) | FlightMessage::Schema(_) => {
-                                yield IllegalFlightMessagesSnafu {reason: format!("A Schema message must be succeeded exclusively by a set of RecordBatch messages, flight_message: {:?}", flight_message)}
-                                        .fail()
-                                        .map_err(BoxedError::new)
-                                        .context(ExternalSnafu);
-                                break;
-                            }
-                        }
-                    }
-                }));
-                let record_batch_stream = RecordBatchStreamWrapper {
-                    schema,
-                    stream,
-                    output_ordering: None,
-                    metrics: Default::default(),
-                    span: Span::current(),
-                };
-                Ok(Output::new_with_stream(Box::pin(record_batch_stream)))
-            }
-        }
+        output_from_flight_message_stream(flight_message_stream).await
     }
 
     /// Ingest a stream of [RecordBatch]es that belong to a table, using Arrow Flight's "`DoPut`"
@@ -512,16 +830,104 @@ struct FlightContext {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::sync::Arc;
+    use std::task::{Context, Poll};
 
     use api::v1::auth_header::AuthScheme;
     use api::v1::{AuthHeader, Basic};
     use common_error::status_code::StatusCode;
+    use common_query::OutputData;
+    use common_recordbatch::{OrderOption, RecordBatch, RecordBatchStream};
+    use datatypes::prelude::{ConcreteDataType, VectorRef};
+    use datatypes::schema::{ColumnSchema, Schema};
+    use datatypes::vectors::Int32Vector;
+    use futures_util::StreamExt;
     use tonic::{Code, Status};
 
     use super::*;
     use crate::error::TonicSnafu;
 
+    struct MockMetricsStream {
+        schema: datatypes::schema::SchemaRef,
+        batch: Option<RecordBatch>,
+        metrics: RecordBatchMetrics,
+        terminal_metrics_only: bool,
+    }
+
+    impl Stream for MockMetricsStream {
+        type Item = common_recordbatch::error::Result<RecordBatch>;
+
+        fn poll_next(mut self: Pin<&mut Self>, _cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+            Poll::Ready(self.batch.take().map(Ok))
+        }
+    }
+
+    impl RecordBatchStream for MockMetricsStream {
+        fn name(&self) -> &str {
+            "MockMetricsStream"
+        }
+
+        fn schema(&self) -> datatypes::schema::SchemaRef {
+            self.schema.clone()
+        }
+
+        fn output_ordering(&self) -> Option<&[OrderOption]> {
+            None
+        }
+
+        fn metrics(&self) -> Option<RecordBatchMetrics> {
+            if self.terminal_metrics_only && self.batch.is_some() {
+                return None;
+            }
+            Some(self.metrics.clone())
+        }
+    }
+
+    fn terminal_metrics_json() -> String {
+        terminal_metrics_json_with_seq(42)
+    }
+
+    fn terminal_metrics_json_with_seq(seq: u64) -> String {
+        serde_json::to_string(&RecordBatchMetrics {
+            region_watermarks: vec![common_recordbatch::adapter::RegionWatermarkEntry {
+                region_id: 7,
+                watermark: Some(seq),
+            }],
+            ..Default::default()
+        })
+        .unwrap()
+    }
+
+    #[test]
+    fn test_put_flow_extensions_preserves_comma_bearing_values() {
+        let mut metadata = MetadataMap::new();
+        Database::put_flow_extensions(
+            &mut metadata,
+            &[
+                ("flow.return_region_seq", "true"),
+                ("flow.incremental_after_seqs", r#"{"1":10,"2":20}"#),
+            ],
+        )
+        .unwrap();
+
+        let value = metadata
+            .get(FLOW_EXTENSIONS_METADATA_KEY)
+            .unwrap()
+            .to_str()
+            .unwrap();
+        let decoded: Vec<(String, String)> = serde_json::from_str(value).unwrap();
+        assert_eq!(
+            decoded,
+            vec![
+                ("flow.return_region_seq".to_string(), "true".to_string()),
+                (
+                    "flow.incremental_after_seqs".to_string(),
+                    r#"{"1":10,"2":20}"#.to_string()
+                ),
+            ]
+        );
+    }
+
     #[test]
     fn test_flight_ctx() {
         let mut ctx = FlightContext::default();
@@ -536,12 +942,12 @@ mod tests {
             auth_scheme: Some(basic),
         });
 
-        assert_matches!(
+        assert!(matches!(
             ctx.auth_header,
             Some(AuthHeader {
                 auth_scheme: Some(AuthScheme::Basic(_)),
             })
-        )
+        ));
     }
 
     #[test]
@@ -558,4 +964,198 @@ mod tests {
 
         assert_eq!(expected.to_string(), actual.to_string());
     }
+
+    #[tokio::test]
+    async fn test_query_with_terminal_metrics_tracks_terminal_only_metrics() {
+        let schema = Arc::new(Schema::new(vec![ColumnSchema::new(
+            "v",
+            ConcreteDataType::int32_datatype(),
+            false,
+        )]));
+        let batch = RecordBatch::new(
+            schema.clone(),
+            vec![Arc::new(Int32Vector::from_slice([1, 2])) as VectorRef],
+        )
+        .unwrap();
+        let output = Output::new_with_stream(Box::pin(MockMetricsStream {
+            schema,
+            batch: Some(batch),
+            metrics: RecordBatchMetrics {
+                region_watermarks: vec![common_recordbatch::adapter::RegionWatermarkEntry {
+                    region_id: 7,
+                    watermark: Some(42),
+                }],
+                ..Default::default()
+            },
+            terminal_metrics_only: true,
+        }));
+
+        let result = OutputWithMetrics::from_output(output);
+        let terminal_metrics = result.metrics.clone();
+        assert!(!terminal_metrics.is_ready());
+        assert!(terminal_metrics.get().is_none());
+
+        let OutputData::Stream(mut stream) = result.output.data else {
+            panic!("expected stream output");
+        };
+        while stream.next().await.is_some() {}
+
+        assert!(terminal_metrics.is_ready());
+        assert_eq!(
+            terminal_metrics.participating_regions(),
+            Some(std::collections::BTreeSet::from([7_u64]))
+        );
+        assert_eq!(
+            terminal_metrics.region_watermark_map(),
+            Some(std::collections::HashMap::from([(7_u64, 42_u64)]))
+        );
+    }
+
+    #[test]
+    fn test_parse_terminal_metrics_rejects_invalid_json() {
+        assert!(parse_terminal_metrics("{not-json}").is_err());
+    }
+
+    #[tokio::test]
+    async fn test_affected_rows_inline_metrics_are_parsed() {
+        let output = output_from_flight_message_stream(futures_util::stream::iter(vec![Ok(
+            FlightMessage::AffectedRows {
+                rows: 3,
+                metrics: Some(terminal_metrics_json()),
+            },
+        )]
+            as Vec<Result<FlightMessage>>))
+        .await
+        .unwrap();
+
+        assert!(matches!(output.output.data, OutputData::AffectedRows(3)));
+        assert!(output.metrics.is_ready());
+        assert_eq!(
+            output.metrics.region_watermark_map(),
+            Some(std::collections::HashMap::from([(7, 42)]))
+        );
+    }
+
+    #[tokio::test]
+    async fn test_affected_rows_inline_metrics_rejects_trailing_metrics() {
+        let metrics_json = terminal_metrics_json();
+        let err = output_from_flight_message_stream(futures_util::stream::iter(vec![
+            Ok(FlightMessage::AffectedRows {
+                rows: 3,
+                metrics: Some(metrics_json.clone()),
+            }),
+            Ok(FlightMessage::Metrics(metrics_json)),
+        ]
+            as Vec<Result<FlightMessage>>))
+        .await
+        .unwrap_err();
+
+        assert!(
+            err.to_string().contains("already carries Metrics"),
+            "unexpected error: {err:?}"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_invalid_terminal_metrics_after_record_batch_yields_batch_then_error() {
+        let schema = Arc::new(Schema::new(vec![ColumnSchema::new(
+            "v",
+            ConcreteDataType::int32_datatype(),
+            false,
+        )]));
+        let batch = RecordBatch::new(
+            schema.clone(),
+            vec![Arc::new(Int32Vector::from_slice([1])) as VectorRef],
+        )
+        .unwrap();
+        let output = output_from_flight_message_stream(futures_util::stream::iter(vec![
+            Ok(FlightMessage::Schema(schema.arrow_schema().clone())),
+            Ok(FlightMessage::RecordBatch(batch.into_df_record_batch())),
+            Ok(FlightMessage::Metrics("{not-json}".to_string())),
+        ]
+            as Vec<Result<FlightMessage>>))
+        .await
+        .unwrap();
+        let terminal_metrics = output.metrics.clone();
+        let OutputData::Stream(mut record_batch_stream) = output.output.data else {
+            panic!("expected stream output");
+        };
+
+        let batch = record_batch_stream.next().await.unwrap().unwrap();
+        assert_eq!(batch.num_rows(), 1);
+
+        let err = record_batch_stream.next().await.unwrap().unwrap_err();
+        assert_eq!("External error", err.to_string());
+        assert!(
+            format!("{err:?}").contains("Invalid terminal metrics message"),
+            "unexpected error: {err:?}"
+        );
+        assert!(record_batch_stream.next().await.is_none());
+        assert!(terminal_metrics.is_ready());
+        assert!(terminal_metrics.get().is_none());
+    }
+
+    #[tokio::test]
+    async fn test_record_batch_stream_continues_after_partial_metrics() {
+        let schema = Arc::new(Schema::new(vec![ColumnSchema::new(
+            "v",
+            ConcreteDataType::int32_datatype(),
+            false,
+        )]));
+        let first_batch = RecordBatch::new(
+            schema.clone(),
+            vec![Arc::new(Int32Vector::from_slice([1])) as VectorRef],
+        )
+        .unwrap();
+        let second_batch = RecordBatch::new(
+            schema.clone(),
+            vec![Arc::new(Int32Vector::from_slice([2])) as VectorRef],
+        )
+        .unwrap();
+        let output = output_from_flight_message_stream(futures_util::stream::iter(vec![
+            Ok(FlightMessage::Schema(schema.arrow_schema().clone())),
+            Ok(FlightMessage::RecordBatch(
+                first_batch.into_df_record_batch(),
+            )),
+            Ok(FlightMessage::Metrics(terminal_metrics_json_with_seq(1))),
+            Ok(FlightMessage::RecordBatch(
+                second_batch.into_df_record_batch(),
+            )),
+            Ok(FlightMessage::Metrics(terminal_metrics_json_with_seq(2))),
+        ]
+            as Vec<Result<FlightMessage>>))
+        .await
+        .unwrap();
+        let terminal_metrics = output.metrics.clone();
+        let OutputData::Stream(mut record_batch_stream) = output.output.data else {
+            panic!("expected stream output");
+        };
+
+        let first_batch = record_batch_stream.next().await.unwrap().unwrap();
+        assert_eq!(first_batch.num_rows(), 1);
+        let second_batch = record_batch_stream.next().await.unwrap().unwrap();
+        assert_eq!(second_batch.num_rows(), 1);
+        assert!(record_batch_stream.next().await.is_none());
+
+        assert!(terminal_metrics.is_ready());
+        assert_eq!(
+            terminal_metrics.region_watermark_map(),
+            Some(std::collections::HashMap::from([(7, 2)]))
+        );
+    }
+
+    #[test]
+    fn test_output_metrics_distinguishes_empty_region_watermarks_from_absence() {
+        let metrics = OutputMetrics::default();
+        metrics.update(Some(RecordBatchMetrics::default()));
+
+        assert_eq!(
+            metrics.participating_regions(),
+            Some(std::collections::BTreeSet::new())
+        );
+        assert_eq!(
+            metrics.region_watermark_map(),
+            Some(std::collections::HashMap::new())
+        );
+    }
 }
diff --git a/src/client/src/error.rs b/src/client/src/error.rs
index 5fdb383358..aa6940dd8b 100644
--- a/src/client/src/error.rs
+++ b/src/client/src/error.rs
@@ -173,20 +173,31 @@ impl ErrorExt for Error {
 define_from_tonic_status!(Error, Tonic);
 
 impl Error {
-    pub fn should_retry(&self) -> bool {
-        // TODO(weny): figure out each case of these codes.
-        matches!(
-            self,
-            Self::RegionServer {
-                code: Code::Cancelled,
-                ..
-            } | Self::RegionServer {
-                code: Code::DeadlineExceeded,
-                ..
-            } | Self::RegionServer {
-                code: Code::Unavailable,
-                ..
+    /// Returns the gRPC status code if this error is caused by a gRPC request failure.
+    pub fn tonic_code(&self) -> Option<Code> {
+        match self {
+            Self::FlightGet { tonic_code, .. }
+            | Self::RegionServer {
+                code: tonic_code, ..
             }
-        )
+            | Self::FlowServer {
+                code: tonic_code, ..
+            }
+            | Self::Tonic { tonic_code, .. } => Some(*tonic_code),
+            _ => None,
+        }
+    }
+
+    /// Returns true if the error is a connection error that may be resolved by retrying the request.
+    pub fn is_connection_error(&self) -> bool {
+        matches!(self.tonic_code(), Some(Code::Unavailable))
+    }
+
+    pub fn should_retry(&self) -> bool {
+        self.is_connection_error()
+            || matches!(
+                self.tonic_code(),
+                Some(Code::Cancelled) | Some(Code::DeadlineExceeded)
+            )
     }
 }
diff --git a/src/client/src/lib.rs b/src/client/src/lib.rs
index bf383acff9..147dffc145 100644
--- a/src/client/src/lib.rs
+++ b/src/client/src/lib.rs
@@ -12,8 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
-
 mod client;
 pub mod client_manager;
 pub mod database;
@@ -34,7 +32,7 @@ pub use common_recordbatch::{RecordBatches, SendableRecordBatchStream};
 use snafu::OptionExt;
 
 pub use self::client::Client;
-pub use self::database::Database;
+pub use self::database::{Database, OutputMetrics, OutputWithMetrics};
 pub use self::error::{Error, Result};
 use crate::error::{IllegalDatabaseResponseSnafu, ServerSnafu};
 
diff --git a/src/cmd/Cargo.toml b/src/cmd/Cargo.toml
index 74309e2024..2a5a599a91 100644
--- a/src/cmd/Cargo.toml
+++ b/src/cmd/Cargo.toml
@@ -30,10 +30,6 @@ base64.workspace = true
 cache.workspace = true
 catalog.workspace = true
 chrono.workspace = true
-datafusion-physical-plan.workspace = true
-datafusion.workspace = true
-datafusion-common.workspace = true
-either = "1.15"
 clap.workspace = true
 cli.workspace = true
 client.workspace = true
@@ -51,14 +47,19 @@ common-procedure.workspace = true
 common-query.workspace = true
 common-recordbatch.workspace = true
 common-runtime.workspace = true
+common-stat.workspace = true
 common-telemetry = { workspace = true, features = [
     "deadlock_detection",
 ] }
 common-time.workspace = true
 common-version.workspace = true
 common-wal.workspace = true
+datafusion.workspace = true
+datafusion-common.workspace = true
+datafusion-physical-plan.workspace = true
 datanode.workspace = true
 datatypes.workspace = true
+either = "1.15"
 etcd-client.workspace = true
 flow.workspace = true
 frontend = { workspace = true, default-features = false }
@@ -71,7 +72,7 @@ meta-client.workspace = true
 meta-srv.workspace = true
 metric-engine.workspace = true
 mito2.workspace = true
-moka.workspace = true
+moka = { workspace = true, features = ["future"] }
 object-store.workspace = true
 parquet = { workspace = true, features = ["object_store"] }
 plugins.workspace = true
@@ -81,21 +82,19 @@ query.workspace = true
 rand.workspace = true
 regex.workspace = true
 reqwest.workspace = true
-standalone.workspace = true
 serde.workspace = true
 serde_json.workspace = true
 servers.workspace = true
 session.workspace = true
-similar-asserts.workspace = true
 snafu.workspace = true
-common-stat.workspace = true
+sqlparser.workspace = true
+standalone.workspace = true
 store-api.workspace = true
 table.workspace = true
 tokio.workspace = true
 toml.workspace = true
 tonic.workspace = true
 tracing-appender.workspace = true
-sqlparser.workspace = true
 
 [target.'cfg(unix)'.dependencies]
 pprof = { version = "0.14", features = [
@@ -110,14 +109,9 @@ api.workspace = true
 client = { workspace = true, features = ["testing"] }
 common-test-util.workspace = true
 common-version.workspace = true
-serde.workspace = true
-temp-env = "0.3"
-tempfile.workspace = true
 file-engine.workspace = true
 mito2.workspace = true
-
-[target.'cfg(not(windows))'.dev-dependencies]
-rexpect = "0.5"
-
-[package.metadata.cargo-udeps.ignore]
-development = ["rexpect"]
+serde.workspace = true
+similar-asserts.workspace = true
+temp-env = "0.3"
+tempfile.workspace = true
diff --git a/src/cmd/src/bin/greptime.rs b/src/cmd/src/bin/greptime.rs
index a34d8e0f38..8c791bddcf 100644
--- a/src/cmd/src/bin/greptime.rs
+++ b/src/cmd/src/bin/greptime.rs
@@ -12,6 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#![recursion_limit = "256"]
 #![doc = include_str!("../../../../README.md")]
 
 use clap::{Parser, Subcommand};
@@ -20,11 +21,11 @@ use cmd::error::{InitTlsProviderSnafu, Result};
 use cmd::options::GlobalOptions;
 use cmd::{App, cli, datanode, flownode, frontend, metasrv, standalone};
 use common_base::Plugins;
-use common_version::{verbose_version, version};
-use servers::install_ring_crypto_provider;
+use common_version::{product_name, verbose_version, version};
+use servers::install_default_crypto_provider;
 
 #[derive(Parser)]
-#[command(name = "greptime", author, version, long_version = verbose_version(), about)]
+#[command(name = product_name(), author, version, long_version = verbose_version(), about)]
 #[command(propagate_version = true)]
 pub(crate) struct Command {
     #[clap(subcommand)]
@@ -52,11 +53,11 @@ enum SubCommand {
     #[clap(name = "metasrv")]
     Metasrv(metasrv::Command),
 
-    /// Run greptimedb as a standalone service.
+    /// Start service in standalone mode.
     #[clap(name = "standalone")]
     Standalone(standalone::Command),
 
-    /// Execute the cli tools for greptimedb.
+    /// Execute the cli tools.
     #[clap(name = "cli")]
     Cli(cli::Command),
 }
@@ -97,7 +98,7 @@ async fn main() -> Result<()> {
 
 async fn main_body() -> Result<()> {
     setup_human_panic();
-    install_ring_crypto_provider().map_err(|msg| InitTlsProviderSnafu { msg }.build())?;
+    install_default_crypto_provider().map_err(|msg| InitTlsProviderSnafu { msg }.build())?;
     start(Command::parse()).await
 }
 
@@ -148,7 +149,7 @@ async fn start(cli: Command) -> Result<()> {
 
 fn setup_human_panic() {
     human_panic::setup_panic!(
-        human_panic::Metadata::new("GreptimeDB", version())
+        human_panic::Metadata::new(product_name(), version())
             .homepage("https://github.com/GreptimeTeam/greptimedb/discussions")
     );
 
diff --git a/src/cmd/src/cli.rs b/src/cmd/src/cli.rs
index 84e797c291..09ec9646d8 100644
--- a/src/cmd/src/cli.rs
+++ b/src/cmd/src/cli.rs
@@ -102,31 +102,79 @@ impl Command {
 
 #[cfg(test)]
 mod tests {
+    use std::net::TcpListener;
+    use std::ops::RangeInclusive;
+
     use clap::Parser;
     use client::{Client, Database};
     use common_catalog::consts::{DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME};
     use common_telemetry::logging::LoggingOptions;
+    use rand::Rng;
 
     use crate::error::Result as CmdResult;
     use crate::options::GlobalOptions;
     use crate::{App, cli, standalone};
 
+    fn random_standalone_addrs() -> (String, String, String, String) {
+        let offset = choose_random_unused_port_offset(14000..=24000, 10);
+
+        (
+            format!("127.0.0.1:{}", 4000 + offset),
+            format!("127.0.0.1:{}", 4001 + offset),
+            format!("127.0.0.1:{}", 4002 + offset),
+            format!("127.0.0.1:{}", 4003 + offset),
+        )
+    }
+
+    fn choose_random_unused_port_offset(
+        port_range: RangeInclusive<u16>,
+        max_attempts: usize,
+    ) -> u16 {
+        let mut rng = rand::rng();
+
+        for _ in 0..max_attempts {
+            let http_port = rng.random_range(port_range.clone());
+            let offset = http_port - 4000;
+            let ports = [4000 + offset, 4001 + offset, 4002 + offset, 4003 + offset];
+
+            let listeners = ports
+                .into_iter()
+                .map(|port| TcpListener::bind(("127.0.0.1", port)))
+                .collect::<Result<Vec<_>, _>>();
+
+            if listeners.is_ok() {
+                return offset;
+            }
+        }
+
+        panic!("failed to find unused standalone test ports");
+    }
+
     #[tokio::test(flavor = "multi_thread")]
     async fn test_export_create_table_with_quoted_names() -> CmdResult<()> {
         let output_dir = tempfile::tempdir().unwrap();
+        let (http_addr, rpc_addr, mysql_addr, postgres_addr) = random_standalone_addrs();
 
         let standalone = standalone::Command::parse_from([
             "standalone",
             "start",
             "--data-home",
             &*output_dir.path().to_string_lossy(),
+            "--http-addr",
+            &http_addr,
+            "--grpc-bind-addr",
+            &rpc_addr,
+            "--mysql-addr",
+            &mysql_addr,
+            "--postgres-addr",
+            &postgres_addr,
         ]);
 
         let standalone_opts = standalone.load_options(&GlobalOptions::default()).unwrap();
         let mut instance = standalone.build(standalone_opts).await?;
         instance.start().await?;
 
-        let client = Client::with_urls(["127.0.0.1:4001"]);
+        let client = Client::with_urls([rpc_addr.as_str()]);
         let database = Database::new(DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME, client);
         database
             .sql(r#"CREATE DATABASE "cli.export.create_table";"#)
@@ -149,7 +197,7 @@ mod tests {
             "data",
             "export",
             "--addr",
-            "127.0.0.1:4000",
+            &http_addr,
             "--output-dir",
             &*output_dir.path().to_string_lossy(),
             "--target",
diff --git a/src/cmd/src/datanode.rs b/src/cmd/src/datanode.rs
index 06e2568b72..3c106bd43f 100644
--- a/src/cmd/src/datanode.rs
+++ b/src/cmd/src/datanode.rs
@@ -197,13 +197,17 @@ pub struct StartCommand {
     #[clap(long)]
     node_id: Option<u64>,
     /// The address to bind the gRPC server.
-    #[clap(long, alias = "rpc-addr")]
-    rpc_bind_addr: Option<String>,
+    #[clap(long = "grpc-bind-addr", alias = "rpc-bind-addr", alias = "rpc-addr")]
+    grpc_bind_addr: Option<String>,
     /// The address advertised to the metasrv, and used for connections from outside the host.
     /// If left empty or unset, the server will automatically use the IP address of the first network interface
-    /// on the host, with the same port number as the one specified in `rpc_bind_addr`.
-    #[clap(long, alias = "rpc-hostname")]
-    rpc_server_addr: Option<String>,
+    /// on the host, with the same port number as the one specified in `grpc_bind_addr`.
+    #[clap(
+        long = "grpc-server-addr",
+        alias = "rpc-server-addr",
+        alias = "rpc-hostname"
+    )]
+    grpc_server_addr: Option<String>,
     #[clap(long, value_delimiter = ',', num_args = 1..)]
     metasrv_addrs: Option<Vec<String>>,
     #[clap(short, long)]
@@ -256,20 +260,20 @@ impl StartCommand {
             tokio_console_addr: global_options.tokio_console_addr.clone(),
         };
 
-        if let Some(addr) = &self.rpc_bind_addr {
+        if let Some(addr) = &self.grpc_bind_addr {
             opts.grpc.bind_addr.clone_from(addr);
         } else if let Some(addr) = &opts.rpc_addr {
             warn!(
-                "Use the deprecated attribute `DatanodeOptions.rpc_addr`, please use `grpc.addr` instead."
+                "Use the deprecated attribute `DatanodeOptions.rpc_addr`, please use `grpc.bind_addr` instead."
             );
             opts.grpc.bind_addr.clone_from(addr);
         }
 
-        if let Some(server_addr) = &self.rpc_server_addr {
+        if let Some(server_addr) = &self.grpc_server_addr {
             opts.grpc.server_addr.clone_from(server_addr);
         } else if let Some(server_addr) = &opts.rpc_hostname {
             warn!(
-                "Use the deprecated attribute `DatanodeOptions.rpc_hostname`, please use `grpc.hostname` instead."
+                "Use the deprecated attribute `DatanodeOptions.rpc_hostname`, please use `grpc.server_addr` instead."
             );
             opts.grpc.server_addr.clone_from(server_addr);
         }
@@ -356,10 +360,11 @@ impl StartCommand {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::io::Write;
     use std::time::Duration;
 
+    use clap::{CommandFactory, Parser};
     use common_config::ENV_VAR_SEP;
     use common_test_util::temp_dir::create_named_temp_file;
     use object_store::config::{FileConfig, GcsConfig, ObjectStoreConfig, S3Config};
@@ -402,8 +407,8 @@ mod tests {
             node_id = 42
 
             [grpc]
-            addr = "127.0.0.1:3001"
-            hostname = "127.0.0.1"
+            bind_addr = "127.0.0.1:3001"
+            server_addr = "127.0.0.1"
             runtime_size = 8
 
             [meta_client]
@@ -449,6 +454,7 @@ mod tests {
         let options = cmd.load_options(&Default::default()).unwrap().component;
 
         assert_eq!("127.0.0.1:3001".to_string(), options.grpc.bind_addr);
+        assert_eq!("127.0.0.1".to_string(), options.grpc.server_addr);
         assert_eq!(Some(42), options.node_id);
 
         let DatanodeWalConfig::RaftEngine(raft_engine_config) = options.wal else {
@@ -661,4 +667,55 @@ mod tests {
             },
         );
     }
+
+    #[test]
+    fn test_parse_grpc_cli_aliases() {
+        let command = StartCommand::try_parse_from([
+            "datanode",
+            "--grpc-bind-addr",
+            "127.0.0.1:13001",
+            "--grpc-server-addr",
+            "10.0.0.1:13001",
+        ])
+        .unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:13001"));
+        assert_eq!(command.grpc_server_addr.as_deref(), Some("10.0.0.1:13001"));
+
+        let command = StartCommand::try_parse_from([
+            "datanode",
+            "--rpc-bind-addr",
+            "127.0.0.1:23001",
+            "--rpc-server-addr",
+            "10.0.0.2:23001",
+        ])
+        .unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:23001"));
+        assert_eq!(command.grpc_server_addr.as_deref(), Some("10.0.0.2:23001"));
+
+        let command = StartCommand::try_parse_from([
+            "datanode",
+            "--rpc-addr",
+            "127.0.0.1:33001",
+            "--rpc-hostname",
+            "10.0.0.3:33001",
+        ])
+        .unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:33001"));
+        assert_eq!(command.grpc_server_addr.as_deref(), Some("10.0.0.3:33001"));
+    }
+
+    #[test]
+    fn test_help_uses_grpc_option_names() {
+        let mut cmd = StartCommand::command();
+        let mut help = Vec::new();
+        cmd.write_long_help(&mut help).unwrap();
+        let help = String::from_utf8(help).unwrap();
+
+        assert!(help.contains("--grpc-bind-addr"));
+        assert!(help.contains("--grpc-server-addr"));
+        assert!(!help.contains("--rpc-bind-addr"));
+        assert!(!help.contains("--rpc-server-addr"));
+        assert!(!help.contains("--rpc-addr"));
+        assert!(!help.contains("--rpc-hostname"));
+    }
 }
diff --git a/src/cmd/src/datanode/builder.rs b/src/cmd/src/datanode/builder.rs
index cfde0c349a..90f38b22bd 100644
--- a/src/cmd/src/datanode/builder.rs
+++ b/src/cmd/src/datanode/builder.rs
@@ -15,7 +15,7 @@
 use std::sync::Arc;
 
 use cache::build_datanode_cache_registry;
-use catalog::kvbackend::MetaKvBackend;
+use catalog::kvbackend::new_read_only_meta_kv_backend;
 use common_base::Plugins;
 use common_meta::cache::LayeredCacheRegistryBuilder;
 use common_telemetry::info;
@@ -99,9 +99,7 @@ impl InstanceBuilder {
         .await
         .context(MetaClientInitSnafu)?;
 
-        let backend = Arc::new(MetaKvBackend {
-            client: client.clone(),
-        });
+        let backend = new_read_only_meta_kv_backend(client.clone());
         let mut builder = DatanodeBuilder::new(dn_opts.clone(), plugins.clone(), backend.clone());
 
         let registry = Arc::new(
diff --git a/src/cmd/src/datanode/objbench.rs b/src/cmd/src/datanode/objbench.rs
index d8f53b9d71..a298430c83 100644
--- a/src/cmd/src/datanode/objbench.rs
+++ b/src/cmd/src/datanode/objbench.rs
@@ -20,13 +20,14 @@ use clap::Parser;
 use colored::Colorize;
 use datanode::config::RegionEngineConfig;
 use datanode::store;
-use either::Either;
+use futures::stream;
 use mito2::access_layer::{
     AccessLayer, AccessLayerRef, Metrics, OperationType, SstWriteRequest, WriteType,
 };
 use mito2::cache::{CacheManager, CacheManagerRef};
 use mito2::config::{FulltextIndexConfig, MitoConfig, Mode};
-use mito2::read::Source;
+use mito2::read::FlatSource;
+use mito2::sst::FormatType;
 use mito2::sst::file::{FileHandle, FileMeta};
 use mito2::sst::file_purger::{FilePurger, FilePurgerRef};
 use mito2::sst::index::intermediate::IntermediateManager;
@@ -189,6 +190,7 @@ impl ObjbenchCommand {
             sequence: None,
             partition_expr: None,
             num_series: 0,
+            ..Default::default()
         };
         let src_handle = FileHandle::new(file_meta, new_noop_file_purger());
 
@@ -231,6 +233,10 @@ impl ObjbenchCommand {
         let reader_build_elapsed = reader_build_start.elapsed();
         let total_rows = reader.parquet_metadata().file_metadata().num_rows();
         println!("{} Reader built in {:?}", "✓".green(), reader_build_elapsed);
+        let reader_stream = Box::pin(stream::try_unfold(reader, |mut reader| async move {
+            let batch = reader.next_record_batch().await?;
+            Ok(batch.map(|batch| (batch, reader)))
+        }));
 
         // Build write request
         let fulltext_index_config = FulltextIndexConfig {
@@ -238,13 +244,16 @@ impl ObjbenchCommand {
             ..Default::default()
         };
 
+        let source =
+            FlatSource::new_stream(region_meta.schema.arrow_schema().clone(), reader_stream);
         let write_req = SstWriteRequest {
             op_type: OperationType::Flush,
             metadata: region_meta,
-            source: Either::Left(Source::Reader(Box::new(reader))),
+            source,
             cache_manager,
             storage: None,
             max_sequence: None,
+            sst_write_format: FormatType::PrimaryKey,
             index_options: Default::default(),
             index_config: mito_engine_config.index.clone(),
             inverted_index_config: MitoConfig::default().inverted_index,
diff --git a/src/cmd/src/datanode/scanbench.rs b/src/cmd/src/datanode/scanbench.rs
index fdda1d97bb..b26705991c 100644
--- a/src/cmd/src/datanode/scanbench.rs
+++ b/src/cmd/src/datanode/scanbench.rs
@@ -53,7 +53,9 @@ use store_api::metadata::RegionMetadata;
 use store_api::path_utils::WAL_DIR;
 use store_api::region_engine::{PrepareRequest, QueryScanContext, RegionEngine};
 use store_api::region_request::{PathType, RegionOpenRequest, RegionRequest};
-use store_api::storage::{RegionId, ScanRequest, TimeSeriesDistribution, TimeSeriesRowSelector};
+use store_api::storage::{
+    ProjectionInput, RegionId, ScanRequest, TimeSeriesDistribution, TimeSeriesRowSelector,
+};
 use tokio::fs;
 
 use crate::datanode::objbench::{build_object_store, parse_config};
@@ -102,10 +104,6 @@ pub struct ScanbenchCommand {
     #[clap(long, value_name = "FILE")]
     pprof_file: Option<PathBuf>,
 
-    /// Force reading the region in flat format.
-    #[clap(long, default_value_t = false)]
-    force_flat_format: bool,
-
     /// Enable WAL replay when opening the region.
     #[clap(long, default_value_t = false)]
     enable_wal: bool,
@@ -580,12 +578,11 @@ impl ScanbenchCommand {
         };
 
         println!(
-            "{} Scanner: {}, Parallelism: {}, Iterations: {}, Force flat format: {}",
+            "{} Scanner: {}, Parallelism: {}, Iterations: {}",
             "ℹ".blue(),
             self.scanner,
             self.parallelism,
             self.iterations,
-            self.force_flat_format,
         );
 
         // Start profiling if pprof_file is specified (unless pprof_after_warmup is set)
@@ -620,13 +617,13 @@ impl ScanbenchCommand {
         let mut total_rows_all = 0u64;
         let mut total_elapsed_all = std::time::Duration::ZERO;
 
+        let projection_input = projection.map(ProjectionInput::new);
         for iteration in 0..self.iterations {
             let request = ScanRequest {
-                projection: projection.clone(),
+                projection_input: projection_input.clone(),
                 filters: filters.clone(),
                 series_row_selector,
                 distribution,
-                force_flat_format: self.force_flat_format,
                 ..Default::default()
             };
 
@@ -662,7 +659,7 @@ impl ScanbenchCommand {
 
                 // Sort ranges within each partition by start time ascending
                 for partition in &mut partitions {
-                    partition.sort_by(|a, b| a.start.cmp(&b.start));
+                    partition.sort_by_key(|a| a.start);
                 }
 
                 scanner
@@ -677,7 +674,9 @@ impl ScanbenchCommand {
 
             // Scan all partitions
             let num_partitions = scanner.properties().partitions.len();
-            let ctx = QueryScanContext::default();
+            let ctx = QueryScanContext {
+                explain_verbose: self.verbose,
+            };
             let metrics_set = ExecutionPlanMetricsSet::new();
 
             let mut scan_futures = FuturesUnordered::new();
diff --git a/src/cmd/src/flownode.rs b/src/cmd/src/flownode.rs
index 3f8458cddf..df7a0725e9 100644
--- a/src/cmd/src/flownode.rs
+++ b/src/cmd/src/flownode.rs
@@ -19,7 +19,9 @@ use std::time::Duration;
 
 use cache::{build_fundamental_cache_registry, with_default_composite_cache_registry};
 use catalog::information_extension::DistributedInformationExtension;
-use catalog::kvbackend::{CachedKvBackendBuilder, KvBackendCatalogManagerBuilder, MetaKvBackend};
+use catalog::kvbackend::{
+    CachedKvBackendBuilder, KvBackendCatalogManagerBuilder, new_read_only_meta_kv_backend,
+};
 use clap::Parser;
 use client::client_manager::NodeClients;
 use common_base::Plugins;
@@ -46,8 +48,8 @@ use snafu::{OptionExt, ResultExt, ensure};
 use tracing_appender::non_blocking::WorkerGuard;
 
 use crate::error::{
-    BuildCacheRegistrySnafu, InitMetadataSnafu, LoadLayeredConfigSnafu, MetaClientInitSnafu,
-    MissingConfigSnafu, OtherSnafu, Result, ShutdownFlownodeSnafu, StartFlownodeSnafu,
+    BuildCacheRegistrySnafu, LoadLayeredConfigSnafu, MetaClientInitSnafu, MissingConfigSnafu,
+    OtherSnafu, Result, ShutdownFlownodeSnafu, StartFlownodeSnafu,
 };
 use crate::options::{GlobalOptions, GreptimeOptions};
 use crate::{App, create_resource_limit_metrics, log_versions, maybe_activate_heap_profile};
@@ -139,13 +141,17 @@ struct StartCommand {
     #[clap(long)]
     node_id: Option<u64>,
     /// Bind address for the gRPC server.
-    #[clap(long, alias = "rpc-addr")]
-    rpc_bind_addr: Option<String>,
+    #[clap(long = "grpc-bind-addr", alias = "rpc-bind-addr", alias = "rpc-addr")]
+    grpc_bind_addr: Option<String>,
     /// The address advertised to the metasrv, and used for connections from outside the host.
     /// If left empty or unset, the server will automatically use the IP address of the first network interface
-    /// on the host, with the same port number as the one specified in `rpc_bind_addr`.
-    #[clap(long, alias = "rpc-hostname")]
-    rpc_server_addr: Option<String>,
+    /// on the host, with the same port number as the one specified in `grpc_bind_addr`.
+    #[clap(
+        long = "grpc-server-addr",
+        alias = "rpc-server-addr",
+        alias = "rpc-hostname"
+    )]
+    grpc_server_addr: Option<String>,
     /// Metasrv address list;
     #[clap(long, value_delimiter = ',', num_args = 1..)]
     metasrv_addrs: Option<Vec<String>>,
@@ -207,11 +213,11 @@ impl StartCommand {
             tokio_console_addr: global_options.tokio_console_addr.clone(),
         };
 
-        if let Some(addr) = &self.rpc_bind_addr {
+        if let Some(addr) = &self.grpc_bind_addr {
             opts.grpc.bind_addr.clone_from(addr);
         }
 
-        if let Some(server_addr) = &self.rpc_server_addr {
+        if let Some(server_addr) = &self.grpc_server_addr {
             opts.grpc.server_addr.clone_from(server_addr);
         }
 
@@ -296,13 +302,14 @@ impl StartCommand {
         let cache_ttl = meta_config.metadata_cache_ttl;
         let cache_tti = meta_config.metadata_cache_tti;
 
+        let readonly_meta_backend = new_read_only_meta_kv_backend(meta_client.clone());
+
         // TODO(discord9): add helper function to ease the creation of cache registry&such
-        let cached_meta_backend =
-            CachedKvBackendBuilder::new(Arc::new(MetaKvBackend::new(meta_client.clone())))
-                .cache_max_capacity(cache_max_capacity)
-                .cache_ttl(cache_ttl)
-                .cache_tti(cache_tti)
-                .build();
+        let cached_meta_backend = CachedKvBackendBuilder::new(readonly_meta_backend.clone())
+            .cache_max_capacity(cache_max_capacity)
+            .cache_ttl(cache_ttl)
+            .cache_tti(cache_tti)
+            .build();
         let cached_meta_backend = Arc::new(cached_meta_backend);
 
         // Builds cache registry
@@ -312,7 +319,7 @@ impl StartCommand {
                 .build(),
         );
         let fundamental_cache_registry =
-            build_fundamental_cache_registry(Arc::new(MetaKvBackend::new(meta_client.clone())));
+            build_fundamental_cache_registry(readonly_meta_backend.clone());
         let layered_cache_registry = Arc::new(
             with_default_composite_cache_registry(
                 layered_cache_builder.add_cache_registry(fundamental_cache_registry),
@@ -342,10 +349,6 @@ impl StartCommand {
 
         let table_metadata_manager =
             Arc::new(TableMetadataManager::new(cached_meta_backend.clone()));
-        table_metadata_manager
-            .init()
-            .await
-            .context(InitMetadataSnafu)?;
 
         let executor = HandlerGroupExecutor::new(vec![
             Arc::new(ParseMailboxMessageHandler),
@@ -432,3 +435,61 @@ impl StartCommand {
         Ok(Instance::new(flownode, guard))
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use clap::{CommandFactory, Parser};
+
+    use super::*;
+
+    #[test]
+    fn test_parse_grpc_cli_aliases() {
+        let command = StartCommand::try_parse_from([
+            "flownode",
+            "--grpc-bind-addr",
+            "127.0.0.1:14004",
+            "--grpc-server-addr",
+            "10.0.0.1:14004",
+        ])
+        .unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:14004"));
+        assert_eq!(command.grpc_server_addr.as_deref(), Some("10.0.0.1:14004"));
+
+        let command = StartCommand::try_parse_from([
+            "flownode",
+            "--rpc-bind-addr",
+            "127.0.0.1:24004",
+            "--rpc-server-addr",
+            "10.0.0.2:24004",
+        ])
+        .unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:24004"));
+        assert_eq!(command.grpc_server_addr.as_deref(), Some("10.0.0.2:24004"));
+
+        let command = StartCommand::try_parse_from([
+            "flownode",
+            "--rpc-addr",
+            "127.0.0.1:34004",
+            "--rpc-hostname",
+            "10.0.0.3:34004",
+        ])
+        .unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:34004"));
+        assert_eq!(command.grpc_server_addr.as_deref(), Some("10.0.0.3:34004"));
+    }
+
+    #[test]
+    fn test_help_uses_grpc_option_names() {
+        let mut cmd = StartCommand::command();
+        let mut help = Vec::new();
+        cmd.write_long_help(&mut help).unwrap();
+        let help = String::from_utf8(help).unwrap();
+
+        assert!(help.contains("--grpc-bind-addr"));
+        assert!(help.contains("--grpc-server-addr"));
+        assert!(!help.contains("--rpc-bind-addr"));
+        assert!(!help.contains("--rpc-server-addr"));
+        assert!(!help.contains("--rpc-addr"));
+        assert!(!help.contains("--rpc-hostname"));
+    }
+}
diff --git a/src/cmd/src/frontend.rs b/src/cmd/src/frontend.rs
index cb802791c5..5e19e90f89 100644
--- a/src/cmd/src/frontend.rs
+++ b/src/cmd/src/frontend.rs
@@ -23,7 +23,7 @@ use catalog::information_extension::DistributedInformationExtension;
 use catalog::information_schema::InformationExtensionRef;
 use catalog::kvbackend::{
     CachedKvBackendBuilder, CatalogManagerConfiguratorRef, KvBackendCatalogManagerBuilder,
-    MetaKvBackend,
+    new_read_only_meta_kv_backend,
 };
 use catalog::process_manager::ProcessManager;
 use clap::Parser;
@@ -152,21 +152,33 @@ impl SubCommand {
 #[derive(Debug, Default, Parser)]
 pub struct StartCommand {
     /// The address to bind the gRPC server.
-    #[clap(long, alias = "rpc-addr")]
-    rpc_bind_addr: Option<String>,
+    #[clap(long = "grpc-bind-addr", alias = "rpc-bind-addr", alias = "rpc-addr")]
+    grpc_bind_addr: Option<String>,
     /// The address advertised to the metasrv, and used for connections from outside the host.
     /// If left empty or unset, the server will automatically use the IP address of the first network interface
-    /// on the host, with the same port number as the one specified in `rpc_bind_addr`.
-    #[clap(long, alias = "rpc-hostname")]
-    rpc_server_addr: Option<String>,
+    /// on the host, with the same port number as the one specified in `grpc_bind_addr`.
+    #[clap(
+        long = "grpc-server-addr",
+        alias = "rpc-server-addr",
+        alias = "rpc-hostname"
+    )]
+    grpc_server_addr: Option<String>,
     /// The address to bind the internal gRPC server.
-    #[clap(long, alias = "internal-rpc-addr")]
-    internal_rpc_bind_addr: Option<String>,
+    #[clap(
+        long = "internal-grpc-bind-addr",
+        alias = "internal-rpc-bind-addr",
+        alias = "internal-rpc-addr"
+    )]
+    internal_grpc_bind_addr: Option<String>,
     /// The address advertised to the metasrv, and used for connections from outside the host.
     /// If left empty or unset, the server will automatically use the IP address of the first network interface
-    /// on the host, with the same port number as the one specified in `internal_rpc_bind_addr`.
-    #[clap(long, alias = "internal-rpc-hostname")]
-    internal_rpc_server_addr: Option<String>,
+    /// on the host, with the same port number as the one specified in `internal_grpc_bind_addr`.
+    #[clap(
+        long = "internal-grpc-server-addr",
+        alias = "internal-rpc-server-addr",
+        alias = "internal-rpc-hostname"
+    )]
+    internal_grpc_server_addr: Option<String>,
     #[clap(long)]
     http_addr: Option<String>,
     #[clap(long)]
@@ -258,16 +270,16 @@ impl StartCommand {
             opts.http.disable_dashboard = disable_dashboard;
         }
 
-        if let Some(addr) = &self.rpc_bind_addr {
+        if let Some(addr) = &self.grpc_bind_addr {
             opts.grpc.bind_addr.clone_from(addr);
             opts.grpc.tls = merge_tls_option(&opts.grpc.tls, tls_opts.clone());
         }
 
-        if let Some(addr) = &self.rpc_server_addr {
+        if let Some(addr) = &self.grpc_server_addr {
             opts.grpc.server_addr.clone_from(addr);
         }
 
-        if let Some(addr) = &self.internal_rpc_bind_addr {
+        if let Some(addr) = &self.internal_grpc_bind_addr {
             if let Some(internal_grpc) = &mut opts.internal_grpc {
                 internal_grpc.bind_addr = addr.clone();
             } else {
@@ -280,7 +292,7 @@ impl StartCommand {
             }
         }
 
-        if let Some(addr) = &self.internal_rpc_server_addr {
+        if let Some(addr) = &self.internal_grpc_server_addr {
             if let Some(internal_grpc) = &mut opts.internal_grpc {
                 internal_grpc.server_addr = addr.clone();
             } else {
@@ -381,13 +393,14 @@ impl StartCommand {
             .await
             .context(error::StartFrontendSnafu)?;
 
+        let readonly_meta_backend = new_read_only_meta_kv_backend(meta_client.clone());
+
         // TODO(discord9): add helper function to ease the creation of cache registry&such
-        let cached_meta_backend =
-            CachedKvBackendBuilder::new(Arc::new(MetaKvBackend::new(meta_client.clone())))
-                .cache_max_capacity(cache_max_capacity)
-                .cache_ttl(cache_ttl)
-                .cache_tti(cache_tti)
-                .build();
+        let cached_meta_backend = CachedKvBackendBuilder::new(readonly_meta_backend.clone())
+            .cache_max_capacity(cache_max_capacity)
+            .cache_ttl(cache_ttl)
+            .cache_tti(cache_tti)
+            .build();
         let cached_meta_backend = Arc::new(cached_meta_backend);
 
         // Builds cache registry
@@ -397,7 +410,7 @@ impl StartCommand {
                 .build(),
         );
         let fundamental_cache_registry =
-            build_fundamental_cache_registry(Arc::new(MetaKvBackend::new(meta_client.clone())));
+            build_fundamental_cache_registry(readonly_meta_backend.clone());
         let layered_cache_registry = Arc::new(
             with_default_composite_cache_registry(
                 layered_cache_builder.add_cache_registry(fundamental_cache_registry),
@@ -515,6 +528,7 @@ mod tests {
     use std::time::Duration;
 
     use auth::{Identity, Password, UserProviderRef};
+    use clap::{CommandFactory, Parser};
     use common_base::readable_size::ReadableSize;
     use common_config::ENV_VAR_SEP;
     use common_test_util::temp_dir::create_named_temp_file;
@@ -530,8 +544,8 @@ mod tests {
             http_addr: Some("127.0.0.1:1234".to_string()),
             mysql_addr: Some("127.0.0.1:5678".to_string()),
             postgres_addr: Some("127.0.0.1:5432".to_string()),
-            internal_rpc_bind_addr: Some("127.0.0.1:4010".to_string()),
-            internal_rpc_server_addr: Some("10.0.0.24:4010".to_string()),
+            internal_grpc_bind_addr: Some("127.0.0.1:4010".to_string()),
+            internal_grpc_server_addr: Some("10.0.0.24:4010".to_string()),
             influxdb_enable: Some(false),
             disable_dashboard: Some(false),
             ..Default::default()
@@ -609,6 +623,7 @@ mod tests {
                 disable_dashboard: false,
                 ..Default::default()
             },
+            meta_client: Some(MetaClientOptions::default()),
             user_provider: Some("static_user_provider:cmd:test=test".to_string()),
             ..Default::default()
         };
@@ -744,4 +759,97 @@ mod tests {
             },
         );
     }
+
+    #[test]
+    fn test_parse_grpc_cli_aliases() {
+        let command = StartCommand::try_parse_from([
+            "frontend",
+            "--grpc-bind-addr",
+            "127.0.0.1:14001",
+            "--grpc-server-addr",
+            "10.0.0.1:14001",
+            "--internal-grpc-bind-addr",
+            "127.0.0.1:14010",
+            "--internal-grpc-server-addr",
+            "10.0.0.1:14010",
+        ])
+        .unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:14001"));
+        assert_eq!(command.grpc_server_addr.as_deref(), Some("10.0.0.1:14001"));
+        assert_eq!(
+            command.internal_grpc_bind_addr.as_deref(),
+            Some("127.0.0.1:14010")
+        );
+        assert_eq!(
+            command.internal_grpc_server_addr.as_deref(),
+            Some("10.0.0.1:14010")
+        );
+
+        let command = StartCommand::try_parse_from([
+            "frontend",
+            "--rpc-bind-addr",
+            "127.0.0.1:24001",
+            "--rpc-server-addr",
+            "10.0.0.2:24001",
+            "--internal-rpc-bind-addr",
+            "127.0.0.1:24010",
+            "--internal-rpc-server-addr",
+            "10.0.0.2:24010",
+        ])
+        .unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:24001"));
+        assert_eq!(command.grpc_server_addr.as_deref(), Some("10.0.0.2:24001"));
+        assert_eq!(
+            command.internal_grpc_bind_addr.as_deref(),
+            Some("127.0.0.1:24010")
+        );
+        assert_eq!(
+            command.internal_grpc_server_addr.as_deref(),
+            Some("10.0.0.2:24010")
+        );
+
+        let command = StartCommand::try_parse_from([
+            "frontend",
+            "--rpc-addr",
+            "127.0.0.1:34001",
+            "--rpc-hostname",
+            "10.0.0.3:34001",
+            "--internal-rpc-addr",
+            "127.0.0.1:34010",
+            "--internal-rpc-hostname",
+            "10.0.0.3:34010",
+        ])
+        .unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:34001"));
+        assert_eq!(command.grpc_server_addr.as_deref(), Some("10.0.0.3:34001"));
+        assert_eq!(
+            command.internal_grpc_bind_addr.as_deref(),
+            Some("127.0.0.1:34010")
+        );
+        assert_eq!(
+            command.internal_grpc_server_addr.as_deref(),
+            Some("10.0.0.3:34010")
+        );
+    }
+
+    #[test]
+    fn test_help_uses_grpc_option_names() {
+        let mut cmd = StartCommand::command();
+        let mut help = Vec::new();
+        cmd.write_long_help(&mut help).unwrap();
+        let help = String::from_utf8(help).unwrap();
+
+        assert!(help.contains("--grpc-bind-addr"));
+        assert!(help.contains("--grpc-server-addr"));
+        assert!(help.contains("--internal-grpc-bind-addr"));
+        assert!(help.contains("--internal-grpc-server-addr"));
+        assert!(!help.contains("--rpc-bind-addr"));
+        assert!(!help.contains("--rpc-server-addr"));
+        assert!(!help.contains("--rpc-addr"));
+        assert!(!help.contains("--rpc-hostname"));
+        assert!(!help.contains("--internal-rpc-bind-addr"));
+        assert!(!help.contains("--internal-rpc-server-addr"));
+        assert!(!help.contains("--internal-rpc-addr"));
+        assert!(!help.contains("--internal-rpc-hostname"));
+    }
 }
diff --git a/src/cmd/src/lib.rs b/src/cmd/src/lib.rs
index 46ca4c8a76..27564597d7 100644
--- a/src/cmd/src/lib.rs
+++ b/src/cmd/src/lib.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
+#![recursion_limit = "256"]
 
 use async_trait::async_trait;
 use common_error::ext::ErrorExt;
diff --git a/src/cmd/src/metasrv.rs b/src/cmd/src/metasrv.rs
index 2ce5fb3a02..bf3cb2f5e7 100644
--- a/src/cmd/src/metasrv.rs
+++ b/src/cmd/src/metasrv.rs
@@ -21,8 +21,8 @@ use clap::Parser;
 use common_base::Plugins;
 use common_config::Configurable;
 use common_meta::distributed_time_constants::init_distributed_time_constants;
-use common_telemetry::info;
 use common_telemetry::logging::{DEFAULT_LOGGING_DIR, TracingOptions};
+use common_telemetry::{info, warn};
 use common_version::{short_version, verbose_version};
 use meta_srv::bootstrap::{MetasrvInstance, metasrv_builder};
 use meta_srv::metasrv::BackendImpl;
@@ -141,13 +141,17 @@ impl SubCommand {
 #[derive(Default, Parser)]
 pub struct StartCommand {
     /// The address to bind the gRPC server.
-    #[clap(long, alias = "bind-addr")]
-    rpc_bind_addr: Option<String>,
+    #[clap(long = "grpc-bind-addr", alias = "rpc-bind-addr", alias = "bind-addr")]
+    grpc_bind_addr: Option<String>,
     /// The communication server address for the frontend and datanode to connect to metasrv.
     /// If left empty or unset, the server will automatically use the IP address of the first network interface
-    /// on the host, with the same port number as the one specified in `rpc_bind_addr`.
-    #[clap(long, alias = "server-addr")]
-    rpc_server_addr: Option<String>,
+    /// on the host, with the same port number as the one specified in `grpc_bind_addr`.
+    #[clap(
+        long = "grpc-server-addr",
+        alias = "rpc-server-addr",
+        alias = "server-addr"
+    )]
+    grpc_server_addr: Option<String>,
     #[clap(long, alias = "store-addr", value_delimiter = ',', num_args = 1..)]
     store_addrs: Option<Vec<String>>,
     #[clap(short, long)]
@@ -179,8 +183,8 @@ pub struct StartCommand {
 impl Debug for StartCommand {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         f.debug_struct("StartCommand")
-            .field("rpc_bind_addr", &self.rpc_bind_addr)
-            .field("rpc_server_addr", &self.rpc_server_addr)
+            .field("grpc_bind_addr", &self.grpc_bind_addr)
+            .field("grpc_server_addr", &self.grpc_server_addr)
             .field("store_addrs", &self.sanitize_store_addrs())
             .field("config_file", &self.config_file)
             .field("selector", &self.selector)
@@ -240,18 +244,24 @@ impl StartCommand {
         };
 
         #[allow(deprecated)]
-        if let Some(addr) = &self.rpc_bind_addr {
+        if let Some(addr) = &self.grpc_bind_addr {
             opts.bind_addr.clone_from(addr);
             opts.grpc.bind_addr.clone_from(addr);
         } else if !opts.bind_addr.is_empty() {
+            warn!(
+                "Use the deprecated attribute `MetasrvOptions.bind_addr`, please use `grpc.bind_addr` instead."
+            );
             opts.grpc.bind_addr.clone_from(&opts.bind_addr);
         }
 
         #[allow(deprecated)]
-        if let Some(addr) = &self.rpc_server_addr {
+        if let Some(addr) = &self.grpc_server_addr {
             opts.server_addr.clone_from(addr);
             opts.grpc.server_addr.clone_from(addr);
         } else if !opts.server_addr.is_empty() {
+            warn!(
+                "Use the deprecated attribute `MetasrvOptions.server_addr`, please use `grpc.server_addr` instead."
+            );
             opts.grpc.server_addr.clone_from(&opts.server_addr);
         }
 
@@ -353,6 +363,7 @@ impl StartCommand {
 mod tests {
     use std::io::Write;
 
+    use clap::{CommandFactory, Parser};
     use common_base::readable_size::ReadableSize;
     use common_config::ENV_VAR_SEP;
     use common_test_util::temp_dir::create_named_temp_file;
@@ -363,8 +374,8 @@ mod tests {
     #[test]
     fn test_read_from_cmd() {
         let cmd = StartCommand {
-            rpc_bind_addr: Some("127.0.0.1:3002".to_string()),
-            rpc_server_addr: Some("127.0.0.1:3002".to_string()),
+            grpc_bind_addr: Some("127.0.0.1:3002".to_string()),
+            grpc_server_addr: Some("127.0.0.1:3002".to_string()),
             store_addrs: Some(vec!["127.0.0.1:2380".to_string()]),
             selector: Some("LoadBased".to_string()),
             ..Default::default()
@@ -432,8 +443,8 @@ mod tests {
     #[test]
     fn test_load_log_options_from_cli() {
         let cmd = StartCommand {
-            rpc_bind_addr: Some("127.0.0.1:3002".to_string()),
-            rpc_server_addr: Some("127.0.0.1:3002".to_string()),
+            grpc_bind_addr: Some("127.0.0.1:3002".to_string()),
+            grpc_server_addr: Some("127.0.0.1:3002".to_string()),
             store_addrs: Some(vec!["127.0.0.1:2380".to_string()]),
             selector: Some("LoadBased".to_string()),
             ..Default::default()
@@ -520,4 +531,55 @@ mod tests {
             },
         );
     }
+
+    #[test]
+    fn test_parse_grpc_cli_aliases() {
+        let command = StartCommand::try_parse_from([
+            "metasrv",
+            "--grpc-bind-addr",
+            "127.0.0.1:13002",
+            "--grpc-server-addr",
+            "10.0.0.1:13002",
+        ])
+        .unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:13002"));
+        assert_eq!(command.grpc_server_addr.as_deref(), Some("10.0.0.1:13002"));
+
+        let command = StartCommand::try_parse_from([
+            "metasrv",
+            "--rpc-bind-addr",
+            "127.0.0.1:23002",
+            "--rpc-server-addr",
+            "10.0.0.2:23002",
+        ])
+        .unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:23002"));
+        assert_eq!(command.grpc_server_addr.as_deref(), Some("10.0.0.2:23002"));
+
+        let command = StartCommand::try_parse_from([
+            "metasrv",
+            "--bind-addr",
+            "127.0.0.1:33002",
+            "--server-addr",
+            "10.0.0.3:33002",
+        ])
+        .unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:33002"));
+        assert_eq!(command.grpc_server_addr.as_deref(), Some("10.0.0.3:33002"));
+    }
+
+    #[test]
+    fn test_help_uses_grpc_option_names() {
+        let mut cmd = StartCommand::command();
+        let mut help = Vec::new();
+        cmd.write_long_help(&mut help).unwrap();
+        let help = String::from_utf8(help).unwrap();
+
+        assert!(help.contains("--grpc-bind-addr"));
+        assert!(help.contains("--grpc-server-addr"));
+        assert!(!help.contains("--rpc-bind-addr"));
+        assert!(!help.contains("--rpc-server-addr"));
+        assert!(!help.contains("--bind-addr"));
+        assert!(!help.contains("--server-addr"));
+    }
 }
diff --git a/src/cmd/src/standalone.rs b/src/cmd/src/standalone.rs
index 92638d3c4a..6f35e74b65 100644
--- a/src/cmd/src/standalone.rs
+++ b/src/cmd/src/standalone.rs
@@ -32,15 +32,17 @@ use common_meta::cache::LayeredCacheRegistryBuilder;
 use common_meta::ddl::flow_meta::FlowMetadataAllocator;
 use common_meta::ddl::table_meta::TableMetadataAllocator;
 use common_meta::ddl::{DdlContext, NoopRegionFailureDetectorControl};
-use common_meta::ddl_manager::{DdlManager, DdlManagerConfiguratorRef};
+use common_meta::ddl_manager::{DdlManager, DdlManagerConfiguratorRef, DdlManagerRef};
 use common_meta::key::flow::FlowMetadataManager;
 use common_meta::key::{TableMetadataManager, TableMetadataManagerRef};
 use common_meta::kv_backend::KvBackendRef;
-use common_meta::procedure_executor::LocalProcedureExecutor;
+use common_meta::node_manager::{FlownodeRef, NodeManagerRef};
+use common_meta::procedure_executor::{LocalProcedureExecutor, ProcedureExecutorRef};
 use common_meta::region_keeper::MemoryRegionKeeper;
 use common_meta::region_registry::LeaderRegionRegistry;
-use common_meta::sequence::SequenceBuilder;
+use common_meta::sequence::{Sequence, SequenceBuilder};
 use common_meta::wal_provider::{WalProviderRef, build_wal_provider};
+use common_options::plugin_options::StandaloneFlag;
 use common_procedure::ProcedureManagerRef;
 use common_query::prelude::set_default_prefix;
 use common_telemetry::info;
@@ -49,6 +51,7 @@ use common_time::timezone::set_default_timezone;
 use common_version::{short_version, verbose_version};
 use datanode::config::DatanodeOptions;
 use datanode::datanode::{Datanode, DatanodeBuilder};
+use datanode::region_server::RegionServer;
 use flow::{
     FlownodeBuilder, FlownodeInstance, FlownodeOptions, FrontendClient, FrontendInvoker,
     GrpcQueryHandlerWithBoxedError,
@@ -58,6 +61,7 @@ use frontend::instance::StandaloneDatanodeManager;
 use frontend::instance::builder::FrontendBuilder;
 use frontend::server::Services;
 use meta_srv::metasrv::{FLOW_ID_SEQ, TABLE_ID_SEQ};
+use plugins::PluginOptions;
 use plugins::frontend::context::{
     CatalogManagerConfigureContext, StandaloneCatalogManagerConfigureContext,
 };
@@ -121,6 +125,7 @@ pub struct Instance {
     flownode: FlownodeInstance,
     procedure_manager: ProcedureManagerRef,
     wal_provider: WalProviderRef,
+    leader_services_controller: Box<dyn StandaloneLeaderServicesController>,
     // Keep the logging guard to prevent the worker from being dropped.
     _guard: Vec<WorkerGuard>,
 }
@@ -130,6 +135,18 @@ impl Instance {
     pub fn server_addr(&self, name: &str) -> Option<SocketAddr> {
         self.frontend.server_handlers().addr(name)
     }
+
+    /// Get the mutable Frontend component of this Standalone instance for externally modification
+    /// by others (might not be in this code base, so don't delete this function).
+    pub fn mut_frontend(&mut self) -> &mut Frontend {
+        &mut self.frontend
+    }
+
+    /// Get the Datanode component of this Standalone instance for externally usage
+    /// by others (might not be in this code base, so don't delete this function).
+    pub fn datanode(&self) -> &Datanode {
+        &self.datanode
+    }
 }
 
 #[async_trait]
@@ -141,15 +158,13 @@ impl App for Instance {
     async fn start(&mut self) -> Result<()> {
         self.datanode.start_telemetry();
 
-        self.procedure_manager
-            .start()
-            .await
-            .context(error::StartProcedureManagerSnafu)?;
-
-        self.wal_provider
-            .start()
-            .await
-            .context(error::StartWalProviderSnafu)?;
+        self.leader_services_controller
+            .start(
+                self.procedure_manager.clone(),
+                self.wal_provider.clone(),
+                self.datanode.region_server(),
+            )
+            .await?;
 
         plugins::start_frontend_plugins(self.frontend.instance.plugins().clone())
             .await
@@ -171,10 +186,12 @@ impl App for Instance {
             .await
             .context(error::ShutdownFrontendSnafu)?;
 
-        self.procedure_manager
-            .stop()
-            .await
-            .context(error::StopProcedureManagerSnafu)?;
+        self.leader_services_controller
+            .stop(
+                self.procedure_manager.clone(),
+                self.datanode.region_server(),
+            )
+            .await?;
 
         self.datanode
             .shutdown()
@@ -196,8 +213,8 @@ impl App for Instance {
 pub struct StartCommand {
     #[clap(long)]
     http_addr: Option<String>,
-    #[clap(long, alias = "rpc-addr")]
-    rpc_bind_addr: Option<String>,
+    #[clap(long = "grpc-bind-addr", alias = "rpc-bind-addr", alias = "rpc-addr")]
+    grpc_bind_addr: Option<String>,
     #[clap(long)]
     mysql_addr: Option<String>,
     #[clap(long)]
@@ -283,7 +300,7 @@ impl StartCommand {
                 .to_string();
         }
 
-        if let Some(addr) = &self.rpc_bind_addr {
+        if let Some(addr) = &self.grpc_bind_addr {
             // frontend grpc addr conflict with datanode default grpc addr
             let datanode_grpc_addr = DatanodeOptions::default().grpc.bind_addr;
             if addr.eq(&datanode_grpc_addr) {
@@ -342,9 +359,19 @@ impl StartCommand {
         info!("Standalone start command: {:#?}", self);
         info!("Standalone options: {opts:#?}");
 
+        let (mut instance, _) =
+            Self::build_with(opts.component, opts.plugins, InstanceCreator::default()).await?;
+        instance._guard.extend(guard);
+        Ok(instance)
+    }
+
+    pub async fn build_with(
+        mut opts: StandaloneOptions,
+        plugin_opts: Vec<PluginOptions>,
+        creator: InstanceCreator,
+    ) -> Result<(Instance, InstanceCreatorResult)> {
         let mut plugins = Plugins::new();
-        let plugin_opts = opts.plugins;
-        let mut opts = opts.component;
+        plugins.insert(StandaloneFlag);
         set_default_prefix(opts.default_column_prefix.as_deref())
             .map_err(BoxedError::new)
             .context(error::BuildCliSnafu)?;
@@ -370,8 +397,10 @@ impl StartCommand {
             .context(error::CreateDirSnafu { dir: data_home })?;
 
         let metadata_dir = metadata_store_dir(data_home);
-        let kv_backend = standalone::build_metadata_kvbackend(metadata_dir, opts.metadata_store)
-            .context(error::BuildMetadataKvbackendSnafu)?;
+        let kv_backend = creator
+            .metadata_kv_backend_creator
+            .create(metadata_dir, &opts)
+            .await?;
         let procedure_manager =
             standalone::build_procedure_manager(kv_backend.clone(), opts.procedure);
 
@@ -392,6 +421,9 @@ impl StartCommand {
 
         let mut builder = DatanodeBuilder::new(dn_opts, plugins.clone(), kv_backend.clone());
         builder.with_cache_registry(layered_cache_registry.clone());
+        if let Some(writable) = creator.open_regions_writable_override {
+            builder.with_open_regions_writable_override(writable);
+        }
         let datanode = builder.build().await.context(error::StartDatanodeSnafu)?;
 
         let information_extension = Arc::new(StandaloneInformationExtension::new(
@@ -462,17 +494,16 @@ impl StartCommand {
                 .await;
         }
 
-        let node_manager = Arc::new(StandaloneDatanodeManager {
-            region_server: datanode.region_server(),
-            flow_server: flownode.flow_engine(),
-        });
+        let node_manager = creator
+            .node_manager_creator
+            .create(
+                &kv_backend,
+                datanode.region_server(),
+                flownode.flow_engine(),
+            )
+            .await?;
 
-        let table_id_allocator = Arc::new(
-            SequenceBuilder::new(TABLE_ID_SEQ, kv_backend.clone())
-                .initial(MIN_USER_TABLE_ID as u64)
-                .step(10)
-                .build(),
-        );
+        let table_id_allocator = creator.table_id_allocator_creator.create(&kv_backend);
         let flow_id_sequence = Arc::new(
             SequenceBuilder::new(FLOW_ID_SEQ, kv_backend.clone())
                 .initial(MIN_USER_FLOW_ID as u64)
@@ -489,7 +520,7 @@ impl StartCommand {
             .context(error::BuildWalProviderSnafu)?;
         let wal_provider = Arc::new(wal_provider);
         let table_metadata_allocator = Arc::new(TableMetadataAllocator::new(
-            table_id_allocator,
+            table_id_allocator.clone(),
             wal_provider.clone(),
         ));
         let flow_metadata_allocator = Arc::new(FlowMetadataAllocator::with_noop_peer_allocator(
@@ -532,10 +563,10 @@ impl StartCommand {
             ddl_manager
         };
 
-        let procedure_executor = Arc::new(LocalProcedureExecutor::new(
-            Arc::new(ddl_manager),
-            procedure_manager.clone(),
-        ));
+        let procedure_executor = creator
+            .procedure_executor_creator
+            .create(Arc::new(ddl_manager), procedure_manager.clone())
+            .await?;
 
         let fe_instance = FrontendBuilder::new(
             fe_opts.clone(),
@@ -568,7 +599,7 @@ impl StartCommand {
             kv_backend.clone(),
             layered_cache_registry.clone(),
             procedure_executor,
-            node_manager,
+            node_manager.clone(),
         )
         .await
         .context(StartFlownodeSnafu)?;
@@ -584,14 +615,21 @@ impl StartCommand {
             heartbeat_task: None,
         };
 
-        Ok(Instance {
+        let instance = Instance {
             datanode,
             frontend,
             flownode,
             procedure_manager,
             wal_provider,
-            _guard: guard,
-        })
+            leader_services_controller: creator.leader_services_controller,
+            _guard: vec![],
+        };
+        let result = InstanceCreatorResult {
+            kv_backend,
+            node_manager,
+            table_id_allocator,
+        };
+        Ok((instance, result))
     }
 
     pub async fn create_table_metadata_manager(
@@ -608,6 +646,258 @@ impl StartCommand {
     }
 }
 
+#[async_trait]
+pub trait NodeManagerCreator: Send + Sync {
+    async fn create(
+        &self,
+        kv_backend: &KvBackendRef,
+        region_server: RegionServer,
+        flow_server: FlownodeRef,
+    ) -> Result<NodeManagerRef>;
+}
+
+pub struct DefaultNodeManagerCreator;
+
+#[async_trait]
+impl NodeManagerCreator for DefaultNodeManagerCreator {
+    async fn create(
+        &self,
+        _: &KvBackendRef,
+        region_server: RegionServer,
+        flow_server: FlownodeRef,
+    ) -> Result<NodeManagerRef> {
+        Ok(Arc::new(StandaloneDatanodeManager {
+            region_server,
+            flow_server,
+        }))
+    }
+}
+
+/// Customizes how standalone opens its metadata KV backend.
+///
+/// The default implementation preserves the built-in raft-engine path. Other
+/// callers can provide a custom implementation without changing standalone
+/// configuration types.
+#[async_trait]
+pub trait MetadataKvBackendCreator: Send + Sync {
+    async fn create(&self, metadata_dir: String, opts: &StandaloneOptions) -> Result<KvBackendRef>;
+}
+
+pub struct DefaultMetadataKvBackendCreator;
+
+#[async_trait]
+impl MetadataKvBackendCreator for DefaultMetadataKvBackendCreator {
+    async fn create(&self, metadata_dir: String, opts: &StandaloneOptions) -> Result<KvBackendRef> {
+        standalone::build_metadata_kvbackend(metadata_dir, opts.metadata_store)
+            .context(error::BuildMetadataKvbackendSnafu)
+    }
+}
+
+pub trait TableIdAllocatorCreator: Send + Sync {
+    fn create(&self, kv_backend: &KvBackendRef) -> Arc<Sequence>;
+}
+
+struct DefaultTableIdAllocatorCreator;
+
+impl TableIdAllocatorCreator for DefaultTableIdAllocatorCreator {
+    fn create(&self, kv_backend: &KvBackendRef) -> Arc<Sequence> {
+        Arc::new(
+            SequenceBuilder::new(TABLE_ID_SEQ, kv_backend.clone())
+                .initial(MIN_USER_TABLE_ID as u64)
+                .step(10)
+                .build(),
+        )
+    }
+}
+
+#[async_trait]
+pub trait ProcedureExecutorCreator: Send + Sync {
+    async fn create(
+        &self,
+        ddl_manager: DdlManagerRef,
+        procedure_manager: ProcedureManagerRef,
+    ) -> Result<ProcedureExecutorRef>;
+}
+
+pub struct DefaultProcedureExecutorCreator;
+
+#[async_trait]
+impl ProcedureExecutorCreator for DefaultProcedureExecutorCreator {
+    async fn create(
+        &self,
+        ddl_manager: DdlManagerRef,
+        procedure_manager: ProcedureManagerRef,
+    ) -> Result<ProcedureExecutorRef> {
+        Ok(Arc::new(LocalProcedureExecutor::new(
+            ddl_manager,
+            procedure_manager,
+        )))
+    }
+}
+
+#[async_trait]
+pub trait StandaloneLeaderServicesController: Send + Sync {
+    /// Starts services that manage standalone metadata or WAL state.
+    ///
+    /// The default implementation starts the procedure manager and WAL provider
+    /// during instance startup.
+    async fn start(
+        &self,
+        procedure_manager: ProcedureManagerRef,
+        wal_provider: WalProviderRef,
+        region_server: RegionServer,
+    ) -> Result<()>;
+
+    /// Stops services started by [`StandaloneLeaderServicesController::start`].
+    async fn stop(
+        &self,
+        procedure_manager: ProcedureManagerRef,
+        region_server: RegionServer,
+    ) -> Result<()>;
+}
+
+pub struct DefaultStandaloneLeaderServicesController;
+
+#[async_trait]
+impl StandaloneLeaderServicesController for DefaultStandaloneLeaderServicesController {
+    async fn start(
+        &self,
+        procedure_manager: ProcedureManagerRef,
+        wal_provider: WalProviderRef,
+        _region_server: RegionServer,
+    ) -> Result<()> {
+        procedure_manager
+            .start()
+            .await
+            .context(error::StartProcedureManagerSnafu)?;
+        wal_provider
+            .start()
+            .await
+            .context(error::StartWalProviderSnafu)
+    }
+
+    async fn stop(
+        &self,
+        procedure_manager: ProcedureManagerRef,
+        _region_server: RegionServer,
+    ) -> Result<()> {
+        procedure_manager
+            .stop()
+            .await
+            .context(error::StopProcedureManagerSnafu)
+    }
+}
+
+/// `InstanceCreator` is used for grouping various component creators for building the
+/// Standalone instance, suitable for customizing how the instance can be built.
+pub struct InstanceCreator {
+    /// Hook for replacing metadata KV construction while reusing the rest of the
+    /// standalone build flow.
+    metadata_kv_backend_creator: Box<dyn MetadataKvBackendCreator>,
+    node_manager_creator: Box<dyn NodeManagerCreator>,
+    table_id_allocator_creator: Box<dyn TableIdAllocatorCreator>,
+    procedure_executor_creator: Box<dyn ProcedureExecutorCreator>,
+    leader_services_controller: Box<dyn StandaloneLeaderServicesController>,
+    open_regions_writable_override: Option<bool>,
+}
+
+impl InstanceCreator {
+    pub fn new(
+        node_manager_creator: Box<dyn NodeManagerCreator>,
+        table_id_allocator_creator: Box<dyn TableIdAllocatorCreator>,
+        procedure_executor_creator: Box<dyn ProcedureExecutorCreator>,
+    ) -> Self {
+        Self {
+            metadata_kv_backend_creator: Box::new(DefaultMetadataKvBackendCreator),
+            node_manager_creator,
+            table_id_allocator_creator,
+            procedure_executor_creator,
+            leader_services_controller: Box::new(DefaultStandaloneLeaderServicesController),
+            open_regions_writable_override: None,
+        }
+    }
+
+    pub fn with_metadata_kv_backend_creator(
+        mut self,
+        metadata_kv_backend_creator: Box<dyn MetadataKvBackendCreator>,
+    ) -> Self {
+        self.metadata_kv_backend_creator = metadata_kv_backend_creator;
+        self
+    }
+
+    /// Wraps the metadata backend creator while retaining the default creator.
+    ///
+    /// This is useful for callers that need to add runtime behavior around
+    /// metadata access without reimplementing backend selection.
+    pub fn map_metadata_kv_backend_creator<F>(mut self, f: F) -> Self
+    where
+        F: FnOnce(Box<dyn MetadataKvBackendCreator>) -> Box<dyn MetadataKvBackendCreator>,
+    {
+        self.metadata_kv_backend_creator = f(self.metadata_kv_backend_creator);
+        self
+    }
+
+    /// Wraps node-manager creation while preserving the selected standalone node manager.
+    pub fn map_node_manager_creator<F>(mut self, f: F) -> Self
+    where
+        F: FnOnce(Box<dyn NodeManagerCreator>) -> Box<dyn NodeManagerCreator>,
+    {
+        self.node_manager_creator = f(self.node_manager_creator);
+        self
+    }
+
+    /// Wraps procedure-executor creation while preserving the current setup.
+    pub fn map_procedure_executor_creator<F>(mut self, f: F) -> Self
+    where
+        F: FnOnce(Box<dyn ProcedureExecutorCreator>) -> Box<dyn ProcedureExecutorCreator>,
+    {
+        self.procedure_executor_creator = f(self.procedure_executor_creator);
+        self
+    }
+
+    /// Replaces startup/shutdown ownership for procedure manager and WAL provider.
+    pub fn with_leader_services_controller(
+        mut self,
+        leader_services_controller: Box<dyn StandaloneLeaderServicesController>,
+    ) -> Self {
+        self.leader_services_controller = leader_services_controller;
+        self
+    }
+
+    /// Overrides whether regions opened during startup should become writable.
+    ///
+    /// `None` keeps the default startup behavior (regions open writable).
+    ///
+    /// Warning: setting this to `false` in standalone mode will leave reopened regions
+    /// permanently read-only. Standalone has no metasrv heartbeat or region-role
+    /// reconciliation, so there is no path to promote regions to Leader after startup.
+    pub fn with_open_regions_writable_override(mut self, writable: bool) -> Self {
+        self.open_regions_writable_override = Some(writable);
+        self
+    }
+}
+
+impl Default for InstanceCreator {
+    fn default() -> Self {
+        Self {
+            metadata_kv_backend_creator: Box::new(DefaultMetadataKvBackendCreator),
+            node_manager_creator: Box::new(DefaultNodeManagerCreator),
+            table_id_allocator_creator: Box::new(DefaultTableIdAllocatorCreator),
+            procedure_executor_creator: Box::new(DefaultProcedureExecutorCreator),
+            leader_services_controller: Box::new(DefaultStandaloneLeaderServicesController),
+            open_regions_writable_override: None,
+        }
+    }
+}
+
+/// `InstanceCreatorResult` is expected to be used paired with [InstanceCreator].
+/// It stores the created and other important components for further reusing.
+pub struct InstanceCreatorResult {
+    pub kv_backend: KvBackendRef,
+    pub node_manager: NodeManagerRef,
+    pub table_id_allocator: Arc<Sequence>,
+}
+
 #[cfg(test)]
 mod tests {
     use std::default::Default;
@@ -615,8 +905,10 @@ mod tests {
     use std::time::Duration;
 
     use auth::{Identity, Password, UserProviderRef};
+    use clap::{CommandFactory, Parser};
     use common_base::readable_size::ReadableSize;
     use common_config::ENV_VAR_SEP;
+    use common_options::plugin_options::StandaloneFlag;
     use common_test_util::temp_dir::create_named_temp_file;
     use common_wal::config::DatanodeWalConfig;
     use frontend::frontend::FrontendOptions;
@@ -634,6 +926,7 @@ mod tests {
         };
 
         let mut plugins = Plugins::new();
+        plugins.insert(StandaloneFlag);
         plugins::setup_frontend_plugins(&mut plugins, &[], &fe_opts)
             .await
             .unwrap();
@@ -860,6 +1153,35 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_parse_grpc_bind_addr_aliases() {
+        let command =
+            StartCommand::try_parse_from(["standalone", "--grpc-bind-addr", "127.0.0.1:14001"])
+                .unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:14001"));
+
+        let command =
+            StartCommand::try_parse_from(["standalone", "--rpc-bind-addr", "127.0.0.1:24001"])
+                .unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:24001"));
+
+        let command =
+            StartCommand::try_parse_from(["standalone", "--rpc-addr", "127.0.0.1:34001"]).unwrap();
+        assert_eq!(command.grpc_bind_addr.as_deref(), Some("127.0.0.1:34001"));
+    }
+
+    #[test]
+    fn test_help_uses_grpc_option_names() {
+        let mut cmd = StartCommand::command();
+        let mut help = Vec::new();
+        cmd.write_long_help(&mut help).unwrap();
+        let help = String::from_utf8(help).unwrap();
+
+        assert!(help.contains("--grpc-bind-addr"));
+        assert!(!help.contains("--rpc-bind-addr"));
+        assert!(!help.contains("--rpc-addr"));
+    }
+
     #[test]
     fn test_load_default_standalone_options() {
         let options =
diff --git a/src/cmd/tests/load_config_test.rs b/src/cmd/tests/load_config_test.rs
index 2300a2250e..6cffcd67c2 100644
--- a/src/cmd/tests/load_config_test.rs
+++ b/src/cmd/tests/load_config_test.rs
@@ -16,7 +16,7 @@ use std::time::Duration;
 
 use cmd::options::GreptimeOptions;
 use common_base::memory_limit::MemoryLimit;
-use common_config::{Configurable, DEFAULT_DATA_HOME};
+use common_config::{Configurable, DEFAULT_DATA_HOME, ENV_VAR_SEP};
 use common_options::datanode::{ClientOptions, DatanodeClientOptions};
 use common_telemetry::logging::{DEFAULT_LOGGING_DIR, DEFAULT_OTLP_HTTP_ENDPOINT, LoggingOptions};
 use common_wal::config::DatanodeWalConfig;
@@ -311,3 +311,51 @@ fn test_load_standalone_example_config() {
     };
     similar_asserts::assert_eq!(options, expected);
 }
+
+#[test]
+fn test_load_standalone_user_provider_from_config() {
+    let config = tempfile::NamedTempFile::new().unwrap();
+    let user_provider = "static_user_provider:file:/tmp/greptimedb-users";
+    std::fs::write(
+        config.path(),
+        format!("user_provider = \"{user_provider}\"\n"),
+    )
+    .unwrap();
+
+    let options =
+        GreptimeOptions::<StandaloneOptions>::load_layered_options(config.path().to_str(), "")
+            .unwrap();
+
+    assert_eq!(
+        options.component.user_provider.as_deref(),
+        Some(user_provider)
+    );
+
+    let frontend_options = options.component.frontend_options();
+    assert_eq!(
+        frontend_options.user_provider.as_deref(),
+        Some(user_provider)
+    );
+}
+
+#[test]
+fn test_load_heartbeat_env_vars_from_env() {
+    let env_prefix = "HEARTBEAT_ENV_VARS_UT";
+    let env_key = [env_prefix, "HEARTBEAT_ENV_VARS"].join(ENV_VAR_SEP);
+
+    temp_env::with_var(env_key, Some("AZ,REGION"), || {
+        let expected = vec!["AZ".to_string(), "REGION".to_string()];
+
+        let datanode =
+            GreptimeOptions::<DatanodeOptions>::load_layered_options(None, env_prefix).unwrap();
+        similar_asserts::assert_eq!(datanode.component.heartbeat_env_vars, expected);
+
+        let frontend =
+            GreptimeOptions::<FrontendOptions>::load_layered_options(None, env_prefix).unwrap();
+        similar_asserts::assert_eq!(frontend.component.heartbeat_env_vars, expected);
+
+        let standalone =
+            GreptimeOptions::<StandaloneOptions>::load_layered_options(None, env_prefix).unwrap();
+        similar_asserts::assert_eq!(standalone.component.heartbeat_env_vars, expected);
+    });
+}
diff --git a/src/common/base/Cargo.toml b/src/common/base/Cargo.toml
index 3ec9e1fa35..44c30cd548 100644
--- a/src/common/base/Cargo.toml
+++ b/src/common/base/Cargo.toml
@@ -16,16 +16,11 @@ anymap2 = "0.13"
 async-trait.workspace = true
 bitvec = "1.0"
 bytes.workspace = true
-common-error.workspace = true
-common-macro.workspace = true
 futures.workspace = true
 lazy_static.workspace = true
-paste.workspace = true
 pin-project.workspace = true
-rand.workspace = true
 regex.workspace = true
 serde = { version = "1.0", features = ["derive"] }
-snafu.workspace = true
 tokio.workspace = true
 zeroize = { version = "1.6", default-features = false, features = ["alloc"] }
 
diff --git a/src/common/catalog/src/lib.rs b/src/common/catalog/src/lib.rs
index 1e9534532e..592e0df8fe 100644
--- a/src/common/catalog/src/lib.rs
+++ b/src/common/catalog/src/lib.rs
@@ -74,9 +74,9 @@ pub fn parse_catalog_and_schema_from_db_string(db: &str) -> (String, String) {
 pub fn parse_optional_catalog_and_schema_from_db_string(db: &str) -> (Option<String>, String) {
     let parts = db.splitn(2, '-').collect::<Vec<&str>>();
     if parts.len() == 2 {
-        (Some(parts[0].to_lowercase()), parts[1].to_lowercase())
+        (Some(parts[0].to_string()), parts[1].to_string())
     } else {
-        (None, db.to_lowercase())
+        (None, db.to_string())
     }
 }
 
@@ -118,7 +118,7 @@ mod tests {
         );
 
         assert_eq!(
-            (Some("catalog".to_string()), "schema".to_string()),
+            (Some("CATALOG".to_string()), "SCHEMA".to_string()),
             parse_optional_catalog_and_schema_from_db_string("CATALOG-SCHEMA")
         );
 
diff --git a/src/common/config/Cargo.toml b/src/common/config/Cargo.toml
index 2737f82a58..27b238add7 100644
--- a/src/common/config/Cargo.toml
+++ b/src/common/config/Cargo.toml
@@ -18,7 +18,6 @@ notify.workspace = true
 object-store.workspace = true
 serde.workspace = true
 serde_json.workspace = true
-serde_with.workspace = true
 snafu.workspace = true
 toml.workspace = true
 
diff --git a/src/common/config/src/config.rs b/src/common/config/src/config.rs
index e25c46a0c0..85ce3d206f 100644
--- a/src/common/config/src/config.rs
+++ b/src/common/config/src/config.rs
@@ -53,7 +53,7 @@ pub trait Configurable: Serialize + DeserializeOwned + Default + Sized {
 
             env.try_parsing(true)
                 .separator(ENV_VAR_SEP)
-                .ignore_empty(true)
+                .ignore_empty(false)
         };
 
         // Workaround: Replacement for `Config::try_from(&default_opts)` due to
@@ -237,4 +237,31 @@ mod tests {
             },
         );
     }
+
+    #[derive(Debug, Serialize, Deserialize, Default)]
+    struct SimpleConfig {
+        name: Option<String>,
+        prefix: Option<String>,
+    }
+
+    impl Configurable for SimpleConfig {}
+
+    #[test]
+    fn test_empty_env_var_is_not_ignored() {
+        let env_prefix = "SIMPLE_CFG_UT";
+        temp_env::with_vars(
+            [(
+                [env_prefix.to_string(), "PREFIX".to_string()].join(ENV_VAR_SEP),
+                Some(""),
+            )],
+            || {
+                let opts = SimpleConfig::load_layered_options(None, env_prefix).unwrap();
+                // With ignore_empty(false), an empty env var should yield Some("")
+                // rather than None (which was the previous behavior with ignore_empty(true)).
+                assert_eq!(opts.prefix, Some("".to_string()));
+                // Unset env var should remain None.
+                assert_eq!(opts.name, None);
+            },
+        );
+    }
 }
diff --git a/src/common/datasource/Cargo.toml b/src/common/datasource/Cargo.toml
index 6ec9a14733..7df293d6d3 100644
--- a/src/common/datasource/Cargo.toml
+++ b/src/common/datasource/Cargo.toml
@@ -28,12 +28,10 @@ common-runtime.workspace = true
 common-telemetry.workspace = true
 datafusion.workspace = true
 datafusion-datasource.workspace = true
-datafusion-orc.workspace = true
 datatypes.workspace = true
 futures.workspace = true
 lazy_static.workspace = true
 object-store.workspace = true
-object_store_opendal.workspace = true
 orc-rust = { version = "0.7", default-features = false, features = ["async"] }
 parquet.workspace = true
 paste.workspace = true
@@ -47,3 +45,4 @@ url.workspace = true
 
 [dev-dependencies]
 common-test-util.workspace = true
+datafusion-orc.workspace = true
diff --git a/src/common/datasource/src/file_format.rs b/src/common/datasource/src/file_format.rs
index 7f4a7c65b4..a6a358c9e4 100644
--- a/src/common/datasource/src/file_format.rs
+++ b/src/common/datasource/src/file_format.rs
@@ -42,7 +42,6 @@ use datafusion::physical_plan::SendableRecordBatchStream;
 use datafusion::physical_plan::metrics::ExecutionPlanMetricsSet;
 use futures::{StreamExt, TryStreamExt};
 use object_store::ObjectStore;
-use object_store_opendal::OpendalStore;
 use snafu::ResultExt;
 use tokio::io::AsyncWriteExt;
 use tokio_util::compat::FuturesAsyncWriteCompatExt;
@@ -317,7 +316,7 @@ pub async fn file_to_stream(
             .with_file_compression_type(df_compression)
             .build();
 
-    let store = Arc::new(OpendalStore::new(store.clone()));
+    let store = Arc::new(object_store::compat::OpendalStore::new(store.clone()));
     let file_opener = config.file_source().create_file_opener(store, &config, 0)?;
     let stream = FileStream::new(&config, 0, file_opener, &ExecutionPlanMetricsSet::new())?;
 
diff --git a/src/common/datasource/src/file_format/parquet.rs b/src/common/datasource/src/file_format/parquet.rs
index c2c14b4680..9c8e8d6ce8 100644
--- a/src/common/datasource/src/file_format/parquet.rs
+++ b/src/common/datasource/src/file_format/parquet.rs
@@ -23,7 +23,9 @@ use datafusion::error::Result as DatafusionResult;
 use datafusion::parquet::arrow::async_reader::AsyncFileReader;
 use datafusion::parquet::arrow::{ArrowWriter, parquet_to_arrow_schema};
 use datafusion::parquet::errors::{ParquetError, Result as ParquetResult};
-use datafusion::parquet::file::metadata::ParquetMetaData;
+use datafusion::parquet::file::metadata::{
+    PageIndexPolicy, ParquetMetaData, ParquetMetaDataReader,
+};
 use datafusion::physical_plan::SendableRecordBatchStream;
 use datafusion::physical_plan::metrics::ExecutionPlanMetricsSet;
 use datafusion_datasource::PartitionedFile;
@@ -94,35 +96,40 @@ impl DefaultParquetFileReaderFactory {
 }
 
 impl ParquetFileReaderFactory for DefaultParquetFileReaderFactory {
-    // TODO(weny): Supports [`metadata_size_hint`].
-    // The upstream has a implementation supports [`metadata_size_hint`],
-    // however it coupled with Box<dyn ObjectStore>.
     fn create_reader(
         &self,
         _partition_index: usize,
         partitioned_file: PartitionedFile,
-        _metadata_size_hint: Option<usize>,
+        metadata_size_hint: Option<usize>,
         _metrics: &ExecutionPlanMetricsSet,
     ) -> DatafusionResult<Box<dyn AsyncFileReader + Send>> {
         let path = partitioned_file.path().to_string();
         let object_store = self.object_store.clone();
 
-        Ok(Box::new(LazyParquetFileReader::new(object_store, path)))
+        Ok(Box::new(LazyParquetFileReader::new(
+            object_store,
+            path,
+            metadata_size_hint,
+        )))
     }
 }
 
 pub struct LazyParquetFileReader {
     object_store: ObjectStore,
     reader: Option<Compat<FuturesAsyncReader>>,
+    file_size: Option<u64>,
+    metadata_size_hint: Option<usize>,
     path: String,
 }
 
 impl LazyParquetFileReader {
-    pub fn new(object_store: ObjectStore, path: String) -> Self {
+    pub fn new(object_store: ObjectStore, path: String, metadata_size_hint: Option<usize>) -> Self {
         LazyParquetFileReader {
             object_store,
             path,
             reader: None,
+            file_size: None,
+            metadata_size_hint,
         }
     }
 
@@ -130,6 +137,7 @@ impl LazyParquetFileReader {
     async fn maybe_initialize(&mut self) -> result::Result<(), object_store::Error> {
         if self.reader.is_none() {
             let meta = self.object_store.stat(&self.path).await?;
+            self.file_size = Some(meta.content_length());
             let reader = self
                 .object_store
                 .reader(&self.path)
@@ -166,8 +174,19 @@ impl AsyncFileReader for LazyParquetFileReader {
             self.maybe_initialize()
                 .await
                 .map_err(|e| ParquetError::External(Box::new(e)))?;
-            // Safety: Must initialized
-            self.reader.as_mut().unwrap().get_metadata(options).await
+
+            let metadata_opts = options.map(|o| o.metadata_options().clone());
+            let metadata_reader = ParquetMetaDataReader::new()
+                .with_metadata_options(metadata_opts)
+                .with_page_index_policy(PageIndexPolicy::from(
+                    options.is_some_and(|o| o.page_index()),
+                ))
+                .with_prefetch_hint(self.metadata_size_hint);
+
+            let metadata = metadata_reader
+                .load_and_finish(self.reader.as_mut().unwrap(), self.file_size.unwrap())
+                .await?;
+            Ok(Arc::new(metadata))
         })
     }
 }
diff --git a/src/common/datasource/src/file_format/tests.rs b/src/common/datasource/src/file_format/tests.rs
index ad54472d33..6ef669ab7b 100644
--- a/src/common/datasource/src/file_format/tests.rs
+++ b/src/common/datasource/src/file_format/tests.rs
@@ -12,10 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
 use std::collections::HashMap;
 use std::sync::Arc;
-use std::vec;
+use std::{assert_matches, vec};
 
 use common_test_util::find_workspace_path;
 use datafusion::assert_batches_eq;
@@ -45,7 +44,7 @@ struct Test<'a> {
 
 impl Test<'_> {
     async fn run(self, store: &ObjectStore) {
-        let store = Arc::new(object_store_opendal::OpendalStore::new(store.clone()));
+        let store = Arc::new(object_store::compat::OpendalStore::new(store.clone()));
         let file_opener = self
             .file_source
             .create_file_opener(store, &self.config, 0)
diff --git a/src/common/datasource/src/lib.rs b/src/common/datasource/src/lib.rs
index 91663ce22c..f4c7fdf120 100644
--- a/src/common/datasource/src/lib.rs
+++ b/src/common/datasource/src/lib.rs
@@ -12,9 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
-#![feature(type_alias_impl_trait)]
-
 pub mod buffered_writer;
 pub mod compressed_writer;
 pub mod compression;
diff --git a/src/common/datasource/src/test_util.rs b/src/common/datasource/src/test_util.rs
index ea2b0c768c..0a13d9c6e8 100644
--- a/src/common/datasource/src/test_util.rs
+++ b/src/common/datasource/src/test_util.rs
@@ -103,7 +103,7 @@ pub async fn setup_stream_to_json_test(origin_path: &str, threshold: impl Fn(usi
         test_util::TEST_BATCH_SIZE,
         schema.clone(),
         FileCompressionType::UNCOMPRESSED,
-        Arc::new(object_store_opendal::OpendalStore::new(store.clone())),
+        Arc::new(object_store::compat::OpendalStore::new(store.clone())),
         true,
     );
 
@@ -157,7 +157,7 @@ pub async fn setup_stream_to_csv_test(
 
     let csv_opener = csv_source
         .create_file_opener(
-            Arc::new(object_store_opendal::OpendalStore::new(store.clone())),
+            Arc::new(object_store::compat::OpendalStore::new(store.clone())),
             &config,
             0,
         )
diff --git a/src/common/frontend/src/error.rs b/src/common/frontend/src/error.rs
index cee8c6df77..19fb29aeb7 100644
--- a/src/common/frontend/src/error.rs
+++ b/src/common/frontend/src/error.rs
@@ -30,7 +30,7 @@ pub enum Error {
 
     #[snafu(display("Failed to list nodes from metasrv"))]
     Meta {
-        source: Box<meta_client::error::Error>,
+        source: Box<common_meta::error::Error>,
         #[snafu(implicit)]
         location: Location,
     },
@@ -52,7 +52,7 @@ pub enum Error {
 
     #[snafu(display("Failed to invoke list process service"))]
     CreateChannel {
-        source: common_grpc::error::Error,
+        source: Box<common_grpc::error::Error>,
         #[snafu(implicit)]
         location: Location,
     },
diff --git a/src/common/frontend/src/selector.rs b/src/common/frontend/src/selector.rs
index 804169d1dd..bba36ef81b 100644
--- a/src/common/frontend/src/selector.rs
+++ b/src/common/frontend/src/selector.rs
@@ -16,7 +16,7 @@ use std::fmt::Debug;
 use std::time::Duration;
 
 use common_grpc::channel_manager::{ChannelConfig, ChannelManager};
-use common_meta::cluster::{ClusterInfo, NodeInfo, Role};
+use common_meta::peer::{Peer, PeerDiscovery};
 use greptime_proto::v1::frontend::{
     KillProcessRequest, KillProcessResponse, ListProcessRequest, ListProcessResponse,
     frontend_client,
@@ -62,7 +62,7 @@ impl FrontendClient for frontend_client::FrontendClient<tonic::transport::channe
 pub trait FrontendSelector {
     async fn select<F>(&self, predicate: F) -> Result<Vec<FrontendClientPtr>>
     where
-        F: Fn(&NodeInfo) -> bool + Send;
+        F: Fn(&Peer) -> bool + Send;
 }
 
 #[derive(Debug, Clone)]
@@ -75,22 +75,23 @@ pub struct MetaClientSelector {
 impl FrontendSelector for MetaClientSelector {
     async fn select<F>(&self, predicate: F) -> Result<Vec<FrontendClientPtr>>
     where
-        F: Fn(&NodeInfo) -> bool + Send,
+        F: Fn(&Peer) -> bool + Send,
     {
-        let nodes = self
+        let peers = self
             .meta_client
-            .list_nodes(Some(Role::Frontend))
+            .active_frontends()
             .await
             .map_err(Box::new)
             .context(MetaSnafu)?;
 
-        nodes
+        peers
             .into_iter()
             .filter(predicate)
-            .map(|node| {
+            .map(|peer| {
                 let channel = self
                     .channel_manager
-                    .get(node.peer.addr)
+                    .get(peer.addr)
+                    .map_err(Box::new)
                     .context(error::CreateChannelSnafu)?;
                 let client = frontend_client::FrontendClient::new(channel);
                 Ok(Box::new(client) as FrontendClientPtr)
diff --git a/src/common/function/Cargo.toml b/src/common/function/Cargo.toml
index d164b9285d..2829f92985 100644
--- a/src/common/function/Cargo.toml
+++ b/src/common/function/Cargo.toml
@@ -47,8 +47,8 @@ geo-types = { version = "0.7", optional = true }
 geohash = { version = "0.13", optional = true }
 h3o = { version = "0.6", optional = true }
 hyperloglogplus = "0.4"
+icu_properties.workspace = true
 jsonb.workspace = true
-jsonpath-rust = "0.7.5"
 memchr = "2.7"
 mito-codec.workspace = true
 nalgebra.workspace = true
diff --git a/src/common/function/src/admin/flush_compact_region.rs b/src/common/function/src/admin/flush_compact_region.rs
index 60fd19ef5a..6dcf92117a 100644
--- a/src/common/function/src/admin/flush_compact_region.rs
+++ b/src/common/function/src/admin/flush_compact_region.rs
@@ -128,7 +128,7 @@ mod tests {
                     };
                     let result = f.invoke_async_with_args(func_args).await.unwrap_err();
                     assert_eq!(
-                        "Execution error: Handler error: Missing TableMutationHandler, not expected",
+                        "Execution error: Missing TableMutationHandler, not expected",
                         result.to_string()
                     );
                 }
diff --git a/src/common/function/src/admin/flush_compact_table.rs b/src/common/function/src/admin/flush_compact_table.rs
index 3298a95061..20b95e4379 100644
--- a/src/common/function/src/admin/flush_compact_table.rs
+++ b/src/common/function/src/admin/flush_compact_table.rs
@@ -355,7 +355,7 @@ mod tests {
                     };
                     let result = f.invoke_async_with_args(func_args).await.unwrap_err();
                     assert_eq!(
-                        "Execution error: Handler error: Missing TableMutationHandler, not expected",
+                        "Execution error: Missing TableMutationHandler, not expected",
                         result.to_string()
                     );
                 }
diff --git a/src/common/function/src/admin/migrate_region.rs b/src/common/function/src/admin/migrate_region.rs
index 91b5540b1a..d958b3cce9 100644
--- a/src/common/function/src/admin/migrate_region.rs
+++ b/src/common/function/src/admin/migrate_region.rs
@@ -173,7 +173,7 @@ mod tests {
         };
         let result = f.invoke_async_with_args(func_args).await.unwrap_err();
         assert_eq!(
-            "Execution error: Handler error: Missing ProcedureServiceHandler, not expected",
+            "Execution error: Missing ProcedureServiceHandler, not expected",
             result.to_string()
         );
     }
diff --git a/src/common/function/src/aggrs/aggr_wrapper.rs b/src/common/function/src/aggrs/aggr_wrapper.rs
index 3780d39582..6242ab9454 100644
--- a/src/common/function/src/aggrs/aggr_wrapper.rs
+++ b/src/common/function/src/aggrs/aggr_wrapper.rs
@@ -25,7 +25,7 @@
 use std::hash::{Hash, Hasher};
 use std::sync::Arc;
 
-use arrow::array::StructArray;
+use arrow::array::{ArrayRef, BooleanArray, StructArray};
 use arrow_schema::{FieldRef, Fields};
 use common_telemetry::debug;
 use datafusion::functions_aggregate::all_default_aggregate_functions;
@@ -38,8 +38,8 @@ use datafusion_common::{Column, ScalarValue};
 use datafusion_expr::expr::{AggregateFunction, AggregateFunctionParams};
 use datafusion_expr::function::StateFieldsArgs;
 use datafusion_expr::{
-    Accumulator, Aggregate, AggregateUDF, AggregateUDFImpl, Expr, ExprSchemable, LogicalPlan,
-    Signature,
+    Accumulator, Aggregate, AggregateUDF, AggregateUDFImpl, EmitTo, Expr, ExprSchemable,
+    GroupsAccumulator, LogicalPlan, Signature,
 };
 use datafusion_physical_expr::aggregate::AggregateFunctionExpr;
 use datatypes::arrow::datatypes::{DataType, Field};
@@ -322,6 +322,14 @@ impl StateWrapper {
             );
         })
     }
+
+    fn fix_inner_acc_args<'b>(
+        &self,
+        mut acc_args: datafusion_expr::function::AccumulatorArgs<'b>,
+    ) -> datafusion_common::Result<datafusion_expr::function::AccumulatorArgs<'b>> {
+        acc_args.return_field = self.deduce_aggr_return_type(&acc_args)?;
+        Ok(acc_args)
+    }
 }
 
 impl AggregateUDFImpl for StateWrapper {
@@ -331,15 +339,32 @@ impl AggregateUDFImpl for StateWrapper {
     ) -> datafusion_common::Result<Box<dyn Accumulator>> {
         // fix and recover proper acc args for the original aggregate function.
         let state_type = acc_args.return_type().clone();
-        let inner = {
-            let mut new_acc_args = acc_args.clone();
-            new_acc_args.return_field = self.deduce_aggr_return_type(&acc_args)?;
-            self.inner.accumulator(new_acc_args)?
-        };
+        let inner = self.inner.accumulator(self.fix_inner_acc_args(acc_args)?)?;
 
         Ok(Box::new(StateAccum::new(inner, state_type)?))
     }
 
+    fn groups_accumulator_supported(
+        &self,
+        acc_args: datafusion_expr::function::AccumulatorArgs,
+    ) -> bool {
+        self.fix_inner_acc_args(acc_args)
+            .map(|args| self.inner.inner().groups_accumulator_supported(args))
+            .unwrap_or(false)
+    }
+
+    fn create_groups_accumulator(
+        &self,
+        acc_args: datafusion_expr::function::AccumulatorArgs,
+    ) -> datafusion_common::Result<Box<dyn GroupsAccumulator>> {
+        let state_type = acc_args.return_type().clone();
+        let inner = self
+            .inner
+            .inner()
+            .create_groups_accumulator(self.fix_inner_acc_args(acc_args)?)?;
+        Ok(Box::new(StateGroupsAccum::new(inner, state_type)?))
+    }
+
     fn as_any(&self) -> &dyn std::any::Any {
         self
     }
@@ -462,6 +487,118 @@ pub struct StateAccum {
     state_fields: Fields,
 }
 
+pub struct StateGroupsAccum {
+    inner: Box<dyn GroupsAccumulator>,
+    state_fields: Fields,
+}
+
+impl StateGroupsAccum {
+    fn new(
+        inner: Box<dyn GroupsAccumulator>,
+        state_type: DataType,
+    ) -> datafusion_common::Result<Self> {
+        let DataType::Struct(fields) = state_type else {
+            return Err(datafusion_common::DataFusionError::Internal(format!(
+                "Expected a struct type for state, got: {:?}",
+                state_type
+            )));
+        };
+        Ok(Self {
+            inner,
+            state_fields: fields,
+        })
+    }
+
+    fn wrap_state_arrays(&self, arrays: Vec<ArrayRef>) -> datafusion_common::Result<ArrayRef> {
+        let array_type = arrays
+            .iter()
+            .map(|array| array.data_type().clone())
+            .collect::<Vec<_>>();
+        let expected_type = self
+            .state_fields
+            .iter()
+            .map(|field| field.data_type().clone())
+            .collect::<Vec<_>>();
+        if array_type != expected_type {
+            debug!(
+                "State mismatch, expected: {}, got: {} for expected fields: {:?} and given array types: {:?}",
+                self.state_fields.len(),
+                arrays.len(),
+                self.state_fields,
+                array_type,
+            );
+            let guess_schema = arrays
+                .iter()
+                .enumerate()
+                .map(|(index, array)| {
+                    Field::new(
+                        format!("col_{index}[mismatch_state]").as_str(),
+                        array.data_type().clone(),
+                        true,
+                    )
+                })
+                .collect::<Fields>();
+            let array = StructArray::try_new(guess_schema, arrays, None)?;
+            return Ok(Arc::new(array));
+        }
+
+        Ok(Arc::new(StructArray::try_new(
+            self.state_fields.clone(),
+            arrays,
+            None,
+        )?))
+    }
+}
+
+impl GroupsAccumulator for StateGroupsAccum {
+    fn update_batch(
+        &mut self,
+        values: &[ArrayRef],
+        group_indices: &[usize],
+        opt_filter: Option<&BooleanArray>,
+        total_num_groups: usize,
+    ) -> datafusion_common::Result<()> {
+        self.inner
+            .update_batch(values, group_indices, opt_filter, total_num_groups)
+    }
+
+    fn merge_batch(
+        &mut self,
+        values: &[ArrayRef],
+        group_indices: &[usize],
+        opt_filter: Option<&BooleanArray>,
+        total_num_groups: usize,
+    ) -> datafusion_common::Result<()> {
+        self.inner
+            .merge_batch(values, group_indices, opt_filter, total_num_groups)
+    }
+
+    fn evaluate(&mut self, emit_to: EmitTo) -> datafusion_common::Result<ArrayRef> {
+        let state = self.inner.state(emit_to)?;
+        self.wrap_state_arrays(state)
+    }
+
+    fn state(&mut self, emit_to: EmitTo) -> datafusion_common::Result<Vec<ArrayRef>> {
+        self.inner.state(emit_to)
+    }
+
+    fn convert_to_state(
+        &self,
+        values: &[ArrayRef],
+        opt_filter: Option<&BooleanArray>,
+    ) -> datafusion_common::Result<Vec<ArrayRef>> {
+        self.inner.convert_to_state(values, opt_filter)
+    }
+
+    fn supports_convert_to_state(&self) -> bool {
+        self.inner.supports_convert_to_state()
+    }
+
+    fn size(&self) -> usize {
+        self.inner.size()
+    }
+}
+
 impl StateAccum {
     pub fn new(
         inner: Box<dyn Accumulator>,
diff --git a/src/common/function/src/aggrs/aggr_wrapper/tests.rs b/src/common/function/src/aggrs/aggr_wrapper/tests.rs
index 8821b9fd24..de3a77df6b 100644
--- a/src/common/function/src/aggrs/aggr_wrapper/tests.rs
+++ b/src/common/function/src/aggrs/aggr_wrapper/tests.rs
@@ -40,10 +40,13 @@ use datafusion_common::arrow::array::AsArray;
 use datafusion_common::arrow::datatypes::{Float64Type, UInt64Type};
 use datafusion_common::{Column, TableReference};
 use datafusion_expr::expr::{AggregateFunction, NullTreatment};
+use datafusion_expr::function::AccumulatorArgs;
 use datafusion_expr::{
-    Aggregate, ColumnarValue, Expr, LogicalPlan, ScalarFunctionArgs, SortExpr, TableScan, lit,
+    Aggregate, AggregateUDFImpl, ColumnarValue, Expr, LogicalPlan, ScalarFunctionArgs, SortExpr,
+    TableScan, lit,
 };
 use datafusion_physical_expr::aggregate::AggregateExprBuilder;
+use datafusion_physical_expr::expressions::col;
 use datafusion_physical_expr::{EquivalenceProperties, Partitioning};
 use datatypes::arrow_array::StringArray;
 use futures::{Stream, StreamExt as _};
@@ -256,6 +259,38 @@ fn dummy_table_scan_with_ts() -> LogicalPlan {
     )
 }
 
+fn create_avg_state_groups_accumulator() -> Box<dyn GroupsAccumulator> {
+    let state_wrapper = StateWrapper::new((*avg_udaf()).clone()).unwrap();
+    let schema = Arc::new(arrow_schema::Schema::new(vec![Field::new(
+        "number",
+        DataType::Float64,
+        true,
+    )]));
+    let expr = col("number", &schema).unwrap();
+    let expr_field = expr.return_field(&schema).unwrap();
+    let return_field = Arc::new(Field::new(
+        "__avg_state(number)",
+        state_wrapper.return_type(&[DataType::Float64]).unwrap(),
+        true,
+    ));
+    let exprs = [expr];
+    let expr_fields = [expr_field];
+    let acc_args = AccumulatorArgs {
+        return_field,
+        schema: &schema,
+        ignore_nulls: false,
+        order_bys: &[],
+        is_reversed: false,
+        name: "__avg_state(number)",
+        is_distinct: false,
+        exprs: &exprs,
+        expr_fields: &expr_fields,
+    };
+
+    assert!(state_wrapper.groups_accumulator_supported(acc_args.clone()));
+    state_wrapper.create_groups_accumulator(acc_args).unwrap()
+}
+
 #[tokio::test]
 async fn test_sum_udaf() {
     let ctx = SessionContext::new();
@@ -796,6 +831,95 @@ async fn test_last_value_order_by_udaf() {
     assert_eq!(merge_eval_res, ScalarValue::Int64(Some(4)));
 }
 
+#[test]
+fn test_avg_state_groups_accumulator_evaluate() {
+    let mut state_accum = create_avg_state_groups_accumulator();
+    let values = vec![Arc::new(Float64Array::from(vec![
+        Some(1.0),
+        Some(2.0),
+        None,
+        Some(3.0),
+        Some(4.0),
+        Some(5.0),
+    ])) as ArrayRef];
+    let group_indices = vec![0, 1, 0, 0, 1, 2];
+
+    state_accum
+        .update_batch(&values, &group_indices, None, 3)
+        .unwrap();
+
+    let result = state_accum.evaluate(EmitTo::All).unwrap();
+    let result = result.as_any().downcast_ref::<StructArray>().unwrap();
+
+    assert_eq!(
+        result
+            .column(0)
+            .as_any()
+            .downcast_ref::<UInt64Array>()
+            .unwrap(),
+        &UInt64Array::from(vec![2, 2, 1])
+    );
+    assert_eq!(
+        result
+            .column(1)
+            .as_any()
+            .downcast_ref::<Float64Array>()
+            .unwrap(),
+        &Float64Array::from(vec![4.0, 6.0, 5.0])
+    );
+}
+
+#[test]
+fn test_avg_state_groups_accumulator_state_merge_evaluate() {
+    let mut source_accum = create_avg_state_groups_accumulator();
+    let source_values = vec![Arc::new(Float64Array::from(vec![
+        Some(1.0),
+        Some(2.0),
+        None,
+        Some(3.0),
+        Some(4.0),
+        Some(5.0),
+    ])) as ArrayRef];
+    let source_group_indices = vec![0, 1, 0, 0, 1, 2];
+
+    source_accum
+        .update_batch(&source_values, &source_group_indices, None, 3)
+        .unwrap();
+    let source_state = source_accum.state(EmitTo::All).unwrap();
+
+    let mut merged_accum = create_avg_state_groups_accumulator();
+    let merged_values =
+        vec![Arc::new(Float64Array::from(vec![Some(10.0), Some(20.0), Some(30.0)])) as ArrayRef];
+    let merged_group_indices = vec![0, 1, 2];
+
+    merged_accum
+        .update_batch(&merged_values, &merged_group_indices, None, 3)
+        .unwrap();
+    merged_accum
+        .merge_batch(&source_state, &[1, 2, 0], None, 3)
+        .unwrap();
+
+    let result = merged_accum.evaluate(EmitTo::All).unwrap();
+    let result = result.as_any().downcast_ref::<StructArray>().unwrap();
+
+    assert_eq!(
+        result
+            .column(0)
+            .as_any()
+            .downcast_ref::<UInt64Array>()
+            .unwrap(),
+        &UInt64Array::from(vec![2, 3, 3])
+    );
+    assert_eq!(
+        result
+            .column(1)
+            .as_any()
+            .downcast_ref::<Float64Array>()
+            .unwrap(),
+        &Float64Array::from(vec![15.0, 24.0, 36.0])
+    );
+}
+
 /// For testing whether the UDAF state fields are correctly implemented.
 /// esp. for our own custom UDAF's state fields.
 /// By compare eval results before and after split to state/merge functions.
diff --git a/src/common/function/src/flush_flow.rs b/src/common/function/src/flush_flow.rs
index c4ea554585..35f347bcb2 100644
--- a/src/common/function/src/flush_flow.rs
+++ b/src/common/function/src/flush_flow.rs
@@ -149,7 +149,7 @@ mod test {
 
         let result = f.invoke_async_with_args(func_args).await.unwrap_err();
         assert_eq!(
-            "Execution error: Handler error: Missing FlowServiceHandler, not expected",
+            "Execution error: Missing FlowServiceHandler, not expected",
             result.to_string()
         );
     }
diff --git a/src/common/function/src/lib.rs b/src/common/function/src/lib.rs
index 36fd27381d..7abd595367 100644
--- a/src/common/function/src/lib.rs
+++ b/src/common/function/src/lib.rs
@@ -13,7 +13,6 @@
 // limitations under the License.
 
 #![feature(try_blocks)]
-#![feature(assert_matches)]
 
 mod admin;
 mod flush_flow;
diff --git a/src/common/function/src/scalars/ip/ipv4.rs b/src/common/function/src/scalars/ip/ipv4.rs
index dfaab40d68..b03fbf0a85 100644
--- a/src/common/function/src/scalars/ip/ipv4.rs
+++ b/src/common/function/src/scalars/ip/ipv4.rs
@@ -20,7 +20,10 @@ use common_query::error::InvalidFuncArgsSnafu;
 use datafusion_common::arrow::array::{Array, AsArray, StringViewBuilder, UInt32Builder};
 use datafusion_common::arrow::compute;
 use datafusion_common::arrow::datatypes::{DataType, UInt32Type};
-use datafusion_expr::{ColumnarValue, ScalarFunctionArgs, Signature, TypeSignature, Volatility};
+use datafusion_expr::{
+    Coercion, ColumnarValue, ScalarFunctionArgs, Signature, TypeSignature, TypeSignatureClass,
+    Volatility,
+};
 use derive_more::Display;
 
 use crate::function::{Function, extract_args};
@@ -44,7 +47,7 @@ impl Default for Ipv4NumToString {
     fn default() -> Self {
         Self {
             signature: Signature::new(
-                TypeSignature::Exact(vec![DataType::UInt32]),
+                TypeSignature::Coercible(vec![Coercion::new_exact(TypeSignatureClass::Integer)]),
                 Volatility::Immutable,
             ),
             aliases: ["inet_ntoa".to_string()],
@@ -70,6 +73,14 @@ impl Function for Ipv4NumToString {
         args: ScalarFunctionArgs,
     ) -> datafusion_common::Result<ColumnarValue> {
         let [arg0] = extract_args(self.name(), &args)?;
+        let arg0 = compute::cast_with_options(
+            &arg0,
+            &DataType::UInt32,
+            &compute::CastOptions {
+                safe: false,
+                ..Default::default()
+            },
+        )?;
         let uint_vec = arg0.as_primitive::<UInt32Type>();
 
         let size = uint_vec.len();
@@ -171,7 +182,7 @@ mod tests {
     use std::sync::Arc;
 
     use arrow_schema::Field;
-    use datafusion_common::arrow::array::{StringViewArray, UInt32Array};
+    use datafusion_common::arrow::array::{Int64Array, StringViewArray, UInt32Array};
 
     use super::*;
 
@@ -200,6 +211,51 @@ mod tests {
         assert_eq!(result.value(3), "255.255.255.255");
     }
 
+    #[test]
+    fn test_ipv4_num_to_string_accepts_int64() {
+        let func = Ipv4NumToString::default();
+
+        // Test data
+        let values = vec![167772161i64, 3232235521i64, 0i64, 4294967295i64];
+        let input = ColumnarValue::Array(Arc::new(Int64Array::from(values)));
+
+        let args = ScalarFunctionArgs {
+            args: vec![input],
+            arg_fields: vec![],
+            number_rows: 4,
+            return_field: Arc::new(Field::new("x", DataType::Utf8View, false)),
+            config_options: Arc::new(Default::default()),
+        };
+        let result = func.invoke_with_args(args).unwrap();
+        let result = result.to_array(4).unwrap();
+        let result = result.as_string_view();
+
+        assert_eq!(result.value(0), "10.0.0.1");
+        assert_eq!(result.value(1), "192.168.0.1");
+        assert_eq!(result.value(2), "0.0.0.0");
+        assert_eq!(result.value(3), "255.255.255.255");
+    }
+
+    #[test]
+    fn test_ipv4_num_to_string_rejects_negative_int64() {
+        let func = Ipv4NumToString::default();
+
+        // Test data
+        let values = vec![-1i64];
+        let input = ColumnarValue::Array(Arc::new(Int64Array::from(values)));
+
+        let args = ScalarFunctionArgs {
+            args: vec![input],
+            arg_fields: vec![],
+            number_rows: 1,
+            return_field: Arc::new(Field::new("x", DataType::Utf8View, false)),
+            config_options: Arc::new(Default::default()),
+        };
+        let result = func.invoke_with_args(args);
+
+        assert!(result.is_err());
+    }
+
     #[test]
     fn test_ipv4_string_to_num() {
         let func = Ipv4StringToNum::default();
diff --git a/src/common/function/src/scalars/json/json_get.rs b/src/common/function/src/scalars/json/json_get.rs
index a7b2243c3d..d081ad917b 100644
--- a/src/common/function/src/scalars/json/json_get.rs
+++ b/src/common/function/src/scalars/json/json_get.rs
@@ -12,24 +12,22 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::str::FromStr;
 use std::sync::Arc;
 
-use arrow::array::{ArrayRef, BinaryViewArray, StringViewArray, StructArray};
+use arrow::array::{ArrayRef, BinaryViewArray, new_null_array};
 use arrow::compute;
-use arrow::datatypes::{Float64Type, Int64Type, UInt64Type};
+use arrow::datatypes::Float64Type;
 use arrow_schema::Field;
 use datafusion_common::arrow::array::{
     Array, AsArray, BinaryViewBuilder, BooleanBuilder, Float64Builder, Int64Builder,
     StringViewBuilder,
 };
 use datafusion_common::arrow::datatypes::DataType;
-use datafusion_common::{DataFusionError, Result};
+use datafusion_common::{DataFusionError, Result, ScalarValue, exec_datafusion_err, exec_err};
 use datafusion_expr::{ColumnarValue, ScalarFunctionArgs, Signature, Volatility};
 use datatypes::arrow_array::{int_array_value_at_index, string_array_value_at_index};
-use datatypes::json::JsonStructureSettings;
+use datatypes::vectors::json::array::JsonArray;
 use derive_more::Display;
-use jsonpath_rust::JsonPath;
 use serde_json::Value;
 
 use crate::function::{Function, extract_args};
@@ -52,6 +50,7 @@ fn get_json_by_path(json: &[u8], path: &str) -> Option<Vec<u8>> {
 
 enum JsonResultValue<'a> {
     Jsonb(Vec<u8>),
+    #[expect(unused)]
     JsonStructByColumn(&'a ArrayRef, usize),
     JsonStructByValue(&'a Value),
 }
@@ -64,57 +63,20 @@ trait JsonGetResultBuilder {
     fn build(&mut self) -> ArrayRef;
 }
 
-/// Common implementation for JSON get scalar functions.
-///
-/// `JsonGet` encapsulates the logic for extracting values from JSON inputs
-/// based on a path expression. Different JSON get functions reuse this
-/// implementation by supplying their own `JsonGetResultBuilder` to control
-/// how the resulting values are materialized into an Arrow array.
-#[derive(Debug)]
-struct JsonGet {
-    signature: Signature,
-}
-
-impl JsonGet {
-    fn invoke<F, B>(&self, args: ScalarFunctionArgs, builder_factory: F) -> Result<ColumnarValue>
-    where
-        F: Fn(usize) -> B,
-        B: JsonGetResultBuilder,
-    {
-        let [arg0, arg1] = extract_args("JSON_GET", &args)?;
-
-        let arg1 = compute::cast(&arg1, &DataType::Utf8View)?;
-        let paths = arg1.as_string_view();
-
-        let mut builder = (builder_factory)(arg0.len());
-        match arg0.data_type() {
-            DataType::Binary | DataType::LargeBinary | DataType::BinaryView => {
-                let arg0 = compute::cast(&arg0, &DataType::BinaryView)?;
-                let jsons = arg0.as_binary_view();
-                jsonb_get(jsons, paths, &mut builder)?;
-            }
-            DataType::Struct(_) => {
-                let jsons = arg0.as_struct();
-                json_struct_get(jsons, paths, &mut builder)?
-            }
-            _ => {
-                return Err(DataFusionError::Execution(format!(
-                    "JSON_GET not supported argument type {}",
-                    arg0.data_type(),
-                )));
-            }
-        };
-
-        Ok(ColumnarValue::Array(builder.build()))
-    }
-}
-
-impl Default for JsonGet {
-    fn default() -> Self {
-        Self {
-            signature: Signature::any(2, Volatility::Immutable),
+fn result_builder(len: usize, with_type: &DataType) -> Result<Box<dyn JsonGetResultBuilder>> {
+    let builder = match with_type {
+        DataType::Utf8 | DataType::LargeUtf8 | DataType::Utf8View => {
+            Box::new(StringResultBuilder(StringViewBuilder::with_capacity(len)))
+                as Box<dyn JsonGetResultBuilder>
         }
-    }
+        DataType::Int64 => Box::new(IntResultBuilder(Int64Builder::with_capacity(len))),
+        DataType::Float64 => Box::new(FloatResultBuilder(Float64Builder::with_capacity(len))),
+        DataType::Boolean => Box::new(BoolResultBuilder(BooleanBuilder::with_capacity(len))),
+        t => {
+            return exec_err!("json_get with unknown type {t}");
+        }
+    };
+    Ok(builder)
 }
 
 // TODO: refactor this to StringLikeArrayBuilder from Arrow 57
@@ -154,7 +116,7 @@ impl JsonGetResultBuilder for StringResultBuilder {
 
 #[derive(Default, Display, Debug)]
 #[display("{}", Self::NAME.to_ascii_uppercase())]
-pub struct JsonGetString(JsonGet);
+pub struct JsonGetString(JsonGetWithType);
 
 impl JsonGetString {
     pub const NAME: &'static str = "json_get_string";
@@ -173,10 +135,10 @@ impl Function for JsonGetString {
         &self.0.signature
     }
 
-    fn invoke_with_args(&self, args: ScalarFunctionArgs) -> Result<ColumnarValue> {
-        self.0.invoke(args, |len: usize| {
-            StringResultBuilder(StringViewBuilder::with_capacity(len))
-        })
+    fn invoke_with_args(&self, mut args: ScalarFunctionArgs) -> Result<ColumnarValue> {
+        args.args
+            .push(ColumnarValue::Scalar(ScalarValue::Utf8View(None)));
+        self.0.invoke_with_args(args)
     }
 }
 
@@ -205,7 +167,7 @@ impl JsonGetResultBuilder for IntResultBuilder {
 
 #[derive(Default, Display, Debug)]
 #[display("{}", Self::NAME.to_ascii_uppercase())]
-pub struct JsonGetInt(JsonGet);
+pub struct JsonGetInt(JsonGetWithType);
 
 impl JsonGetInt {
     pub const NAME: &'static str = "json_get_int";
@@ -224,10 +186,10 @@ impl Function for JsonGetInt {
         &self.0.signature
     }
 
-    fn invoke_with_args(&self, args: ScalarFunctionArgs) -> Result<ColumnarValue> {
-        self.0.invoke(args, |len: usize| {
-            IntResultBuilder(Int64Builder::with_capacity(len))
-        })
+    fn invoke_with_args(&self, mut args: ScalarFunctionArgs) -> Result<ColumnarValue> {
+        args.args
+            .push(ColumnarValue::Scalar(ScalarValue::Int64(None)));
+        self.0.invoke_with_args(args)
     }
 }
 
@@ -264,7 +226,7 @@ impl JsonGetResultBuilder for FloatResultBuilder {
 
 #[derive(Default, Display, Debug)]
 #[display("{}", Self::NAME.to_ascii_uppercase())]
-pub struct JsonGetFloat(JsonGet);
+pub struct JsonGetFloat(JsonGetWithType);
 
 impl JsonGetFloat {
     pub const NAME: &'static str = "json_get_float";
@@ -283,10 +245,10 @@ impl Function for JsonGetFloat {
         &self.0.signature
     }
 
-    fn invoke_with_args(&self, args: ScalarFunctionArgs) -> Result<ColumnarValue> {
-        self.0.invoke(args, |len: usize| {
-            FloatResultBuilder(Float64Builder::with_capacity(len))
-        })
+    fn invoke_with_args(&self, mut args: ScalarFunctionArgs) -> Result<ColumnarValue> {
+        args.args
+            .push(ColumnarValue::Scalar(ScalarValue::Float64(None)));
+        self.0.invoke_with_args(args)
     }
 }
 
@@ -323,7 +285,7 @@ impl JsonGetResultBuilder for BoolResultBuilder {
 
 #[derive(Default, Display, Debug)]
 #[display("{}", Self::NAME.to_ascii_uppercase())]
-pub struct JsonGetBool(JsonGet);
+pub struct JsonGetBool(JsonGetWithType);
 
 impl JsonGetBool {
     pub const NAME: &'static str = "json_get_bool";
@@ -342,24 +304,23 @@ impl Function for JsonGetBool {
         &self.0.signature
     }
 
-    fn invoke_with_args(&self, args: ScalarFunctionArgs) -> Result<ColumnarValue> {
-        self.0.invoke(args, |len: usize| {
-            BoolResultBuilder(BooleanBuilder::with_capacity(len))
-        })
+    fn invoke_with_args(&self, mut args: ScalarFunctionArgs) -> Result<ColumnarValue> {
+        args.args
+            .push(ColumnarValue::Scalar(ScalarValue::Boolean(None)));
+        self.0.invoke_with_args(args)
     }
 }
 
 fn jsonb_get(
     jsons: &BinaryViewArray,
-    paths: &StringViewArray,
+    path: &str,
     builder: &mut dyn JsonGetResultBuilder,
 ) -> Result<()> {
     let size = jsons.len();
     for i in 0..size {
         let json = jsons.is_valid(i).then(|| jsons.value(i));
-        let path = paths.is_valid(i).then(|| paths.value(i));
-        let result = match (json, path) {
-            (Some(json), Some(path)) => get_json_by_path(json, path),
+        let result = match json {
+            Some(json) => get_json_by_path(json, path),
             _ => None,
         };
         if let Some(v) = result {
@@ -371,141 +332,90 @@ fn jsonb_get(
     Ok(())
 }
 
-fn json_struct_get(
-    jsons: &StructArray,
-    paths: &StringViewArray,
-    builder: &mut dyn JsonGetResultBuilder,
-) -> Result<()> {
-    let size = jsons.len();
-    for i in 0..size {
-        if jsons.is_null(i) || paths.is_null(i) {
-            builder.append_null();
-            continue;
+fn json_struct_get(array: &ArrayRef, path: &str, with_type: &DataType) -> Result<ArrayRef> {
+    let path = path.trim_start_matches("$");
+
+    // Fast path: if the JSON array fields can be directly indexed into by the `path`, simply get
+    // the sub-array (`column_by_name`).
+    let mut direct = true;
+    let mut current = array;
+    for segment in path.split(".").filter(|s| !s.is_empty()) {
+        if matches!(current.data_type(), DataType::Binary) {
+            direct = false;
+            break;
         }
-        let path = paths.value(i);
 
-        // naively assume the JSON path is our kind of indexing to the field, by removing its "root"
-        let field_path = path.trim().replace("$.", "");
-        let column = jsons.column_by_name(&field_path);
+        let Some(json) = current.as_struct_opt() else {
+            return exec_err!("unknown JSON array datatype: {}", current.data_type());
+        };
+        let Some(sub_json) = json.column_by_name(segment) else {
+            return Ok(new_null_array(with_type, array.len()));
+        };
+        current = sub_json;
+    }
 
-        if let Some(column) = column {
-            builder.append_value(JsonResultValue::JsonStructByColumn(column, i))?;
-        } else {
-            let Some(raw) = jsons
-                .column_by_name(JsonStructureSettings::RAW_FIELD)
-                .and_then(|x| string_array_value_at_index(x, i))
-            else {
+    // Build the result array with optional value mapper.
+    fn build_with<F>(input: &ArrayRef, with_type: &DataType, value_mapper: F) -> Result<ArrayRef>
+    where
+        for<'a> F: Fn(&'a Value) -> Option<&'a Value>,
+    {
+        let json_array = JsonArray::from(input);
+
+        let mut builder = result_builder(input.len(), with_type)?;
+        for i in 0..input.len() {
+            if input.is_null(i) {
                 builder.append_null();
                 continue;
-            };
+            }
 
-            let path: JsonPath<Value> = JsonPath::try_from(path).map_err(|e| {
-                DataFusionError::Execution(format!("{path} is not a valid JSON path: {e}"))
-            })?;
-            // the wanted field is not retrievable from the JSON struct columns directly, we have
-            // to combine everything (columns and the "_raw") into a complete JSON value to find it
-            let value = json_struct_to_value(raw, jsons, i)?;
+            let value = json_array
+                .try_get_value(i)
+                .map_err(|e| exec_datafusion_err!("{e}"))?;
+            let value = value_mapper(&value);
 
-            match path.find(&value) {
-                Value::Null => builder.append_null(),
-                Value::Array(values) => match values.as_slice() {
-                    [] => builder.append_null(),
-                    [x] => builder.append_value(JsonResultValue::JsonStructByValue(x))?,
-                    _ => builder.append_value(JsonResultValue::JsonStructByValue(&value))?,
-                },
-                value => builder.append_value(JsonResultValue::JsonStructByValue(&value))?,
+            if let Some(value) = value {
+                builder.append_value(JsonResultValue::JsonStructByValue(value))?;
+            } else {
+                builder.append_null();
             }
         }
+        Ok(builder.build())
     }
 
-    Ok(())
-}
-
-fn json_struct_to_value(raw: &str, jsons: &StructArray, i: usize) -> Result<Value> {
-    let Ok(mut json) = Value::from_str(raw) else {
-        return Err(DataFusionError::Internal(format!(
-            "inner field '{}' is not a valid JSON string",
-            JsonStructureSettings::RAW_FIELD
-        )));
-    };
-
-    for (column_name, column) in jsons.column_names().into_iter().zip(jsons.columns()) {
-        if column_name == JsonStructureSettings::RAW_FIELD {
-            continue;
-        }
-
-        let (json_pointer, field) = if let Some((json_object, field)) = column_name.rsplit_once(".")
-        {
-            let json_pointer = format!("/{}", json_object.replace(".", "/"));
-            (json_pointer, field)
-        } else {
-            ("".to_string(), column_name)
-        };
-        let Some(json_object) = json
-            .pointer_mut(&json_pointer)
-            .and_then(|x| x.as_object_mut())
-        else {
-            return Err(DataFusionError::Internal(format!(
-                "value at JSON pointer '{}' is not an object",
-                json_pointer
-            )));
-        };
-
-        macro_rules! insert {
-            ($column: ident, $i: ident, $json_object: ident, $field: ident) => {{
-                if let Some(value) = $column
-                    .is_valid($i)
-                    .then(|| serde_json::Value::from($column.value($i)))
-                {
-                    $json_object.insert($field.to_string(), value);
+    if direct {
+        let casted = if current.data_type() != with_type {
+            match (current.data_type(), with_type) {
+                (DataType::Binary, _) => {
+                    // Fall back to the slow path if the found JSON sub-array is serialized to bytes
+                    // (because of JSON type conflicting)
+                    build_with(current, with_type, |v| Some(v))?
                 }
-            }};
-        }
-
-        match column.data_type() {
-            // boolean => Value::Bool
-            DataType::Boolean => {
-                let column = column.as_boolean();
-                insert!(column, i, json_object, field);
+                (DataType::List(_) | DataType::Struct(_), with_type) if with_type.is_string() => {
+                    // Special handle for wanted array is string (Arrow cast is not working here if
+                    // the datatype is list or struct), because it could be used in displaying the
+                    // result.
+                    build_with(current, with_type, |v| Some(v))?
+                }
+                (_, with_type) if with_type.is_string() => {
+                    // Same special handle for wanted array is string as above, except for simply
+                    // casting by Arrow is more desirable.
+                    arrow_cast::cast(current.as_ref(), with_type)?
+                }
+                _ => new_null_array(with_type, current.len()),
             }
-            // int => Value::Number
-            DataType::Int64 => {
-                let column = column.as_primitive::<Int64Type>();
-                insert!(column, i, json_object, field);
-            }
-            DataType::UInt64 => {
-                let column = column.as_primitive::<UInt64Type>();
-                insert!(column, i, json_object, field);
-            }
-            DataType::Float64 => {
-                let column = column.as_primitive::<Float64Type>();
-                insert!(column, i, json_object, field);
-            }
-            // string => Value::String
-            DataType::Utf8 => {
-                let column = column.as_string::<i32>();
-                insert!(column, i, json_object, field);
-            }
-            DataType::LargeUtf8 => {
-                let column = column.as_string::<i64>();
-                insert!(column, i, json_object, field);
-            }
-            DataType::Utf8View => {
-                let column = column.as_string_view();
-                insert!(column, i, json_object, field);
-            }
-            // other => Value::Array and Value::Object
-            _ => {
-                return Err(DataFusionError::NotImplemented(format!(
-                    "{} is not yet supported to be executed with field {} of datatype {}",
-                    JsonGetString::NAME,
-                    column_name,
-                    column.data_type()
-                )));
-            }
-        }
+        } else {
+            current.clone()
+        };
+        return Ok(casted);
     }
-    Ok(json)
+
+    // Slow path: reconstruct the JSON array from serialized representation of conflicting JSON
+    // values: `serde_json::Value`.
+    let mut pointer = path.replace(".", "/");
+    if !pointer.starts_with("/") {
+        pointer = format!("/{}", pointer);
+    }
+    build_with(array, with_type, |value| value.pointer(&pointer))
 }
 
 /// This function is mostly called as `json_get(value, 'attr')::type` and rewritten by
@@ -513,12 +423,12 @@ fn json_struct_to_value(raw: &str, jsons: &StructArray, i: usize) -> Result<Valu
 /// use the third argument's type to determine the return type.
 #[derive(Debug, Display)]
 #[display("{}", Self::NAME.to_ascii_uppercase())]
-pub(super) struct JsonGetWithType {
+pub struct JsonGetWithType {
     signature: Signature,
 }
 
 impl JsonGetWithType {
-    pub(crate) const NAME: &'static str = "json_get";
+    pub const NAME: &'static str = "json_get";
 }
 
 impl Default for JsonGetWithType {
@@ -565,47 +475,47 @@ impl Function for JsonGetWithType {
         &self,
         args: ScalarFunctionArgs,
     ) -> datafusion_common::Result<ColumnarValue> {
-        let [arg0, arg1, _] = extract_args(self.name(), &args)?;
+        let args_len = args.args.len();
+        if args_len != 2 && args_len != 3 {
+            return exec_err!("json_get expects 2 or 3 arguments, got {args_len}");
+        }
+
+        let arg0 = args.args[0].to_array(args.number_rows)?;
         let len = arg0.len();
 
-        let arg1 = compute::cast(&arg1, &DataType::Utf8View)?;
-        let paths = arg1.as_string_view();
-
-        // mapping datatypes returned from return_field_from_args
-        let mut builder: Box<dyn JsonGetResultBuilder> = match args.return_field.data_type() {
-            DataType::Utf8View => {
-                Box::new(StringResultBuilder(StringViewBuilder::with_capacity(len)))
-            }
-            DataType::Int64 => Box::new(IntResultBuilder(Int64Builder::with_capacity(len))),
-            DataType::Float64 => Box::new(FloatResultBuilder(Float64Builder::with_capacity(len))),
-            DataType::Boolean => Box::new(BoolResultBuilder(BooleanBuilder::with_capacity(len))),
-            _type => {
-                return Err(DataFusionError::Internal(format!(
-                    "Unsupported return type {}",
-                    _type
-                )));
-            }
+        let path = if let ColumnarValue::Scalar(path) = &args.args[1]
+            && let Some(Some(path)) = path.try_as_str()
+        {
+            path
+        } else {
+            return exec_err!(
+                r#"json_get expects a string literal "path" argument, got {}"#,
+                args.args[1]
+            );
         };
 
-        match arg0.data_type() {
+        let with_type = args
+            .args
+            .get(2)
+            .map(|x| x.data_type())
+            .unwrap_or(DataType::Utf8View);
+
+        let result = match arg0.data_type() {
             DataType::Binary | DataType::LargeBinary | DataType::BinaryView => {
                 let arg0 = compute::cast(&arg0, &DataType::BinaryView)?;
                 let jsons = arg0.as_binary_view();
-                jsonb_get(jsons, paths, builder.as_mut())?;
-            }
-            DataType::Struct(_) => {
-                let jsons = arg0.as_struct();
-                json_struct_get(jsons, paths, builder.as_mut())?;
+
+                let mut builder = result_builder(len, &with_type)?;
+                jsonb_get(jsons, path, builder.as_mut())?;
+                builder.build()
             }
+            DataType::Struct(_) => json_struct_get(&arg0, path, &with_type)?,
             _ => {
-                return Err(DataFusionError::Execution(format!(
-                    "JSON_GET not supported argument type {}",
-                    arg0.data_type(),
-                )));
+                return exec_err!("JSON_GET not supported argument type {}", arg0.data_type());
             }
         };
 
-        Ok(ColumnarValue::Array(builder.build()))
+        Ok(ColumnarValue::Array(result))
     }
 }
 
@@ -686,8 +596,8 @@ impl Function for JsonGetObject {
 mod tests {
     use std::sync::Arc;
 
-    use arrow::array::{Float64Array, Int64Array, StructArray};
-    use arrow_schema::Field;
+    use arrow::array::{BooleanArray, Int64Array, StructArray};
+    use arrow_schema::{Field, Fields};
     use datafusion_common::ScalarValue;
     use datafusion_common::arrow::array::{BinaryArray, BinaryViewArray, StringArray};
     use datafusion_common::arrow::datatypes::{Float64Type, Int64Type};
@@ -712,29 +622,35 @@ mod tests {
     /// }
     /// ```
     fn test_json_struct() -> ArrayRef {
+        let payload_fields = Fields::from(vec![
+            Field::new("code", DataType::Int64, true),
+            Field::new("success", DataType::Boolean, true),
+            Field::new("result", DataType::Binary, true),
+        ]);
         Arc::new(StructArray::new(
             vec![
                 Field::new("kind", DataType::Utf8, true),
-                Field::new("payload.code", DataType::Int64, true),
-                Field::new("payload.result.time_cost", DataType::Float64, true),
-                Field::new(JsonStructureSettings::RAW_FIELD, DataType::Utf8View, true),
+                Field::new("payload", DataType::Struct(payload_fields.clone()), true),
             ]
             .into(),
             vec![
                 Arc::new(StringArray::from_iter([Some("foo")])) as ArrayRef,
-                Arc::new(Int64Array::from_iter([Some(404)])),
-                Arc::new(Float64Array::from_iter([Some(1.234)])),
-                Arc::new(StringViewArray::from_iter([Some(
-                    json! ({
-                        "payload": {
-                            "success": false,
-                            "result": {
-                                "error": "not found"
-                            }
-                        }
-                    })
-                    .to_string(),
-                )])),
+                Arc::new(StructArray::new(
+                    payload_fields,
+                    vec![
+                        Arc::new(Int64Array::from_iter([Some(404)])) as ArrayRef,
+                        Arc::new(BooleanArray::from_iter([Some(false)])),
+                        Arc::new(BinaryArray::from_iter([Some(
+                            json!({
+                                "error": "not found",
+                                "time_cost": 1.234
+                            })
+                            .to_string()
+                            .as_bytes(),
+                        )])),
+                    ],
+                    None,
+                )),
             ],
             None,
         ))
@@ -1156,7 +1072,7 @@ mod tests {
                 args: vec![
                     ColumnarValue::Array(json.clone()),
                     ColumnarValue::Scalar(path.into()),
-                    ColumnarValue::Scalar(ScalarValue::Utf8(Some("string".to_string()))),
+                    ColumnarValue::Scalar(ScalarValue::Utf8View(None)),
                 ],
                 arg_fields: vec![],
                 number_rows: 1,
@@ -1194,7 +1110,7 @@ mod tests {
                 args: vec![
                     ColumnarValue::Array(json),
                     ColumnarValue::Scalar((*path).into()),
-                    ColumnarValue::Scalar(ScalarValue::Utf8(Some("int".to_string()))),
+                    ColumnarValue::Scalar(ScalarValue::Int64(None)),
                 ],
                 arg_fields: vec![],
                 number_rows: 1,
@@ -1232,7 +1148,7 @@ mod tests {
                 args: vec![
                     ColumnarValue::Array(json),
                     ColumnarValue::Scalar((*path).into()),
-                    ColumnarValue::Scalar(ScalarValue::Utf8(Some("float".to_string()))),
+                    ColumnarValue::Scalar(ScalarValue::Float64(None)),
                 ],
                 arg_fields: vec![],
                 number_rows: 1,
@@ -1270,7 +1186,7 @@ mod tests {
                 args: vec![
                     ColumnarValue::Array(json),
                     ColumnarValue::Scalar((*path).into()),
-                    ColumnarValue::Scalar(ScalarValue::Utf8(Some("bool".to_string()))),
+                    ColumnarValue::Scalar(ScalarValue::Boolean(None)),
                 ],
                 arg_fields: vec![],
                 number_rows: 1,
diff --git a/src/common/function/src/scalars/json/json_get_rewriter.rs b/src/common/function/src/scalars/json/json_get_rewriter.rs
index 69cea9d443..137b307412 100644
--- a/src/common/function/src/scalars/json/json_get_rewriter.rs
+++ b/src/common/function/src/scalars/json/json_get_rewriter.rs
@@ -40,92 +40,111 @@ impl FunctionRewrite for JsonGetRewriter {
         _schema: &DFSchema,
         _config: &ConfigOptions,
     ) -> Result<Transformed<Expr>> {
-        let transform = match &expr {
-            Expr::Cast(cast) => rewrite_json_get_cast(cast),
-            Expr::ScalarFunction(scalar_func) => rewrite_arrow_cast_json_get(scalar_func),
-            _ => None,
-        };
-        Ok(transform.unwrap_or_else(|| Transformed::no(expr)))
+        Ok(match expr {
+            Expr::Cast(cast) => inject_type_from_cast_expr(cast)?,
+            Expr::ScalarFunction(cast) => inject_type_from_cast_func(cast)?,
+            expr => Transformed::no(expr),
+        })
     }
 }
 
-fn is_json_get_function_call(scalar_func: &ScalarFunction) -> bool {
-    scalar_func.func.name().to_ascii_lowercase() == JsonGetWithType::NAME
-        && scalar_func.args.len() == 2
+// Expr::Cast(
+//   Expr::ScalarFunction(
+//     json_get(column, path),
+//     <data_type>
+//   )
+// )
+// =>
+// Expr::ScalarFunction(
+//   json_get(column, path, <data_type>)
+// )
+fn inject_type_from_cast_expr(cast: Cast) -> Result<Transformed<Expr>> {
+    let Cast { expr, data_type } = cast;
+
+    let mut json_get = match *expr {
+        Expr::ScalarFunction(f)
+            if f.func.name().eq_ignore_ascii_case(JsonGetWithType::NAME) && f.args.len() == 2 =>
+        {
+            f
+        }
+        expr => {
+            return Ok(Transformed::no(Expr::Cast(Cast {
+                expr: Box::new(expr),
+                data_type,
+            })));
+        }
+    };
+
+    let with_type = ScalarValue::try_new_null(&data_type).map(|x| Expr::Literal(x, None))?;
+    json_get.args.push(with_type);
+    Ok(Transformed::yes(Expr::ScalarFunction(json_get)))
 }
 
-fn rewrite_json_get_cast(cast: &Cast) -> Option<Transformed<Expr>> {
-    let scalar_func = extract_scalar_function(&cast.expr)?;
-    if is_json_get_function_call(scalar_func) {
-        let null_expr = Expr::Literal(ScalarValue::Null, None);
-        let null_cast = Expr::Cast(datafusion::logical_expr::expr::Cast {
-            expr: Box::new(null_expr),
-            data_type: cast.data_type.clone(),
-        });
+// Expr::ScalarFunction(
+//   arrow_cast(
+//     Expr::ScalarFunction(
+//       json_get(column, path),
+//     ),
+//     <data_type>
+//   )
+// )
+// =>
+// Expr::ScalarFunction(
+//   json_get(column, path, <data_type>)
+// )
+fn inject_type_from_cast_func(cast: ScalarFunction) -> Result<Transformed<Expr>> {
+    let ScalarFunction { func, args } = cast;
 
-        let mut args = scalar_func.args.clone();
-        args.push(null_cast);
-
-        Some(Transformed::yes(Expr::ScalarFunction(ScalarFunction {
-            func: scalar_func.func.clone(),
-            args,
-        })))
-    } else {
-        None
-    }
-}
-
-// Handle Arrow cast function: cast(json_get(a, 'path'), 'Int64')
-fn rewrite_arrow_cast_json_get(scalar_func: &ScalarFunction) -> Option<Transformed<Expr>> {
     // Check if this is an Arrow cast function
     // The function name might be "arrow_cast" or similar
-    let func_name = scalar_func.func.name().to_ascii_lowercase();
+    let func_name = func.name().to_ascii_lowercase();
     if !func_name.contains("arrow_cast") {
-        return None;
+        let original = Expr::ScalarFunction(ScalarFunction { func, args });
+        return Ok(Transformed::no(original));
     }
 
     // Arrow cast function should have exactly 2 arguments:
     // 1. The expression to cast (could be json_get)
     // 2. The target type as a string literal
-    if scalar_func.args.len() != 2 {
-        return None;
+    if args.len() != 2 {
+        let original = Expr::ScalarFunction(ScalarFunction { func, args });
+        return Ok(Transformed::no(original));
     }
+    let [arg0, arg1] = args.try_into().unwrap_or_else(|_| unreachable!());
 
-    // Extract the inner json_get function
-    let json_get_func = extract_scalar_function(&scalar_func.args[0])?;
-
-    // Check if it's a json_get function
-    if is_json_get_function_call(json_get_func) {
-        // Get the target type from the second argument
-        let target_type = extract_string_literal(&scalar_func.args[1])?;
-        let data_type = parse_data_type_from_string(&target_type)?;
-
-        // Create the null expression with the same type
-        let null_expr = Expr::Literal(ScalarValue::Null, None);
-        let null_cast = Expr::Cast(datafusion::logical_expr::expr::Cast {
-            expr: Box::new(null_expr),
-            data_type,
+    let Some(with_type) = arg1
+        .as_literal()
+        .and_then(|x| x.try_as_str())
+        .flatten()
+        .and_then(parse_data_type_from_string)
+    else {
+        let original = Expr::ScalarFunction(ScalarFunction {
+            func,
+            args: vec![arg0, arg1],
         });
+        return Ok(Transformed::no(original));
+    };
 
-        // Create the new json_get_with_type function with the null parameter
-        let mut args = json_get_func.args.clone();
-        args.push(null_cast);
+    let mut json_get = match arg0 {
+        Expr::ScalarFunction(f)
+            if f.func.name().eq_ignore_ascii_case(JsonGetWithType::NAME) && f.args.len() == 2 =>
+        {
+            f
+        }
+        arg0 => {
+            let original = Expr::ScalarFunction(ScalarFunction {
+                func,
+                args: vec![arg0, arg1],
+            });
+            return Ok(Transformed::no(original));
+        }
+    };
 
-        Some(Transformed::yes(Expr::ScalarFunction(ScalarFunction {
-            func: json_get_func.func.clone(),
-            args,
-        })))
-    } else {
-        None
-    }
-}
+    let with_type = ScalarValue::try_new_null(&with_type).map(|x| Expr::Literal(x, None))?;
+    json_get.args.push(with_type);
 
-// Extract string literal from an expression
-fn extract_string_literal(expr: &Expr) -> Option<String> {
-    match expr {
-        Expr::Literal(ScalarValue::Utf8(Some(s)), _) => Some(s.clone()),
-        _ => None,
-    }
+    let rewritten = Expr::ScalarFunction(json_get);
+    Ok(Transformed::yes(rewritten))
 }
 
 // Parse a data type from a string representation
@@ -149,13 +168,6 @@ fn parse_data_type_from_string(type_str: &str) -> Option<DataType> {
     }
 }
 
-fn extract_scalar_function(expr: &Expr) -> Option<&ScalarFunction> {
-    match expr {
-        Expr::ScalarFunction(func) => Some(func),
-        _ => None,
-    }
-}
-
 #[cfg(test)]
 mod tests {
     use arrow_schema::DataType;
@@ -221,12 +233,8 @@ mod tests {
 
                 // Third argument should be a null cast to Int8
                 match &func.args[2] {
-                    Expr::Cast(Cast { expr, data_type }) => {
-                        assert_eq!(*data_type, DataType::Int8);
-                        match expr.as_ref() {
-                            Expr::Literal(ScalarValue::Null, _) => {}
-                            _ => panic!("Third argument should be a null cast"),
-                        }
+                    Expr::Literal(value, _) => {
+                        assert_eq!(value.data_type(), DataType::Int8);
                     }
                     _ => panic!("Third argument should be a cast expression"),
                 }
@@ -314,12 +322,8 @@ mod tests {
 
                 // Third argument should be a null cast to Int64
                 match &func.args[2] {
-                    Expr::Cast(Cast { expr, data_type }) => {
-                        assert_eq!(*data_type, DataType::Int64);
-                        match expr.as_ref() {
-                            Expr::Literal(ScalarValue::Null, _) => {}
-                            _ => panic!("Third argument should be a null cast"),
-                        }
+                    Expr::Literal(value, _) => {
+                        assert_eq!(value.data_type(), DataType::Int64);
                     }
                     _ => panic!("Third argument should be a cast expression"),
                 }
diff --git a/src/common/function/src/scalars/json/json_to_string.rs b/src/common/function/src/scalars/json/json_to_string.rs
index 6c0cc260b2..6364dff4de 100644
--- a/src/common/function/src/scalars/json/json_to_string.rs
+++ b/src/common/function/src/scalars/json/json_to_string.rs
@@ -19,6 +19,7 @@ use datafusion_common::DataFusionError;
 use datafusion_common::arrow::array::{Array, AsArray, StringViewBuilder};
 use datafusion_common::arrow::datatypes::DataType;
 use datafusion_expr::{ColumnarValue, ScalarFunctionArgs, Signature, Volatility};
+use datatypes::types::jsonb_to_string;
 
 use crate::function::{Function, extract_args};
 
@@ -74,7 +75,7 @@ impl Function for JsonToStringFunction {
         for i in 0..size {
             let json = jsons.is_valid(i).then(|| jsons.value(i));
             let result = json
-                .map(|json| jsonb::from_slice(json).map(|x| x.to_string()))
+                .map(jsonb_to_string)
                 .transpose()
                 .map_err(|e| DataFusionError::Execution(format!("invalid json binary: {e}")))?;
 
diff --git a/src/common/function/src/scalars/matches.rs b/src/common/function/src/scalars/matches.rs
index 821a1b0581..b5de60dc85 100644
--- a/src/common/function/src/scalars/matches.rs
+++ b/src/common/function/src/scalars/matches.rs
@@ -794,16 +794,12 @@ impl Tokenizer {
                     is_quote_present = true;
                     break;
                 }
-                ' ' => {
-                    if !is_quoted {
-                        break;
-                    }
+                ' ' if !is_quoted => {
+                    break;
                 }
-                '(' | ')' | '+' | '-' => {
-                    if !is_quoted {
-                        self.rewind_one();
-                        break;
-                    }
+                '(' | ')' | '+' | '-' if !is_quoted => {
+                    self.rewind_one();
+                    break;
                 }
                 '\\' => {
                     let Some(next) = self.consume_next(pattern) else {
diff --git a/src/common/function/src/scalars/matches_term.rs b/src/common/function/src/scalars/matches_term.rs
index 8dfb25cbc0..ec1b34d408 100644
--- a/src/common/function/src/scalars/matches_term.rs
+++ b/src/common/function/src/scalars/matches_term.rs
@@ -20,6 +20,8 @@ use datafusion_common::arrow::compute;
 use datafusion_common::arrow::datatypes::DataType;
 use datafusion_common::{DataFusionError, ScalarValue};
 use datafusion_expr::{ColumnarValue, ScalarFunctionArgs, Signature, Volatility};
+use icu_properties::props::Script;
+use icu_properties::{CodePointMapData, CodePointMapDataBorrowed};
 use memchr::memmem;
 
 use crate::function::Function;
@@ -27,10 +29,11 @@ use crate::function_registry::FunctionRegistry;
 
 /// Exact term/phrase matching function for text columns.
 ///
-/// This function checks if a text column contains exact term/phrase matches
-/// with non-alphanumeric boundaries. Designed for:
-/// - Whole-word matching (e.g. "cat" in "cat!" but not in "category")
+/// This function uses script-aware matching rules:
+/// - ASCII-only terms keep whole-word style boundary matching, like Whole-word matching (e.g. "cat" in "cat!" but not in "category")
 /// - Phrase matching (e.g. "hello world" in "note:hello world!")
+/// - Terms containing Han characters match as contiguous substrings
+/// - Mixed-script identifiers and numeric terms remain searchable in Chinese text
 ///
 /// # Signature
 /// `matches_term(text: String, term: String) -> Boolean`
@@ -43,9 +46,8 @@ use crate::function_registry::FunctionRegistry;
 /// BooleanVector where each element indicates if the corresponding text
 /// contains an exact match of the term, following these rules:
 /// 1. Exact substring match found (case-sensitive)
-/// 2. Match boundaries are either:
-///    - Start/end of text
-///    - Any non-alphanumeric character (including spaces, hyphens, punctuation, etc.)
+/// 2. For ASCII-only terms, adjacent ASCII word characters block the match
+/// 3. For Han-containing terms, contiguous substring match is sufficient
 ///
 /// # Examples
 /// ```
@@ -60,6 +62,9 @@ use crate::function_registry::FunctionRegistry;
 /// SELECT matches_term(column, 'critical error') FROM logs;
 /// -- Match in: "ERROR:critical error!"
 /// -- No match: "critical_errors"
+/// -- Chinese substring examples --
+/// SELECT matches_term(column, '手机') FROM table;
+/// -- Text: "登录手机号18888888888的动态key" => true
 ///
 /// -- Empty string handling --
 /// SELECT matches_term(column, '') FROM table;
@@ -204,9 +209,8 @@ impl Function for MatchesTermFunction {
 ///
 /// A term is considered matched when:
 /// 1. The exact sequence appears in the text
-/// 2. It is either:
-///    - At the start/end of text with adjacent non-alphanumeric character
-///    - Surrounded by non-alphanumeric characters
+/// 2. ASCII-only terms are not adjacent to ASCII word characters
+/// 3. Han-containing terms match as contiguous substrings
 ///
 /// # Examples
 /// ```
@@ -215,28 +219,105 @@ impl Function for MatchesTermFunction {
 /// assert!(finder.find("dog,cat"));   // Term preceded by comma
 /// assert!(!finder.find("category")); // Partial match rejected
 ///
-/// let finder = MatchesTermFinder::new("world");
-/// assert!(finder.find("hello-world")); // Hyphen boundary
+/// let finder = MatchesTermFinder::new("手机");
+/// assert!(finder.find("登录手机号18888888888的动态key"));
 /// ```
 #[derive(Clone, Debug)]
 pub struct MatchesTermFinder {
     finder: memmem::Finder<'static>,
     term: String,
-    starts_with_non_alnum: bool,
-    ends_with_non_alnum: bool,
+    term_kind: TermKind,
+    starts_with_other: bool,
+    ends_with_other: bool,
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+enum CharClass {
+    AsciiWord,
+    Han,
+    UnicodeWord,
+    Other,
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+enum TermKind {
+    AsciiLike,
+    UnicodeWord,
+    HanContaining,
+}
+
+fn classify_char(c: char) -> CharClass {
+    if c.is_ascii_alphanumeric() {
+        CharClass::AsciiWord
+    } else if is_han(c) {
+        CharClass::Han
+    } else if c.is_alphanumeric() {
+        CharClass::UnicodeWord
+    } else {
+        CharClass::Other
+    }
+}
+
+static HAN_SCRIPT_DATA: CodePointMapDataBorrowed<'static, Script> =
+    CodePointMapData::<Script>::new();
+
+fn is_han(c: char) -> bool {
+    HAN_SCRIPT_DATA.get(c) == Script::Han
+}
+
+fn classify_term(term: &str) -> TermKind {
+    let mut has_han = false;
+    let mut has_unicode_word = false;
+    for c in term.chars() {
+        match classify_char(c) {
+            CharClass::AsciiWord => {}
+            CharClass::Han => has_han = true,
+            CharClass::UnicodeWord => has_unicode_word = true,
+            CharClass::Other => {}
+        }
+    }
+
+    if has_han {
+        TermKind::HanContaining
+    } else if has_unicode_word {
+        TermKind::UnicodeWord
+    } else {
+        TermKind::AsciiLike
+    }
+}
+
+fn boundary_ok(term_kind: TermKind, neighbor: Option<char>, term_has_other_boundary: bool) -> bool {
+    if term_has_other_boundary {
+        return true;
+    }
+
+    match term_kind {
+        TermKind::AsciiLike => !matches!(neighbor.map(classify_char), Some(CharClass::AsciiWord)),
+        TermKind::UnicodeWord => !matches!(
+            neighbor.map(classify_char),
+            Some(CharClass::AsciiWord | CharClass::UnicodeWord | CharClass::Han)
+        ),
+        TermKind::HanContaining => true,
+    }
 }
 
 impl MatchesTermFinder {
     /// Create a new `MatchesTermFinder` for the given term.
     pub fn new(term: &str) -> Self {
-        let starts_with_non_alnum = term.chars().next().is_some_and(|c| !c.is_alphanumeric());
-        let ends_with_non_alnum = term.chars().last().is_some_and(|c| !c.is_alphanumeric());
-
+        let starts_with_other = term
+            .chars()
+            .next()
+            .is_some_and(|c| classify_char(c) == CharClass::Other);
+        let ends_with_other = term
+            .chars()
+            .last()
+            .is_some_and(|c| classify_char(c) == CharClass::Other);
         Self {
             finder: memmem::Finder::new(term).into_owned(),
             term: term.to_string(),
-            starts_with_non_alnum,
-            ends_with_non_alnum,
+            term_kind: classify_term(term),
+            starts_with_other,
+            ends_with_other,
         }
     }
 
@@ -254,21 +335,17 @@ impl MatchesTermFinder {
         while let Some(found_pos) = self.finder.find(&text.as_bytes()[pos..]) {
             let actual_pos = pos + found_pos;
 
-            let prev_ok = self.starts_with_non_alnum
-                || text[..actual_pos]
-                    .chars()
-                    .last()
-                    .map(|c| !c.is_alphanumeric())
-                    .unwrap_or(true);
+            let prev = text[..actual_pos].chars().last();
+            let prev_ok = self.starts_with_other || boundary_ok(self.term_kind, prev, false);
 
             if prev_ok {
+                if self.term_kind == TermKind::HanContaining {
+                    return true;
+                }
+
                 let next_pos = actual_pos + self.finder.needle().len();
-                let next_ok = self.ends_with_non_alnum
-                    || text[next_pos..]
-                        .chars()
-                        .next()
-                        .map(|c| !c.is_alphanumeric())
-                        .unwrap_or(true);
+                let next = text[next_pos..].chars().next();
+                let next_ok = self.ends_with_other || boundary_ok(self.term_kind, next, false);
 
                 if next_ok {
                     return true;
@@ -369,6 +446,25 @@ mod tests {
         assert!(!MatchesTermFinder::new("v1.0").find("v1.0a"));
     }
 
+    #[test]
+    fn mixed_script_terms_match_inside_chinese_context() {
+        let text = "登录手机号18888888888的动态key";
+        assert!(MatchesTermFinder::new("手机号").find(text));
+        assert!(MatchesTermFinder::new("18888888888").find(text));
+        assert!(MatchesTermFinder::new("手机").find(text));
+        assert!(MatchesTermFinder::new("机号").find(text));
+        assert!(MatchesTermFinder::new("机号1888").find(text));
+        assert!(MatchesTermFinder::new("农业").find("中国农业银行"));
+        assert!(MatchesTermFinder::new("error").find("错误error日志"));
+    }
+
+    #[test]
+    fn underscore_still_counts_as_boundary_for_ascii_terms() {
+        assert!(MatchesTermFinder::new("world").find("hello_world"));
+        assert!(MatchesTermFinder::new("id").find("trace_id=abc"));
+        assert!(!MatchesTermFinder::new("error").find("criticalerrors"));
+    }
+
     #[test]
     fn adjacent_alphanumeric_fails() {
         assert!(!MatchesTermFinder::new("cat").find("cat5"));
@@ -406,4 +502,18 @@ mod tests {
         assert!(MatchesTermFinder::new("中文").find("这是中文测试，中文！"));
         assert!(MatchesTermFinder::new("error").find("错误errorerror日志_error!"));
     }
+
+    #[test]
+    fn han_terms_match_as_contiguous_substrings() {
+        assert!(MatchesTermFinder::new("行账号").find("中国农业银行账号"));
+        assert!(MatchesTermFinder::new("登录").find("登录手机号18888888888的动态key"));
+    }
+
+    #[test]
+    fn han_detection_uses_script_not_all_cjk() {
+        assert!(is_han('汉'));
+        assert!(is_han('\u{30000}'));
+        assert!(!is_han('あ'));
+        assert!(!is_han('한'));
+    }
 }
diff --git a/src/common/function/src/scalars/vector.rs b/src/common/function/src/scalars/vector.rs
index 968231aa0a..2f8772c410 100644
--- a/src/common/function/src/scalars/vector.rs
+++ b/src/common/function/src/scalars/vector.rs
@@ -141,7 +141,7 @@ where
             results.push((self.func)(v0, v1)?);
         }
 
-        let results = ScalarValue::iter_to_array(results.into_iter())?;
+        let results = ScalarValue::iter_to_array(results)?;
         Ok(ColumnarValue::Array(results))
     }
 }
@@ -200,7 +200,7 @@ where
             }
         }
 
-        let results = ScalarValue::iter_to_array(results.into_iter())?;
+        let results = ScalarValue::iter_to_array(results)?;
         Ok(ColumnarValue::Array(results))
     }
 }
@@ -232,7 +232,7 @@ where
             results.push((self.func)(&v)?);
         }
 
-        let results = ScalarValue::iter_to_array(results.into_iter())?;
+        let results = ScalarValue::iter_to_array(results)?;
         Ok(ColumnarValue::Array(results))
     }
 }
diff --git a/src/common/function/src/scalars/vector/convert/parse_vector.rs b/src/common/function/src/scalars/vector/convert/parse_vector.rs
index 0d83f098db..7a112a4453 100644
--- a/src/common/function/src/scalars/vector/convert/parse_vector.rs
+++ b/src/common/function/src/scalars/vector/convert/parse_vector.rs
@@ -167,7 +167,7 @@ mod tests {
             "External error: Invalid vector string: [7.0,hello,9.0]",
         ];
 
-        for (input, expected) in inputs.into_iter().zip(expected.into_iter()) {
+        for (input, expected) in inputs.into_iter().zip(expected) {
             let args = ScalarFunctionArgs {
                 args: vec![ColumnarValue::Array(Arc::new(input))],
                 arg_fields: vec![],
diff --git a/src/common/function/src/system/version.rs b/src/common/function/src/system/version.rs
index 1c148bd7ab..9170fb29f1 100644
--- a/src/common/function/src/system/version.rs
+++ b/src/common/function/src/system/version.rs
@@ -43,14 +43,19 @@ impl Function for VersionFunction {
         let version = match func_ctx.query_ctx.channel() {
             Channel::Mysql => {
                 format!(
-                    "{}-greptimedb-{}",
+                    "{}-{}-{}",
                     std::env::var("GREPTIMEDB_MYSQL_SERVER_VERSION")
                         .unwrap_or_else(|_| "8.4.2".to_string()),
+                    common_version::product_name(),
                     common_version::version()
                 )
             }
             Channel::Postgres => {
-                format!("PostgreSQL 16.3 GreptimeDB {}", common_version::version())
+                format!(
+                    "PostgreSQL 16.3 {} {}",
+                    common_version::product_name(),
+                    common_version::version()
+                )
             }
             _ => common_version::version().to_string(),
         };
diff --git a/src/common/grpc-expr/src/error.rs b/src/common/grpc-expr/src/error.rs
index 31f19ba4ea..fe478350d8 100644
--- a/src/common/grpc-expr/src/error.rs
+++ b/src/common/grpc-expr/src/error.rs
@@ -90,9 +90,9 @@ pub enum Error {
     },
 
     #[snafu(display(
-        "Fulltext or Skipping index only supports string type, column: {column_name}, unexpected type: {column_type:?}"
+        "Fulltext index only supports string type, column: {column_name}, unexpected type: {column_type:?}"
     ))]
-    InvalidStringIndexColumnType {
+    InvalidFulltextIndexColumnType {
         column_name: String,
         column_type: ColumnDataType,
         #[snafu(implicit)]
@@ -183,7 +183,7 @@ impl ErrorExt for Error {
             Error::InvalidColumnDef { source, .. } => source.status_code(),
             Error::UnknownLocationType { .. } => StatusCode::InvalidArguments,
 
-            Error::UnknownColumnDataType { .. } | Error::InvalidStringIndexColumnType { .. } => {
+            Error::UnknownColumnDataType { .. } | Error::InvalidFulltextIndexColumnType { .. } => {
                 StatusCode::InvalidArguments
             }
             Error::InvalidSetTableOptionRequest { .. }
diff --git a/src/common/grpc-expr/src/util.rs b/src/common/grpc-expr/src/util.rs
index 84e268b687..aab9c2c19e 100644
--- a/src/common/grpc-expr/src/util.rs
+++ b/src/common/grpc-expr/src/util.rs
@@ -15,7 +15,7 @@
 use std::collections::HashSet;
 
 use api::v1::column_data_type_extension::TypeExt;
-use api::v1::column_def::{contains_fulltext, contains_skipping};
+use api::v1::column_def::contains_fulltext;
 use api::v1::{
     AddColumn, AddColumns, Column, ColumnDataType, ColumnDataTypeExtension, ColumnDef,
     ColumnOptions, ColumnSchema, CreateTableExpr, JsonTypeExtension, SemanticType,
@@ -27,7 +27,7 @@ use table::table_reference::TableReference;
 
 use crate::error::{
     self, DuplicatedColumnNameSnafu, DuplicatedTimestampColumnSnafu,
-    InvalidStringIndexColumnTypeSnafu, MissingTimestampColumnSnafu, Result,
+    InvalidFulltextIndexColumnTypeSnafu, MissingTimestampColumnSnafu, Result,
     UnknownColumnDataTypeSnafu,
 };
 pub struct ColumnExpr<'a> {
@@ -152,9 +152,8 @@ pub fn build_create_table_expr(
         let column_type = infer_column_datatype(datatype, datatype_extension)?;
 
         ensure!(
-            (!contains_fulltext(options) && !contains_skipping(options))
-                || column_type == ColumnDataType::String,
-            InvalidStringIndexColumnTypeSnafu {
+            !contains_fulltext(options) || column_type == ColumnDataType::String,
+            InvalidFulltextIndexColumnTypeSnafu {
                 column_name,
                 column_type,
             }
@@ -245,6 +244,7 @@ mod tests {
     use api::helper::ColumnDataTypeWrapper;
     use api::v1::column::Values;
     use api::v1::column_data_type_extension::TypeExt;
+    use api::v1::column_def::{options_from_fulltext, options_from_skipping};
     use api::v1::{
         Column, ColumnDataType, ColumnDataTypeExtension, Decimal128, DecimalTypeExtension,
         IntervalMonthDayNano, SemanticType,
@@ -253,7 +253,7 @@ mod tests {
     use common_time::interval::IntervalUnit;
     use common_time::timestamp::TimeUnit;
     use datatypes::data_type::ConcreteDataType;
-    use datatypes::schema::{ColumnSchema, SchemaBuilder};
+    use datatypes::schema::{ColumnSchema, FulltextOptions, SchemaBuilder, SkippingIndexOptions};
     use snafu::ResultExt;
 
     use super::*;
@@ -295,6 +295,21 @@ mod tests {
         )
     }
 
+    fn build_proto_column_schema(
+        column_name: &str,
+        datatype: ColumnDataType,
+        semantic_type: SemanticType,
+        options: Option<ColumnOptions>,
+    ) -> api::v1::ColumnSchema {
+        api::v1::ColumnSchema {
+            column_name: column_name.to_string(),
+            datatype: datatype as i32,
+            semantic_type: semantic_type as i32,
+            options,
+            ..Default::default()
+        }
+    }
+
     #[test]
     fn test_build_create_table_request() {
         let table_id = Some(10);
@@ -528,6 +543,68 @@ mod tests {
         assert!(host_column.add_if_not_exists);
     }
 
+    #[test]
+    fn test_build_create_table_expr_allows_skipping_index_on_int_column() {
+        let table_name = TableReference::full("", "", "test_metric");
+        let column_schemas = vec![
+            build_proto_column_schema(
+                "value",
+                ColumnDataType::Int64,
+                SemanticType::Field,
+                options_from_skipping(&SkippingIndexOptions::default()).unwrap(),
+            ),
+            build_proto_column_schema(
+                "ts",
+                ColumnDataType::TimestampMillisecond,
+                SemanticType::Timestamp,
+                None,
+            ),
+        ];
+
+        let result = build_create_table_expr(
+            None,
+            &table_name,
+            ColumnExpr::from_column_schemas(&column_schemas),
+            MITO_ENGINE,
+            "Created on insertion",
+        );
+
+        assert!(result.is_ok());
+    }
+
+    #[test]
+    fn test_build_create_table_expr_rejects_fulltext_index_on_non_string_column() {
+        let table_name = TableReference::full("", "", "test_metric");
+        let column_schemas = vec![
+            build_proto_column_schema(
+                "value",
+                ColumnDataType::Int64,
+                SemanticType::Field,
+                options_from_fulltext(&FulltextOptions {
+                    enable: true,
+                    ..Default::default()
+                })
+                .unwrap(),
+            ),
+            build_proto_column_schema(
+                "ts",
+                ColumnDataType::TimestampMillisecond,
+                SemanticType::Timestamp,
+                None,
+            ),
+        ];
+
+        let result = build_create_table_expr(
+            None,
+            &table_name,
+            ColumnExpr::from_column_schemas(&column_schemas),
+            MITO_ENGINE,
+            "Created on insertion",
+        );
+
+        assert!(result.is_err());
+    }
+
     fn mock_insert_batch() -> (Vec<Column>, u32) {
         let row_count = 2;
 
diff --git a/src/common/grpc/src/flight.rs b/src/common/grpc/src/flight.rs
index 5fc115a60e..c7860f5706 100644
--- a/src/common/grpc/src/flight.rs
+++ b/src/common/grpc/src/flight.rs
@@ -37,11 +37,17 @@ use vec1::{Vec1, vec1};
 use crate::error;
 use crate::error::{DecodeFlightDataSnafu, InvalidFlightDataSnafu, Result};
 
+/// Flight metadata key used to carry flow query extensions as JSON pairs.
+pub const FLOW_EXTENSIONS_METADATA_KEY: &str = "x-greptime-flow-extensions";
+
 #[derive(Debug, Clone)]
 pub enum FlightMessage {
     Schema(SchemaRef),
     RecordBatch(DfRecordBatch),
-    AffectedRows(usize),
+    AffectedRows {
+        rows: usize,
+        metrics: Option<String>,
+    },
     Metrics(String),
 }
 
@@ -116,10 +122,12 @@ impl FlightEncoder {
                     encoded_batch.into(),
                 )
             }
-            FlightMessage::AffectedRows(rows) => {
+            FlightMessage::AffectedRows { rows, metrics } => {
                 let metadata = FlightMetadata {
                     affected_rows: Some(AffectedRows { value: rows as _ }),
-                    metrics: None,
+                    metrics: metrics.map(|s| Metrics {
+                        metrics: s.into_bytes(),
+                    }),
                 }
                 .encode_to_vec();
                 vec1![FlightData {
@@ -223,7 +231,12 @@ impl FlightDecoder {
                 let metadata = FlightMetadata::decode(flight_data.app_metadata.clone())
                     .context(DecodeFlightDataSnafu)?;
                 if let Some(AffectedRows { value }) = metadata.affected_rows {
-                    return Ok(Some(FlightMessage::AffectedRows(value as _)));
+                    return Ok(Some(FlightMessage::AffectedRows {
+                        rows: value as _,
+                        metrics: metadata
+                            .metrics
+                            .map(|m| String::from_utf8_lossy(&m.metrics).to_string()),
+                    }));
                 }
                 if let Some(Metrics { metrics }) = metadata.metrics {
                     return Ok(Some(FlightMessage::Metrics(
@@ -426,6 +439,47 @@ mod test {
         Ok(())
     }
 
+    #[test]
+    fn test_affected_rows_metrics_encode_decode() -> Result<()> {
+        let metrics = r#"{"region_watermarks":[{"region_id":42,"watermark":7}]}"#;
+        let mut encoder = FlightEncoder::default();
+        let encoded = encoder.encode(FlightMessage::AffectedRows {
+            rows: 3,
+            metrics: Some(metrics.to_string()),
+        });
+
+        assert_eq!(encoded.len(), 1);
+
+        let mut decoder = FlightDecoder::default();
+        let decoded = decoder.try_decode(encoded.first())?.unwrap();
+        let FlightMessage::AffectedRows {
+            rows,
+            metrics: decoded_metrics,
+        } = decoded
+        else {
+            unreachable!()
+        };
+        assert_eq!(rows, 3);
+        assert_eq!(decoded_metrics.as_deref(), Some(metrics));
+
+        let encoded = encoder.encode(FlightMessage::AffectedRows {
+            rows: 5,
+            metrics: None,
+        });
+        let decoded = decoder.try_decode(encoded.first())?.unwrap();
+        let FlightMessage::AffectedRows {
+            rows,
+            metrics: decoded_metrics,
+        } = decoded
+        else {
+            unreachable!()
+        };
+        assert_eq!(rows, 5);
+        assert!(decoded_metrics.is_none());
+
+        Ok(())
+    }
+
     #[test]
     fn test_flight_messages_to_recordbatches() {
         let schema = Arc::new(Schema::new(vec![Field::new("m", DataType::Int32, true)]));
@@ -548,4 +602,76 @@ mod test {
         assert_eq!(actual, expected.trim());
         Ok(())
     }
+
+    #[test]
+    fn test_affected_rows_roundtrip_through_flight_codec() {
+        // Verify the full FlightEncoder → FlightDecoder pipeline handles
+        // the new FlightMessage::AffectedRows variant with optional inline
+        // metrics without breaking the wire protocol.
+        let mut encoder = FlightEncoder::default();
+        let mut decoder = FlightDecoder::default();
+
+        // Without metrics — same wire format as old `AffectedRows(7)`.
+        let encoded = encoder.encode(FlightMessage::AffectedRows {
+            rows: 7,
+            metrics: None,
+        });
+        let decoded = decoder.try_decode(encoded.first()).unwrap().unwrap();
+        assert!(matches!(
+            decoded,
+            FlightMessage::AffectedRows {
+                rows: 7,
+                metrics: None,
+            }
+        ));
+
+        // With metrics — new capability, row count preserved.
+        let json = r#"{"region_watermarks":[{"region_id":1,"watermark":99}]}"#;
+        let encoded = encoder.encode(FlightMessage::AffectedRows {
+            rows: 42,
+            metrics: Some(json.to_string()),
+        });
+        let decoded = decoder.try_decode(encoded.first()).unwrap().unwrap();
+        assert!(matches!(
+            decoded,
+            FlightMessage::AffectedRows {
+                rows: 42,
+                metrics: Some(_),
+            }
+        ));
+    }
+
+    /// Simulates the wire output of the **old** `FlightMessage::AffectedRows(usize)`
+    /// variant and verifies that the **new** `FlightDecoder` handles it.
+    #[test]
+    fn test_old_affected_rows_format_decoded_by_new_code() {
+        use arrow_flight::FlightData;
+        use prost::bytes::Bytes as ProstBytes;
+
+        // The old encoder produced FlightData whose app_metadata is
+        // FlightMetadata { affected_rows, metrics: None }. The new
+        // `AffectedRows { rows, metrics: Option<String> }` variant with
+        // `metrics: None` produces the exact same wire bytes.
+        let old_wire_bytes = FlightData {
+            flight_descriptor: None,
+            data_header: build_none_flight_msg().into(),
+            app_metadata: FlightMetadata {
+                affected_rows: Some(AffectedRows { value: 99 }),
+                metrics: None, // old format: no metrics field
+            }
+            .encode_to_vec()
+            .into(),
+            data_body: ProstBytes::default(),
+        };
+
+        let mut decoder = FlightDecoder::default();
+        let decoded = decoder.try_decode(&old_wire_bytes).unwrap().unwrap();
+        assert!(matches!(
+            decoded,
+            FlightMessage::AffectedRows {
+                rows: 99,
+                metrics: None,
+            }
+        ));
+    }
 }
diff --git a/src/common/macro/src/admin_fn.rs b/src/common/macro/src/admin_fn.rs
index 651c083ec8..efeb77f1d1 100644
--- a/src/common/macro/src/admin_fn.rs
+++ b/src/common/macro/src/admin_fn.rs
@@ -316,14 +316,14 @@ fn build_struct(
                     .#handler
                     .as_ref()
                     .context(#snafu_type)
-                    .map_err(|e| datafusion_common::DataFusionError::Execution(format!("Handler error: {}", e.output_msg())))?;
+                    .map_err(|e| datafusion_common::DataFusionError::Execution(e.output_msg()))?;
 
                 let mut builder = store_api::storage::ConcreteDataType::#ret()
                     .create_mutable_vector(rows_num);
 
                 if columns_num == 0 {
                     let result = #fn_name(handler, query_ctx, &[]).await
-                        .map_err(|e| datafusion_common::DataFusionError::Execution(format!("Function execution error: {}", e.output_msg())))?;
+                        .map_err(|e| datafusion_common::DataFusionError::Execution(e.output_msg()))?;
 
                     builder.push_value_ref(&result.as_value_ref());
                 } else {
@@ -333,7 +333,7 @@ fn build_struct(
                             .collect();
 
                         let result = #fn_name(handler, query_ctx, &args).await
-                            .map_err(|e| datafusion_common::DataFusionError::Execution(format!("Function execution error: {}", e.output_msg())))?;
+                            .map_err(|e| datafusion_common::DataFusionError::Execution(e.output_msg()))?;
 
                         builder.push_value_ref(&result.as_value_ref());
                     }
diff --git a/src/common/memory-manager/Cargo.toml b/src/common/memory-manager/Cargo.toml
index a6be50f774..6686c98167 100644
--- a/src/common/memory-manager/Cargo.toml
+++ b/src/common/memory-manager/Cargo.toml
@@ -10,7 +10,6 @@ workspace = true
 [dependencies]
 common-error = { workspace = true }
 common-macro = { workspace = true }
-common-telemetry = { workspace = true }
 humantime = { workspace = true }
 serde = { workspace = true }
 snafu = { workspace = true }
diff --git a/src/common/memory-manager/src/guard.rs b/src/common/memory-manager/src/guard.rs
index 770b6dec24..784b72830d 100644
--- a/src/common/memory-manager/src/guard.rs
+++ b/src/common/memory-manager/src/guard.rs
@@ -14,14 +14,13 @@
 
 use std::{fmt, mem};
 
-use common_telemetry::debug;
 use snafu::ensure;
 use tokio::sync::{OwnedSemaphorePermit, TryAcquireError};
 
 use crate::error::{
     MemoryAcquireTimeoutSnafu, MemoryLimitExceededSnafu, MemorySemaphoreClosedSnafu, Result,
 };
-use crate::manager::{MemoryMetrics, MemoryQuota};
+use crate::manager::{MemoryMetrics, MemoryQuota, UnlimitedMemoryQuota};
 use crate::policy::OnExhaustedPolicy;
 
 /// Guard representing a slice of reserved memory.
@@ -30,31 +29,57 @@ pub struct MemoryGuard<M: MemoryMetrics> {
 }
 
 pub(crate) enum GuardState<M: MemoryMetrics> {
-    Unlimited,
+    Released,
+    Unlimited {
+        quota: UnlimitedMemoryQuota<M>,
+        granted_bytes: u64,
+    },
     Limited {
-        permit: OwnedSemaphorePermit,
         quota: MemoryQuota<M>,
+        permit: OwnedSemaphorePermit,
     },
 }
 
+impl<M: MemoryMetrics> GuardState<M> {
+    fn release(self) {
+        match self {
+            GuardState::Released => {}
+            GuardState::Unlimited {
+                quota,
+                granted_bytes,
+            } => {
+                quota.sub_in_use(granted_bytes);
+            }
+            GuardState::Limited { quota, permit } => {
+                quota.release_permit(permit);
+            }
+        }
+    }
+}
+
 impl<M: MemoryMetrics> MemoryGuard<M> {
-    pub(crate) fn unlimited() -> Self {
+    pub(crate) fn unlimited(quota: UnlimitedMemoryQuota<M>, bytes: u64) -> Self {
+        quota.add_in_use(bytes);
         Self {
-            state: GuardState::Unlimited,
+            state: GuardState::Unlimited {
+                quota,
+                granted_bytes: bytes,
+            },
         }
     }
 
-    pub(crate) fn limited(permit: OwnedSemaphorePermit, quota: MemoryQuota<M>) -> Self {
+    pub(crate) fn limited(quota: MemoryQuota<M>, permit: OwnedSemaphorePermit) -> Self {
         Self {
-            state: GuardState::Limited { permit, quota },
+            state: GuardState::Limited { quota, permit },
         }
     }
 
     /// Returns granted quota in bytes.
     pub fn granted_bytes(&self) -> u64 {
         match &self.state {
-            GuardState::Unlimited => 0,
-            GuardState::Limited { permit, quota } => {
+            GuardState::Released => 0,
+            GuardState::Unlimited { granted_bytes, .. } => *granted_bytes,
+            GuardState::Limited { quota, permit } => {
                 quota.permits_to_bytes(permit.num_permits() as u32)
             }
         }
@@ -68,13 +93,24 @@ impl<M: MemoryMetrics> MemoryGuard<M> {
     /// - Returns error if requested bytes would exceed the manager's total limit
     /// - Returns error if the semaphore is unexpectedly closed
     pub async fn acquire_additional(&mut self, bytes: u64) -> Result<()> {
-        match &mut self.state {
-            GuardState::Unlimited => Ok(()),
-            GuardState::Limited { permit, quota } => {
-                if bytes == 0 {
-                    return Ok(());
-                }
+        if bytes == 0 {
+            return Ok(());
+        }
 
+        match &mut self.state {
+            GuardState::Released => {
+                debug_assert!(false, "released memory guard state should not be reused");
+                Ok(())
+            }
+            GuardState::Unlimited {
+                quota,
+                granted_bytes,
+            } => {
+                quota.add_in_use(bytes);
+                *granted_bytes = granted_bytes.saturating_add(bytes);
+                Ok(())
+            }
+            GuardState::Limited { quota, permit } => {
                 let additional_permits = quota.bytes_to_permits(bytes);
                 let current_permits = permit.num_permits() as u32;
 
@@ -95,7 +131,6 @@ impl<M: MemoryMetrics> MemoryGuard<M> {
 
                 permit.merge(additional_permit);
                 quota.update_in_use_metric();
-                debug!("Acquired additional {} bytes", bytes);
                 Ok(())
             }
         }
@@ -106,13 +141,24 @@ impl<M: MemoryMetrics> MemoryGuard<M> {
     /// On success, merges the new memory into this guard and returns true.
     /// On failure, returns false and leaves this guard unchanged.
     pub fn try_acquire_additional(&mut self, bytes: u64) -> bool {
-        match &mut self.state {
-            GuardState::Unlimited => true,
-            GuardState::Limited { permit, quota } => {
-                if bytes == 0 {
-                    return true;
-                }
+        if bytes == 0 {
+            return true;
+        }
 
+        match &mut self.state {
+            GuardState::Released => {
+                debug_assert!(false, "released memory guard state should not be reused");
+                false
+            }
+            GuardState::Unlimited {
+                quota,
+                granted_bytes,
+            } => {
+                quota.add_in_use(bytes);
+                *granted_bytes = granted_bytes.saturating_add(bytes);
+                true
+            }
+            GuardState::Limited { quota, permit } => {
                 let additional_permits = quota.bytes_to_permits(bytes);
 
                 match quota
@@ -123,11 +169,10 @@ impl<M: MemoryMetrics> MemoryGuard<M> {
                     Ok(additional_permit) => {
                         permit.merge(additional_permit);
                         quota.update_in_use_metric();
-                        debug!("Acquired additional {} bytes", bytes);
                         true
                     }
                     Err(TryAcquireError::NoPermits) | Err(TryAcquireError::Closed) => {
-                        quota.metrics.inc_rejected("try_acquire_additional");
+                        quota.metrics.inc_exhausted("try_acquire_additional");
                         false
                     }
                 }
@@ -168,7 +213,8 @@ impl<M: MemoryMetrics> MemoryGuard<M> {
                     MemoryLimitExceededSnafu {
                         requested_bytes: bytes,
                         limit_bytes: match &self.state {
-                            GuardState::Unlimited => 0, // unreachable: unlimited mode always succeeds
+                            GuardState::Released => 0,
+                            GuardState::Unlimited { .. } => 0, // unreachable: unlimited mode always succeeds
                             GuardState::Limited { quota, .. } => {
                                 quota.permits_to_bytes(quota.limit_permits)
                             }
@@ -184,22 +230,30 @@ impl<M: MemoryMetrics> MemoryGuard<M> {
     ///
     /// Returns true if the release succeeds or is a no-op; false if the request exceeds granted.
     pub fn release_partial(&mut self, bytes: u64) -> bool {
+        if bytes == 0 {
+            return true;
+        }
+
         match &mut self.state {
-            GuardState::Unlimited => true,
-            GuardState::Limited { permit, quota } => {
-                if bytes == 0 {
-                    return true;
+            GuardState::Released => true,
+            GuardState::Unlimited {
+                quota,
+                granted_bytes,
+            } => {
+                if bytes > *granted_bytes {
+                    return false;
                 }
 
+                quota.sub_in_use(bytes);
+                *granted_bytes = granted_bytes.saturating_sub(bytes);
+                true
+            }
+            GuardState::Limited { quota, permit } => {
                 let release_permits = quota.bytes_to_permits(bytes);
 
                 match permit.split(release_permits as usize) {
                     Some(released_permit) => {
-                        let released_bytes =
-                            quota.permits_to_bytes(released_permit.num_permits() as u32);
-                        drop(released_permit);
-                        quota.update_in_use_metric();
-                        debug!("Released {} bytes from memory guard", released_bytes);
+                        quota.release_permit(released_permit);
                         true
                     }
                     None => false,
@@ -211,14 +265,7 @@ impl<M: MemoryMetrics> MemoryGuard<M> {
 
 impl<M: MemoryMetrics> Drop for MemoryGuard<M> {
     fn drop(&mut self) {
-        if let GuardState::Limited { permit, quota } =
-            mem::replace(&mut self.state, GuardState::Unlimited)
-        {
-            let bytes = quota.permits_to_bytes(permit.num_permits() as u32);
-            drop(permit);
-            quota.update_in_use_metric();
-            debug!("Released memory: {} bytes", bytes);
-        }
+        mem::replace(&mut self.state, GuardState::Released).release();
     }
 }
 
diff --git a/src/common/memory-manager/src/lib.rs b/src/common/memory-manager/src/lib.rs
index b1d858ef89..983c6ca524 100644
--- a/src/common/memory-manager/src/lib.rs
+++ b/src/common/memory-manager/src/lib.rs
@@ -45,5 +45,5 @@ impl MemoryMetrics for NoOpMetrics {
     fn set_in_use(&self, _: i64) {}
 
     #[inline(always)]
-    fn inc_rejected(&self, _: &str) {}
+    fn inc_exhausted(&self, _: &str) {}
 }
diff --git a/src/common/memory-manager/src/manager.rs b/src/common/memory-manager/src/manager.rs
index 50360d2a31..282f51e315 100644
--- a/src/common/memory-manager/src/manager.rs
+++ b/src/common/memory-manager/src/manager.rs
@@ -13,9 +13,10 @@
 // limitations under the License.
 
 use std::sync::Arc;
+use std::sync::atomic::{AtomicU64, Ordering};
 
 use snafu::ensure;
-use tokio::sync::{Semaphore, TryAcquireError};
+use tokio::sync::{OwnedSemaphorePermit, Semaphore, TryAcquireError};
 
 use crate::error::{
     MemoryAcquireTimeoutSnafu, MemoryLimitExceededSnafu, MemorySemaphoreClosedSnafu, Result,
@@ -28,13 +29,14 @@ use crate::policy::OnExhaustedPolicy;
 pub trait MemoryMetrics: Clone + Send + Sync + 'static {
     fn set_limit(&self, bytes: i64);
     fn set_in_use(&self, bytes: i64);
-    fn inc_rejected(&self, reason: &str);
+    /// Record that immediate memory acquisition failed due to exhausted quota.
+    fn inc_exhausted(&self, reason: &str);
 }
 
 /// Generic memory manager for quota-controlled operations.
 #[derive(Clone)]
 pub struct MemoryManager<M: MemoryMetrics> {
-    quota: Option<MemoryQuota<M>>,
+    quota: MemoryQuotaState<M>,
 }
 
 impl<M: MemoryMetrics + Default> Default for MemoryManager<M> {
@@ -51,6 +53,18 @@ pub(crate) struct MemoryQuota<M: MemoryMetrics> {
     pub(crate) metrics: M,
 }
 
+#[derive(Clone)]
+pub(crate) struct UnlimitedMemoryQuota<M: MemoryMetrics> {
+    pub(crate) current_bytes: Arc<AtomicU64>,
+    pub(crate) metrics: M,
+}
+
+#[derive(Clone)]
+pub(crate) enum MemoryQuotaState<M: MemoryMetrics> {
+    Unlimited(UnlimitedMemoryQuota<M>),
+    Limited(MemoryQuota<M>),
+}
+
 impl<M: MemoryMetrics> MemoryManager<M> {
     /// Creates a new memory manager with the given limit in bytes.
     /// `limit_bytes = 0` disables the limit.
@@ -62,7 +76,12 @@ impl<M: MemoryMetrics> MemoryManager<M> {
     pub fn with_granularity(limit_bytes: u64, granularity: PermitGranularity, metrics: M) -> Self {
         if limit_bytes == 0 {
             metrics.set_limit(0);
-            return Self { quota: None };
+            return Self {
+                quota: MemoryQuotaState::Unlimited(UnlimitedMemoryQuota {
+                    current_bytes: Arc::new(AtomicU64::new(0)),
+                    metrics,
+                }),
+            };
         }
 
         let limit_permits = granularity.bytes_to_permits(limit_bytes);
@@ -70,7 +89,7 @@ impl<M: MemoryMetrics> MemoryManager<M> {
         metrics.set_limit(limit_aligned_bytes as i64);
 
         Self {
-            quota: Some(MemoryQuota {
+            quota: MemoryQuotaState::Limited(MemoryQuota {
                 semaphore: Arc::new(Semaphore::new(limit_permits as usize)),
                 limit_permits,
                 granularity,
@@ -81,26 +100,30 @@ impl<M: MemoryMetrics> MemoryManager<M> {
 
     /// Returns the configured limit in bytes (0 if unlimited).
     pub fn limit_bytes(&self) -> u64 {
-        self.quota
-            .as_ref()
-            .map(|quota| quota.permits_to_bytes(quota.limit_permits))
-            .unwrap_or(0)
+        match &self.quota {
+            MemoryQuotaState::Unlimited(_) => 0,
+            MemoryQuotaState::Limited(quota) => quota.permits_to_bytes(quota.limit_permits),
+        }
     }
 
     /// Returns currently used bytes.
     pub fn used_bytes(&self) -> u64 {
-        self.quota
-            .as_ref()
-            .map(|quota| quota.permits_to_bytes(quota.used_permits()))
-            .unwrap_or(0)
+        match &self.quota {
+            MemoryQuotaState::Unlimited(quota) => quota.current_bytes.load(Ordering::Acquire),
+            MemoryQuotaState::Limited(quota) => quota.permits_to_bytes(quota.used_permits()),
+        }
     }
 
     /// Returns available bytes.
+    ///
+    /// Unlimited managers report `u64::MAX`.
     pub fn available_bytes(&self) -> u64 {
-        self.quota
-            .as_ref()
-            .map(|quota| quota.permits_to_bytes(quota.available_permits_clamped()))
-            .unwrap_or(0)
+        match &self.quota {
+            MemoryQuotaState::Unlimited(_) => u64::MAX,
+            MemoryQuotaState::Limited(quota) => {
+                quota.permits_to_bytes(quota.available_permits_clamped())
+            }
+        }
     }
 
     /// Acquires memory, waiting if necessary until enough is available.
@@ -110,8 +133,8 @@ impl<M: MemoryMetrics> MemoryManager<M> {
     /// - Returns error if the semaphore is unexpectedly closed
     pub async fn acquire(&self, bytes: u64) -> Result<MemoryGuard<M>> {
         match &self.quota {
-            None => Ok(MemoryGuard::unlimited()),
-            Some(quota) => {
+            MemoryQuotaState::Unlimited(quota) => Ok(MemoryGuard::unlimited(quota.clone(), bytes)),
+            MemoryQuotaState::Limited(quota) => {
                 let permits = quota.bytes_to_permits(bytes);
 
                 ensure!(
@@ -129,7 +152,7 @@ impl<M: MemoryMetrics> MemoryManager<M> {
                     .await
                     .map_err(|_| MemorySemaphoreClosedSnafu.build())?;
                 quota.update_in_use_metric();
-                Ok(MemoryGuard::limited(permit, quota.clone()))
+                Ok(MemoryGuard::limited(quota.clone(), permit))
             }
         }
     }
@@ -137,17 +160,19 @@ impl<M: MemoryMetrics> MemoryManager<M> {
     /// Tries to acquire memory. Returns Some(guard) on success, None if insufficient.
     pub fn try_acquire(&self, bytes: u64) -> Option<MemoryGuard<M>> {
         match &self.quota {
-            None => Some(MemoryGuard::unlimited()),
-            Some(quota) => {
+            MemoryQuotaState::Unlimited(quota) => {
+                Some(MemoryGuard::unlimited(quota.clone(), bytes))
+            }
+            MemoryQuotaState::Limited(quota) => {
                 let permits = quota.bytes_to_permits(bytes);
 
                 match quota.semaphore.clone().try_acquire_many_owned(permits) {
                     Ok(permit) => {
                         quota.update_in_use_metric();
-                        Some(MemoryGuard::limited(permit, quota.clone()))
+                        Some(MemoryGuard::limited(quota.clone(), permit))
                     }
                     Err(TryAcquireError::NoPermits) | Err(TryAcquireError::Closed) => {
-                        quota.metrics.inc_rejected("try_acquire");
+                        quota.metrics.inc_exhausted("try_acquire");
                         None
                     }
                 }
@@ -219,4 +244,49 @@ impl<M: MemoryMetrics> MemoryQuota<M> {
         let bytes = self.permits_to_bytes(self.used_permits());
         self.metrics.set_in_use(bytes as i64);
     }
+
+    pub(crate) fn release_permit(&self, permit: OwnedSemaphorePermit) {
+        drop(permit);
+        self.update_in_use_metric();
+    }
+}
+
+impl<M: MemoryMetrics> UnlimitedMemoryQuota<M> {
+    pub(crate) fn add_in_use(&self, bytes: u64) {
+        if bytes == 0 {
+            return;
+        }
+
+        let previous = self
+            .current_bytes
+            .fetch_update(Ordering::AcqRel, Ordering::Acquire, |current| {
+                Some(current.saturating_add(bytes))
+            })
+            .unwrap();
+        let new_total = previous.saturating_add(bytes);
+        debug_assert!(
+            new_total >= previous,
+            "unlimited memory usage counter overflowed"
+        );
+        self.metrics.set_in_use(new_total as i64);
+    }
+
+    pub(crate) fn sub_in_use(&self, bytes: u64) {
+        if bytes == 0 {
+            return;
+        }
+
+        let previous = self
+            .current_bytes
+            .fetch_update(Ordering::AcqRel, Ordering::Acquire, |current| {
+                Some(current.saturating_sub(bytes))
+            })
+            .unwrap();
+        debug_assert!(
+            previous >= bytes,
+            "unlimited memory usage counter underflowed: current={previous}, release={bytes}"
+        );
+        let new_total = previous.saturating_sub(bytes);
+        self.metrics.set_in_use(new_total as i64);
+    }
 }
diff --git a/src/common/memory-manager/src/tests.rs b/src/common/memory-manager/src/tests.rs
index 886eef9dac..fe02703f0b 100644
--- a/src/common/memory-manager/src/tests.rs
+++ b/src/common/memory-manager/src/tests.rs
@@ -24,7 +24,9 @@ fn test_try_acquire_unlimited() {
     let manager = MemoryManager::new(0, NoOpMetrics);
     let guard = manager.try_acquire(10 * PERMIT_GRANULARITY_BYTES).unwrap();
     assert_eq!(manager.limit_bytes(), 0);
-    assert_eq!(guard.granted_bytes(), 0);
+    assert_eq!(manager.available_bytes(), u64::MAX);
+    assert_eq!(guard.granted_bytes(), 10 * PERMIT_GRANULARITY_BYTES);
+    assert_eq!(manager.used_bytes(), 10 * PERMIT_GRANULARITY_BYTES);
 }
 
 #[test]
@@ -136,7 +138,10 @@ fn test_request_additional_unlimited() {
 
     // Should always succeed with unlimited manager
     assert!(guard.try_acquire_additional(100 * PERMIT_GRANULARITY_BYTES));
-    assert_eq!(guard.granted_bytes(), 0);
+    assert_eq!(guard.granted_bytes(), 105 * PERMIT_GRANULARITY_BYTES);
+    assert_eq!(manager.used_bytes(), 105 * PERMIT_GRANULARITY_BYTES);
+
+    drop(guard);
     assert_eq!(manager.used_bytes(), 0);
 }
 
@@ -187,9 +192,10 @@ fn test_early_release_partial_unlimited() {
     let manager = MemoryManager::new(0, NoOpMetrics);
     let mut guard = manager.try_acquire(100 * PERMIT_GRANULARITY_BYTES).unwrap();
 
-    // Unlimited guard - release should succeed (no-op)
+    // Unlimited guard should track and release exact bytes.
     assert!(guard.release_partial(50 * PERMIT_GRANULARITY_BYTES));
-    assert_eq!(guard.granted_bytes(), 0);
+    assert_eq!(guard.granted_bytes(), 50 * PERMIT_GRANULARITY_BYTES);
+    assert_eq!(manager.used_bytes(), 50 * PERMIT_GRANULARITY_BYTES);
 }
 
 #[test]
@@ -406,6 +412,6 @@ async fn test_acquire_additional_unlimited() {
         .acquire_additional(1000 * PERMIT_GRANULARITY_BYTES)
         .await
         .unwrap();
-    assert_eq!(guard.granted_bytes(), 0);
-    assert_eq!(manager.used_bytes(), 0);
+    assert_eq!(guard.granted_bytes(), 1000 * PERMIT_GRANULARITY_BYTES);
+    assert_eq!(manager.used_bytes(), 1000 * PERMIT_GRANULARITY_BYTES);
 }
diff --git a/src/common/meta/Cargo.toml b/src/common/meta/Cargo.toml
index ec000c710d..5134c40a80 100644
--- a/src/common/meta/Cargo.toml
+++ b/src/common/meta/Cargo.toml
@@ -8,7 +8,6 @@ license.workspace = true
 testing = []
 pg_kvbackend = [
     "dep:tokio-postgres",
-    "dep:backon",
     "dep:deadpool-postgres",
     "dep:deadpool",
     "dep:tokio-postgres-rustls",
@@ -16,7 +15,7 @@ pg_kvbackend = [
     "dep:rustls-native-certs",
     "dep:rustls",
 ]
-mysql_kvbackend = ["dep:sqlx", "dep:backon"]
+mysql_kvbackend = ["dep:sqlx", "dep:rustls"]
 enterprise = ["prost-types"]
 
 [lints]
@@ -28,7 +27,7 @@ api.workspace = true
 async-recursion = "1.0"
 async-stream.workspace = true
 async-trait.workspace = true
-backon = { workspace = true, optional = true }
+backon.workspace = true
 base64.workspace = true
 bytes.workspace = true
 chrono.workspace = true
@@ -38,7 +37,6 @@ common-error.workspace = true
 common-grpc-expr.workspace = true
 common-macro.workspace = true
 common-procedure.workspace = true
-common-procedure-test.workspace = true
 common-query.workspace = true
 common-recordbatch.workspace = true
 common-runtime.workspace = true
@@ -61,7 +59,7 @@ hex.workspace = true
 humantime-serde.workspace = true
 itertools.workspace = true
 lazy_static.workspace = true
-moka.workspace = true
+moka = { workspace = true, features = ["future"] }
 object-store.workspace = true
 prometheus.workspace = true
 prost.workspace = true
@@ -69,7 +67,12 @@ prost-types = { workspace = true, optional = true }
 rand.workspace = true
 regex.workspace = true
 rskafka.workspace = true
-rustls = { workspace = true, default-features = false, features = ["ring", "logging", "std", "tls12"], optional = true }
+rustls = { workspace = true, default-features = false, features = [
+    "aws_lc_rs",
+    "logging",
+    "std",
+    "tls12",
+], optional = true }
 rustls-native-certs = { version = "0.7", optional = true }
 rustls-pemfile = { version = "2.0", optional = true }
 serde.workspace = true
@@ -93,6 +96,7 @@ typetag.workspace = true
 [dev-dependencies]
 chrono.workspace = true
 common-procedure = { workspace = true, features = ["testing"] }
+common-procedure-test.workspace = true
 common-test-util.workspace = true
 common-wal = { workspace = true, features = ["testing"] }
 datatypes.workspace = true
diff --git a/src/common/meta/src/cache/container.rs b/src/common/meta/src/cache/container.rs
index 0510476d15..e3a3e13a76 100644
--- a/src/common/meta/src/cache/container.rs
+++ b/src/common/meta/src/cache/container.rs
@@ -15,10 +15,14 @@
 use std::borrow::Borrow;
 use std::hash::Hash;
 use std::sync::Arc;
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::time::Duration;
 
-use futures::future::{BoxFuture, join_all};
+use backon::{BackoffBuilder, ExponentialBuilder};
+use futures::future::BoxFuture;
 use moka::future::Cache;
 use snafu::{OptionExt, ResultExt};
+use tokio::time::sleep;
 
 use crate::cache_invalidator::{CacheInvalidator, Context};
 use crate::error::{self, Error, Result};
@@ -29,12 +33,29 @@ use crate::metrics;
 pub type TokenFilter<CacheToken> = Box<dyn Fn(&CacheToken) -> bool + Send + Sync>;
 
 /// Invalidates cached values by [CacheToken]s.
-pub type Invalidator<K, V, CacheToken> =
-    Box<dyn for<'a> Fn(&'a Cache<K, V>, &'a CacheToken) -> BoxFuture<'a, Result<()>> + Send + Sync>;
+pub type Invalidator<K, V, CacheToken> = Box<
+    dyn for<'a> Fn(&'a Cache<K, V>, &'a [&CacheToken]) -> BoxFuture<'a, Result<()>> + Send + Sync,
+>;
 
 /// Initializes value (i.e., fetches from remote).
 pub type Initializer<K, V> = Arc<dyn Fn(&'_ K) -> BoxFuture<'_, Result<Option<V>>> + Send + Sync>;
 
+#[derive(Debug, Clone, Copy)]
+/// Initialization strategy for cache-miss loading.
+///
+/// This strategy is selected when building [CacheContainer] and remains immutable
+/// for the lifetime of the container instance.
+pub enum InitStrategy {
+    /// Fast path: load once without version conflict retry.
+    ///
+    /// Under concurrent invalidation, callers may observe stale/dirty value.
+    Unchecked,
+    /// Strict path: retry load when version changes during initialization.
+    ///
+    /// This avoids returning dirty value under invalidate/load races.
+    VersionChecked,
+}
+
 /// [CacheContainer] provides ability to:
 /// - Cache value loaded by [Initializer].
 /// - Invalidate caches by [Invalidator].
@@ -44,6 +65,16 @@ pub struct CacheContainer<K, V, CacheToken> {
     invalidator: Invalidator<K, V, CacheToken>,
     initializer: Initializer<K, V>,
     token_filter: fn(&CacheToken) -> bool,
+    version: Arc<AtomicUsize>,
+    init_strategy: InitStrategy,
+}
+
+fn latest_get_backoff() -> impl Iterator<Item = Duration> {
+    ExponentialBuilder::default()
+        .with_min_delay(Duration::from_millis(10))
+        .with_max_delay(Duration::from_millis(100))
+        .with_max_times(3)
+        .build()
 }
 
 impl<K, V, CacheToken> CacheContainer<K, V, CacheToken>
@@ -52,13 +83,37 @@ where
     V: Send + Sync,
     CacheToken: Send + Sync,
 {
-    /// Constructs an [CacheContainer].
+    /// Constructs an [CacheContainer] with [InitStrategy::Unchecked].
+    ///
+    /// This keeps the historical behavior and can return stale/dirty value under
+    /// concurrent invalidation.
     pub fn new(
         name: String,
         cache: Cache<K, V>,
         invalidator: Invalidator<K, V, CacheToken>,
         initializer: Initializer<K, V>,
         token_filter: fn(&CacheToken) -> bool,
+    ) -> Self {
+        Self::with_strategy(
+            name,
+            cache,
+            invalidator,
+            initializer,
+            token_filter,
+            InitStrategy::Unchecked,
+        )
+    }
+
+    /// Constructs an [CacheContainer] with explicit [InitStrategy].
+    ///
+    /// The strategy is fixed at construction time and cannot be changed later.
+    pub fn with_strategy(
+        name: String,
+        cache: Cache<K, V>,
+        invalidator: Invalidator<K, V, CacheToken>,
+        initializer: Initializer<K, V>,
+        token_filter: fn(&CacheToken) -> bool,
+        init_strategy: InitStrategy,
     ) -> Self {
         Self {
             name,
@@ -66,6 +121,8 @@ where
             invalidator,
             initializer,
             token_filter,
+            version: Arc::new(AtomicUsize::new(0)),
+            init_strategy,
         }
     }
 
@@ -75,6 +132,67 @@ where
     }
 }
 
+impl<K, V, CacheToken> CacheContainer<K, V, CacheToken> {
+    fn inc_version(&self) {
+        self.version.fetch_add(1, Ordering::Relaxed);
+    }
+}
+
+async fn init<'a, K, V>(init: Initializer<K, V>, key: K, cache_name: &'a str) -> Result<V>
+where
+    K: Send + Sync + 'a,
+    V: Send + 'a,
+{
+    metrics::CACHE_CONTAINER_CACHE_MISS
+        .with_label_values(&[cache_name])
+        .inc();
+    let _timer = metrics::CACHE_CONTAINER_LOAD_CACHE
+        .with_label_values(&[cache_name])
+        .start_timer();
+    init(&key)
+        .await
+        .transpose()
+        .context(error::ValueNotExistSnafu)?
+}
+
+async fn init_with_retry<'a, K, V>(
+    init: Initializer<K, V>,
+    key: K,
+    mut backoff: impl Iterator<Item = Duration> + 'a,
+    version: Arc<AtomicUsize>,
+    cache_name: &'a str,
+) -> Result<V>
+where
+    K: Send + Sync + 'a,
+    V: Send + 'a,
+{
+    let mut attempts = 1usize;
+    loop {
+        let pre_version = version.load(Ordering::Relaxed);
+        metrics::CACHE_CONTAINER_CACHE_MISS
+            .with_label_values(&[cache_name])
+            .inc();
+        let _timer = metrics::CACHE_CONTAINER_LOAD_CACHE
+            .with_label_values(&[cache_name])
+            .start_timer();
+        let value = init(&key)
+            .await
+            .transpose()
+            .context(error::ValueNotExistSnafu)??;
+
+        if pre_version == version.load(Ordering::Relaxed) {
+            return Ok(value);
+        }
+
+        if let Some(duration) = backoff.next() {
+            sleep(duration).await;
+            attempts += 1;
+        } else {
+            return error::GetLatestCacheRetryExceededSnafu { attempts }.fail();
+        }
+    }
+}
+
 #[async_trait::async_trait]
 impl<K, V> CacheInvalidator for CacheContainer<K, V, CacheIdent>
 where
@@ -82,14 +200,15 @@ where
     V: Send + Sync,
 {
     async fn invalidate(&self, _ctx: &Context, caches: &[CacheIdent]) -> Result<()> {
-        let tasks = caches
+        let idents = caches
             .iter()
             .filter(|token| (self.token_filter)(token))
-            .map(|token| (self.invalidator)(&self.cache, token));
-        join_all(tasks)
-            .await
-            .into_iter()
-            .collect::<Result<Vec<_>>>()?;
+            .collect::<Vec<_>>();
+        if !idents.is_empty() {
+            self.inc_version();
+            (self.invalidator)(&self.cache, &idents).await?;
+        }
+
         Ok(())
     }
 }
@@ -99,27 +218,39 @@ where
     K: Copy + Hash + Eq + Send + Sync + 'static,
     V: Clone + Send + Sync + 'static,
 {
-    /// Returns a _clone_ of the value corresponding to the key.
+    /// Returns a value from cache for copyable keys.
+    ///
+    /// With [InitStrategy::Unchecked], this method prioritizes latency and may
+    /// return stale/dirty value. With [InitStrategy::VersionChecked], this method
+    /// retries initialization on version change and avoids dirty returns.
     pub async fn get(&self, key: K) -> Result<Option<V>> {
         metrics::CACHE_CONTAINER_CACHE_GET
             .with_label_values(&[&self.name])
             .inc();
-        let moved_init = self.initializer.clone();
-        let moved_key = key;
-        let init = async move {
-            metrics::CACHE_CONTAINER_CACHE_MISS
-                .with_label_values(&[&self.name])
-                .inc();
-            let _timer = metrics::CACHE_CONTAINER_LOAD_CACHE
-                .with_label_values(&[&self.name])
-                .start_timer();
-            moved_init(&moved_key)
-                .await
-                .transpose()
-                .context(error::ValueNotExistSnafu)?
+
+        let result = match self.init_strategy {
+            InitStrategy::Unchecked => {
+                self.cache
+                    .try_get_with(key, init(self.initializer.clone(), key, &self.name))
+                    .await
+            }
+            InitStrategy::VersionChecked => {
+                self.cache
+                    .try_get_with(
+                        key,
+                        init_with_retry(
+                            self.initializer.clone(),
+                            key,
+                            latest_get_backoff(),
+                            self.version.clone(),
+                            &self.name,
+                        ),
+                    )
+                    .await
+            }
         };
 
-        match self.cache.try_get_with(key, init).await {
+        match result {
             Ok(value) => Ok(Some(value)),
             Err(err) => match err.as_ref() {
                 Error::ValueNotExist { .. } => Ok(None),
@@ -136,14 +267,15 @@ where
 {
     /// Invalidates cache by [CacheToken].
     pub async fn invalidate(&self, caches: &[CacheToken]) -> Result<()> {
-        let tasks = caches
+        let idents = caches
             .iter()
             .filter(|token| (self.token_filter)(token))
-            .map(|token| (self.invalidator)(&self.cache, token));
-        join_all(tasks)
-            .await
-            .into_iter()
-            .collect::<Result<Vec<_>>>()?;
+            .collect::<Vec<_>>();
+        if !idents.is_empty() {
+            self.inc_version();
+            (self.invalidator)(&self.cache, &idents).await?;
+        }
+
         Ok(())
     }
 
@@ -156,7 +288,11 @@ where
         self.cache.contains_key(key)
     }
 
-    /// Returns a _clone_ of the value corresponding to the key.
+    /// Returns a value from cache by key reference.
+    ///
+    /// With [InitStrategy::Unchecked], this method prioritizes latency and may
+    /// return stale/dirty value. With [InitStrategy::VersionChecked], this method
+    /// retries initialization on version change and avoids dirty returns.
     pub async fn get_by_ref<Q>(&self, key: &Q) -> Result<Option<V>>
     where
         K: Borrow<Q>,
@@ -165,24 +301,32 @@ where
         metrics::CACHE_CONTAINER_CACHE_GET
             .with_label_values(&[&self.name])
             .inc();
-        let moved_init = self.initializer.clone();
-        let moved_key = key.to_owned();
-
-        let init = async move {
-            metrics::CACHE_CONTAINER_CACHE_MISS
-                .with_label_values(&[&self.name])
-                .inc();
-            let _timer = metrics::CACHE_CONTAINER_LOAD_CACHE
-                .with_label_values(&[&self.name])
-                .start_timer();
-
-            moved_init(&moved_key)
-                .await
-                .transpose()
-                .context(error::ValueNotExistSnafu)?
+        let result = match self.init_strategy {
+            InitStrategy::Unchecked => {
+                self.cache
+                    .try_get_with_by_ref(
+                        key,
+                        init(self.initializer.clone(), key.to_owned(), &self.name),
+                    )
+                    .await
+            }
+            InitStrategy::VersionChecked => {
+                self.cache
+                    .try_get_with_by_ref(
+                        key,
+                        init_with_retry(
+                            self.initializer.clone(),
+                            key.to_owned(),
+                            latest_get_backoff(),
+                            self.version.clone(),
+                            &self.name,
+                        ),
+                    )
+                    .await
+            }
         };
 
-        match self.cache.try_get_with_by_ref(key, init).await {
+        match result {
             Ok(value) => Ok(Some(value)),
             Err(err) => match err.as_ref() {
                 Error::ValueNotExist { .. } => Ok(None),
@@ -296,9 +440,11 @@ mod tests {
             moved_counter.fetch_add(1, Ordering::Relaxed);
             Box::pin(async { Ok(Some("hi".to_string())) })
         });
-        let invalidator: Invalidator<String, String, String> = Box::new(|cache, key| {
+        let invalidator: Invalidator<String, String, String> = Box::new(|cache, keys| {
             Box::pin(async move {
-                cache.invalidate(key).await;
+                for key in keys {
+                    cache.invalidate(*key).await;
+                }
                 Ok(())
             })
         });
@@ -323,4 +469,46 @@ mod tests {
         assert_eq!(value, "hi");
         assert_eq!(counter.load(Ordering::Relaxed), 2);
     }
+
+    #[tokio::test(flavor = "multi_thread")]
+    async fn test_get_by_ref_returns_fresh_value_after_invalidate() {
+        let cache: Cache<String, String> = CacheBuilder::new(128).build();
+        let counter = Arc::new(AtomicI32::new(0));
+        let moved_counter = counter.clone();
+        let init: Initializer<String, String> = Arc::new(move |_| {
+            let counter = moved_counter.clone();
+            Box::pin(async move {
+                let n = counter.fetch_add(1, Ordering::Relaxed) + 1;
+                sleep(Duration::from_millis(100)).await;
+                Ok(Some(format!("v{n}")))
+            })
+        });
+        let invalidator: Invalidator<String, String, String> = Box::new(|cache, keys| {
+            Box::pin(async move {
+                for key in keys {
+                    cache.invalidate(*key).await;
+                }
+                Ok(())
+            })
+        });
+
+        let adv_cache = Arc::new(CacheContainer::with_strategy(
+            "test".to_string(),
+            cache,
+            invalidator,
+            init,
+            always_true_filter,
+            InitStrategy::VersionChecked,
+        ));
+
+        let moved_cache = adv_cache.clone();
+        let get_task = tokio::spawn(async move { moved_cache.get_by_ref("foo").await });
+
+        sleep(Duration::from_millis(50)).await;
+        adv_cache.invalidate(&["foo".to_string()]).await.unwrap();
+
+        let value = get_task.await.unwrap().unwrap().unwrap();
+        assert_eq!(value, "v2");
+        assert_eq!(counter.load(Ordering::Relaxed), 2);
+    }
 }
diff --git a/src/common/meta/src/cache/flow/table_flownode.rs b/src/common/meta/src/cache/flow/table_flownode.rs
index a7777f3361..ebe3664202 100644
--- a/src/common/meta/src/cache/flow/table_flownode.rs
+++ b/src/common/meta/src/cache/flow/table_flownode.rs
@@ -170,20 +170,22 @@ async fn handle_drop_flow(
 
 fn invalidator<'a>(
     cache: &'a Cache<TableId, FlownodeFlowSet>,
-    ident: &'a CacheIdent,
+    idents: &'a [&CacheIdent],
 ) -> BoxFuture<'a, Result<()>> {
     Box::pin(async move {
-        match ident {
-            CacheIdent::CreateFlow(create_flow) => handle_create_flow(cache, create_flow).await,
-            CacheIdent::DropFlow(drop_flow) => handle_drop_flow(cache, drop_flow).await,
-            CacheIdent::FlowNodeAddressChange(node_id) => {
-                info!(
-                    "Invalidate flow node cache for node_id in table_flownode: {}",
-                    node_id
-                );
-                cache.invalidate_all();
+        for ident in idents {
+            match ident {
+                CacheIdent::CreateFlow(create_flow) => handle_create_flow(cache, create_flow).await,
+                CacheIdent::DropFlow(drop_flow) => handle_drop_flow(cache, drop_flow).await,
+                CacheIdent::FlowNodeAddressChange(node_id) => {
+                    info!(
+                        "Invalidate flow node cache for node_id in table_flownode: {}",
+                        node_id
+                    );
+                    cache.invalidate_all();
+                }
+                _ => {}
             }
-            _ => {}
         }
         Ok(())
     })
diff --git a/src/common/meta/src/cache/table/schema.rs b/src/common/meta/src/cache/table/schema.rs
index bcf81d4fe6..bd9e8e6dc1 100644
--- a/src/common/meta/src/cache/table/schema.rs
+++ b/src/common/meta/src/cache/table/schema.rs
@@ -58,11 +58,13 @@ fn init_factory(schema_manager: SchemaManager) -> Initializer<SchemaName, Arc<Sc
 
 fn invalidator<'a>(
     cache: &'a Cache<SchemaName, Arc<SchemaNameValue>>,
-    ident: &'a CacheIdent,
+    idents: &'a [&CacheIdent],
 ) -> BoxFuture<'a, crate::error::Result<()>> {
     Box::pin(async move {
-        if let CacheIdent::SchemaName(schema_name) = ident {
-            cache.invalidate(schema_name).await
+        for ident in idents {
+            if let CacheIdent::SchemaName(schema_name) = ident {
+                cache.invalidate(schema_name).await
+            }
         }
         Ok(())
     })
diff --git a/src/common/meta/src/cache/table/table_info.rs b/src/common/meta/src/cache/table/table_info.rs
index b853d908e8..97af5bcdb7 100644
--- a/src/common/meta/src/cache/table/table_info.rs
+++ b/src/common/meta/src/cache/table/table_info.rs
@@ -61,11 +61,13 @@ fn init_factory(table_info_manager: TableInfoManagerRef) -> Initializer<TableId,
 
 fn invalidator<'a>(
     cache: &'a Cache<TableId, Arc<TableInfo>>,
-    ident: &'a CacheIdent,
+    idents: &'a [&CacheIdent],
 ) -> BoxFuture<'a, Result<()>> {
     Box::pin(async move {
-        if let CacheIdent::TableId(table_id) = ident {
-            cache.invalidate(table_id).await
+        for ident in idents {
+            if let CacheIdent::TableId(table_id) = ident {
+                cache.invalidate(table_id).await
+            }
         }
         Ok(())
     })
diff --git a/src/common/meta/src/cache/table/table_name.rs b/src/common/meta/src/cache/table/table_name.rs
index 540da5e5f4..927a5b3480 100644
--- a/src/common/meta/src/cache/table/table_name.rs
+++ b/src/common/meta/src/cache/table/table_name.rs
@@ -71,11 +71,13 @@ fn init_factory(table_name_manager: TableNameManagerRef) -> Initializer<TableNam
 
 fn invalidator<'a>(
     cache: &'a Cache<TableName, TableId>,
-    ident: &'a CacheIdent,
+    idents: &'a [&CacheIdent],
 ) -> BoxFuture<'a, Result<()>> {
     Box::pin(async move {
-        if let CacheIdent::TableName(table_name) = ident {
-            cache.invalidate(table_name).await
+        for ident in idents {
+            if let CacheIdent::TableName(table_name) = ident {
+                cache.invalidate(table_name).await
+            }
         }
         Ok(())
     })
diff --git a/src/common/meta/src/cache/table/table_route.rs b/src/common/meta/src/cache/table/table_route.rs
index 47abdaa728..be820b0c52 100644
--- a/src/common/meta/src/cache/table/table_route.rs
+++ b/src/common/meta/src/cache/table/table_route.rs
@@ -19,6 +19,7 @@ use moka::future::Cache;
 use snafu::OptionExt;
 use store_api::storage::TableId;
 
+use crate::cache::container::InitStrategy;
 use crate::cache::{CacheContainer, Initializer};
 use crate::error;
 use crate::error::Result;
@@ -65,7 +66,14 @@ pub fn new_table_route_cache(
     let table_info_manager = Arc::new(TableRouteManager::new(kv_backend));
     let init = init_factory(table_info_manager);
 
-    CacheContainer::new(name, cache, Box::new(invalidator), init, filter)
+    CacheContainer::with_strategy(
+        name,
+        cache,
+        Box::new(invalidator),
+        init,
+        filter,
+        InitStrategy::VersionChecked,
+    )
 }
 
 fn init_factory(
@@ -92,11 +100,13 @@ fn init_factory(
 
 fn invalidator<'a>(
     cache: &'a Cache<TableId, Arc<TableRoute>>,
-    ident: &'a CacheIdent,
+    idents: &'a [&CacheIdent],
 ) -> BoxFuture<'a, Result<()>> {
     Box::pin(async move {
-        if let CacheIdent::TableId(table_id) = ident {
-            cache.invalidate(table_id).await
+        for ident in idents {
+            if let CacheIdent::TableId(table_id) = ident {
+                cache.invalidate(table_id).await
+            }
         }
         Ok(())
     })
diff --git a/src/common/meta/src/cache/table/table_schema.rs b/src/common/meta/src/cache/table/table_schema.rs
index 99ece65683..33b1773f45 100644
--- a/src/common/meta/src/cache/table/table_schema.rs
+++ b/src/common/meta/src/cache/table/table_schema.rs
@@ -65,7 +65,7 @@ fn init_factory(table_info_manager: TableInfoManager) -> Initializer<TableId, Ar
 /// Never invalidates table id schema cache.
 fn invalidator<'a>(
     _cache: &'a Cache<TableId, Arc<SchemaName>>,
-    _ident: &'a CacheIdent,
+    _idents: &'a [&CacheIdent],
 ) -> BoxFuture<'a, error::Result<()>> {
     Box::pin(std::future::ready(Ok(())))
 }
diff --git a/src/common/meta/src/cache/table/view_info.rs b/src/common/meta/src/cache/table/view_info.rs
index 6a85493d42..d0e1058a7e 100644
--- a/src/common/meta/src/cache/table/view_info.rs
+++ b/src/common/meta/src/cache/table/view_info.rs
@@ -60,11 +60,13 @@ fn init_factory(view_info_manager: ViewInfoManagerRef) -> Initializer<TableId, A
 
 fn invalidator<'a>(
     cache: &'a Cache<TableId, Arc<ViewInfoValue>>,
-    ident: &'a CacheIdent,
+    idents: &'a [&CacheIdent],
 ) -> BoxFuture<'a, Result<()>> {
     Box::pin(async move {
-        if let CacheIdent::TableId(view_id) = ident {
-            cache.invalidate(view_id).await
+        for ident in idents {
+            if let CacheIdent::TableId(view_id) = ident {
+                cache.invalidate(view_id).await
+            }
         }
         Ok(())
     })
diff --git a/src/common/meta/src/cluster.rs b/src/common/meta/src/cluster.rs
index 78af133e8f..5ab2b87399 100644
--- a/src/common/meta/src/cluster.rs
+++ b/src/common/meta/src/cluster.rs
@@ -12,11 +12,12 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+use std::collections::HashMap;
 use std::fmt::{Display, Formatter};
 use std::hash::{DefaultHasher, Hash, Hasher};
 use std::str::FromStr;
 
-use api::v1::meta::{DatanodeWorkloads, HeartbeatRequest};
+use api::v1::meta::{DatanodeWorkloads, FlownodeWorkloads, HeartbeatRequest};
 use common_error::ext::ErrorExt;
 use lazy_static::lazy_static;
 use regex::Regex;
@@ -134,6 +135,9 @@ pub struct NodeInfo {
     // The node build hostname
     #[serde(default)]
     pub hostname: String,
+    /// Environment variables reported by the node.
+    #[serde(default)]
+    pub env_vars: HashMap<String, String>,
 }
 
 #[derive(Debug, Clone, Copy, Eq, Hash, PartialEq, Serialize, Deserialize, PartialOrd, Ord)]
@@ -187,7 +191,11 @@ pub struct FrontendStatus {}
 
 /// The status of a flownode.
 #[derive(Debug, Serialize, Deserialize)]
-pub struct FlownodeStatus {}
+pub struct FlownodeStatus {
+    /// The workloads of the flownode.
+    #[serde(default)]
+    pub workloads: FlownodeWorkloads,
+}
 
 /// The status of a metasrv.
 #[derive(Debug, Serialize, Deserialize)]
@@ -303,13 +311,13 @@ impl TryFrom<i32> for Role {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_workload::DatanodeWorkloadType;
 
     use super::*;
     use crate::cluster::Role::{Datanode, Frontend};
-    use crate::cluster::{DatanodeStatus, NodeInfo, NodeInfoKey, NodeStatus};
+    use crate::cluster::{DatanodeStatus, FlownodeStatus, NodeInfo, NodeInfoKey, NodeStatus};
     use crate::peer::Peer;
 
     #[test]
@@ -351,6 +359,7 @@ mod tests {
             cpu_usage_millicores: 0,
             memory_usage_bytes: 0,
             hostname: "test_hostname".to_string(),
+            env_vars: Default::default(),
         };
 
         let node_info_bytes: Vec<u8> = node_info.try_into().unwrap();
@@ -404,4 +413,116 @@ mod tests {
         let id4 = calculate_node_id(long_addr);
         assert!(id4 > 0);
     }
+
+    #[test]
+    fn test_flownode_status_backward_compatible_without_workloads() {
+        let raw = r#"{
+            "peer":{"id":1,"addr":"127.0.0.1"},
+            "last_activity_ts":123,
+            "status":{"Flownode":{}},
+            "version":"",
+            "git_commit":"",
+            "start_time_ms":1,
+            "total_cpu_millicores":0,
+            "total_memory_bytes":0,
+            "cpu_usage_millicores":0,
+            "memory_usage_bytes":0,
+            "hostname":""
+        }"#;
+
+        let node_info: NodeInfo = raw.parse().unwrap();
+        assert_matches!(
+            node_info.status,
+            NodeStatus::Flownode(FlownodeStatus { workloads }) if workloads.types.is_empty()
+        );
+    }
+
+    #[test]
+    fn test_flownode_status_round_trip_with_workloads() {
+        let node_info = NodeInfo {
+            peer: Peer {
+                id: 1,
+                addr: "127.0.0.1".to_string(),
+            },
+            last_activity_ts: 123,
+            status: NodeStatus::Flownode(FlownodeStatus {
+                workloads: FlownodeWorkloads { types: vec![7] },
+            }),
+            version: "".to_string(),
+            git_commit: "".to_string(),
+            start_time_ms: 1,
+            total_cpu_millicores: 0,
+            total_memory_bytes: 0,
+            cpu_usage_millicores: 0,
+            memory_usage_bytes: 0,
+            hostname: "test_hostname".to_string(),
+            env_vars: Default::default(),
+        };
+
+        let node_info_bytes: Vec<u8> = node_info.try_into().unwrap();
+        let new_node_info: NodeInfo = node_info_bytes.try_into().unwrap();
+
+        assert_matches!(
+            new_node_info,
+            NodeInfo {
+                status: NodeStatus::Flownode(FlownodeStatus { workloads }),
+                ..
+            } if workloads.types == vec![7]
+        );
+    }
+
+    #[test]
+    fn test_node_info_backward_compatible_without_env_vars() {
+        // Simulate a NodeInfo serialized before env_vars was added
+        let raw = r#"{
+            "peer":{"id":1,"addr":"127.0.0.1"},
+            "last_activity_ts":123,
+            "status":{"Datanode":{"rcus":0,"wcus":0,"leader_regions":0,"follower_regions":0,"workloads":{"types":[0]}}},
+            "version":"",
+            "git_commit":"",
+            "start_time_ms":1,
+            "total_cpu_millicores":0,
+            "total_memory_bytes":0,
+            "cpu_usage_millicores":0,
+            "memory_usage_bytes":0,
+            "hostname":"test"
+        }"#;
+
+        let node_info: NodeInfo = raw.parse().unwrap();
+        assert!(node_info.env_vars.is_empty());
+    }
+
+    #[test]
+    fn test_node_info_with_env_vars_round_trip() {
+        let mut env_vars = HashMap::new();
+        env_vars.insert("AZ".to_string(), "us-east-1a".to_string());
+
+        let node_info = NodeInfo {
+            peer: Peer {
+                id: 1,
+                addr: "127.0.0.1".to_string(),
+            },
+            last_activity_ts: 123,
+            status: NodeStatus::Datanode(DatanodeStatus {
+                rcus: 0,
+                wcus: 0,
+                leader_regions: 0,
+                follower_regions: 0,
+                workloads: DatanodeWorkloads { types: vec![] },
+            }),
+            version: "".to_string(),
+            git_commit: "".to_string(),
+            start_time_ms: 1,
+            total_cpu_millicores: 0,
+            total_memory_bytes: 0,
+            cpu_usage_millicores: 0,
+            memory_usage_bytes: 0,
+            hostname: "test".to_string(),
+            env_vars,
+        };
+
+        let node_info_bytes: Vec<u8> = node_info.try_into().unwrap();
+        let new_node_info: NodeInfo = node_info_bytes.try_into().unwrap();
+        assert_eq!(new_node_info.env_vars.get("AZ").unwrap(), "us-east-1a");
+    }
 }
diff --git a/src/common/meta/src/datanode.rs b/src/common/meta/src/datanode.rs
index 8b521d8e43..f3d3a4bed1 100644
--- a/src/common/meta/src/datanode.rs
+++ b/src/common/meta/src/datanode.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::collections::HashSet;
+use std::collections::{HashMap, HashSet};
 use std::str::FromStr;
 
 use api::v1::meta::{DatanodeWorkloads, HeartbeatRequest, RequestHeader};
@@ -371,6 +371,48 @@ impl GcStat {
     }
 }
 
+/// Environment variables reported by a node in heartbeat messages.
+#[derive(Debug, Clone, Serialize, Deserialize, Default)]
+pub struct EnvVars {
+    pub vars: HashMap<String, String>,
+}
+
+impl EnvVars {
+    pub const ENV_VARS_KEY: &str = "__env_vars";
+
+    pub fn new(vars: HashMap<String, String>) -> Self {
+        Self { vars }
+    }
+
+    /// Read the configured env var keys from the environment and build an EnvVars.
+    pub fn from_config(keys: &[String]) -> Self {
+        let vars = keys
+            .iter()
+            .filter_map(|key| std::env::var(key).ok().map(|value| (key.clone(), value)))
+            .collect();
+        Self { vars }
+    }
+
+    pub fn into_extensions(&self, extensions: &mut HashMap<String, Vec<u8>>) {
+        if self.vars.is_empty() {
+            return;
+        }
+        let bytes = serde_json::to_vec(self).unwrap_or_default();
+        extensions.insert(Self::ENV_VARS_KEY.to_string(), bytes);
+    }
+
+    pub fn from_extensions(extensions: &HashMap<String, Vec<u8>>) -> Result<Option<Self>> {
+        extensions
+            .get(Self::ENV_VARS_KEY)
+            .map(|bytes| {
+                serde_json::from_slice(bytes).with_context(|_| DeserializeFromJsonSnafu {
+                    input: String::from_utf8_lossy(bytes).to_string(),
+                })
+            })
+            .transpose()
+    }
+}
+
 /// The key of the datanode stat in the memory store.
 ///
 /// The format is `__meta_datanode_stat-0-{node_id}`.
@@ -573,4 +615,57 @@ mod tests {
         let region_num = stat_val.region_num().unwrap();
         assert_eq!(2, region_num);
     }
+
+    #[test]
+    fn test_region_stat_from_heartbeat_preserves_staging_leader_role() {
+        let request = HeartbeatRequest {
+            header: Some(RequestHeader::default()),
+            peer: Some(api::v1::meta::Peer {
+                id: 1,
+                addr: "127.0.0.1:3001".to_string(),
+            }),
+            region_stats: vec![api::v1::meta::RegionStat {
+                region_id: RegionId::new(1024, 1).as_u64(),
+                engine: "mito".to_string(),
+                role: api::v1::meta::RegionRole::StagingLeader.into(),
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+
+        let stat = Stat::try_from(&request).unwrap();
+
+        assert_eq!(stat.region_stats.len(), 1);
+        assert_eq!(stat.region_stats[0].role, RegionRole::StagingLeader);
+    }
+
+    #[test]
+    fn test_env_vars_round_trip() {
+        let mut vars = HashMap::new();
+        vars.insert("AZ".to_string(), "us-east-1a".to_string());
+        vars.insert("REGION".to_string(), "us-east-1".to_string());
+        let env_vars = EnvVars::new(vars);
+
+        let mut extensions = HashMap::new();
+        env_vars.into_extensions(&mut extensions);
+
+        let extracted = EnvVars::from_extensions(&extensions).unwrap().unwrap();
+        assert_eq!(extracted.vars.get("AZ").unwrap(), "us-east-1a");
+        assert_eq!(extracted.vars.get("REGION").unwrap(), "us-east-1");
+    }
+
+    #[test]
+    fn test_env_vars_empty_not_written() {
+        let env_vars = EnvVars::default();
+        let mut extensions = HashMap::new();
+        env_vars.into_extensions(&mut extensions);
+        assert!(extensions.is_empty());
+    }
+
+    #[test]
+    fn test_env_vars_from_extensions_missing() {
+        let extensions = HashMap::new();
+        let result = EnvVars::from_extensions(&extensions).unwrap();
+        assert!(result.is_none());
+    }
 }
diff --git a/src/common/meta/src/ddl/create_table.rs b/src/common/meta/src/ddl/create_table.rs
index a5b642f1a2..650a706296 100644
--- a/src/common/meta/src/ddl/create_table.rs
+++ b/src/common/meta/src/ddl/create_table.rs
@@ -45,7 +45,7 @@ use crate::lock_key::{CatalogLock, SchemaLock, TableNameLock};
 use crate::metrics;
 use crate::region_keeper::OperatingRegionGuard;
 use crate::rpc::ddl::CreateTableTask;
-use crate::rpc::router::{RegionRoute, operating_leader_regions};
+use crate::rpc::router::{RegionRoute, operating_leader_region_roles};
 
 pub struct CreateTableProcedure {
     pub context: DdlContext,
@@ -172,8 +172,24 @@ impl CreateTableProcedure {
     ///   - [Code::Cancelled](tonic::status::Code::Cancelled)
     ///   - [Code::DeadlineExceeded](tonic::status::Code::DeadlineExceeded)
     ///   - [Code::Unavailable](tonic::status::Code::Unavailable)
-    pub async fn on_datanode_create_regions(&mut self) -> Result<Status> {
-        let table_route = self.table_route()?.clone();
+    pub async fn on_datanode_create_regions(&mut self, retrying: bool) -> Result<Status> {
+        let mut table_route = self.table_route()?.clone();
+        if retrying {
+            info!(
+                "Remapping region routes addresses for retrying create regions for table: {}",
+                self.data.table_ref()
+            );
+            let storage = self
+                .context
+                .table_metadata_manager
+                .table_route_manager()
+                .table_route_storage();
+            // The peer addresses may change during retries,
+            // so we always remap the region routes.
+            storage
+                .remap_region_routes(&mut table_route.region_routes)
+                .await?;
+        }
         // Registers opening regions
         let guards = self.register_opening_regions(&self.context, &table_route.region_routes)?;
         if !guards.is_empty() {
@@ -240,17 +256,17 @@ impl CreateTableProcedure {
         context: &DdlContext,
         region_routes: &[RegionRoute],
     ) -> Result<Vec<OperatingRegionGuard>> {
-        let opening_regions = operating_leader_regions(region_routes);
+        let opening_regions = operating_leader_region_roles(region_routes);
         if self.opening_regions.len() == opening_regions.len() {
             return Ok(vec![]);
         }
 
         let mut opening_region_guards = Vec::with_capacity(opening_regions.len());
 
-        for (region_id, datanode_id) in opening_regions {
+        for (region_id, datanode_id, role) in opening_regions {
             let guard = context
                 .memory_region_keeper
-                .register(datanode_id, region_id)
+                .register_with_role(datanode_id, region_id, role)
                 .context(error::RegionOperatingRaceSnafu {
                     region_id,
                     peer_id: datanode_id,
@@ -301,7 +317,10 @@ impl Procedure for CreateTableProcedure {
 
         match state {
             CreateTableState::Prepare => self.on_prepare().await,
-            CreateTableState::DatanodeCreateRegions => self.on_datanode_create_regions().await,
+            CreateTableState::DatanodeCreateRegions => {
+                let retrying = ctx.is_retrying().await.unwrap_or(false);
+                self.on_datanode_create_regions(retrying).await
+            }
             CreateTableState::CreateMetadata => self.on_create_metadata(ctx.procedure_id).await,
         }
         .map_err(map_to_procedure_error)
@@ -339,7 +358,7 @@ pub struct CreateTableData {
     #[serde(default)]
     pub column_metadatas: Vec<ColumnMetadata>,
     /// None stands for not allocated yet.
-    table_route: Option<PhysicalTableRouteValue>,
+    pub(crate) table_route: Option<PhysicalTableRouteValue>,
     /// None stands for not allocated yet.
     pub region_wal_options: Option<HashMap<RegionNumber, String>>,
 }
diff --git a/src/common/meta/src/ddl/drop_database.rs b/src/common/meta/src/ddl/drop_database.rs
index d3c5c5831e..53b6190ff5 100644
--- a/src/common/meta/src/ddl/drop_database.rs
+++ b/src/common/meta/src/ddl/drop_database.rs
@@ -58,6 +58,7 @@ pub(crate) struct DropDatabaseContext {
     schema: String,
     drop_if_exists: bool,
     tables: Option<BoxStream<'static, Result<(String, TableNameValue)>>>,
+    retrying: bool,
 }
 
 #[async_trait::async_trait]
@@ -90,6 +91,7 @@ impl DropDatabaseProcedure {
                 schema,
                 drop_if_exists,
                 tables: None,
+                retrying: false,
             },
             state: Box::new(DropDatabaseStart),
         }
@@ -110,6 +112,7 @@ impl DropDatabaseProcedure {
                 schema,
                 drop_if_exists,
                 tables: None,
+                retrying: false,
             },
             state,
         })
@@ -136,9 +139,10 @@ impl Procedure for DropDatabaseProcedure {
             })
     }
 
-    async fn execute(&mut self, _ctx: &ProcedureContext) -> ProcedureResult<Status> {
+    async fn execute(&mut self, ctx: &ProcedureContext) -> ProcedureResult<Status> {
         let state = &mut self.state;
 
+        self.context.retrying = ctx.is_retrying().await.unwrap_or(false);
         let (next, status) = state
             .next(&self.runtime_context, &mut self.context)
             .await
diff --git a/src/common/meta/src/ddl/drop_database/cursor.rs b/src/common/meta/src/ddl/drop_database/cursor.rs
index 0a1180a52e..e060567be5 100644
--- a/src/common/meta/src/ddl/drop_database/cursor.rs
+++ b/src/common/meta/src/ddl/drop_database/cursor.rs
@@ -224,6 +224,7 @@ mod tests {
             schema: DEFAULT_SCHEMA_NAME.to_string(),
             drop_if_exists: false,
             tables: None,
+            retrying: false,
         };
         // Ticks
         let (mut state, status) = state.next(&ddl_context, &mut ctx).await.unwrap();
@@ -259,6 +260,7 @@ mod tests {
             schema: DEFAULT_SCHEMA_NAME.to_string(),
             drop_if_exists: false,
             tables: None,
+            retrying: false,
         };
         // Ticks
         let (state, status) = state.next(&ddl_context, &mut ctx).await.unwrap();
@@ -287,6 +289,7 @@ mod tests {
             schema: DEFAULT_SCHEMA_NAME.to_string(),
             drop_if_exists: false,
             tables: None,
+            retrying: false,
         };
         // Ticks
         let (state, status) = state.next(&ddl_context, &mut ctx).await.unwrap();
diff --git a/src/common/meta/src/ddl/drop_database/executor.rs b/src/common/meta/src/ddl/drop_database/executor.rs
index c478cdb746..1c01060c5e 100644
--- a/src/common/meta/src/ddl/drop_database/executor.rs
+++ b/src/common/meta/src/ddl/drop_database/executor.rs
@@ -29,7 +29,7 @@ use crate::ddl::utils::get_region_wal_options;
 use crate::error::{self, Result};
 use crate::key::table_route::TableRouteValue;
 use crate::region_keeper::OperatingRegionGuard;
-use crate::rpc::router::{RegionRoute, operating_leader_regions};
+use crate::rpc::router::{RegionRoute, operating_leader_region_roles};
 
 #[derive(Debug, Serialize, Deserialize)]
 pub(crate) struct DropDatabaseExecutor {
@@ -69,12 +69,12 @@ impl DropDatabaseExecutor {
         if !self.dropping_regions.is_empty() {
             return Ok(());
         }
-        let dropping_regions = operating_leader_regions(&self.physical_region_routes);
+        let dropping_regions = operating_leader_region_roles(&self.physical_region_routes);
         let mut dropping_region_guards = Vec::with_capacity(dropping_regions.len());
-        for (region_id, datanode_id) in dropping_regions {
+        for (region_id, datanode_id, role) in dropping_regions {
             let guard = ddl_ctx
                 .memory_region_keeper
-                .register(datanode_id, region_id)
+                .register_with_role(datanode_id, region_id, role)
                 .context(error::RegionOperatingRaceSnafu {
                     region_id,
                     peer_id: datanode_id,
@@ -96,10 +96,25 @@ impl State for DropDatabaseExecutor {
     async fn next(
         &mut self,
         ddl_ctx: &DdlContext,
-        _ctx: &mut DropDatabaseContext,
+        ctx: &mut DropDatabaseContext,
     ) -> Result<(Box<dyn State>, Status)> {
         self.register_dropping_regions(ddl_ctx)?;
         let executor = DropTableExecutor::new(self.table_name.clone(), self.table_id, true);
+        if ctx.retrying {
+            info!(
+                "Remapping region routes addresses for retrying drop regions for table_id: {}",
+                self.table_id
+            );
+            let storage = ddl_ctx
+                .table_metadata_manager
+                .table_route_manager()
+                .table_route_storage();
+            // The peer addresses may change during retries,
+            // so we always remap the region routes.
+            storage
+                .remap_region_routes(&mut self.physical_region_routes)
+                .await?;
+        }
         // Deletes metadata for table permanently.
         let table_route_value = TableRouteValue::new(
             self.table_id,
@@ -144,6 +159,7 @@ impl State for DropDatabaseExecutor {
 
 #[cfg(test)]
 mod tests {
+    use std::collections::HashSet;
     use std::sync::Arc;
 
     use api::region::RegionResponse;
@@ -152,16 +168,21 @@ mod tests {
     use common_error::ext::BoxedError;
     use common_query::request::QueryRequest;
     use common_recordbatch::SendableRecordBatchStream;
+    use store_api::region_engine::RegionRole;
+    use store_api::storage::RegionId;
     use table::table_name::TableName;
 
     use crate::ddl::drop_database::cursor::DropDatabaseCursor;
     use crate::ddl::drop_database::executor::DropDatabaseExecutor;
     use crate::ddl::drop_database::{DropDatabaseContext, DropTableTarget, State};
-    use crate::ddl::test_util::{create_logical_table, create_physical_table};
+    use crate::ddl::test_util::datanode_handler::DatanodeWatcher;
+    use crate::ddl::test_util::{
+        create_logical_table, create_physical_table, put_datanode_address,
+    };
     use crate::error::{self, Error, Result};
     use crate::key::datanode_table::DatanodeTableKey;
     use crate::peer::Peer;
-    use crate::rpc::router::region_distribution;
+    use crate::rpc::router::{LeaderState, Region, RegionRoute, region_distribution};
     use crate::test_util::{MockDatanodeHandler, MockDatanodeManager, new_ddl_context};
 
     #[derive(Clone)]
@@ -206,6 +227,7 @@ mod tests {
                 schema: DEFAULT_SCHEMA_NAME.to_string(),
                 drop_if_exists: false,
                 tables: None,
+                retrying: false,
             };
             let (state, status) = state.next(&ddl_context, &mut ctx).await.unwrap();
             assert!(!status.need_persist());
@@ -218,6 +240,7 @@ mod tests {
             schema: DEFAULT_SCHEMA_NAME.to_string(),
             drop_if_exists: false,
             tables: None,
+            retrying: false,
         };
         let mut state = DropDatabaseExecutor::new(
             physical_table_id,
@@ -258,6 +281,7 @@ mod tests {
                 schema: DEFAULT_SCHEMA_NAME.to_string(),
                 drop_if_exists: false,
                 tables: None,
+                retrying: false,
             };
             let (state, status) = state.next(&ddl_context, &mut ctx).await.unwrap();
             assert!(!status.need_persist());
@@ -270,6 +294,7 @@ mod tests {
             schema: DEFAULT_SCHEMA_NAME.to_string(),
             drop_if_exists: false,
             tables: None,
+            retrying: false,
         };
         let mut state = DropDatabaseExecutor::new(
             logical_table_id,
@@ -360,6 +385,7 @@ mod tests {
             schema: DEFAULT_SCHEMA_NAME.to_string(),
             drop_if_exists: false,
             tables: None,
+            retrying: false,
         };
         let err = state.next(&ddl_context, &mut ctx).await.unwrap_err();
         assert!(err.is_retry_later());
@@ -389,6 +415,7 @@ mod tests {
                 schema: DEFAULT_SCHEMA_NAME.to_string(),
                 drop_if_exists: false,
                 tables: None,
+                retrying: false,
             };
             state.recover(&ddl_context).unwrap();
             assert_eq!(state.dropping_regions.len(), 1);
@@ -398,4 +425,73 @@ mod tests {
             assert_eq!(cursor.target, DropTableTarget::Physical);
         }
     }
+
+    #[tokio::test]
+    async fn test_recover_registers_region_role_from_routes() {
+        let node_manager = Arc::new(MockDatanodeManager::new(NaiveDatanodeHandler));
+        let ddl_context = new_ddl_context(node_manager);
+        let region_id = RegionId::new(1024, 1);
+        let mut state = DropDatabaseExecutor::new(
+            1024,
+            1024,
+            TableName::new(DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME, "phy"),
+            vec![RegionRoute {
+                region: Region::new_test(region_id),
+                leader_peer: Some(Peer::empty(7)),
+                follower_peers: vec![],
+                leader_state: Some(LeaderState::Downgrading),
+                leader_down_since: None,
+                write_route_policy: None,
+            }],
+            DropTableTarget::Physical,
+        );
+
+        state.recover(&ddl_context).unwrap();
+
+        let roles = ddl_context
+            .memory_region_keeper
+            .extract_operating_region_roles(7, &HashSet::from([region_id]));
+        assert_eq!(roles.get(&region_id), Some(&RegionRole::DowngradingLeader));
+    }
+
+    #[tokio::test]
+    async fn test_next_remaps_addresses_when_retrying() {
+        let (tx, mut rx) = tokio::sync::mpsc::channel(8);
+        let node_manager = Arc::new(MockDatanodeManager::new(DatanodeWatcher::new(tx)));
+        let ddl_context = new_ddl_context(node_manager);
+        let physical_table_id = create_physical_table(&ddl_context, "phy").await;
+        let (_, table_route) = ddl_context
+            .table_metadata_manager
+            .table_route_manager()
+            .get_physical_table_route(physical_table_id)
+            .await
+            .unwrap();
+
+        let mut state = DropDatabaseExecutor::new(
+            physical_table_id,
+            physical_table_id,
+            TableName::new(DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME, "phy"),
+            table_route.region_routes,
+            DropTableTarget::Physical,
+        );
+        state.physical_region_routes[0]
+            .leader_peer
+            .as_mut()
+            .unwrap()
+            .addr = "old-addr".to_string();
+        let mut ctx = DropDatabaseContext {
+            catalog: DEFAULT_CATALOG_NAME.to_string(),
+            schema: DEFAULT_SCHEMA_NAME.to_string(),
+            drop_if_exists: false,
+            tables: None,
+            retrying: true,
+        };
+
+        put_datanode_address(&ddl_context, 0, "new-addr").await;
+
+        state.next(&ddl_context, &mut ctx).await.unwrap();
+
+        let (peer, _) = rx.try_recv().unwrap();
+        assert_eq!(peer.addr, "new-addr");
+    }
 }
diff --git a/src/common/meta/src/ddl/drop_database/metadata.rs b/src/common/meta/src/ddl/drop_database/metadata.rs
index c71eded7af..0306dd2a3b 100644
--- a/src/common/meta/src/ddl/drop_database/metadata.rs
+++ b/src/common/meta/src/ddl/drop_database/metadata.rs
@@ -122,6 +122,7 @@ mod tests {
             schema: "bar".to_string(),
             drop_if_exists: true,
             tables: None,
+            retrying: false,
         };
         let (state, status) = state.next(&ddl_context, &mut ctx).await.unwrap();
         state
@@ -150,6 +151,7 @@ mod tests {
             schema: "bar".to_string(),
             drop_if_exists: true,
             tables: None,
+            retrying: false,
         };
         let (state, status) = state.next(&ddl_context, &mut ctx).await.unwrap();
         state
diff --git a/src/common/meta/src/ddl/drop_database/start.rs b/src/common/meta/src/ddl/drop_database/start.rs
index 775071d684..6b3e149877 100644
--- a/src/common/meta/src/ddl/drop_database/start.rs
+++ b/src/common/meta/src/ddl/drop_database/start.rs
@@ -72,7 +72,7 @@ impl State for DropDatabaseStart {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::sync::Arc;
 
     use crate::ddl::drop_database::cursor::DropDatabaseCursor;
@@ -93,6 +93,7 @@ mod tests {
             schema: "bar".to_string(),
             drop_if_exists: false,
             tables: None,
+            retrying: false,
         };
         let err = step.next(&ddl_context, &mut ctx).await.unwrap_err();
         assert_matches!(err, error::Error::SchemaNotFound { .. });
@@ -108,6 +109,7 @@ mod tests {
             schema: "bar".to_string(),
             drop_if_exists: true,
             tables: None,
+            retrying: false,
         };
         let (state, status) = state.next(&ddl_context, &mut ctx).await.unwrap();
         state.as_any().downcast_ref::<DropDatabaseEnd>().unwrap();
@@ -130,6 +132,7 @@ mod tests {
             schema: "bar".to_string(),
             drop_if_exists: false,
             tables: None,
+            retrying: false,
         };
         let (state, status) = state.next(&ddl_context, &mut ctx).await.unwrap();
         state.as_any().downcast_ref::<DropDatabaseCursor>().unwrap();
diff --git a/src/common/meta/src/ddl/drop_table.rs b/src/common/meta/src/ddl/drop_table.rs
index 55c33330c4..4f4e2d6053 100644
--- a/src/common/meta/src/ddl/drop_table.rs
+++ b/src/common/meta/src/ddl/drop_table.rs
@@ -43,7 +43,7 @@ use crate::lock_key::{CatalogLock, SchemaLock, TableLock};
 use crate::metrics;
 use crate::region_keeper::OperatingRegionGuard;
 use crate::rpc::ddl::DropTableTask;
-use crate::rpc::router::{RegionRoute, operating_leader_regions};
+use crate::rpc::router::{RegionRoute, operating_leader_region_roles};
 
 pub struct DropTableProcedure {
     /// The context of procedure runtime.
@@ -94,7 +94,7 @@ impl DropTableProcedure {
 
     /// Register dropping regions if doesn't exist.
     fn register_dropping_regions(&mut self) -> Result<()> {
-        let dropping_regions = operating_leader_regions(&self.data.physical_region_routes);
+        let dropping_regions = operating_leader_region_roles(&self.data.physical_region_routes);
 
         if !self.dropping_regions.is_empty() {
             return Ok(());
@@ -102,11 +102,11 @@ impl DropTableProcedure {
 
         let mut dropping_region_guards = Vec::with_capacity(dropping_regions.len());
 
-        for (region_id, datanode_id) in dropping_regions {
+        for (region_id, datanode_id, role) in dropping_regions {
             let guard = self
                 .context
                 .memory_region_keeper
-                .register(datanode_id, region_id)
+                .register_with_role(datanode_id, region_id, role)
                 .context(error::RegionOperatingRaceSnafu {
                     region_id,
                     peer_id: datanode_id,
@@ -154,7 +154,24 @@ impl DropTableProcedure {
         Ok(Status::executing(true))
     }
 
-    pub async fn on_datanode_drop_regions(&mut self) -> Result<Status> {
+    pub async fn on_datanode_drop_regions(&mut self, retrying: bool) -> Result<Status> {
+        if retrying {
+            info!(
+                "Remapping region routes addresses for retrying drop regions for table_id: {}",
+                self.data.table_id()
+            );
+            let storage = self
+                .context
+                .table_metadata_manager
+                .table_route_manager()
+                .table_route_storage();
+            // The peer addresses may change during retries,
+            // so we always remap the region routes.
+            storage
+                .remap_region_routes(&mut self.data.physical_region_routes)
+                .await?;
+        }
+
         self.executor
             .on_drop_regions(
                 &self.context.node_manager,
@@ -215,7 +232,7 @@ impl Procedure for DropTableProcedure {
         Ok(())
     }
 
-    async fn execute(&mut self, _ctx: &ProcedureContext) -> ProcedureResult<Status> {
+    async fn execute(&mut self, ctx: &ProcedureContext) -> ProcedureResult<Status> {
         let state = &self.data.state;
         let _timer = metrics::METRIC_META_PROCEDURE_DROP_TABLE
             .with_label_values(&[state.as_ref()])
@@ -225,7 +242,10 @@ impl Procedure for DropTableProcedure {
             DropTableState::Prepare => self.on_prepare().await,
             DropTableState::DeleteMetadata => self.on_delete_metadata().await,
             DropTableState::InvalidateTableCache => self.on_broadcast().await,
-            DropTableState::DatanodeDropRegions => self.on_datanode_drop_regions().await,
+            DropTableState::DatanodeDropRegions => {
+                let retrying = ctx.is_retrying().await.unwrap_or(false);
+                self.on_datanode_drop_regions(retrying).await
+            }
             DropTableState::DeleteTombstone => self.on_delete_metadata_tombstone().await,
         }
         .map_err(map_to_procedure_error)
diff --git a/src/common/meta/src/ddl/drop_table/executor.rs b/src/common/meta/src/ddl/drop_table/executor.rs
index c342487365..271bdbfede 100644
--- a/src/common/meta/src/ddl/drop_table/executor.rs
+++ b/src/common/meta/src/ddl/drop_table/executor.rs
@@ -322,7 +322,7 @@ impl DropTableExecutor {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
     use std::sync::Arc;
 
diff --git a/src/common/meta/src/ddl/test_util.rs b/src/common/meta/src/ddl/test_util.rs
index 1dd1f783dc..7d8cdd895b 100644
--- a/src/common/meta/src/ddl/test_util.rs
+++ b/src/common/meta/src/ddl/test_util.rs
@@ -19,7 +19,7 @@ pub mod datanode_handler;
 pub mod flownode_handler;
 pub mod region_metadata;
 
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 use std::collections::HashMap;
 
 use api::v1::meta::Partition;
@@ -41,8 +41,12 @@ use crate::ddl::test_util::create_table::{
     TestCreateTableExprBuilder, build_raw_table_info_from_expr,
 };
 use crate::ddl::{DdlContext, TableMetadata};
+use crate::key::node_address::{NodeAddressKey, NodeAddressValue};
 use crate::key::table_route::TableRouteValue;
+use crate::key::{MetadataKey, MetadataValue};
+use crate::peer::Peer;
 use crate::rpc::ddl::CreateTableTask;
+use crate::rpc::store::PutRequest;
 
 pub async fn create_physical_table_metadata(
     ddl_context: &DdlContext,
@@ -56,6 +60,21 @@ pub async fn create_physical_table_metadata(
         .unwrap();
 }
 
+pub async fn put_datanode_address(ddl_context: &DdlContext, node_id: u64, addr: &str) {
+    ddl_context
+        .table_metadata_manager
+        .kv_backend()
+        .put(PutRequest {
+            key: NodeAddressKey::with_datanode(node_id).to_bytes(),
+            value: NodeAddressValue::new(Peer::new(node_id, addr))
+                .try_as_raw_value()
+                .unwrap(),
+            ..Default::default()
+        })
+        .await
+        .unwrap();
+}
+
 pub async fn create_physical_table(ddl_context: &DdlContext, name: &str) -> TableId {
     // Prepares physical table metadata.
     let mut create_physical_table_task = test_create_physical_table_task(name);
diff --git a/src/common/meta/src/ddl/tests/alter_logical_tables.rs b/src/common/meta/src/ddl/tests/alter_logical_tables.rs
index e6bb7676f5..bca56bafc2 100644
--- a/src/common/meta/src/ddl/tests/alter_logical_tables.rs
+++ b/src/common/meta/src/ddl/tests/alter_logical_tables.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 use std::sync::Arc;
 
 use api::region::RegionResponse;
diff --git a/src/common/meta/src/ddl/tests/alter_table.rs b/src/common/meta/src/ddl/tests/alter_table.rs
index d935aa6a15..14ee71b3e6 100644
--- a/src/common/meta/src/ddl/tests/alter_table.rs
+++ b/src/common/meta/src/ddl/tests/alter_table.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 use std::collections::HashMap;
 use std::sync::Arc;
 
@@ -256,7 +256,7 @@ async fn test_on_submit_alter_request() {
         results.push(result);
     }
     rx.try_recv().unwrap_err();
-    results.sort_unstable_by(|(a, _), (b, _)| a.id.cmp(&b.id));
+    results.sort_unstable_by_key(|(a, _)| a.id);
 
     let (peer, request) = results.remove(0);
     assert_alter_request(peer, request, 1, RegionId::new(table_id, 1));
@@ -310,7 +310,7 @@ async fn test_on_submit_alter_request_without_sync_request() {
         results.push(result);
     }
     rx.try_recv().unwrap_err();
-    results.sort_unstable_by(|(a, _), (b, _)| a.id.cmp(&b.id));
+    results.sort_unstable_by_key(|(a, _)| a.id);
 
     let (peer, request) = results.remove(0);
     assert_alter_request(peer, request, 1, RegionId::new(table_id, 1));
diff --git a/src/common/meta/src/ddl/tests/create_flow.rs b/src/common/meta/src/ddl/tests/create_flow.rs
index 8803f98e0d..344fc05024 100644
--- a/src/common/meta/src/ddl/tests/create_flow.rs
+++ b/src/common/meta/src/ddl/tests/create_flow.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 use std::collections::HashMap;
 use std::sync::Arc;
 
@@ -37,6 +37,8 @@ fn test_query_context() -> QueryContext {
         timezone: "UTC".to_string(),
         extensions: HashMap::new(),
         channel: 0,
+        snapshot_seqs: HashMap::new(),
+        sst_min_sequences: HashMap::new(),
     }
 }
 
@@ -251,6 +253,10 @@ fn test_create_flow_data_new_format_serialization() {
         catalog: "new_catalog".to_string(),
         schema: "new_schema".to_string(),
         timezone: "America/New_York".to_string(),
+        extensions: HashMap::new(),
+        channel: 0,
+        snapshot_seqs: HashMap::new(),
+        sst_min_sequences: HashMap::new(),
     };
 
     let data = CreateFlowData {
@@ -272,6 +278,9 @@ fn test_create_flow_data_new_format_serialization() {
     assert_eq!(deserialized.flow_context.catalog, "new_catalog");
     assert_eq!(deserialized.flow_context.schema, "new_schema");
     assert_eq!(deserialized.flow_context.timezone, "America/New_York");
+    assert_eq!(deserialized.flow_context.channel, 0);
+    assert_eq!(deserialized.flow_context.snapshot_seqs, HashMap::new());
+    assert_eq!(deserialized.flow_context.sst_min_sequences, HashMap::new());
 }
 
 #[test]
@@ -286,6 +295,8 @@ fn test_flow_query_context_conversion_from_query_context() {
         ]
         .into(),
         channel: 99,
+        snapshot_seqs: HashMap::from([(1, 10)]),
+        sst_min_sequences: HashMap::from([(1, 8)]),
     };
 
     let flow_context: FlowQueryContext = query_context.into();
@@ -293,6 +304,9 @@ fn test_flow_query_context_conversion_from_query_context() {
     assert_eq!(flow_context.catalog, "prod_catalog");
     assert_eq!(flow_context.schema, "public");
     assert_eq!(flow_context.timezone, "America/Los_Angeles");
+    assert_eq!(flow_context.channel, 99);
+    assert_eq!(flow_context.snapshot_seqs, HashMap::from([(1, 10)]));
+    assert_eq!(flow_context.sst_min_sequences, HashMap::from([(1, 8)]));
 }
 
 #[test]
@@ -301,6 +315,10 @@ fn test_flow_info_conversion_with_flow_context() {
         catalog: "info_catalog".to_string(),
         schema: "info_schema".to_string(),
         timezone: "Europe/Berlin".to_string(),
+        extensions: HashMap::new(),
+        channel: 0,
+        snapshot_seqs: HashMap::new(),
+        sst_min_sequences: HashMap::new(),
     };
 
     let data = CreateFlowData {
@@ -349,6 +367,10 @@ fn test_mixed_serialization_format_support() {
         catalog: "test".to_string(),
         schema: "test".to_string(),
         timezone: "UTC".to_string(),
+        extensions: HashMap::new(),
+        channel: 0,
+        snapshot_seqs: HashMap::new(),
+        sst_min_sequences: HashMap::new(),
     };
     assert_eq!(ctx_from_new, expected_new);
 }
diff --git a/src/common/meta/src/ddl/tests/create_logical_tables.rs b/src/common/meta/src/ddl/tests/create_logical_tables.rs
index f7dd397f9f..e2927e8df6 100644
--- a/src/common/meta/src/ddl/tests/create_logical_tables.rs
+++ b/src/common/meta/src/ddl/tests/create_logical_tables.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 use std::sync::Arc;
 
 use api::region::RegionResponse;
diff --git a/src/common/meta/src/ddl/tests/create_table.rs b/src/common/meta/src/ddl/tests/create_table.rs
index 5cc1db71cb..da91dedcaf 100644
--- a/src/common/meta/src/ddl/tests/create_table.rs
+++ b/src/common/meta/src/ddl/tests/create_table.rs
@@ -12,8 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
-use std::collections::HashMap;
+use std::assert_matches;
+use std::collections::{HashMap, HashSet};
 use std::sync::Arc;
 
 use api::region::RegionResponse;
@@ -30,6 +30,7 @@ use datatypes::prelude::ConcreteDataType;
 use datatypes::schema::ColumnSchema;
 use store_api::metadata::ColumnMetadata;
 use store_api::metric_engine_consts::TABLE_COLUMN_METADATA_EXTENSION_KEY;
+use store_api::region_engine::RegionRole;
 use store_api::storage::RegionId;
 use tokio::sync::mpsc;
 
@@ -42,7 +43,7 @@ use crate::ddl::test_util::datanode_handler::{
     DatanodeWatcher, NaiveDatanodeHandler, RetryErrorDatanodeHandler,
     UnexpectedErrorDatanodeHandler,
 };
-use crate::ddl::test_util::{assert_column_name, get_raw_table_info};
+use crate::ddl::test_util::{assert_column_name, get_raw_table_info, put_datanode_address};
 use crate::error::{Error, Result};
 use crate::key::table_route::TableRouteValue;
 use crate::kv_backend::memory::MemoryKvBackend;
@@ -244,6 +245,27 @@ async fn test_on_datanode_create_regions_should_not_retry() {
     assert!(!error.is_retry_later());
 }
 
+#[tokio::test]
+async fn test_on_datanode_create_regions_remaps_addresses_when_retrying() {
+    let (tx, mut rx) = mpsc::channel(8);
+    let datanode_handler = DatanodeWatcher::new(tx).with_handler(create_request_handler);
+    let node_manager = Arc::new(MockDatanodeManager::new(datanode_handler));
+    let ddl_context = new_ddl_context(node_manager);
+    let task = test_create_table_task("foo");
+    let mut procedure = CreateTableProcedure::new(task, ddl_context.clone()).unwrap();
+    procedure.on_prepare().await.unwrap();
+
+    let table_route = procedure.data.table_route.as_mut().unwrap();
+    let leader = table_route.region_routes[0].leader_peer.as_mut().unwrap();
+    leader.addr = "old-addr".to_string();
+    put_datanode_address(&ddl_context, leader.id, "new-addr").await;
+
+    procedure.on_datanode_create_regions(true).await.unwrap();
+
+    let (peer, _) = rx.try_recv().unwrap();
+    assert_eq!(peer.addr, "new-addr");
+}
+
 #[tokio::test]
 async fn test_on_create_metadata_error() {
     common_telemetry::init_default_ut_logging();
@@ -330,6 +352,10 @@ async fn test_memory_region_keeper_guard_dropped_on_procedure_done() {
             .memory_region_keeper
             .contains(datanode_id, region_id)
     );
+    let roles = ddl_context
+        .memory_region_keeper
+        .extract_operating_region_roles(datanode_id, &HashSet::from([region_id]));
+    assert_eq!(roles.get(&region_id), Some(&RegionRole::Leader));
 
     execute_procedure_until_done(&mut procedure).await;
 
diff --git a/src/common/meta/src/ddl/tests/create_view.rs b/src/common/meta/src/ddl/tests/create_view.rs
index e4fefa8944..cc98bb1bae 100644
--- a/src/common/meta/src/ddl/tests/create_view.rs
+++ b/src/common/meta/src/ddl/tests/create_view.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 use std::collections::HashSet;
 use std::sync::Arc;
 
diff --git a/src/common/meta/src/ddl/tests/drop_flow.rs b/src/common/meta/src/ddl/tests/drop_flow.rs
index 8de42f5c96..af34da4809 100644
--- a/src/common/meta/src/ddl/tests/drop_flow.rs
+++ b/src/common/meta/src/ddl/tests/drop_flow.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 use std::collections::HashMap;
 use std::sync::Arc;
 
diff --git a/src/common/meta/src/ddl/tests/drop_table.rs b/src/common/meta/src/ddl/tests/drop_table.rs
index 65c3915adc..c518bb36a2 100644
--- a/src/common/meta/src/ddl/tests/drop_table.rs
+++ b/src/common/meta/src/ddl/tests/drop_table.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 use std::sync::Arc;
 
 use api::v1::region::{RegionRequest, region_request};
@@ -23,6 +23,7 @@ use common_procedure::Procedure;
 use common_procedure_test::{
     execute_procedure_until, execute_procedure_until_done, new_test_procedure_context,
 };
+use store_api::region_engine::RegionRole;
 use store_api::storage::RegionId;
 use table::metadata::TableId;
 use tokio::sync::mpsc;
@@ -34,7 +35,7 @@ use crate::ddl::test_util::create_table::test_create_table_task;
 use crate::ddl::test_util::datanode_handler::{DatanodeWatcher, NaiveDatanodeHandler};
 use crate::ddl::test_util::{
     create_logical_table, create_physical_table, create_physical_table_metadata,
-    test_create_logical_table_task, test_create_physical_table_task,
+    put_datanode_address, test_create_logical_table_task, test_create_physical_table_task,
 };
 use crate::key::table_route::TableRouteValue;
 use crate::kv_backend::memory::MemoryKvBackend;
@@ -146,7 +147,7 @@ async fn test_on_datanode_drop_regions() {
     // Drop table
     let mut procedure = DropTableProcedure::new(task, ddl_context);
     procedure.on_prepare().await.unwrap();
-    procedure.on_datanode_drop_regions().await.unwrap();
+    procedure.on_datanode_drop_regions(false).await.unwrap();
 
     let check = |peer: Peer,
                  request: RegionRequest,
@@ -172,7 +173,7 @@ async fn test_on_datanode_drop_regions() {
         let result = rx.try_recv().unwrap();
         results.push(result);
     }
-    results.sort_unstable_by(|(a, _), (b, _)| a.id.cmp(&b.id));
+    results.sort_unstable_by_key(|(a, _)| a.id);
 
     let (peer, request) = results.remove(0);
     check(peer, request, 1, RegionId::new(table_id, 1), false);
@@ -186,6 +187,50 @@ async fn test_on_datanode_drop_regions() {
     check(peer, request, 5, RegionId::new(table_id, 1), true);
 }
 
+#[tokio::test]
+async fn test_on_datanode_drop_regions_remaps_addresses_when_retrying() {
+    let (tx, mut rx) = mpsc::channel(8);
+    let datanode_handler = DatanodeWatcher::new(tx);
+    let node_manager = Arc::new(MockDatanodeManager::new(datanode_handler));
+    let ddl_context = new_ddl_context(node_manager);
+    let table_id = 1024;
+    let table_name = "foo";
+    let task = test_create_table_task(table_name, table_id);
+    ddl_context
+        .table_metadata_manager
+        .create_table_metadata(
+            task.table_info.clone(),
+            TableRouteValue::physical(vec![RegionRoute {
+                region: Region::new_test(RegionId::new(table_id, 1)),
+                leader_peer: Some(Peer::new(1, "old-leader")),
+                follower_peers: vec![Peer::new(5, "old-follower")],
+                leader_state: None,
+                leader_down_since: None,
+                write_route_policy: None,
+            }]),
+            HashMap::new(),
+        )
+        .await
+        .unwrap();
+
+    let task = new_drop_table_task(table_name, table_id, false);
+    let mut procedure = DropTableProcedure::new(task, ddl_context.clone());
+    procedure.on_prepare().await.unwrap();
+
+    put_datanode_address(&ddl_context, 1, "new-leader").await;
+    put_datanode_address(&ddl_context, 5, "new-follower").await;
+
+    procedure.on_datanode_drop_regions(true).await.unwrap();
+
+    let mut peers = Vec::new();
+    for _ in 0..2 {
+        peers.push(rx.try_recv().unwrap().0);
+    }
+    peers.sort_unstable_by_key(|p| p.id);
+    assert_eq!(peers[0].addr, "new-leader");
+    assert_eq!(peers[1].addr, "new-follower");
+}
+
 #[tokio::test]
 async fn test_on_rollback() {
     let node_manager = Arc::new(MockDatanodeManager::new(NaiveDatanodeHandler));
@@ -284,6 +329,10 @@ async fn test_memory_region_keeper_guard_dropped_on_procedure_done() {
                 .memory_region_keeper
                 .contains(datanode_id, region_id)
         );
+        let roles = ddl_context
+            .memory_region_keeper
+            .extract_operating_region_roles(datanode_id, &HashSet::from([region_id]));
+        assert_eq!(roles.get(&region_id), Some(&RegionRole::Leader));
 
         execute_procedure_until_done(&mut procedure).await;
 
diff --git a/src/common/meta/src/ddl/truncate_table.rs b/src/common/meta/src/ddl/truncate_table.rs
index 40e8980a1c..21491b6230 100644
--- a/src/common/meta/src/ddl/truncate_table.rs
+++ b/src/common/meta/src/ddl/truncate_table.rs
@@ -42,7 +42,7 @@ use crate::key::table_name::TableNameKey;
 use crate::lock_key::{CatalogLock, SchemaLock, TableLock};
 use crate::metrics;
 use crate::rpc::ddl::TruncateTableTask;
-use crate::rpc::router::{RegionRoute, find_leader_regions, find_leaders};
+use crate::rpc::router::{find_leader_regions, find_leaders};
 
 pub struct TruncateTableProcedure {
     context: DdlContext,
@@ -94,12 +94,11 @@ impl TruncateTableProcedure {
     pub(crate) fn new(
         task: TruncateTableTask,
         table_info_value: DeserializedValueWithBytes<TableInfoValue>,
-        region_routes: Vec<RegionRoute>,
         context: DdlContext,
     ) -> Self {
         Self {
             context,
-            data: TruncateTableData::new(task, table_info_value, region_routes),
+            data: TruncateTableData::new(task, table_info_value),
         }
     }
 
@@ -138,13 +137,18 @@ impl TruncateTableProcedure {
     async fn on_datanode_truncate_regions(&mut self) -> Result<Status> {
         let table_id = self.data.table_id();
 
-        let region_routes = &self.data.region_routes;
-        let leaders = find_leaders(region_routes);
+        let (_, physical_table_route) = self
+            .context
+            .table_metadata_manager
+            .table_route_manager()
+            .get_physical_table_route(table_id)
+            .await?;
+        let leaders = find_leaders(&physical_table_route.region_routes);
         let mut truncate_region_tasks = Vec::with_capacity(leaders.len());
 
         for datanode in leaders {
             let requester = self.context.node_manager.datanode(&datanode).await;
-            let regions = find_leader_regions(region_routes, &datanode);
+            let regions = find_leader_regions(&physical_table_route.region_routes, &datanode);
 
             for region in regions {
                 let region_id = RegionId::new(table_id, region);
@@ -201,20 +205,17 @@ pub struct TruncateTableData {
     state: TruncateTableState,
     task: TruncateTableTask,
     table_info_value: DeserializedValueWithBytes<TableInfoValue>,
-    region_routes: Vec<RegionRoute>,
 }
 
 impl TruncateTableData {
     pub fn new(
         task: TruncateTableTask,
         table_info_value: DeserializedValueWithBytes<TableInfoValue>,
-        region_routes: Vec<RegionRoute>,
     ) -> Self {
         Self {
             state: TruncateTableState::Prepare,
             task,
             table_info_value,
-            region_routes,
         }
     }
 
diff --git a/src/common/meta/src/ddl_manager.rs b/src/common/meta/src/ddl_manager.rs
index 0106add32f..d0619ca74f 100644
--- a/src/common/meta/src/ddl_manager.rs
+++ b/src/common/meta/src/ddl_manager.rs
@@ -45,7 +45,7 @@ use crate::ddl::drop_view::DropViewProcedure;
 use crate::ddl::truncate_table::TruncateTableProcedure;
 use crate::ddl::{DdlContext, utils};
 use crate::error::{
-    CreateRepartitionProcedureSnafu, EmptyDdlTasksSnafu, ProcedureOutputSnafu,
+    self, CreateRepartitionProcedureSnafu, EmptyDdlTasksSnafu, ProcedureOutputSnafu,
     RegisterProcedureLoaderSnafu, RegisterRepartitionProcedureLoaderSnafu, Result,
     SubmitProcedureSnafu, TableInfoNotFoundSnafu, TableNotFoundSnafu, TableRouteNotFoundSnafu,
     UnexpectedLogicalRouteTableSnafu, WaitProcedureSnafu,
@@ -72,7 +72,6 @@ use crate::rpc::ddl::{
     CreateTableTask, CreateViewTask, DropDatabaseTask, DropFlowTask, DropTableTask, DropViewTask,
     QueryContext, SubmitDdlTaskRequest, SubmitDdlTaskResponse, TruncateTableTask,
 };
-use crate::rpc::router::RegionRoute;
 
 /// A configurator that customizes or enhances a [`DdlManager`].
 #[async_trait::async_trait]
@@ -521,15 +520,9 @@ impl DdlManager {
         &self,
         truncate_table_task: TruncateTableTask,
         table_info_value: DeserializedValueWithBytes<TableInfoValue>,
-        region_routes: Vec<RegionRoute>,
     ) -> Result<(ProcedureId, Option<Output>)> {
         let context = self.create_context();
-        let procedure = TruncateTableProcedure::new(
-            truncate_table_task,
-            table_info_value,
-            region_routes,
-            context,
-        );
+        let procedure = TruncateTableProcedure::new(truncate_table_task, table_info_value, context);
 
         let procedure_with_id = ProcedureWithId::with_random_id(Box::new(procedure));
 
@@ -658,19 +651,26 @@ async fn handle_truncate_table_task(
     let table_metadata_manager = &ddl_manager.table_metadata_manager();
     let table_ref = truncate_table_task.table_ref();
 
-    let (table_info_value, table_route_value) =
-        table_metadata_manager.get_full_table_info(table_id).await?;
-
-    let table_info_value = table_info_value.with_context(|| TableInfoNotFoundSnafu {
-        table: table_ref.to_string(),
-    })?;
-
-    let table_route_value = table_route_value.context(TableRouteNotFoundSnafu { table_id })?;
-
-    let table_route = table_route_value.into_inner().region_routes()?.clone();
+    let table_info_value = table_metadata_manager
+        .table_info_manager()
+        .get(table_id)
+        .await?
+        .with_context(|| TableInfoNotFoundSnafu {
+            table: table_ref.to_string(),
+        })?;
+    let physical_table_id = table_metadata_manager
+        .table_route_manager()
+        .get_physical_table_id(table_id)
+        .await?;
+    ensure!(
+        physical_table_id == table_id,
+        error::UnexpectedSnafu {
+            err_msg: "Truncate table is only supported for physical tables."
+        }
+    );
 
     let (id, _) = ddl_manager
-        .submit_truncate_table_task(truncate_table_task, table_info_value, table_route)
+        .submit_truncate_table_task(truncate_table_task, table_info_value)
         .await?;
 
     info!("Table: {table_id} is truncated via procedure_id {id:?}");
diff --git a/src/meta-srv/src/election.rs b/src/common/meta/src/election.rs
similarity index 67%
rename from src/meta-srv/src/election.rs
rename to src/common/meta/src/election.rs
index 2d2826b286..12173beda8 100644
--- a/src/meta-srv/src/election.rs
+++ b/src/common/meta/src/election.rs
@@ -21,15 +21,85 @@ use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, Ordering};
 
 use common_telemetry::{error, info, warn};
+use serde::{Deserialize, Serialize};
 use tokio::sync::broadcast::error::RecvError;
 use tokio::sync::broadcast::{self, Receiver, Sender};
 
 use crate::error::Result;
-use crate::metasrv::MetasrvNodeInfo;
 
-pub(crate) const CANDIDATE_LEASE_SECS: u64 = 600;
+pub const CANDIDATE_LEASE_SECS: u64 = 600;
 const KEEP_ALIVE_INTERVAL_SECS: u64 = CANDIDATE_LEASE_SECS / 2;
 
+/// The value of the leader. It is used to store the leader's address.
+pub struct LeaderValue(pub String);
+
+impl<T: AsRef<[u8]>> From<T> for LeaderValue {
+    fn from(value: T) -> Self {
+        let string = String::from_utf8_lossy(value.as_ref());
+        Self(string.to_string())
+    }
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct MetasrvNodeInfo {
+    // The metasrv's address
+    pub addr: String,
+    // The node build version
+    pub version: String,
+    // The node build git commit hash
+    pub git_commit: String,
+    // The node start timestamp in milliseconds
+    pub start_time_ms: u64,
+    // The node total cpu millicores
+    #[serde(default)]
+    pub total_cpu_millicores: i64,
+    // The node total memory bytes
+    #[serde(default)]
+    pub total_memory_bytes: i64,
+    /// The node build cpu usage millicores
+    #[serde(default)]
+    pub cpu_usage_millicores: i64,
+    /// The node build memory usage bytes
+    #[serde(default)]
+    pub memory_usage_bytes: i64,
+    // The node hostname
+    #[serde(default)]
+    pub hostname: String,
+}
+
+// TODO(zyy17): Allow deprecated fields for backward compatibility. Remove this when the deprecated top-level fields are removed from the proto.
+#[allow(deprecated)]
+impl From<MetasrvNodeInfo> for api::v1::meta::MetasrvNodeInfo {
+    fn from(node_info: MetasrvNodeInfo) -> Self {
+        Self {
+            peer: Some(api::v1::meta::Peer {
+                addr: node_info.addr,
+                ..Default::default()
+            }),
+            // TODO(zyy17): The following top-level fields are deprecated. They are kept for backward compatibility and will be removed in a future version.
+            // New code should use the fields in `info.NodeInfo` instead.
+            version: node_info.version.clone(),
+            git_commit: node_info.git_commit.clone(),
+            start_time_ms: node_info.start_time_ms,
+            cpus: node_info.total_cpu_millicores as u32,
+            memory_bytes: node_info.total_memory_bytes as u64,
+            // The canonical location for node information.
+            info: Some(api::v1::meta::NodeInfo {
+                version: node_info.version,
+                git_commit: node_info.git_commit,
+                start_time_ms: node_info.start_time_ms,
+                total_cpu_millicores: node_info.total_cpu_millicores,
+                total_memory_bytes: node_info.total_memory_bytes,
+                cpu_usage_millicores: node_info.cpu_usage_millicores,
+                memory_usage_bytes: node_info.memory_usage_bytes,
+                cpus: node_info.total_cpu_millicores as u32,
+                memory_bytes: node_info.total_memory_bytes as u64,
+                hostname: node_info.hostname,
+            }),
+        }
+    }
+}
+
 /// Messages sent when the leader changes.
 #[derive(Debug, Clone)]
 pub enum LeaderChangeMessage {
@@ -168,3 +238,5 @@ pub trait Election: Send + Sync {
 
     fn subscribe_leader_change(&self) -> Receiver<LeaderChangeMessage>;
 }
+
+pub type ElectionRef = Arc<dyn Election<Leader = LeaderValue>>;
diff --git a/src/meta-srv/src/election/etcd.rs b/src/common/meta/src/election/etcd.rs
similarity index 94%
rename from src/meta-srv/src/election/etcd.rs
rename to src/common/meta/src/election/etcd.rs
index 883f723d74..affad31ef4 100644
--- a/src/meta-srv/src/election/etcd.rs
+++ b/src/common/meta/src/election/etcd.rs
@@ -16,8 +16,6 @@ use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, Ordering};
 use std::time::Duration;
 
-use common_meta::distributed_time_constants::{META_KEEP_ALIVE_INTERVAL_SECS, META_LEASE_SECS};
-use common_meta::key::{CANDIDATES_ROOT, ELECTION_KEY};
 use common_telemetry::{error, info, warn};
 use etcd_client::{
     Client, GetOptions, LeaderKey as EtcdLeaderKey, LeaseKeepAliveStream, LeaseKeeper, PutOptions,
@@ -27,13 +25,15 @@ use tokio::sync::broadcast;
 use tokio::sync::broadcast::Receiver;
 use tokio::time::{MissedTickBehavior, timeout};
 
+use crate::distributed_time_constants::{META_KEEP_ALIVE_INTERVAL_SECS, META_LEASE_SECS};
 use crate::election::{
-    CANDIDATE_LEASE_SECS, Election, KEEP_ALIVE_INTERVAL_SECS, LeaderChangeMessage, LeaderKey,
-    listen_leader_change, send_leader_change_and_set_flags,
+    CANDIDATE_LEASE_SECS, Election, ElectionRef, KEEP_ALIVE_INTERVAL_SECS, LeaderChangeMessage,
+    LeaderKey, LeaderValue, MetasrvNodeInfo, listen_leader_change,
+    send_leader_change_and_set_flags,
 };
 use crate::error;
 use crate::error::Result;
-use crate::metasrv::{ElectionRef, LeaderValue, MetasrvNodeInfo};
+use crate::key::{CANDIDATES_ROOT, ELECTION_KEY};
 
 impl LeaderKey for EtcdLeaderKey {
     fn name(&self) -> &[u8] {
@@ -253,7 +253,7 @@ impl Election for EtcdElection {
                 .leader(self.election_key())
                 .await
                 .context(error::EtcdFailedSnafu)?;
-            let leader_value = res.kv().context(error::NoLeaderSnafu)?.value();
+            let leader_value = res.kv().context(error::ElectionNoLeaderSnafu)?.value();
             Ok(leader_value.into())
         }
     }
@@ -279,7 +279,7 @@ impl EtcdElection {
             ensure!(
                 res.ttl() > 0,
                 error::UnexpectedSnafu {
-                    violated: "Failed to refresh the lease",
+                    err_msg: "Failed to refresh the lease".to_string(),
                 }
             );
 
diff --git a/src/meta-srv/src/election/rds.rs b/src/common/meta/src/election/rds.rs
similarity index 96%
rename from src/meta-srv/src/election/rds.rs
rename to src/common/meta/src/election/rds.rs
index 16e113415a..6ee529ee02 100644
--- a/src/meta-srv/src/election/rds.rs
+++ b/src/common/meta/src/election/rds.rs
@@ -36,7 +36,7 @@ fn parse_value_and_expire_time(value: &str) -> Result<(String, Timestamp)> {
             .split(LEASE_SEP)
             .collect_tuple()
             .with_context(|| UnexpectedSnafu {
-                violated: format!(
+                err_msg: format!(
                     "Invalid value {}, expect node info || {} || expire time",
                     value, LEASE_SEP
                 ),
@@ -45,7 +45,7 @@ fn parse_value_and_expire_time(value: &str) -> Result<(String, Timestamp)> {
     let expire_time = match Timestamp::from_str(expire_time, None) {
         Ok(ts) => ts,
         Err(_) => UnexpectedSnafu {
-            violated: format!("Invalid timestamp: {}", expire_time),
+            err_msg: format!("Invalid timestamp: {}", expire_time),
         }
         .fail()?,
     };
diff --git a/src/meta-srv/src/election/rds/mysql.rs b/src/common/meta/src/election/rds/mysql.rs
similarity index 97%
rename from src/meta-srv/src/election/rds/mysql.rs
rename to src/common/meta/src/election/rds/mysql.rs
index 20051a2610..bd694e4ae1 100644
--- a/src/meta-srv/src/election/rds/mysql.rs
+++ b/src/common/meta/src/election/rds/mysql.rs
@@ -16,7 +16,6 @@ use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, Ordering};
 use std::time::Duration;
 
-use common_meta::key::{CANDIDATES_ROOT, ELECTION_KEY};
 use common_telemetry::{error, info, warn};
 use common_time::Timestamp;
 use snafu::{OptionExt, ResultExt, ensure};
@@ -29,14 +28,15 @@ use tokio::time::MissedTickBehavior;
 
 use crate::election::rds::{LEASE_SEP, Lease, RdsLeaderKey, parse_value_and_expire_time};
 use crate::election::{
-    Election, LeaderChangeMessage, listen_leader_change, send_leader_change_and_set_flags,
+    Election, ElectionRef, LeaderChangeMessage, LeaderValue, MetasrvNodeInfo, listen_leader_change,
+    send_leader_change_and_set_flags,
 };
 use crate::error::{
     AcquireMySqlClientSnafu, DecodeSqlValueSnafu, DeserializeFromJsonSnafu,
-    LeaderLeaseChangedSnafu, LeaderLeaseExpiredSnafu, MySqlExecutionSnafu, NoLeaderSnafu, Result,
-    SerializeToJsonSnafu, SqlExecutionTimeoutSnafu, UnexpectedSnafu,
+    ElectionLeaderLeaseChangedSnafu, ElectionLeaderLeaseExpiredSnafu, ElectionNoLeaderSnafu,
+    MySqlExecutionSnafu, Result, SerializeToJsonSnafu, SqlExecutionTimeoutSnafu, UnexpectedSnafu,
 };
-use crate::metasrv::{ElectionRef, LeaderValue, MetasrvNodeInfo};
+use crate::key::{CANDIDATES_ROOT, ELECTION_KEY};
 
 struct ElectionSqlFactory<'a> {
     table_name: &'a str,
@@ -592,7 +592,7 @@ impl Election for MySqlElection {
             ensure!(
                 lease.expire_time > lease.current,
                 UnexpectedSnafu {
-                    violated: format!(
+                    err_msg: format!(
                         "Candidate lease expired at {:?} (current time: {:?}), key: {:?}",
                         lease.expire_time,
                         lease.current,
@@ -667,10 +667,10 @@ impl Election for MySqlElection {
             let client = self.client.lock().await;
             let mut executor = Executor::Default(client);
             if let Some(lease) = self.get_value_with_lease(&key, &mut executor).await? {
-                ensure!(lease.expire_time > lease.current, NoLeaderSnafu);
+                ensure!(lease.expire_time > lease.current, ElectionNoLeaderSnafu);
                 Ok(lease.leader_value.as_bytes().into())
             } else {
-                NoLeaderSnafu.fail()
+                ElectionNoLeaderSnafu.fail()
             }
         }
     }
@@ -705,7 +705,7 @@ impl MySqlElection {
         let current_time = match Timestamp::from_str(&current_time_str, None) {
             Ok(ts) => ts,
             Err(_) => UnexpectedSnafu {
-                violated: format!("Invalid timestamp: {}", current_time_str),
+                err_msg: format!("Invalid timestamp: {}", current_time_str),
             }
             .fail()?,
         };
@@ -740,7 +740,7 @@ impl MySqlElection {
             current = match Timestamp::from_str(current_time_str, None) {
                 Ok(ts) => ts,
                 Err(_) => UnexpectedSnafu {
-                    violated: format!("Invalid timestamp: {}", current_time_str),
+                    err_msg: format!("Invalid timestamp: {}", current_time_str),
                 }
                 .fail()?,
             };
@@ -777,7 +777,7 @@ impl MySqlElection {
         ensure!(
             res == 1,
             UnexpectedSnafu {
-                violated: format!("Failed to update key: {}", String::from_utf8_lossy(key)),
+                err_msg: format!("Failed to update key: {}", String::from_utf8_lossy(key)),
             }
         );
 
@@ -920,9 +920,12 @@ impl MySqlElection {
     ///   will be released.
     /// - **Case 2**: If all checks pass, the function returns without performing any actions.
     fn lease_check(&self, lease: &Option<Lease>) -> Result<Lease> {
-        let lease = lease.as_ref().context(NoLeaderSnafu)?;
+        let lease = lease.as_ref().context(ElectionNoLeaderSnafu)?;
         // Case 1: Lease expired
-        ensure!(lease.expire_time > lease.current, LeaderLeaseExpiredSnafu);
+        ensure!(
+            lease.expire_time > lease.current,
+            ElectionLeaderLeaseExpiredSnafu
+        );
         // Case 2: Everything is fine
         Ok(lease.clone())
     }
@@ -960,7 +963,7 @@ impl MySqlElection {
         let remote_lease = self.get_value_with_lease(&key, &mut executor).await?;
         ensure!(
             expected_lease.map(|lease| lease.origin) == remote_lease.map(|lease| lease.origin),
-            LeaderLeaseChangedSnafu
+            ElectionLeaderLeaseChangedSnafu
         );
         self.delete_value(&key, &mut executor).await?;
         self.put_value_with_lease(
@@ -984,15 +987,13 @@ impl MySqlElection {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
-    use std::env;
+    use std::{assert_matches, env};
 
-    use common_meta::maybe_skip_mysql_integration_test;
     use common_telemetry::init_default_ut_logging;
+    use sqlx::MySqlPool;
 
     use super::*;
-    use crate::error;
-    use crate::utils::mysql::create_mysql_pool;
+    use crate::{error, maybe_skip_mysql_integration_test};
 
     async fn create_mysql_client(
         table_name: Option<&str>,
@@ -1003,11 +1004,11 @@ mod tests {
         let endpoint = env::var("GT_MYSQL_ENDPOINTS").unwrap_or_default();
         if endpoint.is_empty() {
             return UnexpectedSnafu {
-                violated: "MySQL endpoint is empty".to_string(),
+                err_msg: "MySQL endpoint is empty".to_string(),
             }
             .fail();
         }
-        let pool = create_mysql_pool(&[endpoint], None).await.unwrap();
+        let pool = MySqlPool::connect(&endpoint).await.unwrap();
         let mut client = ElectionMysqlClient::new(
             pool,
             execution_timeout,
@@ -1302,7 +1303,7 @@ mod tests {
         let err = elected(&leader_mysql_election, table_name, Some(incorrect_lease))
             .await
             .unwrap_err();
-        assert_matches!(err, error::Error::LeaderLeaseChanged { .. });
+        assert_matches!(err, error::Error::ElectionLeaderLeaseChanged { .. });
         let lease = get_lease(&leader_mysql_election).await;
         assert!(lease.is_none());
         drop_table(&leader_mysql_election.client, table_name).await;
diff --git a/src/meta-srv/src/election/rds/postgres.rs b/src/common/meta/src/election/rds/postgres.rs
similarity index 97%
rename from src/meta-srv/src/election/rds/postgres.rs
rename to src/common/meta/src/election/rds/postgres.rs
index c21efd780b..220b33bb60 100644
--- a/src/meta-srv/src/election/rds/postgres.rs
+++ b/src/common/meta/src/election/rds/postgres.rs
@@ -16,7 +16,6 @@ use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, Ordering};
 use std::time::Duration;
 
-use common_meta::key::{CANDIDATES_ROOT, ELECTION_KEY};
 use common_telemetry::{error, info, warn};
 use common_time::Timestamp;
 use deadpool_postgres::{Manager, Pool};
@@ -28,13 +27,15 @@ use tokio_postgres::types::ToSql;
 
 use crate::election::rds::{LEASE_SEP, Lease, RdsLeaderKey, parse_value_and_expire_time};
 use crate::election::{
-    Election, LeaderChangeMessage, listen_leader_change, send_leader_change_and_set_flags,
+    Election, ElectionRef, LeaderChangeMessage, LeaderValue, MetasrvNodeInfo, listen_leader_change,
+    send_leader_change_and_set_flags,
 };
 use crate::error::{
-    DeserializeFromJsonSnafu, GetPostgresClientSnafu, NoLeaderSnafu, PostgresExecutionSnafu,
-    Result, SerializeToJsonSnafu, SqlExecutionTimeoutSnafu, UnexpectedSnafu,
+    DeserializeFromJsonSnafu, ElectionNoLeaderSnafu, GetPostgresClientSnafu,
+    PostgresExecutionSnafu, Result, SerializeToJsonSnafu, SqlExecutionTimeoutSnafu,
+    UnexpectedSnafu,
 };
-use crate::metasrv::{ElectionRef, LeaderValue, MetasrvNodeInfo};
+use crate::key::{CANDIDATES_ROOT, ELECTION_KEY};
 
 struct ElectionSqlFactory<'a> {
     lock_id: u64,
@@ -404,13 +405,13 @@ impl Election for PgElection {
                 .get_value_with_lease(&key)
                 .await?
                 .context(UnexpectedSnafu {
-                    violated: format!("Failed to get lease for key: {:?}", key),
+                    err_msg: format!("Failed to get lease for key: {:?}", key),
                 })?;
 
             ensure!(
                 lease.expire_time > lease.current,
                 UnexpectedSnafu {
-                    violated: format!(
+                    err_msg: format!(
                         "Candidate lease expired at {:?} (current time {:?}), key: {:?}",
                         lease.expire_time, lease.current, key
                     ),
@@ -464,11 +465,11 @@ impl Election for PgElection {
                 .query(&self.sql_set.campaign, &[])
                 .await?;
             let row = res.first().context(UnexpectedSnafu {
-                violated: "Failed to get the result of acquiring advisory lock",
+                err_msg: "Failed to get the result of acquiring advisory lock".to_string(),
             })?;
             let is_leader = row.try_get(0).map_err(|_| {
                 UnexpectedSnafu {
-                    violated: "Failed to get the result of get lock",
+                    err_msg: "Failed to get the result of get lock".to_string(),
                 }
                 .build()
             })?;
@@ -500,10 +501,10 @@ impl Election for PgElection {
         } else {
             let key = self.election_key();
             if let Some(lease) = self.get_value_with_lease(&key).await? {
-                ensure!(lease.expire_time > lease.current, NoLeaderSnafu);
+                ensure!(lease.expire_time > lease.current, ElectionNoLeaderSnafu);
                 Ok(lease.leader_value.as_bytes().into())
             } else {
-                NoLeaderSnafu.fail()
+                ElectionNoLeaderSnafu.fail()
             }
         }
     }
@@ -537,7 +538,7 @@ impl PgElection {
             let current_time = match Timestamp::from_str(current_time_str, None) {
                 Ok(ts) => ts,
                 Err(_) => UnexpectedSnafu {
-                    violated: format!("Invalid timestamp: {}", current_time_str),
+                    err_msg: format!("Invalid timestamp: {}", current_time_str),
                 }
                 .fail()?,
             };
@@ -576,7 +577,7 @@ impl PgElection {
             current = match Timestamp::from_str(current_time_str, None) {
                 Ok(ts) => ts,
                 Err(_) => UnexpectedSnafu {
-                    violated: format!("Invalid timestamp: {}", current_time_str),
+                    err_msg: format!("Invalid timestamp: {}", current_time_str),
                 }
                 .fail()?,
             };
@@ -613,7 +614,7 @@ impl PgElection {
         ensure!(
             res == 1,
             UnexpectedSnafu {
-                violated: format!("Failed to update key: {}", String::from_utf8_lossy(key)),
+                err_msg: format!("Failed to update key: {}", String::from_utf8_lossy(key)),
             }
         );
 
@@ -742,9 +743,9 @@ impl PgElection {
         let lease = self
             .get_value_with_lease(&key)
             .await?
-            .context(NoLeaderSnafu)?;
+            .context(ElectionNoLeaderSnafu)?;
         // Case 2
-        ensure!(lease.expire_time > lease.current, NoLeaderSnafu);
+        ensure!(lease.expire_time > lease.current, ElectionNoLeaderSnafu);
         // Case 3
         Ok(())
     }
@@ -828,14 +829,13 @@ impl PgElection {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
-    use std::env;
+    use std::{assert_matches, env};
 
-    use common_meta::maybe_skip_postgres_integration_test;
+    use deadpool_postgres::{Config, Runtime};
+    use tokio_postgres::NoTls;
 
     use super::*;
-    use crate::error;
-    use crate::utils::postgres::create_postgres_pool;
+    use crate::{error, maybe_skip_postgres_integration_test};
 
     async fn create_postgres_client(
         table_name: Option<&str>,
@@ -846,11 +846,13 @@ mod tests {
         let endpoint = env::var("GT_POSTGRES_ENDPOINTS").unwrap_or_default();
         if endpoint.is_empty() {
             return UnexpectedSnafu {
-                violated: "Postgres endpoint is empty".to_string(),
+                err_msg: "Postgres endpoint is empty".to_string(),
             }
             .fail();
         }
-        let pool = create_postgres_pool(&[endpoint], None, None).await.unwrap();
+        let mut cfg = Config::new();
+        cfg.url = Some(endpoint);
+        let pool = cfg.create_pool(Some(Runtime::Tokio1), NoTls).unwrap();
         let mut pg_client = ElectionPgClient::new(
             pool,
             execution_timeout,
diff --git a/src/common/meta/src/error.rs b/src/common/meta/src/error.rs
index c6613af828..54a5c30fed 100644
--- a/src/common/meta/src/error.rs
+++ b/src/common/meta/src/error.rs
@@ -155,6 +155,13 @@ pub enum Error {
         location: Location,
     },
 
+    #[snafu(display("Trying to write to a read-only kv backend: {}", name))]
+    ReadOnlyKvBackend {
+        name: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
     #[snafu(display("Failed to get procedure state receiver, procedure id: {procedure_id}"))]
     ProcedureStateReceiver {
         procedure_id: ProcedureId,
@@ -338,6 +345,24 @@ pub enum Error {
         location: Location,
     },
 
+    #[snafu(display("Metasrv election has no leader at this moment"))]
+    ElectionNoLeader {
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Metasrv election leader lease expired"))]
+    ElectionLeaderLeaseExpired {
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Metasrv election leader lease changed during election"))]
+    ElectionLeaderLeaseChanged {
+        #[snafu(implicit)]
+        location: Location,
+    },
+
     #[snafu(display("Table already exists, table: {}", table_name))]
     TableAlreadyExists {
         table_name: String,
@@ -714,6 +739,16 @@ pub enum Error {
     #[snafu(display("Failed to get cache"))]
     GetCache { source: Arc<Error> },
 
+    #[snafu(display(
+        "Failed to get latest cache value after {} attempts due to concurrent invalidation",
+        attempts
+    ))]
+    GetLatestCacheRetryExceeded {
+        attempts: usize,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
     #[cfg(feature = "pg_kvbackend")]
     #[snafu(display("Failed to execute via Postgres, sql: {}", sql))]
     PostgresExecution {
@@ -741,6 +776,15 @@ pub enum Error {
         location: Location,
     },
 
+    #[cfg(feature = "pg_kvbackend")]
+    #[snafu(display("Failed to get Postgres client"))]
+    GetPostgresClient {
+        #[snafu(source)]
+        error: deadpool::managed::PoolError<tokio_postgres::Error>,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
     #[cfg(feature = "pg_kvbackend")]
     #[snafu(display("Failed to {} Postgres transaction", operation))]
     PostgresTransaction {
@@ -795,6 +839,24 @@ pub enum Error {
         location: Location,
     },
 
+    #[cfg(feature = "mysql_kvbackend")]
+    #[snafu(display("Failed to decode sql value"))]
+    DecodeSqlValue {
+        #[snafu(source)]
+        error: sqlx::error::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[cfg(feature = "mysql_kvbackend")]
+    #[snafu(display("Failed to acquire mysql client from pool"))]
+    AcquireMySqlClient {
+        #[snafu(source)]
+        error: sqlx::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
     #[cfg(feature = "mysql_kvbackend")]
     #[snafu(display("Failed to {} MySql transaction", operation))]
     MySqlTransaction {
@@ -812,6 +874,15 @@ pub enum Error {
         location: Location,
     },
 
+    #[cfg(any(feature = "pg_kvbackend", feature = "mysql_kvbackend"))]
+    #[snafu(display("Sql execution timeout, sql: {}, duration: {:?}", sql, duration))]
+    SqlExecutionTimeout {
+        sql: String,
+        duration: std::time::Duration,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
     #[snafu(display(
         "Datanode table info not found, table id: {}, datanode id: {}",
         table_id,
@@ -1063,8 +1134,12 @@ impl ErrorExt for Error {
             | ConnectEtcd { .. }
             | MoveValues { .. }
             | GetCache { .. }
+            | GetLatestCacheRetryExceeded { .. }
             | SerializeToJson { .. }
-            | DeserializeFromJson { .. } => StatusCode::Internal,
+            | DeserializeFromJson { .. }
+            | ElectionNoLeader { .. }
+            | ElectionLeaderLeaseExpired { .. }
+            | ElectionLeaderLeaseChanged { .. } => StatusCode::Internal,
 
             NoLeader { .. } => StatusCode::TableUnavailable,
             ValueNotExist { .. }
@@ -1078,7 +1153,7 @@ impl ErrorExt for Error {
             | ColumnIdMismatch { .. }
             | TimestampMismatch { .. } => StatusCode::Unexpected,
 
-            Unsupported { .. } => StatusCode::Unsupported,
+            Unsupported { .. } | ReadOnlyKvBackend { .. } => StatusCode::Unsupported,
             WriteObject { .. } | ReadObject { .. } => StatusCode::StorageUnavailable,
 
             SerdeJson { .. }
@@ -1187,15 +1262,18 @@ impl ErrorExt for Error {
             PostgresExecution { .. }
             | CreatePostgresPool { .. }
             | GetPostgresConnection { .. }
+            | GetPostgresClient { .. }
             | PostgresTransaction { .. }
             | PostgresTlsConfig { .. }
             | InvalidTlsConfig { .. } => StatusCode::Internal,
             #[cfg(feature = "mysql_kvbackend")]
-            MySqlExecution { .. } | CreateMySqlPool { .. } | MySqlTransaction { .. } => {
-                StatusCode::Internal
-            }
+            MySqlExecution { .. }
+            | CreateMySqlPool { .. }
+            | DecodeSqlValue { .. }
+            | AcquireMySqlClient { .. }
+            | MySqlTransaction { .. } => StatusCode::Internal,
             #[cfg(any(feature = "pg_kvbackend", feature = "mysql_kvbackend"))]
-            RdsTransactionRetryFailed { .. } => StatusCode::Internal,
+            RdsTransactionRetryFailed { .. } | SqlExecutionTimeout { .. } => StatusCode::Internal,
             DatanodeTableInfoNotFound { .. } => StatusCode::Internal,
         }
     }
@@ -1243,7 +1321,10 @@ impl Error {
 
     /// Determine whether it is a retry later type through [StatusCode]
     pub fn is_retry_later(&self) -> bool {
-        matches!(self, Error::RetryLater { .. })
+        matches!(
+            self,
+            Error::RetryLater { .. } | Error::GetLatestCacheRetryExceeded { .. }
+        )
     }
 
     /// Determine whether it needs to clean poisons.
diff --git a/src/common/meta/src/heartbeat/utils.rs b/src/common/meta/src/heartbeat/utils.rs
index c037f6306f..f1c3eec1af 100644
--- a/src/common/meta/src/heartbeat/utils.rs
+++ b/src/common/meta/src/heartbeat/utils.rs
@@ -14,7 +14,7 @@
 
 use api::v1::meta::heartbeat_request::NodeWorkloads;
 use api::v1::meta::mailbox_message::Payload;
-use api::v1::meta::{DatanodeWorkloads, MailboxMessage};
+use api::v1::meta::{DatanodeWorkloads, FlownodeWorkloads, MailboxMessage};
 use common_telemetry::tracing_context::TracingContext;
 use common_telemetry::warn;
 use common_time::util::current_time_millis;
@@ -90,6 +90,16 @@ pub fn get_datanode_workloads(node_workloads: Option<&NodeWorkloads>) -> Datanod
     }
 }
 
+/// Extracts flownode workloads from the provided optional `NodeWorkloads`.
+///
+/// Returns empty flownode workloads if the input is `None` or not a flownode payload.
+pub fn get_flownode_workloads(node_workloads: Option<&NodeWorkloads>) -> FlownodeWorkloads {
+    match node_workloads {
+        Some(NodeWorkloads::Flownode(flownode_workloads)) => flownode_workloads.clone(),
+        _ => FlownodeWorkloads { types: vec![] },
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -102,4 +112,16 @@ mod tests {
         let workloads = get_datanode_workloads(node_workloads.as_ref());
         assert_eq!(workloads.types, vec![DatanodeWorkloadType::Hybrid.to_i32()]);
     }
+
+    #[test]
+    fn test_get_flownode_workloads() {
+        let node_workloads = Some(NodeWorkloads::Flownode(FlownodeWorkloads {
+            types: vec![7],
+        }));
+        let workloads = get_flownode_workloads(node_workloads.as_ref());
+        assert_eq!(workloads.types, vec![7]);
+
+        let workloads = get_flownode_workloads(None);
+        assert!(workloads.types.is_empty());
+    }
 }
diff --git a/src/common/meta/src/instruction.rs b/src/common/meta/src/instruction.rs
index 4056597bf5..fd8d17bdff 100644
--- a/src/common/meta/src/instruction.rs
+++ b/src/common/meta/src/instruction.rs
@@ -18,6 +18,7 @@ use std::time::Duration;
 
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
 use store_api::region_engine::SyncRegionFromRequest;
+use store_api::region_request::RegionFlushReason;
 use store_api::storage::{FileRefsManifest, GcReport, RegionId, RegionNumber};
 use strum::Display;
 use table::metadata::TableId;
@@ -338,14 +339,17 @@ pub struct FlushRegions {
     /// Error handling strategy for batch operations (only applies when multiple regions and sync strategy).
     #[serde(default)]
     pub error_strategy: FlushErrorStrategy,
+    /// The source that triggered this flush.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub reason: Option<RegionFlushReason>,
 }
 
 impl Display for FlushRegions {
     fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
         write!(
             f,
-            "FlushRegions(region_ids={:?}, strategy={:?}, error_strategy={:?})",
-            self.region_ids, self.strategy, self.error_strategy
+            "FlushRegions(region_ids={:?}, strategy={:?}, error_strategy={:?}, reason={:?})",
+            self.region_ids, self.strategy, self.error_strategy, self.reason
         )
     }
 }
@@ -357,6 +361,7 @@ impl FlushRegions {
             region_ids: vec![region_id],
             strategy: FlushStrategy::Sync,
             error_strategy: FlushErrorStrategy::FailFast,
+            reason: None,
         }
     }
 
@@ -366,6 +371,7 @@ impl FlushRegions {
             region_ids,
             strategy: FlushStrategy::Async,
             error_strategy: FlushErrorStrategy::TryAll,
+            reason: None,
         }
     }
 
@@ -375,9 +381,15 @@ impl FlushRegions {
             region_ids,
             strategy: FlushStrategy::Sync,
             error_strategy,
+            reason: None,
         }
     }
 
+    pub fn with_reason(mut self, reason: RegionFlushReason) -> Self {
+        self.reason = Some(reason);
+        self
+    }
+
     /// Check if this is a single region flush.
     pub fn is_single_region(&self) -> bool {
         self.region_ids.len() == 1
@@ -1363,6 +1375,7 @@ mod tests {
         assert!(!single_sync.is_hint());
         assert!(single_sync.is_sync());
         assert_eq!(single_sync.error_strategy, FlushErrorStrategy::FailFast);
+        assert_eq!(single_sync.reason, None);
         assert!(single_sync.is_single_region());
         assert_eq!(single_sync.single_region_id(), Some(region_id));
 
@@ -1374,6 +1387,7 @@ mod tests {
         assert!(batch_async.is_hint());
         assert!(!batch_async.is_sync());
         assert_eq!(batch_async.error_strategy, FlushErrorStrategy::TryAll);
+        assert_eq!(batch_async.reason, None);
         assert!(!batch_async.is_single_region());
         assert_eq!(batch_async.single_region_id(), None);
 
@@ -1384,6 +1398,10 @@ mod tests {
         assert!(!batch_sync.is_hint());
         assert!(batch_sync.is_sync());
         assert_eq!(batch_sync.error_strategy, FlushErrorStrategy::FailFast);
+        assert_eq!(batch_sync.reason, None);
+
+        let with_reason = batch_sync.with_reason(RegionFlushReason::RemoteWalPrune);
+        assert_eq!(with_reason.reason, Some(RegionFlushReason::RemoteWalPrune));
     }
 
     #[test]
@@ -1401,6 +1419,7 @@ mod tests {
             region_ids: vec![region_id],
             strategy: FlushStrategy::Async,
             error_strategy: FlushErrorStrategy::TryAll,
+            reason: None,
         };
         assert_eq!(flush_regions.region_ids, vec![region_id]);
         assert_eq!(flush_regions.strategy, FlushStrategy::Async);
@@ -1450,6 +1469,7 @@ mod tests {
         let instruction = Instruction::FlushRegions(flush_regions.clone());
 
         let serialized = serde_json::to_string(&instruction).unwrap();
+        assert!(!serialized.contains("reason"));
         let deserialized: Instruction = serde_json::from_str(&serialized).unwrap();
 
         match deserialized {
@@ -1457,6 +1477,32 @@ mod tests {
                 assert_eq!(fr.region_ids, vec![region_id]);
                 assert_eq!(fr.strategy, FlushStrategy::Sync);
                 assert_eq!(fr.error_strategy, FlushErrorStrategy::FailFast);
+                assert_eq!(fr.reason, None);
+            }
+            _ => panic!("Expected FlushRegions instruction"),
+        }
+
+        let legacy = r#"{"FlushRegions":{"region_ids":[4398046511105],"strategy":"Sync","error_strategy":"FailFast"}}"#;
+        let deserialized: Instruction = serde_json::from_str(legacy).unwrap();
+        match deserialized {
+            Instruction::FlushRegions(fr) => {
+                assert_eq!(fr.region_ids, vec![region_id]);
+                assert_eq!(fr.strategy, FlushStrategy::Sync);
+                assert_eq!(fr.error_strategy, FlushErrorStrategy::FailFast);
+                assert_eq!(fr.reason, None);
+            }
+            _ => panic!("Expected FlushRegions instruction"),
+        }
+
+        let flush_regions = FlushRegions::async_batch(vec![region_id])
+            .with_reason(RegionFlushReason::RemoteWalPrune);
+        let instruction = Instruction::FlushRegions(flush_regions);
+        let serialized = serde_json::to_string(&instruction).unwrap();
+        assert!(serialized.contains(r#""reason":"RemoteWalPrune""#));
+        let deserialized: Instruction = serde_json::from_str(&serialized).unwrap();
+        match deserialized {
+            Instruction::FlushRegions(fr) => {
+                assert_eq!(fr.reason, Some(RegionFlushReason::RemoteWalPrune));
             }
             _ => panic!("Expected FlushRegions instruction"),
         }
@@ -1479,6 +1525,7 @@ mod tests {
                 assert!(!fr.is_hint());
                 assert!(fr.is_sync());
                 assert_eq!(fr.error_strategy, FlushErrorStrategy::TryAll);
+                assert_eq!(fr.reason, None);
             }
             _ => panic!("Expected FlushRegions instruction"),
         }
diff --git a/src/common/meta/src/key.rs b/src/common/meta/src/key.rs
index 97b68f9b04..44ed11148a 100644
--- a/src/common/meta/src/key.rs
+++ b/src/common/meta/src/key.rs
@@ -131,6 +131,7 @@ use common_wal::options::WalOptions;
 use datanode_table::{DatanodeTableKey, DatanodeTableManager, DatanodeTableValue};
 use flow::flow_route::FlowRouteValue;
 use flow::table_flow::TableFlowValue;
+use futures_util::TryStreamExt;
 use lazy_static::lazy_static;
 use regex::Regex;
 pub use schema_metadata_manager::{SchemaMetadataManager, SchemaMetadataManagerRef};
@@ -432,6 +433,28 @@ pub struct DeserializedValueWithBytes<T: DeserializeOwned + Serialize> {
     inner: T,
 }
 
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct DroppedTableName {
+    /// Table id stored in the tombstoned table-name mapping.
+    pub table_id: TableId,
+    /// Original fully qualified table name.
+    pub table_name: TableName,
+}
+
+#[derive(Debug, Clone)]
+pub struct DroppedTableMetadata {
+    /// Table id of the dropped table.
+    pub table_id: TableId,
+    /// Original fully qualified table name.
+    pub table_name: TableName,
+    /// Tombstoned table info value.
+    pub table_info_value: TableInfoValue,
+    /// Tombstoned table route value.
+    pub table_route_value: TableRouteValue,
+    /// Per-region WAL options recovered from tombstoned datanode metadata.
+    pub region_wal_options: HashMap<RegionNumber, WalOptions>,
+}
+
 impl<T: DeserializeOwned + Serialize> Deref for DeserializedValueWithBytes<T> {
     type Target = T;
 
@@ -663,7 +686,7 @@ impl TableMetadataManager {
         if let Some(table_route_value) = &mut table_route_value {
             self.table_route_manager()
                 .table_route_storage()
-                .remap_route_address(table_route_value)
+                .remap_table_route(table_route_value)
                 .await?;
         }
         Ok((table_info_value, table_route_value))
@@ -708,7 +731,7 @@ impl TableMetadataManager {
 
         // Creates view info
         let view_info_value = ViewInfoValue::new(
-            raw_logical_plan,
+            raw_logical_plan.into(),
             table_names,
             columns,
             plan_columns,
@@ -980,6 +1003,51 @@ impl TableMetadataManager {
         self.tombstone_manager.create(keys).await.map(|_| ())
     }
 
+    /// Lists dropped tables from tombstoned table-name entries.
+    pub async fn list_dropped_tables(&self) -> Result<Vec<DroppedTableName>> {
+        let mut stream = self.tombstone_manager.tombstoned_table_names();
+        let mut dropped_tables = Vec::new();
+
+        while let Some(kv) = stream.try_next().await? {
+            let raw_key = self.tombstone_manager.strip_tombstone_prefix(&kv.key)?;
+            let table_name = TableNameKey::from_bytes(raw_key)?.into();
+            let table_id = TableNameValue::try_from_raw_value(&kv.value)?.table_id();
+            dropped_tables.push(DroppedTableName {
+                table_id,
+                table_name,
+            });
+        }
+
+        Ok(dropped_tables)
+    }
+
+    /// Gets dropped table metadata by its original full table name.
+    pub async fn get_dropped_table(
+        &self,
+        table_name: &TableName,
+    ) -> Result<Option<DroppedTableMetadata>> {
+        let table_name_key = TableNameKey::from(table_name);
+        let Some(kv) = self
+            .tombstone_manager
+            .get(&table_name_key.to_bytes())
+            .await?
+        else {
+            return Ok(None);
+        };
+
+        let table_id = TableNameValue::try_from_raw_value(&kv.value)?.table_id();
+        self.get_dropped_table_metadata(table_id, table_name.clone())
+            .await
+    }
+
+    /// Gets dropped table metadata by table id.
+    pub async fn get_dropped_table_by_id(
+        &self,
+        table_id: TableId,
+    ) -> Result<Option<DroppedTableMetadata>> {
+        self.get_dropped_table_metadata(table_id, None).await
+    }
+
     /// Deletes metadata tombstone for table **permanently**.
     /// The caller MUST ensure it has the exclusive access to `TableNameKey`.
     pub async fn delete_table_metadata_tombstone(
@@ -1029,6 +1097,94 @@ impl TableMetadataManager {
         Ok(())
     }
 
+    /// Rebuilds dropped table metadata from tombstoned keys.
+    async fn get_dropped_table_metadata<T>(
+        &self,
+        table_id: TableId,
+        table_name: T,
+    ) -> Result<Option<DroppedTableMetadata>>
+    where
+        T: Into<Option<TableName>>,
+    {
+        let table_info_key = TableInfoKey::new(table_id);
+        let Some(table_info_kv) = self
+            .tombstone_manager
+            .get(&table_info_key.to_bytes())
+            .await?
+        else {
+            return Ok(None);
+        };
+
+        let table_info_value = TableInfoValue::try_from_raw_value(&table_info_kv.value)?;
+        let table_name = table_name
+            .into()
+            .unwrap_or_else(|| table_info_value.table_name());
+
+        let table_route_key = TableRouteKey::new(table_id);
+        let table_route_kv = self
+            .tombstone_manager
+            .get(&table_route_key.to_bytes())
+            .await?
+            .with_context(|| error::UnexpectedSnafu {
+                err_msg: format!("Missing tombstoned table route metadata for table id {table_id}"),
+            })?;
+        let mut table_route_value = TableRouteValue::try_from_raw_value(&table_route_kv.value)?;
+        self.table_route_manager
+            .table_route_storage()
+            .remap_table_route(&mut table_route_value)
+            .await?;
+
+        let region_wal_options = self
+            .dropped_region_wal_options(table_id, &table_route_value)
+            .await?;
+
+        Ok(Some(DroppedTableMetadata {
+            table_id,
+            table_name,
+            table_info_value,
+            table_route_value,
+            region_wal_options,
+        }))
+    }
+
+    /// Rebuilds region WAL options from tombstoned datanode-table entries.
+    async fn dropped_region_wal_options(
+        &self,
+        table_id: TableId,
+        table_route_value: &TableRouteValue,
+    ) -> Result<HashMap<RegionNumber, WalOptions>> {
+        let mut region_wal_options = HashMap::new();
+        let datanode_table_keys = region_distribution(table_route_value.region_routes()?)
+            .into_keys()
+            .map(|datanode_id| DatanodeTableKey::new(datanode_id, table_id))
+            .collect::<Vec<_>>();
+        let datanode_table_key_bytes = datanode_table_keys
+            .iter()
+            .map(|key| key.to_bytes())
+            .collect::<Vec<_>>();
+        let datanode_table_values = self
+            .tombstone_manager
+            .batch_get(&datanode_table_key_bytes)
+            .await?;
+
+        for datanode_table_key in datanode_table_keys {
+            let Some(kv) = datanode_table_values.get(&datanode_table_key.to_bytes()) else {
+                continue;
+            };
+
+            let datanode_table_value = DatanodeTableValue::try_from_raw_value(&kv.value)?;
+            for (region_number, wal_options) in &datanode_table_value.region_info.region_wal_options
+            {
+                region_wal_options.insert(
+                    *region_number,
+                    serde_json::from_str(wal_options).context(error::SerdeJsonSnafu)?,
+                );
+            }
+        }
+
+        Ok(region_wal_options)
+    }
+
     fn view_info_keys(&self, view_id: TableId, view_name: &TableName) -> Result<Vec<Vec<u8>>> {
         let mut keys = Vec::with_capacity(3);
         let view_name = TableNameKey::new(
@@ -1184,7 +1340,7 @@ impl TableMetadataManager {
         definition: String,
     ) -> Result<()> {
         let new_view_info_value = current_view_info_value.update(
-            new_view_info,
+            new_view_info.into(),
             table_names,
             columns,
             plan_columns,
@@ -1504,6 +1660,7 @@ mod tests {
     };
     use crate::kv_backend::KvBackend;
     use crate::kv_backend::memory::MemoryKvBackend;
+    use crate::kv_backend::read_only::ReadOnlyKvBackend;
     use crate::peer::Peer;
     use crate::rpc::router::{LeaderState, Region, RegionRoute, region_distribution};
     use crate::rpc::store::{PutRequest, RangeRequest};
@@ -1610,6 +1767,25 @@ mod tests {
             .collect()
     }
 
+    fn create_mixed_region_wal_options() -> HashMap<RegionNumber, WalOptions> {
+        HashMap::from([
+            (
+                0,
+                WalOptions::Kafka(KafkaWalOptions {
+                    topic: "greptimedb_topic0".to_string(),
+                }),
+            ),
+            (1, WalOptions::RaftEngine),
+            (2, WalOptions::Noop),
+            (
+                3,
+                WalOptions::Kafka(KafkaWalOptions {
+                    topic: "greptimedb_topic1".to_string(),
+                }),
+            ),
+        ])
+    }
+
     #[tokio::test]
     async fn test_raft_engine_topic_region_map() {
         let mem_kv = Arc::new(MemoryKvBackend::default());
@@ -1773,6 +1949,46 @@ mod tests {
         assert_eq!(region_routes[0].follower_peers[0].addr, "new-a3");
     }
 
+    #[tokio::test]
+    async fn test_get_full_table_info_with_read_only_kv_backend() {
+        let mem_kv = Arc::new(MemoryKvBackend::default());
+        let writable_manager = TableMetadataManager::new(mem_kv.clone());
+
+        let region_routes = vec![new_test_region_route()];
+        let table_info = new_test_table_info();
+        let table_id = table_info.ident.table_id;
+
+        create_physical_table_metadata(
+            &writable_manager,
+            table_info.clone(),
+            region_routes.clone(),
+            HashMap::new(),
+        )
+        .await
+        .unwrap();
+
+        let read_only_kv = Arc::new(ReadOnlyKvBackend::new(mem_kv));
+        let read_only_manager = TableMetadataManager::new(read_only_kv);
+
+        let (remote_table_info, remote_table_route) = read_only_manager
+            .get_full_table_info(table_id)
+            .await
+            .unwrap();
+
+        assert_eq!(
+            remote_table_info.unwrap().into_inner().table_info,
+            table_info
+        );
+        assert_eq!(
+            remote_table_route
+                .unwrap()
+                .into_inner()
+                .region_routes()
+                .unwrap(),
+            &region_routes
+        );
+    }
+
     #[tokio::test]
     async fn test_create_logic_tables_metadata() {
         let mem_kv = Arc::new(MemoryKvBackend::default());
@@ -2551,7 +2767,7 @@ mod tests {
         let table_id = 1025;
         let table_name = "foo";
         let task = test_create_table_task(table_name, table_id);
-        let options = create_mock_region_wal_options();
+        let options = create_mixed_region_wal_options();
         let serialized_options = options
             .iter()
             .map(|(k, v)| (*k, serde_json::to_string(v).unwrap()))
@@ -2611,7 +2827,7 @@ mod tests {
         let table_id = 1025;
         let table_name = "foo";
         let task = test_create_table_task(table_name, table_id);
-        let options = create_mock_region_wal_options();
+        let options = create_mixed_region_wal_options();
         let serialized_options = options
             .iter()
             .map(|(k, v)| (*k, serde_json::to_string(v).unwrap()))
@@ -2679,6 +2895,287 @@ mod tests {
         assert_eq!(kvs, expected_result);
     }
 
+    #[tokio::test]
+    async fn test_dropped_table_metadata_enumeration_and_lookup() {
+        let mem_kv = Arc::new(MemoryKvBackend::default());
+        let table_metadata_manager = TableMetadataManager::new(mem_kv.clone());
+        let table_id = 1025;
+        let table_name = "foo";
+        let task = test_create_table_task(table_name, table_id);
+        let table_info = task.table_info.clone();
+        let options = create_mixed_region_wal_options();
+        let serialized_options = options
+            .iter()
+            .map(|(k, v)| (*k, serde_json::to_string(v).unwrap()))
+            .collect::<HashMap<_, _>>();
+        table_metadata_manager
+            .create_table_metadata(
+                table_info.clone(),
+                TableRouteValue::physical(vec![
+                    RegionRoute {
+                        region: Region::new_test(RegionId::new(table_id, 1)),
+                        leader_peer: Some(Peer::empty(1)),
+                        follower_peers: vec![Peer::empty(5)],
+                        leader_state: None,
+                        leader_down_since: None,
+                        write_route_policy: None,
+                    },
+                    RegionRoute {
+                        region: Region::new_test(RegionId::new(table_id, 2)),
+                        leader_peer: Some(Peer::empty(2)),
+                        follower_peers: vec![Peer::empty(4)],
+                        leader_state: None,
+                        leader_down_since: None,
+                        write_route_policy: None,
+                    },
+                    RegionRoute {
+                        region: Region::new_test(RegionId::new(table_id, 3)),
+                        leader_peer: Some(Peer::empty(3)),
+                        follower_peers: vec![],
+                        leader_state: None,
+                        leader_down_since: None,
+                        write_route_policy: None,
+                    },
+                ]),
+                serialized_options,
+            )
+            .await
+            .unwrap();
+        let table_route_value = table_metadata_manager
+            .table_route_manager
+            .table_route_storage()
+            .get_with_raw_bytes(table_id)
+            .await
+            .unwrap()
+            .unwrap();
+        let region_routes = table_route_value.region_routes().unwrap();
+        let table_name = TableName::new(DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME, table_name);
+        let table_route_value = TableRouteValue::physical(region_routes.clone());
+
+        table_metadata_manager
+            .delete_table_metadata(table_id, &table_name, &table_route_value, &options)
+            .await
+            .unwrap();
+
+        let dropped_tables = table_metadata_manager.list_dropped_tables().await.unwrap();
+        assert_eq!(dropped_tables.len(), 1);
+        assert_eq!(dropped_tables[0].table_id, table_id);
+        assert_eq!(dropped_tables[0].table_name, table_name);
+
+        let dropped_table = table_metadata_manager
+            .get_dropped_table(&table_name)
+            .await
+            .unwrap()
+            .unwrap();
+        assert_eq!(dropped_table.table_id, table_id);
+        assert_eq!(dropped_table.table_name, table_name);
+        assert_eq!(dropped_table.table_info_value.table_info, table_info);
+        assert_eq!(
+            dropped_table.table_route_value.region_routes().unwrap(),
+            region_routes
+        );
+        assert_eq!(dropped_table.region_wal_options, options);
+
+        let dropped_table_by_id = table_metadata_manager
+            .get_dropped_table_by_id(table_id)
+            .await
+            .unwrap()
+            .unwrap();
+        assert_eq!(dropped_table_by_id.table_id, table_id);
+        assert_eq!(dropped_table_by_id.table_name, table_name);
+        assert_eq!(dropped_table_by_id.table_info_value.table_info, table_info);
+        assert_eq!(
+            dropped_table_by_id
+                .table_route_value
+                .region_routes()
+                .unwrap(),
+            region_routes
+        );
+        assert_eq!(dropped_table_by_id.region_wal_options, options);
+    }
+
+    #[tokio::test]
+    async fn test_dropped_table_lookup_survives_live_name_recreation() {
+        let mem_kv = Arc::new(MemoryKvBackend::default());
+        let table_metadata_manager = TableMetadataManager::new(mem_kv.clone());
+        let dropped_table_id = 1025;
+        let recreated_table_id = 1026;
+        let table_name = "foo";
+        let dropped_task = test_create_table_task(table_name, dropped_table_id);
+        let dropped_table_info = dropped_task.table_info.clone();
+        let options = create_mock_region_wal_options();
+        let serialized_options = options
+            .iter()
+            .map(|(k, v)| (*k, serde_json::to_string(v).unwrap()))
+            .collect::<HashMap<_, _>>();
+        table_metadata_manager
+            .create_table_metadata(
+                dropped_table_info.clone(),
+                TableRouteValue::physical(vec![
+                    RegionRoute {
+                        region: Region::new_test(RegionId::new(dropped_table_id, 1)),
+                        leader_peer: Some(Peer::empty(1)),
+                        follower_peers: vec![Peer::empty(5)],
+                        leader_state: None,
+                        leader_down_since: None,
+                        write_route_policy: None,
+                    },
+                    RegionRoute {
+                        region: Region::new_test(RegionId::new(dropped_table_id, 2)),
+                        leader_peer: Some(Peer::empty(2)),
+                        follower_peers: vec![Peer::empty(4)],
+                        leader_state: None,
+                        leader_down_since: None,
+                        write_route_policy: None,
+                    },
+                ]),
+                serialized_options.clone(),
+            )
+            .await
+            .unwrap();
+
+        let dropped_table_name =
+            TableName::new(DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME, table_name);
+        let dropped_table_route = table_metadata_manager
+            .table_route_manager
+            .table_route_storage()
+            .get_with_raw_bytes(dropped_table_id)
+            .await
+            .unwrap()
+            .unwrap();
+        let dropped_table_route =
+            TableRouteValue::physical(dropped_table_route.region_routes().unwrap().clone());
+        table_metadata_manager
+            .delete_table_metadata(
+                dropped_table_id,
+                &dropped_table_name,
+                &dropped_table_route,
+                &options,
+            )
+            .await
+            .unwrap();
+
+        let recreated_task = test_create_table_task(table_name, recreated_table_id);
+        table_metadata_manager
+            .create_table_metadata(
+                recreated_task.table_info,
+                TableRouteValue::physical(vec![RegionRoute {
+                    region: Region::new_test(RegionId::new(recreated_table_id, 1)),
+                    leader_peer: Some(Peer::empty(4)),
+                    follower_peers: vec![],
+                    leader_state: None,
+                    leader_down_since: None,
+                    write_route_policy: None,
+                }]),
+                serialized_options,
+            )
+            .await
+            .unwrap();
+
+        assert_eq!(
+            table_metadata_manager
+                .table_name_manager()
+                .get(TableNameKey::from(&dropped_table_name))
+                .await
+                .unwrap()
+                .unwrap()
+                .table_id(),
+            recreated_table_id
+        );
+
+        let dropped_table = table_metadata_manager
+            .get_dropped_table(&dropped_table_name)
+            .await
+            .unwrap()
+            .unwrap();
+        assert_eq!(dropped_table.table_id, dropped_table_id);
+        assert_eq!(dropped_table.table_name, dropped_table_name);
+        assert_eq!(
+            dropped_table.table_info_value.table_info,
+            dropped_table_info
+        );
+
+        let dropped_tables = table_metadata_manager.list_dropped_tables().await.unwrap();
+        assert_eq!(dropped_tables.len(), 1);
+        assert_eq!(dropped_tables[0].table_id, dropped_table_id);
+        assert_eq!(dropped_tables[0].table_name, dropped_table_name);
+    }
+
+    #[tokio::test]
+    async fn test_dropped_table_lookup_ignores_unrelated_malformed_datanode_tombstones() {
+        let mem_kv = Arc::new(MemoryKvBackend::default());
+        let table_metadata_manager = TableMetadataManager::new(mem_kv.clone());
+        let table_id = 1025;
+        let table_name = "foo";
+        let task = test_create_table_task(table_name, table_id);
+        let table_info = task.table_info.clone();
+        let options = create_mixed_region_wal_options();
+        let serialized_options = options
+            .iter()
+            .map(|(k, v)| (*k, serde_json::to_string(v).unwrap()))
+            .collect::<HashMap<_, _>>();
+        table_metadata_manager
+            .create_table_metadata(
+                table_info.clone(),
+                TableRouteValue::physical(vec![
+                    RegionRoute {
+                        region: Region::new_test(RegionId::new(table_id, 1)),
+                        leader_peer: Some(Peer::empty(1)),
+                        follower_peers: vec![Peer::empty(5)],
+                        leader_state: None,
+                        leader_down_since: None,
+                        write_route_policy: None,
+                    },
+                    RegionRoute {
+                        region: Region::new_test(RegionId::new(table_id, 2)),
+                        leader_peer: Some(Peer::empty(2)),
+                        follower_peers: vec![Peer::empty(4)],
+                        leader_state: None,
+                        leader_down_since: None,
+                        write_route_policy: None,
+                    },
+                ]),
+                serialized_options,
+            )
+            .await
+            .unwrap();
+
+        let table_route_value = table_metadata_manager
+            .table_route_manager
+            .table_route_storage()
+            .get_with_raw_bytes(table_id)
+            .await
+            .unwrap()
+            .unwrap();
+        let region_routes = table_route_value.region_routes().unwrap();
+        let table_name = TableName::new(DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME, table_name);
+        let table_route_value = TableRouteValue::physical(region_routes.clone());
+
+        table_metadata_manager
+            .delete_table_metadata(table_id, &table_name, &table_route_value, &options)
+            .await
+            .unwrap();
+
+        mem_kv
+            .put(
+                PutRequest::new()
+                    .with_key("__tombstone/__dn_table/not-a-datanode-table-key")
+                    .with_value("malformed"),
+            )
+            .await
+            .unwrap();
+
+        let dropped_table = table_metadata_manager
+            .get_dropped_table(&table_name)
+            .await
+            .unwrap()
+            .unwrap();
+        assert_eq!(dropped_table.table_id, table_id);
+        assert_eq!(dropped_table.table_name, table_name);
+        assert_eq!(dropped_table.table_info_value.table_info, table_info);
+        assert_eq!(dropped_table.region_wal_options, options);
+    }
+
     #[tokio::test]
     async fn test_create_update_view_info() {
         let mem_kv = Arc::new(MemoryKvBackend::default());
@@ -2752,7 +3249,7 @@ mod tests {
         let new_definition = "CREATE VIEW test AS SELECT * FROM b_table join c_table";
 
         let current_view_info_value = DeserializedValueWithBytes::from_inner(ViewInfoValue::new(
-            logical_plan.clone(),
+            logical_plan.clone().into(),
             table_names,
             columns,
             plan_columns,
@@ -2803,7 +3300,7 @@ mod tests {
         let wrong_definition = "wrong_definition";
         let wrong_view_info_value =
             DeserializedValueWithBytes::from_inner(current_view_info_value.update(
-                wrong_view_info,
+                wrong_view_info.into(),
                 new_table_names.clone(),
                 new_columns.clone(),
                 new_plan_columns.clone(),
diff --git a/src/common/meta/src/key/flow.rs b/src/common/meta/src/key/flow.rs
index 546071f2a0..d581b92685 100644
--- a/src/common/meta/src/key/flow.rs
+++ b/src/common/meta/src/key/flow.rs
@@ -19,6 +19,7 @@ pub mod flow_state;
 mod flownode_addr_helper;
 pub(crate) mod flownode_flow;
 pub(crate) mod table_flow;
+use std::collections::BTreeMap;
 use std::ops::Deref;
 use std::sync::Arc;
 
@@ -44,6 +45,48 @@ use crate::kv_backend::KvBackendRef;
 use crate::kv_backend::txn::Txn;
 use crate::rpc::store::BatchDeleteRequest;
 
+pub const FLOW_KEY_PREFIX: &str = "__flow";
+
+/// Returns a flow metadata key under the `__flow` scope.
+pub fn scoped_flow_key(inner_key: &str) -> String {
+    format!("{FLOW_KEY_PREFIX}/{inner_key}")
+}
+
+/// Returns a flow metadata key prefix under the `__flow` scope.
+pub fn scoped_flow_key_prefix(inner_prefix: &str) -> String {
+    scoped_flow_key(inner_prefix)
+}
+
+/// Returns the full prefix of flow info keys.
+pub fn flow_info_key_prefix() -> String {
+    scoped_flow_key_prefix(flow_info::FLOW_INFO_KEY_PREFIX)
+}
+
+/// Returns the full prefix of flow name keys.
+pub fn flow_name_key_prefix() -> String {
+    scoped_flow_key_prefix(flow_name::FLOW_NAME_KEY_PREFIX)
+}
+
+/// Returns the full prefix of flow route keys.
+pub fn flow_route_key_prefix() -> String {
+    scoped_flow_key_prefix(flow_route::FLOW_ROUTE_KEY_PREFIX)
+}
+
+/// Returns the full prefix of table flow keys.
+pub fn table_flow_key_prefix() -> String {
+    scoped_flow_key_prefix(table_flow::TABLE_FLOW_KEY_PREFIX)
+}
+
+/// Returns the full prefix of flownode flow keys.
+pub fn flownode_flow_key_prefix() -> String {
+    scoped_flow_key_prefix(flownode_flow::FLOWNODE_FLOW_KEY_PREFIX)
+}
+
+/// Returns the full key of flow state.
+pub fn flow_state_full_key() -> String {
+    scoped_flow_key(flow_state::FLOW_STATE_KEY)
+}
+
 /// The key of `__flow/` scope.
 #[derive(Debug, Clone, PartialEq)]
 pub struct FlowScoped<T> {
@@ -151,6 +194,22 @@ impl FlowMetadataManager {
         &self.table_flow_manager
     }
 
+    /// Returns a best-effort mapping from flow partition to flownode address.
+    pub async fn flownode_addrs(
+        &self,
+        flow_id: FlowId,
+    ) -> Result<BTreeMap<FlowPartitionId, String>> {
+        let routes = self.flow_route_manager.routes(flow_id).await?;
+
+        Ok(routes
+            .into_iter()
+            .filter_map(|(key, route)| {
+                let addr = route.peer.addr;
+                (!addr.is_empty()).then_some((key.partition_id(), addr))
+            })
+            .collect())
+    }
+
     /// Creates metadata for flow and returns an error if different metadata exists.
     pub async fn create_flow_metadata(
         &self,
@@ -390,7 +449,7 @@ impl std::fmt::Debug for FlowMetadataManager {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::BTreeMap;
     use std::sync::Arc;
 
@@ -400,10 +459,13 @@ mod tests {
 
     use super::*;
     use crate::FlownodeId;
-    use crate::key::FlowPartitionId;
     use crate::key::flow::table_flow::TableFlowKey;
+    use crate::key::node_address::{NodeAddressKey, NodeAddressValue};
+    use crate::key::{FlowPartitionId, MetadataValue};
+    use crate::kv_backend::KvBackend;
     use crate::kv_backend::memory::MemoryKvBackend;
     use crate::peer::Peer;
+    use crate::rpc::store::PutRequest;
 
     #[derive(Debug)]
     struct MockKey {
@@ -567,6 +629,83 @@ mod tests {
         }
     }
 
+    #[tokio::test]
+    async fn test_flownode_addrs_remaps_to_latest_address() {
+        let mem_kv = Arc::new(MemoryKvBackend::default());
+        let flow_metadata_manager = FlowMetadataManager::new(mem_kv.clone());
+        let flow_id = 10;
+        let flow_value = test_flow_info_value("flow", [(0, 1u64)].into(), vec![1024]);
+        let flow_routes = vec![(
+            0u32,
+            FlowRouteValue {
+                peer: Peer::new(1, "old-addr"),
+            },
+        )];
+
+        flow_metadata_manager
+            .create_flow_metadata(flow_id, flow_value, flow_routes)
+            .await
+            .unwrap();
+
+        mem_kv
+            .put(PutRequest {
+                key: NodeAddressKey::with_flownode(1).to_bytes(),
+                value: NodeAddressValue::new(Peer::new(1, "new-addr"))
+                    .try_as_raw_value()
+                    .unwrap(),
+                ..Default::default()
+            })
+            .await
+            .unwrap();
+
+        let addrs = flow_metadata_manager.flownode_addrs(flow_id).await.unwrap();
+        assert_eq!(addrs, BTreeMap::from([(0, "new-addr".to_string())]));
+    }
+
+    #[tokio::test]
+    async fn test_flownode_addrs_falls_back_to_route_address() {
+        let mem_kv = Arc::new(MemoryKvBackend::default());
+        let flow_metadata_manager = FlowMetadataManager::new(mem_kv);
+        let flow_id = 10;
+        let flow_value = test_flow_info_value("flow", [(0, 1u64)].into(), vec![1024]);
+        let flow_routes = vec![(
+            0u32,
+            FlowRouteValue {
+                peer: Peer::new(1, "route-addr"),
+            },
+        )];
+
+        flow_metadata_manager
+            .create_flow_metadata(flow_id, flow_value, flow_routes)
+            .await
+            .unwrap();
+
+        let addrs = flow_metadata_manager.flownode_addrs(flow_id).await.unwrap();
+        assert_eq!(addrs, BTreeMap::from([(0, "route-addr".to_string())]));
+    }
+
+    #[tokio::test]
+    async fn test_flownode_addrs_skips_empty_addresses() {
+        let mem_kv = Arc::new(MemoryKvBackend::default());
+        let flow_metadata_manager = FlowMetadataManager::new(mem_kv);
+        let flow_id = 10;
+        let flow_value = test_flow_info_value("flow", [(0, 1u64)].into(), vec![1024]);
+        let flow_routes = vec![(
+            0u32,
+            FlowRouteValue {
+                peer: Peer::empty(1),
+            },
+        )];
+
+        flow_metadata_manager
+            .create_flow_metadata(flow_id, flow_value, flow_routes)
+            .await
+            .unwrap();
+
+        let addrs = flow_metadata_manager.flownode_addrs(flow_id).await.unwrap();
+        assert!(addrs.is_empty());
+    }
+
     #[tokio::test]
     async fn test_create_flow_metadata_flow_exists_err() {
         let mem_kv = Arc::new(MemoryKvBackend::default());
diff --git a/src/common/meta/src/key/flow/flow_info.rs b/src/common/meta/src/key/flow/flow_info.rs
index 7b277e36ef..d501822c3c 100644
--- a/src/common/meta/src/key/flow/flow_info.rs
+++ b/src/common/meta/src/key/flow/flow_info.rs
@@ -31,7 +31,7 @@ use crate::key::{DeserializedValueWithBytes, FlowId, FlowPartitionId, MetadataKe
 use crate::kv_backend::KvBackendRef;
 use crate::kv_backend::txn::{Compare, CompareOp, Txn, TxnOp};
 
-const FLOW_INFO_KEY_PREFIX: &str = "info";
+pub const FLOW_INFO_KEY_PREFIX: &str = "info";
 
 lazy_static! {
     static ref FLOW_INFO_KEY_PATTERN: Regex =
diff --git a/src/common/meta/src/key/flow/flow_name.rs b/src/common/meta/src/key/flow/flow_name.rs
index d0967d2b8a..76a1b58d6f 100644
--- a/src/common/meta/src/key/flow/flow_name.rs
+++ b/src/common/meta/src/key/flow/flow_name.rs
@@ -31,7 +31,7 @@ use crate::range_stream::{DEFAULT_PAGE_SIZE, PaginationStream};
 use crate::rpc::KeyValue;
 use crate::rpc::store::RangeRequest;
 
-const FLOW_NAME_KEY_PREFIX: &str = "name";
+pub const FLOW_NAME_KEY_PREFIX: &str = "name";
 
 lazy_static! {
     static ref FLOW_NAME_KEY_PATTERN: Regex = Regex::new(&format!(
diff --git a/src/common/meta/src/key/flow/flow_route.rs b/src/common/meta/src/key/flow/flow_route.rs
index ed82151b44..d402e1acef 100644
--- a/src/common/meta/src/key/flow/flow_route.rs
+++ b/src/common/meta/src/key/flow/flow_route.rs
@@ -32,7 +32,7 @@ use crate::range_stream::{DEFAULT_PAGE_SIZE, PaginationStream};
 use crate::rpc::KeyValue;
 use crate::rpc::store::RangeRequest;
 
-const FLOW_ROUTE_KEY_PREFIX: &str = "route";
+pub const FLOW_ROUTE_KEY_PREFIX: &str = "route";
 
 lazy_static! {
     static ref FLOW_ROUTE_KEY_PATTERN: Regex =
diff --git a/src/common/meta/src/key/flow/flow_state.rs b/src/common/meta/src/key/flow/flow_state.rs
index 12db46dfb0..1b161929fe 100644
--- a/src/common/meta/src/key/flow/flow_state.rs
+++ b/src/common/meta/src/key/flow/flow_state.rs
@@ -24,7 +24,7 @@ use crate::kv_backend::KvBackendRef;
 use crate::rpc::store::PutRequest;
 
 /// The entire FlowId to Flow Size's Map is stored directly in the value part of the key.
-const FLOW_STATE_KEY: &str = "state";
+pub const FLOW_STATE_KEY: &str = "state";
 
 /// The key of flow state.
 #[derive(Debug, Clone, Copy, PartialEq)]
diff --git a/src/common/meta/src/key/flow/flownode_flow.rs b/src/common/meta/src/key/flow/flownode_flow.rs
index 38ab07457e..2ba690c167 100644
--- a/src/common/meta/src/key/flow/flownode_flow.rs
+++ b/src/common/meta/src/key/flow/flownode_flow.rs
@@ -36,7 +36,7 @@ lazy_static! {
     .unwrap();
 }
 
-const FLOWNODE_FLOW_KEY_PREFIX: &str = "flownode";
+pub const FLOWNODE_FLOW_KEY_PREFIX: &str = "flownode";
 
 /// The key of mapping [FlownodeId] to [FlowId].
 ///
diff --git a/src/common/meta/src/key/flow/table_flow.rs b/src/common/meta/src/key/flow/table_flow.rs
index a35d6178ae..d334ca0218 100644
--- a/src/common/meta/src/key/flow/table_flow.rs
+++ b/src/common/meta/src/key/flow/table_flow.rs
@@ -36,7 +36,7 @@ use crate::range_stream::{DEFAULT_PAGE_SIZE, PaginationStream};
 use crate::rpc::KeyValue;
 use crate::rpc::store::RangeRequest;
 
-const TABLE_FLOW_KEY_PREFIX: &str = "source_table";
+pub const TABLE_FLOW_KEY_PREFIX: &str = "source_table";
 
 lazy_static! {
     static ref TABLE_FLOW_KEY_PATTERN: Regex = Regex::new(&format!(
diff --git a/src/common/meta/src/key/table_route.rs b/src/common/meta/src/key/table_route.rs
index a409b75fce..4c618addda 100644
--- a/src/common/meta/src/key/table_route.rs
+++ b/src/common/meta/src/key/table_route.rs
@@ -675,7 +675,7 @@ impl TableRouteStorage {
     pub async fn get(&self, table_id: TableId) -> Result<Option<TableRouteValue>> {
         let mut table_route = self.get_inner(table_id).await?;
         if let Some(table_route) = &mut table_route {
-            self.remap_route_address(table_route).await?;
+            self.remap_table_route(table_route).await?;
         };
 
         Ok(table_route)
@@ -697,7 +697,7 @@ impl TableRouteStorage {
     ) -> Result<Option<DeserializedValueWithBytes<TableRouteValue>>> {
         let mut table_route = self.get_with_raw_bytes_inner(table_id).await?;
         if let Some(table_route) = &mut table_route {
-            self.remap_route_address(table_route).await?;
+            self.remap_table_route(table_route).await?;
         };
 
         Ok(table_route)
@@ -791,10 +791,7 @@ impl TableRouteStorage {
         Ok(())
     }
 
-    pub(crate) async fn remap_route_address(
-        &self,
-        table_route: &mut TableRouteValue,
-    ) -> Result<()> {
+    pub(crate) async fn remap_table_route(&self, table_route: &mut TableRouteValue) -> Result<()> {
         let keys = extract_address_keys(table_route).into_iter().collect();
         let node_addrs = self.get_node_addresses(keys).await?;
         set_addresses(&node_addrs, table_route)?;
@@ -802,6 +799,17 @@ impl TableRouteStorage {
         Ok(())
     }
 
+    pub(crate) async fn remap_region_routes(
+        &self,
+        region_routes: &mut [RegionRoute],
+    ) -> Result<()> {
+        let keys = extract_address_keys_from_region_routes(region_routes)
+            .into_iter()
+            .collect();
+        let node_addrs = self.get_node_addresses(keys).await?;
+        set_addresses_for_region_routes(&node_addrs, region_routes)
+    }
+
     async fn get_node_addresses(
         &self,
         keys: Vec<Vec<u8>>,
@@ -824,15 +832,11 @@ impl TableRouteStorage {
     }
 }
 
-fn set_addresses(
+fn set_addresses_for_region_routes(
     node_addrs: &HashMap<u64, NodeAddressValue>,
-    table_route: &mut TableRouteValue,
+    region_routes: &mut [RegionRoute],
 ) -> Result<()> {
-    let TableRouteValue::Physical(physical_table_route) = table_route else {
-        return Ok(());
-    };
-
-    for region_route in &mut physical_table_route.region_routes {
+    for region_route in region_routes {
         if let Some(leader) = &mut region_route.leader_peer
             && let Some(node_addr) = node_addrs.get(&leader.id)
         {
@@ -848,13 +852,18 @@ fn set_addresses(
     Ok(())
 }
 
-fn extract_address_keys(table_route: &TableRouteValue) -> HashSet<Vec<u8>> {
+fn set_addresses(
+    node_addrs: &HashMap<u64, NodeAddressValue>,
+    table_route: &mut TableRouteValue,
+) -> Result<()> {
     let TableRouteValue::Physical(physical_table_route) = table_route else {
-        return HashSet::default();
+        return Ok(());
     };
+    set_addresses_for_region_routes(node_addrs, &mut physical_table_route.region_routes)
+}
 
-    physical_table_route
-        .region_routes
+fn extract_address_keys_from_region_routes(region_routes: &[RegionRoute]) -> HashSet<Vec<u8>> {
+    region_routes
         .iter()
         .flat_map(|region_route| {
             region_route
@@ -871,6 +880,14 @@ fn extract_address_keys(table_route: &TableRouteValue) -> HashSet<Vec<u8>> {
         .collect()
 }
 
+fn extract_address_keys(table_route: &TableRouteValue) -> HashSet<Vec<u8>> {
+    let TableRouteValue::Physical(physical_table_route) = table_route else {
+        return HashSet::default();
+    };
+
+    extract_address_keys_from_region_routes(&physical_table_route.region_routes)
+}
+
 #[cfg(test)]
 mod tests {
     use std::sync::Arc;
@@ -1104,7 +1121,7 @@ mod tests {
         .unwrap();
 
         table_route_storage
-            .remap_route_address(&mut table_route)
+            .remap_table_route(&mut table_route)
             .await
             .unwrap();
 
diff --git a/src/common/meta/src/key/tombstone.rs b/src/common/meta/src/key/tombstone.rs
index ed1d2f28c2..a6cb6efa89 100644
--- a/src/common/meta/src/key/tombstone.rs
+++ b/src/common/meta/src/key/tombstone.rs
@@ -15,17 +15,41 @@
 use std::collections::HashMap;
 
 use common_telemetry::debug;
+use futures_util::stream::BoxStream;
 use snafu::ensure;
 
 use crate::error::{self, Result};
+use crate::key::TABLE_NAME_KEY_PREFIX;
 use crate::key::txn_helper::TxnOpGetResponseSet;
 use crate::kv_backend::KvBackendRef;
 use crate::kv_backend::txn::{Compare, CompareOp, Txn, TxnOp};
-use crate::rpc::store::{BatchDeleteRequest, BatchGetRequest};
+use crate::range_stream::{DEFAULT_PAGE_SIZE, PaginationStream};
+use crate::rpc::KeyValue;
+use crate::rpc::store::{BatchDeleteRequest, BatchGetRequest, RangeRequest};
 
 /// [TombstoneManager] provides the ability to:
 /// - logically delete values
 /// - restore the deleted values
+///
+/// The tombstone mechanism is primarily used for table metadata deletion.
+/// When a table is logically deleted, its associated metadata keys are moved
+/// to a tombstone area by prepending a prefix (default is `__tombstone/`).
+///
+/// All involved keys in the tombstone mechanism:
+/// - `__table_name/{catalog}/{schema}/{table_name}`: Maps table name to table ID.
+/// - `__table_info/{table_id}`: Stores table information/metadata.
+/// - `__table_route/{table_id}`: Stores table routing information.
+/// - `__table_repart/{table_id}`: Stores table repartition information.
+/// - `__dn_table/{datanode_id}/{table_id}`: Maps datanodes to table regions.
+/// - `__topic_region/{topic_name}/{region_id}`: Maps regions to WAL topics.
+///
+/// These keys are moved to:
+/// - `__tombstone/__table_name/...`
+/// - `__tombstone/__table_info/...`
+/// - `__tombstone/__table_route/...`
+/// - `__tombstone/__table_repart/...`
+/// - `__tombstone/__dn_table/...`
+/// - `__tombstone/__topic_region/...`
 pub struct TombstoneManager {
     kv_backend: KvBackendRef,
     tombstone_prefix: String,
@@ -56,6 +80,66 @@ impl TombstoneManager {
         [self.tombstone_prefix.as_bytes(), key].concat()
     }
 
+    /// Removes the tombstone prefix from a tombstoned key.
+    pub fn strip_tombstone_prefix<'a>(&self, tombstone_key: &'a [u8]) -> Result<&'a [u8]> {
+        ensure!(
+            tombstone_key.starts_with(self.tombstone_prefix.as_bytes()),
+            error::UnexpectedSnafu {
+                err_msg: format!(
+                    "The key '{}' does not start with tombstone prefix '{}'.",
+                    String::from_utf8_lossy(tombstone_key),
+                    self.tombstone_prefix
+                ),
+            }
+        );
+
+        Ok(&tombstone_key[self.tombstone_prefix.len()..])
+    }
+
+    /// Gets a single tombstoned value by its original key.
+    pub async fn get(&self, key: &[u8]) -> Result<Option<KeyValue>> {
+        let tombstone_key = self.to_tombstone(key);
+        self.kv_backend.get(&tombstone_key).await
+    }
+
+    /// Gets tombstoned values by their original keys.
+    pub async fn batch_get(&self, keys: &[Vec<u8>]) -> Result<HashMap<Vec<u8>, KeyValue>> {
+        let tombstone_keys = keys
+            .iter()
+            .map(|key| self.to_tombstone(key))
+            .collect::<Vec<_>>();
+        let resp = self
+            .kv_backend
+            .batch_get(BatchGetRequest::new().with_keys(tombstone_keys))
+            .await?;
+
+        resp.kvs
+            .into_iter()
+            .map(|kv| Ok((self.strip_tombstone_prefix(&kv.key)?.to_vec(), kv)))
+            .collect::<Result<HashMap<_, _>>>()
+    }
+
+    /// Streams all tombstoned key-value pairs.
+    pub fn tombstones(&self) -> BoxStream<'static, Result<KeyValue>> {
+        self.scan_prefix(self.tombstone_prefix.as_bytes().to_vec())
+    }
+
+    /// Streams tombstoned table-name entries only.
+    pub fn tombstoned_table_names(&self) -> BoxStream<'static, Result<KeyValue>> {
+        self.scan_prefix(
+            format!("{}{}/", self.tombstone_prefix, TABLE_NAME_KEY_PREFIX).into_bytes(),
+        )
+    }
+
+    /// Streams tombstoned entries under the provided prefix.
+    fn scan_prefix(&self, prefix: Vec<u8>) -> BoxStream<'static, Result<KeyValue>> {
+        let req = RangeRequest::new().with_prefix(prefix);
+        let stream = PaginationStream::new(self.kv_backend.clone(), req, DEFAULT_PAGE_SIZE, Ok)
+            .into_stream();
+
+        Box::pin(stream)
+    }
+
     #[cfg(test)]
     pub fn set_max_txn_ops(&mut self, max_txn_ops: usize) {
         self.max_txn_ops = Some(max_txn_ops);
@@ -401,6 +485,35 @@ mod tests {
         assert!(kv_backend.is_empty());
     }
 
+    #[tokio::test]
+    async fn test_batch_get_tombstones() {
+        let kv_backend = Arc::new(MemoryKvBackend::default());
+        let tombstone_manager = TombstoneManager::new(kv_backend.clone());
+        kv_backend
+            .put(PutRequest::new().with_key("bar").with_value("baz"))
+            .await
+            .unwrap();
+        kv_backend
+            .put(PutRequest::new().with_key("foo").with_value("hi"))
+            .await
+            .unwrap();
+
+        tombstone_manager
+            .create(vec![b"bar".to_vec(), b"foo".to_vec()])
+            .await
+            .unwrap();
+
+        let kvs = tombstone_manager
+            .batch_get(&[b"bar".to_vec(), b"foo".to_vec(), b"missing".to_vec()])
+            .await
+            .unwrap();
+
+        assert_eq!(kvs.len(), 2);
+        assert_eq!(kvs.get(b"bar".as_slice()).unwrap().value, b"baz");
+        assert_eq!(kvs.get(b"foo".as_slice()).unwrap().value, b"hi");
+        assert!(!kvs.contains_key(b"missing".as_slice()));
+    }
+
     #[tokio::test]
     async fn test_move_values() {
         let kv_backend = Arc::new(MemoryKvBackend::default());
diff --git a/src/common/meta/src/key/topic_name.rs b/src/common/meta/src/key/topic_name.rs
index 5497fbe478..99ae631a72 100644
--- a/src/common/meta/src/key/topic_name.rs
+++ b/src/common/meta/src/key/topic_name.rs
@@ -237,7 +237,7 @@ impl TopicNameManager {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::sync::Arc;
 
     use super::*;
diff --git a/src/common/meta/src/key/view_info.rs b/src/common/meta/src/key/view_info.rs
index 82be00b26d..5efc2c9bb7 100644
--- a/src/common/meta/src/key/view_info.rs
+++ b/src/common/meta/src/key/view_info.rs
@@ -16,6 +16,7 @@ use std::collections::{HashMap, HashSet};
 use std::fmt::Display;
 use std::sync::Arc;
 
+use common_base::bytes::Bytes;
 use serde::{Deserialize, Serialize};
 use snafu::OptionExt;
 use table::metadata::TableId;
@@ -31,9 +32,6 @@ use crate::kv_backend::KvBackendRef;
 use crate::kv_backend::txn::Txn;
 use crate::rpc::store::BatchGetRequest;
 
-/// The VIEW logical plan encoded bytes
-type RawViewLogicalPlan = Vec<u8>;
-
 /// The key stores the metadata of the view.
 ///
 /// The layout: `__view_info/{view_id}`.
@@ -86,7 +84,7 @@ impl MetadataKey<'_, ViewInfoKey> for ViewInfoKey {
 #[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
 pub struct ViewInfoValue {
     // The encoded logical plan
-    pub view_info: RawViewLogicalPlan,
+    pub view_info: Bytes,
     // The resolved fully table names in logical plan
     pub table_names: HashSet<TableName>,
     // The view columns
@@ -100,7 +98,7 @@ pub struct ViewInfoValue {
 
 impl ViewInfoValue {
     pub fn new(
-        view_info: RawViewLogicalPlan,
+        view_info: Bytes,
         table_names: HashSet<TableName>,
         columns: Vec<String>,
         plan_columns: Vec<String>,
@@ -118,7 +116,7 @@ impl ViewInfoValue {
 
     pub(crate) fn update(
         &self,
-        new_view_info: RawViewLogicalPlan,
+        new_view_info: Bytes,
         table_names: HashSet<TableName>,
         columns: Vec<String>,
         plan_columns: Vec<String>,
@@ -305,7 +303,7 @@ mod tests {
         };
 
         let value = ViewInfoValue {
-            view_info: vec![1, 2, 3],
+            view_info: Bytes::from([1, 2, 3].as_ref()),
             version: 1,
             table_names,
             columns: vec!["a".to_string()],
@@ -316,4 +314,53 @@ mod tests {
         let deserialized = ViewInfoValue::try_from_raw_value(&serialized).unwrap();
         assert_eq!(value, deserialized);
     }
+
+    #[test]
+    fn test_deserialize_view_info_value_with_vec_u8() {
+        #[derive(Serialize)]
+        struct OldViewInfoValue {
+            view_info: Vec<u8>,
+            table_names: HashSet<TableName>,
+            columns: Vec<String>,
+            plan_columns: Vec<String>,
+            definition: String,
+            version: u64,
+        }
+
+        let table_names = {
+            let mut set = HashSet::new();
+            set.insert(TableName {
+                catalog_name: "greptime".to_string(),
+                schema_name: "public".to_string(),
+                table_name: "a_table".to_string(),
+            });
+            set.insert(TableName {
+                catalog_name: "greptime".to_string(),
+                schema_name: "public".to_string(),
+                table_name: "b_table".to_string(),
+            });
+            set
+        };
+
+        let old_value = OldViewInfoValue {
+            view_info: vec![1, 2, 3],
+            table_names: table_names.clone(),
+            columns: vec!["a".to_string()],
+            plan_columns: vec!["number".to_string()],
+            definition: "CREATE VIEW test AS SELECT * FROM numbers".to_string(),
+            version: 1,
+        };
+
+        let serialized = serde_json::to_vec(&old_value).unwrap();
+        let deserialized = ViewInfoValue::try_from_raw_value(&serialized).unwrap();
+
+        assert_eq!(deserialized.view_info, vec![1, 2, 3]);
+        assert_eq!(deserialized.table_names, table_names);
+        assert_eq!(deserialized.columns, vec!["a".to_string()]);
+        assert_eq!(deserialized.plan_columns, vec!["number".to_string()]);
+        assert_eq!(
+            deserialized.definition,
+            "CREATE VIEW test AS SELECT * FROM numbers"
+        );
+    }
 }
diff --git a/src/common/meta/src/kv_backend.rs b/src/common/meta/src/kv_backend.rs
index 7f747508d4..10e5157850 100644
--- a/src/common/meta/src/kv_backend.rs
+++ b/src/common/meta/src/kv_backend.rs
@@ -33,6 +33,7 @@ pub mod etcd;
 pub mod memory;
 #[cfg(any(feature = "mysql_kvbackend", feature = "pg_kvbackend"))]
 pub mod rds;
+pub mod read_only;
 pub mod test;
 #[cfg(any(test, feature = "testing"))]
 pub mod test_util;
diff --git a/src/common/meta/src/kv_backend/rds.rs b/src/common/meta/src/kv_backend/rds.rs
index 8acab1eb65..1e94741dd8 100644
--- a/src/common/meta/src/kv_backend/rds.rs
+++ b/src/common/meta/src/kv_backend/rds.rs
@@ -15,12 +15,14 @@
 use std::any::Any;
 use std::collections::HashMap;
 use std::marker::PhantomData;
+#[cfg(any(feature = "pg_kvbackend", feature = "mysql_kvbackend"))]
+use std::sync::OnceLock;
 use std::time::Duration;
 
 use backon::{BackoffBuilder, ExponentialBuilder};
-use common_telemetry::debug;
+use common_telemetry::{debug, info};
 
-use crate::error::{Error, RdsTransactionRetryFailedSnafu, Result};
+use crate::error::{Error, RdsTransactionRetryFailedSnafu, Result, UnexpectedSnafu};
 use crate::kv_backend::txn::{
     Compare, Txn as KvTxn, TxnOp, TxnOpResponse, TxnResponse as KvTxnResponse,
 };
@@ -51,6 +53,38 @@ pub use mysql::MySqlStore;
 
 const RDS_STORE_TXN_RETRY_COUNT: usize = 3;
 
+#[cfg(any(feature = "pg_kvbackend", feature = "mysql_kvbackend"))]
+static RUSTLS_CRYPTO_PROVIDER_INIT: OnceLock<std::result::Result<(), String>> = OnceLock::new();
+
+#[cfg(any(feature = "pg_kvbackend", feature = "mysql_kvbackend"))]
+pub(crate) fn ensure_rustls_crypto_provider_installed() -> Result<()> {
+    RUSTLS_CRYPTO_PROVIDER_INIT
+        .get_or_init(|| {
+            if rustls::crypto::CryptoProvider::get_default().is_some() {
+                return Ok(());
+            }
+
+            match rustls::crypto::CryptoProvider::install_default(
+                rustls::crypto::aws_lc_rs::default_provider(),
+            ) {
+                Ok(()) => Ok(()),
+                Err(_provider) if rustls::crypto::CryptoProvider::get_default().is_some() => {
+                    Ok(())
+                }
+                Err(provider) => Err(format!(
+                    "Failed to install rustls CryptoProvider, existing default: {:?}, attempted provider: {:?}",
+                    rustls::crypto::CryptoProvider::get_default(),
+                    provider
+                )),
+            }
+        })
+        .clone()
+        .map_err(|err_msg| {
+            info!("Failed to install rustls crypto provider: {err_msg}");
+            UnexpectedSnafu { err_msg }.build()
+        })
+}
+
 /// Query executor for rds. It can execute queries or generate a transaction executor.
 #[async_trait::async_trait]
 pub trait Executor: Send + Sync {
@@ -575,12 +609,12 @@ macro_rules! record_rds_sql_execute_elapsed {
             .inspect(|_| {
                 $crate::metrics::RDS_SQL_EXECUTE_ELAPSED
                     .with_label_values(&[$label_store, "success", $label_op, $label_type])
-                    .observe(timer.elapsed().as_millis_f64())
+                    .observe(timer.elapsed().as_millis() as f64)
             })
             .inspect_err(|_| {
                 $crate::metrics::RDS_SQL_EXECUTE_ELAPSED
                     .with_label_values(&[$label_store, "error", $label_op, $label_type])
-                    .observe(timer.elapsed().as_millis_f64());
+                    .observe(timer.elapsed().as_millis() as f64);
             })
     }};
 }
diff --git a/src/common/meta/src/kv_backend/rds/mysql.rs b/src/common/meta/src/kv_backend/rds/mysql.rs
index ee776239bf..6791d58f93 100644
--- a/src/common/meta/src/kv_backend/rds/mysql.rs
+++ b/src/common/meta/src/kv_backend/rds/mysql.rs
@@ -15,19 +15,24 @@
 use std::marker::PhantomData;
 use std::sync::Arc;
 
-use common_telemetry::debug;
-use snafu::ResultExt;
+use common_telemetry::{debug, info, warn};
+use lazy_static::lazy_static;
+use regex::Regex;
+use snafu::{OptionExt, ResultExt};
 use sqlx::mysql::MySqlRow;
 use sqlx::pool::Pool;
 use sqlx::{MySql, MySqlPool, Row, Transaction as MySqlTransaction};
 use strum::AsRefStr;
 
-use crate::error::{CreateMySqlPoolSnafu, MySqlExecutionSnafu, MySqlTransactionSnafu, Result};
+use crate::error::{
+    CreateMySqlPoolSnafu, MySqlExecutionSnafu, MySqlTransactionSnafu, Result, UnexpectedSnafu,
+};
 use crate::kv_backend::KvBackendRef;
 use crate::kv_backend::rds::{
     Executor, ExecutorFactory, ExecutorImpl, KvQueryExecutor, RDS_STORE_OP_BATCH_DELETE,
     RDS_STORE_OP_BATCH_GET, RDS_STORE_OP_BATCH_PUT, RDS_STORE_OP_RANGE_DELETE,
     RDS_STORE_OP_RANGE_QUERY, RDS_STORE_TXN_RETRY_COUNT, RdsStore, Transaction,
+    ensure_rustls_crypto_provider_installed,
 };
 use crate::rpc::KeyValue;
 use crate::rpc::store::{
@@ -37,6 +42,18 @@ use crate::rpc::store::{
 
 const MYSQL_STORE_NAME: &str = "mysql_store";
 
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum ValueBlobType {
+    Blob,
+    MediumBlob,
+    LongBlob,
+}
+
+lazy_static! {
+    static ref VALUE_COLUMN_BLOB_TYPE_RE: Regex =
+        Regex::new(r#"(?i)(?:\(|,)\s*[`"]?v[`"]?\s+(longblob|mediumblob|blob)\b"#).unwrap();
+}
+
 type MySqlClient = Arc<Pool<MySql>>;
 pub struct MySqlTxnClient(MySqlTransaction<'static, MySql>);
 
@@ -161,7 +178,11 @@ impl<'a> MySqlTemplateFactory<'a> {
             table_name: table_name.to_string(),
             create_table_statement: format!(
                 // Cannot be more than 3072 bytes in PRIMARY KEY
-                "CREATE TABLE IF NOT EXISTS `{table_name}`(k VARBINARY(3072) PRIMARY KEY, v BLOB);",
+                "CREATE TABLE IF NOT EXISTS `{table_name}`(k VARBINARY(3072) PRIMARY KEY, v MEDIUMBLOB);",
+            ),
+            show_create_table_statement: format!("SHOW CREATE TABLE `{table_name}`"),
+            alter_value_column_statement: format!(
+                "ALTER TABLE `{table_name}` MODIFY COLUMN v MEDIUMBLOB;"
             ),
             range_template: RangeTemplate {
                 point: format!("SELECT k, v FROM `{table_name}` WHERE k = ?"),
@@ -186,6 +207,8 @@ impl<'a> MySqlTemplateFactory<'a> {
 pub struct MySqlTemplateSet {
     table_name: String,
     create_table_statement: String,
+    show_create_table_statement: String,
+    alter_value_column_statement: String,
     range_template: RangeTemplate,
     delete_template: RangeTemplate,
 }
@@ -534,8 +557,71 @@ impl KvQueryExecutor<MySqlClient> for MySqlStore {
 }
 
 impl MySqlStore {
+    /// Reads the current table definition for best-effort schema upgrades.
+    async fn fetch_create_table_sql(
+        pool: &Pool<MySql>,
+        sql_template_set: &MySqlTemplateSet,
+    ) -> Result<Option<String>> {
+        let row = sqlx::query(&sql_template_set.show_create_table_statement)
+            .fetch_optional(pool)
+            .await
+            .with_context(|_| MySqlExecutionSnafu {
+                sql: sql_template_set.show_create_table_statement.clone(),
+            })?;
+        Ok(row.map(|row| row.get(1)))
+    }
+
+    /// Parses the blob type of the `v` column from `SHOW CREATE TABLE` output.
+    fn parse_value_column_blob_type(create_table_sql: &str) -> Option<ValueBlobType> {
+        // `SHOW CREATE TABLE` returns MySQL-specific DDL. A minimal parser keeps the
+        // upgrade check small and avoids introducing a SQL parser just for one column.
+        let captures = VALUE_COLUMN_BLOB_TYPE_RE.captures(create_table_sql)?;
+        match captures.get(1)?.as_str().to_ascii_lowercase().as_str() {
+            "blob" => Some(ValueBlobType::Blob),
+            "mediumblob" => Some(ValueBlobType::MediumBlob),
+            "longblob" => Some(ValueBlobType::LongBlob),
+            _ => None,
+        }
+    }
+
+    /// Upgrades the metadata value column to `MEDIUMBLOB` when an old table still uses `BLOB`.
+    async fn maybe_upgrade_value_column_to_mediumblob(
+        pool: &Pool<MySql>,
+        sql_template_set: &MySqlTemplateSet,
+    ) -> Result<()> {
+        let table_name = &sql_template_set.table_name;
+        let create_table_sql = Self::fetch_create_table_sql(pool, sql_template_set)
+            .await?
+            .context(UnexpectedSnafu {
+                err_msg: format!("Failed to fetch CREATE TABLE SQL for `{table_name}`"),
+            })?;
+
+        match Self::parse_value_column_blob_type(&create_table_sql) {
+            Some(ValueBlobType::Blob) => {
+                sqlx::query(&sql_template_set.alter_value_column_statement)
+                    .execute(pool)
+                    .await
+                    .with_context(|_| MySqlExecutionSnafu {
+                        sql: sql_template_set.alter_value_column_statement.clone(),
+                    })?;
+                info!("Upgraded MySQL metadata value column to MEDIUMBLOB for `{table_name}`");
+            }
+            Some(ValueBlobType::MediumBlob | ValueBlobType::LongBlob) => {
+                debug!("MySQL metadata value column for `{table_name}` is already compatible");
+            }
+            None => {
+                warn!(
+                    "Failed to determine MySQL metadata value column type from table definition for `{table_name}`, skip automatic MEDIUMBLOB upgrade"
+                );
+            }
+        }
+
+        Ok(())
+    }
+
     /// Create [MySqlStore] impl of [KvBackendRef] from url.
     pub async fn with_url(url: &str, table_name: &str, max_txn_ops: usize) -> Result<KvBackendRef> {
+        ensure_rustls_crypto_provider_installed()?;
         let pool = MySqlPool::connect(url)
             .await
             .context(CreateMySqlPoolSnafu)?;
@@ -558,6 +644,7 @@ impl MySqlStore {
             .with_context(|_| MySqlExecutionSnafu {
                 sql: sql_template_set.create_table_statement.clone(),
             })?;
+        Self::maybe_upgrade_value_column_to_mediumblob(&pool, &sql_template_set).await?;
         Ok(Arc::new(MySqlStore {
             max_txn_ops,
             sql_template_set,
@@ -574,6 +661,7 @@ impl MySqlStore {
 mod tests {
     use common_telemetry::init_default_ut_logging;
     use sqlx::mysql::{MySqlConnectOptions, MySqlSslMode};
+    use uuid::Uuid;
 
     use super::*;
     use crate::kv_backend::test::{
@@ -585,15 +673,46 @@ mod tests {
         text_txn_multi_compare_op, unprepare_kv,
     };
     use crate::maybe_skip_mysql_integration_test;
+    use crate::rpc::store::{PutRequest, RangeRequest};
     use crate::test_util::test_certs_dir;
 
-    async fn build_mysql_kv_backend(table_name: &str) -> Option<MySqlStore> {
+    fn new_test_table_name(prefix: &str) -> String {
+        let uuid = Uuid::new_v4().simple().to_string();
+        let max_prefix_len = 63usize.saturating_sub(uuid.len() + 1);
+        let prefix = &prefix[..prefix.len().min(max_prefix_len)];
+        format!("{prefix}_{uuid}")
+    }
+
+    async fn mysql_pool() -> Option<MySqlPool> {
         init_default_ut_logging();
         let endpoints = std::env::var("GT_MYSQL_ENDPOINTS").unwrap_or_default();
         if endpoints.is_empty() {
             return None;
         }
-        let pool = MySqlPool::connect(&endpoints).await.unwrap();
+        ensure_rustls_crypto_provider_installed().unwrap();
+        Some(MySqlPool::connect(&endpoints).await.unwrap())
+    }
+
+    async fn show_create_table(pool: &MySqlPool, table_name: &str) -> String {
+        let sql = format!("SHOW CREATE TABLE `{table_name}`");
+        let row = sqlx::query(&sql).fetch_one(pool).await.unwrap();
+        row.get::<String, _>(1)
+    }
+
+    async fn create_legacy_blob_table(pool: &MySqlPool, table_name: &str) {
+        let sql = format!(
+            "CREATE TABLE IF NOT EXISTS `{table_name}`(k VARBINARY(3072) PRIMARY KEY, v BLOB);"
+        );
+        sqlx::query(&sql).execute(pool).await.unwrap();
+    }
+
+    async fn drop_table(pool: &MySqlPool, table_name: &str) {
+        let sql = format!("DROP TABLE IF EXISTS `{table_name}`;");
+        sqlx::query(&sql).execute(pool).await.unwrap();
+    }
+
+    async fn build_mysql_kv_backend(table_name: &str) -> Option<MySqlStore> {
+        let pool = mysql_pool().await?;
         let sql_templates = MySqlTemplateFactory::new(table_name).build();
         sqlx::query(&sql_templates.create_table_statement)
             .execute(&pool)
@@ -610,6 +729,156 @@ mod tests {
         })
     }
 
+    #[test]
+    fn test_parse_value_column_blob_type() {
+        let sql = r#"CREATE TABLE `greptime_metakv` (
+  `k` varbinary(3072) NOT NULL,
+  `v` MEDIUMBLOB,
+  PRIMARY KEY (`k`)
+) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4"#;
+        assert_eq!(
+            Some(ValueBlobType::MediumBlob),
+            MySqlStore::parse_value_column_blob_type(sql)
+        );
+
+        let sql = r#"CREATE TABLE `greptime_metakv` (`k` varbinary(3072) NOT NULL, `v` blob, PRIMARY KEY (`k`))"#;
+        assert_eq!(
+            Some(ValueBlobType::Blob),
+            MySqlStore::parse_value_column_blob_type(sql)
+        );
+
+        let sql = r#"CREATE TABLE `greptime_metakv` (`k` varbinary(3072) NOT NULL, `v` longblob, PRIMARY KEY (`k`))"#;
+        assert_eq!(
+            Some(ValueBlobType::LongBlob),
+            MySqlStore::parse_value_column_blob_type(sql)
+        );
+
+        let sql = "CREATE TABLE `greptime_metakv` (`k` varbinary(3072) NOT NULL, `v`    BLOB NOT NULL, PRIMARY KEY (`k`))";
+        assert_eq!(
+            Some(ValueBlobType::Blob),
+            MySqlStore::parse_value_column_blob_type(sql)
+        );
+
+        let sql = "CREATE TABLE `greptime_metakv` (\n  `k` varbinary(3072) NOT NULL,\n  \"v\" MediumBlob,\n  PRIMARY KEY (`k`)\n)";
+        assert_eq!(
+            Some(ValueBlobType::MediumBlob),
+            MySqlStore::parse_value_column_blob_type(sql)
+        );
+
+        let sql = "CREATE TABLE `greptime_metakv` (`k` varbinary(3072) NOT NULL, vv blob, `v` longblob, PRIMARY KEY (`k`))";
+        assert_eq!(
+            Some(ValueBlobType::LongBlob),
+            MySqlStore::parse_value_column_blob_type(sql)
+        );
+
+        let sql = "CREATE TABLE `greptime_metakv` (`k` varbinary(3072) NOT NULL, value blob, PRIMARY KEY (`k`))";
+        assert_eq!(None, MySqlStore::parse_value_column_blob_type(sql));
+
+        let sql = "CREATE TABLE `greptime_metakv` (`k` varbinary(3072) NOT NULL, `v` varchar(255), PRIMARY KEY (`k`))";
+        assert_eq!(None, MySqlStore::parse_value_column_blob_type(sql));
+
+        let sql =
+            "CREATE TABLE `greptime_metakv` (`k` varbinary(3072) NOT NULL, PRIMARY KEY (`k`))";
+        assert_eq!(None, MySqlStore::parse_value_column_blob_type(sql));
+    }
+
+    #[tokio::test]
+    async fn test_mysql_new_metadata_table_uses_mediumblob() {
+        maybe_skip_mysql_integration_test!();
+        let pool = mysql_pool().await.unwrap();
+        let table_name = new_test_table_name("test_mysql_mediumblob_schema");
+
+        MySqlStore::with_mysql_pool(pool.clone(), &table_name, 128)
+            .await
+            .unwrap();
+
+        let create_table = show_create_table(&pool, &table_name).await;
+        assert!(create_table.to_ascii_uppercase().contains("MEDIUMBLOB"));
+
+        drop_table(&pool, &table_name).await;
+    }
+
+    #[tokio::test]
+    async fn test_mysql_legacy_blob_metadata_table_is_upgraded() {
+        maybe_skip_mysql_integration_test!();
+        let pool = mysql_pool().await.unwrap();
+        let table_name = new_test_table_name("test_mysql_legacy_blob_upgrade");
+
+        create_legacy_blob_table(&pool, &table_name).await;
+        MySqlStore::with_mysql_pool(pool.clone(), &table_name, 128)
+            .await
+            .unwrap();
+
+        let create_table = show_create_table(&pool, &table_name).await;
+        assert!(create_table.to_ascii_uppercase().contains("MEDIUMBLOB"));
+
+        drop_table(&pool, &table_name).await;
+    }
+
+    #[tokio::test]
+    async fn test_mysql_metadata_table_stores_large_values() {
+        maybe_skip_mysql_integration_test!();
+        let pool = mysql_pool().await.unwrap();
+        let table_name = new_test_table_name("test_mysql_large_metadata_value");
+        let kv_backend = MySqlStore::with_mysql_pool(pool.clone(), &table_name, 128)
+            .await
+            .unwrap();
+        let key = b"large-value".to_vec();
+        let value = vec![b'x'; 70 * 1024];
+
+        kv_backend
+            .put(
+                PutRequest::new()
+                    .with_key(key.clone())
+                    .with_value(value.clone()),
+            )
+            .await
+            .unwrap();
+        let response = kv_backend
+            .range(RangeRequest::new().with_key(key.clone()))
+            .await
+            .unwrap();
+
+        assert_eq!(1, response.kvs.len());
+        assert_eq!(key, response.kvs[0].key);
+        assert_eq!(value, response.kvs[0].value);
+
+        drop_table(&pool, &table_name).await;
+    }
+
+    #[tokio::test]
+    async fn test_mysql_upgraded_metadata_table_stores_large_values() {
+        maybe_skip_mysql_integration_test!();
+        let pool = mysql_pool().await.unwrap();
+        let table_name = new_test_table_name("test_mysql_upgraded_large_metadata_value");
+
+        create_legacy_blob_table(&pool, &table_name).await;
+        let kv_backend = MySqlStore::with_mysql_pool(pool.clone(), &table_name, 128)
+            .await
+            .unwrap();
+        let key = b"large-value".to_vec();
+        let value = vec![b'y'; 70 * 1024];
+
+        kv_backend
+            .put(
+                PutRequest::new()
+                    .with_key(key.clone())
+                    .with_value(value.clone()),
+            )
+            .await
+            .unwrap();
+        let response = kv_backend
+            .range(RangeRequest::new().with_key(key.clone()))
+            .await
+            .unwrap();
+
+        assert_eq!(1, response.kvs.len());
+        assert_eq!(key, response.kvs[0].key);
+        assert_eq!(value, response.kvs[0].value);
+
+        drop_table(&pool, &table_name).await;
+    }
+
     #[tokio::test]
     async fn test_mysql_put() {
         maybe_skip_mysql_integration_test!();
@@ -718,6 +987,7 @@ mod tests {
     async fn test_mysql_with_tls() {
         common_telemetry::init_default_ut_logging();
         maybe_skip_mysql_integration_test!();
+        ensure_rustls_crypto_provider_installed().unwrap();
         let endpoint = std::env::var("GT_MYSQL_ENDPOINTS").unwrap();
 
         let opts = endpoint
@@ -732,6 +1002,7 @@ mod tests {
     async fn test_mysql_with_mtls() {
         common_telemetry::init_default_ut_logging();
         maybe_skip_mysql_integration_test!();
+        ensure_rustls_crypto_provider_installed().unwrap();
         let endpoint = std::env::var("GT_MYSQL_ENDPOINTS").unwrap();
         let certs_dir = test_certs_dir();
 
@@ -749,6 +1020,7 @@ mod tests {
     async fn test_mysql_with_tls_verify_ca() {
         common_telemetry::init_default_ut_logging();
         maybe_skip_mysql_integration_test!();
+        ensure_rustls_crypto_provider_installed().unwrap();
         let endpoint = std::env::var("GT_MYSQL_ENDPOINTS").unwrap();
         let certs_dir = test_certs_dir();
 
@@ -767,6 +1039,7 @@ mod tests {
     async fn test_mysql_with_tls_verify_ident() {
         common_telemetry::init_default_ut_logging();
         maybe_skip_mysql_integration_test!();
+        ensure_rustls_crypto_provider_installed().unwrap();
         let endpoint = std::env::var("GT_MYSQL_ENDPOINTS").unwrap();
         let certs_dir = test_certs_dir();
 
diff --git a/src/common/meta/src/kv_backend/rds/postgres.rs b/src/common/meta/src/kv_backend/rds/postgres.rs
index b8d0cd4a6d..4a2f8b1356 100644
--- a/src/common/meta/src/kv_backend/rds/postgres.rs
+++ b/src/common/meta/src/kv_backend/rds/postgres.rs
@@ -41,6 +41,7 @@ use crate::kv_backend::rds::{
     Executor, ExecutorFactory, ExecutorImpl, KvQueryExecutor, RDS_STORE_OP_BATCH_DELETE,
     RDS_STORE_OP_BATCH_GET, RDS_STORE_OP_BATCH_PUT, RDS_STORE_OP_RANGE_DELETE,
     RDS_STORE_OP_RANGE_QUERY, RDS_STORE_TXN_RETRY_COUNT, RdsStore, Transaction,
+    ensure_rustls_crypto_provider_installed,
 };
 use crate::rpc::KeyValue;
 use crate::rpc::store::{
@@ -422,6 +423,7 @@ pub fn create_postgres_tls_connector(tls_config: &TlsOption) -> Result<MakeRustl
         "Creating PostgreSQL TLS connector with mode: {:?}",
         tls_config.mode
     );
+    ensure_rustls_crypto_provider_installed()?;
 
     let config_builder = match tls_config.mode {
         TlsMode::Disable => {
@@ -516,7 +518,7 @@ impl ServerCertVerifier for AcceptAnyVerifier {
 
     fn supported_verify_schemes(&self) -> Vec<SignatureScheme> {
         // Support all signature schemes
-        rustls::crypto::ring::default_provider()
+        rustls::crypto::aws_lc_rs::default_provider()
             .signature_verification_algorithms
             .supported_schemes()
     }
@@ -544,7 +546,7 @@ impl ServerCertVerifier for NoHostnameVerification {
             &self.roots,
             intermediates,
             now,
-            rustls::crypto::ring::default_provider()
+            rustls::crypto::aws_lc_rs::default_provider()
                 .signature_verification_algorithms
                 .all,
         )?;
@@ -562,7 +564,7 @@ impl ServerCertVerifier for NoHostnameVerification {
             message,
             cert,
             dss,
-            &rustls::crypto::ring::default_provider().signature_verification_algorithms,
+            &rustls::crypto::aws_lc_rs::default_provider().signature_verification_algorithms,
         )
     }
 
@@ -576,13 +578,13 @@ impl ServerCertVerifier for NoHostnameVerification {
             message,
             cert,
             dss,
-            &rustls::crypto::ring::default_provider().signature_verification_algorithms,
+            &rustls::crypto::aws_lc_rs::default_provider().signature_verification_algorithms,
         )
     }
 
     fn supported_verify_schemes(&self) -> Vec<SignatureScheme> {
         // Support all signature schemes
-        rustls::crypto::ring::default_provider()
+        rustls::crypto::aws_lc_rs::default_provider()
             .signature_verification_algorithms
             .supported_schemes()
     }
diff --git a/src/common/meta/src/kv_backend/read_only.rs b/src/common/meta/src/kv_backend/read_only.rs
new file mode 100644
index 0000000000..43c8d2067a
--- /dev/null
+++ b/src/common/meta/src/kv_backend/read_only.rs
@@ -0,0 +1,348 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::any::Any;
+
+use crate::error::{ReadOnlyKvBackendSnafu, Result};
+use crate::kv_backend::txn::{Txn, TxnOp, TxnResponse};
+use crate::kv_backend::{KvBackend, KvBackendRef, TxnService};
+use crate::rpc::KeyValue;
+use crate::rpc::store::{
+    BatchDeleteRequest, BatchDeleteResponse, BatchGetRequest, BatchGetResponse, BatchPutRequest,
+    BatchPutResponse, CompareAndPutRequest, CompareAndPutResponse, DeleteRangeRequest,
+    DeleteRangeResponse, PutRequest, PutResponse, RangeRequest, RangeResponse,
+};
+
+/// A [`KvBackend`] wrapper that forwards reads and rejects writes.
+pub struct ReadOnlyKvBackend {
+    inner: KvBackendRef,
+    name: String,
+}
+
+impl ReadOnlyKvBackend {
+    pub fn new(inner: KvBackendRef) -> Self {
+        let name = format!("ReadOnlyKvBackend({})", inner.name());
+        Self { inner, name }
+    }
+
+    fn read_only<T>(&self) -> Result<T> {
+        ReadOnlyKvBackendSnafu {
+            name: self.name.clone(),
+        }
+        .fail()
+    }
+
+    fn validate_read_only_ops(&self, ops: &[TxnOp]) -> Result<()> {
+        if ops
+            .iter()
+            .any(|op| matches!(op, TxnOp::Put(_, _) | TxnOp::Delete(_)))
+        {
+            self.read_only()
+        } else {
+            Ok(())
+        }
+    }
+
+    fn validate_read_only_txn(&self, txn: &Txn) -> Result<()> {
+        self.validate_read_only_ops(&txn.req.success)?;
+        self.validate_read_only_ops(&txn.req.failure)
+    }
+}
+
+#[async_trait::async_trait]
+impl TxnService for ReadOnlyKvBackend {
+    type Error = crate::error::Error;
+
+    async fn txn(&self, txn: Txn) -> Result<TxnResponse> {
+        self.validate_read_only_txn(&txn)?;
+        self.inner.txn(txn).await
+    }
+
+    fn max_txn_ops(&self) -> usize {
+        self.inner.max_txn_ops()
+    }
+}
+
+#[async_trait::async_trait]
+impl KvBackend for ReadOnlyKvBackend {
+    fn name(&self) -> &str {
+        &self.name
+    }
+
+    fn as_any(&self) -> &dyn Any {
+        self
+    }
+
+    async fn range(&self, req: RangeRequest) -> Result<RangeResponse> {
+        self.inner.range(req).await
+    }
+
+    async fn put(&self, _req: PutRequest) -> Result<PutResponse> {
+        self.read_only()
+    }
+
+    async fn batch_put(&self, _req: BatchPutRequest) -> Result<BatchPutResponse> {
+        self.read_only()
+    }
+
+    async fn batch_get(&self, req: BatchGetRequest) -> Result<BatchGetResponse> {
+        self.inner.batch_get(req).await
+    }
+
+    async fn compare_and_put(&self, _req: CompareAndPutRequest) -> Result<CompareAndPutResponse> {
+        self.read_only()
+    }
+
+    async fn delete_range(&self, _req: DeleteRangeRequest) -> Result<DeleteRangeResponse> {
+        self.read_only()
+    }
+
+    async fn batch_delete(&self, _req: BatchDeleteRequest) -> Result<BatchDeleteResponse> {
+        self.read_only()
+    }
+
+    async fn get(&self, key: &[u8]) -> Result<Option<KeyValue>> {
+        self.inner.get(key).await
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use common_error::ext::ErrorExt;
+    use common_error::status_code::StatusCode;
+
+    use super::*;
+    use crate::error::Error;
+    use crate::kv_backend::memory::MemoryKvBackend;
+    use crate::kv_backend::txn::{Compare, CompareOp, TxnOpResponse};
+    use crate::rpc::store::{
+        BatchDeleteRequest, BatchGetRequest, BatchPutRequest, CompareAndPutRequest,
+        DeleteRangeRequest, PutRequest,
+    };
+
+    async fn read_only_backend() -> ReadOnlyKvBackend {
+        let inner = Arc::new(MemoryKvBackend::<Error>::new());
+        inner
+            .put(PutRequest::new().with_key(b"k1").with_value(b"v1"))
+            .await
+            .unwrap();
+        inner
+            .put(PutRequest::new().with_key(b"k2").with_value(b"v2"))
+            .await
+            .unwrap();
+
+        ReadOnlyKvBackend::new(inner)
+    }
+
+    fn assert_read_only<T>(result: Result<T>) {
+        let err = match result {
+            Ok(_) => panic!("expected read-only error"),
+            Err(err) => err,
+        };
+        assert!(matches!(err, Error::ReadOnlyKvBackend { .. }));
+        assert_eq!(err.status_code(), StatusCode::Unsupported);
+    }
+
+    struct TxnOnlyBackend;
+
+    #[async_trait::async_trait]
+    impl TxnService for TxnOnlyBackend {
+        type Error = Error;
+
+        async fn txn(&self, _txn: Txn) -> Result<TxnResponse> {
+            Ok(TxnResponse {
+                succeeded: true,
+                responses: vec![TxnOpResponse::ResponseGet(RangeResponse {
+                    kvs: vec![KeyValue {
+                        key: b"k1".to_vec(),
+                        value: b"v1".to_vec(),
+                    }],
+                    more: false,
+                })],
+            })
+        }
+
+        fn max_txn_ops(&self) -> usize {
+            7
+        }
+    }
+
+    #[async_trait::async_trait]
+    impl KvBackend for TxnOnlyBackend {
+        fn name(&self) -> &str {
+            "TxnOnlyBackend"
+        }
+
+        fn as_any(&self) -> &dyn Any {
+            self
+        }
+
+        async fn range(&self, _req: RangeRequest) -> Result<RangeResponse> {
+            unimplemented!("read-only txn should delegate to inner txn")
+        }
+
+        async fn put(&self, _req: PutRequest) -> Result<PutResponse> {
+            unimplemented!("read-only txn should delegate to inner txn")
+        }
+
+        async fn batch_put(&self, _req: BatchPutRequest) -> Result<BatchPutResponse> {
+            unimplemented!("read-only txn should delegate to inner txn")
+        }
+
+        async fn batch_get(&self, _req: BatchGetRequest) -> Result<BatchGetResponse> {
+            unimplemented!("read-only txn should delegate to inner txn")
+        }
+
+        async fn delete_range(&self, _req: DeleteRangeRequest) -> Result<DeleteRangeResponse> {
+            unimplemented!("read-only txn should delegate to inner txn")
+        }
+
+        async fn batch_delete(&self, _req: BatchDeleteRequest) -> Result<BatchDeleteResponse> {
+            unimplemented!("read-only txn should delegate to inner txn")
+        }
+    }
+
+    #[tokio::test]
+    async fn test_read_only_backend_forwards_reads() {
+        let backend = read_only_backend().await;
+
+        let range = backend
+            .range(RangeRequest::new().with_key(b"k1"))
+            .await
+            .unwrap();
+        assert_eq!(range.kvs.len(), 1);
+        assert_eq!(range.kvs[0].value, b"v1");
+
+        let kv = backend.get(b"k2").await.unwrap().unwrap();
+        assert_eq!(kv.value, b"v2");
+
+        let batch = backend
+            .batch_get(BatchGetRequest::new().add_key(b"k1").add_key(b"k2"))
+            .await
+            .unwrap();
+        assert_eq!(batch.kvs.len(), 2);
+    }
+
+    #[tokio::test]
+    async fn test_read_only_backend_rejects_writes() {
+        let backend = read_only_backend().await;
+
+        assert_read_only(
+            backend
+                .put(PutRequest::new().with_key(b"k3").with_value(b"v3"))
+                .await,
+        );
+        assert_read_only(
+            backend
+                .batch_put(BatchPutRequest::new().add_kv(b"k3", b"v3"))
+                .await,
+        );
+        assert_read_only(
+            backend
+                .compare_and_put(
+                    CompareAndPutRequest::new()
+                        .with_key(b"k1")
+                        .with_expect(b"v1")
+                        .with_value(b"v3"),
+                )
+                .await,
+        );
+        assert_read_only(
+            backend
+                .delete_range(DeleteRangeRequest::new().with_key(b"k1"))
+                .await,
+        );
+        assert_read_only(
+            backend
+                .batch_delete(BatchDeleteRequest::new().add_key(b"k1"))
+                .await,
+        );
+    }
+
+    #[tokio::test]
+    async fn test_read_only_backend_rejects_write_txn() {
+        let backend = read_only_backend().await;
+
+        assert_eq!(backend.max_txn_ops(), usize::MAX);
+        assert_read_only(
+            backend
+                .txn(Txn::put_if_not_exists(b"k3".to_vec(), b"v3".to_vec()))
+                .await,
+        );
+    }
+
+    #[tokio::test]
+    async fn test_read_only_backend_delegates_read_txn() {
+        let backend = ReadOnlyKvBackend::new(Arc::new(TxnOnlyBackend));
+
+        assert_eq!(backend.max_txn_ops(), 7);
+        let resp = backend
+            .txn(Txn::new().and_then(vec![TxnOp::Get(b"k1".to_vec())]))
+            .await
+            .unwrap();
+
+        assert!(resp.succeeded);
+        let TxnOpResponse::ResponseGet(range) = &resp.responses[0] else {
+            panic!("expected get response");
+        };
+        assert_eq!(range.kvs[0].value, b"v1");
+    }
+
+    #[tokio::test]
+    async fn test_read_only_backend_allows_get_only_txn() {
+        let backend = read_only_backend().await;
+
+        let resp = backend
+            .txn(Txn::new().and_then(vec![TxnOp::Get(b"k1".to_vec()), TxnOp::Get(b"k2".to_vec())]))
+            .await
+            .unwrap();
+
+        assert!(resp.succeeded);
+        assert_eq!(resp.responses.len(), 2);
+        let TxnOpResponse::ResponseGet(range) = &resp.responses[0] else {
+            panic!("expected get response");
+        };
+        assert_eq!(range.kvs.len(), 1);
+        assert_eq!(range.kvs[0].value, b"v1");
+        let TxnOpResponse::ResponseGet(range) = &resp.responses[1] else {
+            panic!("expected get response");
+        };
+        assert_eq!(range.kvs.len(), 1);
+        assert_eq!(range.kvs[0].value, b"v2");
+    }
+
+    #[tokio::test]
+    async fn test_read_only_backend_allows_compare_and_get_txn() {
+        let backend = read_only_backend().await;
+
+        let txn = Txn::new()
+            .when(vec![Compare::with_value(
+                b"k1".to_vec(),
+                CompareOp::Equal,
+                b"v1".to_vec(),
+            )])
+            .and_then(vec![TxnOp::Get(b"k2".to_vec())])
+            .or_else(vec![TxnOp::Get(b"k1".to_vec())]);
+        let resp = backend.txn(txn).await.unwrap();
+
+        assert!(resp.succeeded);
+        let TxnOpResponse::ResponseGet(range) = &resp.responses[0] else {
+            panic!("expected get response");
+        };
+        assert_eq!(range.kvs.len(), 1);
+        assert_eq!(range.kvs[0].value, b"v2");
+    }
+}
diff --git a/src/common/meta/src/lib.rs b/src/common/meta/src/lib.rs
index 93cd229b16..ef5109dc03 100644
--- a/src/common/meta/src/lib.rs
+++ b/src/common/meta/src/lib.rs
@@ -12,9 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
-#![feature(duration_millis_float)]
-
 pub mod cache;
 pub mod cache_invalidator;
 pub mod cluster;
@@ -22,6 +19,7 @@ pub mod datanode;
 pub mod ddl;
 pub mod ddl_manager;
 pub mod distributed_time_constants;
+pub mod election;
 pub mod error;
 pub mod flow_name;
 pub mod heartbeat;
diff --git a/src/common/meta/src/range_stream.rs b/src/common/meta/src/range_stream.rs
index 2fc333064e..4bcd5de7db 100644
--- a/src/common/meta/src/range_stream.rs
+++ b/src/common/meta/src/range_stream.rs
@@ -187,7 +187,7 @@ impl<T> PaginationStream<T> {
 #[cfg(test)]
 mod tests {
 
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::BTreeMap;
     use std::sync::Arc;
 
diff --git a/src/common/meta/src/reconciliation/reconcile_logical_tables/update_table_infos.rs b/src/common/meta/src/reconciliation/reconcile_logical_tables/update_table_infos.rs
index d0de7a06fb..e8597ccd36 100644
--- a/src/common/meta/src/reconciliation/reconcile_logical_tables/update_table_infos.rs
+++ b/src/common/meta/src/reconciliation/reconcile_logical_tables/update_table_infos.rs
@@ -84,7 +84,7 @@ impl State for UpdateTableInfos {
             .persistent_ctx
             .update_table_infos
             .iter()
-            .zip(table_info_values.into_iter())
+            .zip(table_info_values)
         {
             let new_table_info = Self::build_new_table_info(
                 *table_id,
diff --git a/src/common/meta/src/reconciliation/utils.rs b/src/common/meta/src/reconciliation/utils.rs
index 4debc6de4d..6ddc084596 100644
--- a/src/common/meta/src/reconciliation/utils.rs
+++ b/src/common/meta/src/reconciliation/utils.rs
@@ -949,7 +949,7 @@ impl Display for ReconcileTableMetrics {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
     use std::sync::Arc;
 
diff --git a/src/common/meta/src/region_keeper.rs b/src/common/meta/src/region_keeper.rs
index 605aaba893..aca0b95838 100644
--- a/src/common/meta/src/region_keeper.rs
+++ b/src/common/meta/src/region_keeper.rs
@@ -12,9 +12,11 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::collections::HashSet;
+use std::collections::hash_map::Entry;
+use std::collections::{HashMap, HashSet};
 use std::sync::{Arc, RwLock};
 
+use store_api::region_engine::RegionRole;
 use store_api::storage::RegionId;
 
 use crate::DatanodeId;
@@ -24,7 +26,7 @@ use crate::DatanodeId;
 pub struct OperatingRegionGuard {
     datanode_id: DatanodeId,
     region_id: RegionId,
-    inner: Arc<RwLock<HashSet<(DatanodeId, RegionId)>>>,
+    inner: Arc<RwLock<HashMap<(DatanodeId, RegionId), RegionRole>>>,
 }
 
 impl Drop for OperatingRegionGuard {
@@ -50,7 +52,7 @@ pub type MemoryRegionKeeperRef = Arc<MemoryRegionKeeper>;
 /// - Tracks the deleting regions after the corresponding metadata is deleted.
 #[derive(Debug, Clone, Default)]
 pub struct MemoryRegionKeeper {
-    inner: Arc<RwLock<HashSet<(DatanodeId, RegionId)>>>,
+    inner: Arc<RwLock<HashMap<(DatanodeId, RegionId), RegionRole>>>,
 }
 
 impl MemoryRegionKeeper {
@@ -59,40 +61,48 @@ impl MemoryRegionKeeper {
     }
 
     /// Returns [OperatingRegionGuard] if Region(`region_id`) on Peer(`datanode_id`) does not exist.
-    pub fn register(
+    pub fn register_with_role(
         &self,
         datanode_id: DatanodeId,
         region_id: RegionId,
+        role: RegionRole,
     ) -> Option<OperatingRegionGuard> {
         let mut inner = self.inner.write().unwrap();
 
-        if inner.insert((datanode_id, region_id)) {
-            Some(OperatingRegionGuard {
-                datanode_id,
-                region_id,
-                inner: self.inner.clone(),
-            })
-        } else {
-            None
+        match inner.entry((datanode_id, region_id)) {
+            Entry::Occupied(_) => None,
+            Entry::Vacant(vacant_entry) => {
+                vacant_entry.insert(role);
+                Some(OperatingRegionGuard {
+                    datanode_id,
+                    region_id,
+                    inner: self.inner.clone(),
+                })
+            }
         }
     }
 
     /// Returns true if the keeper contains a (`datanoe_id`, `region_id`) tuple.
     pub fn contains(&self, datanode_id: DatanodeId, region_id: RegionId) -> bool {
         let inner = self.inner.read().unwrap();
-        inner.contains(&(datanode_id, region_id))
+        inner.contains_key(&(datanode_id, region_id))
     }
 
-    /// Extracts all operating regions from `region_ids` and returns operating regions.
-    pub fn extract_operating_regions(
+    /// Extracts all operating regions with roles from `region_ids`.
+    pub fn extract_operating_region_roles(
         &self,
         datanode_id: DatanodeId,
-        region_ids: &mut HashSet<RegionId>,
-    ) -> HashSet<RegionId> {
+        region_ids: &HashSet<RegionId>,
+    ) -> HashMap<RegionId, RegionRole> {
         let inner = self.inner.read().unwrap();
         region_ids
-            .extract_if(|region_id| inner.contains(&(datanode_id, *region_id)))
-            .collect::<HashSet<_>>()
+            .iter()
+            .filter_map(|region_id| {
+                inner
+                    .get(&(datanode_id, *region_id))
+                    .map(|role| (*region_id, *role))
+            })
+            .collect()
     }
 
     /// Returns number of element in tracking set.
@@ -115,8 +125,9 @@ impl MemoryRegionKeeper {
 
 #[cfg(test)]
 mod tests {
-    use std::collections::HashSet;
+    use std::collections::{HashMap, HashSet};
 
+    use store_api::region_engine::RegionRole;
     use store_api::storage::RegionId;
 
     use crate::region_keeper::MemoryRegionKeeper;
@@ -125,20 +136,43 @@ mod tests {
     fn test_opening_region_keeper() {
         let keeper = MemoryRegionKeeper::new();
 
-        let guard = keeper.register(1, RegionId::from_u64(1)).unwrap();
-        assert!(keeper.register(1, RegionId::from_u64(1)).is_none());
-        let guard2 = keeper.register(1, RegionId::from_u64(2)).unwrap();
+        let guard = keeper
+            .register_with_role(1, RegionId::from_u64(1), RegionRole::Leader)
+            .unwrap();
+        assert!(
+            keeper
+                .register_with_role(1, RegionId::from_u64(1), RegionRole::Leader)
+                .is_none()
+        );
+        let guard2 = keeper
+            .register_with_role(1, RegionId::from_u64(2), RegionRole::Follower)
+            .unwrap();
 
-        let mut regions = HashSet::from([
+        let regions = HashSet::from([
             RegionId::from_u64(1),
             RegionId::from_u64(2),
             RegionId::from_u64(3),
         ]);
-        let output = keeper.extract_operating_regions(1, &mut regions);
+        let output = keeper.extract_operating_region_roles(1, &regions);
         assert_eq!(output.len(), 2);
 
-        assert!(output.contains(&RegionId::from_u64(1)));
-        assert!(output.contains(&RegionId::from_u64(2)));
+        assert!(output.contains_key(&RegionId::from_u64(1)));
+        assert!(output.contains_key(&RegionId::from_u64(2)));
+        assert_eq!(keeper.len(), 2);
+
+        let regions = HashSet::from([
+            RegionId::from_u64(1),
+            RegionId::from_u64(2),
+            RegionId::from_u64(3),
+        ]);
+        let output = keeper.extract_operating_region_roles(1, &regions);
+        assert_eq!(
+            output,
+            HashMap::from([
+                (RegionId::from_u64(1), RegionRole::Leader),
+                (RegionId::from_u64(2), RegionRole::Follower),
+            ])
+        );
         assert_eq!(keeper.len(), 2);
 
         drop(guard);
diff --git a/src/common/meta/src/rpc/ddl.rs b/src/common/meta/src/rpc/ddl.rs
index 1e94a8f092..ed6f78154a 100644
--- a/src/common/meta/src/rpc/ddl.rs
+++ b/src/common/meta/src/rpc/ddl.rs
@@ -1432,6 +1432,12 @@ pub struct QueryContext {
     pub timezone: String,
     pub extensions: HashMap<String, String>,
     pub channel: u8,
+    /// Maps region id -> snapshot upper bound sequence for that region.
+    #[serde(default)]
+    pub snapshot_seqs: HashMap<u64, u64>,
+    /// Maps region id -> minimal SST sequence allowed for that region.
+    #[serde(default)]
+    pub sst_min_sequences: HashMap<u64, u64>,
 }
 
 impl QueryContext {
@@ -1459,6 +1465,14 @@ impl QueryContext {
     pub fn channel(&self) -> u8 {
         self.channel
     }
+
+    pub fn snapshot_seqs(&self) -> &HashMap<u64, u64> {
+        &self.snapshot_seqs
+    }
+
+    pub fn sst_min_sequences(&self) -> &HashMap<u64, u64> {
+        &self.sst_min_sequences
+    }
 }
 
 /// Lightweight query context for flow operations containing only essential fields.
@@ -1466,12 +1480,24 @@ impl QueryContext {
 /// for flow creation and execution.
 #[derive(Debug, Clone, Serialize, PartialEq)]
 pub struct FlowQueryContext {
-    /// Current catalog name - needed for flow metadata and recovery
+    /// Current catalog name used for flow metadata and execution.
     pub catalog: String,
-    /// Current schema name - needed for table resolution during flow execution
+    /// Current schema name used for table resolution during flow execution.
     pub schema: String,
-    /// Timezone for timestamp operations in the flow
+    /// Timezone used for timestamp evaluation in the flow.
     pub timezone: String,
+    /// Query extensions carried into flow execution.
+    #[serde(default)]
+    pub extensions: HashMap<String, String>,
+    /// Request channel propagated from the original query context.
+    #[serde(default)]
+    pub channel: u8,
+    /// Per-region snapshot upper bounds bound during query planning/execution.
+    #[serde(default)]
+    pub snapshot_seqs: HashMap<u64, u64>,
+    /// Per-region lower SST scan bounds carried with the flow context.
+    #[serde(default)]
+    pub sst_min_sequences: HashMap<u64, u64>,
 }
 
 impl<'de> Deserialize<'de> for FlowQueryContext {
@@ -1492,6 +1518,14 @@ impl<'de> Deserialize<'de> for FlowQueryContext {
             catalog: String,
             schema: String,
             timezone: String,
+            #[serde(default)]
+            extensions: HashMap<String, String>,
+            #[serde(default)]
+            channel: u8,
+            #[serde(default)]
+            snapshot_seqs: HashMap<u64, u64>,
+            #[serde(default)]
+            sst_min_sequences: HashMap<u64, u64>,
         }
 
         match ContextCompat::deserialize(deserializer)? {
@@ -1499,6 +1533,10 @@ impl<'de> Deserialize<'de> for FlowQueryContext {
                 catalog: helper.catalog,
                 schema: helper.schema,
                 timezone: helper.timezone,
+                extensions: helper.extensions,
+                channel: helper.channel,
+                snapshot_seqs: helper.snapshot_seqs,
+                sst_min_sequences: helper.sst_min_sequences,
             }),
             ContextCompat::Full(full_ctx) => Ok(full_ctx.into()),
         }
@@ -1507,12 +1545,19 @@ impl<'de> Deserialize<'de> for FlowQueryContext {
 
 impl From<PbQueryContext> for QueryContext {
     fn from(pb_ctx: PbQueryContext) -> Self {
+        let (snapshot_seqs, sst_min_sequences) = pb_ctx
+            .snapshot_seqs
+            .map(|seqs| (seqs.snapshot_seqs, seqs.sst_min_sequences))
+            .unwrap_or_default();
+
         Self {
             current_catalog: pb_ctx.current_catalog,
             current_schema: pb_ctx.current_schema,
             timezone: pb_ctx.timezone,
             extensions: pb_ctx.extensions,
             channel: pb_ctx.channel as u8,
+            snapshot_seqs,
+            sst_min_sequences,
         }
     }
 }
@@ -1525,6 +1570,8 @@ impl From<QueryContext> for PbQueryContext {
             timezone,
             extensions,
             channel,
+            snapshot_seqs,
+            sst_min_sequences,
         }: QueryContext,
     ) -> Self {
         PbQueryContext {
@@ -1533,7 +1580,12 @@ impl From<QueryContext> for PbQueryContext {
             timezone,
             extensions,
             channel: channel as u32,
-            snapshot_seqs: None,
+            snapshot_seqs: (!snapshot_seqs.is_empty() || !sst_min_sequences.is_empty()).then_some(
+                api::v1::SnapshotSequences {
+                    snapshot_seqs,
+                    sst_min_sequences,
+                },
+            ),
             explain: None,
         }
     }
@@ -1545,6 +1597,10 @@ impl From<QueryContext> for FlowQueryContext {
             catalog: ctx.current_catalog,
             schema: ctx.current_schema,
             timezone: ctx.timezone,
+            extensions: ctx.extensions,
+            channel: ctx.channel,
+            snapshot_seqs: ctx.snapshot_seqs,
+            sst_min_sequences: ctx.sst_min_sequences,
         }
     }
 }
@@ -1555,8 +1611,10 @@ impl From<FlowQueryContext> for QueryContext {
             current_catalog: flow_ctx.catalog,
             current_schema: flow_ctx.schema,
             timezone: flow_ctx.timezone,
-            extensions: HashMap::new(),
-            channel: 0, // Use default channel for flows
+            extensions: flow_ctx.extensions,
+            channel: flow_ctx.channel,
+            snapshot_seqs: flow_ctx.snapshot_seqs,
+            sst_min_sequences: flow_ctx.sst_min_sequences,
         }
     }
 }
@@ -1720,6 +1778,8 @@ mod tests {
             timezone: "UTC".to_string(),
             extensions,
             channel: 5,
+            snapshot_seqs: HashMap::from([(10, 100)]),
+            sst_min_sequences: HashMap::from([(10, 90)]),
         };
 
         let flow_ctx: FlowQueryContext = query_ctx.into();
@@ -1727,6 +1787,9 @@ mod tests {
         assert_eq!(flow_ctx.catalog, "test_catalog");
         assert_eq!(flow_ctx.schema, "test_schema");
         assert_eq!(flow_ctx.timezone, "UTC");
+        assert_eq!(flow_ctx.channel, 5);
+        assert_eq!(flow_ctx.snapshot_seqs, HashMap::from([(10, 100)]));
+        assert_eq!(flow_ctx.sst_min_sequences, HashMap::from([(10, 90)]));
     }
 
     #[test]
@@ -1735,6 +1798,10 @@ mod tests {
             catalog: "prod_catalog".to_string(),
             schema: "public".to_string(),
             timezone: "America/New_York".to_string(),
+            extensions: HashMap::from([("k".to_string(), "v".to_string())]),
+            channel: 7,
+            snapshot_seqs: HashMap::from([(11, 111)]),
+            sst_min_sequences: HashMap::from([(11, 101)]),
         };
 
         let query_ctx: QueryContext = flow_ctx.clone().into();
@@ -1742,8 +1809,13 @@ mod tests {
         assert_eq!(query_ctx.current_catalog, "prod_catalog");
         assert_eq!(query_ctx.current_schema, "public");
         assert_eq!(query_ctx.timezone, "America/New_York");
-        assert!(query_ctx.extensions.is_empty());
-        assert_eq!(query_ctx.channel, 0);
+        assert_eq!(
+            query_ctx.extensions,
+            HashMap::from([("k".to_string(), "v".to_string())])
+        );
+        assert_eq!(query_ctx.channel, 7);
+        assert_eq!(query_ctx.snapshot_seqs, HashMap::from([(11, 111)]));
+        assert_eq!(query_ctx.sst_min_sequences, HashMap::from([(11, 101)]));
 
         // Test roundtrip conversion
         let flow_ctx_roundtrip: FlowQueryContext = query_ctx.into();
@@ -1756,6 +1828,10 @@ mod tests {
             catalog: "test_catalog".to_string(),
             schema: "test_schema".to_string(),
             timezone: "UTC".to_string(),
+            extensions: HashMap::new(),
+            channel: 0,
+            snapshot_seqs: HashMap::new(),
+            sst_min_sequences: HashMap::new(),
         };
 
         let serialized = serde_json::to_string(&flow_ctx).unwrap();
@@ -1776,6 +1852,10 @@ mod tests {
             catalog: "pb_catalog".to_string(),
             schema: "pb_schema".to_string(),
             timezone: "Asia/Tokyo".to_string(),
+            extensions: HashMap::from([("x".to_string(), "y".to_string())]),
+            channel: 6,
+            snapshot_seqs: HashMap::from([(3, 30)]),
+            sst_min_sequences: HashMap::from([(3, 21)]),
         };
 
         let pb_ctx: PbQueryContext = flow_ctx.into();
@@ -1783,9 +1863,44 @@ mod tests {
         assert_eq!(pb_ctx.current_catalog, "pb_catalog");
         assert_eq!(pb_ctx.current_schema, "pb_schema");
         assert_eq!(pb_ctx.timezone, "Asia/Tokyo");
-        assert!(pb_ctx.extensions.is_empty());
-        assert_eq!(pb_ctx.channel, 0);
-        assert!(pb_ctx.snapshot_seqs.is_none());
+        assert_eq!(
+            pb_ctx.extensions,
+            HashMap::from([("x".to_string(), "y".to_string())])
+        );
+        assert_eq!(pb_ctx.channel, 6);
+        assert_eq!(
+            pb_ctx.snapshot_seqs,
+            Some(api::v1::SnapshotSequences {
+                snapshot_seqs: HashMap::from([(3, 30)]),
+                sst_min_sequences: HashMap::from([(3, 21)]),
+            })
+        );
         assert!(pb_ctx.explain.is_none());
     }
+
+    #[test]
+    fn test_pb_query_context_roundtrip_with_snapshot_sequences() {
+        let pb = PbQueryContext {
+            current_catalog: "c1".to_string(),
+            current_schema: "s1".to_string(),
+            timezone: "UTC".to_string(),
+            extensions: HashMap::from([("flow.return_region_seq".to_string(), "true".to_string())]),
+            channel: 3,
+            snapshot_seqs: Some(api::v1::SnapshotSequences {
+                snapshot_seqs: HashMap::from([(1, 100)]),
+                sst_min_sequences: HashMap::from([(1, 90)]),
+            }),
+            explain: None,
+        };
+
+        let query_ctx: QueryContext = pb.clone().into();
+        let pb_roundtrip: PbQueryContext = query_ctx.into();
+
+        assert_eq!(pb_roundtrip.current_catalog, pb.current_catalog);
+        assert_eq!(pb_roundtrip.current_schema, pb.current_schema);
+        assert_eq!(pb_roundtrip.timezone, pb.timezone);
+        assert_eq!(pb_roundtrip.extensions, pb.extensions);
+        assert_eq!(pb_roundtrip.channel, pb.channel);
+        assert_eq!(pb_roundtrip.snapshot_seqs, pb.snapshot_seqs);
+    }
 }
diff --git a/src/common/meta/src/rpc/router.rs b/src/common/meta/src/rpc/router.rs
index 92627c1d8f..519b12e081 100644
--- a/src/common/meta/src/rpc/router.rs
+++ b/src/common/meta/src/rpc/router.rs
@@ -23,6 +23,7 @@ use derive_builder::Builder;
 use serde::ser::SerializeSeq;
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
 use snafu::OptionExt;
+use store_api::region_engine::RegionRole;
 use store_api::storage::{RegionId, RegionNumber};
 use strum::AsRefStr;
 use table::table_name::TableName;
@@ -99,6 +100,20 @@ pub fn operating_leader_regions(region_routes: &[RegionRoute]) -> Vec<(RegionId,
         .collect::<Vec<_>>()
 }
 
+/// Returns the operating leader regions with corresponding [DatanodeId] and [RegionRole].
+pub fn operating_leader_region_roles(
+    region_routes: &[RegionRoute],
+) -> Vec<(RegionId, DatanodeId, RegionRole)> {
+    region_routes
+        .iter()
+        .filter_map(|route| {
+            let role = route.leader_region_role()?;
+            let leader = route.leader_peer.as_ref()?;
+            Some((route.region.id, leader.id, role))
+        })
+        .collect()
+}
+
 /// Returns the HashMap<[RegionNumber], &[Peer]>;
 ///
 /// If the region doesn't have a leader peer, the [Region] will be omitted.
@@ -342,6 +357,19 @@ impl RegionRoute {
         matches!(self.leader_state, Some(LeaderState::Staging))
     }
 
+    /// Returns the role of the leader region.
+    pub fn leader_region_role(&self) -> Option<RegionRole> {
+        self.leader_peer.as_ref().map(|_| {
+            if self.is_leader_staging() {
+                RegionRole::StagingLeader
+            } else if self.is_leader_downgrading() {
+                RegionRole::DowngradingLeader
+            } else {
+                RegionRole::Leader
+            }
+        })
+    }
+
     /// Marks the Leader [`Region`] as [`RegionState::Downgrading`].
     ///
     /// We should downgrade a [`Region`] before deactivating it:
@@ -577,6 +605,17 @@ mod tests {
     use super::*;
     use crate::key::RegionRoleSet;
 
+    fn new_test_region_route(region_id: RegionId) -> RegionRoute {
+        RegionRoute {
+            region: Region::new_test(region_id),
+            leader_peer: Some(Peer::new(1, "a1")),
+            follower_peers: vec![Peer::new(2, "a2")],
+            leader_state: None,
+            leader_down_since: None,
+            write_route_policy: None,
+        }
+    }
+
     #[test]
     fn test_leader_is_downgraded() {
         let mut region_route = RegionRoute {
@@ -740,6 +779,65 @@ mod tests {
         assert!(!region_route.is_ignore_all_writes());
     }
 
+    #[test]
+    fn test_leader_region_role_without_leader_peer_returns_none() {
+        let region_route = RegionRoute {
+            leader_peer: None,
+            ..new_test_region_route(RegionId::new(1, 1))
+        };
+
+        assert_eq!(region_route.leader_region_role(), None);
+    }
+
+    #[test]
+    fn test_leader_region_role_variants() {
+        let normal = new_test_region_route(RegionId::new(1, 1));
+        let mut downgrading = new_test_region_route(RegionId::new(1, 2));
+        downgrading.leader_state = Some(LeaderState::Downgrading);
+        let mut staging = new_test_region_route(RegionId::new(1, 3));
+        staging.leader_state = Some(LeaderState::Staging);
+
+        assert_eq!(normal.leader_region_role(), Some(RegionRole::Leader));
+        assert_eq!(
+            downgrading.leader_region_role(),
+            Some(RegionRole::DowngradingLeader)
+        );
+        assert_eq!(
+            staging.leader_region_role(),
+            Some(RegionRole::StagingLeader)
+        );
+    }
+
+    #[test]
+    fn test_operating_leader_region_roles_returns_expected_roles() {
+        let no_leader_region = RegionRoute {
+            leader_peer: None,
+            ..new_test_region_route(RegionId::new(1, 4))
+        };
+        let mut downgrading = new_test_region_route(RegionId::new(1, 2));
+        downgrading.leader_peer = Some(Peer::new(2, "a2"));
+        downgrading.leader_state = Some(LeaderState::Downgrading);
+        let mut staging = new_test_region_route(RegionId::new(1, 3));
+        staging.leader_peer = Some(Peer::new(3, "a3"));
+        staging.leader_state = Some(LeaderState::Staging);
+
+        let roles = operating_leader_region_roles(&[
+            new_test_region_route(RegionId::new(1, 1)),
+            downgrading,
+            staging,
+            no_leader_region,
+        ]);
+
+        assert_eq!(
+            roles,
+            vec![
+                (RegionId::new(1, 1), 1, RegionRole::Leader),
+                (RegionId::new(1, 2), 2, RegionRole::DowngradingLeader),
+                (RegionId::new(1, 3), 3, RegionRole::StagingLeader),
+            ]
+        );
+    }
+
     #[test]
     fn test_region_distribution() {
         let region_routes = vec![
diff --git a/src/common/meta/src/sequence.rs b/src/common/meta/src/sequence.rs
index d186446fda..cd0a8ebf88 100644
--- a/src/common/meta/src/sequence.rs
+++ b/src/common/meta/src/sequence.rs
@@ -337,7 +337,7 @@ impl Inner {
 #[cfg(test)]
 mod tests {
     use std::any::Any;
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashSet;
     use std::sync::Arc;
 
diff --git a/src/common/meta/src/snapshot.rs b/src/common/meta/src/snapshot.rs
index 8f4818e33a..e66156b255 100644
--- a/src/common/meta/src/snapshot.rs
+++ b/src/common/meta/src/snapshot.rs
@@ -355,7 +355,7 @@ impl MetadataSnapshotManager {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::sync::Arc;
 
     use common_test_util::temp_dir::{TempDir, create_temp_dir};
diff --git a/src/common/meta/src/state_store.rs b/src/common/meta/src/state_store.rs
index 0ecbd5a8e4..d98a286581 100644
--- a/src/common/meta/src/state_store.rs
+++ b/src/common/meta/src/state_store.rs
@@ -380,9 +380,8 @@ impl PoisonStore for KvStateStore {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
-    use std::env;
     use std::sync::Arc;
+    use std::{assert_matches, env};
 
     use common_procedure::store::state_store::KeyValue;
     use common_telemetry::info;
diff --git a/src/common/meta/src/wal_provider.rs b/src/common/meta/src/wal_provider.rs
index cd599b58ba..bf2cb9ba32 100644
--- a/src/common/meta/src/wal_provider.rs
+++ b/src/common/meta/src/wal_provider.rs
@@ -172,7 +172,7 @@ pub fn extract_topic_from_wal_options(
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_wal::config::kafka::MetasrvKafkaConfig;
     use common_wal::config::kafka::common::KafkaTopicConfig;
diff --git a/src/common/meta/src/wal_provider/topic_pool.rs b/src/common/meta/src/wal_provider/topic_pool.rs
index 919f0b2abe..f9b4863e52 100644
--- a/src/common/meta/src/wal_provider/topic_pool.rs
+++ b/src/common/meta/src/wal_provider/topic_pool.rs
@@ -136,7 +136,7 @@ impl KafkaTopicPool {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_wal::maybe_skip_kafka_integration_test;
     use common_wal::test_util::get_kafka_endpoints;
diff --git a/src/common/options/src/plugin_options.rs b/src/common/options/src/plugin_options.rs
index e3163d3e54..07b7e061f8 100644
--- a/src/common/options/src/plugin_options.rs
+++ b/src/common/options/src/plugin_options.rs
@@ -27,3 +27,10 @@ pub type PluginOptionsSerializerRef = Arc<dyn PluginOptionsSerializer>;
 pub trait PluginOptionsDeserializer<T: DeserializeOwned>: Send + Sync {
     fn deserialize(&self, payload: &str) -> Result<T, serde_json::Error>;
 }
+
+/// A flag for stating the standalone mode in the plugins.
+///
+/// The standalone build and start process calls `setup_frontend_plugins` and `setup_datanode_plugins`,
+/// so we add a flag to the plugins to indicate that the plugins are running in the standalone mode.
+#[derive(Clone, Copy, Debug)]
+pub struct StandaloneFlag;
diff --git a/src/common/procedure/src/lib.rs b/src/common/procedure/src/lib.rs
index e322765558..156a0ab78c 100644
--- a/src/common/procedure/src/lib.rs
+++ b/src/common/procedure/src/lib.rs
@@ -14,8 +14,6 @@
 
 //! Common traits and structures for the procedure framework.
 
-#![feature(assert_matches)]
-
 pub mod error;
 pub mod event;
 pub mod local;
diff --git a/src/common/procedure/src/local.rs b/src/common/procedure/src/local.rs
index fe86cd7993..9e8536308c 100644
--- a/src/common/procedure/src/local.rs
+++ b/src/common/procedure/src/local.rs
@@ -920,7 +920,7 @@ pub(crate) mod test_util {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_error::mock::MockError;
     use common_error::status_code::StatusCode;
diff --git a/src/common/procedure/src/local/runner.rs b/src/common/procedure/src/local/runner.rs
index 46dcef11d4..ca3e221f43 100644
--- a/src/common/procedure/src/local/runner.rs
+++ b/src/common/procedure/src/local/runner.rs
@@ -704,7 +704,7 @@ impl Runner {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::sync::Arc;
     use std::sync::atomic::{AtomicU64, Ordering};
 
@@ -1293,6 +1293,60 @@ mod tests {
         .await;
     }
 
+    #[tokio::test]
+    async fn test_retrying_state_visible_in_context_on_retry() {
+        let retrying_states = Arc::new(std::sync::Mutex::new(Vec::new()));
+        let captured = retrying_states.clone();
+        let mut times = 0;
+
+        let exec_fn = move |ctx: Context| {
+            times += 1;
+            let captured = captured.clone();
+            async move {
+                let is_retrying = ctx.is_retrying().await;
+                captured.lock().unwrap().push(is_retrying);
+                if times == 1 {
+                    Err(Error::retry_later(MockError::new(StatusCode::Unexpected)))
+                } else {
+                    Ok(Status::done())
+                }
+            }
+            .boxed()
+        };
+
+        let procedure = ProcedureAdapter {
+            data: "retrying_state".to_string(),
+            lock_key: LockKey::single_exclusive("catalog.schema.table"),
+            poison_keys: PoisonKeys::default(),
+            exec_fn,
+            rollback_fn: None,
+        };
+
+        let dir = create_temp_dir("retrying_state");
+        let meta = procedure.new_meta(ROOT_ID);
+        let object_store = test_util::new_object_store(&dir);
+        let procedure_store = Arc::new(ProcedureStore::from_object_store(object_store));
+        let mut runner = new_runner(meta.clone(), Box::new(procedure), procedure_store);
+        let ctx = context_with_provider(
+            meta.id,
+            runner.manager_ctx.clone() as Arc<dyn ContextProvider>,
+        );
+
+        runner
+            .manager_ctx
+            .procedures
+            .write()
+            .unwrap()
+            .insert(meta.id, runner.meta.clone());
+        runner.manager_ctx.start();
+
+        runner.execute_once(&ctx).await;
+        runner.execute_once(&ctx).await;
+
+        let states = retrying_states.lock().unwrap().clone();
+        assert_eq!(states, vec![Some(false), Some(true)]);
+    }
+
     #[tokio::test(flavor = "multi_thread")]
     async fn test_execute_on_retry_later_error_with_child() {
         common_telemetry::init_default_ut_logging();
diff --git a/src/common/procedure/src/procedure.rs b/src/common/procedure/src/procedure.rs
index 843052ed97..8e34f4bb30 100644
--- a/src/common/procedure/src/procedure.rs
+++ b/src/common/procedure/src/procedure.rs
@@ -177,6 +177,18 @@ pub struct Context {
     pub provider: ContextProviderRef,
 }
 
+impl Context {
+    /// Returns true if current procedure state is retrying.
+    pub async fn is_retrying(&self) -> Option<bool> {
+        self.provider
+            .procedure_state(self.procedure_id)
+            .await
+            .ok()
+            .flatten()
+            .map(|s| s.is_retrying())
+    }
+}
+
 /// A `Procedure` represents an operation or a set of operations to be performed step-by-step.
 #[async_trait]
 pub trait Procedure: Send {
diff --git a/src/common/procedure/src/store/util.rs b/src/common/procedure/src/store/util.rs
index e6ef5b62ec..c44e5f4712 100644
--- a/src/common/procedure/src/store/util.rs
+++ b/src/common/procedure/src/store/util.rs
@@ -57,7 +57,7 @@ fn merge_multiple_values(
         let (key, value) = pairs.into_iter().next().unwrap();
         let prefix = KeySet::with_prefix(&key);
         let mut parsed_segments = parse_segments(segments, &prefix)?;
-        parsed_segments.sort_unstable_by(|a, b| a.0.cmp(&b.0));
+        parsed_segments.sort_unstable_by_key(|a| a.0);
 
         // Safety: `parsed_segments` must larger than 0.
         let segment_num = parsed_segments.last().unwrap().0;
@@ -133,7 +133,7 @@ pub fn multiple_value_stream(
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use futures::TryStreamExt;
     use futures::stream::{self};
diff --git a/src/common/query/src/prelude.rs b/src/common/query/src/prelude.rs
index c27b94294e..50668bbbb1 100644
--- a/src/common/query/src/prelude.rs
+++ b/src/common/query/src/prelude.rs
@@ -27,7 +27,16 @@ static GREPTIME_TIMESTAMP_CELL: OnceCell<String> = OnceCell::new();
 static GREPTIME_VALUE_CELL: OnceCell<String> = OnceCell::new();
 
 pub fn set_default_prefix(prefix: Option<&str>) -> Result<()> {
-    match prefix {
+    // Strip surrounding double quotes as a defensive measure against upstream
+    // sources (scripts, CI, template engines, incorrect shell escaping) that may
+    // pass literal `""` as the value instead of an empty string.
+    let stripped = prefix.map(|s| {
+        s.strip_prefix('"')
+            .and_then(|s| s.strip_suffix('"'))
+            .unwrap_or(s)
+    });
+
+    match stripped {
         None => {
             // use default greptime prefix
             GREPTIME_TIMESTAMP_CELL.get_or_init(|| GREPTIME_TIMESTAMP.to_string());
@@ -70,3 +79,45 @@ const GREPTIME_VALUE: &str = "greptime_value";
 pub const GREPTIME_COUNT: &str = "greptime_count";
 /// Default physical table name
 pub const GREPTIME_PHYSICAL_TABLE: &str = "greptime_physical_table";
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    // Each test runs in a separate process via `cargo nextest`, so OnceCell
+    // state does not leak between tests.
+
+    #[test]
+    fn test_set_default_prefix_none() {
+        set_default_prefix(None).unwrap();
+        assert_eq!(greptime_timestamp(), "greptime_timestamp");
+        assert_eq!(greptime_value(), "greptime_value");
+    }
+
+    #[test]
+    fn test_set_default_prefix_empty_string() {
+        set_default_prefix(Some("")).unwrap();
+        assert_eq!(greptime_timestamp(), "timestamp");
+        assert_eq!(greptime_value(), "value");
+    }
+
+    #[test]
+    fn test_set_default_prefix_quoted_empty() {
+        // Handles upstream sources that pass literal `""` instead of an empty string
+        set_default_prefix(Some("\"\"")).unwrap();
+        assert_eq!(greptime_timestamp(), "timestamp");
+        assert_eq!(greptime_value(), "value");
+    }
+
+    #[test]
+    fn test_set_default_prefix_custom() {
+        set_default_prefix(Some("mydb")).unwrap();
+        assert_eq!(greptime_timestamp(), "mydb_timestamp");
+        assert_eq!(greptime_value(), "mydb_value");
+    }
+
+    #[test]
+    fn test_set_default_prefix_invalid() {
+        assert!(set_default_prefix(Some("invalid prefix!")).is_err());
+    }
+}
diff --git a/src/common/recordbatch/Cargo.toml b/src/common/recordbatch/Cargo.toml
index 5887dc31c5..efc6b6f60e 100644
--- a/src/common/recordbatch/Cargo.toml
+++ b/src/common/recordbatch/Cargo.toml
@@ -12,6 +12,7 @@ arc-swap = "1.6"
 common-base.workspace = true
 common-error.workspace = true
 common-macro.workspace = true
+common-memory-manager.workspace = true
 common-telemetry.workspace = true
 common-time.workspace = true
 datafusion.workspace = true
diff --git a/src/common/recordbatch/src/adapter.rs b/src/common/recordbatch/src/adapter.rs
index fc12d87dcf..a69eedf7eb 100644
--- a/src/common/recordbatch/src/adapter.rs
+++ b/src/common/recordbatch/src/adapter.rs
@@ -446,6 +446,27 @@ pub struct RecordBatchMetrics {
     // Detailed per-plan metrics
     /// An ordered list of plan metrics, from top to bottom in post-order.
     pub plan_metrics: Vec<PlanMetrics>,
+    /// Per-region watermark for incremental-read checkpoint advancement.
+    ///
+    /// The watermark is the latest sequence (`seq`) this query round safely read
+    /// for each participating region. Flow uses it to decide where the next
+    /// incremental round can resume.
+    ///
+    /// - `Some(seq)`: the query proved it safely read up to `seq`; downstream
+    ///   may advance the checkpoint to this value.
+    /// - `None`: the region participated but the query could not prove a safe
+    ///   read upper-bound, so the checkpoint must not advance for this region.
+    ///
+    /// Omitted when empty for backward compatibility.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub region_watermarks: Vec<RegionWatermarkEntry>,
+}
+
+#[derive(serde::Serialize, serde::Deserialize, Debug, Clone, PartialEq, Eq, PartialOrd, Ord)]
+pub struct RegionWatermarkEntry {
+    pub region_id: u64,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub watermark: Option<u64>,
 }
 
 /// Determines if a metric name represents a time measurement that should be formatted.
@@ -706,6 +727,7 @@ mod test {
     use datatypes::prelude::ConcreteDataType;
     use datatypes::schema::ColumnSchema;
     use datatypes::vectors::Int32Vector;
+    use serde_json::json;
     use snafu::IntoError;
 
     use super::*;
@@ -879,4 +901,53 @@ mod test {
             }
         }
     }
+
+    #[test]
+    fn test_recordbatch_metrics_deserializes_without_region_watermarks() {
+        let metrics: RecordBatchMetrics = serde_json::from_value(json!({
+            "elapsed_compute": 12,
+            "memory_usage": 34,
+            "plan_metrics": []
+        }))
+        .unwrap();
+
+        assert!(metrics.region_watermarks.is_empty());
+        assert_eq!(metrics.elapsed_compute, 12);
+        assert_eq!(metrics.memory_usage, 34);
+    }
+
+    #[test]
+    fn test_recordbatch_metrics_region_watermarks_serde_roundtrip() {
+        let metrics = RecordBatchMetrics {
+            region_watermarks: vec![
+                RegionWatermarkEntry {
+                    region_id: 1,
+                    watermark: Some(100),
+                },
+                RegionWatermarkEntry {
+                    region_id: 2,
+                    watermark: None,
+                },
+            ],
+            ..Default::default()
+        };
+
+        let value = serde_json::to_value(&metrics).unwrap();
+        assert_eq!(
+            value.get("region_watermarks").unwrap(),
+            &json!([
+                { "region_id": 1, "watermark": 100 },
+                { "region_id": 2 }
+            ])
+        );
+
+        let decoded: RecordBatchMetrics = serde_json::from_value(value).unwrap();
+        assert_eq!(decoded.region_watermarks, metrics.region_watermarks);
+    }
+
+    #[test]
+    fn test_recordbatch_metrics_skips_empty_region_watermarks_on_serialize() {
+        let value = serde_json::to_value(RecordBatchMetrics::default()).unwrap();
+        assert!(value.get("region_watermarks").is_none());
+    }
 }
diff --git a/src/common/recordbatch/src/error.rs b/src/common/recordbatch/src/error.rs
index 6d794463a0..00d4291ead 100644
--- a/src/common/recordbatch/src/error.rs
+++ b/src/common/recordbatch/src/error.rs
@@ -188,13 +188,6 @@ pub enum Error {
         #[snafu(implicit)]
         location: Location,
     },
-
-    #[snafu(display("Failed to align JSON array, reason: {reason}"))]
-    AlignJsonArray {
-        reason: String,
-        #[snafu(implicit)]
-        location: Location,
-    },
 }
 
 impl ErrorExt for Error {
@@ -210,8 +203,7 @@ impl ErrorExt for Error {
             | Error::ToArrowScalar { .. }
             | Error::ProjectArrowRecordBatch { .. }
             | Error::PhysicalExpr { .. }
-            | Error::RecordBatchSliceIndexOverflow { .. }
-            | Error::AlignJsonArray { .. } => StatusCode::Internal,
+            | Error::RecordBatchSliceIndexOverflow { .. } => StatusCode::Internal,
 
             Error::PollStream { .. } => StatusCode::EngineExecuteQuery,
 
diff --git a/src/common/recordbatch/src/filter.rs b/src/common/recordbatch/src/filter.rs
index d7c522e656..9f1b596a49 100644
--- a/src/common/recordbatch/src/filter.rs
+++ b/src/common/recordbatch/src/filter.rs
@@ -48,7 +48,7 @@ use crate::error::{ArrowComputeSnafu, Result, ToArrowScalarSnafu, UnsupportedOpe
 ///
 /// This struct contains normalized predicate expr. In the form of
 /// `col` `op` `literal` where the `col` is provided from input.
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 pub struct SimpleFilterEvaluator {
     /// Name of the referenced column.
     column_name: String,
diff --git a/src/common/recordbatch/src/lib.rs b/src/common/recordbatch/src/lib.rs
index 85e0d5c496..629efd6d84 100644
--- a/src/common/recordbatch/src/lib.rs
+++ b/src/common/recordbatch/src/lib.rs
@@ -22,13 +22,17 @@ pub mod recordbatch;
 pub mod util;
 
 use std::fmt;
+use std::future::Future;
 use std::pin::Pin;
 use std::sync::Arc;
-use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
 
 use adapter::RecordBatchMetrics;
 use arc_swap::ArcSwapOption;
 use common_base::readable_size::ReadableSize;
+use common_error::ext::BoxedError;
+use common_memory_manager::{
+    MemoryGuard, MemoryManager, MemoryMetrics, OnExhaustedPolicy, PermitGranularity,
+};
 use common_telemetry::tracing::Span;
 pub use datafusion::physical_plan::SendableRecordBatchStream as DfSendableRecordBatchStream;
 use datatypes::arrow::array::{ArrayRef, AsArray, StringBuilder};
@@ -42,7 +46,7 @@ use error::Result;
 use futures::task::{Context, Poll};
 use futures::{Stream, TryStreamExt};
 pub use recordbatch::RecordBatch;
-use snafu::{ResultExt, ensure};
+use snafu::{IntoError, ResultExt, ensure};
 
 use crate::error::NewDfRecordBatchSnafu;
 
@@ -416,205 +420,100 @@ impl<S: Stream<Item = Result<RecordBatch>> + Unpin> Stream for RecordBatchStream
     }
 }
 
-/// Memory permit for a stream, providing privileged access or rate limiting.
-///
-/// The permit tracks whether this stream has privileged Top-K status.
-/// When dropped, it automatically releases any privileged slot it holds.
-pub struct MemoryPermit {
-    tracker: QueryMemoryTracker,
-    is_privileged: AtomicBool,
-}
-
-impl MemoryPermit {
-    /// Check if this permit currently has privileged status.
-    pub fn is_privileged(&self) -> bool {
-        self.is_privileged.load(Ordering::Acquire)
-    }
-
-    /// Ensure this permit has privileged status by acquiring a slot if available.
-    /// Returns true if privileged (either already privileged or just acquired privilege).
-    fn ensure_privileged(&self) -> bool {
-        if self.is_privileged.load(Ordering::Acquire) {
-            return true;
-        }
-
-        // Try to claim a privileged slot
-        self.tracker
-            .privileged_count
-            .fetch_update(Ordering::AcqRel, Ordering::Acquire, |count| {
-                if count < self.tracker.privileged_slots {
-                    Some(count + 1)
-                } else {
-                    None
-                }
-            })
-            .map(|_| {
-                self.is_privileged.store(true, Ordering::Release);
-                true
-            })
-            .unwrap_or(false)
-    }
-
-    /// Track additional memory usage with this permit.
-    /// Returns error if limit is exceeded.
-    ///
-    /// # Arguments
-    /// * `additional` - Additional memory size to track in bytes
-    /// * `stream_tracked` - Total memory already tracked by this stream
-    ///
-    /// # Behavior
-    /// - Privileged streams: Can push global memory usage up to full limit
-    /// - Standard-tier streams: Can push global memory usage up to limit * standard_tier_memory_fraction (default: 0.7)
-    /// - Standard-tier streams automatically attempt to acquire privilege if slots become available
-    /// - The configured limit is absolute hard limit - no stream can exceed it
-    pub fn track(&self, additional: usize, stream_tracked: usize) -> Result<()> {
-        // Ensure privileged status if possible
-        let is_privileged = self.ensure_privileged();
-
-        self.tracker
-            .track_internal(additional, is_privileged, stream_tracked)
-    }
-
-    /// Release tracked memory.
-    ///
-    /// # Arguments
-    /// * `amount` - Amount of memory to release in bytes
-    pub fn release(&self, amount: usize) {
-        self.tracker.release(amount);
-    }
-}
-
-impl Drop for MemoryPermit {
-    fn drop(&mut self) {
-        // Release privileged slot if we had one
-        if self.is_privileged.load(Ordering::Acquire) {
-            self.tracker
-                .privileged_count
-                .fetch_sub(1, Ordering::Release);
-        }
-    }
-}
-
 /// Memory tracker for RecordBatch streams. Clone to share the same limit across queries.
 ///
-/// Implements a two-tier memory allocation strategy:
-/// - **Privileged tier**: First N streams (default: 20) can use up to the full memory limit
-/// - **Standard tier**: Remaining streams are restricted to a fraction of the limit (default: 70%)
-/// - Privilege is granted on a first-come-first-served basis
-/// - The configured limit is an absolute hard cap - no stream can exceed it
+/// Each stream acquires quota independently from this tracker.
 #[derive(Clone)]
 pub struct QueryMemoryTracker {
-    current: Arc<AtomicUsize>,
-    limit: usize,
-    standard_tier_memory_fraction: f64,
-    privileged_count: Arc<AtomicUsize>,
-    privileged_slots: usize,
-    on_update: Option<Arc<dyn Fn(usize) + Send + Sync>>,
-    on_reject: Option<Arc<dyn Fn() + Send + Sync>>,
+    manager: MemoryManager<CallbackMemoryMetrics>,
+    metrics: CallbackMemoryMetrics,
+    on_exhausted_policy: OnExhaustedPolicy,
 }
 
 impl fmt::Debug for QueryMemoryTracker {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         f.debug_struct("QueryMemoryTracker")
-            .field("current", &self.current.load(Ordering::Acquire))
-            .field("limit", &self.limit)
-            .field(
-                "standard_tier_memory_fraction",
-                &self.standard_tier_memory_fraction,
-            )
-            .field(
-                "privileged_count",
-                &self.privileged_count.load(Ordering::Acquire),
-            )
-            .field("privileged_slots", &self.privileged_slots)
-            .field("on_update", &self.on_update.is_some())
-            .field("on_reject", &self.on_reject.is_some())
+            .field("current", &self.current())
+            .field("limit", &self.limit())
+            .field("on_exhausted_policy", &self.on_exhausted_policy)
+            .field("on_update", &self.metrics.has_on_update())
+            .field("on_exhausted", &self.metrics.has_on_exhausted())
+            .field("on_rejected", &self.metrics.has_on_rejected())
             .finish()
     }
 }
 
 impl QueryMemoryTracker {
-    // Default privileged slots when max_concurrent_queries is 0.
-    const DEFAULT_PRIVILEGED_SLOTS: usize = 20;
-    // Ratio for privileged tier: 70% queries get privileged access, standard tier uses 70% memory.
-    const DEFAULT_PRIVILEGED_TIER_RATIO: f64 = 0.7;
-
-    /// Create a new memory tracker with the given limit and max_concurrent_queries.
-    /// Calculates privileged slots as 70% of max_concurrent_queries (or 20 if max_concurrent_queries is 0).
-    ///
-    /// # Arguments
-    /// * `limit` - Maximum memory usage in bytes (hard limit for all streams). 0 means unlimited.
-    /// * `max_concurrent_queries` - Maximum number of concurrent queries (0 = unlimited).
-    pub fn new(limit: usize, max_concurrent_queries: usize) -> Self {
-        let privileged_slots = Self::calculate_privileged_slots(max_concurrent_queries);
-        Self::with_privileged_slots(limit, privileged_slots)
-    }
-
-    /// Create a new memory tracker with custom privileged slots limit.
-    pub fn with_privileged_slots(limit: usize, privileged_slots: usize) -> Self {
-        Self::with_config(limit, privileged_slots, Self::DEFAULT_PRIVILEGED_TIER_RATIO)
-    }
-
-    /// Create a new memory tracker with full configuration.
-    ///
-    /// # Arguments
-    /// * `limit` - Maximum memory usage in bytes (hard limit for all streams). 0 means unlimited.
-    /// * `privileged_slots` - Maximum number of streams that can get privileged status.
-    /// * `standard_tier_memory_fraction` - Memory fraction for standard-tier streams (range: [0.0, 1.0]).
-    ///
-    /// # Panics
-    /// Panics if `standard_tier_memory_fraction` is not in the range [0.0, 1.0].
-    pub fn with_config(
+    /// Create a builder for a query memory tracker.
+    pub fn builder(
         limit: usize,
-        privileged_slots: usize,
-        standard_tier_memory_fraction: f64,
-    ) -> Self {
-        assert!(
-            (0.0..=1.0).contains(&standard_tier_memory_fraction),
-            "standard_tier_memory_fraction must be in [0.0, 1.0], got {}",
-            standard_tier_memory_fraction
-        );
-
-        Self {
-            current: Arc::new(AtomicUsize::new(0)),
+        on_exhausted_policy: OnExhaustedPolicy,
+    ) -> QueryMemoryTrackerBuilder {
+        QueryMemoryTrackerBuilder {
             limit,
-            standard_tier_memory_fraction,
-            privileged_count: Arc::new(AtomicUsize::new(0)),
-            privileged_slots,
+            on_exhausted_policy,
             on_update: None,
+            on_exhausted: None,
             on_reject: None,
         }
     }
 
-    /// Register a new permit for memory tracking.
-    /// The first `privileged_slots` permits get privileged status automatically.
-    /// The returned permit can be shared across multiple streams of the same query.
-    pub fn register_permit(&self) -> MemoryPermit {
-        // Try to claim a privileged slot
-        let is_privileged = self
-            .privileged_count
-            .fetch_update(Ordering::AcqRel, Ordering::Acquire, |count| {
-                if count < self.privileged_slots {
-                    Some(count + 1)
-                } else {
-                    None
-                }
-            })
-            .is_ok();
-
-        MemoryPermit {
+    fn new_stream_tracker(&self) -> StreamMemoryTracker {
+        StreamMemoryTracker {
             tracker: self.clone(),
-            is_privileged: AtomicBool::new(is_privileged),
+            guard: self.manager.try_acquire(0).unwrap(),
+            tracked_bytes: 0,
         }
     }
+    /// Get the current memory usage in bytes.
+    pub fn current(&self) -> usize {
+        self.manager.used_bytes() as usize
+    }
 
+    fn limit(&self) -> usize {
+        self.manager.limit_bytes() as usize
+    }
+
+    fn reject_error(
+        &self,
+        current: usize,
+        additional: usize,
+        stream_tracked: usize,
+    ) -> error::Error {
+        let limit = self.limit();
+        let msg = format!(
+            "{} requested, {} used globally ({}%), {} used by this stream, hard limit: {}",
+            ReadableSize(additional as u64),
+            ReadableSize(current as u64),
+            (current * 100).checked_div(limit).unwrap_or(0),
+            ReadableSize(stream_tracked as u64),
+            ReadableSize(limit as u64)
+        );
+        error::ExceedMemoryLimitSnafu { msg }.build()
+    }
+
+    fn inc_rejected(&self) {
+        self.metrics.inc_rejected();
+    }
+}
+
+/// Builder for constructing a [`QueryMemoryTracker`] with optional callbacks.
+pub struct QueryMemoryTrackerBuilder {
+    limit: usize,
+    on_exhausted_policy: OnExhaustedPolicy,
+    on_update: Option<UpdateCallback>,
+    on_exhausted: Option<UnitCallback>,
+    on_reject: Option<RejectCallback>,
+}
+
+impl QueryMemoryTrackerBuilder {
     /// Set a callback to be called whenever the usage changes successfully.
     /// The callback receives the new total usage in bytes.
     ///
     /// # Note
-    /// The callback is called after both successful `track()` and `release()` operations.
-    /// It is called even when `limit == 0` (unlimited mode) to track actual usage.
-    pub fn with_on_update<F>(mut self, on_update: F) -> Self
+    /// The callback is called after both successful `track()` and stream drop.
+    /// Usage is exact in unlimited mode and 1KB-aligned in limited mode.
+    pub fn on_update<F>(mut self, on_update: F) -> Self
     where
         F: Fn(usize) + Send + Sync + 'static,
     {
@@ -622,12 +521,22 @@ impl QueryMemoryTracker {
         self
     }
 
-    /// Set a callback to be called when memory allocation is rejected.
+    /// Set a callback to be called when memory is unavailable for immediate acquisition.
     ///
     /// # Note
-    /// This is only called when `track()` fails due to exceeding the limit.
+    /// This is called when the non-blocking allocation fast path fails.
+    /// Requests using `OnExhaustedPolicy::Wait` may still succeed after waiting.
     /// It is never called when `limit == 0` (unlimited mode).
-    pub fn with_on_reject<F>(mut self, on_reject: F) -> Self
+    pub fn on_exhausted<F>(mut self, on_exhausted: F) -> Self
+    where
+        F: Fn() + Send + Sync + 'static,
+    {
+        self.on_exhausted = Some(Arc::new(on_exhausted));
+        self
+    }
+
+    /// Set a callback to be called when the request ultimately fails due to memory pressure.
+    pub fn on_reject<F>(mut self, on_reject: F) -> Self
     where
         F: Fn() + Send + Sync + 'static,
     {
@@ -635,105 +544,151 @@ impl QueryMemoryTracker {
         self
     }
 
-    /// Get the current memory usage in bytes.
-    pub fn current(&self) -> usize {
-        self.current.load(Ordering::Acquire)
+    /// Build a [`QueryMemoryTracker`] from this builder.
+    pub fn build(self) -> QueryMemoryTracker {
+        let metrics = CallbackMemoryMetrics::new(self.on_update, self.on_exhausted, self.on_reject);
+        let manager = MemoryManager::with_granularity(
+            self.limit as u64,
+            PermitGranularity::Kilobyte,
+            metrics.clone(),
+        );
+
+        QueryMemoryTracker {
+            manager,
+            metrics,
+            on_exhausted_policy: self.on_exhausted_policy,
+        }
+    }
+}
+
+struct StreamMemoryTracker {
+    tracker: QueryMemoryTracker,
+    guard: MemoryGuard<CallbackMemoryMetrics>,
+    tracked_bytes: usize,
+}
+
+type MemoryAcquireResult = std::result::Result<(), common_memory_manager::Error>;
+
+impl StreamMemoryTracker {
+    fn inc_rejected(&self) {
+        self.tracker.inc_rejected();
     }
 
-    fn calculate_privileged_slots(max_concurrent_queries: usize) -> usize {
-        if max_concurrent_queries == 0 {
-            Self::DEFAULT_PRIVILEGED_SLOTS
+    fn try_track(&mut self, additional: usize) -> Result<()> {
+        if self.guard.try_acquire_additional(additional as u64) {
+            self.tracked_bytes = self.tracked_bytes.saturating_add(additional);
+            Ok(())
         } else {
-            ((max_concurrent_queries as f64 * Self::DEFAULT_PRIVILEGED_TIER_RATIO) as usize).max(1)
+            Err(self.reject_error(additional))
         }
     }
 
-    /// Internal method to track additional memory usage.
-    ///
-    /// Called by `MemoryPermit::track()`. Use `MemoryPermit::track()` instead of calling this directly.
-    fn track_internal(
-        &self,
-        additional: usize,
-        is_privileged: bool,
-        stream_tracked: usize,
-    ) -> Result<()> {
-        // Calculate effective global limit based on stream privilege
-        // Privileged streams: can push global usage up to full limit
-        // Standard-tier streams: can only push global usage up to fraction of limit
-        let effective_limit = if is_privileged {
-            self.limit
-        } else {
-            (self.limit as f64 * self.standard_tier_memory_fraction) as usize
-        };
-
-        let mut new_total = 0;
+    async fn track_with_policy(mut self, additional: usize) -> (Self, MemoryAcquireResult) {
         let result = self
-            .current
-            .fetch_update(Ordering::AcqRel, Ordering::Acquire, |current| {
-                new_total = current.saturating_add(additional);
+            .guard
+            .acquire_additional_with_policy(additional as u64, self.tracker.on_exhausted_policy)
+            .await;
+        if result.is_ok() {
+            self.tracked_bytes = self.tracked_bytes.saturating_add(additional);
+        }
+        (self, result)
+    }
 
-                if self.limit == 0 {
-                    // Unlimited mode
-                    return Some(new_total);
-                }
+    fn reject_error(&self, additional: usize) -> error::Error {
+        let current = self.tracker.current();
+        self.tracker
+            .reject_error(current, additional, self.tracked_bytes)
+    }
 
-                // Check if new global total exceeds effective limit
-                // The configured limit is absolute hard limit - no stream can exceed it
-                if new_total <= effective_limit {
-                    Some(new_total)
-                } else {
-                    None
-                }
-            });
-
-        match result {
-            Ok(_) => {
-                if let Some(callback) = &self.on_update {
-                    callback(new_total);
-                }
-                Ok(())
+    fn wait_error(&self, additional: usize, source: common_memory_manager::Error) -> error::Error {
+        match source {
+            common_memory_manager::Error::MemoryLimitExceeded { .. } => {
+                self.reject_error(additional)
             }
-            Err(current) => {
-                if let Some(callback) = &self.on_reject {
-                    callback();
-                }
+            common_memory_manager::Error::MemoryAcquireTimeout { waited, .. } => {
+                let current = self.tracker.current();
+                let limit = self.tracker.limit();
                 let msg = format!(
-                    "{} requested, {} used globally ({}%), {} used by this stream (privileged: {}), effective limit: {} ({}%), hard limit: {}",
+                    "timed out waiting {:?} for {}, {} used globally ({}%), {} used by this stream, hard limit: {}",
+                    waited,
                     ReadableSize(additional as u64),
                     ReadableSize(current as u64),
-                    if self.limit > 0 {
-                        current * 100 / self.limit
-                    } else {
-                        0
-                    },
-                    ReadableSize(stream_tracked as u64),
-                    is_privileged,
-                    ReadableSize(effective_limit as u64),
-                    if self.limit > 0 {
-                        effective_limit * 100 / self.limit
-                    } else {
-                        0
-                    },
-                    ReadableSize(self.limit as u64)
+                    (current * 100).checked_div(limit).unwrap_or(0),
+                    ReadableSize(self.tracked_bytes as u64),
+                    ReadableSize(limit as u64)
                 );
-                error::ExceedMemoryLimitSnafu { msg }.fail()
+                error::ExceedMemoryLimitSnafu { msg }.build()
             }
+            error => error::ExternalSnafu.into_error(BoxedError::new(error)),
+        }
+    }
+}
+
+type PendingTrackFuture = Pin<
+    Box<dyn Future<Output = (StreamMemoryTracker, RecordBatch, usize, MemoryAcquireResult)> + Send>,
+>;
+
+#[derive(Clone)]
+struct CallbackMemoryMetrics {
+    inner: Arc<CallbackMemoryMetricsInner>,
+}
+
+type UpdateCallback = Arc<dyn Fn(usize) + Send + Sync>;
+type UnitCallback = Arc<dyn Fn() + Send + Sync>;
+type RejectCallback = UnitCallback;
+
+struct CallbackMemoryMetricsInner {
+    on_update: Option<UpdateCallback>,
+    on_exhausted: Option<UnitCallback>,
+    on_reject: Option<RejectCallback>,
+}
+
+impl CallbackMemoryMetrics {
+    fn new(
+        on_update: Option<UpdateCallback>,
+        on_exhausted: Option<UnitCallback>,
+        on_reject: Option<RejectCallback>,
+    ) -> Self {
+        Self {
+            inner: Arc::new(CallbackMemoryMetricsInner {
+                on_update,
+                on_exhausted,
+                on_reject,
+            }),
         }
     }
 
-    /// Release tracked memory.
-    ///
-    /// # Arguments
-    /// * `amount` - Amount of memory to release in bytes
-    pub fn release(&self, amount: usize) {
-        if let Ok(old_value) =
-            self.current
-                .fetch_update(Ordering::AcqRel, Ordering::Acquire, |current| {
-                    Some(current.saturating_sub(amount))
-                })
-            && let Some(callback) = &self.on_update
-        {
-            callback(old_value.saturating_sub(amount));
+    fn has_on_update(&self) -> bool {
+        self.inner.on_update.is_some()
+    }
+
+    fn has_on_exhausted(&self) -> bool {
+        self.inner.on_exhausted.is_some()
+    }
+
+    fn has_on_rejected(&self) -> bool {
+        self.inner.on_reject.is_some()
+    }
+
+    fn inc_rejected(&self) {
+        if let Some(callback) = &self.inner.on_reject {
+            callback();
+        }
+    }
+}
+
+impl MemoryMetrics for CallbackMemoryMetrics {
+    fn set_limit(&self, _: i64) {}
+
+    fn set_in_use(&self, bytes: i64) {
+        if let Some(callback) = &self.inner.on_update {
+            callback(bytes.max(0) as usize);
+        }
+    }
+
+    fn inc_exhausted(&self, _: &str) {
+        if let Some(callback) = &self.inner.on_exhausted {
+            callback();
         }
     }
 }
@@ -741,38 +696,113 @@ impl QueryMemoryTracker {
 /// A wrapper stream that tracks memory usage of RecordBatches.
 pub struct MemoryTrackedStream {
     inner: SendableRecordBatchStream,
-    permit: Arc<MemoryPermit>,
-    // Total tracked size, released when stream drops.
-    total_tracked: usize,
+    tracker: Option<StreamMemoryTracker>,
+    // Waiting stores a batch that has already been pulled from the inner stream but has not yet
+    // acquired additional quota. This keeps `poll_next()` non-blocking and allows bounded waits,
+    // at the cost of temporarily holding one untracked batch per blocked stream in memory.
+    waiting: Option<PendingTrackFuture>,
 }
 
 impl MemoryTrackedStream {
-    pub fn new(inner: SendableRecordBatchStream, permit: Arc<MemoryPermit>) -> Self {
+    pub fn new(inner: SendableRecordBatchStream, tracker: QueryMemoryTracker) -> Self {
         Self {
             inner,
-            permit,
-            total_tracked: 0,
+            tracker: Some(tracker.new_stream_tracker()),
+            waiting: None,
         }
     }
+
+    fn ready_tracker_mut(&mut self) -> &mut StreamMemoryTracker {
+        debug_assert!(
+            self.waiting.is_none(),
+            "a ready tracker must not coexist with a waiting future"
+        );
+        self.tracker.as_mut().unwrap()
+    }
+
+    fn enter_waiting(&mut self, batch: RecordBatch, additional: usize) {
+        debug_assert!(
+            self.waiting.is_none(),
+            "enter_waiting should only be called from the ready state"
+        );
+        debug_assert!(
+            self.tracker.is_some(),
+            "enter_waiting requires a tracker in the ready state"
+        );
+        let tracker = self.tracker.take().unwrap();
+        self.waiting = Some(Self::start_waiting(tracker, batch, additional));
+    }
+
+    fn start_waiting(
+        tracker: StreamMemoryTracker,
+        batch: RecordBatch,
+        additional: usize,
+    ) -> PendingTrackFuture {
+        Box::pin(async move {
+            let (tracker, result) = tracker.track_with_policy(additional).await;
+            (tracker, batch, additional, result)
+        })
+    }
+
+    fn poll_waiting(&mut self, cx: &mut Context<'_>) -> Poll<Option<Result<RecordBatch>>> {
+        let future = self.waiting.as_mut().unwrap();
+        match future.as_mut().poll(cx) {
+            Poll::Ready((tracker, batch, additional, result)) => {
+                let output = match result {
+                    Ok(()) => Ok(batch),
+                    Err(error) => {
+                        tracker.inc_rejected();
+                        Err(tracker.wait_error(additional, error))
+                    }
+                };
+                self.waiting = None;
+                self.tracker = Some(tracker);
+                Poll::Ready(Some(output))
+            }
+            Poll::Pending => Poll::Pending,
+        }
+    }
+
+    fn poll_batch(
+        &mut self,
+        batch: RecordBatch,
+        cx: &mut Context<'_>,
+    ) -> Poll<Option<Result<RecordBatch>>> {
+        let additional = batch.buffer_memory_size();
+        let tracker = self.ready_tracker_mut();
+
+        if let Err(error) = tracker.try_track(additional) {
+            match tracker.tracker.on_exhausted_policy {
+                OnExhaustedPolicy::Fail => {
+                    tracker.inc_rejected();
+                    return Poll::Ready(Some(Err(error)));
+                }
+                // `Wait` is a deliberate tradeoff: the batch has already been materialized, so we
+                // keep it in memory while waiting for quota instead of failing immediately. Under
+                // contention, real memory usage can therefore exceed `scan_memory_limit` by up to
+                // one buffered batch per blocked stream.
+                OnExhaustedPolicy::Wait { .. } => {
+                    self.enter_waiting(batch, additional);
+                    return self.poll_waiting(cx);
+                }
+            }
+        }
+
+        Poll::Ready(Some(Ok(batch)))
+    }
 }
 
 impl Stream for MemoryTrackedStream {
     type Item = Result<RecordBatch>;
 
     fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        if self.waiting.is_some() {
+            return self.poll_waiting(cx);
+        }
+
         match Pin::new(&mut self.inner).poll_next(cx) {
-            Poll::Ready(Some(Ok(batch))) => {
-                let additional = batch.buffer_memory_size();
-
-                if let Err(e) = self.permit.track(additional, self.total_tracked) {
-                    return Poll::Ready(Some(Err(e)));
-                }
-
-                self.total_tracked += additional;
-
-                Poll::Ready(Some(Ok(batch)))
-            }
-            Poll::Ready(Some(Err(e))) => Poll::Ready(Some(Err(e))),
+            Poll::Ready(Some(Ok(batch))) => self.poll_batch(batch, cx),
+            Poll::Ready(Some(Err(error))) => Poll::Ready(Some(Err(error))),
             Poll::Ready(None) => Poll::Ready(None),
             Poll::Pending => Poll::Pending,
         }
@@ -783,14 +813,6 @@ impl Stream for MemoryTrackedStream {
     }
 }
 
-impl Drop for MemoryTrackedStream {
-    fn drop(&mut self) {
-        if self.total_tracked > 0 {
-            self.permit.release(self.total_tracked);
-        }
-    }
-}
-
 impl RecordBatchStream for MemoryTrackedStream {
     fn schema(&self) -> SchemaRef {
         self.inner.schema()
@@ -808,13 +830,35 @@ impl RecordBatchStream for MemoryTrackedStream {
 #[cfg(test)]
 mod tests {
     use std::sync::Arc;
+    use std::sync::atomic::{AtomicUsize, Ordering};
+    use std::time::Duration;
 
+    use common_memory_manager::{OnExhaustedPolicy, PermitGranularity};
     use datatypes::prelude::{ConcreteDataType, VectorRef};
     use datatypes::schema::{ColumnSchema, Schema};
     use datatypes::vectors::{BooleanVector, Int32Vector, StringVector};
+    use futures::StreamExt;
+    use tokio::time::{sleep, timeout};
 
     use super::*;
 
+    fn large_string_batch(bytes: usize) -> RecordBatch {
+        let schema = Arc::new(Schema::new(vec![ColumnSchema::new(
+            "payload",
+            ConcreteDataType::string_datatype(),
+            false,
+        )]));
+        let payload = "x".repeat(bytes);
+        let vector: VectorRef = Arc::new(StringVector::from(vec![payload]));
+        RecordBatch::new(schema, vec![vector]).unwrap()
+    }
+
+    fn aligned_tracked_bytes(bytes: usize) -> usize {
+        PermitGranularity::Kilobyte
+            .permits_to_bytes(PermitGranularity::Kilobyte.bytes_to_permits(bytes as u64))
+            as usize
+    }
+
     #[test]
     fn test_recordbatches_try_from_columns() {
         let schema = Arc::new(Schema::new(vec![ColumnSchema::new(
@@ -896,156 +940,229 @@ mod tests {
         assert_eq!(collected[1], batch2);
     }
 
+    const MB: usize = 1024 * 1024;
+
     #[test]
     fn test_query_memory_tracker_basic() {
-        let tracker = Arc::new(QueryMemoryTracker::new(1000, 0));
+        let tracker =
+            Arc::new(QueryMemoryTracker::builder(10 * MB, OnExhaustedPolicy::Fail).build());
 
-        // Register first stream - should get privileged status
-        let permit1 = tracker.register_permit();
-        assert!(permit1.is_privileged());
+        let mut stream1 = tracker.new_stream_tracker();
+        assert!(stream1.try_track(5 * MB).is_ok());
+        assert_eq!(tracker.current(), 5 * MB);
 
-        // Privileged stream can use up to limit
-        assert!(permit1.track(500, 0).is_ok());
-        assert_eq!(tracker.current(), 500);
+        let mut stream2 = tracker.new_stream_tracker();
+        assert!(stream2.try_track(4 * MB).is_ok());
+        assert_eq!(tracker.current(), 9 * MB);
 
-        // Register second stream - also privileged
-        let permit2 = tracker.register_permit();
-        assert!(permit2.is_privileged());
-        // Can add more but cannot exceed hard limit (1000)
-        assert!(permit2.track(400, 0).is_ok());
-        assert_eq!(tracker.current(), 900);
-
-        permit1.release(500);
-        permit2.release(400);
+        drop(stream1);
+        drop(stream2);
         assert_eq!(tracker.current(), 0);
     }
 
     #[test]
-    fn test_query_memory_tracker_privileged_limit() {
-        // Privileged slots = 2 for easy testing
-        // Limit: 1000, standard-tier fraction: 0.7 (default)
-        // Privileged can push global to 1000, standard-tier can push global to 700
-        let tracker = Arc::new(QueryMemoryTracker::with_privileged_slots(1000, 2));
+    fn test_query_memory_tracker_shared_global_limit() {
+        let tracker =
+            Arc::new(QueryMemoryTracker::builder(10 * MB, OnExhaustedPolicy::Fail).build());
+        let mut stream1 = tracker.new_stream_tracker();
+        let mut stream2 = tracker.new_stream_tracker();
 
-        // First 2 streams are privileged
-        let permit1 = tracker.register_permit();
-        let permit2 = tracker.register_permit();
-        assert!(permit1.is_privileged());
-        assert!(permit2.is_privileged());
+        assert!(stream1.try_track(3 * MB).is_ok());
+        assert_eq!(tracker.current(), 3 * MB);
+        assert!(stream2.try_track(6 * MB).is_ok());
+        assert_eq!(tracker.current(), 9 * MB);
 
-        // Third stream is standard-tier (not privileged)
-        let permit3 = tracker.register_permit();
-        assert!(!permit3.is_privileged());
-
-        // Privileged stream uses some memory
-        assert!(permit1.track(300, 0).is_ok());
-        assert_eq!(tracker.current(), 300);
-
-        // Standard-tier can add up to 400 (total becomes 700, its effective limit)
-        assert!(permit3.track(400, 0).is_ok());
-        assert_eq!(tracker.current(), 700);
-
-        // Standard-tier stream cannot push global beyond 700
-        let err = permit3.track(100, 400).unwrap_err();
+        let err = stream2.try_track(2 * MB).unwrap_err();
         let err_msg = err.to_string();
-        assert!(err_msg.contains("400B used by this stream"));
-        assert!(err_msg.contains("effective limit: 700B (70%)"));
-        assert!(err_msg.contains("700B used globally (70%)"));
-        assert_eq!(tracker.current(), 700);
+        assert!(err_msg.contains("6.0MiB used by this stream"));
+        assert!(err_msg.contains("9.0MiB used globally (90%)"));
+        assert!(err_msg.contains("hard limit: 10.0MiB"));
+        assert_eq!(tracker.current(), 9 * MB);
 
-        permit1.release(300);
-        permit3.release(400);
+        drop(stream1);
+        assert_eq!(tracker.current(), 6 * MB);
+        drop(stream2);
         assert_eq!(tracker.current(), 0);
     }
 
     #[test]
-    fn test_query_memory_tracker_promotion() {
-        // Privileged slots = 1 for easy testing
-        let tracker = Arc::new(QueryMemoryTracker::with_privileged_slots(1000, 1));
+    fn test_query_memory_tracker_hard_limit() {
+        let tracker =
+            Arc::new(QueryMemoryTracker::builder(10 * MB, OnExhaustedPolicy::Fail).build());
+        let mut stream = tracker.new_stream_tracker();
 
-        // First stream is privileged
-        let permit1 = tracker.register_permit();
-        assert!(permit1.is_privileged());
+        assert!(stream.try_track(9 * MB).is_ok());
+        assert_eq!(tracker.current(), 9 * MB);
 
-        // Second stream is standard-tier (can only use 500)
-        let permit2 = tracker.register_permit();
-        assert!(!permit2.is_privileged());
+        assert!(stream.try_track(2 * MB).is_err());
+        assert_eq!(tracker.current(), 9 * MB);
 
-        // Standard-tier can only track 500
-        assert!(permit2.track(400, 0).is_ok());
-        assert_eq!(tracker.current(), 400);
+        assert!(stream.try_track(MB).is_ok());
+        assert_eq!(tracker.current(), 10 * MB);
 
-        // Drop first permit to release privileged slot
-        drop(permit1);
+        assert!(stream.try_track(MB).is_err());
+        assert_eq!(tracker.current(), 10 * MB);
 
-        // Second stream can now be promoted and use more memory
-        assert!(permit2.track(500, 400).is_ok());
-        assert!(permit2.is_privileged());
-        assert_eq!(tracker.current(), 900);
-
-        permit2.release(900);
+        drop(stream);
         assert_eq!(tracker.current(), 0);
     }
 
     #[test]
-    fn test_query_memory_tracker_privileged_hard_limit() {
-        // Test that the configured limit is absolute hard limit for all streams
-        // Privileged: can use full limit (1000)
-        // Standard-tier: can use 0.7x limit (700 with defaults)
-        let tracker = Arc::new(QueryMemoryTracker::new(1000, 0));
+    fn test_query_memory_tracker_unlimited() {
+        let tracker = Arc::new(QueryMemoryTracker::builder(0, OnExhaustedPolicy::Fail).build());
+        let mut stream = tracker.new_stream_tracker();
 
-        let permit1 = tracker.register_permit();
-        assert!(permit1.is_privileged());
-
-        // Privileged can use up to full limit (1000)
-        assert!(permit1.track(900, 0).is_ok());
-        assert_eq!(tracker.current(), 900);
-
-        // Privileged cannot exceed hard limit (1000)
-        assert!(permit1.track(200, 900).is_err());
-        assert_eq!(tracker.current(), 900);
-
-        // Can add within hard limit
-        assert!(permit1.track(100, 900).is_ok());
-        assert_eq!(tracker.current(), 1000);
-
-        // Cannot exceed even by 1 byte
-        assert!(permit1.track(1, 1000).is_err());
-        assert_eq!(tracker.current(), 1000);
-
-        permit1.release(1000);
+        assert!(stream.try_track(10 * MB).is_ok());
+        assert_eq!(tracker.current(), 10 * MB);
+        drop(stream);
         assert_eq!(tracker.current(), 0);
     }
 
     #[test]
-    fn test_query_memory_tracker_standard_tier_fraction() {
-        // Test standard-tier streams use fraction of limit
-        // Limit: 1000, default fraction: 0.7, so standard-tier can use 700
-        let tracker = Arc::new(QueryMemoryTracker::with_privileged_slots(1000, 1));
+    fn test_query_memory_tracker_rounds_to_kilobytes() {
+        let tracker =
+            Arc::new(QueryMemoryTracker::builder(10 * MB, OnExhaustedPolicy::Fail).build());
+        let mut stream = tracker.new_stream_tracker();
 
-        let permit1 = tracker.register_permit();
-        assert!(permit1.is_privileged());
+        assert!(stream.try_track(1_537).is_ok());
+        assert_eq!(tracker.current(), 2 * 1024);
 
-        let permit2 = tracker.register_permit();
-        assert!(!permit2.is_privileged());
-
-        // Standard-tier can use up to 700 (1000 * 0.7 default)
-        assert!(permit2.track(600, 0).is_ok());
-        assert_eq!(tracker.current(), 600);
-
-        // Cannot exceed standard-tier limit (700)
-        assert!(permit2.track(200, 600).is_err());
-        assert_eq!(tracker.current(), 600);
-
-        // Can add within standard-tier limit
-        assert!(permit2.track(100, 600).is_ok());
-        assert_eq!(tracker.current(), 700);
-
-        // Cannot exceed standard-tier limit
-        assert!(permit2.track(1, 700).is_err());
-        assert_eq!(tracker.current(), 700);
-
-        permit2.release(700);
+        drop(stream);
         assert_eq!(tracker.current(), 0);
     }
+
+    #[tokio::test]
+    async fn test_memory_tracked_stream_waits_for_capacity() {
+        let exhausted = Arc::new(AtomicUsize::new(0));
+        let rejected = Arc::new(AtomicUsize::new(0));
+        let exhausted_counter = exhausted.clone();
+        let rejected_counter = rejected.clone();
+        let tracker = QueryMemoryTracker::builder(
+            MB,
+            OnExhaustedPolicy::Wait {
+                timeout: Duration::from_millis(200),
+            },
+        )
+        .on_exhausted(move || {
+            exhausted_counter.fetch_add(1, Ordering::Relaxed);
+        })
+        .on_reject(move || {
+            rejected_counter.fetch_add(1, Ordering::Relaxed);
+        })
+        .build();
+        let batch = large_string_batch(700 * 1024);
+        let expected_bytes = aligned_tracked_bytes(batch.buffer_memory_size());
+
+        let mut stream1 = MemoryTrackedStream::new(
+            RecordBatches::try_new(batch.schema.clone(), vec![batch.clone()])
+                .unwrap()
+                .as_stream(),
+            tracker.clone(),
+        );
+        let first = stream1.next().await.unwrap().unwrap();
+        assert_eq!(first.num_rows(), 1);
+        assert_eq!(tracker.current(), expected_bytes);
+
+        let stream2 = MemoryTrackedStream::new(
+            RecordBatches::try_new(batch.schema.clone(), vec![batch])
+                .unwrap()
+                .as_stream(),
+            tracker.clone(),
+        );
+        let waiter = tokio::spawn(async move {
+            let mut stream2 = stream2;
+            stream2.next().await.unwrap()
+        });
+
+        sleep(Duration::from_millis(50)).await;
+        assert!(!waiter.is_finished());
+
+        drop(stream1);
+        let second = waiter.await.unwrap().unwrap();
+        assert_eq!(second.num_rows(), 1);
+        assert_eq!(exhausted.load(Ordering::Relaxed), 1);
+        assert_eq!(rejected.load(Ordering::Relaxed), 0);
+    }
+
+    #[tokio::test]
+    async fn test_memory_tracked_stream_wait_times_out() {
+        let exhausted = Arc::new(AtomicUsize::new(0));
+        let rejected = Arc::new(AtomicUsize::new(0));
+        let exhausted_counter = exhausted.clone();
+        let rejected_counter = rejected.clone();
+        let tracker = QueryMemoryTracker::builder(
+            MB,
+            OnExhaustedPolicy::Wait {
+                timeout: Duration::from_millis(50),
+            },
+        )
+        .on_exhausted(move || {
+            exhausted_counter.fetch_add(1, Ordering::Relaxed);
+        })
+        .on_reject(move || {
+            rejected_counter.fetch_add(1, Ordering::Relaxed);
+        })
+        .build();
+        let batch = large_string_batch(700 * 1024);
+
+        let mut stream1 = MemoryTrackedStream::new(
+            RecordBatches::try_new(batch.schema.clone(), vec![batch.clone()])
+                .unwrap()
+                .as_stream(),
+            tracker.clone(),
+        );
+        let first = stream1.next().await.unwrap().unwrap();
+        assert_eq!(first.num_rows(), 1);
+
+        let mut stream2 = MemoryTrackedStream::new(
+            RecordBatches::try_new(batch.schema.clone(), vec![batch])
+                .unwrap()
+                .as_stream(),
+            tracker,
+        );
+        let result = timeout(Duration::from_secs(1), stream2.next())
+            .await
+            .unwrap();
+        let error = result.unwrap().unwrap_err();
+        assert!(error.to_string().contains("timed out waiting"));
+        assert_eq!(exhausted.load(Ordering::Relaxed), 1);
+        assert_eq!(rejected.load(Ordering::Relaxed), 1);
+    }
+
+    #[tokio::test]
+    async fn test_memory_tracked_stream_fail_policy_rejects_immediately() {
+        let exhausted = Arc::new(AtomicUsize::new(0));
+        let rejected = Arc::new(AtomicUsize::new(0));
+        let exhausted_counter = exhausted.clone();
+        let rejected_counter = rejected.clone();
+        let tracker = QueryMemoryTracker::builder(MB, OnExhaustedPolicy::Fail)
+            .on_exhausted(move || {
+                exhausted_counter.fetch_add(1, Ordering::Relaxed);
+            })
+            .on_reject(move || {
+                rejected_counter.fetch_add(1, Ordering::Relaxed);
+            })
+            .build();
+        let batch = large_string_batch(700 * 1024);
+
+        let mut stream1 = MemoryTrackedStream::new(
+            RecordBatches::try_new(batch.schema.clone(), vec![batch.clone()])
+                .unwrap()
+                .as_stream(),
+            tracker.clone(),
+        );
+        let first = stream1.next().await.unwrap().unwrap();
+        assert_eq!(first.num_rows(), 1);
+
+        let mut stream2 = MemoryTrackedStream::new(
+            RecordBatches::try_new(batch.schema.clone(), vec![batch])
+                .unwrap()
+                .as_stream(),
+            tracker,
+        );
+        let result = stream2.next().await.unwrap();
+        assert!(result.is_err());
+        assert_eq!(exhausted.load(Ordering::Relaxed), 1);
+        assert_eq!(rejected.load(Ordering::Relaxed), 1);
+    }
 }
diff --git a/src/common/recordbatch/src/recordbatch.rs b/src/common/recordbatch/src/recordbatch.rs
index f6e0aeed93..4289714afd 100644
--- a/src/common/recordbatch/src/recordbatch.rs
+++ b/src/common/recordbatch/src/recordbatch.rs
@@ -20,10 +20,11 @@ use datafusion::arrow::util::pretty::pretty_format_batches;
 use datafusion_common::arrow::array::ArrayRef;
 use datafusion_common::arrow::compute;
 use datafusion_common::arrow::datatypes::{DataType as ArrowDataType, SchemaRef as ArrowSchemaRef};
-use datatypes::arrow::array::{Array, AsArray, RecordBatchOptions, StructArray, new_null_array};
+use datatypes::arrow::array::{Array, AsArray, RecordBatchOptions};
 use datatypes::extension::json::is_json_extension_type;
 use datatypes::prelude::DataType;
 use datatypes::schema::SchemaRef;
+use datatypes::vectors::json::array::JsonArray;
 use datatypes::vectors::{Helper, VectorRef};
 use serde::ser::{Error, SerializeStruct};
 use serde::{Serialize, Serializer};
@@ -31,8 +32,8 @@ use snafu::{OptionExt, ResultExt, ensure};
 
 use crate::DfRecordBatch;
 use crate::error::{
-    self, AlignJsonArraySnafu, ArrowComputeSnafu, ColumnNotExistsSnafu, DataTypesSnafu,
-    NewDfRecordBatchSnafu, ProjectArrowRecordBatchSnafu, Result,
+    self, ArrowComputeSnafu, ColumnNotExistsSnafu, DataTypesSnafu, ProjectArrowRecordBatchSnafu,
+    Result,
 };
 
 /// A two-dimensional batch of column-oriented data with a defined schema.
@@ -354,80 +355,6 @@ pub fn merge_record_batches(schema: SchemaRef, batches: &[RecordBatch]) -> Resul
     Ok(RecordBatch::from_df_record_batch(schema, record_batch))
 }
 
-/// Align a json array `json_array` to the json type `schema_type`. The `schema_type` is often the
-/// "largest" json type after some insertions in the table schema, while the json array previously
-/// written in the SST could be lagged behind it. So it's important to "amend" the json array's
-/// missing fields with null arrays, to align the array's data type with the provided one.
-///
-/// # Panics
-///
-/// - The json array is not an Arrow [StructArray], or the provided data type `schema_type` is not
-///   of Struct type. Both of which shouldn't happen unless we switch our implementation of how
-///   json array is physically stored.
-pub fn align_json_array(json_array: &ArrayRef, schema_type: &ArrowDataType) -> Result<ArrayRef> {
-    let json_type = json_array.data_type();
-    if json_type == schema_type {
-        return Ok(json_array.clone());
-    }
-
-    let json_array = json_array.as_struct();
-    let array_fields = json_array.fields();
-    let array_columns = json_array.columns();
-    let ArrowDataType::Struct(schema_fields) = schema_type else {
-        unreachable!()
-    };
-    let mut aligned = Vec::with_capacity(schema_fields.len());
-
-    // Compare the fields in the json array and the to-be-aligned schema, amending with null arrays
-    // on the way. It's very important to note that fields in the json array and in the json type
-    // are both SORTED.
-
-    let mut i = 0; // point to the schema fields
-    let mut j = 0; // point to the array fields
-    while i < schema_fields.len() && j < array_fields.len() {
-        let schema_field = &schema_fields[i];
-        let array_field = &array_fields[j];
-        if schema_field.name() == array_field.name() {
-            if matches!(schema_field.data_type(), ArrowDataType::Struct(_)) {
-                // A `StructArray`s in a json array must be another json array. (Like a nested json
-                // object in a json value.)
-                aligned.push(align_json_array(
-                    &array_columns[j],
-                    schema_field.data_type(),
-                )?);
-            } else {
-                aligned.push(array_columns[j].clone());
-            }
-            j += 1;
-        } else {
-            aligned.push(new_null_array(schema_field.data_type(), json_array.len()));
-        }
-        i += 1;
-    }
-    if i < schema_fields.len() {
-        for field in &schema_fields[i..] {
-            aligned.push(new_null_array(field.data_type(), json_array.len()));
-        }
-    }
-    ensure!(
-        j == array_fields.len(),
-        AlignJsonArraySnafu {
-            reason: format!(
-                "this json array has more fields {:?}",
-                array_fields[j..]
-                    .iter()
-                    .map(|x| x.name())
-                    .collect::<Vec<_>>(),
-            )
-        }
-    );
-
-    let json_array =
-        StructArray::try_new(schema_fields.clone(), aligned, json_array.nulls().cloned())
-            .context(NewDfRecordBatchSnafu)?;
-    Ok(Arc::new(json_array))
-}
-
 fn maybe_align_json_array_with_schema(
     schema: &ArrowSchemaRef,
     arrays: Vec<ArrayRef>,
@@ -437,13 +364,15 @@ fn maybe_align_json_array_with_schema(
     }
 
     let mut aligned = Vec::with_capacity(arrays.len());
-    for (field, array) in schema.fields().iter().zip(arrays.into_iter()) {
+    for (field, array) in schema.fields().iter().zip(arrays) {
         if !is_json_extension_type(field) {
             aligned.push(array);
             continue;
         }
 
-        let json_array = align_json_array(&array, field.data_type())?;
+        let json_array = JsonArray::from(&array)
+            .try_align(field.data_type())
+            .context(DataTypesSnafu)?;
         aligned.push(json_array);
     }
     Ok(aligned)
@@ -453,12 +382,8 @@ fn maybe_align_json_array_with_schema(
 mod tests {
     use std::sync::Arc;
 
-    use datatypes::arrow::array::{
-        AsArray, BooleanArray, Float64Array, Int64Array, ListArray, UInt32Array,
-    };
-    use datatypes::arrow::datatypes::{
-        DataType, Field, Fields, Int64Type, Schema as ArrowSchema, UInt32Type,
-    };
+    use datatypes::arrow::array::{AsArray, UInt32Array};
+    use datatypes::arrow::datatypes::{DataType, Field, Schema as ArrowSchema, UInt32Type};
     use datatypes::arrow_array::StringArray;
     use datatypes::data_type::ConcreteDataType;
     use datatypes::schema::{ColumnSchema, Schema};
@@ -466,165 +391,6 @@ mod tests {
 
     use super::*;
 
-    #[test]
-    fn test_align_json_array() -> Result<()> {
-        struct TestCase {
-            json_array: ArrayRef,
-            schema_type: DataType,
-            expected: std::result::Result<ArrayRef, String>,
-        }
-
-        impl TestCase {
-            fn new(
-                json_array: StructArray,
-                schema_type: Fields,
-                expected: std::result::Result<Vec<ArrayRef>, String>,
-            ) -> Self {
-                Self {
-                    json_array: Arc::new(json_array),
-                    schema_type: DataType::Struct(schema_type.clone()),
-                    expected: expected
-                        .map(|x| Arc::new(StructArray::new(schema_type, x, None)) as ArrayRef),
-                }
-            }
-
-            fn test(self) -> Result<()> {
-                let result = align_json_array(&self.json_array, &self.schema_type);
-                match (result, self.expected) {
-                    (Ok(json_array), Ok(expected)) => assert_eq!(&json_array, &expected),
-                    (Ok(json_array), Err(e)) => {
-                        panic!("expecting error {e} but actually get: {json_array:?}")
-                    }
-                    (Err(e), Err(expected)) => assert_eq!(e.to_string(), expected),
-                    (Err(e), Ok(_)) => return Err(e),
-                }
-                Ok(())
-            }
-        }
-
-        // Test empty json array can be aligned with a complex json type.
-        TestCase::new(
-            StructArray::new_empty_fields(2, None),
-            Fields::from(vec![
-                Field::new("int", DataType::Int64, true),
-                Field::new_struct(
-                    "nested",
-                    vec![Field::new("bool", DataType::Boolean, true)],
-                    true,
-                ),
-                Field::new("string", DataType::Utf8, true),
-            ]),
-            Ok(vec![
-                Arc::new(Int64Array::new_null(2)) as ArrayRef,
-                Arc::new(StructArray::new_null(
-                    Fields::from(vec![Arc::new(Field::new("bool", DataType::Boolean, true))]),
-                    2,
-                )),
-                Arc::new(StringArray::new_null(2)),
-            ]),
-        )
-        .test()?;
-
-        // Test simple json array alignment.
-        TestCase::new(
-            StructArray::from(vec![(
-                Arc::new(Field::new("float", DataType::Float64, true)),
-                Arc::new(Float64Array::from(vec![1.0, 2.0, 3.0])) as ArrayRef,
-            )]),
-            Fields::from(vec![
-                Field::new("float", DataType::Float64, true),
-                Field::new("string", DataType::Utf8, true),
-            ]),
-            Ok(vec![
-                Arc::new(Float64Array::from(vec![1.0, 2.0, 3.0])) as ArrayRef,
-                Arc::new(StringArray::new_null(3)),
-            ]),
-        )
-        .test()?;
-
-        // Test complex json array alignment.
-        TestCase::new(
-            StructArray::from(vec![
-                (
-                    Arc::new(Field::new_list(
-                        "list",
-                        Field::new_list_field(DataType::Int64, true),
-                        true,
-                    )),
-                    Arc::new(ListArray::from_iter_primitive::<Int64Type, _, _>(vec![
-                        Some(vec![Some(1)]),
-                        None,
-                        Some(vec![Some(2), Some(3)]),
-                    ])) as ArrayRef,
-                ),
-                (
-                    Arc::new(Field::new_struct(
-                        "nested",
-                        vec![Field::new("int", DataType::Int64, true)],
-                        true,
-                    )),
-                    Arc::new(StructArray::from(vec![(
-                        Arc::new(Field::new("int", DataType::Int64, true)),
-                        Arc::new(Int64Array::from(vec![-1, -2, -3])) as ArrayRef,
-                    )])),
-                ),
-                (
-                    Arc::new(Field::new("string", DataType::Utf8, true)),
-                    Arc::new(StringArray::from(vec!["a", "b", "c"])),
-                ),
-            ]),
-            Fields::from(vec![
-                Field::new("bool", DataType::Boolean, true),
-                Field::new_list("list", Field::new_list_field(DataType::Int64, true), true),
-                Field::new_struct(
-                    "nested",
-                    vec![
-                        Field::new("float", DataType::Float64, true),
-                        Field::new("int", DataType::Int64, true),
-                    ],
-                    true,
-                ),
-                Field::new("string", DataType::Utf8, true),
-            ]),
-            Ok(vec![
-                Arc::new(BooleanArray::new_null(3)) as ArrayRef,
-                Arc::new(ListArray::from_iter_primitive::<Int64Type, _, _>(vec![
-                    Some(vec![Some(1)]),
-                    None,
-                    Some(vec![Some(2), Some(3)]),
-                ])),
-                Arc::new(StructArray::from(vec![
-                    (
-                        Arc::new(Field::new("float", DataType::Float64, true)),
-                        Arc::new(Float64Array::new_null(3)) as ArrayRef,
-                    ),
-                    (
-                        Arc::new(Field::new("int", DataType::Int64, true)),
-                        Arc::new(Int64Array::from(vec![-1, -2, -3])),
-                    ),
-                ])),
-                Arc::new(StringArray::from(vec!["a", "b", "c"])),
-            ]),
-        )
-        .test()?;
-
-        // Test align failed.
-        TestCase::new(
-            StructArray::try_from(vec![
-                ("i", Arc::new(Int64Array::from(vec![1])) as ArrayRef),
-                ("j", Arc::new(Int64Array::from(vec![2])) as ArrayRef),
-            ])
-            .unwrap(),
-            Fields::from(vec![Field::new("i", DataType::Int64, true)]),
-            Err(
-                r#"Failed to align JSON array, reason: this json array has more fields ["j"]"#
-                    .to_string(),
-            ),
-        )
-        .test()?;
-        Ok(())
-    }
-
     #[test]
     fn test_record_batch() {
         let arrow_schema = Arc::new(ArrowSchema::new(vec![
diff --git a/src/common/sql/src/convert.rs b/src/common/sql/src/convert.rs
index bd9a1d0769..be9701f832 100644
--- a/src/common/sql/src/convert.rs
+++ b/src/common/sql/src/convert.rs
@@ -306,7 +306,7 @@ pub(crate) fn parse_string_to_value(
                 let v = parse_string_to_jsonb(&s).context(DatatypeSnafu)?;
                 Ok(Value::Binary(v.into()))
             }
-            JsonFormat::Native(_) => {
+            JsonFormat::Json2(_) => {
                 let extension_type: Option<JsonExtensionType> =
                     column_schema.extension_type().context(DatatypeSnafu)?;
                 let json_structure_settings = extension_type
diff --git a/src/common/sql/src/default_constraint.rs b/src/common/sql/src/default_constraint.rs
index 0084320835..bad9f374fd 100644
--- a/src/common/sql/src/default_constraint.rs
+++ b/src/common/sql/src/default_constraint.rs
@@ -122,7 +122,7 @@ pub fn parse_column_default_constraint(
 
 #[cfg(test)]
 mod test {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use datatypes::prelude::{ConcreteDataType, Value};
     use datatypes::types::BooleanType;
diff --git a/src/common/sql/src/lib.rs b/src/common/sql/src/lib.rs
index abca883124..8835cae6b4 100644
--- a/src/common/sql/src/lib.rs
+++ b/src/common/sql/src/lib.rs
@@ -12,8 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
-
 pub mod convert;
 pub mod default_constraint;
 pub mod error;
diff --git a/src/common/telemetry/src/lib.rs b/src/common/telemetry/src/lib.rs
index cd60d61645..26bf5d53b3 100644
--- a/src/common/telemetry/src/lib.rs
+++ b/src/common/telemetry/src/lib.rs
@@ -21,10 +21,12 @@ mod panic_hook;
 pub mod tracing_context;
 mod tracing_sampler;
 
+pub use common_error;
 pub use logging::{
     LOG_RELOAD_HANDLE, TRACE_RELOAD_HANDLE, get_or_init_tracer, init_default_ut_logging,
     init_global_logging,
 };
 pub use metric::dump_metrics;
 pub use panic_hook::set_panic_hook;
-pub use {common_error, tracing, tracing_subscriber};
+pub use tracing;
+pub use tracing_subscriber;
diff --git a/src/common/telemetry/src/panic_hook.rs b/src/common/telemetry/src/panic_hook.rs
index 47e7b83bd7..4a7a9a1498 100644
--- a/src/common/telemetry/src/panic_hook.rs
+++ b/src/common/telemetry/src/panic_hook.rs
@@ -36,6 +36,7 @@ pub fn set_panic_hook() {
     panic::set_hook(Box::new(move |panic| {
         let backtrace = Backtrace::new();
         let backtrace = format!("{backtrace:?}");
+        let build_info = common_version::build_info();
         if let Some(location) = panic.location() {
             tracing::error!(
                 message = %panic,
@@ -43,9 +44,18 @@ pub fn set_panic_hook() {
                 panic.file = location.file(),
                 panic.line = location.line(),
                 panic.column = location.column(),
+                branch = %build_info.branch,
+                version = %build_info.version,
+                commit = %build_info.commit,
             );
         } else {
-            tracing::error!(message = %panic, backtrace = %backtrace);
+            tracing::error!(
+                message = %panic,
+                backtrace = %backtrace,
+                branch = %build_info.branch,
+                version = %build_info.version,
+                commit = %build_info.commit,
+            );
         }
         PANIC_COUNTER.inc();
         default_hook(panic);
diff --git a/src/common/version/build.rs b/src/common/version/build.rs
index f147918c21..81c805484b 100644
--- a/src/common/version/build.rs
+++ b/src/common/version/build.rs
@@ -29,6 +29,7 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
     let refresh = profile == "release";
 
     println!("cargo:rerun-if-env-changed=RUSTC");
+    println!("cargo:rerun-if-env-changed=GREPTIME_PRODUCT_NAME");
 
     // The "CARGO_WORKSPACE_DIR" is set manually (not by Rust itself) in Cargo config file, to
     // solve the problem where the "CARGO_MANIFEST_DIR" is not what we want when this repo is
@@ -44,6 +45,9 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
     let product_version = load_product_version(&workspace_root);
     println!("cargo:rustc-env=GREPTIME_PRODUCT_VERSION={product_version}");
 
+    let product_name = load_product_name();
+    println!("cargo:rustc-env=GREPTIME_PRODUCT_NAME={product_name}");
+
     let repository = open_repository(&workspace_root);
 
     if refresh {
@@ -100,6 +104,10 @@ fn load_product_version(workspace_root: &Path) -> String {
         .unwrap_or_else(|| env::var("CARGO_PKG_VERSION").unwrap())
 }
 
+fn load_product_name() -> String {
+    env::var("GREPTIME_PRODUCT_NAME").unwrap_or_else(|_| "GreptimeDB".to_string())
+}
+
 fn emit_workspace_watch_list(
     workspace_root: &Path,
     repository: Option<&Repository>,
diff --git a/src/common/version/src/lib.rs b/src/common/version/src/lib.rs
index 82d64f532f..e651225d68 100644
--- a/src/common/version/src/lib.rs
+++ b/src/common/version/src/lib.rs
@@ -109,6 +109,10 @@ pub const fn version() -> &'static str {
     BUILD_INFO.version
 }
 
+pub const fn product_name() -> &'static str {
+    env!("GREPTIME_PRODUCT_NAME")
+}
+
 pub const fn verbose_version() -> &'static str {
     const_format::formatcp!(
         "\nbranch: {}\ncommit: {}\nclean: {}\nversion: {}",
diff --git a/src/common/wal/Cargo.toml b/src/common/wal/Cargo.toml
index 8332f2341f..b9ca39efca 100644
--- a/src/common/wal/Cargo.toml
+++ b/src/common/wal/Cargo.toml
@@ -19,7 +19,7 @@ futures-util.workspace = true
 humantime-serde.workspace = true
 num_cpus.workspace = true
 rskafka.workspace = true
-rustls = { workspace = true, default-features = false, features = ["ring", "logging", "std", "tls12"] }
+rustls = { workspace = true, default-features = false, features = ["aws_lc_rs", "logging", "std", "tls12"] }
 rustls-native-certs = "0.7"
 rustls-pemfile = "2.1"
 serde.workspace = true
diff --git a/src/common/wal/src/lib.rs b/src/common/wal/src/lib.rs
index 659a045f57..a0b1dc99f9 100644
--- a/src/common/wal/src/lib.rs
+++ b/src/common/wal/src/lib.rs
@@ -12,8 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
-
 use std::net::SocketAddr;
 
 use error::{EndpointIPV4NotFoundSnafu, ResolveEndpointSnafu, Result};
@@ -59,7 +57,7 @@ async fn resolve_to_ipv4_one<T: AsRef<str>>(endpoint: T) -> Result<String> {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_telemetry::warn;
     use rskafka::client::{Credentials, SaslConfig};
diff --git a/src/datanode/src/alive_keeper.rs b/src/datanode/src/alive_keeper.rs
index 57f4e00aa2..dbf99fdb28 100644
--- a/src/datanode/src/alive_keeper.rs
+++ b/src/datanode/src/alive_keeper.rs
@@ -503,6 +503,7 @@ mod test {
     use mito2::config::MitoConfig;
     use mito2::test_util::{CreateRequestBuilder, TestEnv};
     use store_api::region_engine::RegionEngine;
+    use store_api::region_request::{EnterStagingRequest, StagingPartitionDirective};
 
     use super::*;
     use crate::tests::mock_region_server;
@@ -621,4 +622,141 @@ mod test {
                 > Instant::now() + Duration::from_millis(heartbeat_interval_millis * 4)
         );
     }
+
+    #[tokio::test(flavor = "multi_thread")]
+    async fn renew_staging_leader_keeps_region_in_staging() {
+        let mut region_server = mock_region_server();
+        let mut engine_env = TestEnv::with_prefix("region-alive-keeper-staging").await;
+        let engine = engine_env.create_engine(MitoConfig::default()).await;
+        let engine = Arc::new(engine);
+        region_server.register_engine(engine.clone());
+
+        let alive_keeper = Arc::new(RegionAliveKeeper::new(
+            region_server.clone(),
+            None,
+            Duration::from_millis(100),
+        ));
+
+        let region_id = RegionId::new(1024, 2);
+        region_server
+            .handle_request(
+                region_id,
+                RegionRequest::Create(CreateRequestBuilder::new().build()),
+            )
+            .await
+            .unwrap();
+        region_server
+            .handle_request(
+                region_id,
+                RegionRequest::EnterStaging(EnterStagingRequest {
+                    partition_directive: StagingPartitionDirective::RejectAllWrites,
+                }),
+            )
+            .await
+            .unwrap();
+
+        alive_keeper.register_region(region_id).await;
+        alive_keeper
+            .renew_region_leases(
+                &[GrantedRegion {
+                    region_id: region_id.as_u64(),
+                    role: api::v1::meta::RegionRole::StagingLeader.into(),
+                    extensions: HashMap::new(),
+                }],
+                Instant::now() + Duration::from_millis(3000),
+            )
+            .await;
+
+        assert_eq!(engine.role(region_id).unwrap(), RegionRole::StagingLeader);
+    }
+
+    #[tokio::test(flavor = "multi_thread")]
+    async fn renew_staging_leader_exit_into_leader() {
+        common_telemetry::init_default_ut_logging();
+        let mut region_server = mock_region_server();
+        let mut engine_env = TestEnv::with_prefix("region-alive-keeper-staging-exit").await;
+        let engine = engine_env.create_engine(MitoConfig::default()).await;
+        let engine = Arc::new(engine);
+        region_server.register_engine(engine.clone());
+
+        let alive_keeper = Arc::new(RegionAliveKeeper::new(
+            region_server.clone(),
+            None,
+            Duration::from_millis(100),
+        ));
+
+        let region_id = RegionId::new(1024, 2);
+        region_server
+            .handle_request(
+                region_id,
+                RegionRequest::Create(CreateRequestBuilder::new().build()),
+            )
+            .await
+            .unwrap();
+        region_server
+            .handle_request(
+                region_id,
+                RegionRequest::EnterStaging(EnterStagingRequest {
+                    partition_directive: StagingPartitionDirective::RejectAllWrites,
+                }),
+            )
+            .await
+            .unwrap();
+
+        alive_keeper.register_region(region_id).await;
+        alive_keeper
+            .renew_region_leases(
+                &[GrantedRegion {
+                    region_id: region_id.as_u64(),
+                    role: api::v1::meta::RegionRole::Leader.into(),
+                    extensions: HashMap::new(),
+                }],
+                Instant::now() + Duration::from_millis(3000),
+            )
+            .await;
+
+        tokio::time::sleep(Duration::from_millis(100)).await;
+        assert_eq!(engine.role(region_id).unwrap(), RegionRole::Leader);
+    }
+
+    #[tokio::test(flavor = "multi_thread")]
+    async fn renew_staging_leader_does_not_promote_normal_leader_into_staging() {
+        let mut region_server = mock_region_server();
+        let mut engine_env = TestEnv::with_prefix("region-alive-keeper-non-staging").await;
+        let engine = engine_env.create_engine(MitoConfig::default()).await;
+        let engine = Arc::new(engine);
+        region_server.register_engine(engine.clone());
+
+        let alive_keeper = Arc::new(RegionAliveKeeper::new(
+            region_server.clone(),
+            None,
+            Duration::from_millis(100),
+        ));
+
+        let region_id = RegionId::new(1024, 4);
+        region_server
+            .handle_request(
+                region_id,
+                RegionRequest::Create(CreateRequestBuilder::new().build()),
+            )
+            .await
+            .unwrap();
+        region_server
+            .set_region_role(region_id, RegionRole::Leader)
+            .unwrap();
+
+        alive_keeper.register_region(region_id).await;
+        alive_keeper
+            .renew_region_leases(
+                &[GrantedRegion {
+                    region_id: region_id.as_u64(),
+                    role: api::v1::meta::RegionRole::StagingLeader.into(),
+                    extensions: HashMap::new(),
+                }],
+                Instant::now() + Duration::from_millis(3000),
+            )
+            .await;
+        tokio::time::sleep(Duration::from_millis(100)).await;
+        assert_eq!(engine.role(region_id).unwrap(), RegionRole::Leader);
+    }
 }
diff --git a/src/datanode/src/config.rs b/src/datanode/src/config.rs
index 6f6815a869..2ce306006b 100644
--- a/src/datanode/src/config.rs
+++ b/src/datanode/src/config.rs
@@ -83,10 +83,14 @@ pub struct DatanodeOptions {
     pub query: QueryOptions,
     pub memory: MemoryOptions,
 
+    /// Environment variable keys to read and report in heartbeat messages.
+    /// The values of these env vars at startup will be sent to metasrv.
+    pub heartbeat_env_vars: Vec<String>,
+
     /// Deprecated options, please use the new options instead.
-    #[deprecated(note = "Please use `grpc.addr` instead.")]
+    #[deprecated(note = "Please use `grpc.bind_addr` instead.")]
     pub rpc_addr: Option<String>,
-    #[deprecated(note = "Please use `grpc.hostname` instead.")]
+    #[deprecated(note = "Please use `grpc.server_addr` instead.")]
     pub rpc_hostname: Option<String>,
     #[deprecated(note = "Please use `grpc.runtime_size` instead.")]
     pub rpc_runtime_size: Option<usize>,
@@ -136,6 +140,7 @@ impl Default for DatanodeOptions {
             tracing: TracingOptions::default(),
             query: QueryOptions::default(),
             memory: MemoryOptions::default(),
+            heartbeat_env_vars: vec![],
 
             // Deprecated options
             rpc_addr: None,
@@ -149,7 +154,11 @@ impl Default for DatanodeOptions {
 
 impl Configurable for DatanodeOptions {
     fn env_list_keys() -> Option<&'static [&'static str]> {
-        Some(&["meta_client.metasrv_addrs", "wal.broker_endpoints"])
+        Some(&[
+            "heartbeat_env_vars",
+            "meta_client.metasrv_addrs",
+            "wal.broker_endpoints",
+        ])
     }
 }
 
diff --git a/src/datanode/src/datanode.rs b/src/datanode/src/datanode.rs
index 3c62015179..9a2fe3d982 100644
--- a/src/datanode/src/datanode.rs
+++ b/src/datanode/src/datanode.rs
@@ -163,6 +163,7 @@ pub struct DatanodeBuilder {
     kv_backend: KvBackendRef,
     cache_registry: Option<Arc<LayeredCacheRegistry>>,
     topic_stats_reporter: Option<Box<dyn TopicStatsReporter>>,
+    open_regions_writable_override: Option<bool>,
     #[cfg(feature = "enterprise")]
     extension_range_provider_factory: Option<mito2::extension::BoxedExtensionRangeProviderFactory>,
 }
@@ -176,6 +177,7 @@ impl DatanodeBuilder {
             meta_client: None,
             kv_backend,
             cache_registry: None,
+            open_regions_writable_override: None,
             #[cfg(feature = "enterprise")]
             extension_range_provider_factory: None,
             topic_stats_reporter: None,
@@ -205,6 +207,20 @@ impl DatanodeBuilder {
         self
     }
 
+    /// Overrides whether regions opened during datanode startup should become writable.
+    ///
+    /// When unset, the builder uses its default writable policy for reopened regions
+    /// (writable only when no metasrv client is configured).
+    ///
+    /// Warning: setting this to `true` on a metasrv-controlled datanode (one built
+    /// with `with_meta_client`) will promote regions to Leader before heartbeat and
+    /// lease coordination begin, bypassing the metasrv safety contract and creating a
+    /// potential split-brain window during startup.
+    pub fn with_open_regions_writable_override(&mut self, writable: bool) -> &mut Self {
+        self.open_regions_writable_override = Some(writable);
+        self
+    }
+
     #[cfg(feature = "enterprise")]
     pub fn with_extension_range_provider(
         &mut self,
@@ -274,10 +290,13 @@ impl DatanodeBuilder {
 
         let region_open_requests =
             build_region_open_requests(node_id, self.kv_backend.clone()).await?;
+        let open_with_writable = self
+            .open_regions_writable_override
+            .unwrap_or(!controlled_by_metasrv);
         let open_all_regions = open_all_regions(
             region_server.clone(),
             region_open_requests,
-            !controlled_by_metasrv,
+            open_with_writable,
             self.opts.init_regions_parallelism,
             // Ignore nonexistent regions in recovery mode.
             is_recovery_mode,
@@ -538,7 +557,6 @@ impl DatanodeBuilder {
                     file_ref_manager,
                     partition_expr_fetcher.clone(),
                     plugins,
-                    opts.max_concurrent_queries,
                 );
 
                 #[cfg(feature = "enterprise")]
@@ -552,14 +570,15 @@ impl DatanodeBuilder {
                 if kafka_config.create_index && opts.node_id.is_none() {
                     warn!("The WAL index creation only available in distributed mode.")
                 }
-                let global_index_collector = if kafka_config.create_index && opts.node_id.is_some()
+                let global_index_collector = if kafka_config.create_index
+                    && let Some(node_id) = opts.node_id
                 {
                     let operator = new_object_store_without_cache(
                         &opts.storage.store,
                         &opts.storage.data_home,
                     )
                     .await?;
-                    let path = default_index_file(opts.node_id.unwrap());
+                    let path = default_index_file(node_id);
                     Some(Self::build_global_index_collector(
                         kafka_config.dump_index_interval,
                         operator,
@@ -581,7 +600,6 @@ impl DatanodeBuilder {
                     file_ref_manager,
                     partition_expr_fetcher,
                     plugins,
-                    opts.max_concurrent_queries,
                 );
 
                 #[cfg(feature = "enterprise")]
@@ -603,7 +621,6 @@ impl DatanodeBuilder {
                     file_ref_manager,
                     partition_expr_fetcher.clone(),
                     plugins,
-                    opts.max_concurrent_queries,
                 );
 
                 #[cfg(feature = "enterprise")]
@@ -785,7 +802,7 @@ async fn open_all_regions(
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::{BTreeMap, HashMap};
     use std::sync::Arc;
 
diff --git a/src/datanode/src/heartbeat.rs b/src/datanode/src/heartbeat.rs
index be662dfe94..3d77251953 100644
--- a/src/datanode/src/heartbeat.rs
+++ b/src/datanode/src/heartbeat.rs
@@ -21,7 +21,7 @@ use api::v1::meta::heartbeat_request::NodeWorkloads;
 use api::v1::meta::{DatanodeWorkloads, HeartbeatRequest, NodeInfo, Peer, RegionRole, RegionStat};
 use common_base::Plugins;
 use common_meta::cache_invalidator::CacheInvalidatorRef;
-use common_meta::datanode::REGION_STATISTIC_KEY;
+use common_meta::datanode::{EnvVars, REGION_STATISTIC_KEY};
 use common_meta::distributed_time_constants::BASE_HEARTBEAT_INTERVAL;
 use common_meta::heartbeat::handler::invalidate_table_cache::InvalidateCacheHandler;
 use common_meta::heartbeat::handler::parse_mailbox_message::ParseMailboxMessageHandler;
@@ -66,6 +66,7 @@ pub struct HeartbeatTask {
     resp_handler_executor: HeartbeatResponseHandlerExecutorRef,
     region_alive_keeper: Arc<RegionAliveKeeper>,
     resource_stat: ResourceStatRef,
+    env_vars: EnvVars,
 }
 
 impl Drop for HeartbeatTask {
@@ -114,6 +115,7 @@ impl HeartbeatTask {
             resp_handler_executor,
             region_alive_keeper,
             resource_stat,
+            env_vars: EnvVars::from_config(&opts.heartbeat_env_vars),
         })
     }
 
@@ -148,9 +150,9 @@ impl HeartbeatTask {
                     let mut follower_region_lease_count = 0;
                     for lease in &lease.regions {
                         match lease.role() {
-                            RegionRole::Leader | RegionRole::DowngradingLeader => {
-                                leader_region_lease_count += 1
-                            }
+                            RegionRole::Leader
+                            | RegionRole::StagingLeader
+                            | RegionRole::DowngradingLeader => leader_region_lease_count += 1,
                             RegionRole::Follower => follower_region_lease_count += 1,
                         }
                     }
@@ -258,6 +260,8 @@ impl HeartbeatTask {
             .mito_engine()
             .context(RegionEngineNotFoundSnafu { name: "mito" })?
             .gc_limiter();
+        let mut env_var_extensions = HashMap::new();
+        self.env_vars.into_extensions(&mut env_var_extensions);
 
         common_runtime::spawn_hb(async move {
             let sleep = tokio::time::sleep(Duration::from_millis(0));
@@ -300,7 +304,7 @@ impl HeartbeatTask {
                         if let Some(message) = message {
                             match outgoing_message_to_mailbox_message(message) {
                                 Ok(message) => {
-                                    let mut extensions = heartbeat_request.extensions.clone();
+                                    let mut extensions = env_var_extensions.clone();
                                     let gc_stat = gc_limiter.gc_stat();
                                     gc_stat.into_extensions(&mut extensions);
 
@@ -328,7 +332,7 @@ impl HeartbeatTask {
                         let now = Instant::now();
                         let duration_since_epoch = (now - epoch).as_millis() as u64;
 
-                        let mut extensions = heartbeat_request.extensions.clone();
+                        let mut extensions = env_var_extensions.clone();
                         let gc_stat = gc_limiter.gc_stat();
                         gc_stat.into_extensions(&mut extensions);
 
diff --git a/src/datanode/src/heartbeat/handler.rs b/src/datanode/src/heartbeat/handler.rs
index defc910573..10948a3e7c 100644
--- a/src/datanode/src/heartbeat/handler.rs
+++ b/src/datanode/src/heartbeat/handler.rs
@@ -295,7 +295,7 @@ impl HeartbeatResponseHandler for RegionHeartbeatResponseHandler {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
     use std::sync::Arc;
     use std::time::Duration;
diff --git a/src/datanode/src/heartbeat/handler/close_region.rs b/src/datanode/src/heartbeat/handler/close_region.rs
index 819fcc4880..484edf5ff4 100644
--- a/src/datanode/src/heartbeat/handler/close_region.rs
+++ b/src/datanode/src/heartbeat/handler/close_region.rs
@@ -47,7 +47,7 @@ impl InstructionHandler for CloseRegionsHandler {
         let results = join_all(futs).await;
 
         let mut errors = vec![];
-        for (region_id, result) in region_ids.into_iter().zip(results.into_iter()) {
+        for (region_id, result) in region_ids.into_iter().zip(results) {
             match result {
                 Ok(_) => (),
                 Err(error::Error::RegionNotFound { .. }) => {
@@ -79,7 +79,6 @@ mod tests {
     use std::assert_matches;
     use std::sync::Arc;
 
-    use assert_matches::assert_matches;
     use common_meta::RegionIdent;
     use common_meta::heartbeat::handler::{HandleControl, HeartbeatResponseHandler};
     use common_meta::heartbeat::mailbox::MessageMeta;
diff --git a/src/datanode/src/heartbeat/handler/downgrade_region.rs b/src/datanode/src/heartbeat/handler/downgrade_region.rs
index 40d9765ca2..98ace5661d 100644
--- a/src/datanode/src/heartbeat/handler/downgrade_region.rs
+++ b/src/datanode/src/heartbeat/handler/downgrade_region.rs
@@ -113,9 +113,7 @@ impl DowngradeRegionsHandler {
                     region_server_moved
                         .handle_request(
                             region_id,
-                            RegionRequest::Flush(RegionFlushRequest {
-                                row_group_size: None,
-                            }),
+                            RegionRequest::Flush(RegionFlushRequest::default()),
                         )
                         .await?;
 
@@ -225,7 +223,7 @@ impl HandlerContext {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::sync::Arc;
     use std::time::Duration;
 
diff --git a/src/datanode/src/heartbeat/handler/flush_region.rs b/src/datanode/src/heartbeat/handler/flush_region.rs
index dafc92ff0f..ca5eac1c15 100644
--- a/src/datanode/src/heartbeat/handler/flush_region.rs
+++ b/src/datanode/src/heartbeat/handler/flush_region.rs
@@ -18,7 +18,7 @@ use common_meta::instruction::{
     FlushErrorStrategy, FlushRegionReply, FlushRegions, FlushStrategy, InstructionReply,
 };
 use common_telemetry::{debug, warn};
-use store_api::region_request::{RegionFlushRequest, RegionRequest};
+use store_api::region_request::{RegionFlushReason, RegionFlushRequest, RegionRequest};
 use store_api::storage::RegionId;
 
 use crate::error::{self, RegionNotFoundSnafu, RegionNotReadySnafu, Result, UnexpectedSnafu};
@@ -39,14 +39,17 @@ impl InstructionHandler for FlushRegionsHandler {
         let strategy = flush_regions.strategy;
         let region_ids = flush_regions.region_ids;
         let error_strategy = flush_regions.error_strategy;
+        let reason = flush_regions.reason;
 
         let reply = if matches!(strategy, FlushStrategy::Async) {
             // Asynchronous hint mode: fire-and-forget, no reply expected
-            ctx.handle_flush_hint(region_ids).await;
+            ctx.handle_flush_hint(region_ids, reason).await;
             None
         } else {
             // Synchronous mode: return reply with results
-            let reply = ctx.handle_flush_sync(region_ids, error_strategy).await;
+            let reply = ctx
+                .handle_flush_sync(region_ids, error_strategy, reason)
+                .await;
             Some(InstructionReply::FlushRegions(reply))
         };
 
@@ -62,9 +65,14 @@ impl InstructionHandler for FlushRegionsHandler {
 
 impl HandlerContext {
     /// Performs the actual region flush operation.
-    async fn perform_region_flush(&self, region_id: RegionId) -> Result<()> {
+    async fn perform_region_flush(
+        &self,
+        region_id: RegionId,
+        reason: Option<RegionFlushReason>,
+    ) -> Result<()> {
         let request = RegionRequest::Flush(RegionFlushRequest {
-            row_group_size: None,
+            reason,
+            ..Default::default()
         });
         self.region_server
             .handle_request(region_id, request)
@@ -73,10 +81,14 @@ impl HandlerContext {
     }
 
     /// Handles asynchronous flush hints (fire-and-forget).
-    async fn handle_flush_hint(&self, region_ids: Vec<RegionId>) {
+    async fn handle_flush_hint(
+        &self,
+        region_ids: Vec<RegionId>,
+        reason: Option<RegionFlushReason>,
+    ) {
         let start_time = Instant::now();
         for region_id in &region_ids {
-            let result = self.perform_region_flush(*region_id).await;
+            let result = self.perform_region_flush(*region_id, reason).await;
             match result {
                 Ok(_) => {}
                 Err(error::Error::RegionNotFound { .. }) => {
@@ -102,11 +114,12 @@ impl HandlerContext {
         &self,
         region_ids: Vec<RegionId>,
         error_strategy: FlushErrorStrategy,
+        reason: Option<RegionFlushReason>,
     ) -> FlushRegionReply {
         let mut results = Vec::with_capacity(region_ids.len());
 
         for region_id in region_ids {
-            let result = self.flush_single_region_sync(region_id).await;
+            let result = self.flush_single_region_sync(region_id, reason).await;
 
             match &result {
                 Ok(_) => results.push((region_id, Ok(()))),
@@ -127,7 +140,11 @@ impl HandlerContext {
     }
 
     /// Flushes a single region synchronously with proper error handling.
-    async fn flush_single_region_sync(&self, region_id: RegionId) -> Result<()> {
+    async fn flush_single_region_sync(
+        &self,
+        region_id: RegionId,
+        reason: Option<RegionFlushReason>,
+    ) -> Result<()> {
         // Check if region is leader and writable
         let Some(writable) = self.region_server.is_region_leader(region_id) else {
             return Err(RegionNotFoundSnafu { region_id }.build());
@@ -148,7 +165,8 @@ impl HandlerContext {
                         .handle_request(
                             region_id,
                             RegionRequest::Flush(RegionFlushRequest {
-                                row_group_size: None,
+                                reason,
+                                ..Default::default()
                             }),
                         )
                         .await?;
@@ -184,19 +202,27 @@ mod tests {
     use super::*;
     use crate::tests::{MockRegionEngine, mock_region_server};
 
+    type FlushedRequests = Arc<RwLock<Vec<(RegionId, Option<RegionFlushReason>)>>>;
+
     #[tokio::test]
     async fn test_handle_flush_region_hint() {
-        let flushed_region_ids: Arc<RwLock<Vec<RegionId>>> = Arc::new(RwLock::new(Vec::new()));
+        let flushed_requests: FlushedRequests = Arc::new(RwLock::new(Vec::new()));
 
         let mock_region_server = mock_region_server();
         let region_ids = (0..16).map(|i| RegionId::new(1024, i)).collect::<Vec<_>>();
         for region_id in &region_ids {
-            let flushed_region_ids_ref = flushed_region_ids.clone();
+            let flushed_requests_ref = flushed_requests.clone();
             let (mock_engine, _) =
                 MockRegionEngine::with_custom_apply_fn(MITO_ENGINE_NAME, move |region_engine| {
                     region_engine.handle_request_mock_fn =
-                        Some(Box::new(move |region_id, _request| {
-                            flushed_region_ids_ref.write().unwrap().push(region_id);
+                        Some(Box::new(move |region_id, request| {
+                            let RegionRequest::Flush(request) = request else {
+                                panic!("Expected flush request");
+                            };
+                            flushed_requests_ref
+                                .write()
+                                .unwrap()
+                                .push((region_id, request.reason));
                             Ok(0)
                         }))
                 });
@@ -206,36 +232,47 @@ mod tests {
         let handler_context = HandlerContext::new_for_test(mock_region_server, kv_backend);
 
         // Async hint mode
-        let flush_instruction = FlushRegions::async_batch(region_ids.clone());
+        let flush_instruction = FlushRegions::async_batch(region_ids.clone())
+            .with_reason(RegionFlushReason::RemoteWalPrune);
         let reply = FlushRegionsHandler
             .handle(&handler_context, flush_instruction)
             .await;
         assert!(reply.is_none()); // Hint mode returns no reply
-        assert_eq!(*flushed_region_ids.read().unwrap(), region_ids);
+        let expected = region_ids
+            .iter()
+            .map(|region_id| (*region_id, Some(RegionFlushReason::RemoteWalPrune)))
+            .collect::<Vec<_>>();
+        assert_eq!(*flushed_requests.read().unwrap(), expected);
 
         // Non-existent regions
-        flushed_region_ids.write().unwrap().clear();
+        flushed_requests.write().unwrap().clear();
         let not_found_region_ids = (0..2).map(|i| RegionId::new(2048, i)).collect::<Vec<_>>();
         let flush_instruction = FlushRegions::async_batch(not_found_region_ids);
         let reply = FlushRegionsHandler
             .handle(&handler_context, flush_instruction)
             .await;
         assert!(reply.is_none());
-        assert!(flushed_region_ids.read().unwrap().is_empty());
+        assert!(flushed_requests.read().unwrap().is_empty());
     }
 
     #[tokio::test]
     async fn test_handle_flush_region_sync_single() {
-        let flushed_region_ids: Arc<RwLock<Vec<RegionId>>> = Arc::new(RwLock::new(Vec::new()));
+        let flushed_requests: FlushedRequests = Arc::new(RwLock::new(Vec::new()));
 
         let mock_region_server = mock_region_server();
         let region_id = RegionId::new(1024, 1);
 
-        let flushed_region_ids_ref = flushed_region_ids.clone();
+        let flushed_requests_ref = flushed_requests.clone();
         let (mock_engine, _) =
             MockRegionEngine::with_custom_apply_fn(MITO_ENGINE_NAME, move |region_engine| {
-                region_engine.handle_request_mock_fn = Some(Box::new(move |region_id, _request| {
-                    flushed_region_ids_ref.write().unwrap().push(region_id);
+                region_engine.handle_request_mock_fn = Some(Box::new(move |region_id, request| {
+                    let RegionRequest::Flush(request) = request else {
+                        panic!("Expected flush request");
+                    };
+                    flushed_requests_ref
+                        .write()
+                        .unwrap()
+                        .push((region_id, request.reason));
                     Ok(0)
                 }))
             });
@@ -243,7 +280,8 @@ mod tests {
         let kv_backend = Arc::new(MemoryKvBackend::new());
         let handler_context = HandlerContext::new_for_test(mock_region_server, kv_backend);
 
-        let flush_instruction = FlushRegions::sync_single(region_id);
+        let flush_instruction =
+            FlushRegions::sync_single(region_id).with_reason(RegionFlushReason::Repartition);
         let reply = FlushRegionsHandler
             .handle(&handler_context, flush_instruction)
             .await;
@@ -252,7 +290,10 @@ mod tests {
         assert_eq!(flush_reply.results.len(), 1);
         assert_eq!(flush_reply.results[0].0, region_id);
         assert!(flush_reply.results[0].1.is_ok());
-        assert_eq!(*flushed_region_ids.read().unwrap(), vec![region_id]);
+        assert_eq!(
+            *flushed_requests.read().unwrap(),
+            vec![(region_id, Some(RegionFlushReason::Repartition))]
+        );
     }
 
     #[tokio::test]
@@ -333,7 +374,7 @@ mod tests {
         let display = format!("{}", flush_regions);
         assert_eq!(
             display,
-            "FlushRegions(region_ids=[4398046511105(1024, 1)], strategy=Sync, error_strategy=FailFast)"
+            "FlushRegions(region_ids=[4398046511105(1024, 1)], strategy=Sync, error_strategy=FailFast, reason=None)"
         );
     }
 }
diff --git a/src/datanode/src/heartbeat/handler/open_region.rs b/src/datanode/src/heartbeat/handler/open_region.rs
index 91ba618d9a..56c07a3efe 100644
--- a/src/datanode/src/heartbeat/handler/open_region.rs
+++ b/src/datanode/src/heartbeat/handler/open_region.rs
@@ -72,7 +72,7 @@ impl InstructionHandler for OpenRegionsHandler {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
     use std::sync::Arc;
 
diff --git a/src/datanode/src/heartbeat/handler/upgrade_region.rs b/src/datanode/src/heartbeat/handler/upgrade_region.rs
index b2036a6ef4..c06e8aa845 100644
--- a/src/datanode/src/heartbeat/handler/upgrade_region.rs
+++ b/src/datanode/src/heartbeat/handler/upgrade_region.rs
@@ -183,9 +183,10 @@ impl UpgradeRegionsHandler {
         .await
         {
             Ok(responses) => {
-                replies.extend(
-                    Self::convert_responses_to_replies(responses, &catchup_regions).into_iter(),
-                );
+                replies.extend(Self::convert_responses_to_replies(
+                    responses,
+                    &catchup_regions,
+                ));
             }
             Err(_) => {
                 replies.extend(catchup_regions.iter().map(|region_id| UpgradeRegionReply {
diff --git a/src/datanode/src/lib.rs b/src/datanode/src/lib.rs
index 55d2b1796d..7e0db3cabc 100644
--- a/src/datanode/src/lib.rs
+++ b/src/datanode/src/lib.rs
@@ -12,8 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
-
 pub mod alive_keeper;
 pub mod config;
 pub mod datanode;
diff --git a/src/datanode/src/region_server.rs b/src/datanode/src/region_server.rs
index 9d675aa276..ac8cf4df26 100644
--- a/src/datanode/src/region_server.rs
+++ b/src/datanode/src/region_server.rs
@@ -17,8 +17,10 @@ mod catalog;
 use std::collections::HashMap;
 use std::fmt::Debug;
 use std::ops::Deref;
+use std::pin::Pin;
 use std::sync::atomic::{AtomicBool, Ordering};
 use std::sync::{Arc, RwLock};
+use std::task::{Context, Poll};
 use std::time::Duration;
 
 use api::region::RegionResponse;
@@ -36,7 +38,8 @@ use common_error::status_code::StatusCode;
 use common_meta::datanode::TopicStatsReporter;
 use common_query::OutputData;
 use common_query::request::QueryRequest;
-use common_recordbatch::SendableRecordBatchStream;
+use common_recordbatch::adapter::RecordBatchMetrics;
+use common_recordbatch::{OrderOption, RecordBatch, RecordBatchStream, SendableRecordBatchStream};
 use common_runtime::Runtime;
 use common_telemetry::tracing::{self, info_span};
 use common_telemetry::tracing_context::{FutureExt, TracingContext};
@@ -45,6 +48,7 @@ use dashmap::DashMap;
 use datafusion::datasource::TableProvider;
 use datafusion_common::tree_node::TreeNode;
 use either::Either;
+use futures_util::Stream;
 use futures_util::future::try_join_all;
 use metric_engine::engine::MetricEngine;
 use mito2::engine::{MITO_ENGINE_NAME, MitoEngine};
@@ -53,6 +57,7 @@ use query::QueryEngineRef;
 pub use query::dummy_catalog::{
     DummyCatalogList, DummyTableProviderFactory, TableProviderFactoryRef,
 };
+use query::options::should_collect_region_watermark_from_extensions;
 use serde_json;
 use servers::error::{
     self as servers_error, ExecuteGrpcRequestSnafu, Result as ServerResult, SuspendedSnafu,
@@ -278,16 +283,21 @@ impl RegionServer {
             .await
             .context(DecodeLogicalPlanSnafu)?;
 
-        self.inner
+        let stream = self
+            .inner
             .handle_read(
                 QueryRequest {
                     header: request.header,
                     region_id,
                     plan,
                 },
-                query_ctx,
+                query_ctx.clone(),
             )
-            .await
+            .await?;
+
+        Ok(wrap_flow_region_watermark_stream(
+            stream, region_id, &query_ctx,
+        ))
     }
 
     #[tracing::instrument(skip_all)]
@@ -313,9 +323,15 @@ impl RegionServer {
             .context(DataFusionSnafu)?
             .data;
 
-        self.inner
-            .handle_read(QueryRequest { plan, ..request }, query_ctx)
-            .await
+        let region_id = request.region_id;
+        let stream = self
+            .inner
+            .handle_read(QueryRequest { plan, ..request }, query_ctx.clone())
+            .await?;
+
+        Ok(wrap_flow_region_watermark_stream(
+            stream, region_id, &query_ctx,
+        ))
     }
 
     /// Returns all opened and reportable regions.
@@ -360,6 +376,7 @@ impl RegionServer {
             engine.role(region_id).map(|role| match role {
                 RegionRole::Follower => false,
                 RegionRole::Leader => true,
+                RegionRole::StagingLeader => true,
                 RegionRole::DowngradingLeader => true,
             })
         })
@@ -748,6 +765,99 @@ impl RegionServer {
     }
 }
 
+fn wrap_flow_region_watermark_stream(
+    stream: SendableRecordBatchStream,
+    region_id: RegionId,
+    query_ctx: &QueryContextRef,
+) -> SendableRecordBatchStream {
+    let Some(seq) = should_collect_region_watermark_from_extensions(&query_ctx.extensions())
+        .then(|| query_ctx.get_snapshot(region_id.as_u64()))
+        .flatten()
+    else {
+        return stream;
+    };
+
+    Box::pin(RegionWatermarkStream::new(stream, region_id, seq))
+}
+
+/// Wraps a region read stream so terminal metrics can carry the scan-open watermark.
+struct RegionWatermarkStream {
+    stream: SendableRecordBatchStream,
+    region_id: u64,
+    snapshot_seq: u64,
+    finished: bool,
+}
+
+impl RegionWatermarkStream {
+    fn new(stream: SendableRecordBatchStream, region_id: RegionId, snapshot_seq: u64) -> Self {
+        Self {
+            stream,
+            region_id: region_id.as_u64(),
+            snapshot_seq,
+            finished: false,
+        }
+    }
+
+    fn merged_metrics(&self, mut metrics: RecordBatchMetrics) -> RecordBatchMetrics {
+        if metrics
+            .region_watermarks
+            .iter()
+            .any(|entry| entry.region_id == self.region_id)
+        {
+            return metrics;
+        }
+
+        metrics
+            .region_watermarks
+            .push(common_recordbatch::adapter::RegionWatermarkEntry {
+                region_id: self.region_id,
+                watermark: Some(self.snapshot_seq),
+            });
+        metrics
+    }
+}
+
+impl RecordBatchStream for RegionWatermarkStream {
+    fn name(&self) -> &str {
+        self.stream.name()
+    }
+
+    fn schema(&self) -> datatypes::schema::SchemaRef {
+        self.stream.schema()
+    }
+
+    fn output_ordering(&self) -> Option<&[OrderOption]> {
+        self.stream.output_ordering()
+    }
+
+    fn metrics(&self) -> Option<RecordBatchMetrics> {
+        let base = self.stream.metrics();
+        if !self.finished {
+            return base;
+        }
+
+        Some(self.merged_metrics(base.unwrap_or_default()))
+    }
+}
+
+impl Stream for RegionWatermarkStream {
+    type Item = common_recordbatch::error::Result<RecordBatch>;
+
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        self.stream.size_hint()
+    }
+
+    fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        match Pin::new(&mut self.stream).poll_next(cx) {
+            Poll::Ready(None) => {
+                self.finished = true;
+                Poll::Ready(None)
+            }
+            other => other,
+        }
+    }
+}
+
 #[async_trait]
 impl RegionServerHandler for RegionServer {
     async fn handle(&self, request: region_request::Body) -> ServerResult<RegionResponseV1> {
@@ -1667,12 +1777,20 @@ impl RegionAttribute {
 #[cfg(test)]
 mod tests {
 
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
+    use std::collections::HashMap;
+    use std::sync::Arc;
 
     use api::v1::SemanticType;
     use common_error::ext::ErrorExt;
-    use datatypes::prelude::ConcreteDataType;
+    use common_recordbatch::RecordBatches;
+    use common_recordbatch::adapter::{RecordBatchMetrics, RegionWatermarkEntry};
+    use datatypes::prelude::{ConcreteDataType, VectorRef};
+    use datatypes::schema::{ColumnSchema, Schema};
+    use datatypes::vectors::Int32Vector;
+    use futures_util::StreamExt;
     use mito2::test_util::CreateRequestBuilder;
+    use query::options::FLOW_RETURN_REGION_SEQ;
     use store_api::metadata::{ColumnMetadata, RegionMetadata, RegionMetadataBuilder};
     use store_api::region_engine::RegionEngine;
     use store_api::region_request::{
@@ -1684,6 +1802,144 @@ mod tests {
     use crate::error::Result;
     use crate::tests::{MockRegionEngine, mock_region_server};
 
+    fn single_value_stream() -> SendableRecordBatchStream {
+        let schema = Arc::new(Schema::new(vec![ColumnSchema::new(
+            "v",
+            ConcreteDataType::int32_datatype(),
+            false,
+        )]));
+        let values: VectorRef = Arc::new(Int32Vector::from_slice([1]));
+        let batch = RecordBatch::new(schema.clone(), vec![values]).unwrap();
+        RecordBatches::try_new(schema, vec![batch])
+            .unwrap()
+            .as_stream()
+    }
+
+    #[tokio::test]
+    async fn test_region_watermark_stream_only_sets_terminal_metrics() {
+        let schema = Arc::new(Schema::new(vec![ColumnSchema::new(
+            "v",
+            ConcreteDataType::int32_datatype(),
+            false,
+        )]));
+        let values: VectorRef = Arc::new(Int32Vector::from_slice([1, 2]));
+        let batch = RecordBatch::new(schema.clone(), vec![values]).unwrap();
+        let stream = RecordBatches::try_new(schema, vec![batch])
+            .unwrap()
+            .as_stream();
+
+        let region_id = RegionId::new(42, 7);
+        let wrapped = RegionWatermarkStream::new(stream, region_id, 99);
+        let mut pinned = Box::pin(wrapped);
+
+        assert!(pinned.as_ref().get_ref().metrics().is_none());
+        while pinned.next().await.is_some() {}
+
+        let metrics = pinned.as_ref().get_ref().metrics().unwrap();
+        assert_eq!(
+            metrics.region_watermarks,
+            vec![RegionWatermarkEntry {
+                region_id: region_id.as_u64(),
+                watermark: Some(99),
+            }]
+        );
+    }
+
+    #[test]
+    fn test_region_watermark_stream_preserves_unproved_watermark() {
+        let schema = Arc::new(Schema::new(vec![ColumnSchema::new(
+            "v",
+            ConcreteDataType::int32_datatype(),
+            false,
+        )]));
+        let values: VectorRef = Arc::new(Int32Vector::from_slice([1]));
+        let batch = RecordBatch::new(schema.clone(), vec![values]).unwrap();
+        let stream = RecordBatches::try_new(schema, vec![batch])
+            .unwrap()
+            .as_stream();
+
+        let region_id = RegionId::new(42, 7);
+        let wrapped = RegionWatermarkStream::new(stream, region_id, 99);
+        let metrics = RecordBatchMetrics {
+            region_watermarks: vec![RegionWatermarkEntry {
+                region_id: region_id.as_u64(),
+                watermark: None,
+            }],
+            ..Default::default()
+        };
+
+        let merged = wrapped.merged_metrics(metrics);
+        assert_eq!(
+            merged.region_watermarks,
+            vec![RegionWatermarkEntry {
+                region_id: region_id.as_u64(),
+                watermark: None,
+            }]
+        );
+    }
+
+    #[tokio::test]
+    async fn test_wrap_flow_region_watermark_stream_adds_terminal_metrics() {
+        let region_id = RegionId::new(42, 7);
+        let query_ctx = Arc::new(
+            QueryContextBuilder::default()
+                .extensions(HashMap::from([(
+                    FLOW_RETURN_REGION_SEQ.to_string(),
+                    "true".to_string(),
+                )]))
+                .build(),
+        );
+        query_ctx.set_snapshot(region_id.as_u64(), 99);
+
+        let wrapped =
+            wrap_flow_region_watermark_stream(single_value_stream(), region_id, &query_ctx);
+        let mut pinned = Box::pin(wrapped);
+        while pinned.next().await.is_some() {}
+
+        let metrics = pinned.as_ref().get_ref().metrics().unwrap();
+        assert_eq!(
+            metrics.region_watermarks,
+            vec![RegionWatermarkEntry {
+                region_id: region_id.as_u64(),
+                watermark: Some(99),
+            }]
+        );
+    }
+
+    #[tokio::test]
+    async fn test_wrap_flow_region_watermark_stream_skips_without_extension() {
+        let region_id = RegionId::new(42, 7);
+        let query_ctx = Arc::new(QueryContextBuilder::default().build());
+        query_ctx.set_snapshot(region_id.as_u64(), 99);
+
+        let wrapped =
+            wrap_flow_region_watermark_stream(single_value_stream(), region_id, &query_ctx);
+        let mut pinned = Box::pin(wrapped);
+        while pinned.next().await.is_some() {}
+
+        assert!(pinned.as_ref().get_ref().metrics().is_none());
+    }
+
+    #[tokio::test]
+    async fn test_wrap_flow_region_watermark_stream_skips_without_snapshot() {
+        let region_id = RegionId::new(42, 7);
+        let query_ctx = Arc::new(
+            QueryContextBuilder::default()
+                .extensions(HashMap::from([(
+                    FLOW_RETURN_REGION_SEQ.to_string(),
+                    "true".to_string(),
+                )]))
+                .build(),
+        );
+
+        let wrapped =
+            wrap_flow_region_watermark_stream(single_value_stream(), region_id, &query_ctx);
+        let mut pinned = Box::pin(wrapped);
+        while pinned.next().await.is_some() {}
+
+        assert!(pinned.as_ref().get_ref().metrics().is_none());
+    }
+
     #[tokio::test]
     async fn test_region_registering() {
         common_telemetry::init_default_ut_logging();
diff --git a/src/datatypes/src/data_type.rs b/src/datatypes/src/data_type.rs
index e39d31d3d1..56bc0ad249 100644
--- a/src/datatypes/src/data_type.rs
+++ b/src/datatypes/src/data_type.rs
@@ -276,10 +276,6 @@ impl ConcreteDataType {
         matches!(self, ConcreteDataType::Null(NullType))
     }
 
-    pub(crate) fn is_struct(&self) -> bool {
-        matches!(self, ConcreteDataType::Struct(_))
-    }
-
     /// Try to cast the type as a [`ListType`].
     pub fn as_list(&self) -> Option<&ListType> {
         match self {
@@ -688,7 +684,7 @@ impl ConcreteDataType {
     }
 
     pub fn json_native_datatype(inner_type: ConcreteDataType) -> ConcreteDataType {
-        ConcreteDataType::Json(JsonType::new_native((&inner_type).into()))
+        ConcreteDataType::Json(JsonType::new_json2((&inner_type).into()))
     }
 }
 
diff --git a/src/datatypes/src/error.rs b/src/datatypes/src/error.rs
index 65aca699ec..4ca765222a 100644
--- a/src/datatypes/src/error.rs
+++ b/src/datatypes/src/error.rs
@@ -274,6 +274,20 @@ pub enum Error {
         #[snafu(implicit)]
         location: Location,
     },
+
+    #[snafu(display("Failed to align JSON value, reason: {}", reason))]
+    AlignJsonValue {
+        reason: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Failed to align JSON array, reason: {reason}"))]
+    AlignJsonArray {
+        reason: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
 }
 
 impl ErrorExt for Error {
@@ -316,7 +330,9 @@ impl ErrorExt for Error {
             | ConvertScalarToArrowArray { .. }
             | ParseExtendedType { .. }
             | InconsistentStructFieldsAndItems { .. }
-            | ArrowMetadata { .. } => StatusCode::Internal,
+            | ArrowMetadata { .. }
+            | AlignJsonValue { .. }
+            | AlignJsonArray { .. } => StatusCode::Internal,
         }
     }
 
diff --git a/src/datatypes/src/json.rs b/src/datatypes/src/json.rs
index 3bebbf89aa..db657abbcb 100644
--- a/src/datatypes/src/json.rs
+++ b/src/datatypes/src/json.rs
@@ -426,7 +426,7 @@ fn decode_struct_with_context<'a>(
 
     let (items, fields) = struct_value.into_parts();
 
-    for (field, field_value) in fields.fields().iter().zip(items.into_iter()) {
+    for (field, field_value) in fields.fields().iter().zip(items) {
         let field_context = context.with_key(field.name());
         let json_value = decode_value_with_context(field_value, &field_context)?;
         json_object.insert(field.name().to_string(), json_value);
@@ -561,7 +561,7 @@ fn decode_struct_with_settings<'a>(
 
     // Process each field in the struct value
     let (struct_data, fields) = struct_value.into_parts();
-    for (field, value) in fields.fields().iter().zip(struct_data.into_iter()) {
+    for (field, value) in fields.fields().iter().zip(struct_data) {
         let field_context = context.with_key(field.name());
 
         // Check if this field should be treated as unstructured
diff --git a/src/datatypes/src/json/value.rs b/src/datatypes/src/json/value.rs
index f8cf71e936..e2422ca78d 100644
--- a/src/datatypes/src/json/value.rs
+++ b/src/datatypes/src/json/value.rs
@@ -21,9 +21,12 @@ use num_traits::ToPrimitive;
 use ordered_float::OrderedFloat;
 use serde::{Deserialize, Serialize};
 use serde_json::Number;
+use snafu::{OptionExt, ResultExt, ensure};
 
+use crate::Result;
 use crate::data_type::ConcreteDataType;
-use crate::types::json_type::JsonNativeType;
+use crate::error::{AlignJsonValueSnafu, SerializeSnafu};
+use crate::types::json_type::{JsonNativeType, JsonNumberType};
 use crate::types::{JsonType, StructField, StructType};
 use crate::value::{ListValue, ListValueRef, StructValue, StructValueRef, Value, ValueRef};
 
@@ -105,19 +108,17 @@ impl Display for JsonNumber {
 }
 
 /// Variants of json.
-///
-/// This follows how [serde_json::Value] designs except that we only choose to use [BTreeMap] to
-/// preserve the fields order by their names in the json object. (By default `serde_json` uses
-/// [BTreeMap], too. But it additionally supports "IndexMap" which preserves the order by insertion
-/// times of fields.)
-#[derive(Debug, Clone, PartialEq, Eq, Hash, Serialize, Deserialize)]
+#[derive(Debug, Clone, PartialEq, Eq, Hash, Serialize, Deserialize, Default)]
 pub enum JsonVariant {
+    #[default]
     Null,
     Bool(bool),
     Number(JsonNumber),
     String(String),
     Array(Vec<JsonVariant>),
     Object(BTreeMap<String, JsonVariant>),
+    /// A special "variant" value of JSON, to represent a union result of conflict JSON type values.
+    Variant(Vec<u8>),
 }
 
 impl JsonVariant {
@@ -166,11 +167,12 @@ impl JsonVariant {
                     .map(|(k, v)| (k.clone(), v.native_type()))
                     .collect(),
             ),
+            JsonVariant::Variant(_) => JsonNativeType::Variant,
         }
     }
 
     fn json_type(&self) -> JsonType {
-        JsonType::new_native(self.native_type())
+        JsonType::new_json2(self.native_type())
     }
 
     fn as_ref(&self) -> JsonVariantRef<'_> {
@@ -192,6 +194,7 @@ impl JsonVariant {
                     .map(|(k, v)| (k.as_str(), v.as_ref()))
                     .collect(),
             ),
+            JsonVariant::Variant(v) => JsonVariantRef::Variant(v),
         }
     }
 }
@@ -291,6 +294,13 @@ impl Display for JsonVariant {
                         .join(", ")
                 )
             }
+            Self::Variant(x) => {
+                let result: serde_json::Result<serde_json::Value> = serde_json::from_slice(x);
+                match result {
+                    Ok(v) => write!(f, "{v}"),
+                    Err(_) => write!(f, "{x:?}"),
+                }
+            }
         }
     }
 }
@@ -403,10 +413,110 @@ impl JsonValue {
                     }
                     Value::Struct(StructValue::new(items, StructType::new(Arc::new(fields))))
                 }
+                JsonVariant::Variant(x) => Value::Binary(x.into()),
             }
         }
         helper(self.json_variant)
     }
+
+    /// Recursively aligns this JSON value to `expected` in place. This is to make JSON values fill
+    /// into a [StructArray], which requires a unified static datatype.
+    ///
+    /// Alignment follows these rules:
+    /// - `Null` aligns to any type, and any value aligns to `Null` as `Null`.
+    /// - Numbers are converted only within compatible number categories.
+    /// - Arrays align each element recursively to the expected item type.
+    /// - Objects require `expected` to contain all fields from the current value. Missing expected
+    ///   fields are filled with `Null`.
+    /// - `Variant` preserves the original JSON payload as serialized bytes.
+    ///
+    /// Returns an error if the value cannot be aligned without losing existing object fields or
+    /// when a scalar type conversion is incompatible.
+    pub(crate) fn try_align(&mut self, expected: &JsonType) -> Result<()> {
+        if self.json_type() == expected {
+            return Ok(());
+        }
+
+        fn helper(value: JsonVariant, expected: &JsonNativeType) -> Result<JsonVariant> {
+            Ok(match (value, expected) {
+                (JsonVariant::Null, _) | (_, JsonNativeType::Null) => JsonVariant::Null,
+                (JsonVariant::Bool(v), JsonNativeType::Bool) => JsonVariant::Bool(v),
+                (JsonVariant::Number(v), JsonNativeType::Number(n)) => {
+                    return match n {
+                        JsonNumberType::U64 => v
+                            .as_u64()
+                            .map(|x| JsonVariant::Number(JsonNumber::PosInt(x))),
+                        JsonNumberType::I64 => v
+                            .as_i64()
+                            .map(|x| JsonVariant::Number(JsonNumber::NegInt(x))),
+                        JsonNumberType::F64 => {
+                            Some(JsonVariant::Number(JsonNumber::Float(v.as_f64().into())))
+                        }
+                    }
+                    .with_context(|| AlignJsonValueSnafu {
+                        reason: format!("unable to align number ‘{}’ to type {}", v, expected),
+                    });
+                }
+                (JsonVariant::String(v), JsonNativeType::String) => JsonVariant::String(v),
+
+                (JsonVariant::Array(items), JsonNativeType::Array(expected)) => JsonVariant::Array(
+                    items
+                        .into_iter()
+                        .map(|item| helper(item, expected.as_ref()))
+                        .collect::<Result<_>>()?,
+                ),
+
+                (JsonVariant::Object(mut kvs), JsonNativeType::Object(expected)) => {
+                    ensure!(
+                        expected.keys().len() >= kvs.keys().len()
+                            && kvs.keys().all(|k| expected.contains_key(k)),
+                        AlignJsonValueSnafu {
+                            reason: format!(
+                                "aligned type '{}' should be superset of value '{}'",
+                                JsonNativeType::Object(expected.clone()),
+                                JsonVariant::from(kvs),
+                            )
+                        }
+                    );
+
+                    let mut object = BTreeMap::new();
+                    for (field, field_type) in expected {
+                        if let Some((k, v)) = kvs.remove_entry(field) {
+                            object.insert(k, helper(v, field_type)?);
+                        } else {
+                            object.insert(field.clone(), JsonVariant::Null);
+                        }
+                    }
+                    JsonVariant::Object(object)
+                }
+
+                (v, JsonNativeType::Variant) => {
+                    let json: serde_json::Value =
+                        JsonValue::new(v).try_into().context(SerializeSnafu)?;
+                    serde_json::to_vec(&json)
+                        .map(JsonVariant::Variant)
+                        .context(SerializeSnafu)?
+                }
+
+                (value, expected) => {
+                    return AlignJsonValueSnafu {
+                        reason: format!(
+                            "unable to align '{}' of type {} to type {}",
+                            value,
+                            value.native_type(),
+                            expected,
+                        ),
+                    }
+                    .fail();
+                }
+            })
+        }
+
+        let x = std::mem::take(&mut self.json_variant);
+        self.json_variant = helper(x, expected.native_type())?;
+        self.json_type = OnceLock::from(expected.clone());
+        Ok(())
+    }
 }
 
 impl<T: Into<JsonVariant>> From<T> for JsonValue {
@@ -418,10 +528,12 @@ impl<T: Into<JsonVariant>> From<T> for JsonValue {
     }
 }
 
-impl From<JsonValue> for serde_json::Value {
-    fn from(v: JsonValue) -> Self {
-        fn helper(v: JsonVariant) -> serde_json::Value {
-            match v {
+impl TryFrom<JsonValue> for serde_json::Value {
+    type Error = serde_json::Error;
+
+    fn try_from(v: JsonValue) -> serde_json::Result<Self> {
+        fn helper(v: JsonVariant) -> serde_json::Result<serde_json::Value> {
+            Ok(match v {
                 JsonVariant::Null => serde_json::Value::Null,
                 JsonVariant::Bool(x) => serde_json::Value::Bool(x),
                 JsonVariant::Number(x) => match x {
@@ -436,13 +548,21 @@ impl From<JsonValue> for serde_json::Value {
                     }
                 },
                 JsonVariant::String(x) => serde_json::Value::String(x),
-                JsonVariant::Array(array) => {
-                    serde_json::Value::Array(array.into_iter().map(helper).collect())
-                }
-                JsonVariant::Object(object) => serde_json::Value::Object(
-                    object.into_iter().map(|(k, v)| (k, helper(v))).collect(),
+                JsonVariant::Array(array) => serde_json::Value::Array(
+                    array
+                        .into_iter()
+                        .map(helper)
+                        .collect::<serde_json::Result<Vec<_>>>()?,
                 ),
-            }
+                JsonVariant::Object(object) => {
+                    let mut map = serde_json::Map::with_capacity(object.len());
+                    for (k, v) in object {
+                        map.insert(k, helper(v)?);
+                    }
+                    serde_json::Value::Object(map)
+                }
+                JsonVariant::Variant(x) => serde_json::from_slice(&x)?,
+            })
         }
         helper(v.json_variant)
     }
@@ -496,6 +616,7 @@ pub enum JsonVariantRef<'a> {
     String(&'a str),
     Array(Vec<JsonVariantRef<'a>>),
     Object(BTreeMap<&'a str, JsonVariantRef<'a>>),
+    Variant(&'a [u8]),
 }
 
 impl JsonVariantRef<'_> {
@@ -524,9 +645,10 @@ impl JsonVariantRef<'_> {
                         .map(|(k, v)| (k.to_string(), native_type(v)))
                         .collect(),
                 ),
+                JsonVariantRef::Variant(_) => JsonNativeType::Variant,
             }
         }
-        JsonType::new_native(native_type(self))
+        JsonType::new_json2(native_type(self))
     }
 }
 
@@ -596,10 +718,17 @@ impl From<JsonVariantRef<'_>> for JsonVariant {
                     .map(|(k, v)| (k.to_string(), v.into()))
                     .collect(),
             ),
+            JsonVariantRef::Variant(x) => Self::Variant(x.to_vec()),
         }
     }
 }
 
+impl<'a> From<&'a [u8]> for JsonVariantRef<'a> {
+    fn from(value: &'a [u8]) -> Self {
+        Self::Variant(value)
+    }
+}
+
 /// Reference to representation of any valid JSON value.
 #[derive(Debug, Serialize)]
 pub struct JsonValueRef<'a> {
@@ -647,7 +776,7 @@ impl<'a> JsonValueRef<'a> {
         }
     }
 
-    pub fn as_value_ref(&self) -> ValueRef<'_> {
+    fn as_value_ref(&self) -> ValueRef<'_> {
         fn helper<'a>(v: &'a JsonVariantRef) -> ValueRef<'a> {
             match v {
                 JsonVariantRef::Null => ValueRef::Null,
@@ -683,6 +812,7 @@ impl<'a> JsonValueRef<'a> {
                         fields: StructType::new(Arc::new(fields)),
                     })
                 }
+                JsonVariantRef::Variant(x) => ValueRef::Binary(x),
             }
         }
         helper(&self.json_variant)
@@ -691,6 +821,21 @@ impl<'a> JsonValueRef<'a> {
     pub(crate) fn data_size(&self) -> usize {
         size_of_val(self)
     }
+
+    pub(crate) fn variant(&self) -> &JsonVariantRef<'a> {
+        &self.json_variant
+    }
+
+    pub(crate) fn as_struct_value(&self) -> ValueRef<'_> {
+        if self.is_object() {
+            return self.as_value_ref();
+        }
+
+        ValueRef::Struct(StructValueRef::RefList {
+            val: vec![self.as_value_ref()],
+            fields: self.json_type().as_struct_type(),
+        })
+    }
 }
 
 impl<'a, T: Into<JsonVariantRef<'a>>> From<T> for JsonValueRef<'a> {
@@ -735,3 +880,100 @@ impl Clone for JsonValueRef<'_> {
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::types::json_type::JsonObjectType;
+
+    #[test]
+    fn test_align_json_value() -> Result<()> {
+        fn parse_json_value(json: &str) -> JsonValue {
+            let value: serde_json::Value = serde_json::from_str(json).unwrap();
+            value.into()
+        }
+
+        // Root type can be aligned to Null, and the cached json_type must be refreshed.
+        let mut value = JsonValue::from(true);
+        assert_eq!(
+            value.json_type(),
+            &JsonType::new_json2(JsonNativeType::Bool)
+        );
+        value.try_align(&JsonType::null())?;
+        assert_eq!(value, JsonValue::null());
+        assert_eq!(value.json_type(), &JsonType::null());
+
+        // Object alignment now requires the expected type to be a superset of the
+        // value fields, while still filling missing expected fields with null.
+        let expected = JsonType::new_json2(JsonNativeType::Object(JsonObjectType::from([
+            ("extra".to_string(), JsonNativeType::u64()),
+            (
+                "items".to_string(),
+                JsonNativeType::Array(Box::new(JsonNativeType::Object(JsonObjectType::from([
+                    ("id".to_string(), JsonNativeType::u64()),
+                    ("payload".to_string(), JsonNativeType::Variant),
+                    ("note".to_string(), JsonNativeType::String),
+                ])))),
+            ),
+            ("name".to_string(), JsonNativeType::String),
+        ])));
+        let mut value = parse_json_value(r#"{"items":[{"id":1,"payload":{"k":"v"}}],"extra":1}"#);
+        assert_ne!(value.json_type(), &expected);
+        value.try_align(&expected)?;
+        assert_eq!(
+            value,
+            JsonValue::from(JsonVariant::Object(BTreeMap::from([
+                ("extra".to_string(), JsonVariant::from(1_u64)),
+                (
+                    "items".to_string(),
+                    JsonVariant::Array(vec![JsonVariant::Object(BTreeMap::from([
+                        ("id".to_string(), JsonVariant::from(1_u64)),
+                        ("note".to_string(), JsonVariant::Null),
+                        (
+                            "payload".to_string(),
+                            JsonVariant::Variant(br#"{"k":"v"}"#.to_vec()),
+                        ),
+                    ]))]),
+                ),
+                ("name".to_string(), JsonVariant::Null),
+            ])))
+        );
+        assert_eq!(value.json_type(), &expected);
+
+        // Object alignment should fail if the expected type misses any field from the value.
+        let expected = JsonType::new_json2(JsonNativeType::Object(JsonObjectType::from([(
+            "items".to_string(),
+            JsonNativeType::Array(Box::new(JsonNativeType::Object(JsonObjectType::from([
+                ("id".to_string(), JsonNativeType::u64()),
+                ("payload".to_string(), JsonNativeType::Variant),
+            ])))),
+        )])));
+        let mut value =
+            parse_json_value(r#"{"items":[{"id":1,"payload":{"k":"v"},"extra":true}]}"#);
+        let err = value.try_align(&expected).unwrap_err();
+        assert_eq!(
+            err.to_string(),
+            r#"Failed to align JSON value, reason: aligned type '{"id":"<Number>","payload":"<Variant>"}' should be superset of value '{ extra: true, id: 1, payload: { k: v } }'"#
+        );
+
+        // Root-level Variant alignment should preserve the original JSON payload.
+        let mut value = parse_json_value(r#"{"foo":[1,true,null]}"#);
+        value.try_align(&JsonType::new_json2(JsonNativeType::Variant))?;
+        assert_eq!(
+            value,
+            JsonValue::from(JsonVariant::Variant(br#"{"foo":[1,true,null]}"#.to_vec()))
+        );
+
+        // Incompatible scalar alignment should fail instead of coercing the value.
+        let mut value = JsonValue::from("hello");
+        let err = value
+            .try_align(&JsonType::new_json2(JsonNativeType::Bool))
+            .unwrap_err();
+        assert_eq!(
+            err.to_string(),
+            r#"Failed to align JSON value, reason: unable to align 'hello' of type "<String>" to type "<Bool>""#
+        );
+
+        Ok(())
+    }
+}
diff --git a/src/datatypes/src/lib.rs b/src/datatypes/src/lib.rs
index 2c3d4c23bf..0dcce16857 100644
--- a/src/datatypes/src/lib.rs
+++ b/src/datatypes/src/lib.rs
@@ -12,9 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
-#![feature(box_patterns)]
-
 pub mod arrow_array;
 pub mod data_type;
 pub mod duration;
diff --git a/src/datatypes/src/schema.rs b/src/datatypes/src/schema.rs
index 9070e2babe..d53b42e3e7 100644
--- a/src/datatypes/src/schema.rs
+++ b/src/datatypes/src/schema.rs
@@ -14,10 +14,11 @@
 
 mod column_schema;
 pub mod constraint;
+pub mod ext;
 
 use std::collections::HashMap;
-use std::fmt;
 use std::sync::Arc;
+use std::{fmt, mem};
 
 use arrow::datatypes::{Field, Schema as ArrowSchema};
 use datafusion_common::DFSchemaRef;
@@ -177,6 +178,26 @@ impl Schema {
         &self.arrow_schema.metadata
     }
 
+    /// Returns the estimated memory footprint of this schema.
+    pub fn estimated_size(&self) -> usize {
+        mem::size_of_val(self)
+            + mem::size_of::<ColumnSchema>() * self.column_schemas.capacity()
+            + self
+                .column_schemas
+                .iter()
+                .map(|column_schema| {
+                    column_schema.estimated_size() - mem::size_of::<ColumnSchema>()
+                })
+                .sum::<usize>()
+            + mem::size_of::<(String, usize)>() * self.name_to_index.capacity()
+            + self
+                .name_to_index
+                .keys()
+                .map(|name| name.capacity())
+                .sum::<usize>()
+            + arrow_schema_size(self.arrow_schema.as_ref())
+    }
+
     /// Generate a new projected schema
     ///
     /// # Panic
@@ -213,6 +234,17 @@ impl Schema {
     }
 }
 
+fn arrow_schema_size(schema: &ArrowSchema) -> usize {
+    mem::size_of_val(schema)
+        + schema.fields.size()
+        + mem::size_of::<(String, String)>() * schema.metadata.capacity()
+        + schema
+            .metadata
+            .iter()
+            .map(|(key, value)| key.capacity() + value.capacity())
+            .sum::<usize>()
+}
+
 #[derive(Default)]
 pub struct SchemaBuilder {
     column_schemas: Vec<ColumnSchema>,
diff --git a/src/datatypes/src/schema/column_schema.rs b/src/datatypes/src/schema/column_schema.rs
index 183cf05da8..2479f4fc41 100644
--- a/src/datatypes/src/schema/column_schema.rs
+++ b/src/datatypes/src/schema/column_schema.rs
@@ -13,8 +13,8 @@
 // limitations under the License.
 
 use std::collections::HashMap;
-use std::fmt;
 use std::str::FromStr;
+use std::{fmt, mem};
 
 use arrow::datatypes::Field;
 use arrow_schema::extension::{
@@ -178,6 +178,19 @@ impl ColumnSchema {
         self
     }
 
+    /// Returns the estimated memory footprint of this schema.
+    pub fn estimated_size(&self) -> usize {
+        mem::size_of_val(self) - mem::size_of_val(&self.data_type)
+            + self.data_type.as_arrow_type().size()
+            + self.name.capacity()
+            + self
+                .default_constraint
+                .as_ref()
+                .map(column_default_constraint_size)
+                .unwrap_or_default()
+            + metadata_size(&self.metadata)
+    }
+
     /// Set the inverted index for the column.
     /// Similar to [with_inverted_index] but don't take the ownership.
     ///
@@ -493,6 +506,21 @@ impl ColumnSchema {
     }
 }
 
+fn metadata_size(metadata: &Metadata) -> usize {
+    mem::size_of::<(String, String)>() * metadata.capacity()
+        + metadata
+            .iter()
+            .map(|(key, value)| key.capacity() + value.capacity())
+            .sum::<usize>()
+}
+
+fn column_default_constraint_size(default_constraint: &ColumnDefaultConstraint) -> usize {
+    match default_constraint {
+        ColumnDefaultConstraint::Function(expr) => expr.capacity(),
+        ColumnDefaultConstraint::Value(value) => value.as_value_ref().data_size(),
+    }
+}
+
 /// Column extended type set in column schema's metadata.
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub enum ColumnExtType {
diff --git a/src/datatypes/src/schema/ext.rs b/src/datatypes/src/schema/ext.rs
new file mode 100644
index 0000000000..16ffcc6901
--- /dev/null
+++ b/src/datatypes/src/schema/ext.rs
@@ -0,0 +1,27 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use crate::extension::json;
+
+/// Add some useful utilities upon Arrow's [Schema](arrow_schema::Schema).
+pub trait ArrowSchemaExt {
+    /// Check if this [Schema](arrow_schema::Schema) has JSON extension field.
+    fn has_json_extension_field(&self) -> bool;
+}
+
+impl ArrowSchemaExt for arrow_schema::Schema {
+    fn has_json_extension_field(&self) -> bool {
+        self.fields().iter().any(json::is_json_extension_type)
+    }
+}
diff --git a/src/datatypes/src/types/json_type.rs b/src/datatypes/src/types/json_type.rs
index 61586fc460..cd479ecd37 100644
--- a/src/datatypes/src/types/json_type.rs
+++ b/src/datatypes/src/types/json_type.rs
@@ -23,9 +23,11 @@ use regex::{Captures, Regex};
 use serde::{Deserialize, Serialize};
 use snafu::ResultExt;
 
+use crate::Error;
 use crate::data_type::DataType;
 use crate::error::{
     DeserializeSnafu, InvalidJsonSnafu, InvalidJsonbSnafu, MergeJsonDatatypeSnafu, Result,
+    UnsupportedArrowTypeSnafu,
 };
 use crate::prelude::ConcreteDataType;
 use crate::scalars::ScalarVectorBuilder;
@@ -36,6 +38,7 @@ use crate::vectors::json::builder::JsonVectorBuilder;
 use crate::vectors::{BinaryVectorBuilder, MutableVector};
 
 pub const JSON_TYPE_NAME: &str = "Json";
+const JSON2_TYPE_NAME: &str = "Json2";
 const JSON_PLAIN_FIELD_NAME: &str = "__json_plain__";
 const JSON_PLAIN_FIELD_METADATA_KEY: &str = "is_plain_json";
 
@@ -48,14 +51,19 @@ pub enum JsonNumberType {
     F64,
 }
 
-#[derive(Debug, Clone, PartialEq, Eq, Hash, PartialOrd, Ord, Serialize, Deserialize)]
+#[derive(Debug, Clone, PartialEq, Eq, Hash, PartialOrd, Ord, Serialize, Deserialize, Default)]
 pub enum JsonNativeType {
+    #[default]
     Null,
     Bool,
     Number(JsonNumberType),
     String,
     Array(Box<JsonNativeType>),
     Object(JsonObjectType),
+    /// A special (not in the JSON official specification) JSON type to indicate the "resolved" or
+    /// "lifted" type of two conflicting JSON types. For example, when merging JSON types of "Bool"
+    /// and "Number".
+    Variant,
 }
 
 impl JsonNativeType {
@@ -74,6 +82,41 @@ impl JsonNativeType {
     pub fn f64() -> Self {
         Self::Number(JsonNumberType::F64)
     }
+
+    /// Merge other [JsonNativeType] into this.
+    /// Conflicting fields will be resolved to the "Variant" type.
+    pub fn merge(&mut self, other: &JsonNativeType) {
+        if self == other {
+            return;
+        }
+
+        fn merge_object(this: &mut JsonObjectType, that: &JsonObjectType) {
+            // merge "that" into "this" directly:
+            for (type_name, that_type) in that {
+                if let Some(this_type) = this.get_mut(type_name) {
+                    this_type.merge(that_type);
+                } else {
+                    this.insert(type_name.clone(), that_type.clone());
+                }
+            }
+        }
+
+        let zelf = std::mem::take(self);
+        *self = match (zelf, other) {
+            (JsonNativeType::Object(mut this), JsonNativeType::Object(that)) => {
+                merge_object(&mut this, that);
+                JsonNativeType::Object(this)
+            }
+            (JsonNativeType::Array(mut this), JsonNativeType::Array(that)) => {
+                this.merge(that);
+                JsonNativeType::Array(this)
+            }
+            (JsonNativeType::Null, that) => that.clone(),
+            (this, JsonNativeType::Null) => this,
+            (this, that) if this == *that => this,
+            _ => JsonNativeType::Variant,
+        };
+    }
 }
 
 impl From<&JsonNativeType> for ConcreteDataType {
@@ -97,6 +140,7 @@ impl From<&JsonNativeType> for ConcreteDataType {
                     .collect();
                 ConcreteDataType::Struct(StructType::new(Arc::new(fields)))
             }
+            JsonNativeType::Variant => ConcreteDataType::binary_datatype(),
         }
     }
 }
@@ -127,11 +171,62 @@ impl From<&ConcreteDataType> for JsonNativeType {
                     .collect(),
             ),
             ConcreteDataType::Json(json_type) => json_type.native_type().clone(),
+            ConcreteDataType::Binary(_) => JsonNativeType::Variant,
             _ => unreachable!(),
         }
     }
 }
 
+impl TryFrom<&ArrowDataType> for JsonNativeType {
+    type Error = Error;
+
+    fn try_from(t: &ArrowDataType) -> Result<Self> {
+        let t = match t {
+            ArrowDataType::Null => JsonNativeType::Null,
+            ArrowDataType::Boolean => JsonNativeType::Bool,
+            ArrowDataType::Int8
+            | ArrowDataType::Int16
+            | ArrowDataType::Int32
+            | ArrowDataType::Int64 => JsonNativeType::i64(),
+            ArrowDataType::UInt8
+            | ArrowDataType::UInt16
+            | ArrowDataType::UInt32
+            | ArrowDataType::UInt64 => JsonNativeType::u64(),
+            ArrowDataType::Float16 | ArrowDataType::Float32 | ArrowDataType::Float64 => {
+                JsonNativeType::f64()
+            }
+            ArrowDataType::Binary
+            | ArrowDataType::FixedSizeBinary(_)
+            | ArrowDataType::LargeBinary
+            | ArrowDataType::BinaryView => JsonNativeType::Variant,
+            ArrowDataType::Utf8 | ArrowDataType::LargeUtf8 | ArrowDataType::Utf8View => {
+                JsonNativeType::String
+            }
+            ArrowDataType::List(field)
+            | ArrowDataType::ListView(field)
+            | ArrowDataType::FixedSizeList(field, _)
+            | ArrowDataType::LargeList(field)
+            | ArrowDataType::LargeListView(field) => {
+                JsonNativeType::Array(Box::new(Self::try_from(field.data_type())?))
+            }
+            ArrowDataType::Struct(fields) => {
+                let mut object = JsonObjectType::new();
+                for field in fields {
+                    object.insert(field.name().clone(), Self::try_from(field.data_type())?);
+                }
+                JsonNativeType::Object(object)
+            }
+            t => {
+                return UnsupportedArrowTypeSnafu {
+                    arrow_type: t.clone(),
+                }
+                .fail();
+            }
+        };
+        Ok(t)
+    }
+}
+
 impl Display for JsonNativeType {
     fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
         match self {
@@ -155,6 +250,7 @@ impl Display for JsonNativeType {
                         .join(",")
                 )
             }
+            JsonNativeType::Variant => write!(f, r#""<Variant>""#),
         }
     }
 }
@@ -163,7 +259,7 @@ impl Display for JsonNativeType {
 pub enum JsonFormat {
     #[default]
     Jsonb,
-    Native(Box<JsonNativeType>),
+    Json2(Box<JsonNativeType>),
 }
 
 /// JsonType is a data type for JSON data. It is stored as binary data of jsonb format.
@@ -178,26 +274,26 @@ impl JsonType {
         Self { format }
     }
 
-    pub(crate) fn new_native(native: JsonNativeType) -> Self {
+    pub(crate) fn new_json2(native: JsonNativeType) -> Self {
         Self {
-            format: JsonFormat::Native(Box::new(native)),
+            format: JsonFormat::Json2(Box::new(native)),
         }
     }
 
-    pub fn is_native_type(&self) -> bool {
-        matches!(self.format, JsonFormat::Native(_))
+    pub fn is_json2(&self) -> bool {
+        matches!(self.format, JsonFormat::Json2(_))
     }
 
-    pub fn native_type(&self) -> &JsonNativeType {
+    pub(crate) fn native_type(&self) -> &JsonNativeType {
         match &self.format {
             JsonFormat::Jsonb => &JsonNativeType::String,
-            JsonFormat::Native(x) => x.as_ref(),
+            JsonFormat::Json2(x) => x.as_ref(),
         }
     }
 
     pub fn null() -> Self {
         Self {
-            format: JsonFormat::Native(Box::new(JsonNativeType::Null)),
+            format: JsonFormat::Json2(Box::new(JsonNativeType::Null)),
         }
     }
 
@@ -208,7 +304,7 @@ impl JsonType {
     pub(crate) fn as_struct_type(&self) -> StructType {
         match &self.format {
             JsonFormat::Jsonb => StructType::default(),
-            JsonFormat::Native(inner) => match ConcreteDataType::from(inner.as_ref()) {
+            JsonFormat::Json2(inner) => match ConcreteDataType::from(inner.as_ref()) {
                 ConcreteDataType::Struct(t) => t.clone(),
                 x => plain_json_struct_type(x),
             },
@@ -217,11 +313,14 @@ impl JsonType {
 
     /// Try to merge this json type with others, error on datatype conflict.
     pub fn merge(&mut self, other: &JsonType) -> Result<()> {
-        match (&self.format, &other.format) {
+        if self == other {
+            return Ok(());
+        }
+
+        match (&mut self.format, &other.format) {
             (JsonFormat::Jsonb, JsonFormat::Jsonb) => Ok(()),
-            (JsonFormat::Native(this), JsonFormat::Native(that)) => {
-                let merged = merge(this.as_ref(), that.as_ref())?;
-                self.format = JsonFormat::Native(Box::new(merged));
+            (JsonFormat::Json2(this), JsonFormat::Json2(that)) => {
+                this.merge(that);
                 Ok(())
             }
             _ => MergeJsonDatatypeSnafu {
@@ -231,22 +330,11 @@ impl JsonType {
         }
     }
 
-    /// Check if it can merge with `other` json type.
-    pub fn is_mergeable(&self, other: &JsonType) -> bool {
-        match (&self.format, &other.format) {
-            (JsonFormat::Jsonb, JsonFormat::Jsonb) => true,
-            (JsonFormat::Native(this), JsonFormat::Native(that)) => {
-                is_mergeable(this.as_ref(), that.as_ref())
-            }
-            _ => false,
-        }
-    }
-
     /// Check if it includes all fields in `other` json type.
     pub fn is_include(&self, other: &JsonType) -> bool {
         match (&self.format, &other.format) {
             (JsonFormat::Jsonb, JsonFormat::Jsonb) => true,
-            (JsonFormat::Native(this), JsonFormat::Native(that)) => {
+            (JsonFormat::Json2(this), JsonFormat::Json2(that)) => {
                 is_include(this.as_ref(), that.as_ref())
             }
             _ => false,
@@ -288,59 +376,9 @@ pub(crate) fn plain_json_struct_type(item_type: ConcreteDataType) -> StructType
     StructType::new(Arc::new(vec![field]))
 }
 
-fn is_mergeable(this: &JsonNativeType, that: &JsonNativeType) -> bool {
-    fn is_mergeable_object(this: &JsonObjectType, that: &JsonObjectType) -> bool {
-        for (type_name, that_type) in that {
-            if let Some(this_type) = this.get(type_name)
-                && !is_mergeable(this_type, that_type)
-            {
-                return false;
-            }
-        }
-        true
-    }
-
-    match (this, that) {
-        (this, that) if this == that => true,
-        (JsonNativeType::Array(this), JsonNativeType::Array(that)) => {
-            is_mergeable(this.as_ref(), that.as_ref())
-        }
-        (JsonNativeType::Object(this), JsonNativeType::Object(that)) => {
-            is_mergeable_object(this, that)
-        }
-        (JsonNativeType::Null, _) | (_, JsonNativeType::Null) => true,
-        _ => false,
-    }
-}
-
-fn merge(this: &JsonNativeType, that: &JsonNativeType) -> Result<JsonNativeType> {
-    fn merge_object(this: &JsonObjectType, that: &JsonObjectType) -> Result<JsonObjectType> {
-        let mut this = this.clone();
-        // merge "that" into "this" directly:
-        for (type_name, that_type) in that {
-            if let Some(this_type) = this.get_mut(type_name) {
-                let merged_type = merge(this_type, that_type)?;
-                *this_type = merged_type;
-            } else {
-                this.insert(type_name.clone(), that_type.clone());
-            }
-        }
-        Ok(this)
-    }
-
-    match (this, that) {
-        (this, that) if this == that => Ok(this.clone()),
-        (JsonNativeType::Array(this), JsonNativeType::Array(that)) => {
-            merge(this.as_ref(), that.as_ref()).map(|x| JsonNativeType::Array(Box::new(x)))
-        }
-        (JsonNativeType::Object(this), JsonNativeType::Object(that)) => {
-            merge_object(this, that).map(JsonNativeType::Object)
-        }
-        (JsonNativeType::Null, x) | (x, JsonNativeType::Null) => Ok(x.clone()),
-        _ => MergeJsonDatatypeSnafu {
-            reason: format!("datatypes have conflict, this: {this}, that: {that}"),
-        }
-        .fail(),
+impl From<&ArrowDataType> for JsonType {
+    fn from(t: &ArrowDataType) -> Self {
+        JsonType::new_json2(JsonNativeType::from(&ConcreteDataType::from_arrow_type(t)))
     }
 }
 
@@ -348,7 +386,14 @@ impl DataType for JsonType {
     fn name(&self) -> String {
         match &self.format {
             JsonFormat::Jsonb => JSON_TYPE_NAME.to_string(),
-            JsonFormat::Native(x) => format!("Json<{x}>"),
+            JsonFormat::Json2(x) => format!(
+                "{JSON2_TYPE_NAME}{}",
+                if x.is_null() {
+                    "".to_string()
+                } else {
+                    x.to_string()
+                }
+            ),
         }
     }
 
@@ -363,14 +408,14 @@ impl DataType for JsonType {
     fn as_arrow_type(&self) -> ArrowDataType {
         match self.format {
             JsonFormat::Jsonb => ArrowDataType::Binary,
-            JsonFormat::Native(_) => self.as_struct_type().as_arrow_type(),
+            JsonFormat::Json2(_) => self.as_struct_type().as_arrow_type(),
         }
     }
 
     fn create_mutable_vector(&self, capacity: usize) -> Box<dyn MutableVector> {
         match &self.format {
             JsonFormat::Jsonb => Box::new(BinaryVectorBuilder::with_capacity(capacity)),
-            JsonFormat::Native(x) => Box::new(JsonVectorBuilder::new(*x.clone(), capacity)),
+            JsonFormat::Json2(x) => Box::new(JsonVectorBuilder::new(*x.clone(), capacity)),
         }
     }
 
@@ -396,7 +441,7 @@ pub fn jsonb_to_string(val: &[u8]) -> Result<String> {
     match jsonb::from_slice(val) {
         Ok(jsonb_value) => {
             let serialized = jsonb_value.to_string();
-            Ok(serialized)
+            fix_unicode_point(&serialized)
         }
         Err(e) => InvalidJsonbSnafu { error: e }.fail(),
     }
@@ -405,18 +450,12 @@ pub fn jsonb_to_string(val: &[u8]) -> Result<String> {
 /// Converts a json type value to serde_json::Value
 pub fn jsonb_to_serde_json(val: &[u8]) -> Result<serde_json::Value> {
     let json_string = jsonb_to_string(val)?;
-    jsonb_string_to_serde_value(&json_string)
+    serde_json::Value::from_str(&json_string).context(DeserializeSnafu { json: json_string })
 }
 
-/// Attempts to deserialize a JSON text into `serde_json::Value`, with a best-effort
-/// fallback for Rust-style Unicode escape sequences.
+/// Normalizes a JSON string by converting Rust-style Unicode escape sequences to JSON-compatible format.
 ///
-/// This function is intended to be used on JSON strings produced from the internal
-/// JSONB representation (e.g. via [`jsonb_to_string`]). It first calls
-/// `serde_json::Value::from_str` directly. If that succeeds, the parsed value is
-/// returned as-is.
-///
-/// If the initial parse fails, the input is scanned for Rust-style Unicode code
+/// The input is scanned for Rust-style Unicode code
 /// point escapes of the form `\\u{H...}` (a backslash, `u`, an opening brace,
 /// followed by 1–6 hexadecimal digits, and a closing brace). Each such escape is
 /// converted into JSON-compatible UTF‑16 escape sequences:
@@ -427,59 +466,44 @@ pub fn jsonb_to_serde_json(val: &[u8]) -> Result<serde_json::Value> {
 ///   the code point is encoded as a UTF‑16 surrogate pair and emitted as two consecutive
 ///   `\\uXXXX` sequences (as JSON format required).
 ///
-/// After this normalization, the function retries parsing the resulting string as
-/// JSON and returns the deserialized value or a `DeserializeSnafu` error if it
-/// still cannot be parsed.
-fn jsonb_string_to_serde_value(json: &str) -> Result<serde_json::Value> {
-    match serde_json::Value::from_str(json) {
-        Ok(v) => Ok(v),
-        Err(e) => {
-            // If above deserialization is failed, the JSON string might contain some Rust chars
-            // that are somehow incorrectly represented as Unicode code point literal. For example,
-            // "\u{fe0f}". We have to convert them to JSON compatible format, like "\uFE0F", then
-            // try to deserialize the JSON string again.
-            if !e.is_syntax() || !e.to_string().contains("invalid escape") {
-                return Err(e).context(DeserializeSnafu { json });
-            }
+/// After this normalization, the function returns the normalized string
+fn fix_unicode_point(json: &str) -> Result<String> {
+    static UNICODE_CODE_POINT_PATTERN: LazyLock<Regex> = LazyLock::new(|| {
+        // Match literal "\u{...}" sequences, capturing 1–6 (code point range) hex digits
+        // inside braces.
+        Regex::new(r"\\u\{([0-9a-fA-F]{1,6})}").unwrap_or_else(|e| panic!("{}", e))
+    });
 
-            static UNICODE_CODE_POINT_PATTERN: LazyLock<Regex> = LazyLock::new(|| {
-                // Match literal "\u{...}" sequences, capturing 1–6 (code point range) hex digits
-                // inside braces.
-                Regex::new(r"\\u\{([0-9a-fA-F]{1,6})}").unwrap_or_else(|e| panic!("{}", e))
-            });
+    let v = UNICODE_CODE_POINT_PATTERN.replace_all(json, |caps: &Captures| {
+        // Extract the hex payload (without braces) and parse to a code point.
+        let hex = &caps[1];
+        let Ok(code) = u32::from_str_radix(hex, 16) else {
+            // On parse failure, leave the original escape sequence unchanged.
+            return caps[0].to_string();
+        };
 
-            let v = UNICODE_CODE_POINT_PATTERN.replace_all(json, |caps: &Captures| {
-                // Extract the hex payload (without braces) and parse to a code point.
-                let hex = &caps[1];
-                let Ok(code) = u32::from_str_radix(hex, 16) else {
-                    // On parse failure, leave the original escape sequence unchanged.
-                    return caps[0].to_string();
-                };
+        if code <= 0xFFFF {
+            // Basic Multilingual Plane: JSON can represent this directly as \uXXXX.
+            format!("\\u{:04X}", code)
+        } else if code > 0x10FFFF {
+            // Beyond max Unicode code point
+            caps[0].to_string()
+        } else {
+            // Supplementary planes: JSON needs UTF-16 surrogate pairs.
+            // Convert the code point to a 20-bit value.
+            let code = code - 0x10000;
 
-                if code <= 0xFFFF {
-                    // Basic Multilingual Plane: JSON can represent this directly as \uXXXX.
-                    format!("\\u{:04X}", code)
-                } else if code > 0x10FFFF {
-                    // Beyond max Unicode code point
-                    caps[0].to_string()
-                } else {
-                    // Supplementary planes: JSON needs UTF-16 surrogate pairs.
-                    // Convert the code point to a 20-bit value.
-                    let code = code - 0x10000;
+            // High surrogate: top 10 bits, offset by 0xD800.
+            let high = 0xD800 + ((code >> 10) & 0x3FF);
 
-                    // High surrogate: top 10 bits, offset by 0xD800.
-                    let high = 0xD800 + ((code >> 10) & 0x3FF);
+            // Low surrogate: bottom 10 bits, offset by 0xDC00.
+            let low = 0xDC00 + (code & 0x3FF);
 
-                    // Low surrogate: bottom 10 bits, offset by 0xDC00.
-                    let low = 0xDC00 + (code & 0x3FF);
-
-                    // Emit two \uXXXX escapes in sequence.
-                    format!("\\u{:04X}\\u{:04X}", high, low)
-                }
-            });
-            serde_json::Value::from_str(&v).context(DeserializeSnafu { json })
+            // Emit two \uXXXX escapes in sequence.
+            format!("\\u{:04X}\\u{:04X}", high, low)
         }
-    }
+    });
+    Ok(v.to_string())
 }
 
 /// Parses a string to a json type value
@@ -495,45 +519,54 @@ mod tests {
     use crate::json::JsonStructureSettings;
 
     #[test]
-    fn test_jsonb_string_to_serde_value() -> Result<()> {
+    fn test_fix_unicode_point() -> Result<()> {
         let valid_cases = vec![
-            (r#"{"data": "simple ascii"}"#, r#"{"data":"simple ascii"}"#),
+            (r#"{"data": "simple ascii"}"#, r#"{"data": "simple ascii"}"#),
             (
-                r#"{"data": "Greek sigma: \u{03a3}"}"#,
-                r#"{"data":"Greek sigma: Σ"}"#,
+                r#"{"data":"Greek sigma: \u{03a3}"}"#,
+                r#"{"data":"Greek sigma: \u03A3"}"#,
             ),
             (
-                r#"{"data": "Joker card: \u{1f0df}"}"#,
-                r#"{"data":"Joker card: 🃟"}"#,
+                r#"{"data":"Joker card: \u{1f0df}"}"#,
+                r#"{"data":"Joker card: \uD83C\uDCDF"}"#,
             ),
             (
-                r#"{"data": "BMP boundary: \u{ffff}"}"#,
-                r#"{"data":"BMP boundary: ￿"}"#,
+                r#"{"data":"BMP boundary: \u{ffff}"}"#,
+                r#"{"data":"BMP boundary: \uFFFF"}"#,
             ),
             (
-                r#"{"data": "Supplementary min: \u{10000}"}"#,
-                r#"{"data":"Supplementary min: 𐀀"}"#,
+                r#"{"data":"Supplementary min: \u{10000}"}"#,
+                r#"{"data":"Supplementary min: \uD800\uDC00"}"#,
             ),
             (
-                r#"{"data": "Supplementary max: \u{10ffff}"}"#,
-                r#"{"data":"Supplementary max: 􏿿"}"#,
+                r#"{"data":"Supplementary max: \u{10ffff}"}"#,
+                r#"{"data":"Supplementary max: \uDBFF\uDFFF"}"#,
             ),
         ];
         for (input, expect) in valid_cases {
-            let v = jsonb_string_to_serde_value(input)?;
-            assert_eq!(v.to_string(), expect);
+            let v = fix_unicode_point(input)?;
+            assert_eq!(v, expect);
         }
 
-        let invalid_cases = vec![
-            r#"{"data": "Invalid hex: \u{gggg}"}"#,
-            r#"{"data": "Beyond max Unicode code point: \u{110000}"}"#,
-            r#"{"data": "Out of range: \u{1100000}"}"#, // 7 digit
-            r#"{"data": "Empty braces: \u{}"}"#,
+        let invalid_escape_cases = vec![
+            (
+                r#"{"data": "Invalid hex: \u{gggg}"}"#,
+                r#"{"data": "Invalid hex: \u{gggg}"}"#,
+            ),
+            (
+                r#"{"data": "Empty braces: \u{}"}"#,
+                r#"{"data": "Empty braces: \u{}"}"#,
+            ),
+            (
+                r#"{"data": "Out of range: \u{1100000}"}"#,
+                r#"{"data": "Out of range: \u{1100000}"}"#,
+            ),
         ];
-        for input in invalid_cases {
-            let result = jsonb_string_to_serde_value(input);
-            assert!(result.is_err());
+        for (input, expect) in invalid_escape_cases {
+            let v = fix_unicode_point(input)?;
+            assert_eq!(v, expect);
         }
+
         Ok(())
     }
 
@@ -754,68 +787,93 @@ mod tests {
             let result = json_type.merge(other);
             match (result, expected) {
                 (Ok(()), Ok(expected)) => {
-                    assert_eq!(json_type.name(), expected);
-                    assert!(json_type.is_mergeable(other));
+                    assert_eq!(json_type.native_type().to_string(), expected);
                 }
                 (Err(err), Err(expected)) => {
                     assert_eq!(err.to_string(), expected);
-                    assert!(!json_type.is_mergeable(other));
                 }
                 _ => unreachable!(),
             }
             Ok(())
         }
 
-        let json_type = &mut JsonType::new_native(JsonNativeType::Null);
+        // Null should be absorbed by a concrete scalar type.
+        test("true", &mut JsonType::null(), Ok(r#""<Bool>""#))?;
 
-        // can merge with json object:
-        let json = r#"{
-            "hello": "world",
-            "list": [1, 2, 3],
-            "object": {"a": 1}
-        }"#;
-        let expected =
-            r#"Json<{"hello":"<String>","list":["<Number>"],"object":{"a":"<Number>"}}>"#;
-        test(json, json_type, Ok(expected))?;
+        // Merging a null value into an existing concrete type should keep the type unchanged.
+        test(
+            "null",
+            &mut JsonType::new_json2(JsonNativeType::Bool),
+            Ok(r#""<Bool>""#),
+        )?;
 
-        // cannot merge with other non-object json values:
-        let jsons = [r#""s""#, "1", "[1]"];
-        let expects = [
-            r#"Failed to merge JSON datatype: datatypes have conflict, this: {"hello":"<String>","list":["<Number>"],"object":{"a":"<Number>"}}, that: "<String>""#,
-            r#"Failed to merge JSON datatype: datatypes have conflict, this: {"hello":"<String>","list":["<Number>"],"object":{"a":"<Number>"}}, that: "<Number>""#,
-            r#"Failed to merge JSON datatype: datatypes have conflict, this: {"hello":"<String>","list":["<Number>"],"object":{"a":"<Number>"}}, that: ["<Number>"]"#,
-        ];
-        for (json, expect) in jsons.into_iter().zip(expects.into_iter()) {
-            test(json, json_type, Err(expect))?;
-        }
+        // Identical number categories should stay as Number.
+        test(
+            "1",
+            &mut JsonType::new_json2(JsonNativeType::i64()),
+            Ok(r#""<Number>""#),
+        )?;
 
-        // cannot merge with other json object with conflict field datatype:
-        let json = r#"{
-            "hello": 1,
-            "float": 0.123,
-            "no": 42
-        }"#;
-        let expected = r#"Failed to merge JSON datatype: datatypes have conflict, this: "<String>", that: "<Number>""#;
-        test(json, json_type, Err(expected))?;
+        // Conflicting number categories should be lifted to Variant.
+        test(
+            "1.5",
+            &mut JsonType::new_json2(JsonNativeType::i64()),
+            Ok(r#""<Variant>""#),
+        )?;
 
-        // can merge with another json object:
-        let json = r#"{
-            "hello": "greptime",
-            "float": 0.123,
-            "int": 42
-        }"#;
-        let expected = r#"Json<{"float":"<Number>","hello":"<String>","int":"<Number>","list":["<Number>"],"object":{"a":"<Number>"}}>"#;
-        test(json, json_type, Ok(expected))?;
+        // Object merge should preserve existing fields and append missing fields.
+        test(
+            r#"{"foo":"x"}"#,
+            &mut JsonType::new_json2(JsonNativeType::Object(JsonObjectType::from([(
+                "bar".to_string(),
+                JsonNativeType::i64(),
+            )]))),
+            Ok(r#"{"bar":"<Number>","foo":"<String>"}"#),
+        )?;
 
-        // can merge with some complex nested json object:
-        let json = r#"{
-            "list": [4],
-            "object": {"foo": "bar", "l": ["x"], "o": {"key": "value"}},
-            "float": 0.456,
-            "int": 0
-        }"#;
-        let expected = r#"Json<{"float":"<Number>","hello":"<String>","int":"<Number>","list":["<Number>"],"object":{"a":"<Number>","foo":"<String>","l":["<String>"],"o":{"key":"<String>"}}}>"#;
-        test(json, json_type, Ok(expected))?;
+        // Conflicting object field types should only lift that field to Variant.
+        test(
+            r#"{"foo":1}"#,
+            &mut JsonType::new_json2(JsonNativeType::Object(JsonObjectType::from([(
+                "foo".to_string(),
+                JsonNativeType::Bool,
+            )]))),
+            Ok(r#"{"foo":"<Variant>"}"#),
+        )?;
+
+        // Nested objects should merge recursively.
+        test(
+            r#"{"nested":{"foo":"bar"}}"#,
+            &mut JsonType::new_json2(JsonNativeType::Object(JsonObjectType::from([(
+                "nested".to_string(),
+                JsonNativeType::Object(JsonObjectType::from([(
+                    "bar".to_string(),
+                    JsonNativeType::Bool,
+                )])),
+            )]))),
+            Ok(r#"{"nested":{"bar":"<Bool>","foo":"<String>"}}"#),
+        )?;
+
+        // Arrays should merge their element types recursively.
+        test(
+            r#"["foo"]"#,
+            &mut JsonType::new_json2(JsonNativeType::Array(Box::new(JsonNativeType::u64()))),
+            Ok(r#"["<Variant>"]"#),
+        )?;
+
+        // Root-level incompatible types should be lifted to Variant.
+        test(
+            r#"{"foo":"bar"}"#,
+            &mut JsonType::new_json2(JsonNativeType::Bool),
+            Ok(r#""<Variant>""#),
+        )?;
+
+        // Jsonb and Json2 should not be mergeable.
+        test(
+            "true",
+            &mut JsonType::new(JsonFormat::Jsonb),
+            Err("Failed to merge JSON datatype: json format not match"),
+        )?;
 
         Ok(())
     }
diff --git a/src/datatypes/src/value.rs b/src/datatypes/src/value.rs
index ab64c801e5..13193fe0b0 100644
--- a/src/datatypes/src/value.rs
+++ b/src/datatypes/src/value.rs
@@ -922,7 +922,7 @@ impl TryFrom<Value> for serde_json::Value {
                 let map = struct_type
                     .fields()
                     .iter()
-                    .zip(items.into_iter())
+                    .zip(items)
                     .map(|(field, value)| {
                         Ok((
                             field.name().to_string(),
@@ -932,7 +932,7 @@ impl TryFrom<Value> for serde_json::Value {
                     .collect::<serde_json::Result<Map<String, serde_json::Value>>>()?;
                 serde_json::Value::Object(map)
             }
-            Value::Json(v) => (*v).into(),
+            Value::Json(v) => (*v).try_into()?,
         };
 
         Ok(json_value)
@@ -2723,26 +2723,26 @@ pub(crate) mod tests {
                 .unwrap()
         );
         assert_eq!(
-            ScalarValue::UInt8(Some(u8::MIN + 1)),
-            Value::UInt8(u8::MIN + 1)
+            ScalarValue::UInt8(Some(1)),
+            Value::UInt8(1)
                 .try_to_scalar_value(&ConcreteDataType::uint8_datatype())
                 .unwrap()
         );
         assert_eq!(
-            ScalarValue::UInt16(Some(u16::MIN + 2)),
-            Value::UInt16(u16::MIN + 2)
+            ScalarValue::UInt16(Some(2)),
+            Value::UInt16(2)
                 .try_to_scalar_value(&ConcreteDataType::uint16_datatype())
                 .unwrap()
         );
         assert_eq!(
-            ScalarValue::UInt32(Some(u32::MIN + 3)),
-            Value::UInt32(u32::MIN + 3)
+            ScalarValue::UInt32(Some(3)),
+            Value::UInt32(3)
                 .try_to_scalar_value(&ConcreteDataType::uint32_datatype())
                 .unwrap()
         );
         assert_eq!(
-            ScalarValue::UInt64(Some(u64::MIN + 4)),
-            Value::UInt64(u64::MIN + 4)
+            ScalarValue::UInt64(Some(4)),
+            Value::UInt64(4)
                 .try_to_scalar_value(&ConcreteDataType::uint64_datatype())
                 .unwrap()
         );
diff --git a/src/datatypes/src/vectors.rs b/src/datatypes/src/vectors.rs
index 5355b35ff4..5f116e0952 100644
--- a/src/datatypes/src/vectors.rs
+++ b/src/datatypes/src/vectors.rs
@@ -35,7 +35,7 @@ mod duration;
 mod eq;
 mod helper;
 mod interval;
-pub(crate) mod json;
+pub mod json;
 mod list;
 mod null;
 pub(crate) mod operations;
@@ -307,10 +307,11 @@ macro_rules! impl_extend_for_builder {
     }};
 }
 
-pub(crate) use {
-    impl_extend_for_builder, impl_get_for_vector, impl_get_ref_for_vector,
-    impl_try_from_arrow_array_for_vector, impl_validity_for_vector,
-};
+pub(crate) use impl_extend_for_builder;
+pub(crate) use impl_get_for_vector;
+pub(crate) use impl_get_ref_for_vector;
+pub(crate) use impl_try_from_arrow_array_for_vector;
+pub(crate) use impl_validity_for_vector;
 
 #[cfg(test)]
 pub mod tests {
diff --git a/src/datatypes/src/vectors/binary.rs b/src/datatypes/src/vectors/binary.rs
index 21d188199c..6d9954fb80 100644
--- a/src/datatypes/src/vectors/binary.rs
+++ b/src/datatypes/src/vectors/binary.rs
@@ -458,7 +458,7 @@ impl BinaryVector {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use arrow::datatypes::DataType as ArrowDataType;
     use common_base::bytes::Bytes;
diff --git a/src/datatypes/src/vectors/eq.rs b/src/datatypes/src/vectors/eq.rs
index 372b3ceee4..e8e942d3ff 100644
--- a/src/datatypes/src/vectors/eq.rs
+++ b/src/datatypes/src/vectors/eq.rs
@@ -67,7 +67,7 @@ fn equal(lhs: &dyn Vector, rhs: &dyn Vector) -> bool {
                 .downcast_ref::<ConstantVector>()
                 .unwrap()
                 .inner(),
-            &**lhs
+            &**rhs
                 .as_any()
                 .downcast_ref::<ConstantVector>()
                 .unwrap()
@@ -259,6 +259,22 @@ mod tests {
         ])));
     }
 
+    // Regression: second arm must downcast `rhs` (was `lhs`), or same-length ConstantVectors
+    // with different inners compare equal.
+    #[test]
+    fn test_constant_vector_eq_compares_both_inners() {
+        assert_vector_ref_ne(
+            Arc::new(ConstantVector::new(
+                Arc::new(BooleanVector::from(vec![true])),
+                5,
+            )),
+            Arc::new(ConstantVector::new(
+                Arc::new(BooleanVector::from(vec![false])),
+                5,
+            )),
+        );
+    }
+
     #[test]
     fn test_vector_ne() {
         assert_vector_ref_ne(
diff --git a/src/datatypes/src/vectors/helper.rs b/src/datatypes/src/vectors/helper.rs
index 4907bbf380..21bb371617 100644
--- a/src/datatypes/src/vectors/helper.rs
+++ b/src/datatypes/src/vectors/helper.rs
@@ -462,14 +462,6 @@ impl Helper {
     }
 }
 
-#[cfg(test)]
-pub(crate) fn pretty_print(vector: VectorRef) -> String {
-    let array = vector.to_arrow_array();
-    arrow::util::pretty::pretty_format_columns(&vector.vector_type_name(), &[array])
-        .map(|x| x.to_string())
-        .unwrap_or_else(|e| e.to_string())
-}
-
 #[cfg(test)]
 mod tests {
     use arrow::array::{
diff --git a/src/datatypes/src/vectors/json.rs b/src/datatypes/src/vectors/json.rs
index 83aa1dd2aa..3dcb8d7bc0 100644
--- a/src/datatypes/src/vectors/json.rs
+++ b/src/datatypes/src/vectors/json.rs
@@ -12,4 +12,5 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+pub mod array;
 pub(crate) mod builder;
diff --git a/src/datatypes/src/vectors/json/array.rs b/src/datatypes/src/vectors/json/array.rs
new file mode 100644
index 0000000000..75779821c5
--- /dev/null
+++ b/src/datatypes/src/vectors/json/array.rs
@@ -0,0 +1,462 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::cmp::Ordering;
+use std::sync::Arc;
+
+use arrow::compute;
+use arrow::util::display::{ArrayFormatter, FormatOptions};
+use arrow_array::cast::AsArray;
+use arrow_array::types::{Float64Type, Int64Type, UInt64Type};
+use arrow_array::{Array, ArrayRef, GenericListArray, ListArray, StructArray, new_null_array};
+use arrow_schema::{DataType, FieldRef};
+use serde_json::Value;
+use snafu::{OptionExt, ResultExt};
+
+use crate::arrow_array::{StringArray, binary_array_value, string_array_value};
+use crate::error::{
+    AlignJsonArraySnafu, ArrowComputeSnafu, DeserializeSnafu, InvalidJsonSnafu, Result,
+};
+
+pub struct JsonArray<'a> {
+    inner: &'a ArrayRef,
+}
+
+impl JsonArray<'_> {
+    /// Try to get the value (as a [Value]) at the index `i`.
+    pub fn try_get_value(&self, i: usize) -> Result<Value> {
+        let array = self.inner;
+        if array.is_null(i) {
+            return Ok(Value::Null);
+        }
+
+        let value = match array.data_type() {
+            DataType::Null => Value::Null,
+            DataType::Boolean => Value::Bool(array.as_boolean().value(i)),
+            DataType::Int64 => Value::from(array.as_primitive::<Int64Type>().value(i)),
+            DataType::UInt64 => Value::from(array.as_primitive::<UInt64Type>().value(i)),
+            DataType::Float64 => Value::from(array.as_primitive::<Float64Type>().value(i)),
+            DataType::Utf8 | DataType::LargeUtf8 | DataType::Utf8View => {
+                Value::String(string_array_value(array, i).to_string())
+            }
+            DataType::Binary | DataType::LargeBinary | DataType::BinaryView => {
+                let bytes = binary_array_value(array, i);
+                serde_json::from_slice(bytes).with_context(|_| DeserializeSnafu {
+                    json: String::from_utf8_lossy(bytes),
+                })?
+            }
+            DataType::Struct(_) => {
+                let structs = array.as_struct();
+                let object = structs
+                    .fields()
+                    .iter()
+                    .zip(structs.columns())
+                    .map(|(field, column)| {
+                        JsonArray::from(column)
+                            .try_get_value(i)
+                            .map(|v| (field.name().clone(), v))
+                    })
+                    .collect::<Result<_>>()?;
+                Value::Object(object)
+            }
+            DataType::List(_) => {
+                let lists = array.as_list::<i32>();
+                let list = lists.value(i);
+                let list = JsonArray::from(&list);
+                let mut values = Vec::with_capacity(list.inner.len());
+                for i in 0..list.inner.len() {
+                    values.push(list.try_get_value(i)?);
+                }
+                Value::Array(values)
+            }
+            t => {
+                return InvalidJsonSnafu {
+                    value: format!("unknown JSON type {t}"),
+                }
+                .fail();
+            }
+        };
+        Ok(value)
+    }
+
+    /// Align a JSON array to the `expect` data type. The alignment mostly does three things:
+    ///
+    /// 1. set the missing fields with null arrays;
+    /// 2. discard the fields that are not in the `expect` data type;
+    /// 3. cast the fields to the ones with same names in the `expect` if their data types are not
+    ///    matched.
+    pub fn try_align(&self, expect: &DataType) -> Result<ArrayRef> {
+        if self.inner.data_type() == expect {
+            return Ok(self.inner.clone());
+        }
+
+        let struct_array = self.inner.as_struct_opt().context(AlignJsonArraySnafu {
+            reason: "expect struct array",
+        })?;
+        let array_fields = struct_array.fields();
+        let array_columns = struct_array.columns();
+        let DataType::Struct(expect_fields) = expect else {
+            return AlignJsonArraySnafu {
+                reason: "expect struct datatype",
+            }
+            .fail();
+        };
+        let mut aligned = Vec::with_capacity(expect_fields.len());
+
+        // Compare the fields in the JSON array and the to-be-aligned schema, amending with null
+        // arrays on the way. It's very important to note that fields in the JSON array and those
+        // in the JSON type are both **SORTED**, which can be guaranteed because the fields in the
+        // JSON type implementation are sorted.
+        debug_assert!(expect_fields.iter().map(|f| f.name()).is_sorted());
+        debug_assert!(array_fields.iter().map(|f| f.name()).is_sorted());
+
+        let mut i = 0; // point to the expect fields
+        let mut j = 0; // point to the array fields
+        while i < expect_fields.len() && j < array_fields.len() {
+            let expect_field = &expect_fields[i];
+            let array_field = &array_fields[j];
+            match expect_field.name().cmp(array_field.name()) {
+                Ordering::Equal => {
+                    if expect_field.data_type() == array_field.data_type() {
+                        aligned.push(array_columns[j].clone());
+                    } else {
+                        let expect_type = expect_field.data_type();
+                        let array_type = array_field.data_type();
+                        let array = match (expect_type, array_type) {
+                            (DataType::Struct(_), DataType::Struct(_)) => {
+                                JsonArray::from(&array_columns[j]).try_align(expect_type)?
+                            }
+                            (DataType::List(expect_item), DataType::List(array_item)) => {
+                                let list_array = array_columns[j].as_list::<i32>();
+                                try_align_list(list_array, expect_item, array_item)?
+                            }
+                            _ => JsonArray::from(&array_columns[j]).try_cast(expect_type)?,
+                        };
+                        aligned.push(array);
+                    }
+                    i += 1;
+                    j += 1;
+                }
+                Ordering::Less => {
+                    aligned.push(new_null_array(expect_field.data_type(), struct_array.len()));
+                    i += 1;
+                }
+                Ordering::Greater => {
+                    j += 1;
+                }
+            }
+        }
+        if i < expect_fields.len() {
+            for field in &expect_fields[i..] {
+                aligned.push(new_null_array(field.data_type(), struct_array.len()));
+            }
+        }
+
+        let json_array = StructArray::try_new(
+            expect_fields.clone(),
+            aligned,
+            struct_array.nulls().cloned(),
+        )
+        .map_err(|e| {
+            AlignJsonArraySnafu {
+                reason: e.to_string(),
+            }
+            .build()
+        })?;
+        Ok(Arc::new(json_array))
+    }
+
+    fn try_cast(&self, to_type: &DataType) -> Result<ArrayRef> {
+        if compute::can_cast_types(self.inner.data_type(), to_type) {
+            return compute::cast(&self.inner, to_type).context(ArrowComputeSnafu);
+        }
+
+        // TODO(LFC): Cast according to `to_type` instead of formatting to String here.
+        let formatter = ArrayFormatter::try_new(&self.inner, &FormatOptions::default())
+            .context(ArrowComputeSnafu)?;
+        let values = (0..self.inner.len())
+            .map(|i| {
+                self.inner
+                    .is_valid(i)
+                    .then(|| formatter.value(i).to_string())
+            })
+            .collect::<Vec<_>>();
+        Ok(Arc::new(StringArray::from(values)))
+    }
+}
+
+fn try_align_list(
+    list_array: &ListArray,
+    expect_item: &FieldRef,
+    array_item: &FieldRef,
+) -> Result<ArrayRef> {
+    let item_aligned = match (expect_item.data_type(), array_item.data_type()) {
+        (DataType::Struct(_), DataType::Struct(_)) => {
+            JsonArray::from(list_array.values()).try_align(expect_item.data_type())?
+        }
+        (DataType::List(expect_item), DataType::List(array_item)) => {
+            let list_array = list_array.values().as_list::<i32>();
+            try_align_list(list_array, expect_item, array_item)?
+        }
+        _ => JsonArray::from(list_array.values()).try_cast(expect_item.data_type())?,
+    };
+    Ok(Arc::new(
+        GenericListArray::<i32>::try_new(
+            expect_item.clone(),
+            list_array.offsets().clone(),
+            item_aligned,
+            list_array.nulls().cloned(),
+        )
+        .context(ArrowComputeSnafu)?,
+    ))
+}
+
+impl<'a> From<&'a ArrayRef> for JsonArray<'a> {
+    fn from(inner: &'a ArrayRef) -> Self {
+        Self { inner }
+    }
+}
+
+#[cfg(test)]
+mod test {
+    use arrow_array::types::Int64Type;
+    use arrow_array::{BinaryArray, BooleanArray, Float64Array, Int32Array, Int64Array, ListArray};
+    use arrow_schema::{Field, Fields};
+    use serde_json::json;
+
+    use super::*;
+
+    #[test]
+    fn test_try_get_value() -> Result<()> {
+        let nulls = new_null_array(&DataType::Null, 2);
+        assert_eq!(JsonArray::from(&nulls).try_get_value(0)?, Value::Null);
+
+        let bools: ArrayRef = Arc::new(BooleanArray::from(vec![Some(true), None]));
+        assert_eq!(JsonArray::from(&bools).try_get_value(0)?, json!(true));
+        assert_eq!(JsonArray::from(&bools).try_get_value(1)?, Value::Null);
+
+        let ints: ArrayRef = Arc::new(Int64Array::from(vec![Some(-7), None]));
+        assert_eq!(JsonArray::from(&ints).try_get_value(0)?, json!(-7));
+        assert_eq!(JsonArray::from(&ints).try_get_value(1)?, Value::Null);
+
+        let floats: ArrayRef = Arc::new(Float64Array::from(vec![Some(1.5)]));
+        assert_eq!(JsonArray::from(&floats).try_get_value(0)?, json!(1.5));
+
+        let strings: ArrayRef = Arc::new(StringArray::from(vec![Some("hello"), None]));
+        assert_eq!(JsonArray::from(&strings).try_get_value(0)?, json!("hello"));
+        assert_eq!(JsonArray::from(&strings).try_get_value(1)?, Value::Null);
+
+        let binaries: ArrayRef = Arc::new(BinaryArray::from(vec![
+            br#"{"nested":[1,null,"x"]}"#.as_slice(),
+            b"null".as_slice(),
+        ]));
+        assert_eq!(
+            JsonArray::from(&binaries).try_get_value(0)?,
+            json!({"nested": [1, null, "x"]})
+        );
+        assert_eq!(JsonArray::from(&binaries).try_get_value(1)?, Value::Null);
+
+        let lists: ArrayRef = Arc::new(ListArray::from_iter_primitive::<Int64Type, _, _>(vec![
+            Some(vec![Some(1), None, Some(3)]),
+            None,
+        ]));
+        assert_eq!(
+            JsonArray::from(&lists).try_get_value(0)?,
+            json!([1, null, 3])
+        );
+        assert_eq!(JsonArray::from(&lists).try_get_value(1)?, Value::Null);
+
+        let structs: ArrayRef = Arc::new(StructArray::from(vec![
+            (
+                Arc::new(Field::new("flag", DataType::Boolean, true)),
+                Arc::new(BooleanArray::from(vec![Some(true), None])) as ArrayRef,
+            ),
+            (
+                Arc::new(Field::new_list(
+                    "items",
+                    Field::new_list_field(DataType::Int64, true),
+                    true,
+                )),
+                Arc::new(ListArray::from_iter_primitive::<Int64Type, _, _>(vec![
+                    Some(vec![Some(1), None]),
+                    Some(vec![Some(2)]),
+                ])) as ArrayRef,
+            ),
+        ]));
+        assert_eq!(
+            JsonArray::from(&structs).try_get_value(0)?,
+            json!({"flag": true, "items": [1, null]})
+        );
+        assert_eq!(
+            JsonArray::from(&structs).try_get_value(1)?,
+            json!({"flag": null, "items": [2]})
+        );
+
+        let unsupported: ArrayRef = Arc::new(Int32Array::from(vec![1]));
+        assert_eq!(
+            JsonArray::from(&unsupported)
+                .try_get_value(0)
+                .unwrap_err()
+                .to_string(),
+            "Invalid JSON: unknown JSON type Int32"
+        );
+
+        Ok(())
+    }
+
+    #[test]
+    fn test_align_json_array() -> Result<()> {
+        struct TestCase {
+            json_array: ArrayRef,
+            schema_type: DataType,
+            expected: std::result::Result<ArrayRef, String>,
+        }
+
+        impl TestCase {
+            fn new(
+                json_array: StructArray,
+                schema_type: Fields,
+                expected: std::result::Result<Vec<ArrayRef>, String>,
+            ) -> Self {
+                Self {
+                    json_array: Arc::new(json_array),
+                    schema_type: DataType::Struct(schema_type.clone()),
+                    expected: expected
+                        .map(|x| Arc::new(StructArray::new(schema_type, x, None)) as ArrayRef),
+                }
+            }
+
+            fn test(self) -> Result<()> {
+                let result = JsonArray::from(&self.json_array).try_align(&self.schema_type);
+                match (result, self.expected) {
+                    (Ok(json_array), Ok(expected)) => assert_eq!(&json_array, &expected),
+                    (Ok(json_array), Err(e)) => {
+                        panic!("expecting error {e} but actually get: {json_array:?}")
+                    }
+                    (Err(e), Err(expected)) => assert_eq!(e.to_string(), expected),
+                    (Err(e), Ok(_)) => return Err(e),
+                }
+                Ok(())
+            }
+        }
+
+        // Test empty json array can be aligned with a complex json type.
+        TestCase::new(
+            StructArray::new_empty_fields(2, None),
+            Fields::from(vec![
+                Field::new("int", DataType::Int64, true),
+                Field::new_struct(
+                    "nested",
+                    vec![Field::new("bool", DataType::Boolean, true)],
+                    true,
+                ),
+                Field::new("string", DataType::Utf8, true),
+            ]),
+            Ok(vec![
+                Arc::new(Int64Array::new_null(2)) as ArrayRef,
+                Arc::new(StructArray::new_null(
+                    Fields::from(vec![Arc::new(Field::new("bool", DataType::Boolean, true))]),
+                    2,
+                )),
+                Arc::new(StringArray::new_null(2)),
+            ]),
+        )
+        .test()?;
+
+        // Test simple json array alignment.
+        TestCase::new(
+            StructArray::from(vec![(
+                Arc::new(Field::new("float", DataType::Float64, true)),
+                Arc::new(Float64Array::from(vec![1.0, 2.0, 3.0])) as ArrayRef,
+            )]),
+            Fields::from(vec![
+                Field::new("float", DataType::Float64, true),
+                Field::new("string", DataType::Utf8, true),
+            ]),
+            Ok(vec![
+                Arc::new(Float64Array::from(vec![1.0, 2.0, 3.0])) as ArrayRef,
+                Arc::new(StringArray::new_null(3)),
+            ]),
+        )
+        .test()?;
+
+        // Test complex json array alignment.
+        TestCase::new(
+            StructArray::from(vec![
+                (
+                    Arc::new(Field::new_list(
+                        "list",
+                        Field::new_list_field(DataType::Int64, true),
+                        true,
+                    )),
+                    Arc::new(ListArray::from_iter_primitive::<Int64Type, _, _>(vec![
+                        Some(vec![Some(1)]),
+                        None,
+                        Some(vec![Some(2), Some(3)]),
+                    ])) as ArrayRef,
+                ),
+                (
+                    Arc::new(Field::new_struct(
+                        "nested",
+                        vec![Field::new("int", DataType::Int64, true)],
+                        true,
+                    )),
+                    Arc::new(StructArray::from(vec![(
+                        Arc::new(Field::new("int", DataType::Int64, true)),
+                        Arc::new(Int64Array::from(vec![-1, -2, -3])) as ArrayRef,
+                    )])),
+                ),
+                (
+                    Arc::new(Field::new("string", DataType::Utf8, true)),
+                    Arc::new(StringArray::from(vec!["a", "b", "c"])),
+                ),
+            ]),
+            Fields::from(vec![
+                Field::new("bool", DataType::Boolean, true),
+                Field::new_list("list", Field::new_list_field(DataType::Int64, true), true),
+                Field::new_struct(
+                    "nested",
+                    vec![
+                        Field::new("float", DataType::Float64, true),
+                        Field::new("int", DataType::Int64, true),
+                    ],
+                    true,
+                ),
+                Field::new("string", DataType::Utf8, true),
+            ]),
+            Ok(vec![
+                Arc::new(BooleanArray::new_null(3)) as ArrayRef,
+                Arc::new(ListArray::from_iter_primitive::<Int64Type, _, _>(vec![
+                    Some(vec![Some(1)]),
+                    None,
+                    Some(vec![Some(2), Some(3)]),
+                ])),
+                Arc::new(StructArray::from(vec![
+                    (
+                        Arc::new(Field::new("float", DataType::Float64, true)),
+                        Arc::new(Float64Array::new_null(3)) as ArrayRef,
+                    ),
+                    (
+                        Arc::new(Field::new("int", DataType::Int64, true)),
+                        Arc::new(Int64Array::from(vec![-1, -2, -3])),
+                    ),
+                ])),
+                Arc::new(StringArray::from(vec!["a", "b", "c"])),
+            ]),
+        )
+        .test()?;
+
+        Ok(())
+    }
+}
diff --git a/src/datatypes/src/vectors/json/builder.rs b/src/datatypes/src/vectors/json/builder.rs
index ecc19f4bdd..be79a921c7 100644
--- a/src/datatypes/src/vectors/json/builder.rs
+++ b/src/datatypes/src/vectors/json/builder.rs
@@ -13,191 +13,46 @@
 // limitations under the License.
 
 use std::any::Any;
-use std::collections::HashMap;
-use std::sync::LazyLock;
 
 use crate::data_type::ConcreteDataType;
 use crate::error::{Result, TryFromValueSnafu, UnsupportedOperationSnafu};
-use crate::json::value::JsonValueRef;
+use crate::json::value::{JsonValue, JsonVariant};
 use crate::prelude::{ValueRef, Vector, VectorRef};
+use crate::types::JsonType;
 use crate::types::json_type::JsonNativeType;
-use crate::types::{JsonType, json_type};
-use crate::value::StructValueRef;
 use crate::vectors::{MutableVector, StructVectorBuilder};
 
-struct JsonStructsBuilder {
-    json_type: JsonType,
-    inner: StructVectorBuilder,
-}
-
-impl JsonStructsBuilder {
-    fn new(json_type: JsonType, capacity: usize) -> Self {
-        let struct_type = json_type.as_struct_type();
-        let inner = StructVectorBuilder::with_type_and_capacity(struct_type, capacity);
-        Self { json_type, inner }
-    }
-
-    fn len(&self) -> usize {
-        self.inner.len()
-    }
-
-    fn push(&mut self, json: &JsonValueRef) -> Result<()> {
-        let mut value = json.as_value_ref();
-        if !json.is_object() {
-            let fields = json_type::plain_json_struct_type(value.data_type());
-            value = ValueRef::Struct(StructValueRef::RefList {
-                val: vec![value],
-                fields,
-            })
-        }
-        self.inner.try_push_value_ref(&value)
-    }
-
-    /// Try to merge (and consume the data of) other json vector builder into this one.
-    /// Note that the other builder's json type must be able to be merged with this one's
-    /// (this one's json type has all the fields in other one's, and no datatypes conflict).
-    /// Normally this is guaranteed, as long as json values are pushed through [JsonVectorBuilder].
-    fn try_merge(&mut self, other: &mut JsonStructsBuilder) -> Result<()> {
-        debug_assert!(self.json_type.is_mergeable(&other.json_type));
-
-        fn helper(this: &mut StructVectorBuilder, that: &mut StructVectorBuilder) -> Result<()> {
-            let that_len = that.len();
-            if let Some(x) = that.mut_null_buffer().finish() {
-                this.mut_null_buffer().append_buffer(&x)
-            } else {
-                this.mut_null_buffer().append_n_non_nulls(that_len);
-            }
-
-            let that_fields = that.struct_type().fields();
-            let mut that_builders = that_fields
-                .iter()
-                .zip(that.mut_value_builders().iter_mut())
-                .map(|(field, builder)| (field.name(), builder))
-                .collect::<HashMap<_, _>>();
-
-            for (field, this_builder) in this
-                .struct_type()
-                .fields()
-                .iter()
-                .zip(this.mut_value_builders().iter_mut())
-            {
-                if let Some(that_builder) = that_builders.get_mut(field.name()) {
-                    if field.data_type().is_struct() {
-                        let this = this_builder
-                            .as_mut_any()
-                            .downcast_mut::<StructVectorBuilder>()
-                            // Safety: a struct datatype field must be corresponding to a struct vector builder.
-                            .unwrap();
-
-                        let that = that_builder
-                            .as_mut_any()
-                            .downcast_mut::<StructVectorBuilder>()
-                            // Safety: other builder with same field name must have same datatype,
-                            // ensured because the two json types are mergeable.
-                            .unwrap();
-                        helper(this, that)?;
-                    } else {
-                        let vector = that_builder.to_vector();
-                        this_builder.extend_slice_of(vector.as_ref(), 0, vector.len())?;
-                    }
-                } else {
-                    this_builder.push_nulls(that_len);
-                }
-            }
-            Ok(())
-        }
-        helper(&mut self.inner, &mut other.inner)
-    }
-
-    /// Same as [JsonStructsBuilder::try_merge], but does not consume the other builder's data.
-    fn try_merge_cloned(&mut self, other: &JsonStructsBuilder) -> Result<()> {
-        debug_assert!(self.json_type.is_mergeable(&other.json_type));
-
-        fn helper(this: &mut StructVectorBuilder, that: &StructVectorBuilder) -> Result<()> {
-            let that_len = that.len();
-            if let Some(x) = that.null_buffer().finish_cloned() {
-                this.mut_null_buffer().append_buffer(&x)
-            } else {
-                this.mut_null_buffer().append_n_non_nulls(that_len);
-            }
-
-            let that_fields = that.struct_type().fields();
-            let that_builders = that_fields
-                .iter()
-                .zip(that.value_builders().iter())
-                .map(|(field, builder)| (field.name(), builder))
-                .collect::<HashMap<_, _>>();
-
-            for (field, this_builder) in this
-                .struct_type()
-                .fields()
-                .iter()
-                .zip(this.mut_value_builders().iter_mut())
-            {
-                if let Some(that_builder) = that_builders.get(field.name()) {
-                    if field.data_type().is_struct() {
-                        let this = this_builder
-                            .as_mut_any()
-                            .downcast_mut::<StructVectorBuilder>()
-                            // Safety: a struct datatype field must be corresponding to a struct vector builder.
-                            .unwrap();
-
-                        let that = that_builder
-                            .as_any()
-                            .downcast_ref::<StructVectorBuilder>()
-                            // Safety: other builder with same field name must have same datatype,
-                            // ensured because the two json types are mergeable.
-                            .unwrap();
-                        helper(this, that)?;
-                    } else {
-                        let vector = that_builder.to_vector_cloned();
-                        this_builder.extend_slice_of(vector.as_ref(), 0, vector.len())?;
-                    }
-                } else {
-                    this_builder.push_nulls(that_len);
-                }
-            }
-            Ok(())
-        }
-        helper(&mut self.inner, &other.inner)
-    }
-}
-
-/// The vector builder for json type values.
-///
-/// Json type are dynamic, to some degree (as long as they can be merged into each other). So are
-/// json values. Json values are physically stored in struct vectors, which require the types of
-/// struct values to be fixed inside a certain struct vector. So to resolve "dynamic" vs "fixed"
-/// datatype problem, in this builder, each type of json value gets its own struct vector builder.
-/// Once new json type value is pushing into this builder, it creates a new "child" builder for it.
-///
-/// Given the "mixed" nature of the values stored in this builder, to produce the json vector, a
-/// "merge" operation is performed. The "merge" is to iterate over all the "child" builders, and fill
-/// nulls for missing json fields. The final vector's json type is fixed to be the "merge" of all
-/// pushed json types.
+#[derive(Clone)]
 pub(crate) struct JsonVectorBuilder {
     merged_type: JsonType,
-    capacity: usize,
-    builders: Vec<JsonStructsBuilder>,
+    values: Vec<JsonValue>,
 }
 
 impl JsonVectorBuilder {
     pub(crate) fn new(json_type: JsonNativeType, capacity: usize) -> Self {
         Self {
-            merged_type: JsonType::new_native(json_type),
-            capacity,
-            builders: vec![],
+            merged_type: JsonType::new_json2(json_type),
+            values: Vec::with_capacity(capacity),
         }
     }
 
-    fn try_create_new_builder(&mut self, json_type: &JsonType) -> Result<&mut JsonStructsBuilder> {
-        self.merged_type.merge(json_type)?;
+    fn try_build(&mut self) -> Result<VectorRef> {
+        let mut builder = StructVectorBuilder::with_type_and_capacity(
+            self.merged_type.as_struct_type(),
+            self.values.len(),
+        );
+        for value in self.values.iter_mut() {
+            value.try_align(&self.merged_type)?;
 
-        let builder = JsonStructsBuilder::new(json_type.clone(), self.capacity);
-        self.builders.push(builder);
+            if value.is_null() {
+                builder.push_null();
+                continue;
+            }
 
-        let len = self.builders.len();
-        Ok(&mut self.builders[len - 1])
+            let value = value.as_ref();
+            builder.try_push_value_ref(&value.as_struct_value())?;
+        }
+        Ok(builder.to_vector())
     }
 }
 
@@ -207,7 +62,7 @@ impl MutableVector for JsonVectorBuilder {
     }
 
     fn len(&self) -> usize {
-        self.builders.iter().map(|x| x.len()).sum()
+        self.values.len()
     }
 
     fn as_any(&self) -> &dyn Any {
@@ -219,37 +74,11 @@ impl MutableVector for JsonVectorBuilder {
     }
 
     fn to_vector(&mut self) -> VectorRef {
-        // Fast path:
-        if self.builders.len() == 1 {
-            return self.builders[0].inner.to_vector();
-        }
-
-        let mut unified_jsons = JsonStructsBuilder::new(self.merged_type.clone(), self.capacity);
-        for builder in self.builders.iter_mut() {
-            unified_jsons
-                .try_merge(builder)
-                // Safety: the "unified_jsons" has the merged json type from all the builders,
-                // so it should merge them without errors.
-                .unwrap_or_else(|e| panic!("failed to merge json builders, error: {e}"));
-        }
-        unified_jsons.inner.to_vector()
+        self.try_build().unwrap_or_else(|e| panic!("{}", e))
     }
 
     fn to_vector_cloned(&self) -> VectorRef {
-        // Fast path:
-        if self.builders.len() == 1 {
-            return self.builders[0].inner.to_vector_cloned();
-        }
-
-        let mut unified_jsons = JsonStructsBuilder::new(self.merged_type.clone(), self.capacity);
-        for builder in self.builders.iter() {
-            unified_jsons
-                .try_merge_cloned(builder)
-                // Safety: the "unified_jsons" has the merged json type from all the builders,
-                // so it should merge them without errors.
-                .unwrap_or_else(|e| panic!("failed to merge json builders, error: {e}"));
-        }
-        unified_jsons.inner.to_vector_cloned()
+        self.clone().to_vector()
     }
 
     fn try_push_value_ref(&mut self, value: &ValueRef) -> Result<()> {
@@ -260,29 +89,15 @@ impl MutableVector for JsonVectorBuilder {
             .fail();
         };
         let json_type = value.json_type();
+        self.merged_type.merge(json_type)?;
 
-        let builder = match self.builders.last_mut() {
-            Some(last) => {
-                // TODO(LFC): use "is_include" and amend json value with nulls
-                if &last.json_type != json_type {
-                    self.try_create_new_builder(json_type)?
-                } else {
-                    last
-                }
-            }
-            None => self.try_create_new_builder(json_type)?,
-        };
-
-        builder.push(value.as_ref())
+        let value = JsonValue::new(JsonVariant::from(value.variant().clone()));
+        self.values.push(value);
+        Ok(())
     }
 
     fn push_null(&mut self) {
-        static NULL_JSON: LazyLock<ValueRef> =
-            LazyLock::new(|| ValueRef::Json(Box::new(JsonValueRef::null())));
-        self.try_push_value_ref(&NULL_JSON)
-            // Safety: learning from the method "try_push_value_ref", a null json value should be
-            // always able to push into any json vectors.
-            .unwrap_or_else(|e| panic!("failed to push null json value, error: {e}"));
+        self.values.push(JsonValue::null())
     }
 
     fn extend_slice_of(&mut self, _: &dyn Vector, _: usize, _: usize) -> Result<()> {
@@ -296,189 +111,104 @@ impl MutableVector for JsonVectorBuilder {
 
 #[cfg(test)]
 mod tests {
+    use common_base::bytes::Bytes;
+
     use super::*;
-    use crate::data_type::DataType;
-    use crate::json::JsonStructureSettings;
-    use crate::vectors::helper::pretty_print;
-
-    fn push(json: &str, builder: &mut JsonVectorBuilder, expected: std::result::Result<(), &str>) {
-        let settings = JsonStructureSettings::Structured(None);
-        let json: serde_json::Value = serde_json::from_str(json).unwrap();
-        let value = settings.encode(json).unwrap();
-
-        let value = value.as_value_ref();
-        let result = builder
-            .try_push_value_ref(&value)
-            .map_err(|e| e.to_string());
-        let expected = expected.map_err(|e| e.to_string());
-        assert_eq!(result, expected);
-    }
+    use crate::data_type::ConcreteDataType;
+    use crate::types::json_type::JsonObjectType;
+    use crate::value::{StructValue, Value, ValueRef};
 
     #[test]
-    fn test_push_plain_jsons() -> Result<()> {
-        let jsons = vec!["1", "2", r#""s""#, "[true]"];
-        let results = vec![
-            Ok(()),
-            Ok(()),
-            Err(
-                r#"Failed to merge JSON datatype: datatypes have conflict, this: "<Number>", that: "<String>""#,
-            ),
-            Err(
-                r#"Failed to merge JSON datatype: datatypes have conflict, this: "<Number>", that: ["<Bool>"]"#,
-            ),
-        ];
-        let mut builder = JsonVectorBuilder::new(JsonNativeType::Null, 1);
-        for (json, result) in jsons.into_iter().zip(results.into_iter()) {
-            push(json, &mut builder, result);
+    fn test_json_vector_builder() -> Result<()> {
+        fn parse_json_value(json: &str) -> Value {
+            let value: serde_json::Value = serde_json::from_str(json).unwrap();
+            Value::Json(Box::new(value.into()))
         }
+
+        // Object inputs should merge into a superset schema, preserve null rows,
+        // and align conflicting nested values into Variant payloads.
+        let mut builder = JsonVectorBuilder::new(JsonNativeType::Object(Default::default()), 3);
+        let first = parse_json_value(r#"{"id":1,"payload":{"name":"foo"}}"#);
+        let second = parse_json_value(r#"{"id":2,"extra":true,"payload":"raw"}"#);
+        builder.try_push_value_ref(&first.as_value_ref())?;
+        builder.push_null();
+        builder.try_push_value_ref(&second.as_value_ref())?;
+
+        let merged_type = JsonType::new_json2(JsonNativeType::Object(JsonObjectType::from([
+            ("extra".to_string(), JsonNativeType::Bool),
+            ("id".to_string(), JsonNativeType::i64()),
+            ("payload".to_string(), JsonNativeType::Variant),
+        ])));
+        assert_eq!(
+            builder.data_type(),
+            ConcreteDataType::Json(merged_type.clone())
+        );
+
+        let merged_struct_type = merged_type.as_struct_type();
         let vector = builder.to_vector();
-        let expected = r#"
-+---------------------+
-| StructVector        |
-+---------------------+
-| {__json_plain__: 1} |
-| {__json_plain__: 2} |
-+---------------------+"#;
-        assert_eq!(pretty_print(vector), expected.trim());
-        Ok(())
-    }
+        assert_eq!(vector.len(), 3);
+        assert_eq!(
+            vector.get(0),
+            Value::Struct(StructValue::new(
+                vec![
+                    Value::Null,
+                    Value::Int64(1),
+                    Value::Binary(Bytes::from(br#"{"name":"foo"}"#.to_vec())),
+                ],
+                merged_struct_type.clone(),
+            ))
+        );
+        assert_eq!(vector.get(1), Value::Null);
+        assert_eq!(
+            vector.get(2),
+            Value::Struct(StructValue::new(
+                vec![
+                    Value::Boolean(true),
+                    Value::Int64(2),
+                    Value::Binary(Bytes::from(br#""raw""#.to_vec())),
+                ],
+                merged_struct_type,
+            ))
+        );
 
-    #[test]
-    fn test_push_json_objects() -> Result<()> {
-        let jsons = vec![
-            r#"{
-            "s": "a",
-            "list": [1, 2, 3]
-        }"#,
-            r#"{
-            "list": [4],
-            "s": "b"
-        }"#,
-            r#"{
-            "s": "c",
-            "float": 0.9
-        }"#,
-            r#"{
-            "float": 0.8,
-            "s": "d"
-        }"#,
-            r#"{
-            "float": 0.7,
-            "int": -1
-        }"#,
-            r#"{
-            "int": 0,
-            "float": 0.6
-        }"#,
-            r#"{
-            "int": 1,
-            "object": {"hello": "world", "timestamp": 1761523200000}
-        }"#,
-            r#"{
-            "object": {"hello": "greptime", "timestamp": 1761523201000},
-            "int": 2
-        }"#,
-            r#"{
-            "object": {"timestamp": 1761523202000},
-            "nested": {"a": {"b": {"b": {"a": "abba"}}}}
-        }"#,
-            r#"{
-            "nested": {"a": {"b": {"a": {"b": "abab"}}}},
-            "object": {"timestamp": 1761523203000}
-        }"#,
-        ];
-        let mut builder = JsonVectorBuilder::new(JsonNativeType::Null, 1);
-        for json in jsons {
-            push(json, &mut builder, Ok(()));
-        }
-        assert_eq!(builder.len(), 10);
+        // Root-level conflicts should be lifted to a plain Variant field that preserves
+        // each original JSON payload.
+        let mut variant_builder = JsonVectorBuilder::new(JsonNativeType::Bool, 2);
+        let object = parse_json_value(r#"{"k":1}"#);
+        let boolean = parse_json_value("true");
+        variant_builder.try_push_value_ref(&boolean.as_value_ref())?;
+        variant_builder.try_push_value_ref(&object.as_value_ref())?;
 
-        // test children builders:
-        assert_eq!(builder.builders.len(), 6);
-        let expect_types = [
-            r#"Json<{"list":["<Number>"],"s":"<String>"}>"#,
-            r#"Json<{"float":"<Number>","s":"<String>"}>"#,
-            r#"Json<{"float":"<Number>","int":"<Number>"}>"#,
-            r#"Json<{"int":"<Number>","object":{"hello":"<String>","timestamp":"<Number>"}}>"#,
-            r#"Json<{"nested":{"a":{"b":{"b":{"a":"<String>"}}}},"object":{"timestamp":"<Number>"}}>"#,
-            r#"Json<{"nested":{"a":{"b":{"a":{"b":"<String>"}}}},"object":{"timestamp":"<Number>"}}>"#,
-        ];
-        let expect_vectors = [
-            r#"
-+-------------------------+
-| StructVector            |
-+-------------------------+
-| {list: [1, 2, 3], s: a} |
-| {list: [4], s: b}       |
-+-------------------------+"#,
-            r#"
-+--------------------+
-| StructVector       |
-+--------------------+
-| {float: 0.9, s: c} |
-| {float: 0.8, s: d} |
-+--------------------+"#,
-            r#"
-+-----------------------+
-| StructVector          |
-+-----------------------+
-| {float: 0.7, int: -1} |
-| {float: 0.6, int: 0}  |
-+-----------------------+"#,
-            r#"
-+---------------------------------------------------------------+
-| StructVector                                                  |
-+---------------------------------------------------------------+
-| {int: 1, object: {hello: world, timestamp: 1761523200000}}    |
-| {int: 2, object: {hello: greptime, timestamp: 1761523201000}} |
-+---------------------------------------------------------------+"#,
-            r#"
-+------------------------------------------------------------------------+
-| StructVector                                                           |
-+------------------------------------------------------------------------+
-| {nested: {a: {b: {b: {a: abba}}}}, object: {timestamp: 1761523202000}} |
-+------------------------------------------------------------------------+"#,
-            r#"
-+------------------------------------------------------------------------+
-| StructVector                                                           |
-+------------------------------------------------------------------------+
-| {nested: {a: {b: {a: {b: abab}}}}, object: {timestamp: 1761523203000}} |
-+------------------------------------------------------------------------+"#,
-        ];
-        for (builder, (expect_type, expect_vector)) in builder
-            .builders
-            .iter()
-            .zip(expect_types.into_iter().zip(expect_vectors.into_iter()))
-        {
-            assert_eq!(builder.json_type.name(), expect_type);
-            let vector = builder.inner.to_vector_cloned();
-            assert_eq!(pretty_print(vector), expect_vector.trim());
-        }
+        let variant_type = JsonType::new_json2(JsonNativeType::Variant);
+        assert_eq!(
+            variant_builder.data_type(),
+            ConcreteDataType::Json(variant_type.clone())
+        );
 
-        // test final merged json type:
-        let expected = r#"Json<{"float":"<Number>","int":"<Number>","list":["<Number>"],"nested":{"a":{"b":{"a":{"b":"<String>"},"b":{"a":"<String>"}}}},"object":{"hello":"<String>","timestamp":"<Number>"},"s":"<String>"}>"#;
-        assert_eq!(builder.data_type().to_string(), expected);
+        let variant_struct_type = variant_type.as_struct_type();
+        let vector = variant_builder.to_vector();
+        assert_eq!(
+            vector.get(0),
+            Value::Struct(StructValue::new(
+                vec![Value::Binary(Bytes::from(b"true".to_vec()))],
+                variant_struct_type.clone(),
+            ))
+        );
+        assert_eq!(
+            vector.get(1),
+            Value::Struct(StructValue::new(
+                vec![Value::Binary(Bytes::from(br#"{"k":1}"#.to_vec()))],
+                variant_struct_type,
+            ))
+        );
+
+        // Non-JSON values should be rejected at push time.
+        let mut invalid_builder = JsonVectorBuilder::new(JsonNativeType::Bool, 1);
+        let err = invalid_builder
+            .try_push_value_ref(&ValueRef::Boolean(true))
+            .unwrap_err();
+        assert!(err.to_string().contains("expected json value"));
 
-        // test final produced vector:
-        let expected = r#"
-+-------------------------------------------------------------------------------------------------------------------+
-| StructVector                                                                                                      |
-+-------------------------------------------------------------------------------------------------------------------+
-| {float: , int: , list: [1, 2, 3], nested: , object: , s: a}                                                       |
-| {float: , int: , list: [4], nested: , object: , s: b}                                                             |
-| {float: 0.9, int: , list: , nested: , object: , s: c}                                                             |
-| {float: 0.8, int: , list: , nested: , object: , s: d}                                                             |
-| {float: 0.7, int: -1, list: , nested: , object: , s: }                                                            |
-| {float: 0.6, int: 0, list: , nested: , object: , s: }                                                             |
-| {float: , int: 1, list: , nested: , object: {hello: world, timestamp: 1761523200000}, s: }                        |
-| {float: , int: 2, list: , nested: , object: {hello: greptime, timestamp: 1761523201000}, s: }                     |
-| {float: , int: , list: , nested: {a: {b: {a: , b: {a: abba}}}}, object: {hello: , timestamp: 1761523202000}, s: } |
-| {float: , int: , list: , nested: {a: {b: {a: {b: abab}, b: }}}, object: {hello: , timestamp: 1761523203000}, s: } |
-+-------------------------------------------------------------------------------------------------------------------+"#;
-        let vector = builder.to_vector_cloned();
-        assert_eq!(pretty_print(vector), expected.trim());
-        let vector = builder.to_vector();
-        assert_eq!(pretty_print(vector), expected.trim());
         Ok(())
     }
 }
diff --git a/src/datatypes/src/vectors/struct_vector.rs b/src/datatypes/src/vectors/struct_vector.rs
index 44de9abf5e..d3c44840aa 100644
--- a/src/datatypes/src/vectors/struct_vector.rs
+++ b/src/datatypes/src/vectors/struct_vector.rs
@@ -122,7 +122,7 @@ impl Vector for StructVector {
             .map(|i| {
                 let field_array = &self.array.column(i);
 
-                if field_array.is_null(i) {
+                if field_array.is_null(index) {
                     Value::Null
                 } else {
                     let scalar_value = ScalarValue::try_from_array(field_array, index).unwrap();
@@ -323,26 +323,6 @@ impl StructVectorBuilder {
         }
         self.null_buffer.append_null();
     }
-
-    pub(crate) fn struct_type(&self) -> &StructType {
-        &self.fields
-    }
-
-    pub(crate) fn value_builders(&self) -> &[Box<dyn MutableVector>] {
-        &self.value_builders
-    }
-
-    pub(crate) fn mut_value_builders(&mut self) -> &mut [Box<dyn MutableVector>] {
-        &mut self.value_builders
-    }
-
-    pub(crate) fn null_buffer(&self) -> &NullBufferBuilder {
-        &self.null_buffer
-    }
-
-    pub(crate) fn mut_null_buffer(&mut self) -> &mut NullBufferBuilder {
-        &mut self.null_buffer
-    }
 }
 
 impl MutableVector for StructVectorBuilder {
diff --git a/src/file-engine/Cargo.toml b/src/file-engine/Cargo.toml
index 1ef7f8ea52..8d69512fa1 100644
--- a/src/file-engine/Cargo.toml
+++ b/src/file-engine/Cargo.toml
@@ -30,7 +30,6 @@ datafusion-orc.workspace = true
 datatypes.workspace = true
 futures.workspace = true
 object-store.workspace = true
-object_store_opendal.workspace = true
 serde = { version = "1.0", features = ["derive"] }
 serde_json.workspace = true
 snafu.workspace = true
diff --git a/src/file-engine/src/engine.rs b/src/file-engine/src/engine.rs
index 693ae325df..175ebef237 100644
--- a/src/file-engine/src/engine.rs
+++ b/src/file-engine/src/engine.rs
@@ -94,7 +94,7 @@ impl RegionEngine for FileRegionEngine {
         let stream = self.handle_query(region_id, request).await?;
         let metadata = self.get_metadata(region_id).await?;
         // We don't support enabling append mode for file engine.
-        let scanner = Box::new(SinglePartitionScanner::new(stream, false, metadata));
+        let scanner = Box::new(SinglePartitionScanner::new(stream, false, metadata, None));
         Ok(scanner)
     }
 
diff --git a/src/file-engine/src/lib.rs b/src/file-engine/src/lib.rs
index cc9bac8c6f..51d13946cd 100644
--- a/src/file-engine/src/lib.rs
+++ b/src/file-engine/src/lib.rs
@@ -12,8 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
-
 pub mod config;
 pub mod engine;
 pub mod error;
diff --git a/src/file-engine/src/query.rs b/src/file-engine/src/query.rs
index 2182978631..0ac563606a 100644
--- a/src/file-engine/src/query.rs
+++ b/src/file-engine/src/query.rs
@@ -44,7 +44,8 @@ impl FileRegion {
     pub fn query(&self, request: ScanRequest) -> Result<SendableRecordBatchStream> {
         let store = build_backend(&self.url, &self.options).context(BuildBackendSnafu)?;
 
-        let file_projection = self.projection_pushdown_to_file(&request.projection)?;
+        let projection = request.projection_indices();
+        let file_projection = self.projection_pushdown_to_file(projection)?;
         let file_filters = self.filters_pushdown_to_file(&request.filters)?;
         let file_schema = Arc::new(Schema::new(self.file_options.file_column_schemas.clone()));
 
@@ -70,7 +71,7 @@ impl FileRegion {
             },
         )?;
 
-        let scan_schema = self.scan_schema(&request.projection)?;
+        let scan_schema = self.scan_schema(projection)?;
 
         Ok(Box::pin(FileToScanRegionStream::new(
             scan_schema,
@@ -81,9 +82,9 @@ impl FileRegion {
 
     fn projection_pushdown_to_file(
         &self,
-        req_projection: &Option<Vec<usize>>,
+        req_projection: Option<&[usize]>,
     ) -> Result<Option<Vec<usize>>> {
-        let Some(scan_projection) = req_projection.as_ref() else {
+        let Some(scan_projection) = req_projection else {
             return Ok(None);
         };
 
@@ -136,7 +137,7 @@ impl FileRegion {
         Ok(file_filters)
     }
 
-    fn scan_schema(&self, req_projection: &Option<Vec<usize>>) -> Result<SchemaRef> {
+    fn scan_schema(&self, req_projection: Option<&[usize]>) -> Result<SchemaRef> {
         let schema = if let Some(indices) = req_projection {
             Arc::new(
                 self.metadata
diff --git a/src/file-engine/src/query/file_stream.rs b/src/file-engine/src/query/file_stream.rs
index 73453589dc..931dfabe62 100644
--- a/src/file-engine/src/query/file_stream.rs
+++ b/src/file-engine/src/query/file_stream.rs
@@ -61,7 +61,7 @@ fn build_record_batch_stream(
             .with_file_group(FileGroup::new(files))
             .build();
 
-    let store = Arc::new(object_store_opendal::OpendalStore::new(
+    let store = Arc::new(object_store::compat::OpendalStore::new(
         scan_plan_config.store.clone(),
     ));
 
diff --git a/src/file-engine/src/region.rs b/src/file-engine/src/region.rs
index c17e797966..3808b33a67 100644
--- a/src/file-engine/src/region.rs
+++ b/src/file-engine/src/region.rs
@@ -105,7 +105,7 @@ impl FileRegion {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use store_api::region_request::PathType;
 
diff --git a/src/flow/src/adapter/flownode_impl.rs b/src/flow/src/adapter/flownode_impl.rs
index d1e4600e23..53a3265d7d 100644
--- a/src/flow/src/adapter/flownode_impl.rs
+++ b/src/flow/src/adapter/flownode_impl.rs
@@ -213,7 +213,7 @@ impl FlowDualEngine {
             if !frontend_list.is_empty() {
                 let fe_list = frontend_list
                     .iter()
-                    .map(|(_, info)| &info.peer.addr)
+                    .map(|peer| &peer.addr)
                     .collect::<Vec<_>>();
                 info!("Available frontend found: {:?}", fe_list);
                 return Ok(());
@@ -1060,7 +1060,7 @@ impl StreamingEngine {
 
                 let fetch_order: Vec<FetchFromRow> = table_col_names
                     .iter()
-                    .zip(default_vals.into_iter())
+                    .zip(default_vals)
                     .map(|(col_name, col_default_val)| {
                         name_to_col
                             .get(col_name)
diff --git a/src/flow/src/batching_mode.rs b/src/flow/src/batching_mode.rs
index 7817cfb6d0..4162daa20c 100644
--- a/src/flow/src/batching_mode.rs
+++ b/src/flow/src/batching_mode.rs
@@ -48,10 +48,6 @@ pub struct BatchingModeOptions {
     /// which prevent flownode from starting
     #[serde(with = "humantime_serde")]
     pub experimental_frontend_scan_timeout: Duration,
-    /// Frontend activity timeout
-    /// if frontend is down(not sending heartbeat) for more than frontend_activity_timeout, it will be removed from the list that flownode use to connect
-    #[serde(with = "humantime_serde")]
-    pub experimental_frontend_activity_timeout: Duration,
     /// Maximum number of filters allowed in a single query
     pub experimental_max_filter_num_per_query: usize,
     /// Time window merge distance
@@ -71,7 +67,6 @@ impl Default for BatchingModeOptions {
             grpc_conn_timeout: Duration::from_secs(5),
             experimental_grpc_max_retries: 3,
             experimental_frontend_scan_timeout: Duration::from_secs(30),
-            experimental_frontend_activity_timeout: Duration::from_secs(60),
             experimental_max_filter_num_per_query: 20,
             experimental_time_window_merge_threshold: 3,
             read_preference: Default::default(),
diff --git a/src/flow/src/batching_mode/frontend_client.rs b/src/flow/src/batching_mode/frontend_client.rs
index 8ec6de3a08..c29c52846b 100644
--- a/src/flow/src/batching_mode/frontend_client.rs
+++ b/src/flow/src/batching_mode/frontend_client.rs
@@ -16,7 +16,6 @@
 
 use std::collections::HashMap;
 use std::sync::{Arc, Mutex, Weak};
-use std::time::SystemTime;
 
 use api::v1::greptime_request::Request;
 use api::v1::query_request::Query;
@@ -24,9 +23,7 @@ use api::v1::{CreateTableExpr, QueryRequest};
 use client::{Client, Database};
 use common_error::ext::BoxedError;
 use common_grpc::channel_manager::{ChannelConfig, ChannelManager, load_client_tls_config};
-use common_meta::cluster::{NodeInfo, NodeInfoKey, Role};
-use common_meta::peer::Peer;
-use common_meta::rpc::store::RangeRequest;
+use common_meta::peer::{Peer, PeerDiscovery};
 use common_query::Output;
 use common_telemetry::warn;
 use meta_client::client::MetaClient;
@@ -199,39 +196,23 @@ impl DatabaseWithPeer {
 }
 
 impl FrontendClient {
+    // TODO: support more fine-grained load balancing strategies for frontend
+    // selection, such as AZ (availability zone) awareness, to prefer frontends
+    // in the same zone as the flownode and reduce cross-AZ latency.
     /// scan for available frontend from metadata
-    pub(crate) async fn scan_for_frontend(&self) -> Result<Vec<(NodeInfoKey, NodeInfo)>, Error> {
+    pub(crate) async fn scan_for_frontend(&self) -> Result<Vec<Peer>, Error> {
         let Self::Distributed { meta_client, .. } = self else {
             return Ok(vec![]);
         };
-        let cluster_client = meta_client
-            .cluster_client()
-            .map_err(BoxedError::new)
-            .context(ExternalSnafu)?;
 
-        let prefix = NodeInfoKey::key_prefix_with_role(Role::Frontend);
-        let req = RangeRequest::new().with_prefix(prefix);
-        let resp = cluster_client
-            .range(req)
+        meta_client
+            .active_frontends()
             .await
             .map_err(BoxedError::new)
-            .context(ExternalSnafu)?;
-        let mut res = Vec::with_capacity(resp.kvs.len());
-        for kv in resp.kvs {
-            let key = NodeInfoKey::try_from(kv.key)
-                .map_err(BoxedError::new)
-                .context(ExternalSnafu)?;
-
-            let val = NodeInfo::try_from(kv.value)
-                .map_err(BoxedError::new)
-                .context(ExternalSnafu)?;
-            res.push((key, val));
-        }
-        Ok(res)
+            .context(ExternalSnafu)
     }
 
-    /// Get the frontend with recent enough(less than 1 minute from now) `last_activity_ts`
-    /// and is able to process query
+    /// Get a frontend discovered by metasrv and verified with a query probe.
     async fn get_random_active_frontend(
         &self,
         catalog: &str,
@@ -255,26 +236,11 @@ impl FrontendClient {
         interval.tick().await;
         for retry in 0..batch_opts.experimental_grpc_max_retries {
             let mut frontends = self.scan_for_frontend().await?;
-            let now_in_ms = SystemTime::now()
-                .duration_since(SystemTime::UNIX_EPOCH)
-                .unwrap()
-                .as_millis() as i64;
             // shuffle the frontends to avoid always pick the same one
             frontends.shuffle(&mut rng());
 
-            // found node with maximum last_activity_ts
-            for (_, node_info) in frontends
-                .iter()
-                // filter out frontend that have been down for more than 1 min
-                .filter(|(_, node_info)| {
-                    node_info.last_activity_ts
-                        + batch_opts
-                            .experimental_frontend_activity_timeout
-                            .as_millis() as i64
-                        > now_in_ms
-                })
-            {
-                let addr = &node_info.peer.addr;
+            for peer in frontends {
+                let addr = peer.addr.clone();
                 let client = Client::with_manager_and_urls(chnl_mgr.clone(), vec![addr.clone()]);
                 let database = {
                     let mut db = Database::new(catalog, schema, client);
@@ -283,7 +249,7 @@ impl FrontendClient {
                     }
                     db
                 };
-                let db = DatabaseWithPeer::new(database, node_info.peer.clone());
+                let db = DatabaseWithPeer::new(database, peer);
                 match db.try_select_one().await {
                     Ok(_) => return Ok(db),
                     Err(e) => {
@@ -351,12 +317,7 @@ impl FrontendClient {
                         database_client
                             .handler
                             .lock()
-                            .map_err(|e| {
-                                UnexpectedSnafu {
-                                    reason: format!("Failed to lock database client: {e}"),
-                                }
-                                .build()
-                            })?
+                            .unwrap()
                             .as_ref()
                             .context(UnexpectedSnafu {
                                 reason: "Standalone's frontend instance is not set",
@@ -429,12 +390,7 @@ impl FrontendClient {
                         database_client
                             .handler
                             .lock()
-                            .map_err(|e| {
-                                UnexpectedSnafu {
-                                    reason: format!("Failed to lock database client: {e}"),
-                                }
-                                .build()
-                            })?
+                            .unwrap()
                             .as_ref()
                             .context(UnexpectedSnafu {
                                 reason: "Standalone's frontend instance is not set",
diff --git a/src/flow/src/lib.rs b/src/flow/src/lib.rs
index bd4cbd9f08..fe8a760a07 100644
--- a/src/flow/src/lib.rs
+++ b/src/flow/src/lib.rs
@@ -17,7 +17,6 @@
 //! It also contains definition of expression, adapter and plan, and internal state management.
 
 #![allow(dead_code)]
-#![warn(clippy::missing_docs_in_private_items)]
 #![warn(clippy::too_many_lines)]
 
 // TODO(discord9): enable this lint to handle out of bound access
diff --git a/src/flow/src/utils.rs b/src/flow/src/utils.rs
index 8f28ed23c7..1a9879b996 100644
--- a/src/flow/src/utils.rs
+++ b/src/flow/src/utils.rs
@@ -213,7 +213,7 @@ impl KeyExpiryManager {
         let mut before = self.event_ts_to_key.split_off(&expire_time);
         std::mem::swap(&mut before, &mut self.event_ts_to_key);
 
-        Some(before.into_iter().flat_map(|(_ts, keys)| keys.into_iter()))
+        Some(before.into_values().flat_map(|keys| keys.into_iter()))
     }
 }
 
@@ -409,8 +409,8 @@ impl Arrangement {
         // iter over batches that only have updates of `timestamp>now` and find the first non empty batch, then get the minimum timestamp in that batch
         for (_ts, batch) in self.spine.range((Bound::Excluded(now), Bound::Unbounded)) {
             let min_ts = batch
-                .iter()
-                .flat_map(|(_k, v)| v.iter().map(|(_, ts, _)| *ts).min())
+                .values()
+                .flat_map(|v| v.iter().map(|(_, ts, _)| *ts).min())
                 .min();
 
             if min_ts.is_some() {
diff --git a/src/frontend/Cargo.toml b/src/frontend/Cargo.toml
index 03b0d35130..aa9276a0cb 100644
--- a/src/frontend/Cargo.toml
+++ b/src/frontend/Cargo.toml
@@ -52,11 +52,13 @@ futures.workspace = true
 hostname.workspace = true
 humantime.workspace = true
 humantime-serde.workspace = true
+itertools.workspace = true
 lazy_static.workspace = true
 log-query.workspace = true
 meta-client.workspace = true
 num_cpus.workspace = true
 opentelemetry-proto.workspace = true
+opentelemetry-semantic-conventions = { version = "0.31", features = ["semconv_experimental"] }
 operator.workspace = true
 otel-arrow-rust.workspace = true
 partition.workspace = true
diff --git a/src/frontend/src/error.rs b/src/frontend/src/error.rs
index d148e6aa1b..6f78d23e14 100644
--- a/src/frontend/src/error.rs
+++ b/src/frontend/src/error.rs
@@ -399,7 +399,7 @@ impl ErrorExt for Error {
 
             Error::PrometheusLabelValuesQueryPlan { source, .. } => source.status_code(),
 
-            Error::CollectRecordbatch { .. } => StatusCode::EngineExecuteQuery,
+            Error::CollectRecordbatch { source, .. } => source.status_code(),
 
             Error::SqlExecIntercepted { source, .. } => source.status_code(),
             Error::StartServer { source, .. } => source.status_code(),
diff --git a/src/frontend/src/frontend.rs b/src/frontend/src/frontend.rs
index 69ae59517e..6e862fdce4 100644
--- a/src/frontend/src/frontend.rs
+++ b/src/frontend/src/frontend.rs
@@ -72,6 +72,8 @@ pub struct FrontendOptions {
     pub memory: MemoryOptions,
     /// The event recorder options.
     pub event_recorder: EventRecorderOptions,
+    /// Environment variable keys to read and report in heartbeat messages.
+    pub heartbeat_env_vars: Vec<String>,
 }
 
 impl Default for FrontendOptions {
@@ -101,13 +103,14 @@ impl Default for FrontendOptions {
             slow_query: SlowQueryOptions::default(),
             memory: MemoryOptions::default(),
             event_recorder: EventRecorderOptions::default(),
+            heartbeat_env_vars: vec![],
         }
     }
 }
 
 impl Configurable for FrontendOptions {
     fn env_list_keys() -> Option<&'static [&'static str]> {
-        Some(&["meta_client.metasrv_addrs"])
+        Some(&["heartbeat_env_vars", "meta_client.metasrv_addrs"])
     }
 }
 
diff --git a/src/frontend/src/heartbeat.rs b/src/frontend/src/heartbeat.rs
index add4940214..937c74be19 100644
--- a/src/frontend/src/heartbeat.rs
+++ b/src/frontend/src/heartbeat.rs
@@ -18,6 +18,7 @@ mod tests;
 use std::sync::Arc;
 
 use api::v1::meta::{HeartbeatRequest, NodeInfo, Peer};
+use common_meta::datanode::EnvVars;
 use common_meta::heartbeat::handler::{
     HeartbeatResponseHandlerContext, HeartbeatResponseHandlerExecutorRef,
 };
@@ -45,6 +46,7 @@ pub struct HeartbeatTask {
     resp_handler_executor: HeartbeatResponseHandlerExecutorRef,
     start_time_ms: u64,
     resource_stat: ResourceStatRef,
+    env_vars: EnvVars,
 }
 
 impl HeartbeatTask {
@@ -67,6 +69,7 @@ impl HeartbeatTask {
             resp_handler_executor,
             start_time_ms: common_time::util::current_time_millis() as u64,
             resource_stat,
+            env_vars: EnvVars::from_config(&opts.heartbeat_env_vars),
         }
     }
 
@@ -202,10 +205,14 @@ impl HeartbeatTask {
         let total_cpu_millicores = self.resource_stat.get_total_cpu_millicores();
         let total_memory_bytes = self.resource_stat.get_total_memory_bytes();
         let resource_stat = self.resource_stat.clone();
+        let env_vars = self.env_vars.clone();
         common_runtime::spawn_hb(async move {
             let sleep = tokio::time::sleep(Duration::from_millis(0));
             tokio::pin!(sleep);
 
+            let mut extensions = std::collections::HashMap::new();
+            env_vars.into_extensions(&mut extensions);
+
             let heartbeat_request = HeartbeatRequest {
                 peer: self_peer,
                 info: Self::build_node_info(
@@ -213,6 +220,7 @@ impl HeartbeatTask {
                     total_cpu_millicores,
                     total_memory_bytes,
                 ),
+                extensions,
                 ..Default::default()
             };
 
diff --git a/src/frontend/src/instance.rs b/src/frontend/src/instance.rs
index fa8a74cad2..dbce6a693e 100644
--- a/src/frontend/src/instance.rs
+++ b/src/frontend/src/instance.rs
@@ -13,6 +13,7 @@
 // limitations under the License.
 
 pub mod builder;
+mod dashboard;
 mod grpc;
 mod influxdb;
 mod jaeger;
@@ -194,7 +195,7 @@ impl Instance {
         let query_interceptor = self.plugins.get::<SqlQueryInterceptorRef<Error>>();
         let query_interceptor = query_interceptor.as_ref();
 
-        if should_capture_statement(Some(&stmt)) {
+        if stmt.is_readonly() {
             let slow_query_timer = self
                 .slow_query_options
                 .enable
@@ -482,6 +483,19 @@ fn derive_timeout(stmt: &Statement, query_ctx: &QueryContextRef) -> Option<Durat
     }
 }
 
+/// Derives timeout for plan execution.
+fn derive_timeout_for_plan(plan: &LogicalPlan, query_ctx: &QueryContextRef) -> Option<Duration> {
+    let query_timeout = query_ctx.query_timeout()?;
+    if query_timeout.is_zero() {
+        return None;
+    }
+    match query_ctx.channel() {
+        Channel::Mysql if is_readonly_plan(plan) => Some(query_timeout),
+        Channel::Postgres => Some(query_timeout),
+        _ => None,
+    }
+}
+
 fn attach_timeout(output: Output, mut timeout: Duration) -> Result<Output> {
     if timeout.is_zero() {
         return StatementTimeoutSnafu.fail();
@@ -587,18 +601,53 @@ impl Instance {
         }
     }
 
+    async fn exec_plan(&self, plan: LogicalPlan, query_ctx: QueryContextRef) -> Result<Output> {
+        self.query_engine
+            .execute(plan, query_ctx)
+            .await
+            .context(ExecLogicalPlanSnafu)
+    }
+
+    async fn exec_plan_with_timeout(
+        &self,
+        plan: LogicalPlan,
+        query_ctx: QueryContextRef,
+    ) -> Result<Output> {
+        let timeout = derive_timeout_for_plan(&plan, &query_ctx);
+        match timeout {
+            Some(timeout) => {
+                let start = tokio::time::Instant::now();
+                let output = tokio::time::timeout(timeout, self.exec_plan(plan, query_ctx))
+                    .await
+                    .map_err(|_| StatementTimeoutSnafu.build())??;
+                let remaining_timeout = timeout.checked_sub(start.elapsed()).unwrap_or_default();
+                attach_timeout(output, remaining_timeout)
+            }
+            None => self.exec_plan(plan, query_ctx).await,
+        }
+    }
+
     async fn do_exec_plan_inner(
         &self,
-        stmt: Option<Statement>,
         plan: LogicalPlan,
+        stmt: Option<Statement>,
         query_ctx: QueryContextRef,
     ) -> Result<Output> {
         ensure!(!self.is_suspended(), error::SuspendedSnafu);
 
-        if should_capture_statement(stmt.as_ref()) {
-            // It's safe to unwrap here because we've already checked the type.
-            let stmt = stmt.unwrap();
-            let query = stmt.to_string();
+        let query_interceptor_opt = self.plugins.get::<SqlQueryInterceptorRef<Error>>();
+        let query_interceptor = query_interceptor_opt.as_ref();
+
+        if let Some(ref s) = stmt {
+            query_interceptor.pre_execute(s, Some(&plan), query_ctx.clone())?;
+        }
+
+        let query = stmt
+            .as_ref()
+            .map(|s| s.to_string())
+            .unwrap_or_else(|| plan.display_indent().to_string());
+
+        let result = if is_readonly_plan(&plan) {
             let slow_query_timer = self
                 .slow_query_options
                 .enable
@@ -606,7 +655,7 @@ impl Instance {
                 .flatten()
                 .map(|event_recorder| {
                     SlowQueryTimer::new(
-                        CatalogQueryStatement::Sql(stmt.clone()),
+                        CatalogQueryStatement::Plan(query.clone()),
                         self.slow_query_options.threshold,
                         self.slow_query_options.sample_ratio,
                         self.slow_query_options.record_type,
@@ -623,7 +672,7 @@ impl Instance {
                 slow_query_timer,
             );
 
-            let query_fut = self.query_engine.execute(plan.clone(), query_ctx);
+            let query_fut = self.exec_plan_with_timeout(plan, query_ctx.clone());
 
             CancellableFuture::new(query_fut, ticket.cancellation_handle.clone())
                 .await
@@ -639,15 +688,11 @@ impl Instance {
                     };
                     Output { data, meta }
                 })
-                .context(ExecLogicalPlanSnafu)
         } else {
-            // plan should be prepared before exec
-            // we'll do check there
-            self.query_engine
-                .execute(plan.clone(), query_ctx)
-                .await
-                .context(ExecLogicalPlanSnafu)
-        }
+            self.exec_plan_with_timeout(plan, query_ctx.clone()).await
+        };
+
+        result.and_then(|output| query_interceptor.post_execute(output, query_ctx))
     }
 
     #[tracing::instrument(skip_all, name = "SqlQueryHandler::do_promql_query")]
@@ -676,10 +721,19 @@ impl Instance {
     ) -> Result<Option<DescribeResult>> {
         ensure!(!self.is_suspended(), error::SuspendedSnafu);
 
-        if matches!(
-            stmt,
-            Statement::Insert(_) | Statement::Query(_) | Statement::Delete(_)
-        ) {
+        // EXPLAIN / EXPLAIN ANALYZE wrap an inner statement; describe them when the
+        // wrapped statement is something we already plan (so that bind parameters
+        // in the inner query get their types inferred). See #8029.
+        let is_inner_plannable = |s: &Statement| {
+            matches!(
+                s,
+                Statement::Insert(_) | Statement::Query(_) | Statement::Delete(_)
+            )
+        };
+        let plannable = is_inner_plannable(&stmt)
+            || matches!(&stmt, Statement::Explain(explain) if is_inner_plannable(explain.statement.as_ref()));
+
+        if plannable {
             self.plugins
                 .get::<PermissionCheckerRef>()
                 .as_ref()
@@ -726,11 +780,11 @@ impl SqlQueryHandler for Instance {
 
     async fn do_exec_plan(
         &self,
-        stmt: Option<Statement>,
         plan: LogicalPlan,
+        stmt: Option<Statement>,
         query_ctx: QueryContextRef,
     ) -> server_error::Result<Output> {
-        self.do_exec_plan_inner(stmt, plan, query_ctx)
+        self.do_exec_plan_inner(plan, stmt, query_ctx)
             .await
             .map_err(BoxedError::new)
             .context(server_error::ExecutePlanSnafu)
@@ -1118,13 +1172,8 @@ fn validate_database(name: &ObjectName, query_ctx: &QueryContextRef) -> Result<(
         .context(SqlExecInterceptedSnafu)
 }
 
-// Create a query ticket and slow query timer if the statement is a query or readonly statement.
-fn should_capture_statement(stmt: Option<&Statement>) -> bool {
-    if let Some(stmt) = stmt {
-        matches!(stmt, Statement::Query(_)) || stmt.is_readonly()
-    } else {
-        false
-    }
+fn is_readonly_plan(plan: &LogicalPlan) -> bool {
+    !matches!(plan, LogicalPlan::Dml(_) | LogicalPlan::Ddl(_))
 }
 
 #[cfg(test)]
diff --git a/src/frontend/src/instance/dashboard.rs b/src/frontend/src/instance/dashboard.rs
new file mode 100644
index 0000000000..5b83a31f20
--- /dev/null
+++ b/src/frontend/src/instance/dashboard.rs
@@ -0,0 +1,402 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::HashMap;
+use std::sync::Arc;
+
+use api::v1::value::ValueData;
+use api::v1::{
+    ColumnDataType, ColumnDef, ColumnSchema as PbColumnSchema, Row, RowInsertRequest,
+    RowInsertRequests, Rows, SemanticType,
+};
+use async_trait::async_trait;
+use common_catalog::consts::{DEFAULT_PRIVATE_SCHEMA_NAME, default_engine};
+use common_error::ext::BoxedError;
+use common_query::OutputData;
+use common_recordbatch::util as record_util;
+use common_telemetry::info;
+use common_time::FOREVER;
+use datafusion::datasource::DefaultTableSource;
+use datafusion::logical_expr::col;
+use datafusion::sql::TableReference;
+use datafusion_expr::{DmlStatement, LogicalPlan, lit};
+use datatypes::arrow::array::{Array, AsArray};
+use servers::error::{
+    CollectRecordbatchSnafu, DataFusionSnafu, ExecuteQuerySnafu, NotSupportedSnafu,
+    TableNotFoundSnafu,
+};
+use servers::query_handler::DashboardDefinition;
+use session::context::{QueryContextBuilder, QueryContextRef};
+use snafu::{OptionExt, ResultExt};
+use table::TableRef;
+use table::metadata::TableInfo;
+use table::requests::TTL_KEY;
+use table::table::adapter::DfTableProviderAdapter;
+
+use crate::instance::Instance;
+
+pub const DASHBOARD_TABLE_NAME: &str = "dashboard";
+pub const DASHBOARD_TABLE_NAME_COLUMN_NAME: &str = "name";
+pub const DASHBOARD_TABLE_DEFINITION_COLUMN_NAME: &str = "definition";
+pub const DASHBOARD_TABLE_CREATED_AT_COLUMN_NAME: &str = "created_at";
+
+impl Instance {
+    /// Build a schema for dashboard table.
+    /// Returns the (time index, primary keys, column) definitions.
+    fn build_dashboard_schema() -> (String, Vec<String>, Vec<ColumnDef>) {
+        (
+            DASHBOARD_TABLE_CREATED_AT_COLUMN_NAME.to_string(),
+            vec![DASHBOARD_TABLE_NAME_COLUMN_NAME.to_string()],
+            vec![
+                ColumnDef {
+                    name: DASHBOARD_TABLE_NAME_COLUMN_NAME.to_string(),
+                    data_type: ColumnDataType::String as i32,
+                    is_nullable: false,
+                    default_constraint: vec![],
+                    semantic_type: SemanticType::Tag as i32,
+                    comment: String::new(),
+                    datatype_extension: None,
+                    options: None,
+                },
+                ColumnDef {
+                    name: DASHBOARD_TABLE_DEFINITION_COLUMN_NAME.to_string(),
+                    data_type: ColumnDataType::String as i32,
+                    is_nullable: false,
+                    default_constraint: vec![],
+                    semantic_type: SemanticType::Field as i32,
+                    comment: String::new(),
+                    datatype_extension: None,
+                    options: None,
+                },
+                ColumnDef {
+                    name: DASHBOARD_TABLE_CREATED_AT_COLUMN_NAME.to_string(),
+                    data_type: ColumnDataType::TimestampNanosecond as i32,
+                    is_nullable: false,
+                    default_constraint: vec![],
+                    semantic_type: SemanticType::Timestamp as i32,
+                    comment: String::new(),
+                    datatype_extension: None,
+                    options: None,
+                },
+            ],
+        )
+    }
+
+    /// Build a column schemas for inserting a row into the dashboard table.
+    fn build_dashboard_insert_column_schemas() -> Vec<PbColumnSchema> {
+        vec![
+            PbColumnSchema {
+                column_name: DASHBOARD_TABLE_NAME_COLUMN_NAME.to_string(),
+                datatype: ColumnDataType::String.into(),
+                semantic_type: SemanticType::Tag.into(),
+                ..Default::default()
+            },
+            PbColumnSchema {
+                column_name: DASHBOARD_TABLE_DEFINITION_COLUMN_NAME.to_string(),
+                datatype: ColumnDataType::String.into(),
+                semantic_type: SemanticType::Field.into(),
+                ..Default::default()
+            },
+            PbColumnSchema {
+                column_name: DASHBOARD_TABLE_CREATED_AT_COLUMN_NAME.to_string(),
+                datatype: ColumnDataType::TimestampNanosecond.into(),
+                semantic_type: SemanticType::Timestamp.into(),
+                ..Default::default()
+            },
+        ]
+    }
+
+    fn dashboard_query_ctx(table_info: &TableInfo) -> QueryContextRef {
+        QueryContextBuilder::default()
+            .current_catalog(table_info.catalog_name.clone())
+            .current_schema(table_info.schema_name.clone())
+            .build()
+            .into()
+    }
+
+    async fn create_dashboard_table_if_not_exists(
+        &self,
+        ctx: QueryContextRef,
+    ) -> servers::error::Result<TableRef> {
+        let catalog = ctx.current_catalog();
+
+        if let Some(table) = self
+            .catalog_manager
+            .table(
+                catalog,
+                DEFAULT_PRIVATE_SCHEMA_NAME,
+                DASHBOARD_TABLE_NAME,
+                Some(&ctx),
+            )
+            .await?
+        {
+            return Ok(table);
+        }
+
+        let (time_index, primary_keys, column_defs) = Self::build_dashboard_schema();
+
+        let mut table_options = HashMap::new();
+        table_options.insert(TTL_KEY.to_string(), FOREVER.to_string());
+
+        let mut create_table_expr = api::v1::CreateTableExpr {
+            catalog_name: catalog.to_string(),
+            schema_name: DEFAULT_PRIVATE_SCHEMA_NAME.to_string(),
+            table_name: DASHBOARD_TABLE_NAME.to_string(),
+            desc: "GreptimeDB dashboard table".to_string(),
+            column_defs,
+            time_index,
+            primary_keys,
+            create_if_not_exists: true,
+            table_options,
+            table_id: None,
+            engine: default_engine().to_string(),
+        };
+
+        self.statement_executor
+            .create_table_inner(&mut create_table_expr, None, ctx.clone())
+            .await
+            .map_err(BoxedError::new)
+            .context(ExecuteQuerySnafu)?;
+
+        let table = self
+            .catalog_manager
+            .table(
+                catalog,
+                DEFAULT_PRIVATE_SCHEMA_NAME,
+                DASHBOARD_TABLE_NAME,
+                Some(&ctx),
+            )
+            .await?
+            .context(TableNotFoundSnafu {
+                catalog: catalog.to_string(),
+                schema: DEFAULT_PRIVATE_SCHEMA_NAME.to_string(),
+                table: DASHBOARD_TABLE_NAME.to_string(),
+            })?;
+
+        Ok(table)
+    }
+
+    /// Insert a dashboard into the dashboard table.
+    async fn insert_dashboard(
+        &self,
+        name: &str,
+        definition: &str,
+        query_ctx: QueryContextRef,
+    ) -> servers::error::Result<()> {
+        let table = self
+            .create_dashboard_table_if_not_exists(query_ctx.clone())
+            .await?;
+        let table_info = table.table_info();
+
+        let insert = RowInsertRequest {
+            table_name: DASHBOARD_TABLE_NAME.to_string(),
+            rows: Some(Rows {
+                schema: Self::build_dashboard_insert_column_schemas(),
+                rows: vec![Row {
+                    values: vec![
+                        ValueData::StringValue(name.to_string()).into(),
+                        ValueData::StringValue(definition.to_string()).into(),
+                        ValueData::TimestampNanosecondValue(0).into(),
+                    ],
+                }],
+            }),
+        };
+
+        let requests = RowInsertRequests {
+            inserts: vec![insert],
+        };
+
+        let output = self
+            .inserter
+            .handle_row_inserts(
+                requests,
+                Self::dashboard_query_ctx(&table_info),
+                &self.statement_executor,
+                false,
+                false,
+            )
+            .await
+            .map_err(BoxedError::new)
+            .context(ExecuteQuerySnafu)?;
+
+        info!(
+            "Insert dashboard success, name: {}, table: {}, output: {:?}",
+            name,
+            table_info.full_table_name(),
+            output
+        );
+
+        Ok(())
+    }
+
+    /// List all dashboards.
+    async fn list_dashboards(
+        &self,
+        query_ctx: QueryContextRef,
+    ) -> servers::error::Result<Vec<DashboardDefinition>> {
+        let table = if let Some(table) = self
+            .catalog_manager
+            .table(
+                query_ctx.current_catalog(),
+                DEFAULT_PRIVATE_SCHEMA_NAME,
+                DASHBOARD_TABLE_NAME,
+                Some(&query_ctx),
+            )
+            .await?
+        {
+            table
+        } else {
+            return Ok(vec![]);
+        };
+
+        let table_info = table.table_info();
+
+        let dataframe = self
+            .query_engine
+            .read_table(table.clone())
+            .map_err(BoxedError::new)
+            .context(ExecuteQuerySnafu)?;
+
+        let dataframe = dataframe
+            .select_columns(&[
+                DASHBOARD_TABLE_NAME_COLUMN_NAME,
+                DASHBOARD_TABLE_DEFINITION_COLUMN_NAME,
+            ])
+            .context(DataFusionSnafu)?;
+
+        let plan = dataframe.into_parts().1;
+
+        let output = self
+            .query_engine
+            .execute(plan, Self::dashboard_query_ctx(&table_info))
+            .await
+            .map_err(BoxedError::new)
+            .context(ExecuteQuerySnafu)?;
+
+        let stream = match output.data {
+            OutputData::Stream(stream) => stream,
+            OutputData::RecordBatches(record_batches) => record_batches.as_stream(),
+            _ => unreachable!(),
+        };
+
+        let records = record_util::collect(stream)
+            .await
+            .context(CollectRecordbatchSnafu)?;
+
+        let mut dashboards = Vec::new();
+
+        for r in &records {
+            let name_column = r.column(0);
+            let definition_column = r.column(1);
+
+            let name = name_column
+                .as_string_opt::<i32>()
+                .context(NotSupportedSnafu {
+                    feat: "Invalid data type for greptime_private.dashboard.name",
+                })?;
+
+            let definition =
+                definition_column
+                    .as_string_opt::<i32>()
+                    .context(NotSupportedSnafu {
+                        feat: "Invalid data type for greptime_private.dashboard.definition",
+                    })?;
+
+            for i in 0..name.len() {
+                dashboards.push(DashboardDefinition {
+                    name: name.value(i).to_string(),
+                    definition: definition.value(i).to_string(),
+                });
+            }
+        }
+
+        Ok(dashboards)
+    }
+
+    /// Delete a dashboard by name.
+    async fn delete_dashboard(
+        &self,
+        name: &str,
+        query_ctx: QueryContextRef,
+    ) -> servers::error::Result<()> {
+        let table = self
+            .create_dashboard_table_if_not_exists(query_ctx.clone())
+            .await?;
+        let table_info = table.table_info();
+
+        let dataframe = self
+            .query_engine
+            .read_table(table.clone())
+            .map_err(BoxedError::new)
+            .context(ExecuteQuerySnafu)?;
+
+        let name_condition = col(DASHBOARD_TABLE_NAME_COLUMN_NAME).eq(lit(name));
+
+        let dataframe = dataframe.filter(name_condition).context(DataFusionSnafu)?;
+
+        let table_name = TableReference::full(
+            table_info.catalog_name.clone(),
+            table_info.schema_name.clone(),
+            table_info.name.clone(),
+        );
+
+        let table_provider = Arc::new(DfTableProviderAdapter::new(table.clone()));
+        let table_source = Arc::new(DefaultTableSource::new(table_provider));
+
+        let stmt = DmlStatement::new(
+            table_name,
+            table_source,
+            datafusion_expr::WriteOp::Delete,
+            Arc::new(dataframe.into_parts().1),
+        );
+
+        let plan = LogicalPlan::Dml(stmt);
+
+        let output = self
+            .query_engine
+            .execute(plan, Self::dashboard_query_ctx(&table_info))
+            .await
+            .map_err(BoxedError::new)
+            .context(ExecuteQuerySnafu)?;
+
+        info!(
+            "Delete dashboard success, name: {}, table: {}, output: {:?}",
+            name,
+            table_info.full_table_name(),
+            output
+        );
+
+        Ok(())
+    }
+}
+
+#[async_trait]
+impl servers::query_handler::DashboardHandler for Instance {
+    async fn save(
+        &self,
+        name: &str,
+        definition: &str,
+        ctx: QueryContextRef,
+    ) -> servers::error::Result<()> {
+        self.insert_dashboard(name, definition, ctx).await
+    }
+
+    async fn list(&self, ctx: QueryContextRef) -> servers::error::Result<Vec<DashboardDefinition>> {
+        self.list_dashboards(ctx).await
+    }
+
+    async fn delete(&self, name: &str, ctx: QueryContextRef) -> servers::error::Result<()> {
+        self.delete_dashboard(name, ctx).await
+    }
+}
diff --git a/src/frontend/src/instance/grpc.rs b/src/frontend/src/instance/grpc.rs
index c4191145f8..8d18293cb8 100644
--- a/src/frontend/src/instance/grpc.rs
+++ b/src/frontend/src/instance/grpc.rs
@@ -27,7 +27,6 @@ use api::v1::{
 use async_stream::try_stream;
 use async_trait::async_trait;
 use auth::{PermissionChecker, PermissionCheckerRef, PermissionReq};
-use common_base::AffectedRows;
 use common_error::ext::BoxedError;
 use common_grpc::flight::do_put::DoPutResponse;
 use common_query::Output;
@@ -131,7 +130,7 @@ impl GrpcQueryHandler for Instance {
                                 .await
                                 .context(SubstraitDecodeLogicalPlanSnafu)?;
                             let output =
-                                self.do_exec_plan_inner(None, logical_plan, ctx.clone()).await?;
+                                self.do_exec_plan_inner(logical_plan, None, ctx.clone()).await?;
 
                             attach_timer(output, timer)
                         }
@@ -260,62 +259,6 @@ impl GrpcQueryHandler for Instance {
             .context(server_error::ExecuteGrpcQuerySnafu)
     }
 
-    async fn put_record_batch(
-        &self,
-        request: servers::grpc::flight::PutRecordBatchRequest,
-        table_ref: &mut Option<TableRef>,
-        ctx: QueryContextRef,
-    ) -> server_error::Result<AffectedRows> {
-        let result: Result<AffectedRows> = async {
-            let table = if let Some(table) = table_ref {
-                table.clone()
-            } else {
-                let table = self
-                    .catalog_manager()
-                    .table(
-                        &request.table_name.catalog_name,
-                        &request.table_name.schema_name,
-                        &request.table_name.table_name,
-                        None,
-                    )
-                    .await
-                    .context(CatalogSnafu)?
-                    .with_context(|| TableNotFoundSnafu {
-                        table_name: request.table_name.to_string(),
-                    })?;
-                *table_ref = Some(table.clone());
-                table
-            };
-
-            let interceptor_ref = self.plugins.get::<GrpcQueryInterceptorRef<Error>>();
-            let interceptor = interceptor_ref.as_ref();
-            interceptor.pre_bulk_insert(table.clone(), ctx.clone())?;
-
-            self.plugins
-                .get::<PermissionCheckerRef>()
-                .as_ref()
-                .check_permission(ctx.current_user(), PermissionReq::BulkInsert)
-                .context(PermissionSnafu)?;
-
-            // do we check limit for bulk insert?
-
-            self.inserter
-                .handle_bulk_insert(
-                    table,
-                    request.flight_data,
-                    request.record_batch,
-                    request.schema_bytes,
-                )
-                .await
-                .context(TableOperationSnafu)
-        }
-        .await;
-
-        result
-            .map_err(BoxedError::new)
-            .context(server_error::ExecuteGrpcRequestSnafu)
-    }
-
     fn handle_put_record_batch_stream(
         &self,
         stream: servers::grpc::flight::PutRecordBatchRequestStream,
@@ -524,7 +467,7 @@ impl Instance {
         let optimized_plan = state.optimize(&analyzed_plan).context(DataFusionSnafu)?;
 
         let output = self
-            .do_exec_plan_inner(None, optimized_plan, ctx.clone())
+            .do_exec_plan_inner(optimized_plan, None, ctx.clone())
             .await?;
 
         Ok(attach_timer(output, timer))
diff --git a/src/frontend/src/instance/influxdb.rs b/src/frontend/src/instance/influxdb.rs
index 0c63688262..fe5fdeac77 100644
--- a/src/frontend/src/instance/influxdb.rs
+++ b/src/frontend/src/instance/influxdb.rs
@@ -21,9 +21,7 @@ use client::Output;
 use common_error::ext::BoxedError;
 use common_time::Timestamp;
 use common_time::timestamp::TimeUnit;
-use servers::error::{
-    AuthSnafu, CatalogSnafu, Error, TimestampOverflowSnafu, UnexpectedResultSnafu,
-};
+use servers::error::{AuthSnafu, Error, TimestampOverflowSnafu, UnexpectedResultSnafu};
 use servers::influxdb::InfluxdbRequest;
 use servers::interceptor::{LineProtocolInterceptor, LineProtocolInterceptorRef};
 use servers::query_handler::InfluxdbLineProtocolHandler;
@@ -92,8 +90,7 @@ impl InfluxdbLineTimestampAligner<'_> {
                     &insert.table_name,
                     Some(query_context),
                 )
-                .await
-                .context(CatalogSnafu)?
+                .await?
                 .map(|x| x.schema())
                 .and_then(|schema| {
                     schema.timestamp_column().map(|col| {
diff --git a/src/frontend/src/instance/jaeger.rs b/src/frontend/src/instance/jaeger.rs
index 607ed80098..e500de7a85 100644
--- a/src/frontend/src/instance/jaeger.rs
+++ b/src/frontend/src/instance/jaeger.rs
@@ -38,8 +38,7 @@ use datafusion_expr::{Expr, ExprFunctionExt, SortExpr, col, lit, lit_timestamp_n
 use query::QueryEngineRef;
 use serde_json::Value as JsonValue;
 use servers::error::{
-    CatalogSnafu, CollectRecordbatchSnafu, DataFusionSnafu, Result as ServerResult,
-    TableNotFoundSnafu,
+    CollectRecordbatchSnafu, DataFusionSnafu, Result as ServerResult, TableNotFoundSnafu,
 };
 use servers::http::jaeger::{JAEGER_QUERY_TABLE_NAME_KEY, QueryTraceParams, TraceUserAgent};
 use servers::otlp::trace::{
@@ -336,8 +335,7 @@ async fn query_trace_table(
             table_name,
             Some(&ctx),
         )
-        .await
-        .context(CatalogSnafu)?
+        .await?
         .with_context(|| TableNotFoundSnafu {
             table: table_name,
             catalog: ctx.current_catalog(),
@@ -425,8 +423,7 @@ async fn get_table(
             table_name,
             Some(&ctx),
         )
-        .await
-        .context(CatalogSnafu)?
+        .await?
         .with_context(|| TableNotFoundSnafu {
             table: table_name,
             catalog: ctx.current_catalog(),
diff --git a/src/frontend/src/instance/otlp.rs b/src/frontend/src/instance/otlp.rs
index 52df274780..434a413fed 100644
--- a/src/frontend/src/instance/otlp.rs
+++ b/src/frontend/src/instance/otlp.rs
@@ -14,12 +14,19 @@
 
 use std::sync::Arc;
 
+use api::helper::ColumnDataTypeWrapper;
+use api::v1::alter_table_expr::Kind;
+use api::v1::{
+    AlterTableExpr, ColumnDataType, ModifyColumnType, ModifyColumnTypes, RowInsertRequests,
+};
 use async_trait::async_trait;
 use auth::{PermissionChecker, PermissionCheckerRef, PermissionReq};
 use client::Output;
-use common_error::ext::BoxedError;
+use common_error::ext::{BoxedError, ErrorExt};
+use common_error::status_code::StatusCode;
 use common_query::prelude::GREPTIME_PHYSICAL_TABLE;
-use common_telemetry::tracing;
+use common_telemetry::{tracing, warn};
+use itertools::Itertools;
 use opentelemetry_proto::tonic::collector::logs::v1::ExportLogsServiceRequest;
 use opentelemetry_proto::tonic::collector::trace::v1::ExportTraceServiceRequest;
 use otel_arrow_rust::proto::opentelemetry::collector::metrics::v1::ExportMetricsServiceRequest;
@@ -28,13 +35,62 @@ use servers::error::{self, AuthSnafu, Result as ServerResult};
 use servers::http::prom_store::PHYSICAL_TABLE_PARAM;
 use servers::interceptor::{OpenTelemetryProtocolInterceptor, OpenTelemetryProtocolInterceptorRef};
 use servers::otlp;
-use servers::query_handler::{OpenTelemetryProtocolHandler, PipelineHandlerRef};
+use servers::otlp::trace::TraceAuxData;
+use servers::otlp::trace::coerce::{coerce_value_data, trace_value_datatype};
+use servers::otlp::trace::span::{TraceSpan, TraceSpanGroup};
+use servers::query_handler::{
+    OpenTelemetryProtocolHandler, PipelineHandlerRef, TraceIngestOutcome,
+};
 use session::context::QueryContextRef;
-use snafu::ResultExt;
+use snafu::{IntoError, ResultExt};
 use table::requests::{OTLP_METRIC_COMPAT_KEY, OTLP_METRIC_COMPAT_PROM};
 
 use crate::instance::Instance;
-use crate::metrics::{OTLP_LOGS_ROWS, OTLP_METRICS_ROWS, OTLP_TRACES_ROWS};
+use crate::instance::otlp::trace_semconv::trace_semconv_fixed_type;
+use crate::instance::otlp::trace_types::{
+    PendingTraceColumnRewrite, choose_trace_reconcile_decision, enrich_trace_reconcile_error,
+    is_trace_reconcile_candidate_type, push_observed_trace_type, validate_trace_column_rewrites,
+};
+use crate::metrics::{
+    OTLP_LOGS_ROWS, OTLP_METRICS_ROWS, OTLP_TRACES_FAILURE_COUNT, OTLP_TRACES_ROWS,
+};
+
+pub mod trace_semconv;
+pub mod trace_types;
+
+const TRACE_INGEST_CHUNK_SIZE: usize = 64;
+const TRACE_FAILURE_MESSAGE_LIMIT: usize = 4;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum ChunkFailureReaction {
+    RetryPerSpan,
+    DiscardChunk,
+    Propagate,
+}
+
+impl ChunkFailureReaction {
+    fn as_metric_label(self) -> &'static str {
+        match self {
+            Self::RetryPerSpan => "retry_per_span",
+            Self::DiscardChunk => "discard_chunk",
+            Self::Propagate => "propagate_failure",
+        }
+    }
+}
+
+struct TraceChunkIngestContext<'a> {
+    pipeline_handler: PipelineHandlerRef,
+    pipeline: &'a PipelineWay,
+    pipeline_params: &'a GreptimePipelineParams,
+    table_name: &'a str,
+    is_trace_v1_model: bool,
+}
+
+struct TraceIngestState {
+    aux_data: TraceAuxData,
+    outcome: TraceIngestOutcome,
+    failure_messages: Vec<String>,
+}
 
 #[async_trait]
 impl OpenTelemetryProtocolHandler for Instance {
@@ -110,7 +166,7 @@ impl OpenTelemetryProtocolHandler for Instance {
         pipeline_params: GreptimePipelineParams,
         table_name: String,
         ctx: QueryContextRef,
-    ) -> ServerResult<Output> {
+    ) -> ServerResult<TraceIngestOutcome> {
         self.plugins
             .get::<PermissionCheckerRef>()
             .as_ref()
@@ -122,30 +178,16 @@ impl OpenTelemetryProtocolHandler for Instance {
             .get::<OpenTelemetryProtocolInterceptorRef<servers::error::Error>>();
         interceptor_ref.pre_execute(ctx.clone())?;
 
-        let is_trace_v1_model = matches!(pipeline, PipelineWay::OtlpTraceDirectV1);
-
-        let (requests, rows) = otlp::trace::to_grpc_insert_requests(
-            request,
-            pipeline,
-            pipeline_params,
-            table_name,
-            &ctx,
+        let spans = otlp::trace::span::parse(request);
+        self.ingest_trace_spans(
             pipeline_handler,
-        )?;
-
-        OTLP_TRACES_ROWS.inc_by(rows as u64);
-
-        if is_trace_v1_model {
-            self.handle_trace_inserts(requests, ctx)
-                .await
-                .map_err(BoxedError::new)
-                .context(error::ExecuteGrpcQuerySnafu)
-        } else {
-            self.handle_log_inserts(requests, ctx)
-                .await
-                .map_err(BoxedError::new)
-                .context(error::ExecuteGrpcQuerySnafu)
-        }
+            &pipeline,
+            &pipeline_params,
+            table_name,
+            spans,
+            ctx,
+        )
+        .await
     }
 
     #[tracing::instrument(skip_all)]
@@ -200,3 +242,740 @@ impl OpenTelemetryProtocolHandler for Instance {
         Ok(outputs)
     }
 }
+
+impl Instance {
+    /// Ingest OTLP trace spans with chunk-level writes and span-level fallback on
+    /// deterministic chunk failures.
+    async fn ingest_trace_spans(
+        &self,
+        pipeline_handler: PipelineHandlerRef,
+        pipeline: &PipelineWay,
+        pipeline_params: &GreptimePipelineParams,
+        table_name: String,
+        groups: Vec<TraceSpanGroup>,
+        ctx: QueryContextRef,
+    ) -> ServerResult<TraceIngestOutcome> {
+        let is_trace_v1_model = matches!(pipeline, PipelineWay::OtlpTraceDirectV1);
+        let ingest_ctx = TraceChunkIngestContext {
+            pipeline_handler,
+            pipeline,
+            pipeline_params,
+            table_name: &table_name,
+            is_trace_v1_model,
+        };
+        let mut ingest_state = TraceIngestState {
+            aux_data: TraceAuxData::default(),
+            outcome: TraceIngestOutcome::default(),
+            failure_messages: Vec::new(),
+        };
+
+        for group in groups {
+            let chunks = group
+                .spans
+                .into_iter()
+                .chunks(TRACE_INGEST_CHUNK_SIZE)
+                .into_iter()
+                .map(|chunk| chunk.collect::<Vec<_>>())
+                .collect::<Vec<_>>();
+            for chunk in chunks {
+                self.ingest_trace_chunk(&ingest_ctx, chunk, ctx.clone(), &mut ingest_state)
+                    .await?;
+            }
+        }
+
+        OTLP_TRACES_ROWS.inc_by(ingest_state.outcome.accepted_spans as u64);
+
+        if !ingest_state.aux_data.is_empty() {
+            // Auxiliary trace tables are derived from spans whose main-table
+            // writes are already confirmed, so they never create new accepted
+            // spans and they do not affect rejected span counts.
+            let (aux_requests, _) = otlp::trace::to_grpc_insert_requests_for_aux_tables(
+                std::mem::take(&mut ingest_state.aux_data),
+                ingest_ctx.pipeline,
+                ingest_ctx.table_name,
+            )?;
+
+            if !aux_requests.inserts.is_empty() {
+                match self
+                    .insert_trace_requests(aux_requests, ingest_ctx.is_trace_v1_model, ctx)
+                    .await
+                {
+                    Ok(output) => {
+                        Self::add_trace_write_cost(&mut ingest_state.outcome, output.meta.cost);
+                    }
+                    Err(err) => {
+                        Self::push_trace_failure_message(
+                            &mut ingest_state.failure_messages,
+                            "aux_table_update_failed",
+                            format!(
+                                "Auxiliary trace tables were not fully updated ({})",
+                                err.status_code().as_ref()
+                            ),
+                        );
+                    }
+                }
+            }
+        }
+
+        ingest_state.outcome.error_message = Self::finish_trace_failure_message(
+            ingest_state.outcome.accepted_spans,
+            ingest_state.outcome.rejected_spans,
+            ingest_state.failure_messages,
+        );
+
+        Ok(ingest_state.outcome)
+    }
+
+    /// Ingest one owned trace chunk so successful spans can be moved into the
+    /// accepted set without extra cloning.
+    async fn ingest_trace_chunk(
+        &self,
+        ingest_ctx: &TraceChunkIngestContext<'_>,
+        chunk: Vec<TraceSpan>,
+        ctx: QueryContextRef,
+        ingest_state: &mut TraceIngestState,
+    ) -> ServerResult<()> {
+        // Try the fast path first so healthy batches keep their original
+        // throughput and write amplification stays low.
+        let (requests, chunk_rows) = otlp::trace::to_grpc_insert_requests_from_spans(
+            &chunk,
+            ingest_ctx.pipeline,
+            ingest_ctx.pipeline_params,
+            ingest_ctx.table_name,
+            &ctx,
+            ingest_ctx.pipeline_handler.clone(),
+        )?;
+
+        match self
+            .insert_trace_requests(requests, ingest_ctx.is_trace_v1_model, ctx.clone())
+            .await
+        {
+            Ok(output) => {
+                Self::add_trace_write_cost(&mut ingest_state.outcome, output.meta.cost);
+                ingest_state.outcome.accepted_spans += chunk_rows;
+                for span in &chunk {
+                    ingest_state.aux_data.observe_span(span);
+                }
+            }
+            Err(err) => match Self::classify_trace_chunk_failure(err.status_code()) {
+                ChunkFailureReaction::RetryPerSpan => {
+                    Self::push_trace_failure_message(
+                        &mut ingest_state.failure_messages,
+                        ChunkFailureReaction::RetryPerSpan.as_metric_label(),
+                        format!("Chunk fallback triggered by {}", err.status_code().as_ref()),
+                    );
+                    // Only deterministic failures are retried span by span.
+                    // This includes schemaless table or column creation paths for
+                    // trace ingestion. Ambiguous failures are handled below
+                    // without retrying because the chunk may already have been
+                    // ingested.
+                    self.ingest_trace_chunk_span_by_span(
+                        ingest_ctx,
+                        chunk,
+                        ctx.clone(),
+                        ingest_state,
+                    )
+                    .await?;
+                }
+                ChunkFailureReaction::DiscardChunk => {
+                    ingest_state.outcome.rejected_spans += chunk.len();
+                    Self::push_trace_failure_message(
+                        &mut ingest_state.failure_messages,
+                        ChunkFailureReaction::DiscardChunk.as_metric_label(),
+                        format!(
+                            "Discarded {} spans after ambiguous chunk failure ({})",
+                            chunk.len(),
+                            err.status_code().as_ref()
+                        ),
+                    );
+                    // TODO(shuiyisong): Add an idempotent retry-safe recovery path for
+                    // ambiguous chunk failures such as timeout-like errors.
+                }
+                // Retryable or ambiguous failures must fail the request instead of
+                // becoming partial success. This path is not retry-safe because the
+                // chunk may already have been committed before the error surfaced.
+                ChunkFailureReaction::Propagate => {
+                    Self::push_trace_failure_message(
+                        &mut ingest_state.failure_messages,
+                        ChunkFailureReaction::Propagate.as_metric_label(),
+                        format!(
+                            "Propagating retryable chunk failure ({})",
+                            err.status_code().as_ref()
+                        ),
+                    );
+                    return Err(err);
+                }
+            },
+        }
+
+        Ok(())
+    }
+
+    /// Retry spans one by one only after a deterministic chunk failure.
+    async fn ingest_trace_chunk_span_by_span(
+        &self,
+        ingest_ctx: &TraceChunkIngestContext<'_>,
+        chunk: Vec<TraceSpan>,
+        ctx: QueryContextRef,
+        ingest_state: &mut TraceIngestState,
+    ) -> ServerResult<()> {
+        for span in chunk {
+            let (requests, rows) = otlp::trace::to_grpc_insert_requests_from_spans(
+                std::slice::from_ref(&span),
+                ingest_ctx.pipeline,
+                ingest_ctx.pipeline_params,
+                ingest_ctx.table_name,
+                &ctx,
+                ingest_ctx.pipeline_handler.clone(),
+            )?;
+
+            match self
+                .insert_trace_requests(requests, ingest_ctx.is_trace_v1_model, ctx.clone())
+                .await
+            {
+                Ok(output) => {
+                    Self::add_trace_write_cost(&mut ingest_state.outcome, output.meta.cost);
+                    ingest_state.outcome.accepted_spans += rows;
+                    ingest_state.aux_data.observe_span(&span);
+                }
+                Err(err) => {
+                    if Self::should_propagate_trace_span_failure(err.status_code()) {
+                        Self::push_trace_failure_message(
+                            &mut ingest_state.failure_messages,
+                            ChunkFailureReaction::Propagate.as_metric_label(),
+                            format!(
+                                "Propagating retryable span failure for {}:{} ({})",
+                                span.trace_id,
+                                span.span_id,
+                                err.status_code().as_ref()
+                            ),
+                        );
+                        return Err(err);
+                    }
+
+                    ingest_state.outcome.rejected_spans += 1;
+                    Self::push_trace_failure_message(
+                        &mut ingest_state.failure_messages,
+                        "span_rejected",
+                        format!(
+                            "Rejected span {}:{} ({})",
+                            span.trace_id,
+                            span.span_id,
+                            err.status_code().as_ref()
+                        ),
+                    );
+                }
+            }
+        }
+
+        Ok(())
+    }
+
+    /// Reconcile and insert one trace request batch.
+    async fn insert_trace_requests(
+        &self,
+        mut requests: RowInsertRequests,
+        is_trace_v1_model: bool,
+        ctx: QueryContextRef,
+    ) -> ServerResult<Output> {
+        if is_trace_v1_model {
+            self.reconcile_trace_column_types(&mut requests, &ctx)
+                .await?;
+            self.handle_trace_inserts(requests, ctx)
+                .await
+                .map_err(BoxedError::new)
+                .context(error::ExecuteGrpcQuerySnafu)
+        } else {
+            self.handle_log_inserts(requests, ctx)
+                .await
+                .map_err(BoxedError::new)
+                .context(error::ExecuteGrpcQuerySnafu)
+        }
+    }
+
+    fn classify_trace_chunk_failure(status: StatusCode) -> ChunkFailureReaction {
+        match status {
+            StatusCode::InvalidArguments
+            | StatusCode::InvalidSyntax
+            | StatusCode::Unsupported
+            | StatusCode::TableNotFound
+            | StatusCode::TableColumnNotFound => ChunkFailureReaction::RetryPerSpan,
+            StatusCode::DatabaseNotFound => ChunkFailureReaction::DiscardChunk,
+            StatusCode::Cancelled | StatusCode::DeadlineExceeded => ChunkFailureReaction::Propagate,
+            _ if status.is_retryable() => ChunkFailureReaction::Propagate,
+            _ => ChunkFailureReaction::DiscardChunk,
+        }
+    }
+
+    fn should_propagate_trace_span_failure(status: StatusCode) -> bool {
+        matches!(
+            Self::classify_trace_chunk_failure(status),
+            ChunkFailureReaction::Propagate
+        )
+    }
+
+    fn add_trace_write_cost(outcome: &mut TraceIngestOutcome, cost: usize) {
+        outcome.write_cost += cost;
+    }
+
+    fn push_trace_failure_message(messages: &mut Vec<String>, label: &str, message: String) {
+        OTLP_TRACES_FAILURE_COUNT.with_label_values(&[label]).inc();
+
+        if messages.len() < TRACE_FAILURE_MESSAGE_LIMIT {
+            messages.push(message);
+        } else if messages.len() == TRACE_FAILURE_MESSAGE_LIMIT {
+            tracing::debug!(
+                label,
+                limit = TRACE_FAILURE_MESSAGE_LIMIT,
+                "Trace ingest failure message limit reached; suppressing additional failure details"
+            );
+        }
+    }
+
+    fn finish_trace_failure_message(
+        accepted_spans: usize,
+        rejected_spans: usize,
+        messages: Vec<String>,
+    ) -> Option<String> {
+        if rejected_spans == 0 && messages.is_empty() {
+            return None;
+        }
+
+        let mut summary = format!(
+            "Accepted {} spans, rejected {} spans",
+            accepted_spans, rejected_spans
+        );
+
+        if !messages.is_empty() {
+            summary.push_str(": ");
+            summary.push_str(&messages.join("; "));
+        }
+
+        Some(summary)
+    }
+
+    /// Widen existing trace table columns to Float64 before request rewrite.
+    async fn alter_trace_table_columns_to_float64(
+        &self,
+        ctx: &QueryContextRef,
+        table_name: &str,
+        column_names: &[String],
+    ) -> ServerResult<()> {
+        let catalog_name = ctx.current_catalog().to_string();
+        let schema_name = ctx.current_schema();
+        let alter_expr = AlterTableExpr {
+            catalog_name: catalog_name.clone(),
+            schema_name: schema_name.clone(),
+            table_name: table_name.to_string(),
+            kind: Some(Kind::ModifyColumnTypes(ModifyColumnTypes {
+                modify_column_types: column_names
+                    .iter()
+                    .map(|column_name| ModifyColumnType {
+                        column_name: column_name.clone(),
+                        target_type: ColumnDataType::Float64 as i32,
+                        target_type_extension: None,
+                    })
+                    .collect(),
+            })),
+        };
+
+        if let Err(err) = self
+            .statement_executor
+            .alter_table_inner(alter_expr, ctx.clone())
+            .await
+        {
+            let table = self
+                .catalog_manager
+                .table(&catalog_name, &schema_name, table_name, None)
+                .await
+                .map_err(servers::error::Error::from)?;
+            let alter_already_applied = table
+                .map(|table| {
+                    let table_schema = table.schema();
+                    column_names.iter().all(|column_name| {
+                        table_schema
+                            .column_schema_by_name(column_name)
+                            .and_then(|table_col| {
+                                ColumnDataTypeWrapper::try_from(table_col.data_type.clone())
+                                    .ok()
+                                    .map(|wrapper| wrapper.datatype())
+                            })
+                            == Some(ColumnDataType::Float64)
+                    })
+                })
+                .unwrap_or(false);
+
+            if alter_already_applied {
+                return Ok(());
+            }
+
+            warn!(
+                table_name,
+                columns = ?column_names,
+                error = %err,
+                "failed to widen trace columns before insert"
+            );
+
+            return Err(wrap_trace_alter_failure(err));
+        }
+
+        Ok(())
+    }
+
+    /// Coerce request column types and values to match the existing table schema
+    /// for compatible type pairs. Existing table schema wins when present;
+    /// otherwise the full request batch decides a stable target type.
+    async fn reconcile_trace_column_types(
+        &self,
+        requests: &mut RowInsertRequests,
+        ctx: &QueryContextRef,
+    ) -> ServerResult<()> {
+        let catalog = ctx.current_catalog();
+        let schema = ctx.current_schema();
+
+        for req in &mut requests.inserts {
+            let table = self
+                .catalog_manager
+                .table(catalog, &schema, &req.table_name, None)
+                .await?;
+
+            let Some(rows) = req.rows.as_mut() else {
+                continue;
+            };
+
+            let table_schema = table.map(|table| table.schema());
+            let mut pending_rewrites = Vec::new();
+            let mut pending_alter_columns = Vec::new();
+
+            for (col_idx, col_schema) in rows.schema.iter().enumerate() {
+                let Some(current_type) = ColumnDataType::try_from(col_schema.datatype).ok() else {
+                    continue;
+                };
+
+                let mut observed_types = Vec::new();
+                push_observed_trace_type(&mut observed_types, current_type);
+
+                // Scan the full request first so the final type decision is not affected
+                // by row order inside the batch.
+                for row in &rows.rows {
+                    let Some(value) = row
+                        .values
+                        .get(col_idx)
+                        .and_then(|value| value.value_data.as_ref())
+                    else {
+                        continue;
+                    };
+
+                    let Some(value_type) = trace_value_datatype(value) else {
+                        continue;
+                    };
+                    push_observed_trace_type(&mut observed_types, value_type);
+                }
+
+                let existing_type = table_schema
+                    .as_ref()
+                    .and_then(|schema| schema.column_schema_by_name(&col_schema.column_name))
+                    .and_then(|table_col| {
+                        ColumnDataTypeWrapper::try_from(table_col.data_type.clone())
+                            .ok()
+                            .map(|wrapper| wrapper.datatype())
+                    });
+                let fixed_type = trace_semconv_fixed_type(&col_schema.column_name);
+
+                if !observed_types
+                    .iter()
+                    .copied()
+                    .any(is_trace_reconcile_candidate_type)
+                    && existing_type
+                        .map(|datatype| !is_trace_reconcile_candidate_type(datatype))
+                        .unwrap_or(true)
+                    && fixed_type.is_none()
+                {
+                    continue;
+                }
+
+                // Decide the final type once per column, then rewrite all affected cells
+                // together in one row pass below.
+                let Some(decision) = choose_trace_reconcile_decision(
+                    &col_schema.column_name,
+                    &observed_types,
+                    existing_type,
+                )
+                .map_err(|_| {
+                    enrich_trace_reconcile_error(
+                        &req.table_name,
+                        &col_schema.column_name,
+                        &observed_types,
+                        existing_type,
+                        fixed_type,
+                    )
+                })?
+                else {
+                    continue;
+                };
+                let target_type = decision.target_type();
+
+                if !decision.requires_alter()
+                    && observed_types
+                        .iter()
+                        .all(|observed| *observed == target_type)
+                    && col_schema.datatype == target_type as i32
+                {
+                    continue;
+                }
+
+                if decision.requires_alter()
+                    && !pending_alter_columns.contains(&col_schema.column_name)
+                {
+                    pending_alter_columns.push(col_schema.column_name.clone());
+                }
+
+                pending_rewrites.push(PendingTraceColumnRewrite {
+                    col_idx,
+                    target_type,
+                    column_name: col_schema.column_name.clone(),
+                });
+            }
+
+            if pending_rewrites.is_empty() {
+                continue;
+            }
+
+            validate_trace_column_rewrites(&rows.rows, &pending_rewrites, &req.table_name)?;
+
+            if !pending_alter_columns.is_empty() {
+                self.alter_trace_table_columns_to_float64(
+                    ctx,
+                    &req.table_name,
+                    &pending_alter_columns,
+                )
+                .await?;
+            }
+
+            // Update schema metadata before mutating row values so both stay in sync.
+            for pending_rewrite in &pending_rewrites {
+                rows.schema[pending_rewrite.col_idx].datatype = pending_rewrite.target_type as i32;
+            }
+
+            // Apply all pending column rewrites in one row pass.
+            for row in &mut rows.rows {
+                for pending_rewrite in &pending_rewrites {
+                    let Some(value) = row.values.get_mut(pending_rewrite.col_idx) else {
+                        continue;
+                    };
+                    let Some(request_type) =
+                        value.value_data.as_ref().and_then(trace_value_datatype)
+                    else {
+                        continue;
+                    };
+                    if request_type == pending_rewrite.target_type {
+                        continue;
+                    }
+
+                    value.value_data = coerce_value_data(
+                        &value.value_data,
+                        pending_rewrite.target_type,
+                        request_type,
+                    )
+                    .map_err(|_| {
+                        error::InvalidParameterSnafu {
+                            reason: format!(
+                                "failed to coerce trace column '{}' in table '{}' from {:?} to {:?}",
+                                pending_rewrite.column_name,
+                                req.table_name,
+                                request_type,
+                                pending_rewrite.target_type
+                            ),
+                        }
+                        .build()
+                    })?;
+                }
+            }
+        }
+
+        Ok(())
+    }
+}
+
+/// Preserve the original alter failure status so chunk retry behavior stays correct.
+fn wrap_trace_alter_failure<E>(err: E) -> servers::error::Error
+where
+    E: ErrorExt + Send + Sync + 'static,
+{
+    error::ExecuteGrpcQuerySnafu.into_error(BoxedError::new(err))
+}
+
+#[cfg(test)]
+mod tests {
+    use common_error::ext::ErrorExt;
+    use common_error::status_code::StatusCode;
+    use servers::query_handler::TraceIngestOutcome;
+
+    use super::{ChunkFailureReaction, Instance, wrap_trace_alter_failure};
+    use crate::metrics::OTLP_TRACES_FAILURE_COUNT;
+
+    #[test]
+    fn test_classify_trace_chunk_failure() {
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::InvalidArguments),
+            ChunkFailureReaction::RetryPerSpan
+        );
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::InvalidSyntax),
+            ChunkFailureReaction::RetryPerSpan
+        );
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::Unsupported),
+            ChunkFailureReaction::RetryPerSpan
+        );
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::TableColumnNotFound),
+            ChunkFailureReaction::RetryPerSpan
+        );
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::TableNotFound),
+            ChunkFailureReaction::RetryPerSpan
+        );
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::DatabaseNotFound),
+            ChunkFailureReaction::DiscardChunk
+        );
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::DeadlineExceeded),
+            ChunkFailureReaction::Propagate
+        );
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::Cancelled),
+            ChunkFailureReaction::Propagate
+        );
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::StorageUnavailable),
+            ChunkFailureReaction::Propagate
+        );
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::Internal),
+            ChunkFailureReaction::Propagate
+        );
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::RegionNotReady),
+            ChunkFailureReaction::Propagate
+        );
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::TableUnavailable),
+            ChunkFailureReaction::Propagate
+        );
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::RegionBusy),
+            ChunkFailureReaction::Propagate
+        );
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::RuntimeResourcesExhausted),
+            ChunkFailureReaction::Propagate
+        );
+    }
+
+    #[test]
+    fn test_classify_trace_span_failure() {
+        assert!(Instance::should_propagate_trace_span_failure(
+            StatusCode::DeadlineExceeded
+        ));
+        assert!(Instance::should_propagate_trace_span_failure(
+            StatusCode::StorageUnavailable
+        ));
+        assert!(!Instance::should_propagate_trace_span_failure(
+            StatusCode::InvalidArguments
+        ));
+    }
+
+    #[test]
+    fn test_add_trace_write_cost() {
+        let mut outcome = TraceIngestOutcome::default();
+        Instance::add_trace_write_cost(&mut outcome, 3);
+        Instance::add_trace_write_cost(&mut outcome, 5);
+        assert_eq!(outcome.write_cost, 8);
+    }
+
+    #[test]
+    fn test_finish_trace_failure_message() {
+        let message = Instance::finish_trace_failure_message(
+            3,
+            2,
+            vec!["Rejected span trace:span (InvalidArguments)".to_string()],
+        )
+        .unwrap();
+        assert!(message.contains("Accepted 3 spans, rejected 2 spans"));
+        assert!(message.contains("Rejected span trace:span"));
+
+        assert_eq!(Instance::finish_trace_failure_message(2, 0, vec![]), None);
+    }
+
+    #[test]
+    fn test_finish_trace_failure_message_without_detail_messages() {
+        assert_eq!(
+            Instance::finish_trace_failure_message(0, 2, vec![]),
+            Some("Accepted 0 spans, rejected 2 spans".to_string())
+        );
+    }
+
+    #[test]
+    fn test_push_trace_failure_message_increments_labeled_counter() {
+        let label = "retry_per_span_counter_test";
+        let initial = OTLP_TRACES_FAILURE_COUNT.with_label_values(&[label]).get();
+        let mut messages = Vec::new();
+
+        Instance::push_trace_failure_message(
+            &mut messages,
+            label,
+            "Chunk fallback triggered by InvalidArguments".to_string(),
+        );
+
+        assert_eq!(messages.len(), 1);
+        assert_eq!(
+            OTLP_TRACES_FAILURE_COUNT.with_label_values(&[label]).get(),
+            initial + 1
+        );
+    }
+
+    #[test]
+    fn test_push_trace_failure_message_caps_recorded_messages() {
+        let label = "retry_per_span_limit_test";
+        let mut messages = Vec::new();
+
+        for idx in 0..=4 {
+            Instance::push_trace_failure_message(&mut messages, label, format!("failure-{idx}"));
+        }
+
+        assert_eq!(messages.len(), 4);
+        assert_eq!(
+            messages,
+            vec![
+                "failure-0".to_string(),
+                "failure-1".to_string(),
+                "failure-2".to_string(),
+                "failure-3".to_string()
+            ]
+        );
+    }
+
+    #[test]
+    fn test_classify_trace_chunk_failure_defaults_to_discard() {
+        assert_eq!(
+            Instance::classify_trace_chunk_failure(StatusCode::Unknown),
+            ChunkFailureReaction::DiscardChunk
+        );
+    }
+
+    #[test]
+    fn test_wrap_trace_alter_failure_preserves_status_code() {
+        let err = wrap_trace_alter_failure(
+            servers::error::TableNotFoundSnafu {
+                catalog: "greptime".to_string(),
+                schema: "public".to_string(),
+                table: "trace_type_missing".to_string(),
+            }
+            .build(),
+        );
+
+        assert_eq!(err.status_code(), StatusCode::TableNotFound);
+    }
+}
diff --git a/src/frontend/src/instance/otlp/trace_semconv.rs b/src/frontend/src/instance/otlp/trace_semconv.rs
new file mode 100644
index 0000000000..70ad5a0676
--- /dev/null
+++ b/src/frontend/src/instance/otlp/trace_semconv.rs
@@ -0,0 +1,308 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use api::v1::ColumnDataType;
+use opentelemetry_semantic_conventions::{attribute, resource};
+
+/// Returns fixed scalar types for flattened trace semconv columns.
+///
+/// The mapping is maintained from the official OpenTelemetry semantic
+/// conventions docs under `docs/` and `docs/registry/attributes/` in:
+/// https://github.com/open-telemetry/semantic-conventions/tree/main/docs
+///
+/// Only attributes whose docs mark them as `Stable` or `Release Candidate` are
+/// included here. `Development` and lower-stability attributes must keep the
+/// dynamic reconciliation path instead.
+pub(super) fn trace_semconv_fixed_type(column_name: &str) -> Option<ColumnDataType> {
+    if let Some(resource_attribute) = column_name.strip_prefix("resource_attributes.") {
+        return match resource_attribute {
+            // Service resource attributes:
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/resource/service.md
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/registry/attributes/service.md
+            resource::SERVICE_NAME
+            | resource::SERVICE_INSTANCE_ID
+            | resource::SERVICE_NAMESPACE
+            | resource::SERVICE_VERSION => Some(ColumnDataType::String),
+
+            // Telemetry SDK resource attributes:
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/resource/README.md
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/registry/attributes/telemetry.md
+            resource::TELEMETRY_SDK_LANGUAGE
+            | resource::TELEMETRY_SDK_NAME
+            | resource::TELEMETRY_SDK_VERSION => Some(ColumnDataType::String),
+
+            // Container resource attributes:
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/resource/container.md
+            resource::CONTAINER_ID => Some(ColumnDataType::String),
+
+            // Browser resource attributes:
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/resource/browser.md
+            resource::USER_AGENT_ORIGINAL => Some(ColumnDataType::String),
+
+            // Kubernetes resource attributes:
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/resource/k8s/README.md
+            resource::K8S_CLUSTER_NAME
+            | resource::K8S_CLUSTER_UID
+            | resource::K8S_CONTAINER_NAME
+            | resource::K8S_CRONJOB_NAME
+            | resource::K8S_CRONJOB_UID
+            | resource::K8S_DAEMONSET_NAME
+            | resource::K8S_DAEMONSET_UID
+            | resource::K8S_DEPLOYMENT_NAME
+            | resource::K8S_DEPLOYMENT_UID
+            | resource::K8S_JOB_NAME
+            | resource::K8S_JOB_UID
+            | resource::K8S_NAMESPACE_NAME
+            | resource::K8S_NODE_NAME
+            | resource::K8S_NODE_UID
+            | resource::K8S_POD_NAME
+            | resource::K8S_POD_UID
+            | resource::K8S_REPLICASET_NAME
+            | resource::K8S_REPLICASET_UID
+            | resource::K8S_STATEFULSET_NAME
+            | resource::K8S_STATEFULSET_UID
+            // The current docs include these `k8s.pod.*` keys, but crate 0.31
+            // does not export constants for them yet.
+            | "k8s.pod.hostname"
+            | "k8s.pod.ip"
+            | "k8s.pod.start_time" => Some(ColumnDataType::String),
+            resource::K8S_CONTAINER_RESTART_COUNT => Some(ColumnDataType::Int64),
+
+            _ => None,
+        };
+    }
+
+    if let Some(span_attribute) = column_name.strip_prefix("span_attributes.") {
+        return match span_attribute {
+            // General client, server, and network attributes:
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/general/attributes.md
+            attribute::CLIENT_ADDRESS
+            | attribute::SERVER_ADDRESS
+            | attribute::NETWORK_LOCAL_ADDRESS
+            | attribute::NETWORK_PEER_ADDRESS
+            | attribute::NETWORK_PROTOCOL_NAME
+            | attribute::NETWORK_PROTOCOL_VERSION
+            | attribute::NETWORK_TRANSPORT
+            | attribute::NETWORK_TYPE => Some(ColumnDataType::String),
+            attribute::CLIENT_PORT
+            | attribute::SERVER_PORT
+            | attribute::NETWORK_LOCAL_PORT
+            | attribute::NETWORK_PEER_PORT => Some(ColumnDataType::Int64),
+
+            // HTTP attributes:
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/http/http-spans.md
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/registry/attributes/http.md
+            attribute::HTTP_REQUEST_METHOD
+            | attribute::HTTP_REQUEST_METHOD_ORIGINAL
+            | attribute::HTTP_ROUTE
+            | attribute::URL_FULL
+            | attribute::URL_PATH
+            | attribute::URL_QUERY
+            | attribute::URL_SCHEME
+            | attribute::USER_AGENT_ORIGINAL => Some(ColumnDataType::String),
+            attribute::HTTP_REQUEST_RESEND_COUNT | attribute::HTTP_RESPONSE_STATUS_CODE => {
+                Some(ColumnDataType::Int64)
+            }
+
+            // RPC attributes:
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/rpc/rpc-spans.md
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/registry/attributes/rpc.md
+            attribute::RPC_METHOD
+            | attribute::RPC_SYSTEM
+            // The current docs renamed this attribute to `rpc.system.name`,
+            // but crate 0.31 still exports `RPC_SYSTEM = "rpc.system"`.
+            | "rpc.system.name"
+            | "rpc.method_original"
+            | "rpc.response.status_code" => Some(ColumnDataType::String),
+
+            // General database attributes:
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/db/database-spans.md
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/registry/attributes/db.md
+            attribute::DB_COLLECTION_NAME
+            | attribute::DB_NAMESPACE
+            | attribute::DB_OPERATION_NAME
+            | attribute::DB_QUERY_SUMMARY
+            | attribute::DB_QUERY_TEXT
+            | attribute::DB_RESPONSE_STATUS_CODE
+            | attribute::DB_STORED_PROCEDURE_NAME
+            | attribute::DB_SYSTEM_NAME => Some(ColumnDataType::String),
+            attribute::DB_OPERATION_BATCH_SIZE => Some(ColumnDataType::Int64),
+
+            // Error attributes:
+            // https://github.com/open-telemetry/semantic-conventions/blob/main/docs/registry/attributes/error.md
+            attribute::ERROR_TYPE => Some(ColumnDataType::String),
+
+            _ => None,
+        };
+    }
+
+    None
+}
+
+#[cfg(test)]
+mod tests {
+    use api::v1::ColumnDataType;
+    use opentelemetry_semantic_conventions::{attribute, resource};
+
+    use super::trace_semconv_fixed_type;
+
+    fn resource_column(key: &str) -> String {
+        format!("resource_attributes.{key}")
+    }
+
+    fn span_column(key: &str) -> String {
+        format!("span_attributes.{key}")
+    }
+
+    #[test]
+    fn test_trace_semconv_fixed_type_includes_stable_service_key() {
+        assert_eq!(
+            trace_semconv_fixed_type(&resource_column(resource::SERVICE_NAME)),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&resource_column(resource::SERVICE_VERSION)),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&resource_column(resource::SERVICE_INSTANCE_ID)),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&resource_column(resource::SERVICE_NAMESPACE)),
+            Some(ColumnDataType::String)
+        );
+    }
+
+    #[test]
+    fn test_trace_semconv_fixed_type_includes_http_server_and_error_keys() {
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column(attribute::HTTP_RESPONSE_STATUS_CODE)),
+            Some(ColumnDataType::Int64)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column(attribute::SERVER_PORT)),
+            Some(ColumnDataType::Int64)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column(attribute::ERROR_TYPE)),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column(attribute::CLIENT_ADDRESS)),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column(attribute::CLIENT_PORT)),
+            Some(ColumnDataType::Int64)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column(attribute::URL_FULL)),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column(attribute::URL_PATH)),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column(attribute::URL_QUERY)),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column(attribute::URL_SCHEME)),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column(attribute::USER_AGENT_ORIGINAL)),
+            Some(ColumnDataType::String)
+        );
+    }
+
+    #[test]
+    fn test_trace_semconv_fixed_type_includes_rc_rpc_key() {
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column(attribute::RPC_SYSTEM)),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type("span_attributes.rpc.system.name"),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column("rpc.response.status_code")),
+            Some(ColumnDataType::String)
+        );
+    }
+
+    #[test]
+    fn test_trace_semconv_fixed_type_includes_db_and_network_keys() {
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column(attribute::DB_SYSTEM_NAME)),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column(attribute::DB_OPERATION_BATCH_SIZE)),
+            Some(ColumnDataType::Int64)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column(attribute::NETWORK_PEER_PORT)),
+            Some(ColumnDataType::Int64)
+        );
+    }
+
+    #[test]
+    fn test_trace_semconv_fixed_type_includes_resource_semconv_keys() {
+        assert_eq!(
+            trace_semconv_fixed_type(&resource_column(resource::CONTAINER_ID)),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&resource_column(resource::K8S_CONTAINER_RESTART_COUNT)),
+            Some(ColumnDataType::Int64)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&resource_column(resource::TELEMETRY_SDK_LANGUAGE)),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&resource_column(resource::TELEMETRY_SDK_NAME)),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&resource_column(resource::TELEMETRY_SDK_VERSION)),
+            Some(ColumnDataType::String)
+        );
+    }
+
+    #[test]
+    fn test_trace_semconv_fixed_type_excludes_development_keys() {
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column("messaging.system")),
+            None
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&span_column("rpc.request.metadata.x-request-id")),
+            None
+        );
+        assert_eq!(
+            trace_semconv_fixed_type(&resource_column("k8s.pod.label.app")),
+            None
+        );
+    }
+
+    #[test]
+    fn test_trace_semconv_fixed_type_unknown_key() {
+        assert_eq!(trace_semconv_fixed_type(&span_column("custom.attr")), None);
+    }
+}
diff --git a/src/frontend/src/instance/otlp/trace_types.rs b/src/frontend/src/instance/otlp/trace_types.rs
new file mode 100644
index 0000000000..86eaf9c6b1
--- /dev/null
+++ b/src/frontend/src/instance/otlp/trace_types.rs
@@ -0,0 +1,482 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use api::v1::{ColumnDataType, Row};
+use servers::error::{self, Result as ServerResult};
+use servers::otlp::trace::coerce::{
+    coerce_value_data, is_supported_trace_coercion, resolve_new_trace_column_type,
+    trace_value_datatype,
+};
+
+use crate::instance::otlp::trace_semconv::trace_semconv_fixed_type;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(super) enum TraceReconcileDecision {
+    UseExisting(ColumnDataType),
+    UseRequestLocal(ColumnDataType),
+    AlterExistingTo(ColumnDataType),
+}
+
+impl TraceReconcileDecision {
+    pub(super) fn target_type(self) -> ColumnDataType {
+        match self {
+            Self::UseExisting(target_type)
+            | Self::UseRequestLocal(target_type)
+            | Self::AlterExistingTo(target_type) => target_type,
+        }
+    }
+
+    pub(super) fn requires_alter(self) -> bool {
+        matches!(self, Self::AlterExistingTo(_))
+    }
+}
+
+pub(super) struct PendingTraceColumnRewrite {
+    pub(super) col_idx: usize,
+    pub(super) target_type: ColumnDataType,
+    pub(super) column_name: String,
+}
+
+/// Picks the reconciliation action for one trace column.
+///
+/// Existing table schema is authoritative unless the only incompatible case is
+/// widening an existing Int64 column to Float64 for incoming Int64/Float64 data.
+pub(super) fn choose_trace_reconcile_decision(
+    column_name: &str,
+    observed_types: &[ColumnDataType],
+    existing_type: Option<ColumnDataType>,
+) -> ServerResult<Option<TraceReconcileDecision>> {
+    if let Some(fixed_type) = trace_semconv_fixed_type(column_name) {
+        return choose_fixed_trace_reconcile_decision(fixed_type, observed_types, existing_type);
+    }
+
+    let Some(existing_type) = existing_type else {
+        return resolve_new_trace_column_type(observed_types.iter().copied())
+            .map(|target_type| target_type.map(TraceReconcileDecision::UseRequestLocal))
+            .map_err(|_| {
+                error::InvalidParameterSnafu {
+                    reason: "unsupported trace type mix".to_string(),
+                }
+                .build()
+            });
+    };
+
+    if observed_types.iter().all(|&request_type| {
+        request_type == existing_type || is_supported_trace_coercion(request_type, existing_type)
+    }) {
+        return Ok(Some(TraceReconcileDecision::UseExisting(existing_type)));
+    }
+
+    if existing_type == ColumnDataType::Int64
+        && observed_types.contains(&ColumnDataType::Float64)
+        && observed_types.iter().all(|observed_type| {
+            matches!(
+                observed_type,
+                ColumnDataType::Int64 | ColumnDataType::Float64
+            )
+        })
+    {
+        return Ok(Some(TraceReconcileDecision::AlterExistingTo(
+            ColumnDataType::Float64,
+        )));
+    }
+
+    error::InvalidParameterSnafu {
+        reason: "unsupported trace type mix".to_string(),
+    }
+    .fail()
+}
+
+fn choose_fixed_trace_reconcile_decision(
+    fixed_type: ColumnDataType,
+    observed_types: &[ColumnDataType],
+    existing_type: Option<ColumnDataType>,
+) -> ServerResult<Option<TraceReconcileDecision>> {
+    let Some(existing_type) = existing_type else {
+        return Ok(Some(TraceReconcileDecision::UseRequestLocal(fixed_type)));
+    };
+
+    if existing_type == fixed_type {
+        return Ok(Some(TraceReconcileDecision::UseExisting(fixed_type)));
+    }
+
+    if fixed_type == ColumnDataType::Float64
+        && existing_type == ColumnDataType::Int64
+        && observed_types.iter().all(|observed_type| {
+            matches!(
+                observed_type,
+                ColumnDataType::Int64 | ColumnDataType::Float64
+            )
+        })
+    {
+        return Ok(Some(TraceReconcileDecision::AlterExistingTo(fixed_type)));
+    }
+
+    error::InvalidParameterSnafu {
+        reason: "unsupported trace type mix".to_string(),
+    }
+    .fail()
+}
+
+/// Validate all pending trace column rewrites before any schema mutation happens.
+pub(super) fn validate_trace_column_rewrites(
+    rows: &[Row],
+    pending_rewrites: &[PendingTraceColumnRewrite],
+    table_name: &str,
+) -> ServerResult<()> {
+    for row in rows {
+        for pending_rewrite in pending_rewrites {
+            let Some(value) = row.values.get(pending_rewrite.col_idx) else {
+                continue;
+            };
+            let Some(request_type) = value.value_data.as_ref().and_then(trace_value_datatype)
+            else {
+                continue;
+            };
+            if request_type == pending_rewrite.target_type {
+                continue;
+            }
+
+            coerce_value_data(&value.value_data, pending_rewrite.target_type, request_type)
+                .map_err(|_| {
+                    error::InvalidParameterSnafu {
+                        reason: format!(
+                            "failed to coerce trace column '{}' in table '{}' from {:?} to {:?}",
+                            pending_rewrite.column_name,
+                            table_name,
+                            request_type,
+                            pending_rewrite.target_type
+                        ),
+                    }
+                    .build()
+                })?;
+        }
+    }
+
+    Ok(())
+}
+
+pub(super) fn enrich_trace_reconcile_error(
+    table_name: &str,
+    column_name: &str,
+    observed_types: &[ColumnDataType],
+    existing_type: Option<ColumnDataType>,
+    fixed_type: Option<ColumnDataType>,
+) -> servers::error::Error {
+    let observed_types = observed_types
+        .iter()
+        .map(|datatype| format!("{datatype:?}"))
+        .collect::<Vec<_>>()
+        .join(", ");
+
+    error::InvalidParameterSnafu {
+        reason: match (existing_type, fixed_type) {
+            (Some(existing_type), Some(fixed_type)) => format!(
+                "failed to reconcile trace column '{}' in table '{}' with observed types [{}] against existing {:?} and fixed semconv {:?}",
+                column_name, table_name, observed_types, existing_type, fixed_type
+            ),
+            (Some(existing_type), None) => format!(
+                "failed to reconcile trace column '{}' in table '{}' with observed types [{}] against existing {:?}",
+                column_name, table_name, observed_types, existing_type
+            ),
+            (None, Some(fixed_type)) => format!(
+                "failed to reconcile trace column '{}' in table '{}' with observed types [{}] and fixed semconv {:?}",
+                column_name, table_name, observed_types, fixed_type
+            ),
+            (None, None) => format!(
+                "failed to reconcile trace column '{}' in table '{}' with observed types [{}]",
+                column_name, table_name, observed_types
+            ),
+        },
+    }
+    .build()
+}
+
+/// Only these trace scalar types participate in reconciliation. Other column kinds
+/// such as JSON and binary keep their original write path and schema checks.
+pub(super) fn is_trace_reconcile_candidate_type(datatype: ColumnDataType) -> bool {
+    matches!(
+        datatype,
+        ColumnDataType::String
+            | ColumnDataType::Boolean
+            | ColumnDataType::Int64
+            | ColumnDataType::Float64
+    )
+}
+
+/// Keeps the observed type list small without depending on enum ordering.
+pub(super) fn push_observed_trace_type(
+    observed_types: &mut Vec<ColumnDataType>,
+    datatype: ColumnDataType,
+) {
+    if !observed_types.contains(&datatype) {
+        observed_types.push(datatype);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use api::v1::value::ValueData;
+    use api::v1::{ColumnDataType, Row, Value};
+    use common_error::ext::ErrorExt;
+    use common_error::status_code::StatusCode;
+
+    use super::{
+        PendingTraceColumnRewrite, TraceReconcileDecision, choose_trace_reconcile_decision,
+        enrich_trace_reconcile_error, is_trace_reconcile_candidate_type, push_observed_trace_type,
+        validate_trace_column_rewrites,
+    };
+
+    #[test]
+    fn test_choose_trace_reconcile_decision_existing_int64_keeps_int64() {
+        assert_eq!(
+            choose_trace_reconcile_decision(
+                "span_attributes.attr_int",
+                &[ColumnDataType::Int64],
+                Some(ColumnDataType::Int64)
+            )
+            .unwrap(),
+            Some(TraceReconcileDecision::UseExisting(ColumnDataType::Int64))
+        );
+    }
+
+    #[test]
+    fn test_choose_trace_reconcile_decision_existing_int64_widens_to_float64() {
+        assert_eq!(
+            choose_trace_reconcile_decision(
+                "span_attributes.attr_double",
+                &[ColumnDataType::Int64, ColumnDataType::Float64],
+                Some(ColumnDataType::Int64)
+            )
+            .unwrap(),
+            Some(TraceReconcileDecision::AlterExistingTo(
+                ColumnDataType::Float64
+            ))
+        );
+    }
+
+    #[test]
+    fn test_choose_trace_reconcile_decision_existing_float64_stays_authoritative() {
+        assert_eq!(
+            choose_trace_reconcile_decision(
+                "span_attributes.attr_double",
+                &[ColumnDataType::Int64, ColumnDataType::Float64],
+                Some(ColumnDataType::Float64)
+            )
+            .unwrap(),
+            Some(TraceReconcileDecision::UseExisting(ColumnDataType::Float64))
+        );
+    }
+
+    #[test]
+    fn test_choose_trace_reconcile_decision_existing_int64_with_boolean_is_error() {
+        let err = choose_trace_reconcile_decision(
+            "span_attributes.attr_numeric",
+            &[ColumnDataType::Boolean, ColumnDataType::Int64],
+            Some(ColumnDataType::Int64),
+        )
+        .unwrap_err();
+        assert_eq!(err.status_code(), StatusCode::InvalidArguments);
+    }
+
+    #[test]
+    fn test_choose_trace_reconcile_decision_request_local_prefers_float64() {
+        assert_eq!(
+            choose_trace_reconcile_decision(
+                "span_attributes.attr_numeric",
+                &[ColumnDataType::Int64, ColumnDataType::Float64],
+                None
+            )
+            .unwrap(),
+            Some(TraceReconcileDecision::UseRequestLocal(
+                ColumnDataType::Float64
+            ))
+        );
+    }
+
+    #[test]
+    fn test_choose_trace_reconcile_decision_whitelisted_new_int64_column_uses_fixed_type() {
+        assert_eq!(
+            choose_trace_reconcile_decision(
+                "span_attributes.http.response.status_code",
+                &[ColumnDataType::String, ColumnDataType::Int64],
+                None
+            )
+            .unwrap(),
+            Some(TraceReconcileDecision::UseRequestLocal(
+                ColumnDataType::Int64
+            ))
+        );
+    }
+
+    #[test]
+    fn test_choose_trace_reconcile_decision_new_boolean_column_uses_dynamic_resolution() {
+        assert_eq!(
+            choose_trace_reconcile_decision(
+                "span_attributes.messaging.destination.temporary",
+                &[ColumnDataType::String, ColumnDataType::Boolean],
+                None
+            )
+            .unwrap(),
+            Some(TraceReconcileDecision::UseRequestLocal(
+                ColumnDataType::Boolean
+            ))
+        );
+    }
+
+    #[test]
+    fn test_choose_trace_reconcile_decision_whitelisted_existing_matching_type_uses_fixed_type() {
+        assert_eq!(
+            choose_trace_reconcile_decision(
+                "resource_attributes.service.name",
+                &[ColumnDataType::String],
+                Some(ColumnDataType::String)
+            )
+            .unwrap(),
+            Some(TraceReconcileDecision::UseExisting(ColumnDataType::String))
+        );
+    }
+
+    #[test]
+    fn test_choose_trace_reconcile_decision_whitelisted_existing_conflicting_type_is_error() {
+        let err = choose_trace_reconcile_decision(
+            "span_attributes.server.port",
+            &[ColumnDataType::Int64],
+            Some(ColumnDataType::String),
+        )
+        .unwrap_err();
+        assert_eq!(err.status_code(), StatusCode::InvalidArguments);
+    }
+
+    #[test]
+    fn test_choose_trace_reconcile_decision_non_whitelisted_retains_dynamic_behavior() {
+        assert_eq!(
+            choose_trace_reconcile_decision(
+                "span_attributes.attr_numeric",
+                &[ColumnDataType::Int64, ColumnDataType::Float64],
+                None
+            )
+            .unwrap(),
+            Some(TraceReconcileDecision::UseRequestLocal(
+                ColumnDataType::Float64
+            ))
+        );
+    }
+
+    #[test]
+    fn test_validate_trace_column_rewrites_rejects_invalid_string_parse() {
+        let rows = vec![Row {
+            values: vec![Value {
+                value_data: Some(ValueData::StringValue("not_a_number".to_string())),
+            }],
+        }];
+        let pending_rewrites = vec![PendingTraceColumnRewrite {
+            col_idx: 0,
+            target_type: ColumnDataType::Int64,
+            column_name: "span_attributes.attr_int".to_string(),
+        }];
+
+        let err = validate_trace_column_rewrites(&rows, &pending_rewrites, "trace_type_atomicity")
+            .unwrap_err();
+        assert_eq!(err.status_code(), StatusCode::InvalidArguments);
+    }
+
+    #[test]
+    fn test_validate_trace_column_rewrites_whitelisted_values_validate_against_fixed_type() {
+        let rows = vec![Row {
+            values: vec![Value {
+                value_data: Some(ValueData::StringValue("503".to_string())),
+            }],
+        }];
+        let pending_rewrites = vec![PendingTraceColumnRewrite {
+            col_idx: 0,
+            target_type: ColumnDataType::Int64,
+            column_name: "span_attributes.http.response.status_code".to_string(),
+        }];
+
+        validate_trace_column_rewrites(&rows, &pending_rewrites, "trace_type_atomicity").unwrap();
+    }
+
+    #[test]
+    fn test_validate_trace_column_rewrites_whitelisted_boolean_rejects_invalid_string_parse() {
+        let rows = vec![Row {
+            values: vec![Value {
+                value_data: Some(ValueData::StringValue("not_a_bool".to_string())),
+            }],
+        }];
+        let pending_rewrites = vec![PendingTraceColumnRewrite {
+            col_idx: 0,
+            target_type: ColumnDataType::Boolean,
+            column_name: "span_attributes.messaging.destination.temporary".to_string(),
+        }];
+
+        let err = validate_trace_column_rewrites(&rows, &pending_rewrites, "trace_type_atomicity")
+            .unwrap_err();
+        assert_eq!(err.status_code(), StatusCode::InvalidArguments);
+    }
+
+    #[test]
+    fn test_enrich_trace_reconcile_error_includes_existing_type() {
+        let err = enrich_trace_reconcile_error(
+            "trace_type_atomicity",
+            "span_attributes.attr_int",
+            &[ColumnDataType::String, ColumnDataType::Int64],
+            Some(ColumnDataType::Boolean),
+            None,
+        );
+
+        assert_eq!(err.status_code(), StatusCode::InvalidArguments);
+        assert!(err.to_string().contains("span_attributes.attr_int"));
+        assert!(err.to_string().contains("Boolean"));
+    }
+
+    #[test]
+    fn test_enrich_trace_reconcile_error_includes_fixed_semconv_type() {
+        let err = enrich_trace_reconcile_error(
+            "trace_type_atomicity",
+            "span_attributes.server.port",
+            &[ColumnDataType::String, ColumnDataType::Int64],
+            Some(ColumnDataType::String),
+            Some(ColumnDataType::Int64),
+        );
+
+        assert_eq!(err.status_code(), StatusCode::InvalidArguments);
+        assert!(err.to_string().contains("span_attributes.server.port"));
+        assert!(err.to_string().contains("fixed semconv Int64"));
+    }
+
+    #[test]
+    fn test_is_trace_reconcile_candidate_type_filters_non_scalar_types() {
+        assert!(is_trace_reconcile_candidate_type(ColumnDataType::String));
+        assert!(is_trace_reconcile_candidate_type(ColumnDataType::Boolean));
+        assert!(!is_trace_reconcile_candidate_type(ColumnDataType::Binary));
+        assert!(!is_trace_reconcile_candidate_type(
+            ColumnDataType::TimestampMillisecond
+        ));
+    }
+
+    #[test]
+    fn test_push_observed_trace_type_deduplicates_types() {
+        let mut observed_types = Vec::new();
+
+        push_observed_trace_type(&mut observed_types, ColumnDataType::Int64);
+        push_observed_trace_type(&mut observed_types, ColumnDataType::Int64);
+        push_observed_trace_type(&mut observed_types, ColumnDataType::Float64);
+
+        assert_eq!(
+            observed_types,
+            vec![ColumnDataType::Int64, ColumnDataType::Float64]
+        );
+    }
+}
diff --git a/src/frontend/src/instance/prom_store.rs b/src/frontend/src/instance/prom_store.rs
index 9a323eb989..10fc2f1790 100644
--- a/src/frontend/src/instance/prom_store.rs
+++ b/src/frontend/src/instance/prom_store.rs
@@ -17,7 +17,11 @@ use std::sync::Arc;
 
 use api::prom_store::remote::read_request::ResponseType;
 use api::prom_store::remote::{Query, QueryResult, ReadRequest, ReadResponse};
-use api::v1::RowInsertRequests;
+use api::v1::alter_table_expr::Kind;
+use api::v1::{
+    AddColumn, AddColumns, AlterTableExpr, ColumnDataType, ColumnDef, CreateTableExpr,
+    RowInsertRequests, SemanticType,
+};
 use async_trait::async_trait;
 use auth::{PermissionChecker, PermissionCheckerRef, PermissionReq};
 use client::OutputData;
@@ -27,19 +31,25 @@ use common_query::Output;
 use common_query::prelude::GREPTIME_PHYSICAL_TABLE;
 use common_recordbatch::RecordBatches;
 use common_telemetry::{debug, tracing};
-use operator::insert::InserterRef;
+use operator::insert::{
+    AutoCreateTableType, InserterRef, build_create_table_expr, fill_table_options_for_create,
+};
 use operator::statement::StatementExecutor;
 use prost::Message;
 use servers::error::{self, AuthSnafu, Result as ServerResult};
 use servers::http::header::{CONTENT_ENCODING_SNAPPY, CONTENT_TYPE_PROTOBUF, collect_plan_metrics};
 use servers::http::prom_store::PHYSICAL_TABLE_PARAM;
 use servers::interceptor::{PromStoreProtocolInterceptor, PromStoreProtocolInterceptorRef};
+use servers::pending_rows_batcher::PendingRowsSchemaAlterer;
 use servers::prom_store::{self, Metrics};
 use servers::query_handler::{
     PromStoreProtocolHandler, PromStoreProtocolHandlerRef, PromStoreResponse,
 };
 use session::context::QueryContextRef;
 use snafu::{OptionExt, ResultExt};
+use store_api::metric_engine_consts::{METRIC_ENGINE_NAME, PHYSICAL_TABLE_METADATA_KEY};
+use store_api::mito_engine_options::SST_FORMAT_KEY;
+use table::table_reference::TableReference;
 use tracing::instrument;
 
 use crate::error::{
@@ -50,6 +60,34 @@ use crate::instance::Instance;
 
 const SAMPLES_RESPONSE_TYPE: i32 = ResponseType::Samples as i32;
 
+fn auto_create_table_type_for_prom_remote_write(
+    ctx: &QueryContextRef,
+    with_metric_engine: bool,
+) -> AutoCreateTableType {
+    if with_metric_engine {
+        let physical_table = ctx
+            .extension(PHYSICAL_TABLE_PARAM)
+            .unwrap_or(GREPTIME_PHYSICAL_TABLE)
+            .to_string();
+        AutoCreateTableType::Logical(physical_table)
+    } else {
+        AutoCreateTableType::Physical
+    }
+}
+
+fn required_physical_table_for_create_type(create_type: &AutoCreateTableType) -> Option<&str> {
+    match create_type {
+        AutoCreateTableType::Logical(physical_table) => Some(physical_table.as_str()),
+        _ => None,
+    }
+}
+
+fn fill_metric_physical_table_options(table_options: &mut HashMap<String, String>) {
+    // We always enforce flat format in this ingestion path.
+    table_options.insert(SST_FORMAT_KEY.to_string(), "flat".to_string());
+    table_options.insert(PHYSICAL_TABLE_METADATA_KEY.to_string(), "true".to_string());
+}
+
 #[inline]
 fn is_supported(response_type: i32) -> bool {
     // Only supports samples response right now
@@ -160,13 +198,163 @@ impl Instance {
 }
 
 #[async_trait]
-impl PromStoreProtocolHandler for Instance {
-    async fn write(
+impl PendingRowsSchemaAlterer for Instance {
+    async fn create_tables_if_missing_batch(
         &self,
-        request: RowInsertRequests,
-        ctx: QueryContextRef,
+        catalog: &str,
+        schema: &str,
+        tables: &[(&str, &[api::v1::ColumnSchema])],
         with_metric_engine: bool,
-    ) -> ServerResult<Output> {
+        ctx: QueryContextRef,
+    ) -> ServerResult<()> {
+        if tables.is_empty() {
+            return Ok(());
+        }
+
+        let create_type = auto_create_table_type_for_prom_remote_write(&ctx, with_metric_engine);
+        if let Some(physical_table) = required_physical_table_for_create_type(&create_type) {
+            self.create_metric_physical_table_if_missing(
+                catalog,
+                schema,
+                physical_table,
+                ctx.clone(),
+            )
+            .await?;
+        }
+
+        let engine = if matches!(create_type, AutoCreateTableType::Logical(_)) {
+            METRIC_ENGINE_NAME
+        } else {
+            common_catalog::consts::default_engine()
+        };
+
+        // Check which tables actually still need to be created (may have been
+        // concurrently created by another request).
+        let mut create_exprs: Vec<CreateTableExpr> = Vec::with_capacity(tables.len());
+        for &(table_name, request_schema) in tables {
+            let existing = self
+                .catalog_manager()
+                .table(catalog, schema, table_name, Some(ctx.as_ref()))
+                .await
+                .map_err(BoxedError::new)
+                .context(error::ExecuteGrpcQuerySnafu)?;
+            if existing.is_some() {
+                continue;
+            }
+
+            let table_ref = TableReference::full(catalog, schema, table_name);
+            let mut create_table_expr = build_create_table_expr(&table_ref, request_schema, engine)
+                .map_err(BoxedError::new)
+                .context(error::ExecuteGrpcQuerySnafu)?;
+
+            let mut table_options = std::collections::HashMap::with_capacity(4);
+            fill_table_options_for_create(&mut table_options, &create_type, &ctx);
+            create_table_expr.table_options.extend(table_options);
+            create_exprs.push(create_table_expr);
+        }
+
+        if create_exprs.is_empty() {
+            return Ok(());
+        }
+
+        match create_type {
+            AutoCreateTableType::Logical(_) => {
+                // Use the batch API for logical tables.
+                self.statement_executor
+                    .create_logical_tables(&create_exprs, ctx)
+                    .await
+                    .map_err(BoxedError::new)
+                    .context(error::ExecuteGrpcQuerySnafu)?;
+            }
+            AutoCreateTableType::Physical => {
+                // Physical tables don't have a batch DDL path; create one at a time.
+                for mut expr in create_exprs {
+                    expr.table_options
+                        .insert(SST_FORMAT_KEY.to_string(), "flat".to_string());
+                    self.statement_executor
+                        .create_table_inner(&mut expr, None, ctx.clone())
+                        .await
+                        .map_err(BoxedError::new)
+                        .context(error::ExecuteGrpcQuerySnafu)?;
+                }
+            }
+            create_type => {
+                return error::InvalidPromRemoteRequestSnafu {
+                    msg: format!(
+                        "prom remote write only supports logical or physical auto-create: {}",
+                        create_type.as_str()
+                    ),
+                }
+                .fail();
+            }
+        }
+
+        Ok(())
+    }
+
+    async fn add_missing_prom_tag_columns_batch(
+        &self,
+        catalog: &str,
+        schema: &str,
+        tables: &[(&str, &[String])],
+        ctx: QueryContextRef,
+    ) -> ServerResult<()> {
+        if tables.is_empty() {
+            return Ok(());
+        }
+
+        let alter_exprs: Vec<AlterTableExpr> = tables
+            .iter()
+            .filter(|(_, columns)| !columns.is_empty())
+            .map(|&(table_name, columns)| {
+                let add_columns = AddColumns {
+                    add_columns: columns
+                        .iter()
+                        .map(|column_name| AddColumn {
+                            column_def: Some(ColumnDef {
+                                name: column_name.clone(),
+                                data_type: ColumnDataType::String as i32,
+                                is_nullable: true,
+                                semantic_type: SemanticType::Tag as i32,
+                                comment: String::new(),
+                                ..Default::default()
+                            }),
+                            location: None,
+                            add_if_not_exists: true,
+                        })
+                        .collect(),
+                };
+
+                AlterTableExpr {
+                    catalog_name: catalog.to_string(),
+                    schema_name: schema.to_string(),
+                    table_name: table_name.to_string(),
+                    kind: Some(Kind::AddColumns(add_columns)),
+                }
+            })
+            .collect();
+
+        if alter_exprs.is_empty() {
+            return Ok(());
+        }
+
+        self.statement_executor
+            .alter_logical_tables(alter_exprs, ctx)
+            .await
+            .map_err(BoxedError::new)
+            .context(error::ExecuteGrpcQuerySnafu)?;
+
+        Ok(())
+    }
+}
+
+#[async_trait]
+impl PromStoreProtocolHandler for Instance {
+    async fn pre_write(
+        &self,
+        request: &RowInsertRequests,
+        ctx: QueryContextRef,
+    ) -> ServerResult<()> {
         self.plugins
             .get::<PermissionCheckerRef>()
             .as_ref()
@@ -175,7 +363,17 @@ impl PromStoreProtocolHandler for Instance {
         let interceptor_ref = self
             .plugins
             .get::<PromStoreProtocolInterceptorRef<servers::error::Error>>();
-        interceptor_ref.pre_write(&request, ctx.clone())?;
+        interceptor_ref.pre_write(request, ctx)?;
+        Ok(())
+    }
+
+    async fn write(
+        &self,
+        request: RowInsertRequests,
+        ctx: QueryContextRef,
+        with_metric_engine: bool,
+    ) -> ServerResult<Output> {
+        self.pre_write(&request, ctx.clone()).await?;
 
         let output = if with_metric_engine {
             let physical_table = ctx
@@ -258,6 +456,61 @@ impl PromStoreProtocolHandler for Instance {
     }
 }
 
+impl Instance {
+    async fn create_metric_physical_table_if_missing(
+        &self,
+        catalog: &str,
+        schema: &str,
+        physical_table: &str,
+        ctx: QueryContextRef,
+    ) -> ServerResult<()> {
+        let table = self
+            .catalog_manager()
+            .table(catalog, schema, physical_table, Some(ctx.as_ref()))
+            .await
+            .map_err(BoxedError::new)
+            .context(error::ExecuteGrpcQuerySnafu)?;
+        if table.is_some() {
+            return Ok(());
+        }
+
+        let table_ref = TableReference::full(catalog, schema, physical_table);
+        let default_schema = vec![
+            api::v1::ColumnSchema {
+                column_name: common_query::prelude::greptime_timestamp().to_string(),
+                datatype: api::v1::ColumnDataType::TimestampMillisecond as i32,
+                semantic_type: api::v1::SemanticType::Timestamp as i32,
+                datatype_extension: None,
+                options: None,
+            },
+            api::v1::ColumnSchema {
+                column_name: common_query::prelude::greptime_value().to_string(),
+                datatype: api::v1::ColumnDataType::Float64 as i32,
+                semantic_type: api::v1::SemanticType::Field as i32,
+                datatype_extension: None,
+                options: None,
+            },
+        ];
+        let mut create_table_expr = build_create_table_expr(
+            &table_ref,
+            &default_schema,
+            common_catalog::consts::default_engine(),
+        )
+        .map_err(BoxedError::new)
+        .context(error::ExecuteGrpcQuerySnafu)?;
+        create_table_expr.engine = METRIC_ENGINE_NAME.to_string();
+        fill_metric_physical_table_options(&mut create_table_expr.table_options);
+
+        self.statement_executor
+            .create_table_inner(&mut create_table_expr, None, ctx)
+            .await
+            .map_err(BoxedError::new)
+            .context(error::ExecuteGrpcQuerySnafu)?;
+
+        Ok(())
+    }
+}
+
 /// This handler is mainly used for `frontend` or `standalone` to directly import
 /// the metrics collected by itself, thereby avoiding importing metrics through the network,
 /// thus reducing compression and network transmission overhead,
@@ -311,3 +564,72 @@ impl PromStoreProtocolHandler for ExportMetricHandler {
         unreachable!();
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use session::context::QueryContext;
+
+    use super::*;
+
+    #[test]
+    fn test_auto_create_table_type_for_prom_remote_write_metric_engine() {
+        let mut query_ctx = QueryContext::with(
+            common_catalog::consts::DEFAULT_CATALOG_NAME,
+            common_catalog::consts::DEFAULT_SCHEMA_NAME,
+        );
+        query_ctx.set_extension(PHYSICAL_TABLE_PARAM, "metric_physical".to_string());
+        let ctx = Arc::new(query_ctx);
+
+        let create_type = auto_create_table_type_for_prom_remote_write(&ctx, true);
+        match create_type {
+            AutoCreateTableType::Logical(physical) => assert_eq!(physical, "metric_physical"),
+            _ => panic!("expected logical table create type"),
+        }
+    }
+
+    #[test]
+    fn test_auto_create_table_type_for_prom_remote_write_without_metric_engine() {
+        let ctx = Arc::new(QueryContext::with(
+            common_catalog::consts::DEFAULT_CATALOG_NAME,
+            common_catalog::consts::DEFAULT_SCHEMA_NAME,
+        ));
+
+        let create_type = auto_create_table_type_for_prom_remote_write(&ctx, false);
+        match create_type {
+            AutoCreateTableType::Physical => {}
+            _ => panic!("expected physical table create type"),
+        }
+    }
+
+    #[test]
+    fn test_required_physical_table_for_create_type() {
+        let logical = AutoCreateTableType::Logical("phy_table".to_string());
+        assert_eq!(
+            Some("phy_table"),
+            required_physical_table_for_create_type(&logical)
+        );
+
+        let physical = AutoCreateTableType::Physical;
+        assert_eq!(None, required_physical_table_for_create_type(&physical));
+    }
+
+    #[test]
+    fn test_metric_physical_table_options_forces_flat_sst_format() {
+        let mut table_options = HashMap::new();
+
+        fill_metric_physical_table_options(&mut table_options);
+
+        assert_eq!(
+            Some("flat"),
+            table_options.get(SST_FORMAT_KEY).map(String::as_str)
+        );
+        assert_eq!(
+            Some("true"),
+            table_options
+                .get(PHYSICAL_TABLE_METADATA_KEY)
+                .map(String::as_str)
+        );
+    }
+}
diff --git a/src/frontend/src/instance/promql.rs b/src/frontend/src/instance/promql.rs
index 419be8d96e..3a3aba2307 100644
--- a/src/frontend/src/instance/promql.rs
+++ b/src/frontend/src/instance/promql.rs
@@ -31,7 +31,7 @@ use snafu::{OptionExt, ResultExt};
 use crate::error::{
     CatalogSnafu, CollectRecordbatchSnafu, ExecLogicalPlanSnafu,
     PrometheusLabelValuesQueryPlanSnafu, PrometheusMetricNamesQueryPlanSnafu, ReadTableSnafu,
-    Result, TableNotFoundSnafu,
+    Result, TableNotFoundSnafu, TableSnafu,
 };
 use crate::instance::Instance;
 
@@ -120,20 +120,32 @@ impl Instance {
             })
             .unwrap_or_else(|| ctx.current_schema());
 
+        let full_table_name = format_full_table_name(ctx.current_catalog(), &table_schema, &metric);
         let table = self
             .catalog_manager
             .table(ctx.current_catalog(), &table_schema, &metric, Some(ctx))
             .await
             .context(CatalogSnafu)?
             .with_context(|| TableNotFoundSnafu {
-                table_name: format_full_table_name(ctx.current_catalog(), &table_schema, &metric),
+                table_name: full_table_name.clone(),
             })?;
 
+        // Check label column existence before building the query plan so a missing label can be
+        // reported as `TableColumnNotFound` and handled like Prometheus expects.
+        if table.schema().column_schema_by_name(&label_name).is_none() {
+            return table::error::ColumnNotExistsSnafu {
+                column_name: label_name,
+                table_name: full_table_name,
+            }
+            .fail()
+            .context(TableSnafu);
+        }
+
         let dataframe = self
             .query_engine
             .read_table(table.clone())
             .with_context(|_| ReadTableSnafu {
-                table_name: format_full_table_name(ctx.current_catalog(), &table_schema, &metric),
+                table_name: full_table_name,
             })?;
 
         let scan_plan = dataframe.into_unoptimized_plan();
diff --git a/src/frontend/src/lib.rs b/src/frontend/src/lib.rs
index 16321795b7..c170236073 100644
--- a/src/frontend/src/lib.rs
+++ b/src/frontend/src/lib.rs
@@ -12,8 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
-
 pub mod error;
 pub mod events;
 pub mod frontend;
diff --git a/src/frontend/src/metrics.rs b/src/frontend/src/metrics.rs
index 58ba21476a..aba33637cf 100644
--- a/src/frontend/src/metrics.rs
+++ b/src/frontend/src/metrics.rs
@@ -52,6 +52,14 @@ lazy_static! {
     )
     .unwrap();
 
+    /// The number of OpenTelemetry trace ingest failures on the frontend node.
+    pub static ref OTLP_TRACES_FAILURE_COUNT: IntCounterVec = register_int_counter_vec!(
+        "greptime_frontend_otlp_traces_failure_count",
+        "frontend otlp trace ingest failure count",
+        &["label"]
+    )
+    .unwrap();
+
     /// The number of OpenTelemetry logs send by frontend node.
     pub static ref OTLP_LOGS_ROWS: IntCounter = register_int_counter!(
         "greptime_frontend_otlp_logs_rows",
diff --git a/src/frontend/src/server.rs b/src/frontend/src/server.rs
index 45c3ec3649..e66ae718ba 100644
--- a/src/frontend/src/server.rs
+++ b/src/frontend/src/server.rs
@@ -37,6 +37,7 @@ use servers::interceptor::LogIngestInterceptorRef;
 use servers::metrics_handler::MetricsHandler;
 use servers::mysql::server::{MysqlServer, MysqlSpawnConfig, MysqlSpawnRef};
 use servers::otel_arrow::OtelArrowServiceHandler;
+use servers::pending_rows_batcher::PendingRowsBatcher;
 use servers::postgres::PostgresServer;
 use servers::request_memory_limiter::ServerMemoryLimiter;
 use servers::server::{Server, ServerHandlers};
@@ -124,12 +125,29 @@ where
         }
 
         if opts.prom_store.enable {
+            let pending_rows_batcher = if opts.prom_store.with_metric_engine {
+                PendingRowsBatcher::try_new(
+                    self.instance.partition_manager().clone(),
+                    self.instance.node_manager().clone(),
+                    self.instance.catalog_manager().clone(),
+                    opts.prom_store.with_metric_engine,
+                    self.instance.clone(),
+                    opts.prom_store.pending_rows_flush_interval,
+                    opts.prom_store.max_batch_rows,
+                    opts.prom_store.max_concurrent_flushes,
+                    opts.prom_store.worker_channel_capacity,
+                    opts.prom_store.max_inflight_requests,
+                )
+            } else {
+                None
+            };
             builder = builder
                 .with_prom_handler(
                     self.instance.clone(),
                     Some(self.instance.clone()),
                     opts.prom_store.with_metric_engine,
                     opts.http.prom_validation_mode,
+                    pending_rows_batcher,
                 )
                 .with_prometheus_handler(self.instance.clone());
         }
@@ -143,6 +161,8 @@ where
             builder = builder.with_jaeger_handler(self.instance.clone());
         }
 
+        builder = builder.with_dashboard_handler(self.instance.clone());
+
         if let Some(configurator) = self.plugins.get::<RouterConfigurator>() {
             info!("Adding extra router from plugins");
             builder = builder.with_extra_router(configurator.router());
diff --git a/src/frontend/src/service_config/prom_store.rs b/src/frontend/src/service_config/prom_store.rs
index b3adf889d2..99f1eada6d 100644
--- a/src/frontend/src/service_config/prom_store.rs
+++ b/src/frontend/src/service_config/prom_store.rs
@@ -12,12 +12,40 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+use std::time::Duration;
+
 use serde::{Deserialize, Serialize};
 
 #[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
 pub struct PromStoreOptions {
     pub enable: bool,
     pub with_metric_engine: bool,
+    #[serde(default, with = "humantime_serde")]
+    pub pending_rows_flush_interval: Duration,
+    #[serde(default = "default_max_batch_rows")]
+    pub max_batch_rows: usize,
+    #[serde(default = "default_max_concurrent_flushes")]
+    pub max_concurrent_flushes: usize,
+    #[serde(default = "default_worker_channel_capacity")]
+    pub worker_channel_capacity: usize,
+    #[serde(default = "default_max_inflight_requests")]
+    pub max_inflight_requests: usize,
+}
+
+fn default_max_batch_rows() -> usize {
+    100_000
+}
+
+fn default_max_concurrent_flushes() -> usize {
+    256
+}
+
+fn default_worker_channel_capacity() -> usize {
+    65526
+}
+
+fn default_max_inflight_requests() -> usize {
+    3000
 }
 
 impl Default for PromStoreOptions {
@@ -25,18 +53,43 @@ impl Default for PromStoreOptions {
         Self {
             enable: true,
             with_metric_engine: true,
+            pending_rows_flush_interval: Duration::ZERO,
+            max_batch_rows: default_max_batch_rows(),
+            max_concurrent_flushes: default_max_concurrent_flushes(),
+            worker_channel_capacity: default_worker_channel_capacity(),
+            max_inflight_requests: default_max_inflight_requests(),
         }
     }
 }
 
 #[cfg(test)]
 mod tests {
+    use std::time::Duration;
+
     use super::PromStoreOptions;
+    use crate::service_config::prom_store::{
+        default_max_batch_rows, default_max_concurrent_flushes, default_max_inflight_requests,
+        default_worker_channel_capacity,
+    };
 
     #[test]
     fn test_prom_store_options() {
         let default = PromStoreOptions::default();
         assert!(default.enable);
-        assert!(default.with_metric_engine)
+        assert!(default.with_metric_engine);
+        assert_eq!(default.pending_rows_flush_interval, Duration::ZERO);
+        assert_eq!(default.max_batch_rows, default_max_batch_rows());
+        assert_eq!(
+            default.max_concurrent_flushes,
+            default_max_concurrent_flushes()
+        );
+        assert_eq!(
+            default.worker_channel_capacity,
+            default_worker_channel_capacity()
+        );
+        assert_eq!(
+            default.max_inflight_requests,
+            default_max_inflight_requests()
+        );
     }
 }
diff --git a/src/index/src/fulltext_index/search/tantivy.rs b/src/index/src/fulltext_index/search/tantivy.rs
index d06a5cd329..9fb0091868 100644
--- a/src/index/src/fulltext_index/search/tantivy.rs
+++ b/src/index/src/fulltext_index/search/tantivy.rs
@@ -14,6 +14,7 @@
 
 use std::collections::{BTreeSet, HashMap};
 use std::path::Path;
+use std::sync::Arc;
 use std::time::Instant;
 
 use async_trait::async_trait;
@@ -27,15 +28,19 @@ use tantivy::{Index, IndexReader, ReloadPolicy, TantivyDocument};
 use crate::fulltext_index::Config;
 use crate::fulltext_index::create::{ROWID_FIELD_NAME, TEXT_FIELD_NAME};
 use crate::fulltext_index::error::{
-    Result, TantivyDocNotFoundSnafu, TantivyParserSnafu, TantivySnafu,
+    JoinSnafu, Result, TantivyDocNotFoundSnafu, TantivyParserSnafu, TantivySnafu,
 };
 use crate::fulltext_index::search::{FulltextIndexSearcher, RowId};
 
 /// `TantivyFulltextIndexSearcher` is a searcher using Tantivy.
 pub struct TantivyFulltextIndexSearcher {
-    /// Tanitvy index.
+    inner: Arc<TantivySearcherInner>,
+}
+
+struct TantivySearcherInner {
+    /// Tantivy index.
     index: Index,
-    /// Tanitvy index reader.
+    /// Tantivy index reader.
     reader: IndexReader,
     /// The default field used to build `QueryParser`
     default_field: Field,
@@ -66,63 +71,73 @@ impl TantivyFulltextIndexSearcher {
         );
 
         Ok(Self {
-            index,
-            reader,
-            default_field,
+            inner: Arc::new(TantivySearcherInner {
+                index,
+                reader,
+                default_field,
+            }),
         })
     }
 }
 
+fn search_sync(inner: &TantivySearcherInner, query: &str) -> Result<BTreeSet<RowId>> {
+    let searcher = inner.reader.searcher();
+    let query_parser = QueryParser::for_index(&inner.index, vec![inner.default_field]);
+    let query = query_parser
+        .parse_query(query)
+        .context(TantivyParserSnafu)?;
+    let doc_addrs = searcher
+        .search(&query, &DocSetCollector)
+        .context(TantivySnafu)?;
+
+    let seg_metas = inner
+        .index
+        .searchable_segment_metas()
+        .context(TantivySnafu)?;
+
+    // FAST PATH: only one segment, the doc id is the same as the row id.
+    //            Also for compatibility with the old version.
+    if seg_metas.len() == 1 {
+        return Ok(doc_addrs.into_iter().map(|d| d.doc_id).collect());
+    }
+
+    // SLOW PATH: multiple segments, need to calculate the row id.
+    let rowid_field = searcher
+        .schema()
+        .get_field(ROWID_FIELD_NAME)
+        .context(TantivySnafu)?;
+    let mut seg_offsets = HashMap::with_capacity(seg_metas.len());
+    let mut res = BTreeSet::new();
+    for doc_addr in doc_addrs {
+        let offset = if let Some(offset) = seg_offsets.get(&doc_addr.segment_ord) {
+            *offset
+        } else {
+            // Calculate the offset at the first time meeting the segment and cache it since
+            // the offset is the same for all rows in the same segment.
+            let doc: TantivyDocument = searcher.doc(doc_addr).context(TantivySnafu)?;
+            let rowid = doc
+                .get_first(rowid_field)
+                .and_then(|v| v.as_u64())
+                .context(TantivyDocNotFoundSnafu { doc_addr })?;
+
+            let offset = rowid as u32 - doc_addr.doc_id;
+            seg_offsets.insert(doc_addr.segment_ord, offset);
+            offset
+        };
+
+        res.insert(doc_addr.doc_id + offset);
+    }
+
+    Ok(res)
+}
+
 #[async_trait]
 impl FulltextIndexSearcher for TantivyFulltextIndexSearcher {
     async fn search(&self, query: &str) -> Result<BTreeSet<RowId>> {
-        let searcher = self.reader.searcher();
-        let query_parser = QueryParser::for_index(&self.index, vec![self.default_field]);
-        let query = query_parser
-            .parse_query(query)
-            .context(TantivyParserSnafu)?;
-        let doc_addrs = searcher
-            .search(&query, &DocSetCollector)
-            .context(TantivySnafu)?;
-
-        let seg_metas = self
-            .index
-            .searchable_segment_metas()
-            .context(TantivySnafu)?;
-
-        // FAST PATH: only one segment, the doc id is the same as the row id.
-        //            Also for compatibility with the old version.
-        if seg_metas.len() == 1 {
-            return Ok(doc_addrs.into_iter().map(|d| d.doc_id).collect());
-        }
-
-        // SLOW PATH: multiple segments, need to calculate the row id.
-        let rowid_field = searcher
-            .schema()
-            .get_field(ROWID_FIELD_NAME)
-            .context(TantivySnafu)?;
-        let mut seg_offsets = HashMap::with_capacity(seg_metas.len());
-        let mut res = BTreeSet::new();
-        for doc_addr in doc_addrs {
-            let offset = if let Some(offset) = seg_offsets.get(&doc_addr.segment_ord) {
-                *offset
-            } else {
-                // Calculate the offset at the first time meeting the segment and cache it since
-                // the offset is the same for all rows in the same segment.
-                let doc: TantivyDocument = searcher.doc(doc_addr).context(TantivySnafu)?;
-                let rowid = doc
-                    .get_first(rowid_field)
-                    .and_then(|v| v.as_u64())
-                    .context(TantivyDocNotFoundSnafu { doc_addr })?;
-
-                let offset = rowid as u32 - doc_addr.doc_id;
-                seg_offsets.insert(doc_addr.segment_ord, offset);
-                offset
-            };
-
-            res.insert(doc_addr.doc_id + offset);
-        }
-
-        Ok(res)
+        let inner = self.inner.clone();
+        let query = query.to_string();
+        common_runtime::spawn_blocking_global(move || search_sync(&inner, &query))
+            .await
+            .context(JoinSnafu)?
     }
 }
diff --git a/src/index/src/fulltext_index/tokenizer.rs b/src/index/src/fulltext_index/tokenizer.rs
index b5093afb33..919c497317 100644
--- a/src/index/src/fulltext_index/tokenizer.rs
+++ b/src/index/src/fulltext_index/tokenizer.rs
@@ -82,7 +82,9 @@ impl Tokenizer for EnglishTokenizer {
 
 /// `ChineseTokenizer` tokenizes a Chinese text.
 ///
-/// It uses the Jieba tokenizer to split the text into Chinese words.
+/// It uses Jieba search-mode tokenization to improve recall for Chinese fulltext search.
+/// Enabling HMM also helps merge some unknown fragments into larger tokens, which can reduce
+/// token cardinality versus a fully fragmented output.
 #[derive(Debug, Default)]
 pub struct ChineseTokenizer;
 
@@ -91,11 +93,35 @@ impl Tokenizer for ChineseTokenizer {
         if text.is_ascii() {
             EnglishTokenizer {}.tokenize(text)
         } else {
-            JIEBA.cut(text, false)
+            // Search-mode tokenization emits finer-grained searchable terms, while HMM helps
+            // merge some unknown fragments and avoid excessive token fragmentation.
+            let mut tokens = JIEBA
+                .cut_for_search(text, true)
+                .into_iter()
+                .filter(|s| is_indexable_token(s))
+                .collect::<Vec<_>>();
+
+            let english = EnglishTokenizer {};
+            tokens.extend(
+                english
+                    .tokenize(text)
+                    .into_iter()
+                    .filter(|token| is_ascii_underscore_token(token)),
+            );
+
+            tokens
         }
     }
 }
 
+fn is_indexable_token(token: &str) -> bool {
+    token.chars().any(|c| c.is_alphanumeric() || c == '_')
+}
+
+fn is_ascii_underscore_token(token: &str) -> bool {
+    token.is_ascii() && token.chars().any(|c| c == '_')
+}
+
 /// `Analyzer` analyzes a text into a list of tokens.
 ///
 /// It uses a `Tokenizer` to tokenize the text and optionally lowercases the tokens.
@@ -138,11 +164,26 @@ mod tests {
     #[test]
     fn test_english_tokenizer() {
         let tokenizer = EnglishTokenizer;
-        let text = "Hello, world!!! This is a----++   test012_345+67890";
+        let text = "Hello, world!!! This is a----++   test012_345+67890 ship_ship ship__ship _ __ __IDENTIFIER__ _ship ship_";
         let tokens = tokenizer.tokenize(text);
         assert_eq!(
             tokens,
-            vec!["Hello", "world", "This", "is", "a", "test012_345", "67890"]
+            vec![
+                "Hello",
+                "world",
+                "This",
+                "is",
+                "a",
+                "test012_345",
+                "67890",
+                "ship_ship",
+                "ship__ship",
+                "_",
+                "__",
+                "__IDENTIFIER__",
+                "_ship",
+                "ship_"
+            ]
         );
     }
 
@@ -167,6 +208,331 @@ mod tests {
         assert_eq!(tokens, vec!["我", "喜欢", "苹果"]);
     }
 
+    #[test]
+    fn test_chinese_tokenizer_issue_7943_sample() {
+        let tokenizer = ChineseTokenizer;
+        let text = "[2026/04/09/ 13:56:11.031]2026-04-09 13:56:11.031 - [ trace_id=340a6a44b0bd8e37bb7697ss7da61ff0 span_id=085ff5ttf1e0a23b trace_flags=01] - [http-nio-8081-exec-16] INFO c.h.p.xx.web.service.impl.CCCXForwardKKKServiceImpl.pushout(188) - 登录手机号18888888888的动态key：829889AC8 ship_ship ship__ship _ __ __IDENTIFIER__ _ship ship_ EOF";
+        let tokens = tokenizer.tokenize(text);
+
+        assert_eq!(
+            tokens,
+            vec![
+                "2026",
+                "04",
+                "09",
+                "13",
+                "56",
+                "11.031",
+                "2026-04",
+                "09",
+                "13",
+                "56",
+                "11.031",
+                "trace",
+                "_",
+                "id",
+                "340a6a44b0bd8e37bb7697ss7da61ff0",
+                "span",
+                "_",
+                "id",
+                "085ff5ttf1e0a23b",
+                "trace",
+                "_",
+                "flags",
+                "01",
+                "http",
+                "nio-8081",
+                "exec-16",
+                "INFO",
+                "c",
+                "h",
+                "p",
+                "xx",
+                "web",
+                "service",
+                "impl",
+                "CCCXForwardKKKServiceImpl",
+                "pushout",
+                "188",
+                "登录",
+                "手机",
+                "手机号",
+                "18888888888",
+                "的",
+                "动态",
+                "key",
+                "829889AC8",
+                "ship",
+                "_",
+                "ship",
+                "ship",
+                "__",
+                "ship",
+                "_",
+                "__",
+                "__",
+                "IDENTIFIER",
+                "__",
+                "_",
+                "ship",
+                "ship",
+                "_",
+                "EOF",
+                "trace_id",
+                "span_id",
+                "trace_flags",
+                "ship_ship",
+                "ship__ship",
+                "_",
+                "__",
+                "__IDENTIFIER__",
+                "_ship",
+                "ship_"
+            ]
+        );
+    }
+
+    #[test]
+    fn test_chinese_tokenizer_keeps_ascii_underscore_compounds() {
+        let tokenizer = ChineseTokenizer;
+        let text = "trace_id=abc 登录手机号 dynamic_key=xyz";
+
+        let tokens = tokenizer.tokenize(text);
+
+        assert!(tokens.contains(&"trace_id"));
+        assert!(tokens.contains(&"dynamic_key"));
+        assert!(tokens.contains(&"登录"));
+        assert!(tokens.contains(&"手机号"));
+    }
+
+    #[test]
+    fn test_chinese_tokenizer_skips_non_ascii_underscore_tokens() {
+        let tokenizer = ChineseTokenizer;
+        let text = "登录_id trace_id 手机号_trace";
+
+        let tokens = tokenizer.tokenize(text);
+
+        assert_eq!(
+            tokens,
+            [
+                "登录",
+                "_",
+                "id",
+                "trace",
+                "_",
+                "id",
+                "手机",
+                "手机号",
+                "_",
+                "trace",
+                "trace_id"
+            ]
+        );
+    }
+
+    #[test]
+    fn test_chinese_tokenizer_aggressive_tokenization_probe() {
+        let tokenizer = ChineseTokenizer;
+        let text = "哈基米哦南北绿豆，噢马自立曼波。登录手机号。中国农业银行。装电视台，中国中央广播电视台。压不缩，笑不活。";
+
+        let default_tokens = tokenizer.tokenize(text);
+        let cut_hmm_false = JIEBA.cut(text, false);
+        let cut_hmm_true = JIEBA.cut(text, true);
+        let cut_for_search_hmm_false = JIEBA.cut_for_search(text, false);
+        let cut_for_search_hmm_true = JIEBA.cut_for_search(text, true);
+
+        assert_eq!(
+            default_tokens,
+            [
+                "哈基米",
+                "哦",
+                "南北",
+                "绿豆",
+                "噢",
+                "马",
+                "自立",
+                "曼波",
+                "登录",
+                "手机",
+                "手机号",
+                "中国",
+                "农业",
+                "银行",
+                "中国农业银行",
+                "装",
+                "电视",
+                "电视台",
+                "中国",
+                "中央",
+                "广播",
+                "电视",
+                "电视台",
+                "不缩",
+                "压不缩",
+                "笑",
+                "不活",
+            ]
+        );
+        assert_eq!(
+            cut_hmm_false,
+            [
+                "哈",
+                "基",
+                "米",
+                "哦",
+                "南北",
+                "绿豆",
+                "，",
+                "噢",
+                "马",
+                "自立",
+                "曼",
+                "波",
+                "。",
+                "登录",
+                "手机号",
+                "。",
+                "中国农业银行",
+                "。",
+                "装",
+                "电视台",
+                "，",
+                "中国",
+                "中央",
+                "广播",
+                "电视台",
+                "。",
+                "压",
+                "不",
+                "缩",
+                "，",
+                "笑",
+                "不",
+                "活",
+                "。"
+            ]
+        );
+        assert_eq!(
+            cut_hmm_true,
+            [
+                "哈基米",
+                "哦",
+                "南北",
+                "绿豆",
+                "，",
+                "噢",
+                "马",
+                "自立",
+                "曼波",
+                "。",
+                "登录",
+                "手机号",
+                "。",
+                "中国农业银行",
+                "。",
+                "装",
+                "电视台",
+                "，",
+                "中国",
+                "中央",
+                "广播",
+                "电视台",
+                "。",
+                "压不缩",
+                "，",
+                "笑",
+                "不活",
+                "。"
+            ]
+        );
+        assert_eq!(
+            cut_for_search_hmm_false,
+            [
+                "哈",
+                "基",
+                "米",
+                "哦",
+                "南北",
+                "绿豆",
+                "，",
+                "噢",
+                "马",
+                "自立",
+                "曼",
+                "波",
+                "。",
+                "登录",
+                "手机",
+                "手机号",
+                "。",
+                "中国",
+                "农业",
+                "银行",
+                "中国农业银行",
+                "。",
+                "装",
+                "电视",
+                "电视台",
+                "，",
+                "中国",
+                "中央",
+                "广播",
+                "电视",
+                "电视台",
+                "。",
+                "压",
+                "不",
+                "缩",
+                "，",
+                "笑",
+                "不",
+                "活",
+                "。"
+            ]
+        );
+
+        assert_eq!(
+            cut_for_search_hmm_true,
+            [
+                "哈基米",
+                "哦",
+                "南北",
+                "绿豆",
+                "，",
+                "噢",
+                "马",
+                "自立",
+                "曼波",
+                "。",
+                "登录",
+                "手机",
+                "手机号",
+                "。",
+                "中国",
+                "农业",
+                "银行",
+                "中国农业银行",
+                "。",
+                "装",
+                "电视",
+                "电视台",
+                "，",
+                "中国",
+                "中央",
+                "广播",
+                "电视",
+                "电视台",
+                "。",
+                "不缩",
+                "压不缩",
+                "，",
+                "笑",
+                "不活",
+                "。"
+            ]
+        );
+    }
+
     #[test]
     fn test_valid_ascii_token_lookup_table() {
         // Test all ASCII values in a single loop
diff --git a/src/index/src/inverted_index/format/reader/footer.rs b/src/index/src/inverted_index/format/reader/footer.rs
index 866021c6e6..5d6ac922c9 100644
--- a/src/index/src/inverted_index/format/reader/footer.rs
+++ b/src/index/src/inverted_index/format/reader/footer.rs
@@ -173,7 +173,7 @@ impl<R: RangeReader> InvertedIndexFooterReader<R> {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use prost::Message;
 
diff --git a/src/index/src/inverted_index/search/index_apply/predicates_apply.rs b/src/index/src/inverted_index/search/index_apply/predicates_apply.rs
index 441a4b4304..eaea8cfb95 100644
--- a/src/index/src/inverted_index/search/index_apply/predicates_apply.rs
+++ b/src/index/src/inverted_index/search/index_apply/predicates_apply.rs
@@ -94,15 +94,15 @@ impl IndexApplier for PredicatesIndexApplier {
             .collect::<Vec<_>>();
 
         let mut mapper = ParallelFstValuesMapper::new(reader);
-        let mut bm_vec = mapper.map_values_vec(&value_and_meta_vec, metrics).await?;
+        let bm_vec = mapper.map_values_vec(&value_and_meta_vec, metrics).await?;
 
-        let mut bitmap = bm_vec.pop().unwrap(); // SAFETY: `fst_ranges` is not empty
-        for bm in bm_vec {
-            if bm.count_ones() == 0 {
+        let mut iter = bm_vec.into_iter();
+        let mut bitmap = iter.next().unwrap(); // SAFETY: `fst_ranges` is not empty
+        for bm in iter {
+            bitmap.intersect(bm);
+            if bitmap.count_ones() == 0 {
                 break;
             }
-
-            bitmap.intersect(bm);
         }
 
         output.matched_segment_ids = bitmap;
diff --git a/src/index/src/lib.rs b/src/index/src/lib.rs
index c469acb8c3..7969ece891 100644
--- a/src/index/src/lib.rs
+++ b/src/index/src/lib.rs
@@ -13,7 +13,6 @@
 // limitations under the License.
 
 #![feature(iter_partition_in_place)]
-#![feature(assert_matches)]
 
 pub mod bitmap;
 pub mod bloom_filter;
diff --git a/src/log-query/src/log_query.rs b/src/log-query/src/log_query.rs
index c5b71c6efb..2955b7bded 100644
--- a/src/log-query/src/log_query.rs
+++ b/src/log-query/src/log_query.rs
@@ -199,57 +199,53 @@ impl TimeFilter {
         let mut start_dt = None;
         let mut end_dt = None;
 
-        if self.start.is_some() && self.end.is_none() && self.span.is_none() {
-            // Only 'start' is provided
-            let s = self.start.as_ref().unwrap();
-            let (start, end_opt) = Self::parse_datetime(s)?;
-            if end_opt.is_none() {
+        match (&self.start, &self.end, &self.span) {
+            (Some(start), None, None) => {
+                let (start, end_opt) = Self::parse_datetime(start)?;
+                if end_opt.is_none() {
+                    return Err(InvalidTimeFilterSnafu {
+                        filter: self.clone(),
+                    }
+                    .build());
+                }
+                start_dt = Some(start);
+                end_dt = end_opt;
+            }
+            (Some(start), Some(end), _) => {
+                // Both 'start' and 'end' are provided
+                let (start, _) = Self::parse_datetime(start)?;
+                let (end, _) = Self::parse_datetime(end)?;
+                start_dt = Some(start);
+                end_dt = Some(end);
+            }
+            (Some(start), None, Some(span)) => {
+                let (start, _) = Self::parse_datetime(start)?;
+                let span = Self::parse_span(span)?;
+                let end = start + span;
+                start_dt = Some(start);
+                end_dt = Some(end);
+            }
+            (None, Some(end), Some(span)) => {
+                let (end, _) = Self::parse_datetime(end)?;
+                let span = Self::parse_span(span)?;
+                let start = end - span;
+                start_dt = Some(start);
+                end_dt = Some(end);
+            }
+            (None, None, Some(span)) => {
+                let span = Self::parse_span(span)?;
+                let end = Utc::now();
+                let start = end - span;
+                start_dt = Some(start);
+                end_dt = Some(end);
+            }
+            _ => {
+                // Exception
                 return Err(InvalidTimeFilterSnafu {
                     filter: self.clone(),
                 }
                 .build());
             }
-            start_dt = Some(start);
-            end_dt = end_opt;
-        } else if self.start.is_some() && self.end.is_some() {
-            // Both 'start' and 'end' are provided
-            let (start, _) = Self::parse_datetime(self.start.as_ref().unwrap())?;
-            let (end, _) = Self::parse_datetime(self.end.as_ref().unwrap())?;
-            start_dt = Some(start);
-            end_dt = Some(end);
-        } else if self.span.is_some() && (self.start.is_some() || self.end.is_some()) {
-            // 'span' with 'start' or 'end'
-            let span = Self::parse_span(self.span.as_ref().unwrap())?;
-            if self.start.is_some() {
-                let (start, _) = Self::parse_datetime(self.start.as_ref().unwrap())?;
-                let end = start + span;
-                start_dt = Some(start);
-                end_dt = Some(end);
-            } else {
-                let (end, _) = Self::parse_datetime(self.end.as_ref().unwrap())?;
-                let start = end - span;
-                start_dt = Some(start);
-                end_dt = Some(end);
-            }
-        } else if self.span.is_some() && self.start.is_none() && self.end.is_none() {
-            // Only 'span' is provided
-            let span = Self::parse_span(self.span.as_ref().unwrap())?;
-            let end = Utc::now();
-            let start = end - span;
-            start_dt = Some(start);
-            end_dt = Some(end);
-        } else if self.start.is_some() && self.span.is_some() && self.end.is_some() {
-            // All fields are provided; 'start' and 'end' take priority
-            let (start, _) = Self::parse_datetime(self.start.as_ref().unwrap())?;
-            let (end, _) = Self::parse_datetime(self.end.as_ref().unwrap())?;
-            start_dt = Some(start);
-            end_dt = Some(end);
-        } else {
-            // Exception
-            return Err(InvalidTimeFilterSnafu {
-                filter: self.clone(),
-            }
-            .build());
         }
 
         // Validate that end is after start
diff --git a/src/log-store/src/kafka/index/iterator.rs b/src/log-store/src/kafka/index/iterator.rs
index 9ab350036d..ba8e7273d7 100644
--- a/src/log-store/src/kafka/index/iterator.rs
+++ b/src/log-store/src/kafka/index/iterator.rs
@@ -61,11 +61,7 @@ impl RegionWalRange {
 
     fn next_batch_size(&self) -> Option<u64> {
         if self.current_entry_id < self.end_entry_id {
-            Some(
-                self.end_entry_id
-                    .checked_sub(self.current_entry_id)
-                    .unwrap_or_default(),
-            )
+            Some(self.end_entry_id.saturating_sub(self.current_entry_id))
         } else {
             None
         }
diff --git a/src/log-store/src/kafka/log_store.rs b/src/log-store/src/kafka/log_store.rs
index 702e5bf319..e7fd06816d 100644
--- a/src/log-store/src/kafka/log_store.rs
+++ b/src/log-store/src/kafka/log_store.rs
@@ -550,7 +550,7 @@ fn check_termination(offset: i64, end_offset: i64) -> bool {
 #[cfg(test)]
 mod tests {
 
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
     use std::sync::Arc;
     use std::time::Duration;
diff --git a/src/log-store/src/kafka/util/record.rs b/src/log-store/src/kafka/util/record.rs
index 1e291d9776..720f989139 100644
--- a/src/log-store/src/kafka/util/record.rs
+++ b/src/log-store/src/kafka/util/record.rs
@@ -306,7 +306,7 @@ pub(crate) fn maybe_emit_entry(
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::sync::Arc;
 
     use super::*;
diff --git a/src/log-store/src/lib.rs b/src/log-store/src/lib.rs
index ec8207d5eb..c054fbc1d8 100644
--- a/src/log-store/src/lib.rs
+++ b/src/log-store/src/lib.rs
@@ -12,9 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(io_error_more)]
-#![feature(assert_matches)]
-
 pub mod error;
 pub mod kafka;
 pub mod metrics;
diff --git a/src/meta-client/Cargo.toml b/src/meta-client/Cargo.toml
index 26066c7b96..28c03b3f0d 100644
--- a/src/meta-client/Cargo.toml
+++ b/src/meta-client/Cargo.toml
@@ -17,6 +17,7 @@ common-macro.workspace = true
 common-meta.workspace = true
 common-options.workspace = true
 common-telemetry.workspace = true
+common-time.workspace = true
 futures.workspace = true
 futures-util.workspace = true
 humantime-serde.workspace = true
diff --git a/src/meta-client/examples/meta_client.rs b/src/meta-client/examples/meta_client.rs
index 5db13c9658..e1cea4be4d 100644
--- a/src/meta-client/examples/meta_client.rs
+++ b/src/meta-client/examples/meta_client.rs
@@ -16,10 +16,7 @@ use std::time::Duration;
 
 use api::v1::meta::{HeartbeatRequest, Peer};
 use common_grpc::channel_manager::{ChannelConfig, ChannelManager};
-use common_meta::rpc::store::{
-    BatchDeleteRequest, BatchGetRequest, BatchPutRequest, CompareAndPutRequest, DeleteRangeRequest,
-    PutRequest, RangeRequest,
-};
+use common_meta::rpc::store::{BatchGetRequest, RangeRequest};
 use meta_client::client::MetaClientBuilder;
 use tracing::{Level, event, subscriber};
 use tracing_subscriber::FmtSubscriber;
@@ -67,14 +64,6 @@ async fn run() {
         }
     });
 
-    // put
-    let put = PutRequest::new()
-        .with_key(b"key1".to_vec())
-        .with_value(b"value1".to_vec())
-        .with_prev_kv();
-    let res = meta_client.put(put).await.unwrap();
-    event!(Level::INFO, "put result: {:#?}", res);
-
     // get
     let range = RangeRequest::new().with_key(b"key1".to_vec());
     let res = meta_client.range(range.clone()).await.unwrap();
@@ -85,60 +74,10 @@ async fn run() {
     let res = meta_client.range(range2.clone()).await.unwrap();
     event!(Level::INFO, "get prefix result: {:#?}", res);
 
-    // batch put
-    let batch_put = BatchPutRequest::new()
-        .add_kv(b"batch_put1".to_vec(), b"batch_put_v1".to_vec())
-        .add_kv(b"batch_put2".to_vec(), b"batch_put_v2".to_vec())
-        .with_prev_kv();
-    let res = meta_client.batch_put(batch_put).await.unwrap();
-    event!(Level::INFO, "batch put result: {:#?}", res);
-
-    // cas
-    let cas = CompareAndPutRequest::new()
-        .with_key(b"batch_put1".to_vec())
-        .with_expect(b"batch_put_v_fail".to_vec())
-        .with_value(b"batch_put_v111".to_vec());
-
-    let res = meta_client.compare_and_put(cas).await.unwrap();
-    event!(Level::INFO, "cas 0 result: {:#?}", res);
-
-    let cas = CompareAndPutRequest::new()
-        .with_key(b"batch_put1".to_vec())
-        .with_expect(b"batch_put_v1".to_vec())
-        .with_value(b"batch_put_v111".to_vec());
-
-    let res = meta_client.compare_and_put(cas).await.unwrap();
-    event!(Level::INFO, "cas 1 result: {:#?}", res);
-
-    // delete
-    let delete_range = DeleteRangeRequest::new().with_key(b"key1".to_vec());
-    let res = meta_client.delete_range(delete_range).await.unwrap();
-    event!(Level::INFO, "delete range result: {:#?}", res);
-
-    // get none
-    let res = meta_client.range(range).await.unwrap();
-    event!(Level::INFO, "get range result: {:#?}", res);
-
-    // batch delete
-    // put two
-    let batch_put = BatchPutRequest::new()
-        .add_kv(b"batch_put1".to_vec(), b"batch_put_v1".to_vec())
-        .add_kv(b"batch_put2".to_vec(), b"batch_put_v2".to_vec())
-        .with_prev_kv();
-    let res = meta_client.batch_put(batch_put).await.unwrap();
-    event!(Level::INFO, "batch put result: {:#?}", res);
-
-    // delete one
-    let batch_delete = BatchDeleteRequest::new()
-        .add_key(b"batch_put1".to_vec())
-        .with_prev_kv();
-    let res = meta_client.batch_delete(batch_delete).await.unwrap();
-    event!(Level::INFO, "batch delete result: {:#?}", res);
-
-    // get other one
+    // batch get
     let batch_get = BatchGetRequest::new()
-        .add_key(b"batch_put1".to_vec())
-        .add_key(b"batch_put2".to_vec());
+        .add_key(b"key1".to_vec())
+        .add_key(b"key2".to_vec());
 
     let res = meta_client.batch_get(batch_get).await.unwrap();
     event!(Level::INFO, "batch get result: {:#?}", res);
diff --git a/src/meta-client/src/client.rs b/src/meta-client/src/client.rs
index 4a79105615..46b537a68d 100644
--- a/src/meta-client/src/client.rs
+++ b/src/meta-client/src/client.rs
@@ -26,6 +26,7 @@ use std::fmt::Debug;
 use std::sync::Arc;
 use std::time::Duration;
 
+use api::v1::meta::heartbeat_request::NodeWorkloads;
 use api::v1::meta::{
     MetasrvNodeInfo, ProcedureDetailResponse, ReconcileRequest, ReconcileResponse, Role,
 };
@@ -38,11 +39,13 @@ use common_meta::cluster::{
     ClusterInfo, MetasrvStatus, NodeInfo, NodeInfoKey, NodeStatus, Role as ClusterRole,
 };
 use common_meta::datanode::{DatanodeStatKey, DatanodeStatValue, RegionStat};
+use common_meta::distributed_time_constants::default_distributed_time_constants;
 use common_meta::error::{
     self as meta_error, ExternalSnafu, Result as MetaResult, UnsupportedSnafu,
 };
 use common_meta::key::flow::flow_state::{FlowStat, FlowStateManager};
 use common_meta::kv_backend::KvBackendRef;
+use common_meta::peer::{Peer, PeerDiscovery};
 use common_meta::procedure_executor::{ExecutorContext, ProcedureExecutor};
 use common_meta::range_stream::PaginationStream;
 use common_meta::rpc::KeyValue;
@@ -59,6 +62,7 @@ use common_meta::rpc::store::{
 };
 use common_options::plugin_options::PluginOptionsDeserializer;
 use common_telemetry::info;
+use common_time::util::DefaultSystemTimer;
 use config::Client as ConfigClient;
 use futures::TryStreamExt;
 use heartbeat::{Client as HeartbeatClient, HeartbeatConfig};
@@ -87,6 +91,8 @@ pub struct MetaClientBuilder {
     role: Role,
     enable_heartbeat: bool,
     enable_store: bool,
+    #[cfg(test)]
+    enable_direct_store_writes: bool,
     enable_procedure: bool,
     enable_access_cluster_info: bool,
     region_follower: Option<RegionFollowerClientRef>,
@@ -139,6 +145,10 @@ impl MetaClientBuilder {
         }
     }
 
+    /// Enables the Store client in read-only mode.
+    ///
+    /// Store write methods fail fast by default. Metadata writes from production
+    /// frontend/datanode/flownode clients should go through metasrv procedures.
     pub fn enable_store(self) -> Self {
         Self {
             enable_store: true,
@@ -146,6 +156,18 @@ impl MetaClientBuilder {
         }
     }
 
+    /// Enables direct Store write RPCs for tests.
+    ///
+    /// Production metadata writes should use metasrv-owned write paths instead.
+    #[cfg(test)]
+    pub(super) fn enable_direct_store_writes_for_test(self) -> Self {
+        Self {
+            enable_store: true,
+            enable_direct_store_writes: true,
+            ..self
+        }
+    }
+
     pub fn enable_procedure(self) -> Self {
         Self {
             enable_procedure: true,
@@ -212,9 +234,16 @@ impl MetaClientBuilder {
         let config = self
             .enable_heartbeat
             .then(|| ConfigClient::new(self.id, self.role, mgr.clone()));
-        let store = self
-            .enable_store
-            .then(|| StoreClient::new(self.id, self.role, mgr.clone()));
+        let store = self.enable_store.then(|| {
+            #[cfg(test)]
+            {
+                if self.enable_direct_store_writes {
+                    return StoreClient::new_writable(self.id, self.role, mgr.clone());
+                }
+            }
+
+            StoreClient::new(self.id, self.role, mgr.clone())
+        });
         let procedure = self.enable_procedure.then(|| {
             let mgr = self.ddl_channel_manager.unwrap_or(mgr.clone());
             ProcedureClient::new(
@@ -456,6 +485,7 @@ impl ClusterInfo for MetaClient {
                             cpu_usage_millicores: node_info.cpu_usage_millicores,
                             memory_usage_bytes: node_info.memory_usage_bytes,
                             hostname: node_info.hostname,
+                            env_vars: Default::default(),
                         }
                     } else {
                         // TODO(zyy17): It's for backward compatibility. Remove this when the deprecated fields are removed from the proto.
@@ -471,6 +501,7 @@ impl ClusterInfo for MetaClient {
                             cpu_usage_millicores: 0,
                             memory_usage_bytes: 0,
                             hostname: "".to_string(),
+                            env_vars: Default::default(),
                         }
                     }
                 })
@@ -488,6 +519,7 @@ impl ClusterInfo for MetaClient {
                             cpu_usage_millicores: node_info.cpu_usage_millicores,
                             memory_usage_bytes: node_info.memory_usage_bytes,
                             hostname: node_info.hostname,
+                            env_vars: Default::default(),
                         }
                     } else {
                         // TODO(zyy17): It's for backward compatibility. Remove this when the deprecated fields are removed from the proto.
@@ -503,6 +535,7 @@ impl ClusterInfo for MetaClient {
                             cpu_usage_millicores: 0,
                             memory_usage_bytes: 0,
                             hostname: "".to_string(),
+                            env_vars: Default::default(),
                         }
                     }
                 }))
@@ -553,6 +586,60 @@ impl ClusterInfo for MetaClient {
     }
 }
 
+// TODO(weny): the discovery using client side timestamp may be inaccurate,
+// maybe we need to use the timestamp from metasrv in the future.
+#[async_trait::async_trait]
+impl PeerDiscovery for MetaClient {
+    async fn active_frontends(&self) -> MetaResult<Vec<Peer>> {
+        let nodes = self
+            .list_nodes(Some(ClusterRole::Frontend))
+            .await
+            .map_err(BoxedError::new)
+            .context(ExternalSnafu)?;
+        Ok(util::alive_frontends(
+            &DefaultSystemTimer,
+            nodes,
+            // TODO(weny): the heartbeat interval should be received from metasrv
+            // instead of using the default value.
+            default_distributed_time_constants().frontend_heartbeat_interval,
+        ))
+    }
+
+    async fn active_datanodes(
+        &self,
+        filter: Option<for<'a> fn(&'a NodeWorkloads) -> bool>,
+    ) -> MetaResult<Vec<Peer>> {
+        let nodes = self
+            .list_nodes(Some(ClusterRole::Datanode))
+            .await
+            .map_err(BoxedError::new)
+            .context(ExternalSnafu)?;
+        Ok(util::alive_datanodes(
+            &DefaultSystemTimer,
+            nodes,
+            default_distributed_time_constants().datanode_lease,
+            filter,
+        ))
+    }
+
+    async fn active_flownodes(
+        &self,
+        filter: Option<for<'a> fn(&'a NodeWorkloads) -> bool>,
+    ) -> MetaResult<Vec<Peer>> {
+        let nodes = self
+            .list_nodes(Some(ClusterRole::Flownode))
+            .await
+            .map_err(BoxedError::new)
+            .context(ExternalSnafu)?;
+        Ok(util::alive_flownodes(
+            &DefaultSystemTimer,
+            nodes,
+            default_distributed_time_constants().flownode_lease,
+            filter,
+        ))
+    }
+}
+
 fn decode_stats(kv: KeyValue) -> MetaResult<DatanodeStatValue> {
     DatanodeStatValue::try_from(kv.value)
         .map_err(BoxedError::new)
@@ -927,6 +1014,19 @@ mod tests {
         assert!(matches!(res.err(), Some(error::Error::NotStarted { .. })));
     }
 
+    #[tokio::test]
+    async fn test_store_writes_are_read_only_by_default() {
+        let meta_client = MetaClientBuilder::new(0, Role::Datanode)
+            .enable_store()
+            .build();
+
+        let res = meta_client.put(PutRequest::default()).await;
+        assert!(matches!(
+            res.err(),
+            Some(error::Error::ReadOnlyKvBackend { .. })
+        ));
+    }
+
     #[tokio::test]
     async fn test_ask_leader() {
         let tc = new_client("test_ask_leader").await;
diff --git a/src/meta-client/src/client/cluster.rs b/src/meta-client/src/client/cluster.rs
index 8da45aaa7a..976626b3a0 100644
--- a/src/meta-client/src/client/cluster.rs
+++ b/src/meta-client/src/client/cluster.rs
@@ -212,8 +212,8 @@ impl Inner {
                         }
                     }
                 }
-            } else if let Err(err) = leader_provider.ask_leader().await {
-                return Err(err);
+            } else {
+                leader_provider.ask_leader().await?;
             }
         }
 
diff --git a/src/meta-client/src/client/procedure.rs b/src/meta-client/src/client/procedure.rs
index 93e37511d9..63f45c28fb 100644
--- a/src/meta-client/src/client/procedure.rs
+++ b/src/meta-client/src/client/procedure.rs
@@ -208,8 +208,8 @@ impl Inner {
                         }
                     }
                 }
-            } else if let Err(err) = leader_provider.ask_leader().await {
-                return Err(err);
+            } else {
+                leader_provider.ask_leader().await?;
             }
         }
 
diff --git a/src/meta-client/src/client/store.rs b/src/meta-client/src/client/store.rs
index a7cf449629..2aee8d8323 100644
--- a/src/meta-client/src/client/store.rs
+++ b/src/meta-client/src/client/store.rs
@@ -39,11 +39,27 @@ pub struct Client {
 
 impl Client {
     pub fn new(id: Id, role: Role, channel_manager: ChannelManager) -> Self {
+        Self::new_with_read_only(id, role, channel_manager, true)
+    }
+
+    /// Builds a writable direct Store RPC client for tests.
+    #[cfg(test)]
+    pub(super) fn new_writable(id: Id, role: Role, channel_manager: ChannelManager) -> Self {
+        Self::new_with_read_only(id, role, channel_manager, false)
+    }
+
+    fn new_with_read_only(
+        id: Id,
+        role: Role,
+        channel_manager: ChannelManager,
+        read_only: bool,
+    ) -> Self {
         let inner = Arc::new(RwLock::new(Inner {
             id,
             role,
             channel_manager,
             peers: vec![],
+            read_only,
         }));
 
         Self { inner }
@@ -103,6 +119,7 @@ struct Inner {
     role: Role,
     channel_manager: ChannelManager,
     peers: Vec<String>,
+    read_only: bool,
 }
 
 impl Inner {
@@ -142,6 +159,8 @@ impl Inner {
     }
 
     async fn put(&self, mut req: PutRequest) -> Result<PutResponse> {
+        self.ensure_writable()?;
+
         let mut client = self.random_client()?;
         req.set_header(
             self.id,
@@ -167,6 +186,8 @@ impl Inner {
     }
 
     async fn batch_put(&self, mut req: BatchPutRequest) -> Result<BatchPutResponse> {
+        self.ensure_writable()?;
+
         let mut client = self.random_client()?;
         req.set_header(
             self.id,
@@ -179,6 +200,8 @@ impl Inner {
     }
 
     async fn batch_delete(&self, mut req: BatchDeleteRequest) -> Result<BatchDeleteResponse> {
+        self.ensure_writable()?;
+
         let mut client = self.random_client()?;
         req.set_header(
             self.id,
@@ -194,6 +217,8 @@ impl Inner {
         &self,
         mut req: CompareAndPutRequest,
     ) -> Result<CompareAndPutResponse> {
+        self.ensure_writable()?;
+
         let mut client = self.random_client()?;
         req.set_header(
             self.id,
@@ -209,6 +234,8 @@ impl Inner {
     }
 
     async fn delete_range(&self, mut req: DeleteRangeRequest) -> Result<DeleteRangeResponse> {
+        self.ensure_writable()?;
+
         let mut client = self.random_client()?;
         req.set_header(
             self.id,
@@ -231,6 +258,17 @@ impl Inner {
         self.make_client(peer)
     }
 
+    fn ensure_writable(&self) -> Result<()> {
+        if self.read_only {
+            return error::ReadOnlyKvBackendSnafu {
+                name: "MetaClient Store".to_string(),
+            }
+            .fail();
+        }
+
+        Ok(())
+    }
+
     fn make_client(&self, addr: impl AsRef<str>) -> Result<StoreClient<Channel>> {
         let channel = self
             .channel_manager
@@ -284,4 +322,26 @@ mod test {
             .unwrap();
         assert_eq!(1, client.inner.write().await.peers.len());
     }
+
+    #[tokio::test]
+    async fn test_read_only_store_rejects_writes_before_rpc() {
+        let client = Client::new(0, Role::Frontend, ChannelManager::default());
+
+        fn assert_read_only<T>(result: Result<T>) {
+            assert!(matches!(
+                result,
+                Err(error::Error::ReadOnlyKvBackend { .. })
+            ));
+        }
+
+        assert_read_only(client.put(PutRequest::default()).await);
+        assert_read_only(client.batch_put(BatchPutRequest::default()).await);
+        assert_read_only(client.batch_delete(BatchDeleteRequest::default()).await);
+        assert_read_only(
+            client
+                .compare_and_put(CompareAndPutRequest::default())
+                .await,
+        );
+        assert_read_only(client.delete_range(DeleteRangeRequest::default()).await);
+    }
 }
diff --git a/src/meta-client/src/client/util.rs b/src/meta-client/src/client/util.rs
index 758e1f9de1..aa08871e04 100644
--- a/src/meta-client/src/client/util.rs
+++ b/src/meta-client/src/client/util.rs
@@ -12,7 +12,11 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+use api::v1::meta::heartbeat_request::NodeWorkloads;
 use api::v1::meta::{ErrorCode, ResponseHeader};
+use common_meta::cluster::{NodeInfo, NodeStatus};
+use common_meta::peer::Peer;
+use common_time::util::SystemTimer;
 use tonic::{Code, Status};
 
 pub(crate) fn is_unreachable(status: &Status) -> bool {
@@ -30,3 +34,255 @@ pub(crate) fn is_not_leader(header: &Option<ResponseHeader>) -> bool {
 
     err.code == ErrorCode::NotLeader as i32
 }
+
+fn is_active_node(
+    timer: &impl SystemTimer,
+    last_activity_ts: i64,
+    active_duration: std::time::Duration,
+) -> bool {
+    let now = timer.current_time_millis();
+    let elapsed = now.checked_sub(last_activity_ts).unwrap_or(0) as u64;
+    elapsed < active_duration.as_millis() as u64
+}
+
+pub(crate) fn alive_frontends(
+    timer: &impl SystemTimer,
+    nodes: Vec<NodeInfo>,
+    active_duration: std::time::Duration,
+) -> Vec<Peer> {
+    nodes
+        .into_iter()
+        .filter_map(|node| {
+            if matches!(node.status, NodeStatus::Frontend(_))
+                && is_active_node(timer, node.last_activity_ts, active_duration)
+            {
+                Some(node.peer)
+            } else {
+                None
+            }
+        })
+        .collect()
+}
+
+pub(crate) fn alive_datanodes(
+    timer: &impl SystemTimer,
+    nodes: Vec<NodeInfo>,
+    active_duration: std::time::Duration,
+    filter: Option<for<'a> fn(&'a NodeWorkloads) -> bool>,
+) -> Vec<Peer> {
+    let filter = filter.unwrap_or(|_| true);
+
+    nodes
+        .into_iter()
+        .filter_map(|node| {
+            if let NodeStatus::Datanode(status) = node.status
+                && is_active_node(timer, node.last_activity_ts, active_duration)
+            {
+                let workloads = NodeWorkloads::Datanode(status.workloads);
+                filter(&workloads).then_some(node.peer)
+            } else {
+                None
+            }
+        })
+        .collect()
+}
+
+pub(crate) fn alive_flownodes(
+    timer: &impl SystemTimer,
+    nodes: Vec<NodeInfo>,
+    active_duration: std::time::Duration,
+    filter: Option<for<'a> fn(&'a NodeWorkloads) -> bool>,
+) -> Vec<Peer> {
+    let filter = filter.unwrap_or(|_| true);
+
+    nodes
+        .into_iter()
+        .filter_map(|node| {
+            if let NodeStatus::Flownode(status) = node.status
+                && is_active_node(timer, node.last_activity_ts, active_duration)
+            {
+                let workloads = NodeWorkloads::Flownode(status.workloads);
+                filter(&workloads).then_some(node.peer)
+            } else {
+                None
+            }
+        })
+        .collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use std::time::Duration;
+
+    use api::v1::meta::heartbeat_request::NodeWorkloads;
+    use api::v1::meta::{DatanodeWorkloads, FlownodeWorkloads, Peer};
+    use common_meta::cluster::{
+        DatanodeStatus, FlownodeStatus, FrontendStatus, MetasrvStatus, NodeInfo, NodeStatus, Role,
+    };
+    use common_time::util::SystemTimer;
+
+    use super::*;
+
+    struct MockSystemTimer(i64);
+
+    impl MockSystemTimer {
+        fn new(now: i64) -> Self {
+            Self(now)
+        }
+    }
+
+    impl SystemTimer for MockSystemTimer {
+        fn current_time_millis(&self) -> i64 {
+            self.0
+        }
+
+        fn current_time_rfc3339(&self) -> String {
+            "1970-01-01T00:00:00Z".to_string()
+        }
+    }
+
+    fn node_info(role: Role, id: u64, addr: &str, last_activity_ts: i64) -> NodeInfo {
+        let status = match role {
+            Role::Frontend => NodeStatus::Frontend(FrontendStatus {}),
+            Role::Datanode => NodeStatus::Datanode(DatanodeStatus {
+                rcus: 0,
+                wcus: 0,
+                leader_regions: 0,
+                follower_regions: 0,
+                workloads: DatanodeWorkloads { types: vec![] },
+            }),
+            Role::Flownode => NodeStatus::Flownode(FlownodeStatus {
+                workloads: FlownodeWorkloads { types: vec![] },
+            }),
+            Role::Metasrv => NodeStatus::Metasrv(MetasrvStatus { is_leader: false }),
+        };
+
+        NodeInfo {
+            peer: Peer::new(id, addr),
+            last_activity_ts,
+            status,
+            version: String::new(),
+            git_commit: String::new(),
+            start_time_ms: 0,
+            total_cpu_millicores: 0,
+            total_memory_bytes: 0,
+            cpu_usage_millicores: 0,
+            memory_usage_bytes: 0,
+            hostname: String::new(),
+            env_vars: Default::default(),
+        }
+    }
+
+    fn ingest_only(workloads: &NodeWorkloads) -> bool {
+        matches!(
+            workloads,
+            NodeWorkloads::Datanode(DatanodeWorkloads { types }) if types.as_slice() == [1]
+        )
+    }
+
+    fn empty_flownode_workloads(workloads: &NodeWorkloads) -> bool {
+        matches!(
+            workloads,
+            NodeWorkloads::Flownode(FlownodeWorkloads { types }) if types.is_empty()
+        )
+    }
+
+    #[test]
+    fn test_alive_frontends_filters_by_activity_and_role() {
+        let timer = MockSystemTimer::new(100);
+        let peers = alive_frontends(
+            &timer,
+            vec![
+                node_info(Role::Frontend, 1, "127.0.0.1:3001", 95),
+                node_info(Role::Frontend, 2, "127.0.0.1:3002", 89),
+                node_info(Role::Datanode, 3, "127.0.0.1:4001", 99),
+            ],
+            Duration::from_millis(10),
+        );
+
+        assert_eq!(
+            vec![1],
+            peers.into_iter().map(|peer| peer.id).collect::<Vec<_>>()
+        );
+    }
+
+    #[test]
+    fn test_alive_datanodes_filters_by_activity_and_workload() {
+        let timer = MockSystemTimer::new(100);
+        let mut first = node_info(Role::Datanode, 1, "127.0.0.1:4001", 95);
+        let mut second = node_info(Role::Datanode, 2, "127.0.0.1:4002", 95);
+        let stale = node_info(Role::Datanode, 3, "127.0.0.1:4003", 89);
+
+        if let NodeStatus::Datanode(status) = &mut first.status {
+            status.workloads = DatanodeWorkloads { types: vec![1] };
+        }
+        if let NodeStatus::Datanode(status) = &mut second.status {
+            status.workloads = DatanodeWorkloads { types: vec![2] };
+        }
+
+        let peers = alive_datanodes(
+            &timer,
+            vec![first, second, stale],
+            Duration::from_millis(10),
+            Some(ingest_only),
+        );
+
+        assert_eq!(
+            vec![1],
+            peers.into_iter().map(|peer| peer.id).collect::<Vec<_>>()
+        );
+    }
+
+    #[test]
+    fn test_alive_flownodes_uses_empty_workload_semantics() {
+        let timer = MockSystemTimer::new(100);
+        let peers = alive_flownodes(
+            &timer,
+            vec![
+                node_info(Role::Flownode, 1, "127.0.0.1:5001", 95),
+                node_info(Role::Flownode, 2, "127.0.0.1:5002", 89),
+                node_info(Role::Frontend, 3, "127.0.0.1:3001", 99),
+            ],
+            Duration::from_millis(10),
+            Some(empty_flownode_workloads),
+        );
+
+        assert_eq!(
+            vec![1],
+            peers.into_iter().map(|peer| peer.id).collect::<Vec<_>>()
+        );
+    }
+
+    #[test]
+    fn test_alive_flownodes_filters_by_workloads() {
+        let timer = MockSystemTimer::new(100);
+        let mut first = node_info(Role::Flownode, 1, "127.0.0.1:5001", 95);
+        let mut second = node_info(Role::Flownode, 2, "127.0.0.1:5002", 95);
+
+        if let NodeStatus::Flownode(status) = &mut first.status {
+            status.workloads = FlownodeWorkloads { types: vec![7] };
+        }
+        if let NodeStatus::Flownode(status) = &mut second.status {
+            status.workloads = FlownodeWorkloads { types: vec![8] };
+        }
+
+        fn workload_type_is_7(workloads: &NodeWorkloads) -> bool {
+            matches!(
+                workloads,
+                NodeWorkloads::Flownode(FlownodeWorkloads { types }) if types.as_slice() == [7]
+            )
+        }
+
+        let peers = alive_flownodes(
+            &timer,
+            vec![first, second],
+            Duration::from_millis(10),
+            Some(workload_type_is_7),
+        );
+
+        assert_eq!(
+            vec![1],
+            peers.into_iter().map(|peer| peer.id).collect::<Vec<_>>()
+        );
+    }
+}
diff --git a/src/meta-client/src/error.rs b/src/meta-client/src/error.rs
index 95c820c039..19577f672c 100644
--- a/src/meta-client/src/error.rs
+++ b/src/meta-client/src/error.rs
@@ -144,9 +144,10 @@ impl ErrorExt for Error {
             | Error::CreateHeartbeatStream { .. }
             | Error::CreateChannel { .. }
             | Error::RetryTimesExceeded { .. }
-            | Error::ReadOnlyKvBackend { .. }
             | Error::ConvertMetaConfig { .. } => StatusCode::Internal,
 
+            Error::ReadOnlyKvBackend { .. } => StatusCode::Unsupported,
+
             Error::MetaServer { code, .. } => *code,
 
             Error::InvalidResponseHeader { source, .. }
diff --git a/src/meta-client/src/mocks.rs b/src/meta-client/src/mocks.rs
index 294ac16ef9..2c80091c58 100644
--- a/src/meta-client/src/mocks.rs
+++ b/src/meta-client/src/mocks.rs
@@ -63,6 +63,7 @@ pub async fn mock_client_by(server_addr: String, channel_manager: ChannelManager
     let id = 2000u64;
     let mut meta_client = MetaClientBuilder::datanode_default_options(id)
         .enable_access_cluster_info()
+        .enable_direct_store_writes_for_test()
         .channel_manager(channel_manager)
         .build();
     meta_client.start(&[&server_addr]).await.unwrap();
diff --git a/src/meta-srv/src/bootstrap.rs b/src/meta-srv/src/bootstrap.rs
index 2cfe7d2f7d..a9cc72ec53 100644
--- a/src/meta-srv/src/bootstrap.rs
+++ b/src/meta-srv/src/bootstrap.rs
@@ -24,6 +24,7 @@ use common_base::Plugins;
 use common_config::Configurable;
 #[cfg(any(feature = "pg_kvbackend", feature = "mysql_kvbackend"))]
 use common_meta::distributed_time_constants::META_LEASE_SECS;
+use common_meta::election::etcd::EtcdElection;
 use common_meta::kv_backend::chroot::ChrootKvBackend;
 use common_meta::kv_backend::etcd::EtcdStore;
 use common_meta::kv_backend::memory::MemoryKvBackend;
@@ -42,9 +43,6 @@ use tonic::codec::CompressionEncoding;
 use tonic::transport::server::{Router, TcpIncoming};
 
 use crate::cluster::{MetaPeerClientBuilder, MetaPeerClientRef};
-#[cfg(any(feature = "pg_kvbackend", feature = "mysql_kvbackend"))]
-use crate::election::CANDIDATE_LEASE_SECS;
-use crate::election::etcd::EtcdElection;
 use crate::error::OtherSnafu;
 use crate::metasrv::builder::MetasrvBuilder;
 use crate::metasrv::{
@@ -281,7 +279,8 @@ pub async fn metasrv_builder(
                 etcd_client,
                 opts.store_key_prefix.clone(),
             )
-            .await?;
+            .await
+            .context(error::KvBackendSnafu)?;
 
             (kv_backend, Some(election))
         }
@@ -290,16 +289,12 @@ pub async fn metasrv_builder(
             use std::time::Duration;
 
             use common_meta::distributed_time_constants::POSTGRES_KEEP_ALIVE_SECS;
-            use common_meta::kv_backend::rds::PgStore;
+            use common_meta::election::CANDIDATE_LEASE_SECS;
             use deadpool_postgres::{Config, ManagerConfig, RecyclingMethod};
 
-            use crate::election::rds::postgres::{ElectionPgClient, PgElection};
-            use crate::utils::postgres::create_postgres_pool;
+            use crate::utils::postgres::{build_postgres_election, build_postgres_kv_backend};
 
             let candidate_lease_ttl = Duration::from_secs(CANDIDATE_LEASE_SECS);
-            let execution_timeout = Duration::from_secs(META_LEASE_SECS);
-            let statement_timeout = Duration::from_secs(META_LEASE_SECS);
-            let idle_session_timeout = Duration::from_secs(META_LEASE_SECS);
             let meta_lease_ttl = Duration::from_secs(META_LEASE_SECS);
 
             let mut cfg = Config::new();
@@ -308,23 +303,12 @@ pub async fn metasrv_builder(
             cfg.manager = Some(ManagerConfig {
                 recycling_method: RecyclingMethod::Verified,
             });
-            // Use a dedicated pool for the election client to allow customized session settings.
-            let pool = create_postgres_pool(
+
+            let election = build_postgres_election(
                 &opts.store_addrs,
                 Some(cfg.clone()),
                 opts.backend_tls.clone(),
-            )
-            .await?;
-
-            let election_client = ElectionPgClient::new(
-                pool,
-                execution_timeout,
-                idle_session_timeout,
-                statement_timeout,
-            )?;
-            let election = PgElection::with_pg_client(
                 opts.grpc.server_addr.clone(),
-                election_client,
                 opts.store_key_prefix.clone(),
                 candidate_lease_ttl,
                 meta_lease_ttl,
@@ -334,17 +318,16 @@ pub async fn metasrv_builder(
             )
             .await?;
 
-            let pool = create_postgres_pool(&opts.store_addrs, Some(cfg), opts.backend_tls.clone())
-                .await?;
-            let kv_backend = PgStore::with_pg_pool(
-                pool,
+            let kv_backend = build_postgres_kv_backend(
+                &opts.store_addrs,
+                Some(cfg),
+                opts.backend_tls.clone(),
                 opts.meta_schema_name.as_deref(),
                 &opts.meta_table_name,
                 opts.max_txn_ops,
                 opts.auto_create_schema,
             )
-            .await
-            .context(error::KvBackendSnafu)?;
+            .await?;
 
             (kv_backend, Some(election))
         }
@@ -352,42 +335,32 @@ pub async fn metasrv_builder(
         (None, BackendImpl::MysqlStore) => {
             use std::time::Duration;
 
-            use common_meta::kv_backend::rds::MySqlStore;
+            use common_meta::election::CANDIDATE_LEASE_SECS;
 
-            use crate::election::rds::mysql::{ElectionMysqlClient, MySqlElection};
-            use crate::utils::mysql::create_mysql_pool;
+            use crate::utils::mysql::{build_mysql_election, build_mysql_kv_backend};
 
-            let pool = create_mysql_pool(&opts.store_addrs, opts.backend_tls.as_ref()).await?;
-            let kv_backend =
-                MySqlStore::with_mysql_pool(pool, &opts.meta_table_name, opts.max_txn_ops)
-                    .await
-                    .context(error::KvBackendSnafu)?;
+            let kv_backend = build_mysql_kv_backend(
+                &opts.store_addrs,
+                opts.backend_tls.as_ref(),
+                &opts.meta_table_name,
+                opts.max_txn_ops,
+            )
+            .await?;
             // Since election will acquire a lock of the table, we need a separate table for election.
             let election_table_name = opts.meta_table_name.clone() + "_election";
-            // We use a separate pool for election since we need a different session keep-alive idle time.
-            let pool = create_mysql_pool(&opts.store_addrs, opts.backend_tls.as_ref()).await?;
-            let execution_timeout = Duration::from_secs(META_LEASE_SECS);
-            let statement_timeout = Duration::from_secs(META_LEASE_SECS);
-            let idle_session_timeout = Duration::from_secs(META_LEASE_SECS);
             let innode_lock_wait_timeout = Duration::from_secs(META_LEASE_SECS / 2);
             let meta_lease_ttl = Duration::from_secs(META_LEASE_SECS);
             let candidate_lease_ttl = Duration::from_secs(CANDIDATE_LEASE_SECS);
 
-            let election_client = ElectionMysqlClient::new(
-                pool,
-                execution_timeout,
-                statement_timeout,
-                innode_lock_wait_timeout,
-                idle_session_timeout,
-                &election_table_name,
-            );
-            let election = MySqlElection::with_mysql_client(
+            let election = build_mysql_election(
+                &opts.store_addrs,
+                opts.backend_tls.as_ref(),
                 opts.grpc.server_addr.clone(),
-                election_client,
                 opts.store_key_prefix.clone(),
                 candidate_lease_ttl,
                 meta_lease_ttl,
                 &election_table_name,
+                innode_lock_wait_timeout,
             )
             .await?;
             (kv_backend, Some(election))
diff --git a/src/meta-srv/src/cluster.rs b/src/meta-srv/src/cluster.rs
index 35b15b3b29..ef3ba07702 100644
--- a/src/meta-srv/src/cluster.rs
+++ b/src/meta-srv/src/cluster.rs
@@ -247,7 +247,7 @@ impl MetaPeerClient {
         // Safety: when self.is_leader() == false, election must not empty.
         let election = self.election.as_ref().unwrap();
 
-        let leader_addr = election.leader().await?.0;
+        let leader_addr = election.leader().await.context(error::KvBackendSnafu)?.0;
 
         let channel = self
             .channel_manager
@@ -279,7 +279,7 @@ impl MetaPeerClient {
         // Safety: when self.is_leader() == false, election must not empty.
         let election = self.election.as_ref().unwrap();
 
-        let leader_addr = election.leader().await?.0;
+        let leader_addr = election.leader().await.context(error::KvBackendSnafu)?.0;
 
         let channel = self
             .channel_manager
diff --git a/src/meta-srv/src/discovery/lease.rs b/src/meta-srv/src/discovery/lease.rs
index b8ca8a0ebb..9fcc52f423 100644
--- a/src/meta-srv/src/discovery/lease.rs
+++ b/src/meta-srv/src/discovery/lease.rs
@@ -379,6 +379,7 @@ mod tests {
             cpu_usage_millicores: 0,
             memory_usage_bytes: 0,
             hostname: "test_hostname".to_string(),
+            env_vars: Default::default(),
         };
 
         let key_prefix = NodeInfoKey::key_prefix_with_role(Role::Frontend);
@@ -407,6 +408,7 @@ mod tests {
             cpu_usage_millicores: 0,
             memory_usage_bytes: 0,
             hostname: "test_hostname".to_string(),
+            env_vars: Default::default(),
         };
 
         in_memory
@@ -450,6 +452,7 @@ mod tests {
             cpu_usage_millicores: 0,
             memory_usage_bytes: 0,
             hostname: "test_hostname".to_string(),
+            env_vars: Default::default(),
         };
         let key_prefix = NodeInfoKey::key_prefix_with_role(Role::Frontend);
         in_memory
@@ -492,6 +495,7 @@ mod tests {
             cpu_usage_millicores: 0,
             memory_usage_bytes: 0,
             hostname: "test_hostname".to_string(),
+            env_vars: Default::default(),
         };
 
         let key_prefix = NodeInfoKey::key_prefix_with_role(Role::Frontend);
diff --git a/src/meta-srv/src/error.rs b/src/meta-srv/src/error.rs
index 7b7983b1ba..a0f800f981 100644
--- a/src/meta-srv/src/error.rs
+++ b/src/meta-srv/src/error.rs
@@ -1136,6 +1136,12 @@ impl Error {
             Error::RetryLater { .. }
                 | Error::RetryLaterWithSource { .. }
                 | Error::MailboxTimeout { .. }
+        ) || matches!(
+            self,
+            Error::AllocateRegions { source, .. } if source.is_retry_later()
+        ) || matches!(
+            self,
+            Error::DeallocateRegions { source, .. } if source.is_retry_later()
         )
     }
 }
@@ -1324,3 +1330,35 @@ pub(crate) fn match_for_io_error(err_status: &tonic::Status) -> Option<&std::io:
         err = err.source()?;
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use common_error::mock::MockError;
+    use common_error::status_code::StatusCode;
+    use snafu::ResultExt;
+
+    use super::DeallocateRegionsSnafu;
+
+    #[test]
+    fn test_deallocate_regions_is_retryable_when_source_is_retry_later() {
+        let source = common_meta::error::Error::retry_later(MockError::new(StatusCode::Internal));
+        let err = Err::<(), _>(source)
+            .context(DeallocateRegionsSnafu { table_id: 1024_u32 })
+            .unwrap_err();
+
+        assert!(err.is_retryable());
+    }
+
+    #[test]
+    fn test_deallocate_regions_is_not_retryable_when_source_is_not_retry_later() {
+        let source = common_meta::error::UnexpectedSnafu {
+            err_msg: "mock error",
+        }
+        .build();
+        let err = Err::<(), _>(source)
+            .context(DeallocateRegionsSnafu { table_id: 1024_u32 })
+            .unwrap_err();
+
+        assert!(!err.is_retryable());
+    }
+}
diff --git a/src/meta-srv/src/gc/candidate.rs b/src/meta-srv/src/gc/candidate.rs
index 05fc79ac52..2101318cdf 100644
--- a/src/meta-srv/src/gc/candidate.rs
+++ b/src/meta-srv/src/gc/candidate.rs
@@ -109,7 +109,7 @@ impl GcScheduler {
             }
 
             // Sort candidates by score in descending order and take top N
-            candidates.sort_by(|a, b| b.score.cmp(&a.score));
+            candidates.sort_by_key(|a| std::cmp::Reverse(a.score));
             let top_candidates: Vec<GcCandidate> = candidates
                 .into_iter()
                 .take(self.config.regions_per_table_threshold)
diff --git a/src/meta-srv/src/gc/handler.rs b/src/meta-srv/src/gc/handler.rs
index 105ddca58c..c62e927f89 100644
--- a/src/meta-srv/src/gc/handler.rs
+++ b/src/meta-srv/src/gc/handler.rs
@@ -346,7 +346,7 @@ impl GcScheduler {
                         // Add to need_retry_regions since it failed
                         combined_report
                             .need_retry_regions
-                            .extend(fast_list_regions.clone().into_iter());
+                            .extend(fast_list_regions.clone());
                     }
                 }
             }
diff --git a/src/meta-srv/src/handler.rs b/src/meta-srv/src/handler.rs
index 58dd1a7c98..4b05db4e4c 100644
--- a/src/meta-srv/src/handler.rs
+++ b/src/meta-srv/src/handler.rs
@@ -870,7 +870,7 @@ impl HeartbeatHandlerGroupBuilderCustomizer for DefaultHeartbeatHandlerGroupBuil
 #[cfg(test)]
 mod tests {
 
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::sync::Arc;
     use std::time::Duration;
 
@@ -972,7 +972,7 @@ mod tests {
             "RemapFlowPeerHandler",
         ];
         assert_eq!(names.len(), handlers.len());
-        for (handler, name) in handlers.iter().zip(names.into_iter()) {
+        for (handler, name) in handlers.iter().zip(names) {
             assert_eq!(handler.name, name);
         }
     }
@@ -1009,7 +1009,7 @@ mod tests {
             "RemapFlowPeerHandler",
         ];
         assert_eq!(names.len(), handlers.len());
-        for (handler, name) in handlers.iter().zip(names.into_iter()) {
+        for (handler, name) in handlers.iter().zip(names) {
             assert_eq!(handler.name, name);
         }
     }
@@ -1043,7 +1043,7 @@ mod tests {
             "RemapFlowPeerHandler",
         ];
         assert_eq!(names.len(), handlers.len());
-        for (handler, name) in handlers.iter().zip(names.into_iter()) {
+        for (handler, name) in handlers.iter().zip(names) {
             assert_eq!(handler.name, name);
         }
     }
@@ -1077,7 +1077,7 @@ mod tests {
             "RemapFlowPeerHandler",
         ];
         assert_eq!(names.len(), handlers.len());
-        for (handler, name) in handlers.iter().zip(names.into_iter()) {
+        for (handler, name) in handlers.iter().zip(names) {
             assert_eq!(handler.name, name);
         }
     }
@@ -1111,7 +1111,7 @@ mod tests {
             "RemapFlowPeerHandler",
         ];
         assert_eq!(names.len(), handlers.len());
-        for (handler, name) in handlers.iter().zip(names.into_iter()) {
+        for (handler, name) in handlers.iter().zip(names) {
             assert_eq!(handler.name, name);
         }
     }
@@ -1145,7 +1145,7 @@ mod tests {
         ];
 
         assert_eq!(names.len(), handlers.len());
-        for (handler, name) in handlers.iter().zip(names.into_iter()) {
+        for (handler, name) in handlers.iter().zip(names) {
             assert_eq!(handler.name, name);
         }
     }
@@ -1179,7 +1179,7 @@ mod tests {
         ];
 
         assert_eq!(names.len(), handlers.len());
-        for (handler, name) in handlers.iter().zip(names.into_iter()) {
+        for (handler, name) in handlers.iter().zip(names) {
             assert_eq!(handler.name, name);
         }
     }
@@ -1212,7 +1212,7 @@ mod tests {
             "RemapFlowPeerHandler",
         ];
         assert_eq!(names.len(), handlers.len());
-        for (handler, name) in handlers.iter().zip(names.into_iter()) {
+        for (handler, name) in handlers.iter().zip(names) {
             assert_eq!(handler.name, name);
         }
     }
diff --git a/src/meta-srv/src/handler/collect_cluster_info_handler.rs b/src/meta-srv/src/handler/collect_cluster_info_handler.rs
index c96229f9cf..c4d73477b6 100644
--- a/src/meta-srv/src/handler/collect_cluster_info_handler.rs
+++ b/src/meta-srv/src/handler/collect_cluster_info_handler.rs
@@ -12,12 +12,17 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+use std::collections::HashMap;
+
 use api::v1::meta::{HeartbeatRequest, NodeInfo as PbNodeInfo, Role};
 use common_meta::cluster::{
     DatanodeStatus, FlownodeStatus, FrontendStatus, NodeInfo, NodeInfoKey, NodeStatus,
 };
+use common_meta::datanode::EnvVars;
+use common_meta::heartbeat::utils::get_flownode_workloads;
 use common_meta::peer::Peer;
 use common_meta::rpc::store::PutRequest;
+use common_telemetry::warn;
 use snafu::ResultExt;
 use store_api::region_engine::RegionRole;
 
@@ -41,7 +46,7 @@ impl HeartbeatHandler for CollectFrontendClusterInfoHandler {
         ctx: &mut Context,
         _acc: &mut HeartbeatAccumulator,
     ) -> Result<HandleControl> {
-        let Some((key, peer, info)) = extract_base_info(req) else {
+        let Some((key, peer, info, env_vars)) = extract_base_info(req) else {
             return Ok(HandleControl::Continue);
         };
 
@@ -57,6 +62,7 @@ impl HeartbeatHandler for CollectFrontendClusterInfoHandler {
             cpu_usage_millicores: info.cpu_usage_millicores,
             memory_usage_bytes: info.memory_usage_bytes,
             hostname: info.hostname,
+            env_vars,
         };
 
         put_into_memory_store(ctx, key, value).await?;
@@ -79,14 +85,17 @@ impl HeartbeatHandler for CollectFlownodeClusterInfoHandler {
         ctx: &mut Context,
         _acc: &mut HeartbeatAccumulator,
     ) -> Result<HandleControl> {
-        let Some((key, peer, info)) = extract_base_info(req) else {
+        let Some((key, peer, info, env_vars)) = extract_base_info(req) else {
             return Ok(HandleControl::Continue);
         };
+        let flownode_workloads = get_flownode_workloads(req.node_workloads.as_ref());
 
         let value = NodeInfo {
             peer,
             last_activity_ts: common_time::util::current_time_millis(),
-            status: NodeStatus::Flownode(FlownodeStatus {}),
+            status: NodeStatus::Flownode(FlownodeStatus {
+                workloads: flownode_workloads,
+            }),
             version: info.version,
             git_commit: info.git_commit,
             start_time_ms: info.start_time_ms,
@@ -95,6 +104,7 @@ impl HeartbeatHandler for CollectFlownodeClusterInfoHandler {
             cpu_usage_millicores: info.cpu_usage_millicores,
             memory_usage_bytes: info.memory_usage_bytes,
             hostname: info.hostname,
+            env_vars,
         };
 
         put_into_memory_store(ctx, key, value).await?;
@@ -118,7 +128,7 @@ impl HeartbeatHandler for CollectDatanodeClusterInfoHandler {
         ctx: &mut Context,
         acc: &mut HeartbeatAccumulator,
     ) -> Result<HandleControl> {
-        let Some((key, peer, info)) = extract_base_info(req) else {
+        let Some((key, peer, info, env_vars)) = extract_base_info(req) else {
             return Ok(HandleControl::Continue);
         };
 
@@ -129,7 +139,7 @@ impl HeartbeatHandler for CollectDatanodeClusterInfoHandler {
         let leader_regions = stat
             .region_stats
             .iter()
-            .filter(|s| s.role == RegionRole::Leader)
+            .filter(|s| matches!(s.role, RegionRole::Leader | RegionRole::StagingLeader))
             .count();
         let follower_regions = stat.region_stats.len() - leader_regions;
 
@@ -151,6 +161,7 @@ impl HeartbeatHandler for CollectDatanodeClusterInfoHandler {
             cpu_usage_millicores: info.cpu_usage_millicores,
             memory_usage_bytes: info.memory_usage_bytes,
             hostname: info.hostname,
+            env_vars,
         };
 
         put_into_memory_store(ctx, key, value).await?;
@@ -159,7 +170,9 @@ impl HeartbeatHandler for CollectDatanodeClusterInfoHandler {
     }
 }
 
-fn extract_base_info(request: &HeartbeatRequest) -> Option<(NodeInfoKey, Peer, PbNodeInfo)> {
+fn extract_base_info(
+    request: &HeartbeatRequest,
+) -> Option<(NodeInfoKey, Peer, PbNodeInfo, HashMap<String, String>)> {
     let HeartbeatRequest { peer, info, .. } = request;
     let key = NodeInfoKey::new(request)?;
     let Some(peer) = &peer else {
@@ -169,7 +182,17 @@ fn extract_base_info(request: &HeartbeatRequest) -> Option<(NodeInfoKey, Peer, P
         return None;
     };
 
-    Some((key, peer.clone(), info.clone()))
+    let env_vars = EnvVars::from_extensions(&request.extensions)
+        .inspect_err(|e| {
+            warn!(e;
+                "Failed to deserialize __env_vars from heartbeat extensions, peer: {}", peer
+            );
+        })
+        .unwrap_or_default()
+        .map(|e| e.vars)
+        .unwrap_or_default();
+
+    Some((key, peer.clone(), info.clone(), env_vars))
 }
 
 async fn put_into_memory_store(ctx: &mut Context, key: NodeInfoKey, value: NodeInfo) -> Result<()> {
diff --git a/src/meta-srv/src/handler/collect_leader_region_handler.rs b/src/meta-srv/src/handler/collect_leader_region_handler.rs
index ddb4cd0ea3..95b03e3341 100644
--- a/src/meta-srv/src/handler/collect_leader_region_handler.rs
+++ b/src/meta-srv/src/handler/collect_leader_region_handler.rs
@@ -40,7 +40,7 @@ impl HeartbeatHandler for CollectLeaderRegionHandler {
 
         let mut key_values = Vec::with_capacity(current_stat.region_stats.len());
         for stat in current_stat.region_stats.iter() {
-            if stat.role != RegionRole::Leader {
+            if !matches!(stat.role, RegionRole::Leader | RegionRole::StagingLeader) {
                 continue;
             }
 
diff --git a/src/meta-srv/src/handler/keep_lease_handler.rs b/src/meta-srv/src/handler/keep_lease_handler.rs
index 45570a831d..426ff545c9 100644
--- a/src/meta-srv/src/handler/keep_lease_handler.rs
+++ b/src/meta-srv/src/handler/keep_lease_handler.rs
@@ -13,8 +13,8 @@
 // limitations under the License.
 
 use api::v1::meta::heartbeat_request::NodeWorkloads;
-use api::v1::meta::{FlownodeWorkloads, HeartbeatRequest, Peer, Role};
-use common_meta::heartbeat::utils::get_datanode_workloads;
+use api::v1::meta::{HeartbeatRequest, Peer, Role};
+use common_meta::heartbeat::utils::{get_datanode_workloads, get_flownode_workloads};
 use common_meta::rpc::store::PutRequest;
 use common_telemetry::{trace, warn};
 use common_time::util as time_util;
@@ -85,7 +85,12 @@ impl HeartbeatHandler for FlownodeKeepLeaseHandler {
         ctx: &mut Context,
         _acc: &mut HeartbeatAccumulator,
     ) -> Result<HandleControl> {
-        let HeartbeatRequest { header, peer, .. } = req;
+        let HeartbeatRequest {
+            header,
+            peer,
+            node_workloads,
+            ..
+        } = req;
         let Some(_header) = &header else {
             return Ok(HandleControl::Continue);
         };
@@ -97,7 +102,7 @@ impl HeartbeatHandler for FlownodeKeepLeaseHandler {
         let value = LeaseValue {
             timestamp_millis: time_util::current_time_millis(),
             node_addr: peer.addr.clone(),
-            workloads: NodeWorkloads::Flownode(FlownodeWorkloads { types: vec![] }),
+            workloads: NodeWorkloads::Flownode(get_flownode_workloads(node_workloads.as_ref())),
         };
 
         trace!("Receive a heartbeat from flownode: {key:?}, {value:?}");
diff --git a/src/meta-srv/src/handler/persist_stats_handler.rs b/src/meta-srv/src/handler/persist_stats_handler.rs
index 75281f982a..d863070225 100644
--- a/src/meta-srv/src/handler/persist_stats_handler.rs
+++ b/src/meta-srv/src/handler/persist_stats_handler.rs
@@ -121,7 +121,10 @@ fn to_persisted_if_leader(
     datanode_id: DatanodeId,
     timestamp_millis: i64,
 ) -> Option<(Row, PersistedRegionStat)> {
-    if matches!(region_stat.role, RegionRole::Leader) {
+    if matches!(
+        region_stat.role,
+        RegionRole::Leader | RegionRole::StagingLeader
+    ) {
         let persisted_region_stat = last_persisted_region_stats.get(&region_stat.id).map(|s| *s);
         Some((
             compute_persist_region_stat(
diff --git a/src/meta-srv/src/handler/region_lease_handler.rs b/src/meta-srv/src/handler/region_lease_handler.rs
index fef84ef0db..aa26a037e3 100644
--- a/src/meta-srv/src/handler/region_lease_handler.rs
+++ b/src/meta-srv/src/handler/region_lease_handler.rs
@@ -281,7 +281,7 @@ mod test {
 
         let opening_region_id = RegionId::new(table_id, region_number + 2);
         let _guard = opening_region_keeper
-            .register(follower_peer.id, opening_region_id)
+            .register_with_role(follower_peer.id, opening_region_id, RegionRole::Follower)
             .unwrap();
 
         let acc = &mut HeartbeatAccumulator::default();
@@ -398,6 +398,65 @@ mod test {
         assert_eq!(acc.inactive_region_ids, HashSet::from([no_exist_region_id]));
     }
 
+    #[tokio::test]
+    async fn test_handle_staging_leader() {
+        let datanode_id = 1;
+        let region_number = 1u32;
+        let table_id = 10;
+        let region_id = RegionId::new(table_id, region_number);
+        let peer = Peer::empty(datanode_id);
+        let table_info = new_test_table_info(table_id);
+
+        let region_routes = vec![RegionRoute {
+            region: Region::new_test(region_id),
+            leader_peer: Some(peer.clone()),
+            leader_state: Some(LeaderState::Staging),
+            ..Default::default()
+        }];
+
+        let keeper = new_test_keeper();
+        let table_metadata_manager = keeper.table_metadata_manager();
+
+        table_metadata_manager
+            .create_table_metadata(
+                table_info,
+                TableRouteValue::physical(region_routes),
+                HashMap::default(),
+            )
+            .await
+            .unwrap();
+
+        let builder = MetasrvBuilder::new();
+        let metasrv = builder.build().await.unwrap();
+        let ctx = &mut metasrv.new_ctx();
+
+        let req = HeartbeatRequest {
+            duration_since_epoch: 1234,
+            ..Default::default()
+        };
+
+        let acc = &mut HeartbeatAccumulator::default();
+        acc.stat = Some(Stat {
+            id: peer.id,
+            region_stats: vec![new_empty_region_stat(region_id, RegionRole::StagingLeader)],
+            ..Default::default()
+        });
+
+        let handler = RegionLeaseHandler::new(
+            default_distributed_time_constants().region_lease.as_secs(),
+            table_metadata_manager.clone(),
+            Default::default(),
+            None,
+        );
+
+        handler.handle(&req, ctx, acc).await.unwrap();
+
+        assert_region_lease(
+            acc,
+            vec![GrantedRegion::new(region_id, RegionRole::StagingLeader)],
+        );
+    }
+
     fn assert_region_lease(acc: &HeartbeatAccumulator, expected: Vec<GrantedRegion>) {
         let region_lease = acc.region_lease.as_ref().unwrap().clone();
         let granted: Vec<GrantedRegion> = region_lease
diff --git a/src/meta-srv/src/lib.rs b/src/meta-srv/src/lib.rs
index c67bc32b40..70ce449bba 100644
--- a/src/meta-srv/src/lib.rs
+++ b/src/meta-srv/src/lib.rs
@@ -12,16 +12,13 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
 #![feature(hash_set_entry)]
 #![feature(duration_constructors)]
-#![feature(string_from_utf8_lossy_owned)]
 
 pub mod bootstrap;
 pub mod cache_invalidator;
 pub mod cluster;
 pub mod discovery;
-pub mod election;
 pub mod error;
 pub mod events;
 mod failure_detector;
diff --git a/src/meta-srv/src/metasrv.rs b/src/meta-srv/src/metasrv.rs
index 165efd0555..6544578148 100644
--- a/src/meta-srv/src/metasrv.rs
+++ b/src/meta-srv/src/metasrv.rs
@@ -32,6 +32,8 @@ use common_meta::ddl_manager::DdlManagerRef;
 use common_meta::distributed_time_constants::{
     self, BASE_HEARTBEAT_INTERVAL, default_distributed_time_constants, frontend_heartbeat_interval,
 };
+use common_meta::election::LeaderChangeMessage;
+pub use common_meta::election::{ElectionRef, MetasrvNodeInfo};
 use common_meta::key::TableMetadataManagerRef;
 use common_meta::key::runtime_switch::RuntimeSwitchManagerRef;
 use common_meta::kv_backend::{KvBackendRef, ResettableKvBackend, ResettableKvBackendRef};
@@ -64,7 +66,6 @@ use tokio::sync::broadcast::error::RecvError;
 
 use crate::cluster::MetaPeerClientRef;
 use crate::discovery;
-use crate::election::{Election, LeaderChangeMessage};
 use crate::error::{
     self, InitMetadataSnafu, KvBackendSnafu, Result, StartProcedureManagerSnafu,
     StartTelemetryTaskSnafu, StopProcedureManagerSnafu,
@@ -82,6 +83,7 @@ use crate::selector::{RegionStatAwareSelector, Selector, SelectorType};
 use crate::service::mailbox::MailboxRef;
 use crate::service::store::cached_kv::LeaderCachedKvBackend;
 use crate::state::{StateRef, become_follower, become_leader};
+use crate::utils::database::DatabaseOperatorRef;
 
 pub const TABLE_ID_SEQ: &str = "table_id";
 pub const FLOW_ID_SEQ: &str = "flow_id";
@@ -459,76 +461,6 @@ impl Context {
     }
 }
 
-/// The value of the leader. It is used to store the leader's address.
-pub struct LeaderValue(pub String);
-
-impl<T: AsRef<[u8]>> From<T> for LeaderValue {
-    fn from(value: T) -> Self {
-        let string = String::from_utf8_lossy(value.as_ref());
-        Self(string.to_string())
-    }
-}
-
-#[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct MetasrvNodeInfo {
-    // The metasrv's address
-    pub addr: String,
-    // The node build version
-    pub version: String,
-    // The node build git commit hash
-    pub git_commit: String,
-    // The node start timestamp in milliseconds
-    pub start_time_ms: u64,
-    // The node total cpu millicores
-    #[serde(default)]
-    pub total_cpu_millicores: i64,
-    // The node total memory bytes
-    #[serde(default)]
-    pub total_memory_bytes: i64,
-    /// The node build cpu usage millicores
-    #[serde(default)]
-    pub cpu_usage_millicores: i64,
-    /// The node build memory usage bytes
-    #[serde(default)]
-    pub memory_usage_bytes: i64,
-    // The node hostname
-    #[serde(default)]
-    pub hostname: String,
-}
-
-// TODO(zyy17): Allow deprecated fields for backward compatibility. Remove this when the deprecated top-level fields are removed from the proto.
-#[allow(deprecated)]
-impl From<MetasrvNodeInfo> for api::v1::meta::MetasrvNodeInfo {
-    fn from(node_info: MetasrvNodeInfo) -> Self {
-        Self {
-            peer: Some(api::v1::meta::Peer {
-                addr: node_info.addr,
-                ..Default::default()
-            }),
-            // TODO(zyy17): The following top-level fields are deprecated. They are kept for backward compatibility and will be removed in a future version.
-            // New code should use the fields in `info.NodeInfo` instead.
-            version: node_info.version.clone(),
-            git_commit: node_info.git_commit.clone(),
-            start_time_ms: node_info.start_time_ms,
-            cpus: node_info.total_cpu_millicores as u32,
-            memory_bytes: node_info.total_memory_bytes as u64,
-            // The canonical location for node information.
-            info: Some(api::v1::meta::NodeInfo {
-                version: node_info.version,
-                git_commit: node_info.git_commit,
-                start_time_ms: node_info.start_time_ms,
-                total_cpu_millicores: node_info.total_cpu_millicores,
-                total_memory_bytes: node_info.total_memory_bytes,
-                cpu_usage_millicores: node_info.cpu_usage_millicores,
-                memory_usage_bytes: node_info.memory_usage_bytes,
-                cpus: node_info.total_cpu_millicores as u32,
-                memory_bytes: node_info.total_memory_bytes as u64,
-                hostname: node_info.hostname,
-            }),
-        }
-    }
-}
-
 #[derive(Clone, Copy)]
 pub enum SelectTarget {
     Datanode,
@@ -552,7 +484,6 @@ pub struct SelectorContext {
 pub type SelectorRef = Arc<dyn Selector<Context = SelectorContext, Output = Vec<Peer>>>;
 pub type RegionStatAwareSelectorRef =
     Arc<dyn RegionStatAwareSelector<Context = SelectorContext, Output = Vec<(RegionId, Peer)>>>;
-pub type ElectionRef = Arc<dyn Election<Leader = LeaderValue>>;
 
 pub struct MetaStateHandler {
     subscribe_manager: Option<SubscriptionManagerRef>,
@@ -640,6 +571,7 @@ pub struct Metasrv {
     reconciliation_manager: ReconciliationManagerRef,
     resource_stat: ResourceStatRef,
     gc_ticker: Option<GcTickerRef>,
+    database_operator: DatabaseOperatorRef,
 
     plugins: Plugins,
 }
@@ -939,6 +871,10 @@ impl Metasrv {
         &self.reconciliation_manager
     }
 
+    pub fn database_operator(&self) -> &DatabaseOperatorRef {
+        &self.database_operator
+    }
+
     pub fn plugins(&self) -> &Plugins {
         &self.plugins
     }
diff --git a/src/meta-srv/src/metasrv/builder.rs b/src/meta-srv/src/metasrv/builder.rs
index 58039d2de4..54a94fffbf 100644
--- a/src/meta-srv/src/metasrv/builder.rs
+++ b/src/meta-srv/src/metasrv/builder.rs
@@ -84,6 +84,7 @@ use crate::selector::round_robin::RoundRobinSelector;
 use crate::service::mailbox::MailboxRef;
 use crate::service::store::cached_kv::LeaderCachedKvBackend;
 use crate::state::State;
+use crate::utils::database::DatabaseOperator;
 use crate::utils::insert_forwarder::InsertForwarder;
 
 /// The time window for twcs compaction of the region stats table.
@@ -206,9 +207,10 @@ impl MetasrvBuilder {
 
         let meta_peer_client = meta_peer_client
             .unwrap_or_else(|| build_default_meta_peer_client(&election, &in_memory));
+        let database_operator = Arc::new(DatabaseOperator::new(meta_peer_client.clone()));
 
         let event_inserter = Box::new(InsertForwarder::new(
-            meta_peer_client.clone(),
+            database_operator.clone(),
             Some(InsertOptions {
                 ttl: options.event_recorder.ttl,
                 append_mode: true,
@@ -501,7 +503,7 @@ impl MetasrvBuilder {
 
         let persist_region_stats_handler = if !options.stats_persistence.ttl.is_zero() {
             let inserter = Box::new(InsertForwarder::new(
-                meta_peer_client.clone(),
+                database_operator.clone(),
                 Some(InsertOptions {
                     ttl: options.stats_persistence.ttl,
                     append_mode: true,
@@ -600,6 +602,7 @@ impl MetasrvBuilder {
             topic_stats_registry,
             resource_stat: Arc::new(resource_stat),
             gc_ticker,
+            database_operator,
         })
     }
 }
diff --git a/src/meta-srv/src/procedure/region_migration.rs b/src/meta-srv/src/procedure/region_migration.rs
index b3797860b3..6563da63bc 100644
--- a/src/meta-srv/src/procedure/region_migration.rs
+++ b/src/meta-srv/src/procedure/region_migration.rs
@@ -924,7 +924,7 @@ impl Procedure for RegionMigrationProcedure {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::sync::Arc;
 
     use common_meta::distributed_time_constants::default_distributed_time_constants;
diff --git a/src/meta-srv/src/procedure/region_migration/downgrade_leader_region.rs b/src/meta-srv/src/procedure/region_migration/downgrade_leader_region.rs
index 37cc0dfa96..b7e8315a77 100644
--- a/src/meta-srv/src/procedure/region_migration/downgrade_leader_region.rs
+++ b/src/meta-srv/src/procedure/region_migration/downgrade_leader_region.rs
@@ -372,7 +372,7 @@ impl DowngradeLeaderRegion {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
 
     use common_meta::key::table_route::TableRouteValue;
diff --git a/src/meta-srv/src/procedure/region_migration/flush_leader_region.rs b/src/meta-srv/src/procedure/region_migration/flush_leader_region.rs
index f3dc0ee661..860849ba6e 100644
--- a/src/meta-srv/src/procedure/region_migration/flush_leader_region.rs
+++ b/src/meta-srv/src/procedure/region_migration/flush_leader_region.rs
@@ -17,6 +17,7 @@ use std::any::Any;
 use common_procedure::{Context as ProcedureContext, Status};
 use serde::{Deserialize, Serialize};
 use snafu::OptionExt;
+use store_api::region_request::RegionFlushReason;
 use tokio::time::Instant;
 
 use crate::error::{self, Result};
@@ -86,6 +87,7 @@ impl PreFlushRegion {
             leader,
             operation_timeout,
             utils::ErrorStrategy::Ignore,
+            Some(RegionFlushReason::RegionMigration),
         )
         .await
     }
@@ -93,7 +95,7 @@ impl PreFlushRegion {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use store_api::storage::RegionId;
 
diff --git a/src/meta-srv/src/procedure/region_migration/manager.rs b/src/meta-srv/src/procedure/region_migration/manager.rs
index 70cba21b5f..d18458b08f 100644
--- a/src/meta-srv/src/procedure/region_migration/manager.rs
+++ b/src/meta-srv/src/procedure/region_migration/manager.rs
@@ -620,7 +620,7 @@ impl RegionMigrationManager {
 
 #[cfg(test)]
 mod test {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_meta::key::table_route::LogicalTableRouteValue;
     use common_meta::key::test_utils::new_test_table_info;
diff --git a/src/meta-srv/src/procedure/region_migration/migration_start.rs b/src/meta-srv/src/procedure/region_migration/migration_start.rs
index 17b577501e..b9d4372c47 100644
--- a/src/meta-srv/src/procedure/region_migration/migration_start.rs
+++ b/src/meta-srv/src/procedure/region_migration/migration_start.rs
@@ -187,7 +187,7 @@ impl RegionMigrationStart {
 #[cfg(test)]
 mod tests {
 
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_meta::key::test_utils::new_test_table_info;
     use common_meta::peer::Peer;
diff --git a/src/meta-srv/src/procedure/region_migration/open_candidate_region.rs b/src/meta-srv/src/procedure/region_migration/open_candidate_region.rs
index 8e2e015669..0c0e5de5d7 100644
--- a/src/meta-srv/src/procedure/region_migration/open_candidate_region.rs
+++ b/src/meta-srv/src/procedure/region_migration/open_candidate_region.rs
@@ -25,6 +25,7 @@ use common_telemetry::info;
 use common_telemetry::tracing_context::TracingContext;
 use serde::{Deserialize, Serialize};
 use snafu::{OptionExt, ResultExt};
+use store_api::region_engine::RegionRole;
 use tokio::time::Instant;
 
 use crate::error::{self, Result};
@@ -129,7 +130,7 @@ impl OpenCandidateRegion {
                 // Registers the opening region.
                 let guard = ctx
                     .opening_region_keeper
-                    .register(candidate.id, *region_id)
+                    .register_with_role(candidate.id, *region_id, RegionRole::Follower)
                     .context(error::RegionOperatingRaceSnafu {
                         peer_id: candidate.id,
                         region_id: *region_id,
@@ -208,7 +209,7 @@ impl OpenCandidateRegion {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
 
     use common_catalog::consts::MITO2_ENGINE;
@@ -296,7 +297,7 @@ mod tests {
         let mut ctx = env.context_factory().new_context(persistent_context);
         let opening_region_keeper = env.opening_region_keeper();
         let _guard = opening_region_keeper
-            .register(to_peer_id, region_id)
+            .register_with_role(to_peer_id, region_id, RegionRole::Follower)
             .unwrap();
 
         let open_instruction = new_mock_open_instruction(to_peer_id, region_id);
diff --git a/src/meta-srv/src/procedure/region_migration/test_util.rs b/src/meta-srv/src/procedure/region_migration/test_util.rs
index 4e5624401e..269357bc0a 100644
--- a/src/meta-srv/src/procedure/region_migration/test_util.rs
+++ b/src/meta-srv/src/procedure/region_migration/test_util.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 use std::collections::HashMap;
 use std::sync::Arc;
 use std::sync::atomic::{AtomicUsize, Ordering};
diff --git a/src/meta-srv/src/procedure/region_migration/update_metadata/downgrade_leader_region.rs b/src/meta-srv/src/procedure/region_migration/update_metadata/downgrade_leader_region.rs
index 9e8545bb43..bab79a96bf 100644
--- a/src/meta-srv/src/procedure/region_migration/update_metadata/downgrade_leader_region.rs
+++ b/src/meta-srv/src/procedure/region_migration/update_metadata/downgrade_leader_region.rs
@@ -88,7 +88,7 @@ impl UpdateMetadata {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
     use std::sync::Arc;
 
diff --git a/src/meta-srv/src/procedure/region_migration/update_metadata/rollback_downgraded_region.rs b/src/meta-srv/src/procedure/region_migration/update_metadata/rollback_downgraded_region.rs
index b3ee848fbe..2f0ed0fbe9 100644
--- a/src/meta-srv/src/procedure/region_migration/update_metadata/rollback_downgraded_region.rs
+++ b/src/meta-srv/src/procedure/region_migration/update_metadata/rollback_downgraded_region.rs
@@ -72,7 +72,7 @@ impl UpdateMetadata {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
     use std::sync::Arc;
 
diff --git a/src/meta-srv/src/procedure/region_migration/update_metadata/upgrade_candidate_region.rs b/src/meta-srv/src/procedure/region_migration/update_metadata/upgrade_candidate_region.rs
index db70e3e166..93df5e6e2e 100644
--- a/src/meta-srv/src/procedure/region_migration/update_metadata/upgrade_candidate_region.rs
+++ b/src/meta-srv/src/procedure/region_migration/update_metadata/upgrade_candidate_region.rs
@@ -224,13 +224,14 @@ impl UpdateMetadata {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_meta::key::test_utils::new_test_table_info;
     use common_meta::peer::Peer;
     use common_meta::region_keeper::MemoryRegionKeeper;
     use common_meta::rpc::router::{LeaderState, Region, RegionRoute};
     use common_time::util::current_time_millis;
+    use store_api::region_engine::RegionRole;
     use store_api::storage::RegionId;
 
     use crate::error::Error;
@@ -467,7 +468,7 @@ mod tests {
         }];
 
         let guard = opening_keeper
-            .register(2, RegionId::new(table_id, 1))
+            .register_with_role(2, RegionId::new(table_id, 1), RegionRole::Follower)
             .unwrap();
         ctx.volatile_ctx.opening_region_guards.push(guard);
 
diff --git a/src/meta-srv/src/procedure/region_migration/upgrade_candidate_region.rs b/src/meta-srv/src/procedure/region_migration/upgrade_candidate_region.rs
index 39ff1ed741..4c60215cf7 100644
--- a/src/meta-srv/src/procedure/region_migration/upgrade_candidate_region.rs
+++ b/src/meta-srv/src/procedure/region_migration/upgrade_candidate_region.rs
@@ -353,7 +353,7 @@ impl UpgradeCandidateRegion {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
 
     use common_meta::key::table_route::TableRouteValue;
diff --git a/src/meta-srv/src/procedure/region_migration/utils.rs b/src/meta-srv/src/procedure/region_migration/utils.rs
index df2e8014e2..91af7ccf17 100644
--- a/src/meta-srv/src/procedure/region_migration/utils.rs
+++ b/src/meta-srv/src/procedure/region_migration/utils.rs
@@ -242,7 +242,7 @@ pub async fn analyze_region_migration_task(
 #[cfg(test)]
 mod tests {
 
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::sync::Arc;
     use std::time::Duration;
 
diff --git a/src/meta-srv/src/procedure/repartition.rs b/src/meta-srv/src/procedure/repartition.rs
index 37c7745ae5..f314a40080 100644
--- a/src/meta-srv/src/procedure/repartition.rs
+++ b/src/meta-srv/src/procedure/repartition.rs
@@ -23,7 +23,7 @@ pub mod repartition_start;
 pub mod utils;
 
 use std::any::Any;
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 use std::fmt::{Debug, Display};
 use std::time::{Duration, Instant};
 
@@ -40,15 +40,15 @@ use common_meta::key::table_route::TableRouteValue;
 use common_meta::key::{DeserializedValueWithBytes, TableMetadataManagerRef};
 use common_meta::lock_key::{CatalogLock, SchemaLock, TableLock, TableNameLock};
 use common_meta::node_manager::NodeManagerRef;
-use common_meta::region_keeper::MemoryRegionKeeperRef;
+use common_meta::region_keeper::{MemoryRegionKeeperRef, OperatingRegionGuard};
 use common_meta::region_registry::LeaderRegionRegistryRef;
-use common_meta::rpc::router::RegionRoute;
+use common_meta::rpc::router::{RegionRoute, operating_leader_region_roles};
 use common_procedure::error::{FromJsonSnafu, ToJsonSnafu};
 use common_procedure::{
     BoxedProcedure, Context as ProcedureContext, Error as ProcedureError, LockKey, Procedure,
     ProcedureManagerRef, Result as ProcedureResult, Status, StringKey, UserMetadata,
 };
-use common_telemetry::{error, info};
+use common_telemetry::{error, info, warn};
 use partition::expr::PartitionExpr;
 use serde::{Deserialize, Serialize};
 use snafu::{OptionExt, ResultExt};
@@ -56,12 +56,16 @@ use store_api::storage::{RegionNumber, TableId};
 use table::table_name::TableName;
 
 use crate::error::{self, Result};
+use crate::procedure::repartition::collect::ProcedureMeta;
+use crate::procedure::repartition::deallocate_region::DeallocateRegion;
 use crate::procedure::repartition::group::{
-    Context as RepartitionGroupContext, RepartitionGroupProcedure,
+    Context as RepartitionGroupContext, RepartitionGroupProcedure, region_routes,
 };
 use crate::procedure::repartition::plan::RepartitionPlanEntry;
 use crate::procedure::repartition::repartition_start::RepartitionStart;
-use crate::procedure::repartition::utils::get_datanode_table_value;
+use crate::procedure::repartition::utils::{
+    get_datanode_table_value, rollback_group_metadata_routes,
+};
 use crate::service::mailbox::MailboxRef;
 
 #[cfg(test)]
@@ -74,6 +78,18 @@ pub struct PersistentContext {
     pub table_name: String,
     pub table_id: TableId,
     pub plans: Vec<RepartitionPlanEntry>,
+    #[serde(default)]
+    /// Records failed sub-procedures for parent rollback selection.
+    ///
+    /// The parent repartition procedure uses these entries to decide which plans
+    /// require group-metadata restoration and allocated-region cleanup.
+    pub failed_procedures: Vec<ProcedureMeta>,
+    #[serde(default)]
+    /// Records unknown sub-procedures for parent rollback selection.
+    ///
+    /// Unknown procedures are treated the same as failed ones when selecting the
+    /// plan subset that must be rolled back by the parent procedure.
+    pub unknown_procedures: Vec<ProcedureMeta>,
     /// The timeout for repartition operations.
     #[serde(with = "humantime_serde", default = "default_timeout")]
     pub timeout: Duration,
@@ -102,6 +118,8 @@ impl PersistentContext {
             table_name,
             table_id,
             plans: vec![],
+            failed_procedures: vec![],
+            unknown_procedures: vec![],
             timeout: timeout.unwrap_or_else(default_timeout),
         }
     }
@@ -393,6 +411,23 @@ impl Context {
             .await;
         Ok(())
     }
+
+    pub fn register_operating_regions(
+        memory_region_keeper: &MemoryRegionKeeperRef,
+        region_routes: &[RegionRoute],
+    ) -> Result<Vec<OperatingRegionGuard>> {
+        let mut operating_guards = Vec::with_capacity(region_routes.len());
+        for (region_id, datanode_id, role) in operating_leader_region_roles(region_routes) {
+            let guard = memory_region_keeper
+                .register_with_role(datanode_id, region_id, role)
+                .context(error::RegionOperatingRaceSnafu {
+                    peer_id: datanode_id,
+                    region_id,
+                })?;
+            operating_guards.push(guard);
+        }
+        Ok(operating_guards)
+    }
 }
 
 #[async_trait::async_trait]
@@ -456,6 +491,164 @@ impl RepartitionProcedure {
 
         Ok(Self { state, context })
     }
+
+    /// Returns whether parent rollback should remove this repartition's allocated regions.
+    ///
+    /// This uses an "after AllocateRegion" semantic: once execution reaches
+    /// `AllocateRegion` or any later state, rollback must try to remove this round's
+    /// `allocated_region_ids` from table-route metadata when they exist.
+    ///
+    /// State flow:
+    /// `RepartitionStart -> AllocateRegion -> Dispatch -> Collect -> DeallocateRegion -> RepartitionEnd`
+    ///                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+    ///                     rollback allocated regions in metadata
+    ///
+    /// Notes:
+    /// - `RepartitionStart`: no-op, because allocation has not happened yet.
+    /// - `AllocateRegion` / `Dispatch` / `Collect`  rollback-active.
+    /// - `DeallocateRegion`: is not rollback-active.
+    /// - `RepartitionEnd`: no-op.
+    fn should_rollback_allocated_regions(&self) -> bool {
+        self.state.as_any().is::<allocate_region::AllocateRegion>()
+            || self.state.as_any().is::<dispatch::Dispatch>()
+            || self.state.as_any().is::<collect::Collect>()
+    }
+
+    fn rollback_plan_indices(&self) -> HashSet<usize> {
+        self.context
+            .persistent_ctx
+            .failed_procedures
+            .iter()
+            .chain(self.context.persistent_ctx.unknown_procedures.iter())
+            .map(|procedure_meta| procedure_meta.plan_index)
+            .collect()
+    }
+
+    /// Returns allocated region ids that parent rollback should remove.
+    ///
+    /// Rollback uses an "after AllocateRegion" semantic:
+    /// - in `AllocateRegion` and `Dispatch`, all allocated regions belong to the
+    ///   current repartition attempt and must be cleaned up.
+    /// - in `Collect`, only the plans referenced by failed or unknown
+    ///   sub-procedures should be rolled back.
+    fn rollback_allocated_region_ids(&self) -> HashSet<store_api::storage::RegionId> {
+        if self.state.as_any().is::<allocate_region::AllocateRegion>()
+            || self.state.as_any().is::<dispatch::Dispatch>()
+        {
+            return self
+                .context
+                .persistent_ctx
+                .plans
+                .iter()
+                .flat_map(|plan| plan.allocated_region_ids.iter().copied())
+                .collect();
+        }
+
+        self.rollback_plan_indices()
+            .into_iter()
+            .flat_map(|plan_index| {
+                self.context.persistent_ctx.plans[plan_index]
+                    .allocated_region_ids
+                    .iter()
+                    .copied()
+            })
+            .collect()
+    }
+
+    /// Restores group-level staging metadata for failed/unknown plans.
+    ///
+    /// The helper mutates `region_routes` in memory.
+    async fn rollback_group_metadata_for_selected_plans(
+        &mut self,
+        region_routes: &mut [RegionRoute],
+    ) -> Result<()> {
+        let rollback_plan_indices = self.rollback_plan_indices();
+        if rollback_plan_indices.is_empty() {
+            return Ok(());
+        }
+
+        let mut region_routes_map = region_routes
+            .iter_mut()
+            .map(|route| (route.region.id, route))
+            .collect::<HashMap<_, _>>();
+        for plan_index in rollback_plan_indices {
+            let plan = &self.context.persistent_ctx.plans[plan_index];
+            rollback_group_metadata_routes(
+                plan.group_id,
+                &plan.source_regions,
+                &plan.original_target_routes,
+                &plan.allocated_region_ids,
+                &plan.pending_deallocate_region_ids,
+                &mut region_routes_map,
+            )?;
+        }
+
+        Ok(())
+    }
+
+    async fn rollback_inner(&mut self, procedure_ctx: &ProcedureContext) -> Result<()> {
+        if !self.should_rollback_allocated_regions() {
+            return Ok(());
+        }
+
+        let table_id = self.context.persistent_ctx.table_id;
+        let allocated_region_ids = self.rollback_allocated_region_ids();
+
+        let table_lock = TableLock::Write(table_id).into();
+        let _guard = procedure_ctx.provider.acquire_lock(&table_lock).await;
+        let table_route_value = self.context.get_table_route_value().await?;
+        let original_region_routes = region_routes(table_id, table_route_value.get_inner_ref())?;
+        let mut current_region_routes = original_region_routes.clone();
+        self.rollback_group_metadata_for_selected_plans(&mut current_region_routes)
+            .await?;
+        let allocated_region_routes = DeallocateRegion::filter_deallocatable_region_routes(
+            table_id,
+            &current_region_routes,
+            &allocated_region_ids,
+        );
+        if !allocated_region_routes.is_empty() {
+            let table = TableName {
+                catalog_name: self.context.persistent_ctx.catalog_name.clone(),
+                schema_name: self.context.persistent_ctx.schema_name.clone(),
+                table_name: self.context.persistent_ctx.table_name.clone(),
+            };
+            // Memory guards are not required here,
+            // because the table metadata still contains routes for the deallocating regions.
+            if let Err(err) = DeallocateRegion::deallocate_regions(
+                &self.context.node_manager,
+                &self.context.leader_region_registry,
+                table,
+                table_id,
+                &allocated_region_routes,
+            )
+            .await
+            {
+                warn!(err; "Failed to drop allocated regions during repartition rollback, table_id: {}, regions: {:?}", table_id, allocated_region_ids);
+            }
+        }
+
+        let new_region_routes =
+            DeallocateRegion::generate_region_routes(&current_region_routes, &allocated_region_ids);
+
+        if new_region_routes != *original_region_routes {
+            self.context
+                .update_table_route(&table_route_value, new_region_routes, HashMap::new())
+                .await
+                .map_err(BoxedError::new)
+                .with_context(|_| error::RetryLaterWithSourceSnafu {
+                    reason: format!(
+                        "Failed to rollback allocated region routes for repartition table: {}",
+                        table_id
+                    ),
+                })?;
+        }
+
+        if let Err(err) = self.context.invalidate_table_cache().await {
+            warn!(err; "Failed to invalidate table cache during repartition rollback, table_id: {}", table_id);
+        }
+
+        Ok(())
+    }
 }
 
 #[async_trait::async_trait]
@@ -497,9 +690,14 @@ impl Procedure for RepartitionProcedure {
         }
     }
 
+    async fn rollback(&mut self, ctx: &ProcedureContext) -> ProcedureResult<()> {
+        self.rollback_inner(ctx)
+            .await
+            .map_err(ProcedureError::external)
+    }
+
     fn rollback_supported(&self) -> bool {
-        // TODO(weny): support rollback.
-        false
+        true
     }
 
     fn dump(&self) -> ProcedureResult<String> {
@@ -624,3 +822,1121 @@ impl RepartitionProcedureFactory for DefaultRepartitionProcedureFactory {
         Ok(())
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+    use std::sync::Arc;
+    use std::sync::atomic::{AtomicBool, Ordering};
+
+    use common_error::ext::BoxedError;
+    use common_error::mock::MockError;
+    use common_error::status_code::StatusCode;
+    use common_meta::ddl::test_util::datanode_handler::{
+        DatanodeWatcher, NaiveDatanodeHandler, UnexpectedErrorDatanodeHandler,
+    };
+    use common_meta::error;
+    use common_meta::peer::Peer;
+    use common_meta::region_keeper::MemoryRegionKeeper;
+    use common_meta::rpc::router::{LeaderState, Region, RegionRoute};
+    use common_meta::test_util::MockDatanodeManager;
+    use common_procedure::{Error as ProcedureError, Procedure, ProcedureId, ProcedureState};
+    use store_api::region_engine::RegionRole;
+    use store_api::storage::RegionId;
+    use table::table_name::TableName;
+    use tokio::sync::mpsc;
+    use uuid::Uuid;
+
+    use super::*;
+    use crate::procedure::repartition::allocate_region::AllocateRegion;
+    use crate::procedure::repartition::collect::Collect;
+    use crate::procedure::repartition::deallocate_region::DeallocateRegion;
+    use crate::procedure::repartition::dispatch::Dispatch;
+    use crate::procedure::repartition::group::update_metadata::UpdateMetadata;
+    use crate::procedure::repartition::plan::RegionDescriptor;
+    use crate::procedure::repartition::repartition_end::RepartitionEnd;
+    use crate::procedure::repartition::test_util::{
+        TestingEnv, assert_parent_state, current_parent_region_routes, extract_subprocedure_ids,
+        new_parent_context, procedure_context_with_receivers, procedure_state_receiver, range_expr,
+        test_region_route, test_region_wal_options,
+    };
+
+    fn test_plan(table_id: TableId) -> RepartitionPlanEntry {
+        RepartitionPlanEntry {
+            group_id: uuid::Uuid::new_v4(),
+            source_regions: vec![RegionDescriptor {
+                region_id: RegionId::new(table_id, 1),
+                partition_expr: range_expr("x", 0, 100),
+            }],
+            target_regions: vec![
+                RegionDescriptor {
+                    region_id: RegionId::new(table_id, 1),
+                    partition_expr: range_expr("x", 0, 50),
+                },
+                RegionDescriptor {
+                    region_id: RegionId::new(table_id, 3),
+                    partition_expr: range_expr("x", 50, 100),
+                },
+            ],
+            allocated_region_ids: vec![RegionId::new(table_id, 3)],
+            pending_deallocate_region_ids: vec![],
+            transition_map: vec![vec![0, 1]],
+            original_target_routes: vec![],
+        }
+    }
+
+    fn with_rollback_metadata(
+        mut plan: RepartitionPlanEntry,
+        original_target_routes: Vec<RegionRoute>,
+    ) -> RepartitionPlanEntry {
+        plan.original_target_routes = original_target_routes;
+        plan
+    }
+
+    fn apply_group_staging(
+        plan: &RepartitionPlanEntry,
+        current_region_routes: &[RegionRoute],
+    ) -> Vec<RegionRoute> {
+        UpdateMetadata::apply_staging_region_routes(
+            plan.group_id,
+            &plan.source_regions,
+            &plan.target_regions,
+            &plan.pending_deallocate_region_ids,
+            current_region_routes,
+        )
+        .unwrap()
+    }
+
+    fn exit_group_staging(
+        plan: &RepartitionPlanEntry,
+        current_region_routes: &[RegionRoute],
+    ) -> Vec<RegionRoute> {
+        UpdateMetadata::exit_staging_region_routes(
+            plan.group_id,
+            &plan.source_regions,
+            &plan.target_regions,
+            current_region_routes,
+        )
+        .unwrap()
+    }
+
+    fn region_route_by_id(region_routes: &[RegionRoute], region_id: RegionId) -> &RegionRoute {
+        region_routes
+            .iter()
+            .find(|route| route.region.id == region_id)
+            .unwrap()
+    }
+
+    fn test_procedure(state: Box<dyn State>, context: Context) -> RepartitionProcedure {
+        RepartitionProcedure { state, context }
+    }
+
+    fn test_context(env: &TestingEnv, table_id: TableId) -> Context {
+        let node_manager = Arc::new(MockDatanodeManager::new(UnexpectedErrorDatanodeHandler));
+        let ddl_ctx = env.ddl_context(node_manager);
+        let persistent_ctx = PersistentContext::new(
+            TableName::new("test_catalog", "test_schema", "test_table"),
+            table_id,
+            None,
+        );
+
+        Context::new(
+            &ddl_ctx,
+            env.mailbox_ctx.mailbox().clone(),
+            env.server_addr.clone(),
+            persistent_ctx,
+        )
+    }
+
+    #[test]
+    fn test_filter_allocated_region_routes() {
+        let table_id = 1024;
+        let region_routes = vec![
+            test_region_route(RegionId::new(table_id, 1), "a"),
+            test_region_route(RegionId::new(table_id, 2), "b"),
+        ];
+        let allocated_region_ids = HashSet::from([RegionId::new(table_id, 2)]);
+
+        let new_region_routes =
+            DeallocateRegion::generate_region_routes(&region_routes, &allocated_region_ids);
+
+        assert_eq!(new_region_routes.len(), 1);
+        assert_eq!(new_region_routes[0].region.id, RegionId::new(table_id, 1));
+    }
+
+    #[test]
+    fn test_should_rollback_allocated_regions() {
+        let env = TestingEnv::new();
+        let table_id = 1024;
+
+        let procedure = test_procedure(
+            Box::new(RepartitionStart::new(vec![], vec![])),
+            test_context(&env, table_id),
+        );
+        assert!(!procedure.should_rollback_allocated_regions());
+
+        let procedure = test_procedure(
+            Box::new(AllocateRegion::new(vec![])),
+            test_context(&env, table_id),
+        );
+        assert!(procedure.should_rollback_allocated_regions());
+
+        let procedure = test_procedure(Box::new(Dispatch), test_context(&env, table_id));
+        assert!(procedure.should_rollback_allocated_regions());
+
+        let procedure =
+            test_procedure(Box::new(Collect::new(vec![])), test_context(&env, table_id));
+        assert!(procedure.should_rollback_allocated_regions());
+
+        let procedure = test_procedure(Box::new(DeallocateRegion), test_context(&env, table_id));
+        assert!(!procedure.should_rollback_allocated_regions());
+
+        let procedure = test_procedure(Box::new(RepartitionEnd), test_context(&env, table_id));
+        assert!(!procedure.should_rollback_allocated_regions());
+    }
+
+    #[test]
+    fn test_register_operating_regions_preserves_route_roles() {
+        let keeper = Arc::new(MemoryRegionKeeper::new());
+        let region_routes = vec![
+            RegionRoute {
+                region: Region::new_test(RegionId::new(1024, 1)),
+                leader_peer: Some(Peer::empty(1)),
+                follower_peers: vec![],
+                leader_state: None,
+                leader_down_since: None,
+                write_route_policy: None,
+            },
+            RegionRoute {
+                region: Region::new_test(RegionId::new(1024, 2)),
+                leader_peer: Some(Peer::empty(2)),
+                follower_peers: vec![],
+                leader_state: Some(LeaderState::Staging),
+                leader_down_since: None,
+                write_route_policy: None,
+            },
+            RegionRoute {
+                region: Region::new_test(RegionId::new(1024, 3)),
+                leader_peer: Some(Peer::empty(3)),
+                follower_peers: vec![],
+                leader_state: Some(LeaderState::Downgrading),
+                leader_down_since: None,
+                write_route_policy: None,
+            },
+        ];
+
+        let _guards = Context::register_operating_regions(&keeper, &region_routes).unwrap();
+
+        let leader_roles =
+            keeper.extract_operating_region_roles(1, &HashSet::from([RegionId::new(1024, 1)]));
+        let staging_roles =
+            keeper.extract_operating_region_roles(2, &HashSet::from([RegionId::new(1024, 2)]));
+        let downgrading_roles =
+            keeper.extract_operating_region_roles(3, &HashSet::from([RegionId::new(1024, 3)]));
+
+        assert_eq!(
+            leader_roles.get(&RegionId::new(1024, 1)),
+            Some(&RegionRole::Leader)
+        );
+        assert_eq!(
+            staging_roles.get(&RegionId::new(1024, 2)),
+            Some(&RegionRole::StagingLeader)
+        );
+        assert_eq!(
+            downgrading_roles.get(&RegionId::new(1024, 3)),
+            Some(&RegionRole::DowngradingLeader)
+        );
+    }
+
+    #[tokio::test]
+    async fn test_repartition_rollback_removes_allocated_routes_from_dispatch() {
+        let env = TestingEnv::new();
+        let table_id = 1024;
+        let node_manager = Arc::new(MockDatanodeManager::new(UnexpectedErrorDatanodeHandler));
+        let ddl_ctx = env.ddl_context(node_manager);
+        let original_region_routes = vec![
+            test_region_route(
+                RegionId::new(table_id, 1),
+                &range_expr("x", 0, 100).as_json_str().unwrap(),
+            ),
+            test_region_route(
+                RegionId::new(table_id, 2),
+                &range_expr("x", 50, 100).as_json_str().unwrap(),
+            ),
+            test_region_route(RegionId::new(table_id, 3), ""),
+        ];
+        env.create_physical_table_metadata_with_wal_options(
+            table_id,
+            original_region_routes,
+            test_region_wal_options(&[1, 2]),
+        )
+        .await;
+
+        let mut persistent_ctx = PersistentContext::new(
+            TableName::new("test_catalog", "test_schema", "test_table"),
+            table_id,
+            None,
+        );
+        persistent_ctx.plans = vec![with_rollback_metadata(
+            test_plan(table_id),
+            vec![
+                test_region_route(
+                    RegionId::new(table_id, 1),
+                    &range_expr("x", 0, 100).as_json_str().unwrap(),
+                ),
+                test_region_route(RegionId::new(table_id, 3), ""),
+            ],
+        )];
+        persistent_ctx.failed_procedures = vec![ProcedureMeta {
+            plan_index: 0,
+            group_id: Uuid::new_v4(),
+            procedure_id: ProcedureId::random(),
+        }];
+        let context = Context::new(
+            &ddl_ctx,
+            env.mailbox_ctx.mailbox().clone(),
+            env.server_addr.clone(),
+            persistent_ctx,
+        );
+        let mut procedure = RepartitionProcedure {
+            state: Box::new(Dispatch),
+            context,
+        };
+
+        procedure
+            .rollback(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+
+        let region_routes = current_parent_region_routes(&procedure.context).await;
+        assert_eq!(region_routes.len(), 2);
+        assert_eq!(region_routes[0].region.id, RegionId::new(table_id, 1));
+        assert_eq!(region_routes[1].region.id, RegionId::new(table_id, 2));
+    }
+
+    #[tokio::test]
+    async fn test_repartition_rollback_removes_allocated_routes_from_allocate() {
+        let env = TestingEnv::new();
+        let table_id = 1024;
+        let node_manager = Arc::new(MockDatanodeManager::new(UnexpectedErrorDatanodeHandler));
+        let ddl_ctx = env.ddl_context(node_manager);
+        let original_region_routes = vec![
+            test_region_route(
+                RegionId::new(table_id, 1),
+                &range_expr("x", 0, 100).as_json_str().unwrap(),
+            ),
+            test_region_route(
+                RegionId::new(table_id, 2),
+                &range_expr("x", 50, 100).as_json_str().unwrap(),
+            ),
+            test_region_route(RegionId::new(table_id, 3), ""),
+        ];
+        env.create_physical_table_metadata_with_wal_options(
+            table_id,
+            original_region_routes,
+            test_region_wal_options(&[1, 2]),
+        )
+        .await;
+
+        let mut persistent_ctx = PersistentContext::new(
+            TableName::new("test_catalog", "test_schema", "test_table"),
+            table_id,
+            None,
+        );
+        persistent_ctx.plans = vec![test_plan(table_id)];
+        let context = Context::new(
+            &ddl_ctx,
+            env.mailbox_ctx.mailbox().clone(),
+            env.server_addr.clone(),
+            persistent_ctx,
+        );
+        let mut procedure = RepartitionProcedure {
+            state: Box::new(AllocateRegion::new(vec![])),
+            context,
+        };
+
+        procedure
+            .rollback(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+
+        let region_routes = current_parent_region_routes(&procedure.context).await;
+        assert_eq!(region_routes.len(), 2);
+        assert_eq!(region_routes[0].region.id, RegionId::new(table_id, 1));
+        assert_eq!(region_routes[1].region.id, RegionId::new(table_id, 2));
+    }
+
+    #[tokio::test]
+    async fn test_repartition_rollback_from_collect_only_removes_failed_allocated_routes() {
+        let env = TestingEnv::new();
+        let table_id = 1024;
+        let node_manager = Arc::new(MockDatanodeManager::new(UnexpectedErrorDatanodeHandler));
+        let ddl_ctx = env.ddl_context(node_manager);
+        let original_region_routes = vec![
+            test_region_route(
+                RegionId::new(table_id, 1),
+                &range_expr("x", 0, 100).as_json_str().unwrap(),
+            ),
+            test_region_route(
+                RegionId::new(table_id, 2),
+                &range_expr("x", 100, 200).as_json_str().unwrap(),
+            ),
+            test_region_route(RegionId::new(table_id, 3), ""),
+            test_region_route(RegionId::new(table_id, 4), ""),
+        ];
+        env.create_physical_table_metadata_with_wal_options(
+            table_id,
+            original_region_routes,
+            test_region_wal_options(&[1, 2, 3, 4]),
+        )
+        .await;
+
+        let mut persistent_ctx = PersistentContext::new(
+            TableName::new("test_catalog", "test_schema", "test_table"),
+            table_id,
+            None,
+        );
+        let failed_plan = test_plan(table_id);
+        let failed_plan = with_rollback_metadata(
+            failed_plan,
+            vec![
+                test_region_route(
+                    RegionId::new(table_id, 1),
+                    &range_expr("x", 0, 100).as_json_str().unwrap(),
+                ),
+                test_region_route(RegionId::new(table_id, 3), ""),
+            ],
+        );
+        let succeeded_plan = RepartitionPlanEntry {
+            group_id: Uuid::new_v4(),
+            source_regions: vec![RegionDescriptor {
+                region_id: RegionId::new(table_id, 2),
+                partition_expr: range_expr("x", 100, 200),
+            }],
+            target_regions: vec![
+                RegionDescriptor {
+                    region_id: RegionId::new(table_id, 2),
+                    partition_expr: range_expr("x", 100, 150),
+                },
+                RegionDescriptor {
+                    region_id: RegionId::new(table_id, 4),
+                    partition_expr: range_expr("x", 150, 200),
+                },
+            ],
+            allocated_region_ids: vec![RegionId::new(table_id, 4)],
+            pending_deallocate_region_ids: vec![],
+            transition_map: vec![vec![0]],
+            original_target_routes: vec![
+                test_region_route(
+                    RegionId::new(table_id, 2),
+                    &range_expr("x", 100, 200).as_json_str().unwrap(),
+                ),
+                test_region_route(RegionId::new(table_id, 4), ""),
+            ],
+        };
+        persistent_ctx.plans = vec![failed_plan, succeeded_plan];
+        persistent_ctx.failed_procedures = vec![ProcedureMeta {
+            plan_index: 0,
+            group_id: persistent_ctx.plans[0].group_id,
+            procedure_id: ProcedureId::random(),
+        }];
+
+        let context = Context::new(
+            &ddl_ctx,
+            env.mailbox_ctx.mailbox().clone(),
+            env.server_addr.clone(),
+            persistent_ctx,
+        );
+        let mut procedure = RepartitionProcedure {
+            state: Box::new(Collect::new(vec![])),
+            context,
+        };
+
+        procedure
+            .rollback(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+
+        let region_routes = current_parent_region_routes(&procedure.context).await;
+        assert_eq!(region_routes.len(), 3);
+        assert_eq!(region_routes[0].region.id, RegionId::new(table_id, 1));
+        assert_eq!(region_routes[1].region.id, RegionId::new(table_id, 2));
+        assert_eq!(region_routes[2].region.id, RegionId::new(table_id, 4));
+    }
+
+    #[tokio::test]
+    async fn test_repartition_rollback_from_collect_restores_failed_group_metadata_only() {
+        let env = TestingEnv::new();
+        let table_id = 1024;
+        let node_manager = Arc::new(MockDatanodeManager::new(UnexpectedErrorDatanodeHandler));
+        let ddl_ctx = env.ddl_context(node_manager);
+        let original_region_routes = vec![
+            test_region_route(
+                RegionId::new(table_id, 1),
+                &range_expr("x", 0, 100).as_json_str().unwrap(),
+            ),
+            test_region_route(
+                RegionId::new(table_id, 2),
+                &range_expr("x", 100, 200).as_json_str().unwrap(),
+            ),
+            test_region_route(RegionId::new(table_id, 3), ""),
+            test_region_route(RegionId::new(table_id, 4), ""),
+        ];
+
+        let failed_plan = with_rollback_metadata(
+            test_plan(table_id),
+            vec![
+                original_region_routes[0].clone(),
+                original_region_routes[2].clone(),
+            ],
+        );
+        let succeeded_plan = RepartitionPlanEntry {
+            group_id: Uuid::new_v4(),
+            source_regions: vec![RegionDescriptor {
+                region_id: RegionId::new(table_id, 2),
+                partition_expr: range_expr("x", 100, 200),
+            }],
+            target_regions: vec![
+                RegionDescriptor {
+                    region_id: RegionId::new(table_id, 2),
+                    partition_expr: range_expr("x", 100, 150),
+                },
+                RegionDescriptor {
+                    region_id: RegionId::new(table_id, 4),
+                    partition_expr: range_expr("x", 150, 200),
+                },
+            ],
+            allocated_region_ids: vec![RegionId::new(table_id, 4)],
+            pending_deallocate_region_ids: vec![],
+            transition_map: vec![vec![0, 1]],
+            original_target_routes: vec![
+                original_region_routes[1].clone(),
+                original_region_routes[3].clone(),
+            ],
+        };
+        let current_region_routes = apply_group_staging(&failed_plan, &original_region_routes);
+        let current_region_routes = apply_group_staging(&succeeded_plan, &current_region_routes);
+        let current_region_routes = exit_group_staging(&succeeded_plan, &current_region_routes);
+        env.create_physical_table_metadata_with_wal_options(
+            table_id,
+            current_region_routes,
+            test_region_wal_options(&[1, 2, 3, 4]),
+        )
+        .await;
+
+        let mut persistent_ctx = PersistentContext::new(
+            TableName::new("test_catalog", "test_schema", "test_table"),
+            table_id,
+            None,
+        );
+        persistent_ctx.plans = vec![failed_plan, succeeded_plan.clone()];
+        persistent_ctx.failed_procedures = vec![ProcedureMeta {
+            plan_index: 0,
+            group_id: persistent_ctx.plans[0].group_id,
+            procedure_id: ProcedureId::random(),
+        }];
+
+        let context = Context::new(
+            &ddl_ctx,
+            env.mailbox_ctx.mailbox().clone(),
+            env.server_addr.clone(),
+            persistent_ctx,
+        );
+        let mut procedure = RepartitionProcedure {
+            state: Box::new(Collect::new(vec![])),
+            context,
+        };
+
+        procedure
+            .rollback(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+
+        assert_eq!(
+            current_parent_region_routes(&procedure.context).await,
+            vec![
+                test_region_route(
+                    RegionId::new(table_id, 1),
+                    &range_expr("x", 0, 100).as_json_str().unwrap(),
+                ),
+                RegionRoute {
+                    region: Region {
+                        id: RegionId::new(table_id, 2),
+                        partition_expr: range_expr("x", 100, 150).as_json_str().unwrap(),
+                        ..Default::default()
+                    },
+                    leader_peer: Some(Peer::empty(1)),
+                    ..Default::default()
+                },
+                RegionRoute {
+                    region: Region {
+                        id: RegionId::new(table_id, 4),
+                        partition_expr: range_expr("x", 150, 200).as_json_str().unwrap(),
+                        ..Default::default()
+                    },
+                    leader_peer: Some(Peer::empty(1)),
+                    ..Default::default()
+                },
+            ]
+        );
+    }
+
+    #[tokio::test]
+    async fn test_repartition_rollback_from_collect_restores_unknown_group_metadata() {
+        let env = TestingEnv::new();
+        let table_id = 1024;
+        let node_manager = Arc::new(MockDatanodeManager::new(UnexpectedErrorDatanodeHandler));
+        let ddl_ctx = env.ddl_context(node_manager);
+        let original_region_routes = vec![
+            test_region_route(
+                RegionId::new(table_id, 1),
+                &range_expr("x", 0, 100).as_json_str().unwrap(),
+            ),
+            test_region_route(
+                RegionId::new(table_id, 2),
+                &range_expr("x", 100, 200).as_json_str().unwrap(),
+            ),
+            test_region_route(RegionId::new(table_id, 3), ""),
+        ];
+        let plan = with_rollback_metadata(
+            test_plan(table_id),
+            vec![
+                original_region_routes[0].clone(),
+                original_region_routes[2].clone(),
+            ],
+        );
+        let staged_region_routes = apply_group_staging(&plan, &original_region_routes);
+        assert_eq!(
+            region_route_by_id(&staged_region_routes, RegionId::new(table_id, 1))
+                .region
+                .partition_expr(),
+            range_expr("x", 0, 50).as_json_str().unwrap()
+        );
+        assert!(
+            region_route_by_id(&staged_region_routes, RegionId::new(table_id, 1))
+                .is_leader_staging()
+        );
+        assert_eq!(
+            region_route_by_id(&staged_region_routes, RegionId::new(table_id, 3))
+                .region
+                .partition_expr(),
+            range_expr("x", 50, 100).as_json_str().unwrap()
+        );
+        assert!(
+            region_route_by_id(&staged_region_routes, RegionId::new(table_id, 3))
+                .is_leader_staging()
+        );
+        env.create_physical_table_metadata_with_wal_options(
+            table_id,
+            staged_region_routes,
+            test_region_wal_options(&[1, 2, 3]),
+        )
+        .await;
+
+        let mut persistent_ctx = PersistentContext::new(
+            TableName::new("test_catalog", "test_schema", "test_table"),
+            table_id,
+            None,
+        );
+        persistent_ctx.plans = vec![plan.clone()];
+        persistent_ctx.unknown_procedures = vec![ProcedureMeta {
+            plan_index: 0,
+            group_id: plan.group_id,
+            procedure_id: ProcedureId::random(),
+        }];
+
+        let context = Context::new(
+            &ddl_ctx,
+            env.mailbox_ctx.mailbox().clone(),
+            env.server_addr.clone(),
+            persistent_ctx,
+        );
+        let mut procedure = RepartitionProcedure {
+            state: Box::new(Collect::new(vec![])),
+            context,
+        };
+
+        procedure
+            .rollback(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+
+        assert_eq!(
+            current_parent_region_routes(&procedure.context).await,
+            vec![
+                original_region_routes[0].clone(),
+                original_region_routes[1].clone()
+            ]
+        );
+    }
+
+    #[tokio::test]
+    async fn test_repartition_rollback_is_idempotent() {
+        let env = TestingEnv::new();
+        let table_id = 1024;
+        let node_manager = Arc::new(MockDatanodeManager::new(UnexpectedErrorDatanodeHandler));
+        let ddl_ctx = env.ddl_context(node_manager);
+        let original_region_routes = vec![
+            test_region_route(
+                RegionId::new(table_id, 1),
+                &range_expr("x", 0, 100).as_json_str().unwrap(),
+            ),
+            test_region_route(
+                RegionId::new(table_id, 2),
+                &range_expr("x", 50, 100).as_json_str().unwrap(),
+            ),
+            test_region_route(RegionId::new(table_id, 3), ""),
+        ];
+        env.create_physical_table_metadata_with_wal_options(
+            table_id,
+            original_region_routes,
+            test_region_wal_options(&[1, 2]),
+        )
+        .await;
+
+        let mut persistent_ctx = PersistentContext::new(
+            TableName::new("test_catalog", "test_schema", "test_table"),
+            table_id,
+            None,
+        );
+        persistent_ctx.plans = vec![with_rollback_metadata(
+            test_plan(table_id),
+            vec![
+                test_region_route(
+                    RegionId::new(table_id, 1),
+                    &range_expr("x", 0, 100).as_json_str().unwrap(),
+                ),
+                test_region_route(RegionId::new(table_id, 3), ""),
+            ],
+        )];
+        persistent_ctx.failed_procedures = vec![ProcedureMeta {
+            plan_index: 0,
+            group_id: Uuid::new_v4(),
+            procedure_id: ProcedureId::random(),
+        }];
+        let context = Context::new(
+            &ddl_ctx,
+            env.mailbox_ctx.mailbox().clone(),
+            env.server_addr.clone(),
+            persistent_ctx,
+        );
+        let mut procedure = RepartitionProcedure {
+            state: Box::new(Dispatch),
+            context,
+        };
+
+        procedure
+            .rollback(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+        let once = current_parent_region_routes(&procedure.context).await;
+
+        procedure
+            .rollback(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+        let twice = current_parent_region_routes(&procedure.context).await;
+
+        assert_eq!(once, twice);
+        assert_eq!(once.len(), 2);
+        assert_eq!(once[0].region.id, RegionId::new(table_id, 1));
+        assert_eq!(once[1].region.id, RegionId::new(table_id, 2));
+    }
+
+    #[tokio::test]
+    async fn test_repartition_rollback_from_collect_restores_failed_merge_group_metadata_only() {
+        let env = TestingEnv::new();
+        let table_id = 1024;
+        let node_manager = Arc::new(MockDatanodeManager::new(UnexpectedErrorDatanodeHandler));
+        let ddl_ctx = env.ddl_context(node_manager);
+        let original_region_routes = vec![
+            test_region_route(
+                RegionId::new(table_id, 1),
+                &range_expr("x", 0, 100).as_json_str().unwrap(),
+            ),
+            test_region_route(
+                RegionId::new(table_id, 2),
+                &range_expr("x", 100, 200).as_json_str().unwrap(),
+            ),
+            test_region_route(
+                RegionId::new(table_id, 3),
+                &range_expr("x", 200, 300).as_json_str().unwrap(),
+            ),
+            test_region_route(RegionId::new(table_id, 4), ""),
+        ];
+        let failed_merge_plan = RepartitionPlanEntry {
+            group_id: Uuid::new_v4(),
+            source_regions: vec![
+                RegionDescriptor {
+                    region_id: RegionId::new(table_id, 1),
+                    partition_expr: range_expr("x", 0, 100),
+                },
+                RegionDescriptor {
+                    region_id: RegionId::new(table_id, 2),
+                    partition_expr: range_expr("x", 100, 200),
+                },
+            ],
+            target_regions: vec![RegionDescriptor {
+                region_id: RegionId::new(table_id, 1),
+                partition_expr: range_expr("x", 0, 200),
+            }],
+            allocated_region_ids: vec![],
+            pending_deallocate_region_ids: vec![RegionId::new(table_id, 2)],
+            transition_map: vec![vec![0], vec![0]],
+            original_target_routes: vec![original_region_routes[0].clone()],
+        };
+        let succeeded_split_plan = RepartitionPlanEntry {
+            group_id: Uuid::new_v4(),
+            source_regions: vec![RegionDescriptor {
+                region_id: RegionId::new(table_id, 3),
+                partition_expr: range_expr("x", 200, 300),
+            }],
+            target_regions: vec![
+                RegionDescriptor {
+                    region_id: RegionId::new(table_id, 3),
+                    partition_expr: range_expr("x", 200, 250),
+                },
+                RegionDescriptor {
+                    region_id: RegionId::new(table_id, 4),
+                    partition_expr: range_expr("x", 250, 300),
+                },
+            ],
+            allocated_region_ids: vec![RegionId::new(table_id, 4)],
+            pending_deallocate_region_ids: vec![],
+            transition_map: vec![vec![0, 1]],
+            original_target_routes: vec![
+                original_region_routes[2].clone(),
+                original_region_routes[3].clone(),
+            ],
+        };
+        let current_region_routes =
+            apply_group_staging(&failed_merge_plan, &original_region_routes);
+        let current_region_routes =
+            apply_group_staging(&succeeded_split_plan, &current_region_routes);
+        let staged_region_routes =
+            exit_group_staging(&succeeded_split_plan, &current_region_routes);
+        env.create_physical_table_metadata_with_wal_options(
+            table_id,
+            staged_region_routes,
+            test_region_wal_options(&[1, 2, 3, 4]),
+        )
+        .await;
+
+        let mut persistent_ctx = PersistentContext::new(
+            TableName::new("test_catalog", "test_schema", "test_table"),
+            table_id,
+            None,
+        );
+        persistent_ctx.plans = vec![failed_merge_plan, succeeded_split_plan.clone()];
+        persistent_ctx.failed_procedures = vec![ProcedureMeta {
+            plan_index: 0,
+            group_id: persistent_ctx.plans[0].group_id,
+            procedure_id: ProcedureId::random(),
+        }];
+
+        let context = Context::new(
+            &ddl_ctx,
+            env.mailbox_ctx.mailbox().clone(),
+            env.server_addr.clone(),
+            persistent_ctx,
+        );
+        let mut procedure = RepartitionProcedure {
+            state: Box::new(Collect::new(vec![])),
+            context,
+        };
+
+        procedure
+            .rollback(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+
+        let region_routes = current_parent_region_routes(&procedure.context).await;
+        assert_eq!(
+            region_routes,
+            vec![
+                test_region_route(
+                    RegionId::new(table_id, 1),
+                    &range_expr("x", 0, 100).as_json_str().unwrap(),
+                ),
+                test_region_route(
+                    RegionId::new(table_id, 2),
+                    &range_expr("x", 100, 200).as_json_str().unwrap(),
+                ),
+                RegionRoute {
+                    region: Region {
+                        id: RegionId::new(table_id, 3),
+                        partition_expr: range_expr("x", 200, 250).as_json_str().unwrap(),
+                        ..Default::default()
+                    },
+                    leader_peer: Some(Peer::empty(1)),
+                    ..Default::default()
+                },
+                RegionRoute {
+                    region: Region {
+                        id: RegionId::new(table_id, 4),
+                        partition_expr: range_expr("x", 250, 300).as_json_str().unwrap(),
+                        ..Default::default()
+                    },
+                    leader_peer: Some(Peer::empty(1)),
+                    ..Default::default()
+                },
+            ]
+        );
+    }
+
+    #[tokio::test]
+    async fn test_repartition_procedure_flow_split_failed_and_full_rollback() {
+        let env = TestingEnv::new();
+        let table_id = 1024;
+        let node_manager = Arc::new(MockDatanodeManager::new(NaiveDatanodeHandler));
+
+        env.create_physical_table_metadata_for_repartition(
+            table_id,
+            vec![
+                test_region_route(
+                    RegionId::new(table_id, 1),
+                    &range_expr("x", 0, 100).as_json_str().unwrap(),
+                ),
+                test_region_route(
+                    RegionId::new(table_id, 2),
+                    &range_expr("x", 100, 200).as_json_str().unwrap(),
+                ),
+            ],
+            test_region_wal_options(&[1, 2]),
+        )
+        .await;
+
+        let context = new_parent_context(&env, node_manager, table_id);
+        let mut procedure = RepartitionProcedure::new(
+            vec![range_expr("x", 0, 100)],
+            vec![range_expr("x", 0, 50), range_expr("x", 50, 100)],
+            context,
+        );
+
+        let start_status = procedure
+            .execute(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+        assert!(!start_status.need_persist());
+        let start_status = procedure
+            .execute(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+        assert!(start_status.need_persist());
+        assert_parent_state::<AllocateRegion>(&procedure);
+
+        let allocate_status = procedure
+            .execute(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+        assert!(allocate_status.need_persist());
+        assert_parent_state::<Dispatch>(&procedure);
+        assert_eq!(procedure.context.persistent_ctx.plans.len(), 1);
+        let plan = &procedure.context.persistent_ctx.plans[0];
+        let expected_plan = test_plan(table_id);
+        assert_eq!(plan.source_regions, expected_plan.source_regions);
+        assert_eq!(plan.target_regions, expected_plan.target_regions);
+        assert_eq!(
+            plan.allocated_region_ids,
+            expected_plan.allocated_region_ids
+        );
+        assert_eq!(
+            plan.pending_deallocate_region_ids,
+            expected_plan.pending_deallocate_region_ids
+        );
+        assert_eq!(plan.transition_map, expected_plan.transition_map);
+        assert_eq!(
+            current_parent_region_routes(&procedure.context).await,
+            vec![
+                test_region_route(
+                    RegionId::new(table_id, 1),
+                    &range_expr("x", 0, 100).as_json_str().unwrap(),
+                ),
+                test_region_route(
+                    RegionId::new(table_id, 2),
+                    &range_expr("x", 100, 200).as_json_str().unwrap(),
+                ),
+                RegionRoute {
+                    region: Region {
+                        id: RegionId::new(table_id, 3),
+                        partition_expr: range_expr("x", 50, 100).as_json_str().unwrap(),
+                        ..Default::default()
+                    },
+                    leader_peer: Some(Peer::empty(0)),
+                    ..Default::default()
+                },
+            ]
+        );
+
+        let dispatch_status = procedure
+            .execute(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+        assert!(!dispatch_status.need_persist());
+        let subprocedure_ids = extract_subprocedure_ids(dispatch_status);
+        assert_eq!(subprocedure_ids.len(), 1);
+        assert_parent_state::<Collect>(&procedure);
+
+        let failed_state = ProcedureState::failed(Arc::new(ProcedureError::external(
+            MockError::new(StatusCode::Internal),
+        )));
+        let collect_ctx = procedure_context_with_receivers(HashMap::from([(
+            subprocedure_ids[0],
+            procedure_state_receiver(failed_state),
+        )]));
+
+        let err = procedure.execute(&collect_ctx).await.unwrap_err();
+        assert!(!err.is_retry_later());
+        assert_parent_state::<Collect>(&procedure);
+
+        procedure
+            .rollback(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+
+        let region_routes = current_parent_region_routes(&procedure.context).await;
+        assert_eq!(
+            region_routes,
+            vec![
+                test_region_route(
+                    RegionId::new(table_id, 1),
+                    &range_expr("x", 0, 100).as_json_str().unwrap(),
+                ),
+                test_region_route(
+                    RegionId::new(table_id, 2),
+                    &range_expr("x", 100, 200).as_json_str().unwrap(),
+                ),
+            ]
+        );
+    }
+
+    #[tokio::test]
+    async fn test_repartition_procedure_flow_split_allocate_retryable_then_resume() {
+        common_telemetry::init_default_ut_logging();
+        let env = TestingEnv::new();
+        let table_id = 1024;
+        let (tx, _rx) = mpsc::channel(8);
+        let should_retry = Arc::new(AtomicBool::new(true));
+        let datanode_handler = DatanodeWatcher::new(tx).with_handler(move |_, _| {
+            if should_retry.swap(false, Ordering::SeqCst) {
+                return Err(error::Error::RetryLater {
+                    source: BoxedError::new(
+                        error::UnexpectedSnafu {
+                            err_msg: "retry later",
+                        }
+                        .build(),
+                    ),
+                    clean_poisons: false,
+                });
+            }
+
+            Ok(api::region::RegionResponse::new(0))
+        });
+        let node_manager = Arc::new(MockDatanodeManager::new(datanode_handler));
+
+        env.create_physical_table_metadata_for_repartition(
+            table_id,
+            vec![
+                test_region_route(
+                    RegionId::new(table_id, 1),
+                    &range_expr("x", 0, 100).as_json_str().unwrap(),
+                ),
+                test_region_route(
+                    RegionId::new(table_id, 2),
+                    &range_expr("x", 100, 200).as_json_str().unwrap(),
+                ),
+            ],
+            test_region_wal_options(&[1, 2]),
+        )
+        .await;
+
+        let context = new_parent_context(&env, node_manager, table_id);
+        let mut procedure = RepartitionProcedure::new(
+            vec![range_expr("x", 0, 100)],
+            vec![range_expr("x", 0, 50), range_expr("x", 50, 100)],
+            context,
+        );
+
+        let start_status = procedure
+            .execute(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+        assert!(!start_status.need_persist());
+        let start_status = procedure
+            .execute(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+        assert!(start_status.need_persist());
+        assert_parent_state::<AllocateRegion>(&procedure);
+
+        let err = procedure
+            .execute(&TestingEnv::procedure_context())
+            .await
+            .unwrap_err();
+        assert!(err.is_retry_later());
+        assert_parent_state::<AllocateRegion>(&procedure);
+        assert!(!procedure.context.persistent_ctx.plans.is_empty());
+        assert_eq!(
+            current_parent_region_routes(&procedure.context).await,
+            vec![
+                test_region_route(
+                    RegionId::new(table_id, 1),
+                    &range_expr("x", 0, 100).as_json_str().unwrap(),
+                ),
+                test_region_route(
+                    RegionId::new(table_id, 2),
+                    &range_expr("x", 100, 200).as_json_str().unwrap(),
+                ),
+            ]
+        );
+
+        let allocate_status = procedure
+            .execute(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+        assert!(allocate_status.need_persist());
+        assert_parent_state::<Dispatch>(&procedure);
+
+        assert_eq!(procedure.context.persistent_ctx.plans.len(), 1);
+        let plan = &procedure.context.persistent_ctx.plans[0];
+        let expected_plan = test_plan(table_id);
+        assert_eq!(plan.source_regions, expected_plan.source_regions);
+        assert_eq!(plan.target_regions, expected_plan.target_regions);
+        assert_eq!(
+            plan.allocated_region_ids,
+            expected_plan.allocated_region_ids
+        );
+        assert_eq!(plan.transition_map, expected_plan.transition_map);
+        assert_eq!(
+            current_parent_region_routes(&procedure.context).await,
+            vec![
+                test_region_route(
+                    RegionId::new(table_id, 1),
+                    &range_expr("x", 0, 100).as_json_str().unwrap(),
+                ),
+                test_region_route(
+                    RegionId::new(table_id, 2),
+                    &range_expr("x", 100, 200).as_json_str().unwrap(),
+                ),
+                RegionRoute {
+                    region: Region {
+                        id: RegionId::new(table_id, 3),
+                        partition_expr: range_expr("x", 50, 100).as_json_str().unwrap(),
+                        ..Default::default()
+                    },
+                    leader_peer: Some(Peer::empty(0)),
+                    ..Default::default()
+                },
+            ]
+        );
+
+        let dispatch_status = procedure
+            .execute(&TestingEnv::procedure_context())
+            .await
+            .unwrap();
+        assert!(!dispatch_status.need_persist());
+        let subprocedure_ids = extract_subprocedure_ids(dispatch_status);
+        assert_eq!(subprocedure_ids.len(), 1);
+        assert_parent_state::<Collect>(&procedure);
+    }
+}
diff --git a/src/meta-srv/src/procedure/repartition/allocate_region.rs b/src/meta-srv/src/procedure/repartition/allocate_region.rs
index b1bf93d986..c1f3ca2503 100644
--- a/src/meta-srv/src/procedure/repartition/allocate_region.rs
+++ b/src/meta-srv/src/procedure/repartition/allocate_region.rs
@@ -21,13 +21,12 @@ use common_meta::ddl::create_table::template::{
 };
 use common_meta::lock_key::TableLock;
 use common_meta::node_manager::NodeManagerRef;
-use common_meta::region_keeper::{MemoryRegionKeeperRef, OperatingRegionGuard};
-use common_meta::rpc::router::{RegionRoute, operating_leader_regions};
+use common_meta::rpc::router::RegionRoute;
 use common_procedure::{Context as ProcedureContext, Status};
-use common_telemetry::info;
-use serde::{Deserialize, Serialize};
+use common_telemetry::{debug, info};
+use serde::{Deserialize, Deserializer, Serialize};
 use snafu::{OptionExt, ResultExt};
-use store_api::storage::{RegionNumber, TableId};
+use store_api::storage::{RegionId, RegionNumber, TableId};
 use table::metadata::TableInfo;
 use table::table_reference::TableReference;
 use tokio::time::Instant;
@@ -40,14 +39,104 @@ use crate::procedure::repartition::plan::{
 };
 use crate::procedure::repartition::{Context, State};
 
+#[derive(Debug, Clone, Serialize)]
+pub enum AllocateRegion {
+    Build(BuildPlan),
+    Execute(ExecutePlan),
+}
+
+impl<'de> Deserialize<'de> for AllocateRegion {
+    fn deserialize<D>(deserializer: D) -> std::result::Result<Self, D::Error>
+    where
+        D: Deserializer<'de>,
+    {
+        #[derive(Deserialize)]
+        enum CurrentAllocateRegion {
+            Build(BuildPlan),
+            Execute(ExecutePlan),
+        }
+
+        #[derive(Deserialize)]
+        struct LegacyAllocateRegion {
+            plan_entries: Vec<AllocationPlanEntry>,
+        }
+
+        #[derive(Deserialize)]
+        #[serde(untagged)]
+        enum AllocateRegionRepr {
+            Current(CurrentAllocateRegion),
+            Legacy(LegacyAllocateRegion),
+        }
+
+        match AllocateRegionRepr::deserialize(deserializer)? {
+            AllocateRegionRepr::Current(CurrentAllocateRegion::Build(build_plan)) => {
+                Ok(Self::Build(build_plan))
+            }
+            AllocateRegionRepr::Current(CurrentAllocateRegion::Execute(execute_plan)) => {
+                Ok(Self::Execute(execute_plan))
+            }
+            AllocateRegionRepr::Legacy(legacy) => Ok(Self::Build(BuildPlan {
+                plan_entries: legacy.plan_entries,
+            })),
+        }
+    }
+}
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct AllocateRegion {
+pub struct BuildPlan {
     plan_entries: Vec<AllocationPlanEntry>,
 }
 
-#[async_trait::async_trait]
-#[typetag::serde]
-impl State for AllocateRegion {
+impl BuildPlan {
+    async fn next(
+        &mut self,
+        ctx: &mut Context,
+        _procedure_ctx: &ProcedureContext,
+    ) -> Result<(Box<dyn State>, Status)> {
+        let timer = Instant::now();
+        let table_id = ctx.persistent_ctx.table_id;
+        let table_route_value = ctx.get_table_route_value().await?;
+        let mut next_region_number =
+            AllocateRegion::get_next_region_number(table_route_value.max_region_number().unwrap());
+
+        // Converts allocation plan to repartition plan.
+        let repartition_plan_entries = AllocateRegion::convert_to_repartition_plans(
+            table_id,
+            &mut next_region_number,
+            &self.plan_entries,
+            table_route_value.region_routes().unwrap(),
+        )?;
+        let plan_count = repartition_plan_entries.len();
+        let to_allocate = AllocateRegion::count_regions_to_allocate(&repartition_plan_entries);
+        info!(
+            "Repartition allocate regions start, table_id: {}, groups: {}, regions_to_allocate: {}",
+            table_id, plan_count, to_allocate
+        );
+
+        // If no region to allocate, directly dispatch the plan.
+        if AllocateRegion::count_regions_to_allocate(&repartition_plan_entries) == 0 {
+            ctx.persistent_ctx.plans = repartition_plan_entries;
+            ctx.update_allocate_region_elapsed(timer.elapsed());
+            return Ok((Box::new(Dispatch), Status::executing(true)));
+        }
+
+        ctx.persistent_ctx.plans = repartition_plan_entries;
+        debug!(
+            "Repartition allocate regions build plan completed, table_id: {}, elapsed: {:?}",
+            table_id,
+            timer.elapsed()
+        );
+        Ok((
+            Box::new(AllocateRegion::Execute(ExecutePlan)),
+            Status::executing(true),
+        ))
+    }
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ExecutePlan;
+
+impl ExecutePlan {
     async fn next(
         &mut self,
         ctx: &mut Context,
@@ -55,36 +144,13 @@ impl State for AllocateRegion {
     ) -> Result<(Box<dyn State>, Status)> {
         let timer = Instant::now();
         let table_id = ctx.persistent_ctx.table_id;
+        let allocate_regions = AllocateRegion::collect_allocate_regions(&ctx.persistent_ctx.plans);
+        let region_number_and_partition_exprs =
+            AllocateRegion::prepare_region_allocation_data(&allocate_regions)?;
+        let table_info_value = ctx.get_table_info_value().await?;
         let table_route_value = ctx.get_table_route_value().await?;
         // Safety: it is physical table route value.
         let region_routes = table_route_value.region_routes().unwrap();
-        let mut next_region_number =
-            Self::get_next_region_number(table_route_value.max_region_number().unwrap());
-
-        // Converts allocation plan to repartition plan.
-        let repartition_plan_entries = Self::convert_to_repartition_plans(
-            table_id,
-            &mut next_region_number,
-            &self.plan_entries,
-        );
-        let plan_count = repartition_plan_entries.len();
-        let to_allocate = Self::count_regions_to_allocate(&repartition_plan_entries);
-        info!(
-            "Repartition allocate regions start, table_id: {}, groups: {}, regions_to_allocate: {}",
-            table_id, plan_count, to_allocate
-        );
-
-        // If no region to allocate, directly dispatch the plan.
-        if Self::count_regions_to_allocate(&repartition_plan_entries) == 0 {
-            ctx.persistent_ctx.plans = repartition_plan_entries;
-            ctx.update_allocate_region_elapsed(timer.elapsed());
-            return Ok((Box::new(Dispatch), Status::executing(true)));
-        }
-
-        let allocate_regions = Self::collect_allocate_regions(&repartition_plan_entries);
-        let region_number_and_partition_exprs =
-            Self::prepare_region_allocation_data(&allocate_regions)?;
-        let table_info_value = ctx.get_table_info_value().await?;
         let new_allocated_region_routes = ctx
             .region_routes_allocator
             .allocate(
@@ -122,12 +188,13 @@ impl State for AllocateRegion {
             table_id, new_region_count, new_regions_brief
         );
 
-        let _operating_guards = Self::register_operating_regions(
+        // The table route metadata is not updated yet; register it in memory for region lease renewal.
+        let _operating_guards = Context::register_operating_regions(
             &ctx.memory_region_keeper,
             &new_allocated_region_routes,
         )?;
         // Allocates the regions on datanodes.
-        Self::allocate_regions(
+        AllocateRegion::allocate_regions(
             &ctx.node_manager,
             &table_info_value.table_info,
             &new_allocated_region_routes,
@@ -135,21 +202,33 @@ impl State for AllocateRegion {
         )
         .await?;
 
-        // TODO(weny): for metric engine, sync logical regions from the the central region.
-
         // Updates the table routes.
         let table_lock = TableLock::Write(table_id).into();
         let _guard = procedure_ctx.provider.acquire_lock(&table_lock).await;
         let new_region_routes =
-            Self::generate_region_routes(region_routes, &new_allocated_region_routes);
+            AllocateRegion::generate_region_routes(region_routes, &new_allocated_region_routes);
         ctx.update_table_route(&table_route_value, new_region_routes, wal_options)
             .await?;
         ctx.invalidate_table_cache().await?;
 
-        ctx.persistent_ctx.plans = repartition_plan_entries;
         ctx.update_allocate_region_elapsed(timer.elapsed());
         Ok((Box::new(Dispatch), Status::executing(true)))
     }
+}
+
+#[async_trait::async_trait]
+#[typetag::serde]
+impl State for AllocateRegion {
+    async fn next(
+        &mut self,
+        ctx: &mut Context,
+        procedure_ctx: &ProcedureContext,
+    ) -> Result<(Box<dyn State>, Status)> {
+        match self {
+            AllocateRegion::Build(build_plan) => build_plan.next(ctx, procedure_ctx).await,
+            AllocateRegion::Execute(execute_plan) => execute_plan.next(ctx, procedure_ctx).await,
+        }
+    }
 
     fn as_any(&self) -> &dyn Any {
         self
@@ -158,24 +237,7 @@ impl State for AllocateRegion {
 
 impl AllocateRegion {
     pub fn new(plan_entries: Vec<AllocationPlanEntry>) -> Self {
-        Self { plan_entries }
-    }
-
-    fn register_operating_regions(
-        memory_region_keeper: &MemoryRegionKeeperRef,
-        region_routes: &[RegionRoute],
-    ) -> Result<Vec<OperatingRegionGuard>> {
-        let mut operating_guards = Vec::with_capacity(region_routes.len());
-        for (region_id, datanode_id) in operating_leader_regions(region_routes) {
-            let guard = memory_region_keeper
-                .register(datanode_id, region_id)
-                .context(error::RegionOperatingRaceSnafu {
-                    peer_id: datanode_id,
-                    region_id,
-                })?;
-            operating_guards.push(guard);
-        }
-        Ok(operating_guards)
+        AllocateRegion::Build(BuildPlan { plan_entries })
     }
 
     fn generate_region_routes(
@@ -197,25 +259,64 @@ impl AllocateRegion {
 
     /// Converts allocation plan entries to repartition plan entries.
     ///
-    /// This method takes the allocation plan entries and converts them to repartition plan entries,
-    /// updating `next_region_number` for each newly allocated region.
+    /// This method converts allocation intents into concrete repartition plans,
+    /// updates `next_region_number` for newly allocated regions, and captures
+    /// each plan's `original_target_routes` from the current table-route view.
+    ///
+    /// This also persists each plan's pre-staging target routes for rollback.
     fn convert_to_repartition_plans(
         table_id: TableId,
         next_region_number: &mut RegionNumber,
         plan_entries: &[AllocationPlanEntry],
-    ) -> Vec<RepartitionPlanEntry> {
+        current_region_routes: &[RegionRoute],
+    ) -> Result<Vec<RepartitionPlanEntry>> {
+        let region_routes_map = current_region_routes
+            .iter()
+            .map(|route| (route.region.id, route))
+            .collect::<HashMap<_, _>>();
+
         plan_entries
             .iter()
             .map(|plan_entry| {
-                convert_allocation_plan_to_repartition_plan(
+                let mut plan = convert_allocation_plan_to_repartition_plan(
                     table_id,
                     next_region_number,
                     plan_entry,
-                )
+                );
+                Self::capture_plan_original_target_routes(&mut plan, &region_routes_map)?;
+                Ok(plan)
             })
             .collect()
     }
 
+    fn capture_plan_original_target_routes(
+        plan: &mut RepartitionPlanEntry,
+        region_routes_map: &HashMap<RegionId, &RegionRoute>,
+    ) -> Result<()> {
+        // Persist the pre-staging target-route view on the parent plan.
+        // Newly allocated targets are skipped because rollback deletes their
+        // route metadata rather than restoring an original target route.
+        let mut original_target_routes = Vec::with_capacity(plan.target_regions.len());
+        for target in &plan.target_regions {
+            if plan.allocated_region_ids.contains(&target.region_id) {
+                // This target region is to be allocated, so it doesn't exist in current routes.
+                continue;
+            }
+            let route = region_routes_map.get(&target.region_id).context(
+                error::RepartitionTargetRegionMissingSnafu {
+                    group_id: plan.group_id,
+                    region_id: target.region_id,
+                },
+            )?;
+            {
+                original_target_routes.push((*route).clone());
+            }
+        }
+
+        plan.original_target_routes = original_target_routes;
+        Ok(())
+    }
+
     /// Collects all regions that need to be allocated from the repartition plan entries.
     fn collect_allocate_regions(
         repartition_plan_entries: &[RepartitionPlanEntry],
@@ -296,10 +397,13 @@ impl AllocateRegion {
 
 #[cfg(test)]
 mod tests {
+    use common_meta::peer::Peer;
+    use common_meta::rpc::router::{Region, RegionRoute};
     use store_api::storage::RegionId;
     use uuid::Uuid;
 
     use super::*;
+    use crate::procedure::repartition::State;
     use crate::procedure::repartition::test_util::range_expr;
 
     fn create_region_descriptor(
@@ -343,6 +447,20 @@ mod tests {
         }
     }
 
+    fn create_current_region_routes(table_id: TableId, region_numbers: &[u32]) -> Vec<RegionRoute> {
+        region_numbers
+            .iter()
+            .map(|region_number| RegionRoute {
+                region: Region {
+                    id: RegionId::new(table_id, *region_number),
+                    ..Default::default()
+                },
+                leader_peer: Some(Peer::empty(1)),
+                ..Default::default()
+            })
+            .collect()
+    }
+
     #[test]
     fn test_convert_to_repartition_plans_no_allocation() {
         let table_id = 1024;
@@ -359,7 +477,9 @@ mod tests {
             table_id,
             &mut next_region_number,
             &plan_entries,
-        );
+            &create_current_region_routes(table_id, &[1, 2]),
+        )
+        .unwrap();
 
         assert_eq!(result.len(), 1);
         assert_eq!(result[0].target_regions.len(), 2);
@@ -384,7 +504,9 @@ mod tests {
             table_id,
             &mut next_region_number,
             &plan_entries,
-        );
+            &create_current_region_routes(table_id, &[1, 2]),
+        )
+        .unwrap();
 
         assert_eq!(result.len(), 1);
         assert_eq!(result[0].target_regions.len(), 4);
@@ -417,7 +539,9 @@ mod tests {
             table_id,
             &mut next_region_number,
             &plan_entries,
-        );
+            &create_current_region_routes(table_id, &[1, 2, 3, 4]),
+        )
+        .unwrap();
 
         assert_eq!(result.len(), 3);
         assert_eq!(result[0].allocated_region_ids.len(), 1);
@@ -442,7 +566,9 @@ mod tests {
             table_id,
             &mut next_region_number,
             &plan_entries,
-        );
+            &create_current_region_routes(table_id, &[1, 2, 3, 4]),
+        )
+        .unwrap();
 
         let count = AllocateRegion::count_regions_to_allocate(&repartition_plans);
         assert_eq!(count, 2);
@@ -462,7 +588,9 @@ mod tests {
             table_id,
             &mut next_region_number,
             &plan_entries,
-        );
+            &create_current_region_routes(table_id, &[1, 2]),
+        )
+        .unwrap();
 
         let allocate_regions = AllocateRegion::collect_allocate_regions(&repartition_plans);
         assert_eq!(allocate_regions.len(), 2);
@@ -488,4 +616,71 @@ mod tests {
         assert!(!result[0].1.is_empty());
         assert!(!result[1].1.is_empty());
     }
+
+    #[test]
+    fn test_allocate_region_state_backward_compatibility() {
+        // Arrange
+        let serialized = r#"{"repartition_state":"AllocateRegion","plan_entries":[]}"#;
+
+        // Act
+        let state: Box<dyn State> = serde_json::from_str(serialized).unwrap();
+
+        // Assert
+        let allocate_region = state
+            .as_any()
+            .downcast_ref::<AllocateRegion>()
+            .expect("expected AllocateRegion state");
+        match allocate_region {
+            AllocateRegion::Build(build_plan) => assert!(build_plan.plan_entries.is_empty()),
+            AllocateRegion::Execute(_) => panic!("expected build plan"),
+        }
+    }
+
+    #[test]
+    fn test_allocate_region_state_round_trip() {
+        // Arrange
+        let state: Box<dyn State> = Box::new(AllocateRegion::new(vec![]));
+
+        // Act
+        let serialized = serde_json::to_string(&state).unwrap();
+        let deserialized: Box<dyn State> = serde_json::from_str(&serialized).unwrap();
+
+        // Assert
+        assert_eq!(
+            serialized,
+            r#"{"repartition_state":"AllocateRegion","Build":{"plan_entries":[]}}"#
+        );
+        let allocate_region = deserialized
+            .as_any()
+            .downcast_ref::<AllocateRegion>()
+            .expect("expected AllocateRegion state");
+        match allocate_region {
+            AllocateRegion::Build(build_plan) => assert!(build_plan.plan_entries.is_empty()),
+            AllocateRegion::Execute(_) => panic!("expected build plan"),
+        }
+    }
+
+    #[test]
+    fn test_allocate_region_execute_state_round_trip() {
+        // Arrange
+        let state: Box<dyn State> = Box::new(AllocateRegion::Execute(ExecutePlan));
+
+        // Act
+        let serialized = serde_json::to_string(&state).unwrap();
+        let deserialized: Box<dyn State> = serde_json::from_str(&serialized).unwrap();
+
+        // Assert
+        assert_eq!(
+            serialized,
+            r#"{"repartition_state":"AllocateRegion","Execute":null}"#
+        );
+        let allocate_region = deserialized
+            .as_any()
+            .downcast_ref::<AllocateRegion>()
+            .expect("expected AllocateRegion state");
+        match allocate_region {
+            AllocateRegion::Execute(_) => {}
+            AllocateRegion::Build(_) => panic!("expected execute plan"),
+        }
+    }
 }
diff --git a/src/meta-srv/src/procedure/repartition/collect.rs b/src/meta-srv/src/procedure/repartition/collect.rs
index d413158b94..1a6d0c6257 100644
--- a/src/meta-srv/src/procedure/repartition/collect.rs
+++ b/src/meta-srv/src/procedure/repartition/collect.rs
@@ -94,17 +94,28 @@ impl State for Collect {
             }
         }
 
-        let inflight = self.inflight_procedures.len();
         let succeeded = self.succeeded_procedures.len();
         let failed = self.failed_procedures.len();
         let unknown = self.unknown_procedures.len();
         info!(
-            "Collected repartition group results for table_id: {}, inflight: {}, succeeded: {}, failed: {}, unknown: {}",
-            table_id, inflight, succeeded, failed, unknown
+            "Collected repartition group results for table_id: {}, succeeded: {}, failed: {}, unknown: {}",
+            table_id, succeeded, failed, unknown
         );
 
         if failed > 0 || unknown > 0 {
-            // TODO(weny): retry the failed or unknown procedures.
+            ctx.persistent_ctx
+                .failed_procedures
+                .extend(self.failed_procedures.iter());
+            ctx.persistent_ctx
+                .unknown_procedures
+                .extend(self.unknown_procedures.iter());
+            return crate::error::UnexpectedSnafu {
+                violated: format!(
+                    "Repartition groups failed or became unknown, table_id: {}, failed: {}, unknown: {}",
+                    table_id, failed, unknown
+                ),
+            }
+            .fail();
         }
 
         if let Some(start_time) = ctx.volatile_ctx.dispatch_start_time.take() {
@@ -118,3 +129,139 @@ impl State for Collect {
         self
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use common_error::mock::MockError;
+    use common_error::status_code::StatusCode;
+    use common_meta::test_util::MockDatanodeManager;
+    use common_procedure::{
+        Context as ProcedureContext, ContextProvider, Error as ProcedureError, ProcedureId,
+        ProcedureState,
+    };
+    use common_procedure_test::MockContextProvider;
+    use tokio::sync::watch;
+
+    use super::*;
+    use crate::procedure::repartition::PersistentContext;
+    use crate::procedure::repartition::test_util::TestingEnv;
+
+    struct FailedProcedureContextProvider {
+        receiver: watch::Receiver<ProcedureState>,
+        inner: MockContextProvider,
+    }
+
+    #[async_trait::async_trait]
+    impl ContextProvider for FailedProcedureContextProvider {
+        async fn procedure_state(
+            &self,
+            procedure_id: ProcedureId,
+        ) -> common_procedure::Result<Option<ProcedureState>> {
+            self.inner.procedure_state(procedure_id).await
+        }
+
+        async fn procedure_state_receiver(
+            &self,
+            _procedure_id: ProcedureId,
+        ) -> common_procedure::Result<Option<watch::Receiver<ProcedureState>>> {
+            Ok(Some(self.receiver.clone()))
+        }
+
+        async fn try_put_poison(
+            &self,
+            key: &common_procedure::PoisonKey,
+            procedure_id: ProcedureId,
+        ) -> common_procedure::Result<()> {
+            self.inner.try_put_poison(key, procedure_id).await
+        }
+
+        async fn acquire_lock(
+            &self,
+            key: &common_procedure::StringKey,
+        ) -> common_procedure::local::DynamicKeyLockGuard {
+            self.inner.acquire_lock(key).await
+        }
+    }
+
+    #[tokio::test]
+    async fn test_collect_returns_error_when_unknown_exists() {
+        let env = TestingEnv::new();
+        let ddl_ctx = env.ddl_context(Arc::new(MockDatanodeManager::new(())));
+        let persistent_ctx = PersistentContext::new(
+            table::table_name::TableName::new("test_catalog", "test_schema", "test_table"),
+            1024,
+            None,
+        );
+        let mut ctx = crate::procedure::repartition::Context::new(
+            &ddl_ctx,
+            env.mailbox_ctx.mailbox().clone(),
+            env.server_addr.clone(),
+            persistent_ctx,
+        );
+        let mut state = Collect {
+            inflight_procedures: vec![],
+            succeeded_procedures: vec![],
+            failed_procedures: vec![],
+            unknown_procedures: vec![ProcedureMeta {
+                plan_index: 0,
+                group_id: uuid::Uuid::new_v4(),
+                procedure_id: common_procedure::ProcedureId::random(),
+            }],
+        };
+
+        let err = state
+            .next(&mut ctx, &TestingEnv::procedure_context())
+            .await
+            .unwrap_err();
+
+        assert!(!err.is_retryable());
+    }
+
+    #[tokio::test]
+    async fn test_collect_returns_error_when_failed_exists() {
+        let env = TestingEnv::new();
+        let ddl_ctx = env.ddl_context(Arc::new(MockDatanodeManager::new(())));
+        let persistent_ctx = PersistentContext::new(
+            table::table_name::TableName::new("test_catalog", "test_schema", "test_table"),
+            1024,
+            None,
+        );
+        let mut ctx = crate::procedure::repartition::Context::new(
+            &ddl_ctx,
+            env.mailbox_ctx.mailbox().clone(),
+            env.server_addr.clone(),
+            persistent_ctx,
+        );
+        let procedure_id = common_procedure::ProcedureId::random();
+        let (tx, rx) = watch::channel(ProcedureState::Running);
+        tx.send(ProcedureState::failed(Arc::new(ProcedureError::external(
+            MockError::new(StatusCode::Internal),
+        ))))
+        .unwrap();
+        let procedure_ctx = ProcedureContext {
+            procedure_id: ProcedureId::random(),
+            provider: Arc::new(FailedProcedureContextProvider {
+                receiver: rx,
+                inner: MockContextProvider::default(),
+            }),
+        };
+        let mut state = Collect {
+            inflight_procedures: vec![ProcedureMeta {
+                plan_index: 0,
+                group_id: uuid::Uuid::new_v4(),
+                procedure_id,
+            }],
+            succeeded_procedures: vec![],
+            failed_procedures: vec![],
+            unknown_procedures: vec![],
+        };
+
+        let err = state.next(&mut ctx, &procedure_ctx).await.unwrap_err();
+
+        assert_eq!(state.failed_procedures.len(), 1);
+        assert_eq!(state.unknown_procedures.len(), 0);
+        assert!(!err.is_retryable());
+    }
+}
diff --git a/src/meta-srv/src/procedure/repartition/deallocate_region.rs b/src/meta-srv/src/procedure/repartition/deallocate_region.rs
index 12233c27e7..f3102221fb 100644
--- a/src/meta-srv/src/procedure/repartition/deallocate_region.rs
+++ b/src/meta-srv/src/procedure/repartition/deallocate_region.rs
@@ -88,7 +88,8 @@ impl State for DeallocateRegion {
             &ctx.persistent_ctx.schema_name,
             &ctx.persistent_ctx.table_name,
         );
-        // Deallocates the regions on datanodes.
+        // Memory guards are not required here,
+        // because the table metadata still contains routes for the deallocating regions.
         Self::deallocate_regions(
             &ctx.node_manager,
             &ctx.leader_region_registry,
@@ -116,7 +117,7 @@ impl State for DeallocateRegion {
 }
 
 impl DeallocateRegion {
-    async fn deallocate_regions(
+    pub(crate) async fn deallocate_regions(
         node_manager: &NodeManagerRef,
         leader_region_registry: &LeaderRegionRegistryRef,
         table: TableName,
@@ -141,7 +142,7 @@ impl DeallocateRegion {
         Ok(())
     }
 
-    fn filter_deallocatable_region_routes(
+    pub(crate) fn filter_deallocatable_region_routes(
         table_id: TableId,
         region_routes: &[RegionRoute],
         pending_deallocate_region_ids: &HashSet<RegionId>,
@@ -165,7 +166,7 @@ impl DeallocateRegion {
             .collect::<Vec<_>>()
     }
 
-    fn generate_region_routes(
+    pub(crate) fn generate_region_routes(
         region_routes: &[RegionRoute],
         pending_deallocate_region_ids: &HashSet<RegionId>,
     ) -> Vec<RegionRoute> {
@@ -181,12 +182,21 @@ impl DeallocateRegion {
 #[cfg(test)]
 mod tests {
     use std::collections::HashSet;
+    use std::sync::Arc;
 
+    use common_meta::ddl::test_util::datanode_handler::RetryErrorDatanodeHandler;
     use common_meta::peer::Peer;
     use common_meta::rpc::router::{Region, RegionRoute};
+    use common_meta::test_util::MockDatanodeManager;
     use store_api::storage::{RegionId, TableId};
 
+    use crate::error::Error;
+    use crate::procedure::repartition::State;
     use crate::procedure::repartition::deallocate_region::DeallocateRegion;
+    use crate::procedure::repartition::plan::RepartitionPlanEntry;
+    use crate::procedure::repartition::test_util::{
+        TestingEnv, current_parent_region_routes, new_parent_context,
+    };
 
     fn test_region_routes(table_id: TableId) -> Vec<RegionRoute> {
         vec![
@@ -238,4 +248,37 @@ mod tests {
         assert_eq!(new_region_routes.len(), 1);
         assert_eq!(new_region_routes[0].region.id, RegionId::new(table_id, 2));
     }
+
+    #[tokio::test]
+    async fn test_next_retryable_when_deallocate_regions_retry_later() {
+        let env = TestingEnv::new();
+        let table_id = 1024;
+        let original_routes = test_region_routes(table_id);
+
+        env.create_physical_table_metadata(table_id, original_routes.clone())
+            .await;
+
+        let node_manager = Arc::new(MockDatanodeManager::new(RetryErrorDatanodeHandler));
+        let mut ctx = new_parent_context(&env, node_manager, table_id);
+        ctx.persistent_ctx.plans = vec![RepartitionPlanEntry {
+            group_id: uuid::Uuid::new_v4(),
+            source_regions: vec![],
+            target_regions: vec![],
+            allocated_region_ids: vec![],
+            pending_deallocate_region_ids: vec![RegionId::new(table_id, 1)],
+            transition_map: vec![],
+            original_target_routes: vec![],
+        }];
+
+        let mut state = DeallocateRegion;
+
+        let err = state
+            .next(&mut ctx, &TestingEnv::procedure_context())
+            .await
+            .unwrap_err();
+
+        assert!(matches!(err, Error::DeallocateRegions { .. }));
+        assert!(err.is_retryable());
+        assert_eq!(current_parent_region_routes(&ctx).await, original_routes);
+    }
 }
diff --git a/src/meta-srv/src/procedure/repartition/dispatch.rs b/src/meta-srv/src/procedure/repartition/dispatch.rs
index 02dc73362d..3a9f9376f1 100644
--- a/src/meta-srv/src/procedure/repartition/dispatch.rs
+++ b/src/meta-srv/src/procedure/repartition/dispatch.rs
@@ -31,7 +31,7 @@ use crate::procedure::repartition::{self, Context, State};
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct Dispatch;
 
-fn build_region_mapping(
+pub(crate) fn build_region_mapping(
     source_regions: &[RegionDescriptor],
     target_regions: &[RegionDescriptor],
     transition_map: &[Vec<usize>],
@@ -106,7 +106,11 @@ impl State for Dispatch {
 
         Ok((
             Box::new(Collect::new(procedure_metas)),
-            Status::suspended(procedures, true),
+            // The state is not persisted after sub-procedures are spawned.
+            // If metasrv restarts before all sub-procedures complete,
+            // it restores from the `Dispatch` state and re-dispatches them.
+            // This is safe because the sub-procedures are idempotent.
+            Status::suspended(procedures, false),
         ))
     }
 
diff --git a/src/meta-srv/src/procedure/repartition/group.rs b/src/meta-srv/src/procedure/repartition/group.rs
index 2ef764b634..12374e8ada 100644
--- a/src/meta-srv/src/procedure/repartition/group.rs
+++ b/src/meta-srv/src/procedure/repartition/group.rs
@@ -200,6 +200,12 @@ impl Procedure for RepartitionGroupProcedure {
         Self::TYPE_NAME
     }
 
+    async fn rollback(&mut self, _ctx: &ProcedureContext) -> ProcedureResult<()> {
+        // The parent repartition procedure is responsible for rollback and recovery.
+        // Subprocedures are not recovered after metasrv restarts, so implementing rollback for them is meaningless.
+        Ok(())
+    }
+
     #[tracing::instrument(skip_all, fields(
         state = %self.state.name(),
         table_id = self.context.persistent_ctx.table_id,
@@ -238,6 +244,8 @@ impl Procedure for RepartitionGroupProcedure {
     }
 
     fn rollback_supported(&self) -> bool {
+        // Parent repartition owns rollback and recovery because subprocedures are
+        // not relied on as durable rollback units across metasrv restarts.
         false
     }
 
@@ -304,7 +312,7 @@ impl Context {
 pub struct GroupPrepareResult {
     /// The validated source region routes.
     pub source_routes: Vec<RegionRoute>,
-    /// The validated target region routes.
+    /// Validated target region routes used for metadata rollback (logical rollback).
     pub target_routes: Vec<RegionRoute>,
     /// The primary source region id (first source region), used for retrieving region options.
     pub central_region: RegionId,
@@ -597,7 +605,7 @@ pub(crate) trait State: Sync + Send + Debug {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::sync::Arc;
 
     use common_meta::key::TableMetadataManager;
diff --git a/src/meta-srv/src/procedure/repartition/group/apply_staging_manifest.rs b/src/meta-srv/src/procedure/repartition/group/apply_staging_manifest.rs
index 2020e9e2f4..43e5ee31d9 100644
--- a/src/meta-srv/src/procedure/repartition/group/apply_staging_manifest.rs
+++ b/src/meta-srv/src/procedure/repartition/group/apply_staging_manifest.rs
@@ -332,7 +332,14 @@ impl ApplyStagingManifest {
                 );
 
                 Ok(())
-            }
+            },
+            Err(error::Error::MailboxChannelClosed {..})=> error::RetryLaterSnafu {
+                reason: format!(
+                    "Mailbox closed when sending apply staging manifests to datanode {:?}, elapsed: {:?}",
+                    peer,
+                    now.elapsed()
+                ),
+            }.fail()?,
             Err(error::Error::MailboxTimeout { .. }) => {
                 let reason = format!(
                     "Mailbox received timeout for apply staging manifests on datanode {:?}, elapsed: {:?}",
diff --git a/src/meta-srv/src/procedure/repartition/group/enter_staging_region.rs b/src/meta-srv/src/procedure/repartition/group/enter_staging_region.rs
index 18ecfd4bb2..c1957031d5 100644
--- a/src/meta-srv/src/procedure/repartition/group/enter_staging_region.rs
+++ b/src/meta-srv/src/procedure/repartition/group/enter_staging_region.rs
@@ -28,6 +28,7 @@ use common_telemetry::tracing_context::TracingContext;
 use futures::future::{join_all, try_join_all};
 use serde::{Deserialize, Serialize};
 use snafu::{OptionExt, ResultExt, ensure};
+use store_api::region_request::RegionFlushReason;
 use store_api::storage::RegionId;
 
 use crate::error::{self, Error, Result};
@@ -315,7 +316,14 @@ impl EnterStagingRegion {
                 );
 
                 Ok(())
-            }
+            },
+            Err(error::Error::MailboxChannelClosed {..})=> error::RetryLaterSnafu {
+                reason: format!(
+                    "Mailbox closed when sending enter staging regions to datanode {:?}, elapsed: {:?}",
+                    peer,
+                    now.elapsed()
+                ),
+            }.fail()?,
             Err(error::Error::MailboxTimeout { .. }) => {
                 let reason = format!(
                     "Mailbox received timeout for enter staging regions on datanode {:?}, elapsed: {:?}",
@@ -411,6 +419,7 @@ impl EnterStagingRegion {
                     peer,
                     operation_timeout,
                     ErrorStrategy::Retry,
+                    Some(RegionFlushReason::Repartition),
                 )
             })
             .collect::<Vec<_>>();
@@ -433,7 +442,7 @@ impl EnterStagingRegion {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::time::Duration;
 
     use common_meta::instruction::StagingPartitionDirective;
diff --git a/src/meta-srv/src/procedure/repartition/group/remap_manifest.rs b/src/meta-srv/src/procedure/repartition/group/remap_manifest.rs
index 6e3460c2ce..1d6a75100e 100644
--- a/src/meta-srv/src/procedure/repartition/group/remap_manifest.rs
+++ b/src/meta-srv/src/procedure/repartition/group/remap_manifest.rs
@@ -184,6 +184,14 @@ impl RemapManifest {
 
                 Self::handle_remap_manifest_reply(remap.region_id, reply, &now, peer)
             }
+            Err(error::Error::MailboxChannelClosed { .. }) => error::RetryLaterSnafu {
+                reason: format!(
+                    "Mailbox closed when sending remap manifests to datanode {:?}, elapsed: {:?}",
+                    peer,
+                    now.elapsed()
+                ),
+            }
+            .fail()?,
             Err(error::Error::MailboxTimeout { .. }) => {
                 let reason = format!(
                     "Mailbox received timeout for remap manifests on datanode {:?}, elapsed: {:?}",
diff --git a/src/meta-srv/src/procedure/repartition/group/repartition_start.rs b/src/meta-srv/src/procedure/repartition/group/repartition_start.rs
index 72de2f2934..8b8b5208b4 100644
--- a/src/meta-srv/src/procedure/repartition/group/repartition_start.rs
+++ b/src/meta-srv/src/procedure/repartition/group/repartition_start.rs
@@ -207,7 +207,7 @@ impl State for RepartitionStart {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_meta::peer::Peer;
     use common_meta::rpc::router::{Region, RegionRoute};
diff --git a/src/meta-srv/src/procedure/repartition/group/sync_region.rs b/src/meta-srv/src/procedure/repartition/group/sync_region.rs
index 5e842da2fb..f95b20e95f 100644
--- a/src/meta-srv/src/procedure/repartition/group/sync_region.rs
+++ b/src/meta-srv/src/procedure/repartition/group/sync_region.rs
@@ -27,6 +27,7 @@ use futures::future::join_all;
 use serde::{Deserialize, Serialize};
 use snafu::{OptionExt, ResultExt, ensure};
 use store_api::region_engine::SyncRegionFromRequest;
+use store_api::region_request::RegionFlushReason;
 use store_api::storage::RegionId;
 
 use crate::error::{self, Error, Result};
@@ -85,6 +86,7 @@ impl SyncRegion {
             &prepare_result.central_region_datanode,
             operation_timeout,
             ErrorStrategy::Retry,
+            Some(RegionFlushReason::Repartition),
         )
         .await
     }
@@ -273,6 +275,14 @@ impl SyncRegion {
                 }
                 Ok(())
             }
+            Err(error::Error::MailboxChannelClosed { .. }) => error::RetryLaterSnafu {
+                reason: format!(
+                    "Mailbox closed when sending sync region to datanode {:?}, elapsed: {:?}",
+                    peer,
+                    now.elapsed()
+                ),
+            }
+            .fail()?,
             Err(error::Error::MailboxTimeout { .. }) => {
                 let reason = format!(
                     "Mailbox received timeout for sync regions on datanode {:?}, elapsed: {:?}",
@@ -338,7 +348,7 @@ impl SyncRegion {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_meta::peer::Peer;
     use common_meta::rpc::router::{Region, RegionRoute};
diff --git a/src/meta-srv/src/procedure/repartition/group/update_metadata.rs b/src/meta-srv/src/procedure/repartition/group/update_metadata.rs
index cf3b111971..90fe6c5893 100644
--- a/src/meta-srv/src/procedure/repartition/group/update_metadata.rs
+++ b/src/meta-srv/src/procedure/repartition/group/update_metadata.rs
@@ -14,7 +14,6 @@
 
 pub(crate) mod apply_staging_region;
 pub(crate) mod exit_staging_region;
-pub(crate) mod rollback_staging_region;
 
 use std::any::Any;
 use std::time::Instant;
@@ -34,9 +33,7 @@ use crate::procedure::repartition::group::{Context, State};
 pub enum UpdateMetadata {
     /// Applies the new partition expressions for staging regions.
     ApplyStaging,
-    /// Rolls back the new partition expressions for staging regions.
-    RollbackStaging,
-    /// Exits the staging regions.
+    /// Exits the staging regions after the group finishes its forward path.
     ExitStaging,
 }
 
@@ -64,17 +61,6 @@ impl State for UpdateMetadata {
                 ctx.update_update_metadata_elapsed(timer.elapsed());
                 Ok((Box::new(EnterStagingRegion), Status::executing(false)))
             }
-            UpdateMetadata::RollbackStaging => {
-                self.rollback_staging_regions(ctx).await?;
-
-                if let Err(err) = ctx.invalidate_table_cache().await {
-                    warn!(
-                        err;
-                        "Failed to broadcast the invalidate table cache message during the rollback staging regions"
-                    );
-                };
-                Ok((Box::new(RepartitionEnd), Status::executing(false)))
-            }
             UpdateMetadata::ExitStaging => {
                 self.exit_staging_regions(ctx).await?;
                 if let Err(err) = ctx.invalidate_table_cache().await {
diff --git a/src/meta-srv/src/procedure/repartition/group/update_metadata/apply_staging_region.rs b/src/meta-srv/src/procedure/repartition/group/update_metadata/apply_staging_region.rs
index ecde5f0507..ff01161ff5 100644
--- a/src/meta-srv/src/procedure/repartition/group/update_metadata/apply_staging_region.rs
+++ b/src/meta-srv/src/procedure/repartition/group/update_metadata/apply_staging_region.rs
@@ -30,7 +30,7 @@ impl UpdateMetadata {
     /// Abort:
     /// - Target region not found.
     /// - Source region not found.
-    fn apply_staging_region_routes(
+    pub(crate) fn apply_staging_region_routes(
         group_id: GroupId,
         sources: &[RegionDescriptor],
         targets: &[RegionDescriptor],
@@ -50,10 +50,12 @@ impl UpdateMetadata {
                     region_id: target.region_id,
                 },
             )?;
+            // Set the new partition expression for the target region route.
             region_route.region.partition_expr = target
                 .partition_expr
                 .as_json_str()
                 .context(error::SerializePartitionExprSnafu)?;
+            // Set leader staging state and write route policy for the target region route.
             region_route.set_leader_staging();
             region_route.clear_ignore_all_writes();
         }
@@ -65,6 +67,7 @@ impl UpdateMetadata {
                     region_id: source.region_id,
                 },
             )?;
+            // Set leader staging state for the source region route.
             region_route.set_leader_staging();
             if pending_deallocate_region_ids.contains(&source.region_id) {
                 // When a region is pending deallocation, it should ignore all writes.
diff --git a/src/meta-srv/src/procedure/repartition/group/update_metadata/exit_staging_region.rs b/src/meta-srv/src/procedure/repartition/group/update_metadata/exit_staging_region.rs
index f8e505f0ff..50864daa93 100644
--- a/src/meta-srv/src/procedure/repartition/group/update_metadata/exit_staging_region.rs
+++ b/src/meta-srv/src/procedure/repartition/group/update_metadata/exit_staging_region.rs
@@ -25,7 +25,7 @@ use crate::procedure::repartition::group::{Context, GroupId, region_routes};
 use crate::procedure::repartition::plan::RegionDescriptor;
 
 impl UpdateMetadata {
-    fn exit_staging_region_routes(
+    pub(crate) fn exit_staging_region_routes(
         group_id: GroupId,
         sources: &[RegionDescriptor],
         targets: &[RegionDescriptor],
diff --git a/src/meta-srv/src/procedure/repartition/group/update_metadata/rollback_staging_region.rs b/src/meta-srv/src/procedure/repartition/group/update_metadata/rollback_staging_region.rs
deleted file mode 100644
index e9bef4cf8e..0000000000
--- a/src/meta-srv/src/procedure/repartition/group/update_metadata/rollback_staging_region.rs
+++ /dev/null
@@ -1,199 +0,0 @@
-// Copyright 2023 Greptime Team
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-use std::collections::HashMap;
-
-use common_error::ext::BoxedError;
-use common_meta::rpc::router::RegionRoute;
-use common_telemetry::{error, info};
-use snafu::{OptionExt, ResultExt};
-
-use crate::error::{self, Result};
-use crate::procedure::repartition::group::update_metadata::UpdateMetadata;
-use crate::procedure::repartition::group::{Context, GroupId, region_routes};
-
-impl UpdateMetadata {
-    /// Rolls back the staging regions.
-    ///
-    /// Abort:
-    /// - Source region not found.
-    /// - Target region not found.
-    fn rollback_staging_region_routes(
-        group_id: GroupId,
-        source_routes: &[RegionRoute],
-        target_routes: &[RegionRoute],
-        current_region_routes: &[RegionRoute],
-    ) -> Result<Vec<RegionRoute>> {
-        let mut region_routes = current_region_routes.to_vec();
-        let mut region_routes_map = region_routes
-            .iter_mut()
-            .map(|route| (route.region.id, route))
-            .collect::<HashMap<_, _>>();
-
-        for source in source_routes {
-            let region_route = region_routes_map.get_mut(&source.region.id).context(
-                error::RepartitionSourceRegionMissingSnafu {
-                    group_id,
-                    region_id: source.region.id,
-                },
-            )?;
-            region_route.region.partition_expr = source.region.partition_expr.clone();
-            region_route.clear_leader_staging();
-            region_route.clear_ignore_all_writes();
-        }
-
-        for target in target_routes {
-            let region_route = region_routes_map.get_mut(&target.region.id).context(
-                error::RepartitionTargetRegionMissingSnafu {
-                    group_id,
-                    region_id: target.region.id,
-                },
-            )?;
-            region_route.clear_leader_staging();
-        }
-
-        Ok(region_routes)
-    }
-
-    /// Rolls back the metadata for staging regions.
-    ///
-    /// Abort:
-    /// - Table route is not physical.
-    /// - Source region not found.
-    /// - Target region not found.
-    /// - Failed to update the table route.
-    /// - Central region datanode table value not found.
-    pub(crate) async fn rollback_staging_regions(&self, ctx: &mut Context) -> Result<()> {
-        let table_id = ctx.persistent_ctx.table_id;
-        let group_id = ctx.persistent_ctx.group_id;
-        let current_table_route_value = ctx.get_table_route_value().await?;
-        let region_routes = region_routes(table_id, current_table_route_value.get_inner_ref())?;
-        // Safety: prepare result is set in [RepartitionStart] state.
-        let prepare_result = ctx.persistent_ctx.group_prepare_result.as_ref().unwrap();
-        let new_region_routes = Self::rollback_staging_region_routes(
-            group_id,
-            &prepare_result.source_routes,
-            &prepare_result.target_routes,
-            region_routes,
-        )?;
-
-        let source_count = prepare_result.source_routes.len();
-        let target_count = prepare_result.target_routes.len();
-        info!(
-            "Rollback staging regions for repartition, table_id: {}, group_id: {}, sources: {}, targets: {}",
-            table_id, group_id, source_count, target_count
-        );
-
-        if let Err(err) = ctx
-            .update_table_route(&current_table_route_value, new_region_routes)
-            .await
-        {
-            error!(err; "Failed to update the table route during the updating metadata for repartition: {table_id}, group_id: {group_id}");
-            return Err(BoxedError::new(err)).context(error::RetryLaterWithSourceSnafu {
-                reason: format!(
-                    "Failed to update the table route during the updating metadata for repartition: {table_id}, group_id: {group_id}"
-                ),
-            });
-        };
-
-        Ok(())
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use common_meta::peer::Peer;
-    use common_meta::rpc::router::{LeaderState, Region, RegionRoute};
-    use store_api::storage::RegionId;
-    use uuid::Uuid;
-
-    use crate::procedure::repartition::group::update_metadata::UpdateMetadata;
-    use crate::procedure::repartition::test_util::range_expr;
-
-    #[test]
-    fn test_rollback_staging_region_routes() {
-        let group_id = Uuid::new_v4();
-        let table_id = 1024;
-        let region_routes = vec![
-            {
-                let mut route = RegionRoute {
-                    region: Region {
-                        id: RegionId::new(table_id, 1),
-                        partition_expr: range_expr("x", 0, 100).as_json_str().unwrap(),
-                        ..Default::default()
-                    },
-                    leader_peer: Some(Peer::empty(1)),
-                    leader_state: Some(LeaderState::Staging),
-                    ..Default::default()
-                };
-                route.set_ignore_all_writes();
-                route
-            },
-            RegionRoute {
-                region: Region {
-                    id: RegionId::new(table_id, 2),
-                    partition_expr: String::new(),
-                    ..Default::default()
-                },
-                leader_peer: Some(Peer::empty(1)),
-                leader_state: Some(LeaderState::Staging),
-                ..Default::default()
-            },
-            RegionRoute {
-                region: Region {
-                    id: RegionId::new(table_id, 3),
-                    partition_expr: String::new(),
-                    ..Default::default()
-                },
-                leader_peer: Some(Peer::empty(1)),
-                leader_state: Some(LeaderState::Downgrading),
-                ..Default::default()
-            },
-        ];
-        let source_routes = vec![RegionRoute {
-            region: Region {
-                id: RegionId::new(table_id, 1),
-                partition_expr: range_expr("x", 0, 20).as_json_str().unwrap(),
-                ..Default::default()
-            },
-            leader_peer: Some(Peer::empty(1)),
-            ..Default::default()
-        }];
-        let target_routes = vec![RegionRoute {
-            region: Region {
-                id: RegionId::new(table_id, 2),
-                partition_expr: range_expr("x", 0, 20).as_json_str().unwrap(),
-                ..Default::default()
-            },
-            leader_peer: Some(Peer::empty(1)),
-            ..Default::default()
-        }];
-        let new_region_routes = UpdateMetadata::rollback_staging_region_routes(
-            group_id,
-            &source_routes,
-            &target_routes,
-            &region_routes,
-        )
-        .unwrap();
-        assert!(!new_region_routes[0].is_leader_staging());
-        assert!(!new_region_routes[0].is_ignore_all_writes());
-        assert_eq!(
-            new_region_routes[0].region.partition_expr,
-            range_expr("x", 0, 20).as_json_str().unwrap(),
-        );
-        assert!(!new_region_routes[1].is_leader_staging());
-        assert!(!new_region_routes[1].is_ignore_all_writes());
-        assert!(new_region_routes[2].is_leader_downgrading());
-    }
-}
diff --git a/src/meta-srv/src/procedure/repartition/plan.rs b/src/meta-srv/src/procedure/repartition/plan.rs
index 9eabec7101..063a64341b 100644
--- a/src/meta-srv/src/procedure/repartition/plan.rs
+++ b/src/meta-srv/src/procedure/repartition/plan.rs
@@ -14,6 +14,7 @@
 
 use std::cmp::Ordering;
 
+use common_meta::rpc::router::RegionRoute;
 use partition::expr::PartitionExpr;
 use serde::{Deserialize, Serialize};
 use store_api::storage::{RegionId, RegionNumber, TableId};
@@ -46,7 +47,7 @@ pub struct AllocationPlanEntry {
 
 /// A plan entry for the dispatch phase after region allocation,
 /// with concrete source and target region descriptors.
-#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
 pub struct RepartitionPlanEntry {
     /// The group id for this plan entry.
     pub group_id: GroupId,
@@ -61,6 +62,9 @@ pub struct RepartitionPlanEntry {
     /// For each `source_regions[k]`, the corresponding vector contains global
     /// `target_regions` that overlap with it.
     pub transition_map: Vec<Vec<usize>>,
+    /// Pre-staging target routes persisted for parent rollback and recovery.
+    #[serde(default)]
+    pub original_target_routes: Vec<RegionRoute>,
 }
 
 impl RepartitionPlanEntry {
@@ -138,6 +142,7 @@ pub fn convert_allocation_plan_to_repartition_plan(
                 allocated_region_ids,
                 pending_deallocate_region_ids: vec![],
                 transition_map: transition_map.clone(),
+                original_target_routes: vec![],
             }
         }
         Ordering::Equal => {
@@ -157,6 +162,7 @@ pub fn convert_allocation_plan_to_repartition_plan(
                 allocated_region_ids: vec![],
                 pending_deallocate_region_ids: vec![],
                 transition_map: transition_map.clone(),
+                original_target_routes: vec![],
             }
         }
         Ordering::Greater => {
@@ -184,6 +190,7 @@ pub fn convert_allocation_plan_to_repartition_plan(
                 allocated_region_ids: vec![],
                 pending_deallocate_region_ids,
                 transition_map: transition_map.clone(),
+                original_target_routes: vec![],
             }
         }
     }
diff --git a/src/meta-srv/src/procedure/repartition/repartition_start.rs b/src/meta-srv/src/procedure/repartition/repartition_start.rs
index 1f657d58f2..5c6bcfdb06 100644
--- a/src/meta-srv/src/procedure/repartition/repartition_start.rs
+++ b/src/meta-srv/src/procedure/repartition/repartition_start.rs
@@ -102,7 +102,7 @@ impl State for RepartitionStart {
 }
 
 impl RepartitionStart {
-    fn build_plan(
+    pub(crate) fn build_plan(
         physical_route: &PhysicalTableRouteValue,
         from_exprs: &[PartitionExpr],
         to_exprs: &[PartitionExpr],
diff --git a/src/meta-srv/src/procedure/repartition/test_util.rs b/src/meta-srv/src/procedure/repartition/test_util.rs
index 3cefd4a095..83856a49e6 100644
--- a/src/meta-srv/src/procedure/repartition/test_util.rs
+++ b/src/meta-srv/src/procedure/repartition/test_util.rs
@@ -16,22 +16,41 @@ use std::collections::HashMap;
 use std::sync::Arc;
 use std::time::{Duration, Instant};
 
+use common_meta::ddl::DdlContext;
+use common_meta::key::table_route::TableRouteValue;
+use common_meta::key::test_utils::{new_test_table_info, new_test_table_info_with_name};
 use common_meta::key::{TableMetadataManager, TableMetadataManagerRef};
+use common_meta::kv_backend::KvBackendRef;
 use common_meta::kv_backend::memory::MemoryKvBackend;
+use common_meta::node_manager::NodeManagerRef;
+use common_meta::peer::Peer;
+use common_meta::rpc::router::{Region, RegionRoute};
 use common_meta::sequence::SequenceBuilder;
+use common_meta::test_util::new_ddl_context_with_kv_backend;
+use common_procedure::{
+    Context as ProcedureContext, ContextProvider, ProcedureId, ProcedureState, Status,
+};
+use common_procedure_test::MockContextProvider;
+use common_wal::options::{KafkaWalOptions, WalOptions};
 use datatypes::value::Value;
 use partition::expr::{PartitionExpr, col};
-use store_api::storage::TableId;
+use store_api::storage::{RegionId, RegionNumber, TableId};
+use table::table_name::TableName;
+use tokio::sync::watch;
 use uuid::Uuid;
 
 use crate::cache_invalidator::MetasrvCacheInvalidator;
 use crate::metasrv::MetasrvInfo;
 use crate::procedure::repartition::group::{Context, PersistentContext, VolatileContext};
 use crate::procedure::repartition::plan::RegionDescriptor;
+use crate::procedure::repartition::{
+    Context as ParentContext, PersistentContext as ParentPersistentContext, RepartitionProcedure,
+};
 use crate::procedure::test_util::MailboxContext;
 
 /// `TestingEnv` provides components during the tests.
 pub struct TestingEnv {
+    pub kv_backend: KvBackendRef,
     pub table_metadata_manager: TableMetadataManagerRef,
     pub mailbox_ctx: MailboxContext,
     pub server_addr: String,
@@ -45,13 +64,14 @@ impl Default for TestingEnv {
 
 impl TestingEnv {
     pub fn new() -> Self {
-        let kv_backend = Arc::new(MemoryKvBackend::new());
+        let kv_backend = Arc::new(MemoryKvBackend::new()) as KvBackendRef;
         let table_metadata_manager = Arc::new(TableMetadataManager::new(kv_backend.clone()));
         let mailbox_sequence =
             SequenceBuilder::new("test_heartbeat_mailbox", kv_backend.clone()).build();
         let mailbox_ctx = MailboxContext::new(mailbox_sequence);
 
         Self {
+            kv_backend,
             table_metadata_manager,
             mailbox_ctx,
             server_addr: "localhost".to_string(),
@@ -76,6 +96,65 @@ impl TestingEnv {
             volatile_ctx: VolatileContext::default(),
         }
     }
+
+    pub fn procedure_context() -> ProcedureContext {
+        ProcedureContext {
+            procedure_id: ProcedureId::random(),
+            provider: Arc::new(MockContextProvider::default()),
+        }
+    }
+
+    pub async fn create_physical_table_metadata(
+        &self,
+        table_id: TableId,
+        region_routes: Vec<RegionRoute>,
+    ) {
+        self.create_physical_table_metadata_with_wal_options(
+            table_id,
+            region_routes,
+            HashMap::default(),
+        )
+        .await;
+    }
+
+    pub async fn create_physical_table_metadata_with_wal_options(
+        &self,
+        table_id: TableId,
+        region_routes: Vec<RegionRoute>,
+        region_wal_options: HashMap<RegionNumber, String>,
+    ) {
+        self.table_metadata_manager
+            .create_table_metadata(
+                new_test_table_info(table_id),
+                TableRouteValue::physical(region_routes),
+                region_wal_options,
+            )
+            .await
+            .unwrap();
+    }
+
+    pub async fn create_physical_table_metadata_for_repartition(
+        &self,
+        table_id: TableId,
+        region_routes: Vec<RegionRoute>,
+        region_wal_options: HashMap<RegionNumber, String>,
+    ) {
+        let mut table_info = new_test_table_info_with_name(table_id, "test_table");
+        table_info.meta.column_ids = vec![0, 1, 2];
+
+        self.table_metadata_manager
+            .create_table_metadata(
+                table_info,
+                TableRouteValue::physical(region_routes),
+                region_wal_options,
+            )
+            .await
+            .unwrap();
+    }
+
+    pub fn ddl_context(&self, node_manager: NodeManagerRef) -> DdlContext {
+        new_ddl_context_with_kv_backend(node_manager, self.kv_backend.clone())
+    }
 }
 
 pub fn range_expr(col_name: &str, start: i64, end: i64) -> PartitionExpr {
@@ -84,6 +163,18 @@ pub fn range_expr(col_name: &str, start: i64, end: i64) -> PartitionExpr {
         .and(col(col_name).lt(Value::Int64(end)))
 }
 
+pub fn test_region_wal_options(region_numbers: &[RegionNumber]) -> HashMap<RegionNumber, String> {
+    let wal_options = serde_json::to_string(&WalOptions::Kafka(KafkaWalOptions {
+        topic: "test_topic".to_string(),
+    }))
+    .unwrap();
+
+    region_numbers
+        .iter()
+        .map(|region_number| (*region_number, wal_options.clone()))
+        .collect()
+}
+
 pub fn new_persistent_context(
     table_id: TableId,
     sources: Vec<RegionDescriptor>,
@@ -105,3 +196,110 @@ pub fn new_persistent_context(
         timeout: Duration::from_secs(120),
     }
 }
+
+pub fn test_region_route(region_id: RegionId, partition_expr: &str) -> RegionRoute {
+    RegionRoute {
+        region: Region {
+            id: region_id,
+            partition_expr: partition_expr.to_string(),
+            ..Default::default()
+        },
+        leader_peer: Some(Peer::empty(1)),
+        ..Default::default()
+    }
+}
+
+pub async fn current_parent_region_routes(ctx: &ParentContext) -> Vec<RegionRoute> {
+    let table_route_value = ctx.get_table_route_value().await.unwrap().into_inner();
+    table_route_value.region_routes().unwrap().clone()
+}
+
+pub fn new_parent_context(
+    env: &TestingEnv,
+    node_manager: NodeManagerRef,
+    table_id: TableId,
+) -> ParentContext {
+    let ddl_ctx = env.ddl_context(node_manager);
+    let persistent_ctx = ParentPersistentContext::new(
+        TableName::new("test_catalog", "test_schema", "test_table"),
+        table_id,
+        None,
+    );
+
+    ParentContext::new(
+        &ddl_ctx,
+        env.mailbox_ctx.mailbox().clone(),
+        env.server_addr.clone(),
+        persistent_ctx,
+    )
+}
+
+pub fn assert_parent_state<T: 'static>(procedure: &RepartitionProcedure) {
+    assert!(procedure.state.as_any().is::<T>());
+}
+
+pub fn extract_subprocedure_ids(status: Status) -> Vec<ProcedureId> {
+    let Status::Suspended { subprocedures, .. } = status else {
+        panic!("expected suspended status");
+    };
+
+    subprocedures
+        .into_iter()
+        .map(|procedure| procedure.id)
+        .collect()
+}
+
+pub fn procedure_state_receiver(state: ProcedureState) -> watch::Receiver<ProcedureState> {
+    let (tx, rx) = watch::channel(ProcedureState::Running);
+    tx.send(state).unwrap();
+    rx
+}
+
+pub fn procedure_context_with_receivers(
+    receivers: HashMap<ProcedureId, watch::Receiver<ProcedureState>>,
+) -> ProcedureContext {
+    ProcedureContext {
+        procedure_id: ProcedureId::random(),
+        provider: Arc::new(ProcedureStateReceiverProvider {
+            receivers,
+            inner: MockContextProvider::default(),
+        }),
+    }
+}
+
+struct ProcedureStateReceiverProvider {
+    receivers: HashMap<ProcedureId, watch::Receiver<ProcedureState>>,
+    inner: MockContextProvider,
+}
+
+#[async_trait::async_trait]
+impl ContextProvider for ProcedureStateReceiverProvider {
+    async fn procedure_state(
+        &self,
+        procedure_id: ProcedureId,
+    ) -> common_procedure::Result<Option<ProcedureState>> {
+        self.inner.procedure_state(procedure_id).await
+    }
+
+    async fn procedure_state_receiver(
+        &self,
+        procedure_id: ProcedureId,
+    ) -> common_procedure::Result<Option<watch::Receiver<ProcedureState>>> {
+        Ok(self.receivers.get(&procedure_id).cloned())
+    }
+
+    async fn try_put_poison(
+        &self,
+        key: &common_procedure::PoisonKey,
+        procedure_id: ProcedureId,
+    ) -> common_procedure::Result<()> {
+        self.inner.try_put_poison(key, procedure_id).await
+    }
+
+    async fn acquire_lock(
+        &self,
+        key: &common_procedure::StringKey,
+    ) -> common_procedure::local::DynamicKeyLockGuard {
+        self.inner.acquire_lock(key).await
+    }
+}
diff --git a/src/meta-srv/src/procedure/repartition/utils.rs b/src/meta-srv/src/procedure/repartition/utils.rs
index 1bda85699b..6f274e9596 100644
--- a/src/meta-srv/src/procedure/repartition/utils.rs
+++ b/src/meta-srv/src/procedure/repartition/utils.rs
@@ -22,6 +22,8 @@ use snafu::{OptionExt, ResultExt, ensure};
 use store_api::storage::{RegionId, RegionNumber, TableId};
 
 use crate::error::{self, Result};
+use crate::procedure::repartition::group::GroupId;
+use crate::procedure::repartition::plan::RegionDescriptor;
 
 /// Returns the `datanode_table_value`
 ///
@@ -118,14 +120,79 @@ pub fn merge_and_validate_region_wal_options(
     Ok(new_region_wal_options)
 }
 
+/// Restores group staging metadata in-place for parent repartition rollback.
+///
+/// This helper lives in repartition utilities instead of the group subprocedure
+/// because parent repartition owns crash recovery and rollback selection.
+///
+/// The function mutates `region_routes` in place to avoid rebuilding the route
+/// vector for each selected plan. It restores:
+/// - source-region leader staging flags,
+/// - merge-source `ignore_all_writes` markers for pending-deallocate sources,
+/// - target-region partition expressions,
+/// - target-region write-route policies,
+/// - target-region leader staging flags.
+///
+/// `original_target_routes` contains only pre-existing target routes.
+/// Newly allocated targets are removed by parent rollback instead of being
+/// restored here.
+pub fn rollback_group_metadata_routes(
+    group_id: GroupId,
+    source_regions: &[RegionDescriptor],
+    original_target_routes: &[RegionRoute],
+    allocated_region_ids: &[RegionId],
+    pending_deallocate_region_ids: &[RegionId],
+    region_routes_map: &mut HashMap<RegionId, &mut RegionRoute>,
+) -> Result<()> {
+    for source in source_regions {
+        let region_route = region_routes_map.get_mut(&source.region_id).context(
+            error::RepartitionSourceRegionMissingSnafu {
+                group_id,
+                region_id: source.region_id,
+            },
+        )?;
+        region_route.clear_leader_staging();
+        if pending_deallocate_region_ids.contains(&source.region_id) {
+            region_route.clear_ignore_all_writes();
+        }
+    }
+
+    for target in original_target_routes {
+        let Some(region_route) = region_routes_map.get_mut(&target.region.id) else {
+            // Ignores newly allocated region routes that do not exist in the current region routes.
+            // They may have already been deleted (to ensure idempotency).
+            if allocated_region_ids.contains(&target.region.id) {
+                continue;
+            }
+
+            return error::RepartitionTargetRegionMissingSnafu {
+                group_id,
+                region_id: target.region.id,
+            }
+            .fail();
+        };
+        region_route.region.partition_expr = target.region.partition_expr.clone();
+        region_route.write_route_policy = target.write_route_policy;
+        region_route.clear_leader_staging();
+    }
+
+    Ok(())
+}
+
 #[cfg(test)]
 mod tests {
+    use std::collections::HashSet;
+
     use common_meta::peer::Peer;
-    use common_meta::rpc::router::{Region, RegionRoute};
+    use common_meta::rpc::router::{LeaderState, Region, RegionRoute};
     use common_wal::options::{KafkaWalOptions, WalOptions};
     use store_api::storage::RegionId;
+    use uuid::Uuid;
 
     use super::*;
+    use crate::procedure::repartition::group::update_metadata::UpdateMetadata;
+    use crate::procedure::repartition::plan::RegionDescriptor;
+    use crate::procedure::repartition::test_util::range_expr;
 
     /// Helper function to create a Kafka WAL option string from a topic name.
     fn kafka_wal_option(topic: &str) -> String {
@@ -149,6 +216,45 @@ mod tests {
         }
     }
 
+    fn new_staged_region_route(
+        region_id: RegionId,
+        partition_expr: &str,
+        leader_state: Option<LeaderState>,
+        ignore_all_writes: bool,
+    ) -> RegionRoute {
+        let mut route = RegionRoute {
+            region: Region {
+                id: region_id,
+                partition_expr: partition_expr.to_string(),
+                ..Default::default()
+            },
+            leader_peer: Some(Peer::empty(1)),
+            leader_state,
+            ..Default::default()
+        };
+
+        if ignore_all_writes {
+            route.set_ignore_all_writes();
+        }
+
+        route
+    }
+
+    fn original_target_routes(
+        region_routes: &[RegionRoute],
+        targets: &[RegionDescriptor],
+    ) -> Vec<RegionRoute> {
+        let target_ids = targets
+            .iter()
+            .map(|target| target.region_id)
+            .collect::<HashSet<_>>();
+        region_routes
+            .iter()
+            .filter(|route| target_ids.contains(&route.region.id))
+            .cloned()
+            .collect()
+    }
+
     #[test]
     fn test_merge_and_validate_region_wal_options_success() {
         let table_id = 1;
@@ -254,4 +360,141 @@ mod tests {
         assert!(error_msg.contains("Mismatch"));
         assert!(error_msg.contains(&table_id.to_string()));
     }
+
+    #[test]
+    fn test_rollback_group_metadata_routes_split_case() {
+        let group_id = Uuid::new_v4();
+        let table_id = 1024;
+        let original_region_routes = vec![
+            new_staged_region_route(
+                RegionId::new(table_id, 1),
+                &range_expr("x", 0, 100).as_json_str().unwrap(),
+                None,
+                false,
+            ),
+            new_staged_region_route(
+                RegionId::new(table_id, 2),
+                &range_expr("x", 100, 200).as_json_str().unwrap(),
+                None,
+                false,
+            ),
+            new_staged_region_route(RegionId::new(table_id, 3), "", None, false),
+        ];
+        let sources = vec![RegionDescriptor {
+            region_id: RegionId::new(table_id, 1),
+            partition_expr: range_expr("x", 0, 100),
+        }];
+        let targets = vec![
+            RegionDescriptor {
+                region_id: RegionId::new(table_id, 1),
+                partition_expr: range_expr("x", 0, 50),
+            },
+            RegionDescriptor {
+                region_id: RegionId::new(table_id, 3),
+                partition_expr: range_expr("x", 50, 100),
+            },
+        ];
+        let mut applied_region_routes = UpdateMetadata::apply_staging_region_routes(
+            group_id,
+            &sources,
+            &targets,
+            &[],
+            &original_region_routes,
+        )
+        .unwrap();
+        let target_routes = original_target_routes(&original_region_routes, &targets);
+
+        rollback_group_metadata_routes(
+            group_id,
+            &sources,
+            &target_routes,
+            &[],
+            &[],
+            &mut applied_region_routes
+                .iter_mut()
+                .map(|route| (route.region.id, route))
+                .collect(),
+        )
+        .unwrap();
+
+        assert_eq!(applied_region_routes, original_region_routes);
+    }
+
+    #[test]
+    fn test_rollback_group_metadata_routes_merge_case_is_idempotent() {
+        let group_id = Uuid::new_v4();
+        let table_id = 1024;
+        let original_region_routes = vec![
+            new_staged_region_route(
+                RegionId::new(table_id, 1),
+                &range_expr("x", 0, 100).as_json_str().unwrap(),
+                None,
+                false,
+            ),
+            new_staged_region_route(
+                RegionId::new(table_id, 2),
+                &range_expr("x", 100, 200).as_json_str().unwrap(),
+                None,
+                false,
+            ),
+            new_staged_region_route(
+                RegionId::new(table_id, 3),
+                &range_expr("x", 200, 300).as_json_str().unwrap(),
+                None,
+                false,
+            ),
+        ];
+        let sources = vec![
+            RegionDescriptor {
+                region_id: RegionId::new(table_id, 1),
+                partition_expr: range_expr("x", 0, 100),
+            },
+            RegionDescriptor {
+                region_id: RegionId::new(table_id, 2),
+                partition_expr: range_expr("x", 100, 200),
+            },
+        ];
+        let targets = vec![RegionDescriptor {
+            region_id: RegionId::new(table_id, 1),
+            partition_expr: range_expr("x", 0, 200),
+        }];
+        let target_routes = original_target_routes(&original_region_routes, &targets);
+        let mut once = UpdateMetadata::apply_staging_region_routes(
+            group_id,
+            &sources,
+            &targets,
+            &[RegionId::new(table_id, 2)],
+            &original_region_routes,
+        )
+        .unwrap();
+
+        rollback_group_metadata_routes(
+            group_id,
+            &sources,
+            &target_routes,
+            &[],
+            &[RegionId::new(table_id, 2)],
+            &mut once
+                .iter_mut()
+                .map(|route| (route.region.id, route))
+                .collect(),
+        )
+        .unwrap();
+        let mut twice = once.clone();
+        rollback_group_metadata_routes(
+            group_id,
+            &sources,
+            &target_routes,
+            &[],
+            &[RegionId::new(table_id, 2)],
+            &mut twice
+                .iter_mut()
+                .map(|route| (route.region.id, route))
+                .collect(),
+        )
+        .unwrap();
+
+        assert_eq!(once, original_region_routes);
+        assert_eq!(once, twice);
+    }
 }
diff --git a/src/meta-srv/src/procedure/tests.rs b/src/meta-srv/src/procedure/tests.rs
index 105bfb7bc4..93879d3c08 100644
--- a/src/meta-srv/src/procedure/tests.rs
+++ b/src/meta-srv/src/procedure/tests.rs
@@ -219,7 +219,7 @@ async fn test_on_datanode_create_regions() {
         }
     });
 
-    let status = procedure.on_datanode_create_regions().await.unwrap();
+    let status = procedure.on_datanode_create_regions(false).await.unwrap();
     assert!(matches!(
         status,
         Status::Executing {
diff --git a/src/meta-srv/src/procedure/utils.rs b/src/meta-srv/src/procedure/utils.rs
index bea2195573..0ac48cfc4f 100644
--- a/src/meta-srv/src/procedure/utils.rs
+++ b/src/meta-srv/src/procedure/utils.rs
@@ -20,6 +20,7 @@ use common_meta::peer::Peer;
 use common_telemetry::tracing_context::TracingContext;
 use common_telemetry::{info, warn};
 use snafu::ResultExt;
+use store_api::region_request::RegionFlushReason;
 use store_api::storage::RegionId;
 use tokio::time::Instant;
 
@@ -99,11 +100,12 @@ pub(crate) async fn flush_region(
     datanode: &Peer,
     timeout: Duration,
     error_strategy: ErrorStrategy,
+    reason: Option<RegionFlushReason>,
 ) -> Result<()> {
-    let flush_instruction = Instruction::FlushRegions(FlushRegions::sync_batch(
-        region_ids.to_vec(),
-        FlushErrorStrategy::TryAll,
-    ));
+    let mut flush_regions =
+        FlushRegions::sync_batch(region_ids.to_vec(), FlushErrorStrategy::TryAll);
+    flush_regions.reason = reason;
+    let flush_instruction = Instruction::FlushRegions(flush_regions);
 
     let tracing_ctx = TracingContext::from_current_span();
     let msg = MailboxMessage::json_message(
@@ -190,6 +192,23 @@ pub(crate) async fn flush_region(
             operation: "Flush regions",
         }
         .fail(),
+        Err(error::Error::MailboxChannelClosed { .. }) => match error_strategy {
+            ErrorStrategy::Ignore => {
+                warn!(
+                    "Failed to flush regions({:?}), the datanode({}) is unreachable(MailboxChannelClosed). Skip flush operation.",
+                    region_ids, datanode
+                );
+                Ok(())
+            }
+            ErrorStrategy::Retry => error::RetryLaterSnafu {
+                reason: format!(
+                    "Mailbox closed when sending flush region to datanode {:?}, elapsed: {:?}",
+                    datanode,
+                    now.elapsed()
+                ),
+            }
+            .fail()?,
+        },
         Err(err) => Err(err),
     }
 }
@@ -291,6 +310,70 @@ pub mod mock {
     }
 }
 
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use common_meta::instruction::{FlushStrategy, Instruction};
+    use common_meta::kv_backend::memory::MemoryKvBackend;
+    use common_meta::sequence::SequenceBuilder;
+
+    use super::*;
+    use crate::procedure::test_util::{MailboxContext, new_flush_region_reply_for_region};
+
+    #[tokio::test]
+    async fn test_flush_region_payload_includes_reason() {
+        let kv_backend = Arc::new(MemoryKvBackend::new());
+        let mailbox_sequence = SequenceBuilder::new("test_flush_region_reason", kv_backend).build();
+        let mut mailbox_ctx = MailboxContext::new(mailbox_sequence);
+
+        let datanode = Peer::new(1, "127.0.0.1:4001");
+        let region_id = RegionId::new(1024, 1);
+        let (tx, mut rx) = tokio::sync::mpsc::channel(1);
+        mailbox_ctx
+            .insert_heartbeat_response_receiver(Channel::Datanode(datanode.id), tx)
+            .await;
+
+        let mailbox = mailbox_ctx.mailbox().clone();
+        let reply_mailbox = mailbox.clone();
+        let reply_task = tokio::spawn(async move {
+            let response = rx.recv().await.unwrap().unwrap();
+            let msg = response.mailbox_message.unwrap();
+            let instruction = HeartbeatMailbox::json_instruction(&msg).unwrap();
+            let Instruction::FlushRegions(flush_regions) = instruction else {
+                panic!("Expected FlushRegions instruction");
+            };
+
+            assert_eq!(flush_regions.region_ids, vec![region_id]);
+            assert_eq!(flush_regions.strategy, FlushStrategy::Sync);
+            assert_eq!(flush_regions.reason, Some(RegionFlushReason::Repartition));
+
+            reply_mailbox
+                .on_recv(
+                    msg.id,
+                    Ok(new_flush_region_reply_for_region(
+                        msg.id, region_id, true, None,
+                    )),
+                )
+                .await
+                .unwrap();
+        });
+
+        flush_region(
+            &mailbox,
+            "127.0.0.1:3002",
+            &[region_id],
+            &datanode,
+            Duration::from_secs(5),
+            ErrorStrategy::Retry,
+            Some(RegionFlushReason::Repartition),
+        )
+        .await
+        .unwrap();
+        reply_task.await.unwrap();
+    }
+}
+
 #[cfg(test)]
 pub mod test_data {
     use std::sync::Arc;
diff --git a/src/meta-srv/src/procedure/wal_prune.rs b/src/meta-srv/src/procedure/wal_prune.rs
index 0897441647..f5e74ef543 100644
--- a/src/meta-srv/src/procedure/wal_prune.rs
+++ b/src/meta-srv/src/procedure/wal_prune.rs
@@ -193,7 +193,7 @@ impl Procedure for WalPruneProcedure {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_wal::maybe_skip_kafka_integration_test;
     use common_wal::test_util::get_kafka_endpoints;
diff --git a/src/meta-srv/src/procedure/wal_prune/manager.rs b/src/meta-srv/src/procedure/wal_prune/manager.rs
index e5f45d229a..1f1ab2b2f0 100644
--- a/src/meta-srv/src/procedure/wal_prune/manager.rs
+++ b/src/meta-srv/src/procedure/wal_prune/manager.rs
@@ -287,7 +287,7 @@ impl WalPruneManager {
 
 #[cfg(test)]
 mod test {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::time::Duration;
 
     use common_meta::key::topic_name::TopicNameKey;
diff --git a/src/meta-srv/src/region/flush_trigger.rs b/src/meta-srv/src/region/flush_trigger.rs
index 4000d6dc13..cdaa4339f4 100644
--- a/src/meta-srv/src/region/flush_trigger.rs
+++ b/src/meta-srv/src/region/flush_trigger.rs
@@ -31,6 +31,7 @@ use common_wal::config::kafka::common::{
     DEFAULT_CHECKPOINT_TRIGGER_SIZE, DEFAULT_FLUSH_TRIGGER_SIZE,
 };
 use snafu::{OptionExt, ResultExt};
+use store_api::region_request::RegionFlushReason;
 use store_api::storage::RegionId;
 use tokio::sync::mpsc::{Receiver, Sender};
 
@@ -358,8 +359,10 @@ impl RegionFlushTrigger {
         let flush_instructions = leader_to_region_ids
             .into_iter()
             .map(|(leader, region_ids)| {
-                let flush_instruction =
-                    Instruction::FlushRegions(FlushRegions::async_batch(region_ids));
+                let flush_instruction = Instruction::FlushRegions(
+                    FlushRegions::async_batch(region_ids)
+                        .with_reason(RegionFlushReason::RemoteWalPrune),
+                );
                 (leader, flush_instruction)
             });
 
@@ -502,11 +505,20 @@ fn is_recent(timestamp: i64, now: i64, duration: Duration) -> bool {
 
 #[cfg(test)]
 mod tests {
+    use std::sync::Arc;
+
     use common_base::readable_size::ReadableSize;
-    use common_meta::region_registry::LeaderRegionManifestInfo;
+    use common_meta::instruction::FlushStrategy;
+    use common_meta::key::TableMetadataManager;
+    use common_meta::kv_backend::memory::MemoryKvBackend;
+    use common_meta::region_registry::{LeaderRegionManifestInfo, LeaderRegionRegistry};
+    use common_meta::sequence::SequenceBuilder;
+    use common_meta::stats::topic::TopicStatsRegistry;
     use store_api::storage::RegionId;
 
     use super::*;
+    use crate::handler::HeartbeatMailbox;
+    use crate::procedure::test_util::{MailboxContext, new_wal_prune_metadata};
 
     #[test]
     fn test_is_recent() {
@@ -722,4 +734,58 @@ mod tests {
             should_persist_region_checkpoint(&current, Some(ReplayCheckpoint::new(90, Some(10))));
         assert!(result.is_none());
     }
+
+    #[tokio::test]
+    async fn test_send_flush_instructions_payload_includes_remote_wal_prune_reason() {
+        let kv_backend = Arc::new(MemoryKvBackend::new());
+        let table_metadata_manager = Arc::new(TableMetadataManager::new(kv_backend.clone()));
+        let leader_region_registry = Arc::new(LeaderRegionRegistry::new());
+        let topic_stats_registry = Arc::new(TopicStatsRegistry::default());
+        let mailbox_sequence =
+            SequenceBuilder::new("test_remote_wal_prune_flush_reason", kv_backend).build();
+        let mut mailbox_ctx = MailboxContext::new(mailbox_sequence);
+
+        let (tx, mut rx) = tokio::sync::mpsc::channel(1);
+        mailbox_ctx
+            .insert_heartbeat_response_receiver(Channel::Datanode(1), tx)
+            .await;
+
+        let (trigger, _ticker) = RegionFlushTrigger::new(
+            table_metadata_manager.clone(),
+            leader_region_registry.clone(),
+            topic_stats_registry,
+            mailbox_ctx.mailbox().clone(),
+            "127.0.0.1:3002".to_string(),
+            ReadableSize(1),
+            ReadableSize(1),
+        );
+
+        let topic = "test_topic".to_string();
+        new_wal_prune_metadata(
+            table_metadata_manager,
+            leader_region_registry,
+            1,
+            1,
+            &[0],
+            topic,
+        )
+        .await;
+
+        let region_id = RegionId::new(0, 0);
+        trigger.send_flush_instructions(&[region_id]).await.unwrap();
+
+        let response = rx.recv().await.unwrap().unwrap();
+        let msg = response.mailbox_message.unwrap();
+        let instruction = HeartbeatMailbox::json_instruction(&msg).unwrap();
+        let Instruction::FlushRegions(flush_regions) = instruction else {
+            panic!("Expected FlushRegions instruction");
+        };
+
+        assert_eq!(flush_regions.region_ids, vec![region_id]);
+        assert_eq!(flush_regions.strategy, FlushStrategy::Async);
+        assert_eq!(
+            flush_regions.reason,
+            Some(RegionFlushReason::RemoteWalPrune)
+        );
+    }
 }
diff --git a/src/meta-srv/src/region/lease_keeper.rs b/src/meta-srv/src/region/lease_keeper.rs
index 6d282fb49f..d4a899d002 100644
--- a/src/meta-srv/src/region/lease_keeper.rs
+++ b/src/meta-srv/src/region/lease_keeper.rs
@@ -20,7 +20,7 @@ use common_meta::key::TableMetadataManagerRef;
 use common_meta::key::table_route::TableRouteValue;
 use common_meta::region_keeper::MemoryRegionKeeperRef;
 use common_meta::rpc::router::RegionRoute;
-use common_telemetry::warn;
+use common_telemetry::{info, warn};
 use snafu::ResultExt;
 use store_api::region_engine::RegionRole;
 use store_api::storage::{RegionId, TableId};
@@ -63,13 +63,9 @@ fn renew_region_lease_via_region_route(
     if let Some(leader) = &region_route.leader_peer
         && leader.id == datanode_id
     {
-        let region_role = if region_route.is_leader_downgrading() {
-            RegionRole::DowngradingLeader
-        } else {
-            RegionRole::Leader
-        };
-
-        return Some((region_id, region_role));
+        return region_route
+            .leader_region_role()
+            .map(|region_role| (region_id, region_role));
     }
 
     // If it's a follower region on this datanode.
@@ -81,11 +77,28 @@ fn renew_region_lease_via_region_route(
         return Some((region_id, RegionRole::Follower));
     }
 
-    warn!(
-        "Denied to renew region lease for datanode: {datanode_id}, region_id: {region_id}, region_routes: {:?}",
-        region_route
-    );
-    // The region doesn't belong to this datanode.
+    None
+}
+
+fn renew_region_lease_via_operating_regions(
+    operating_regions: &HashMap<RegionId, RegionRole>,
+    datanode_id: DatanodeId,
+    region_id: RegionId,
+    reported_role: RegionRole,
+) -> Option<RegionLeaseInfo> {
+    // `operating_regions` is filtered by the current datanode in `collect_metadata`,
+    // so looking up by `region_id` is sufficient here.
+    if let Some(role) = operating_regions.get(&region_id) {
+        let region_lease_info = RegionLeaseInfo::operating(region_id, *role);
+        if *role != reported_role {
+            info!(
+                "The region {} on datanode {} is operating with role {:?}, but reported as {:?}",
+                region_id, datanode_id, role, reported_role
+            );
+        }
+        return Some(region_lease_info);
+    }
+
     None
 }
 
@@ -147,49 +160,51 @@ impl RegionLeaseKeeper {
     }
 
     /// Returns [None] if:
-    /// - The region doesn't belong to the datanode.
+    /// - The region doesn't belong to the datanode in metadata or operating regions.
     /// - The region belongs to a logical table.
     fn renew_region_lease(
         &self,
         table_metadata: &HashMap<TableId, TableRouteValue>,
-        operating_regions: &HashSet<RegionId>,
+        operating_regions: &HashMap<RegionId, RegionRole>,
         datanode_id: DatanodeId,
         region_id: RegionId,
-        role: RegionRole,
+        reported_role: RegionRole,
     ) -> Option<RegionLeaseInfo> {
-        if operating_regions.contains(&region_id) {
-            let region_lease_info = RegionLeaseInfo::operating(region_id, role);
+        // First try to renew via region route
+        if let Some(table_route) = table_metadata.get(&region_id.table_id())
+            && let Ok(Some(region_route)) = table_route.region_route(region_id)
+            && let Some(region_lease) =
+                renew_region_lease_via_region_route(&region_route, datanode_id, region_id)
+        {
+            return Some(RegionLeaseInfo::from(region_lease));
+        }
+        // Then try to renew via operating regions, which covers the opening region without region route in metadata.
+        if let Some(region_lease_info) = renew_region_lease_via_operating_regions(
+            operating_regions,
+            datanode_id,
+            region_id,
+            reported_role,
+        ) {
             return Some(region_lease_info);
         }
 
-        if let Some(table_route) = table_metadata.get(&region_id.table_id()) {
-            if let Ok(Some(region_route)) = table_route.region_route(region_id) {
-                return renew_region_lease_via_region_route(&region_route, datanode_id, region_id)
-                    .map(RegionLeaseInfo::from);
-            } else {
-                warn!(
-                    "Denied to renew region lease for datanode: {datanode_id}, region_id: {region_id}, region route is not found in table({})",
-                    region_id.table_id()
-                );
-            }
-        } else {
-            warn!(
-                "Denied to renew region lease for datanode: {datanode_id}, region_id: {region_id}, table({}) is not found",
-                region_id.table_id()
-            );
-        }
+        warn!(
+            "Denied to renew region lease for datanode: {datanode_id}, region_id: {region_id}, no matching metadata or operating region found",
+        );
         None
     }
 
     async fn collect_metadata(
         &self,
         datanode_id: DatanodeId,
-        mut region_ids: HashSet<RegionId>,
-    ) -> Result<(HashMap<TableId, TableRouteValue>, HashSet<RegionId>)> {
-        // Filters out operating region first, improves the cache hit rate(reduce expensive remote fetches).
+        region_ids: HashSet<RegionId>,
+    ) -> Result<(
+        HashMap<TableId, TableRouteValue>,
+        HashMap<RegionId, RegionRole>,
+    )> {
         let operating_regions = self
             .memory_region_keeper
-            .extract_operating_regions(datanode_id, &mut region_ids);
+            .extract_operating_region_roles(datanode_id, &region_ids);
         let table_ids = region_ids
             .into_iter()
             .map(|region_id| region_id.table_id())
@@ -222,13 +237,13 @@ impl RegionLeaseKeeper {
         let mut renewed = HashMap::new();
         let mut non_exists = HashSet::new();
 
-        for &(region, role) in regions {
+        for &(region, reported_role) in regions {
             match self.renew_region_lease(
                 &table_metadata,
                 &operating_regions,
                 datanode_id,
                 region,
-                role,
+                reported_role,
             ) {
                 Some(region_lease_info) => {
                     renewed.insert(region_lease_info.region_id, region_lease_info);
@@ -313,6 +328,12 @@ mod tests {
             renew_region_lease_via_region_route(&region_route, leader_peer_id, region_id),
             Some((region_id, RegionRole::DowngradingLeader))
         );
+
+        region_route.leader_state = Some(LeaderState::Staging);
+        assert_eq!(
+            renew_region_lease_via_region_route(&region_route, leader_peer_id, region_id),
+            Some((region_id, RegionRole::StagingLeader))
+        );
     }
 
     #[tokio::test]
@@ -368,12 +389,16 @@ mod tests {
         let opening_region_id = RegionId::new(1025, 1);
         let _guard = keeper
             .memory_region_keeper
-            .register(leader_peer_id, opening_region_id)
+            .register_with_role(leader_peer_id, opening_region_id, RegionRole::Leader)
             .unwrap();
         let another_opening_region_id = RegionId::new(1025, 2);
         let _guard2 = keeper
             .memory_region_keeper
-            .register(follower_peer_id, another_opening_region_id)
+            .register_with_role(
+                follower_peer_id,
+                another_opening_region_id,
+                RegionRole::Follower,
+            )
             .unwrap();
 
         let (metadata, regions) = keeper
@@ -387,8 +412,10 @@ mod tests {
             metadata.keys().cloned().collect::<Vec<_>>(),
             vec![region_id.table_id()]
         );
-        assert!(regions.contains(&opening_region_id));
-        assert_eq!(regions.len(), 1);
+        assert_eq!(
+            regions,
+            HashMap::from([(opening_region_id, RegionRole::Leader)])
+        );
     }
 
     #[tokio::test]
@@ -473,17 +500,17 @@ mod tests {
         let opening_region_id = RegionId::new(2048, 1);
         let _guard = keeper
             .memory_region_keeper
-            .register(leader_peer_id, opening_region_id)
+            .register_with_role(leader_peer_id, opening_region_id, RegionRole::Leader)
             .unwrap();
 
         // The opening region on the datanode.
         // NOTES: The procedure lock will ensure only one opening leader.
-        for role in [RegionRole::Leader, RegionRole::Follower] {
+        for reported_role in [RegionRole::Leader, RegionRole::Follower] {
             let RenewRegionLeasesResponse {
                 non_exists,
                 renewed,
             } = keeper
-                .renew_region_leases(leader_peer_id, &[(opening_region_id, role)])
+                .renew_region_leases(leader_peer_id, &[(opening_region_id, reported_role)])
                 .await
                 .unwrap();
 
@@ -492,7 +519,7 @@ mod tests {
                 renewed,
                 HashMap::from([(
                     opening_region_id,
-                    RegionLeaseInfo::operating(opening_region_id, role)
+                    RegionLeaseInfo::operating(opening_region_id, RegionRole::Leader)
                 )])
             );
         }
@@ -581,4 +608,213 @@ mod tests {
             );
         }
     }
+
+    #[tokio::test]
+    async fn test_renew_region_leases_reported_staging_expected_leader() {
+        let table_id = 1024;
+        let table_info: TableInfo = new_test_table_info(table_id);
+
+        let region_id = RegionId::new(table_id, 1);
+        let leader_peer_id = 1024;
+        let region_route = RegionRouteBuilder::default()
+            .region(Region::new_test(region_id))
+            .leader_peer(Peer::empty(leader_peer_id))
+            .build()
+            .unwrap();
+
+        let keeper = new_test_keeper();
+        let table_metadata_manager = keeper.table_metadata_manager();
+        table_metadata_manager
+            .create_table_metadata(
+                table_info,
+                TableRouteValue::physical(vec![region_route]),
+                HashMap::default(),
+            )
+            .await
+            .unwrap();
+
+        let RenewRegionLeasesResponse {
+            non_exists,
+            renewed,
+        } = keeper
+            .renew_region_leases(leader_peer_id, &[(region_id, RegionRole::StagingLeader)])
+            .await
+            .unwrap();
+
+        assert!(non_exists.is_empty());
+        assert_eq!(
+            renewed,
+            HashMap::from([(
+                region_id,
+                RegionLeaseInfo::from((region_id, RegionRole::Leader))
+            )])
+        );
+    }
+
+    #[tokio::test]
+    async fn test_renew_region_leases_reported_staging_expected_staging() {
+        let table_id = 1024;
+        let table_info: TableInfo = new_test_table_info(table_id);
+
+        let region_id = RegionId::new(table_id, 1);
+        let leader_peer_id = 1024;
+        let region_route = RegionRouteBuilder::default()
+            .region(Region::new_test(region_id))
+            .leader_peer(Peer::empty(leader_peer_id))
+            .leader_state(LeaderState::Staging)
+            .build()
+            .unwrap();
+
+        let keeper = new_test_keeper();
+        let table_metadata_manager = keeper.table_metadata_manager();
+        table_metadata_manager
+            .create_table_metadata(
+                table_info,
+                TableRouteValue::physical(vec![region_route]),
+                HashMap::default(),
+            )
+            .await
+            .unwrap();
+
+        let RenewRegionLeasesResponse {
+            non_exists,
+            renewed,
+        } = keeper
+            .renew_region_leases(leader_peer_id, &[(region_id, RegionRole::StagingLeader)])
+            .await
+            .unwrap();
+
+        assert!(non_exists.is_empty());
+        assert_eq!(
+            renewed,
+            HashMap::from([(
+                region_id,
+                RegionLeaseInfo::from((region_id, RegionRole::StagingLeader))
+            )])
+        );
+    }
+
+    #[tokio::test]
+    async fn test_renew_region_leases_metadata_role_beats_keeper_role() {
+        let table_id = 2048;
+        let table_info: TableInfo = new_test_table_info(table_id);
+
+        let datanode_id = 1024;
+        let region_id = RegionId::new(table_id, 1);
+        let region_route = RegionRouteBuilder::default()
+            .region(Region::new_test(region_id))
+            .leader_peer(Peer::empty(datanode_id))
+            .build()
+            .unwrap();
+
+        let keeper = new_test_keeper();
+        let table_metadata_manager = keeper.table_metadata_manager();
+        table_metadata_manager
+            .create_table_metadata(
+                table_info,
+                TableRouteValue::physical(vec![region_route]),
+                HashMap::default(),
+            )
+            .await
+            .unwrap();
+
+        let _guard = keeper
+            .memory_region_keeper
+            .register_with_role(datanode_id, region_id, RegionRole::Follower)
+            .unwrap();
+
+        let RenewRegionLeasesResponse {
+            non_exists,
+            renewed,
+        } = keeper
+            .renew_region_leases(datanode_id, &[(region_id, RegionRole::Follower)])
+            .await
+            .unwrap();
+
+        assert!(non_exists.is_empty());
+        assert_eq!(
+            renewed,
+            HashMap::from([(
+                region_id,
+                RegionLeaseInfo::from((region_id, RegionRole::Leader))
+            )])
+        );
+    }
+
+    #[tokio::test]
+    async fn test_renew_region_leases_missing_route_falls_back_to_keeper_role() {
+        let table_id = 2048;
+        let table_info: TableInfo = new_test_table_info(table_id);
+
+        let datanode_id = 1024;
+        let region_id = RegionId::new(table_id, 1);
+        let another_region_id = RegionId::new(table_id, 2);
+        let region_route = RegionRouteBuilder::default()
+            .region(Region::new_test(another_region_id))
+            .leader_peer(Peer::empty(datanode_id))
+            .build()
+            .unwrap();
+
+        let keeper = new_test_keeper();
+        let table_metadata_manager = keeper.table_metadata_manager();
+        table_metadata_manager
+            .create_table_metadata(
+                table_info,
+                TableRouteValue::physical(vec![region_route]),
+                HashMap::default(),
+            )
+            .await
+            .unwrap();
+
+        let _guard = keeper
+            .memory_region_keeper
+            .register_with_role(datanode_id, region_id, RegionRole::DowngradingLeader)
+            .unwrap();
+
+        let RenewRegionLeasesResponse {
+            non_exists,
+            renewed,
+        } = keeper
+            .renew_region_leases(datanode_id, &[(region_id, RegionRole::StagingLeader)])
+            .await
+            .unwrap();
+
+        assert!(non_exists.is_empty());
+        assert_eq!(
+            renewed,
+            HashMap::from([(
+                region_id,
+                RegionLeaseInfo::operating(region_id, RegionRole::DowngradingLeader)
+            )])
+        );
+    }
+
+    #[tokio::test]
+    async fn test_renew_region_leases_operating_region_uses_keeper_role() {
+        let keeper = new_test_keeper();
+        let datanode_id = 1024;
+        let region_id = RegionId::new(2048, 1);
+
+        let _guard = keeper
+            .memory_region_keeper
+            .register_with_role(datanode_id, region_id, RegionRole::DowngradingLeader)
+            .unwrap();
+
+        let RenewRegionLeasesResponse {
+            non_exists,
+            renewed,
+        } = keeper
+            .renew_region_leases(datanode_id, &[(region_id, RegionRole::StagingLeader)])
+            .await
+            .unwrap();
+
+        assert!(non_exists.is_empty());
+        assert_eq!(
+            renewed,
+            HashMap::from([(
+                region_id,
+                RegionLeaseInfo::operating(region_id, RegionRole::DowngradingLeader)
+            )])
+        );
+    }
 }
diff --git a/src/meta-srv/src/region/supervisor.rs b/src/meta-srv/src/region/supervisor.rs
index 90e08992f6..05716e60c8 100644
--- a/src/meta-srv/src/region/supervisor.rs
+++ b/src/meta-srv/src/region/supervisor.rs
@@ -847,7 +847,7 @@ impl RegionSupervisor {
 
 #[cfg(test)]
 pub(crate) mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
     use std::sync::{Arc, Mutex};
     use std::time::Duration;
diff --git a/src/meta-srv/src/service/admin/leader.rs b/src/meta-srv/src/service/admin/leader.rs
index 1fadb4a3ef..17329e7b47 100644
--- a/src/meta-srv/src/service/admin/leader.rs
+++ b/src/meta-srv/src/service/admin/leader.rs
@@ -32,7 +32,7 @@ pub struct LeaderHandler {
 impl LeaderHandler {
     async fn get_leader(&self) -> Result<Option<String>> {
         if let Some(election) = &self.election {
-            let leader_addr = election.leader().await?.0;
+            let leader_addr = election.leader().await.context(error::KvBackendSnafu)?.0;
             return Ok(Some(leader_addr));
         }
         Ok(None)
diff --git a/src/meta-srv/src/service/cluster.rs b/src/meta-srv/src/service/cluster.rs
index 5c0ae4c71f..366a8aa5fb 100644
--- a/src/meta-srv/src/service/cluster.rs
+++ b/src/meta-srv/src/service/cluster.rs
@@ -63,7 +63,10 @@ impl cluster_server::Cluster for Metasrv {
         let leader_addr = &self.options().grpc.server_addr;
         let (leader, followers) = match self.election() {
             Some(election) => {
-                let nodes = election.all_candidates().await?;
+                let nodes = election
+                    .all_candidates()
+                    .await
+                    .context(error::KvBackendSnafu)?;
                 let followers = nodes
                     .into_iter()
                     .filter(|node_info| &node_info.addr != leader_addr)
diff --git a/src/meta-srv/src/service/heartbeat.rs b/src/meta-srv/src/service/heartbeat.rs
index e09073546a..238ed99df2 100644
--- a/src/meta-srv/src/service/heartbeat.rs
+++ b/src/meta-srv/src/service/heartbeat.rs
@@ -23,7 +23,7 @@ use api::v1::meta::{
 use common_telemetry::{debug, error, info, warn};
 use futures::StreamExt;
 use once_cell::sync::OnceCell;
-use snafu::OptionExt;
+use snafu::{OptionExt, ResultExt};
 use tokio::sync::mpsc;
 use tokio::sync::mpsc::Sender;
 use tokio_stream::wrappers::ReceiverStream;
@@ -148,7 +148,7 @@ async fn handle_ask_leader(_req: AskLeaderRequest, ctx: Context) -> Result<AskLe
             if election.is_leader() {
                 ctx.server_addr
             } else {
-                election.leader().await?.0
+                election.leader().await.context(error::KvBackendSnafu)?.0
             }
         }
         None => ctx.server_addr,
diff --git a/src/meta-srv/src/service/mailbox.rs b/src/meta-srv/src/service/mailbox.rs
index 5e569e6d4d..8b37eeaad5 100644
--- a/src/meta-srv/src/service/mailbox.rs
+++ b/src/meta-srv/src/service/mailbox.rs
@@ -214,7 +214,7 @@ pub trait Mailbox: Send + Sync {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_time::util::current_time_millis;
     use tokio::sync::watch;
diff --git a/src/meta-srv/src/state.rs b/src/meta-srv/src/state.rs
index e5edc5f169..12eb708b5c 100644
--- a/src/meta-srv/src/state.rs
+++ b/src/meta-srv/src/state.rs
@@ -117,7 +117,7 @@ pub fn become_follower() -> impl FnOnce(&State) -> State {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use crate::state::{FollowerState, LeaderState, State, become_follower, become_leader};
 
diff --git a/src/meta-srv/src/utils.rs b/src/meta-srv/src/utils.rs
index c3b1e86da4..ddaffd2390 100644
--- a/src/meta-srv/src/utils.rs
+++ b/src/meta-srv/src/utils.rs
@@ -12,6 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+pub mod database;
 pub mod etcd;
 pub mod insert_forwarder;
 #[cfg(feature = "mysql_kvbackend")]
diff --git a/src/meta-srv/src/utils/database.rs b/src/meta-srv/src/utils/database.rs
new file mode 100644
index 0000000000..bf8cad535f
--- /dev/null
+++ b/src/meta-srv/src/utils/database.rs
@@ -0,0 +1,149 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::sync::Arc;
+
+use client::error::{ExternalSnafu, Result as ClientResult};
+use client::{Client, Database, Output};
+use common_error::ext::BoxedError;
+use common_meta::peer::PeerDiscoveryRef;
+use common_telemetry::{debug, warn};
+use snafu::{ResultExt, ensure};
+use tokio::sync::RwLock;
+
+use crate::error::{ListActiveFrontendsSnafu, NoAvailableFrontendSnafu, Result};
+
+pub type DatabaseOperatorRef = Arc<DatabaseOperator>;
+
+#[derive(Debug, Clone, Copy)]
+/// Database-level request context used by metasrv forwarding.
+pub struct DatabaseContext<'a> {
+    /// Catalog name carried in forwarded requests.
+    pub catalog: &'a str,
+    /// Schema name carried in forwarded requests.
+    pub schema: &'a str,
+}
+
+impl<'a> DatabaseContext<'a> {
+    /// Creates a new database context from catalog and schema.
+    pub fn new(catalog: &'a str, schema: &'a str) -> Self {
+        Self { catalog, schema }
+    }
+}
+
+/// A cached frontend database operator used by metasrv.
+pub struct DatabaseOperator {
+    peer_discovery: PeerDiscoveryRef,
+    client: RwLock<Option<Client>>,
+}
+
+impl DatabaseOperator {
+    /// Creates a database operator backed by discovered frontend peers.
+    pub fn new(peer_discovery: PeerDiscoveryRef) -> Self {
+        Self {
+            peer_discovery,
+            client: RwLock::new(None),
+        }
+    }
+
+    /// Forwards row inserts to an available frontend database client.
+    pub async fn insert(
+        &self,
+        ctx: &DatabaseContext<'_>,
+        requests: api::v1::RowInsertRequests,
+        hints: &[(&str, &str)],
+    ) -> ClientResult<u32> {
+        let client = self.maybe_init_client().await?;
+        let database = Database::new(ctx.catalog, ctx.schema, client);
+
+        let result = database.row_inserts_with_hints(requests, hints).await;
+
+        if should_reset_client(&result) {
+            self.reset_client().await;
+        }
+
+        result
+    }
+
+    /// Executes a serialized logical plan on an available frontend.
+    pub async fn logical_plan(
+        &self,
+        ctx: &DatabaseContext<'_>,
+        plan: Vec<u8>,
+    ) -> ClientResult<Output> {
+        let client = self.maybe_init_client().await?;
+        let database = Database::new(ctx.catalog, ctx.schema, client);
+
+        let result = database.logical_plan(plan).await;
+
+        if should_reset_client(&result) {
+            self.reset_client().await;
+        }
+
+        result
+    }
+
+    async fn build_client(&self) -> Result<Client> {
+        let frontends = self
+            .peer_discovery
+            .active_frontends()
+            .await
+            .context(ListActiveFrontendsSnafu)?;
+
+        ensure!(!frontends.is_empty(), NoAvailableFrontendSnafu);
+
+        let urls = frontends
+            .into_iter()
+            .map(|peer| peer.addr)
+            .collect::<Vec<_>>();
+
+        debug!("Available frontend addresses: {:?}", urls);
+
+        Ok(Client::with_urls(urls))
+    }
+
+    async fn maybe_init_client(&self) -> ClientResult<Client> {
+        if let Some(client) = self.client.read().await.as_ref() {
+            return Ok(client.clone());
+        }
+
+        let client = self
+            .build_client()
+            .await
+            .map_err(BoxedError::new)
+            .context(ExternalSnafu)?;
+
+        let mut guard = self.client.write().await;
+        if let Some(client) = guard.as_ref() {
+            return Ok(client.clone());
+        }
+
+        *guard = Some(client.clone());
+        Ok(client)
+    }
+
+    async fn reset_client(&self) {
+        warn!("Resetting the client");
+        let mut guard = self.client.write().await;
+        guard.take();
+    }
+}
+
+fn should_reset_client<T>(result: &client::error::Result<T>) -> bool {
+    result
+        .as_ref()
+        .err()
+        .map(|err| err.is_connection_error())
+        .unwrap_or(false)
+}
diff --git a/src/meta-srv/src/utils/insert_forwarder.rs b/src/meta-srv/src/utils/insert_forwarder.rs
index 78c7c454d4..b0b072dbda 100644
--- a/src/meta-srv/src/utils/insert_forwarder.rs
+++ b/src/meta-srv/src/utils/insert_forwarder.rs
@@ -15,75 +15,29 @@
 use std::sync::Arc;
 
 use api::v1::RowInsertRequests;
-use client::error::{ExternalSnafu, Result as ClientResult};
+use client::error::Result as ClientResult;
 use client::inserter::{Context, InsertOptions, Inserter};
-use client::{Client, Database};
-use common_error::ext::BoxedError;
-use common_meta::peer::PeerDiscoveryRef;
-use common_telemetry::{debug, warn};
-use snafu::{ResultExt, ensure};
-use tokio::sync::RwLock;
 
-use crate::error::{ListActiveFrontendsSnafu, NoAvailableFrontendSnafu};
+use crate::utils::database::{DatabaseContext, DatabaseOperatorRef};
 
 pub type InsertForwarderRef = Arc<InsertForwarder>;
 
 /// [`InsertForwarder`] is the inserter for the metasrv.
 /// It forwards insert requests to available frontend instances.
 pub struct InsertForwarder {
-    peer_discovery: PeerDiscoveryRef,
-    client: RwLock<Option<Client>>,
+    database_operator: DatabaseOperatorRef,
     options: Option<InsertOptions>,
 }
 
 impl InsertForwarder {
-    /// Creates a new InsertForwarder with the given peer lookup service.
-    pub fn new(peer_discovery: PeerDiscoveryRef, options: Option<InsertOptions>) -> Self {
+    /// Creates a new [`InsertForwarder`] with the given shared database operator
+    /// and optional insert options.
+    pub fn new(database_operator: DatabaseOperatorRef, options: Option<InsertOptions>) -> Self {
         Self {
-            peer_discovery,
-            client: RwLock::new(None),
+            database_operator,
             options,
         }
     }
-
-    /// Builds a new client.
-    async fn build_client(&self) -> crate::error::Result<Client> {
-        let frontends = self
-            .peer_discovery
-            .active_frontends()
-            .await
-            .context(ListActiveFrontendsSnafu)?;
-
-        ensure!(!frontends.is_empty(), NoAvailableFrontendSnafu);
-
-        let urls = frontends
-            .into_iter()
-            .map(|peer| peer.addr)
-            .collect::<Vec<_>>();
-
-        debug!("Available frontend addresses: {:?}", urls);
-
-        Ok(Client::with_urls(urls))
-    }
-
-    /// Initializes the client if it hasn't been initialized yet, or returns the cached client.
-    async fn maybe_init_client(&self) -> Result<Client, BoxedError> {
-        let mut guard = self.client.write().await;
-        if guard.is_none() {
-            let client = self.build_client().await.map_err(BoxedError::new)?;
-            *guard = Some(client);
-        }
-
-        // Safety: checked above that the client is Some.
-        Ok(guard.as_ref().unwrap().clone())
-    }
-
-    /// Resets the cached client, forcing a rebuild on the next use.
-    async fn reset_client(&self) {
-        warn!("Resetting the client");
-        let mut guard = self.client.write().await;
-        guard.take();
-    }
 }
 
 #[async_trait::async_trait]
@@ -93,26 +47,19 @@ impl Inserter for InsertForwarder {
         context: &Context<'_>,
         requests: RowInsertRequests,
     ) -> ClientResult<()> {
-        let client = self.maybe_init_client().await.context(ExternalSnafu)?;
-        let database = Database::new(context.catalog, context.schema, client);
+        let ctx = DatabaseContext::new(context.catalog, context.schema);
         let hints = self.options.as_ref().map_or(vec![], |o| o.to_hints());
 
-        if let Err(e) = database
-            .row_inserts_with_hints(
+        self.database_operator
+            .insert(
+                &ctx,
                 requests,
                 &hints
                     .iter()
                     .map(|(k, v)| (*k, v.as_str()))
                     .collect::<Vec<_>>(),
             )
-            .await
-            .map_err(BoxedError::new)
-            .context(ExternalSnafu)
-        {
-            // Resets the client so it will be rebuilt next time.
-            self.reset_client().await;
-            return Err(e);
-        };
+            .await?;
 
         Ok(())
     }
diff --git a/src/meta-srv/src/utils/mysql.rs b/src/meta-srv/src/utils/mysql.rs
index 1daa36a361..326f3880f4 100644
--- a/src/meta-srv/src/utils/mysql.rs
+++ b/src/meta-srv/src/utils/mysql.rs
@@ -12,6 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+use common_meta::election::ElectionRef;
+use common_meta::election::rds::mysql::{ElectionMysqlClient, MySqlElection};
+use common_meta::kv_backend::KvBackendRef;
+use common_meta::kv_backend::rds::MySqlStore;
 use common_telemetry::info;
 use servers::tls::{TlsMode, TlsOption};
 use snafu::{OptionExt, ResultExt};
@@ -83,3 +87,63 @@ pub async fn create_mysql_pool(
 
     Ok(pool)
 }
+
+/// Builds a MySQL-backed metadata [`KvBackendRef`].
+///
+/// * `store_addrs` - MySQL connection URLs; only the first address is used.
+/// * `tls_config` - optional TLS settings for the MySQL connection.
+/// * `table_name` - metadata KV table name.
+/// * `max_txn_ops` - maximum operations allowed in one metadata transaction.
+pub async fn build_mysql_kv_backend(
+    store_addrs: &[String],
+    tls_config: Option<&TlsOption>,
+    table_name: &str,
+    max_txn_ops: usize,
+) -> Result<KvBackendRef> {
+    let pool = create_mysql_pool(store_addrs, tls_config).await?;
+    MySqlStore::with_mysql_pool(pool, table_name, max_txn_ops)
+        .await
+        .context(error::KvBackendSnafu)
+}
+
+/// Builds a MySQL-backed election implementation.
+///
+/// * `store_addrs` - MySQL connection URLs; only the first address is used.
+/// * `tls_config` - optional TLS settings for the MySQL connection.
+/// * `leader_value` - advertised address of this election candidate.
+/// * `store_key_prefix` - prefix for election and candidate keys.
+/// * `candidate_lease_ttl` - TTL for registered candidate metadata.
+/// * `meta_lease_ttl` - TTL for the elected leader metadata.
+/// * `election_table_name` - dedicated table used for election locking and records.
+/// * `innodb_lock_wait_timeout` - session lock wait timeout for election transactions.
+#[allow(clippy::too_many_arguments)]
+pub async fn build_mysql_election(
+    store_addrs: &[String],
+    tls_config: Option<&TlsOption>,
+    leader_value: String,
+    store_key_prefix: String,
+    candidate_lease_ttl: std::time::Duration,
+    meta_lease_ttl: std::time::Duration,
+    election_table_name: &str,
+    innodb_lock_wait_timeout: std::time::Duration,
+) -> Result<ElectionRef> {
+    let pool = create_mysql_pool(store_addrs, tls_config).await?;
+    let election_client = ElectionMysqlClient::new(
+        pool,
+        meta_lease_ttl,
+        meta_lease_ttl,
+        innodb_lock_wait_timeout,
+        meta_lease_ttl,
+        election_table_name,
+    );
+    MySqlElection::with_mysql_client(
+        leader_value,
+        election_client,
+        store_key_prefix,
+        candidate_lease_ttl,
+        meta_lease_ttl,
+        election_table_name,
+    )
+    .await
+    .context(error::KvBackendSnafu)
+}
diff --git a/src/meta-srv/src/utils/postgres.rs b/src/meta-srv/src/utils/postgres.rs
index e7dfd59383..ec46c076e2 100644
--- a/src/meta-srv/src/utils/postgres.rs
+++ b/src/meta-srv/src/utils/postgres.rs
@@ -13,6 +13,10 @@
 // limitations under the License.
 
 use common_error::ext::BoxedError;
+use common_meta::election::ElectionRef;
+use common_meta::election::rds::postgres::{ElectionPgClient, PgElection};
+use common_meta::kv_backend::KvBackendRef;
+use common_meta::kv_backend::rds::PgStore;
 use common_meta::kv_backend::rds::postgres::{
     TlsMode as PgTlsMode, TlsOption as PgTlsOption, create_postgres_tls_connector,
 };
@@ -72,3 +76,77 @@ pub async fn create_postgres_pool(
 
     Ok(pool)
 }
+
+/// Builds a Postgres-backed metadata [`KvBackendRef`].
+///
+/// * `store_addrs` - Postgres connection URLs; only the first address is used.
+/// * `cfg` - optional deadpool config to customize pool/session behavior.
+/// * `tls_config` - optional TLS settings for the Postgres connection.
+/// * `schema_name` - optional schema containing the metadata table.
+/// * `table_name` - metadata KV table name.
+/// * `max_txn_ops` - maximum operations allowed in one metadata transaction.
+/// * `auto_create_schema` - whether to create `schema_name` when it is missing.
+#[allow(clippy::too_many_arguments)]
+pub async fn build_postgres_kv_backend(
+    store_addrs: &[String],
+    cfg: Option<Config>,
+    tls_config: Option<TlsOption>,
+    schema_name: Option<&str>,
+    table_name: &str,
+    max_txn_ops: usize,
+    auto_create_schema: bool,
+) -> Result<KvBackendRef> {
+    let pool = create_postgres_pool(store_addrs, cfg, tls_config).await?;
+    PgStore::with_pg_pool(
+        pool,
+        schema_name,
+        table_name,
+        max_txn_ops,
+        auto_create_schema,
+    )
+    .await
+    .context(error::KvBackendSnafu)
+}
+
+/// Builds a Postgres-backed election implementation.
+///
+/// * `store_addrs` - Postgres connection URLs; only the first address is used.
+/// * `cfg` - optional deadpool config to customize pool/session behavior.
+/// * `tls_config` - optional TLS settings for the Postgres connection.
+/// * `leader_value` - advertised address of this election candidate.
+/// * `store_key_prefix` - prefix for election and candidate keys.
+/// * `candidate_lease_ttl` - TTL for registered candidate metadata.
+/// * `meta_lease_ttl` - TTL for the elected leader metadata.
+/// * `schema_name` - optional schema containing the metadata table.
+/// * `table_name` - metadata KV table name used for election records.
+/// * `lock_id` - Postgres advisory lock id used by the election.
+#[allow(clippy::too_many_arguments)]
+pub async fn build_postgres_election(
+    store_addrs: &[String],
+    cfg: Option<Config>,
+    tls_config: Option<TlsOption>,
+    leader_value: String,
+    store_key_prefix: String,
+    candidate_lease_ttl: std::time::Duration,
+    meta_lease_ttl: std::time::Duration,
+    schema_name: Option<&str>,
+    table_name: &str,
+    lock_id: u64,
+) -> Result<ElectionRef> {
+    let pool = create_postgres_pool(store_addrs, cfg, tls_config).await?;
+    let election_client =
+        ElectionPgClient::new(pool, meta_lease_ttl, meta_lease_ttl, meta_lease_ttl)
+            .context(error::KvBackendSnafu)?;
+    PgElection::with_pg_client(
+        leader_value,
+        election_client,
+        store_key_prefix,
+        candidate_lease_ttl,
+        meta_lease_ttl,
+        schema_name,
+        table_name,
+        lock_id,
+    )
+    .await
+    .context(error::KvBackendSnafu)
+}
diff --git a/src/metric-engine/Cargo.toml b/src/metric-engine/Cargo.toml
index 567210b952..0005a4dc54 100644
--- a/src/metric-engine/Cargo.toml
+++ b/src/metric-engine/Cargo.toml
@@ -17,6 +17,7 @@ bytes.workspace = true
 fxhash = "0.2"
 common-base.workspace = true
 common-error.workspace = true
+common-grpc.workspace = true
 common-macro.workspace = true
 common-query.workspace = true
 common-recordbatch.workspace = true
@@ -32,7 +33,7 @@ itertools.workspace = true
 lazy_static = "1.4"
 mito-codec.workspace = true
 mito2.workspace = true
-moka.workspace = true
+moka = { workspace = true, features = ["future"] }
 object-store.workspace = true
 prometheus.workspace = true
 serde.workspace = true
diff --git a/src/metric-engine/src/batch_modifier.rs b/src/metric-engine/src/batch_modifier.rs
new file mode 100644
index 0000000000..2162215d1b
--- /dev/null
+++ b/src/metric-engine/src/batch_modifier.rs
@@ -0,0 +1,455 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::hash::Hasher;
+use std::sync::Arc;
+
+use datatypes::arrow::array::{Array, BinaryBuilder, StringArray, UInt64Array};
+use datatypes::arrow::datatypes::{DataType, Field, Schema as ArrowSchema};
+use datatypes::arrow::record_batch::RecordBatch;
+use fxhash::FxHasher;
+use mito_codec::row_converter::SparsePrimaryKeyCodec;
+use snafu::ResultExt;
+use store_api::storage::ColumnId;
+use store_api::storage::consts::PRIMARY_KEY_COLUMN_NAME;
+
+use crate::error::{EncodePrimaryKeySnafu, Result, UnexpectedRequestSnafu};
+
+/// Info about a tag column for TSID computation and sparse primary key encoding.
+#[allow(dead_code)]
+pub struct TagColumnInfo {
+    /// Column name (used for label-name hash).
+    pub name: String,
+    /// Column index in the RecordBatch.
+    pub index: usize,
+    /// Column ID in the physical region.
+    pub column_id: ColumnId,
+}
+
+/// Computes the TSID for each row in a [RecordBatch].
+///
+/// The TSID is a stable hash of the set of labels (tags) present in each row.
+/// It accounts for both the names and values of all non-null tag columns.
+///
+/// # Logic
+/// - If a row has no nulls across all `sorted_tag_columns`, it uses a precomputed hash of all label names.
+/// - If a row has nulls, it dynamically computes a hash of the names of labels that are present (non-null).
+/// - In both cases, it then hashes the values of those present labels in the order specified by `sorted_tag_columns`.
+pub fn compute_tsid_array(
+    batch: &RecordBatch,
+    sorted_tag_columns: &[TagColumnInfo],
+    tag_arrays: &[&StringArray],
+) -> UInt64Array {
+    let num_rows = batch.num_rows();
+
+    let label_name_hash = {
+        let mut hasher = FxHasher::default();
+        for tag_col in sorted_tag_columns {
+            hasher.write(tag_col.name.as_bytes());
+            hasher.write_u8(0xff);
+        }
+        hasher.finish()
+    };
+
+    let mut tsid_values = Vec::with_capacity(num_rows);
+    for row in 0..num_rows {
+        let has_null = tag_arrays.iter().any(|arr| arr.is_null(row));
+
+        let tsid = if !has_null {
+            let mut hasher = FxHasher::default();
+            hasher.write_u64(label_name_hash);
+            for arr in tag_arrays {
+                hasher.write(arr.value(row).as_bytes());
+                hasher.write_u8(0xff);
+            }
+            hasher.finish()
+        } else {
+            let mut name_hasher = FxHasher::default();
+            for (tc, arr) in sorted_tag_columns.iter().zip(tag_arrays.iter()) {
+                if !arr.is_null(row) {
+                    name_hasher.write(tc.name.as_bytes());
+                    name_hasher.write_u8(0xff);
+                }
+            }
+            let row_label_hash = name_hasher.finish();
+
+            let mut val_hasher = FxHasher::default();
+            val_hasher.write_u64(row_label_hash);
+            for arr in tag_arrays {
+                if !arr.is_null(row) {
+                    val_hasher.write(arr.value(row).as_bytes());
+                    val_hasher.write_u8(0xff);
+                }
+            }
+            val_hasher.finish()
+        };
+
+        tsid_values.push(tsid);
+    }
+
+    UInt64Array::from(tsid_values)
+}
+
+fn build_tag_arrays<'a>(
+    batch: &'a RecordBatch,
+    sorted_tag_columns: &[TagColumnInfo],
+) -> Vec<&'a StringArray> {
+    sorted_tag_columns
+        .iter()
+        .map(|tc| {
+            batch
+                .column(tc.index)
+                .as_any()
+                .downcast_ref::<StringArray>()
+                .expect("tag column must be utf8")
+        })
+        .collect()
+}
+
+/// Modifies a [RecordBatch] to include a sparse primary key column.
+///
+/// This function transforms the input `batch` into a new `RecordBatch` where the first column
+/// is the generated primary key (named [PRIMARY_KEY_COLUMN_NAME]), followed by columns
+/// indicated by `extra_column_indices`.
+///
+/// The primary key uses a "sparse" encoding, which compactly represents the row's identity
+/// by only including non-null tag values. The encoding, handled by [SparsePrimaryKeyCodec],
+/// consists of:
+/// 1. The `table_id`.
+/// 2. A `tsid` (Time Series ID), which is a hash of the present tags.
+/// 3. The actual non-null tag values paired with their `column_id`.
+///
+/// # Parameters
+/// - `batch`: The source [RecordBatch].
+/// - `table_id`: The ID of the table.
+/// - `sorted_tag_columns`: Metadata for tag columns, used for both TSID computation and PK encoding.
+/// - `extra_column_indices`: Indices of columns from the original batch to keep in the output
+///   (typically the timestamp and value fields).
+pub fn modify_batch_sparse(
+    batch: RecordBatch,
+    table_id: u32,
+    sorted_tag_columns: &[TagColumnInfo],
+    extra_column_indices: &[usize],
+) -> Result<RecordBatch> {
+    let num_rows = batch.num_rows();
+    let codec = SparsePrimaryKeyCodec::schemaless();
+    let tag_arrays: Vec<&StringArray> = build_tag_arrays(&batch, sorted_tag_columns);
+    let tsid_array = compute_tsid_array(&batch, sorted_tag_columns, &tag_arrays);
+
+    let mut pk_builder = BinaryBuilder::with_capacity(num_rows, 0);
+    let mut buffer = Vec::new();
+    for row in 0..num_rows {
+        buffer.clear();
+        codec
+            .encode_internal(table_id, tsid_array.value(row), &mut buffer)
+            .context(EncodePrimaryKeySnafu)?;
+
+        let tags = sorted_tag_columns
+            .iter()
+            .zip(tag_arrays.iter())
+            .filter(|(_, arr)| !arr.is_null(row))
+            .map(|(tc, arr)| (tc.column_id, arr.value(row).as_bytes()));
+        codec
+            .encode_raw_tag_value(tags, &mut buffer)
+            .context(EncodePrimaryKeySnafu)?;
+
+        pk_builder.append_value(&buffer);
+    }
+
+    let pk_array = pk_builder.finish();
+
+    let mut fields = vec![Arc::new(Field::new(
+        PRIMARY_KEY_COLUMN_NAME,
+        DataType::Binary,
+        false,
+    ))];
+    let mut columns: Vec<Arc<dyn Array>> = vec![Arc::new(pk_array)];
+
+    for &idx in extra_column_indices {
+        fields.push(batch.schema().fields()[idx].clone());
+        columns.push(batch.column(idx).clone());
+    }
+
+    let new_schema = Arc::new(ArrowSchema::new(fields));
+    RecordBatch::try_new(new_schema, columns).map_err(|e| {
+        UnexpectedRequestSnafu {
+            reason: format!("Failed to build modified sparse RecordBatch: {e}"),
+        }
+        .build()
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+    use std::sync::Arc;
+
+    use api::v1::value::ValueData;
+    use api::v1::{ColumnDataType, ColumnSchema, Row, Rows, SemanticType, Value};
+    use datatypes::arrow::array::{BinaryArray, Int64Array, StringArray};
+    use datatypes::arrow::datatypes::{DataType, Field, Schema as ArrowSchema};
+    use datatypes::arrow::record_batch::RecordBatch;
+    use store_api::codec::PrimaryKeyEncoding;
+    use store_api::metadata::ColumnMetadata;
+    use store_api::storage::consts::PRIMARY_KEY_COLUMN_NAME;
+
+    use super::*;
+    use crate::row_modifier::{RowModifier, RowsIter, TableIdInput};
+
+    fn build_sparse_test_batch() -> RecordBatch {
+        let schema = Arc::new(ArrowSchema::new(vec![
+            Field::new("greptime_timestamp", DataType::Int64, false),
+            Field::new("greptime_value", DataType::Float64, true),
+            Field::new("namespace", DataType::Utf8, true),
+            Field::new("host", DataType::Utf8, true),
+        ]));
+        RecordBatch::try_new(
+            schema,
+            vec![
+                Arc::new(Int64Array::from(vec![1000])),
+                Arc::new(datatypes::arrow::array::Float64Array::from(vec![42.0])),
+                Arc::new(StringArray::from(vec!["greptimedb"])),
+                Arc::new(StringArray::from(vec!["127.0.0.1"])),
+            ],
+        )
+        .unwrap()
+    }
+
+    fn sparse_tag_columns() -> Vec<TagColumnInfo> {
+        vec![
+            TagColumnInfo {
+                name: "host".to_string(),
+                index: 3,
+                column_id: 3,
+            },
+            TagColumnInfo {
+                name: "namespace".to_string(),
+                index: 2,
+                column_id: 2,
+            },
+        ]
+    }
+
+    #[test]
+    fn test_compute_tsid_basic() {
+        let schema = Arc::new(ArrowSchema::new(vec![
+            Field::new("namespace", DataType::Utf8, true),
+            Field::new("host", DataType::Utf8, true),
+        ]));
+        let batch = RecordBatch::try_new(
+            schema,
+            vec![
+                Arc::new(StringArray::from(vec!["greptimedb"])),
+                Arc::new(StringArray::from(vec!["127.0.0.1"])),
+            ],
+        )
+        .unwrap();
+
+        let tag_columns: Vec<TagColumnInfo> = vec![
+            TagColumnInfo {
+                name: "host".to_string(),
+                index: 1,
+                column_id: 2,
+            },
+            TagColumnInfo {
+                name: "namespace".to_string(),
+                index: 0,
+                column_id: 1,
+            },
+        ];
+        let tag_arrays = build_tag_arrays(&batch, &tag_columns);
+        let tsid_array = compute_tsid_array(&batch, &tag_columns, &tag_arrays);
+
+        assert_eq!(tsid_array.value(0), 2721566936019240841);
+    }
+
+    #[test]
+    fn test_compute_tsid_with_nulls() {
+        let schema = Arc::new(ArrowSchema::new(vec![
+            Field::new("a", DataType::Utf8, true),
+            Field::new("b", DataType::Utf8, true),
+        ]));
+        let batch_no_null = RecordBatch::try_new(
+            schema.clone(),
+            vec![
+                Arc::new(StringArray::from(vec!["A"])),
+                Arc::new(StringArray::from(vec!["B"])),
+            ],
+        )
+        .unwrap();
+        let tag_cols_2: Vec<TagColumnInfo> = vec![
+            TagColumnInfo {
+                name: "a".to_string(),
+                index: 0,
+                column_id: 1,
+            },
+            TagColumnInfo {
+                name: "b".to_string(),
+                index: 1,
+                column_id: 2,
+            },
+        ];
+        let tag_arrays_2 = build_tag_arrays(&batch_no_null, &tag_cols_2);
+        let tsid_no_null = compute_tsid_array(&batch_no_null, &tag_cols_2, &tag_arrays_2);
+
+        let schema3 = Arc::new(ArrowSchema::new(vec![
+            Field::new("a", DataType::Utf8, true),
+            Field::new("b", DataType::Utf8, true),
+            Field::new("c", DataType::Utf8, true),
+        ]));
+        let batch_with_null = RecordBatch::try_new(
+            schema3,
+            vec![
+                Arc::new(StringArray::from(vec!["A"])),
+                Arc::new(StringArray::from(vec!["B"])),
+                Arc::new(StringArray::from(vec![None as Option<&str>])),
+            ],
+        )
+        .unwrap();
+        let tag_cols_3: Vec<TagColumnInfo> = vec![
+            TagColumnInfo {
+                name: "a".to_string(),
+                index: 0,
+                column_id: 1,
+            },
+            TagColumnInfo {
+                name: "b".to_string(),
+                index: 1,
+                column_id: 2,
+            },
+            TagColumnInfo {
+                name: "c".to_string(),
+                index: 2,
+                column_id: 3,
+            },
+        ];
+        let tag_arrays_3 = build_tag_arrays(&batch_with_null, &tag_cols_3);
+        let tsid_with_null = compute_tsid_array(&batch_with_null, &tag_cols_3, &tag_arrays_3);
+
+        assert_eq!(tsid_no_null.value(0), tsid_with_null.value(0));
+    }
+
+    #[test]
+    fn test_modify_batch_sparse() {
+        let batch = build_sparse_test_batch();
+        let tag_columns = sparse_tag_columns();
+        let non_tag_indices = vec![0, 1];
+        let table_id: u32 = 1025;
+
+        let modified =
+            modify_batch_sparse(batch, table_id, &tag_columns, &non_tag_indices).unwrap();
+
+        assert_eq!(modified.num_columns(), 3);
+        assert_eq!(modified.schema().field(0).name(), PRIMARY_KEY_COLUMN_NAME);
+        assert_eq!(modified.schema().field(1).name(), "greptime_timestamp");
+        assert_eq!(modified.schema().field(2).name(), "greptime_value");
+    }
+
+    #[test]
+    fn test_modify_batch_sparse_matches_row_modifier() {
+        let batch = build_sparse_test_batch();
+        let tag_columns = sparse_tag_columns();
+        let non_tag_indices = vec![0, 1];
+        let table_id: u32 = 1025;
+        let modified =
+            modify_batch_sparse(batch, table_id, &tag_columns, &non_tag_indices).unwrap();
+
+        let make_info = |name: &str, column_id: ColumnId| ColumnMetadata {
+            column_schema: datatypes::schema::ColumnSchema::new(
+                name.to_string(),
+                datatypes::prelude::ConcreteDataType::string_datatype(),
+                false,
+            ),
+            semantic_type: SemanticType::Tag,
+            column_id,
+        };
+        let name_to_column_id: HashMap<String, ColumnMetadata> = [
+            (
+                "greptime_timestamp".to_string(),
+                make_info("greptime_timestamp", 0),
+            ),
+            ("greptime_value".to_string(), make_info("greptime_value", 1)),
+            ("namespace".to_string(), make_info("namespace", 2)),
+            ("host".to_string(), make_info("host", 3)),
+        ]
+        .into_iter()
+        .collect();
+
+        let rows = Rows {
+            schema: vec![
+                ColumnSchema {
+                    column_name: "greptime_timestamp".to_string(),
+                    datatype: ColumnDataType::TimestampMillisecond as i32,
+                    semantic_type: SemanticType::Timestamp as i32,
+                    ..Default::default()
+                },
+                ColumnSchema {
+                    column_name: "greptime_value".to_string(),
+                    datatype: ColumnDataType::Float64 as i32,
+                    semantic_type: SemanticType::Field as i32,
+                    ..Default::default()
+                },
+                ColumnSchema {
+                    column_name: "namespace".to_string(),
+                    datatype: ColumnDataType::String as i32,
+                    semantic_type: SemanticType::Tag as i32,
+                    ..Default::default()
+                },
+                ColumnSchema {
+                    column_name: "host".to_string(),
+                    datatype: ColumnDataType::String as i32,
+                    semantic_type: SemanticType::Tag as i32,
+                    ..Default::default()
+                },
+            ],
+            rows: vec![Row {
+                values: vec![
+                    Value {
+                        value_data: Some(ValueData::TimestampMillisecondValue(1000)),
+                    },
+                    Value {
+                        value_data: Some(ValueData::F64Value(42.0)),
+                    },
+                    Value {
+                        value_data: Some(ValueData::StringValue("greptimedb".to_string())),
+                    },
+                    Value {
+                        value_data: Some(ValueData::StringValue("127.0.0.1".to_string())),
+                    },
+                ],
+            }],
+        };
+
+        let row_iter = RowsIter::new(rows, &name_to_column_id);
+        let rows = RowModifier::default()
+            .modify_rows(
+                row_iter,
+                TableIdInput::Single(table_id),
+                PrimaryKeyEncoding::Sparse,
+            )
+            .unwrap();
+        let ValueData::BinaryValue(expected_pk) =
+            rows.rows[0].values[0].value_data.clone().unwrap()
+        else {
+            panic!("expected binary primary key");
+        };
+
+        let actual_array = modified
+            .column(0)
+            .as_any()
+            .downcast_ref::<BinaryArray>()
+            .unwrap();
+        assert_eq!(actual_array.value(0), expected_pk.as_slice());
+    }
+}
diff --git a/src/metric-engine/src/data_region.rs b/src/metric-engine/src/data_region.rs
index 9bc22e1102..b64d61a287 100644
--- a/src/metric-engine/src/data_region.rs
+++ b/src/metric-engine/src/data_region.rs
@@ -13,7 +13,7 @@
 // limitations under the License.
 
 use api::v1::SemanticType;
-use common_telemetry::{debug, info, warn};
+use common_telemetry::{debug, info};
 use datatypes::schema::{SkippingIndexOptions, SkippingIndexType};
 use mito2::engine::MitoEngine;
 use snafu::ResultExt;
@@ -27,8 +27,8 @@ use store_api::storage::{ConcreteDataType, RegionId};
 
 use crate::engine::IndexOptions;
 use crate::error::{
-    ColumnTypeMismatchSnafu, ForbiddenPhysicalAlterSnafu, MitoReadOperationSnafu,
-    MitoWriteOperationSnafu, Result, SetSkippingIndexOptionSnafu,
+    AddingFieldColumnSnafu, ColumnTypeMismatchSnafu, ForbiddenPhysicalAlterSnafu,
+    MitoReadOperationSnafu, MitoWriteOperationSnafu, Result, SetSkippingIndexOptionSnafu,
 };
 use crate::metrics::{FORBIDDEN_OPERATION_COUNT, MITO_DDL_DURATION, PHYSICAL_COLUMN_COUNT};
 use crate::utils;
@@ -132,10 +132,10 @@ impl DataRegion {
                         .fail();
                     }
                 } else {
-                    warn!(
-                        "Column {} in region {region_id} is not a tag",
-                        c.column_schema.name
-                    );
+                    return AddingFieldColumnSnafu {
+                        name: &c.column_schema.name,
+                    }
+                    .fail();
                 };
 
                 c.column_id = new_column_id_start + delta as u32;
diff --git a/src/metric-engine/src/engine.rs b/src/metric-engine/src/engine.rs
index 7a1efedac4..ef4d802cfc 100644
--- a/src/metric-engine/src/engine.rs
+++ b/src/metric-engine/src/engine.rs
@@ -13,6 +13,7 @@
 // limitations under the License.
 
 mod alter;
+mod bulk_insert;
 mod catchup;
 mod close;
 mod create;
@@ -288,9 +289,8 @@ impl RegionEngine for MetricEngine {
                 debug_assert_eq!(region_id, resp_region_id);
                 return response;
             }
-            RegionRequest::BulkInserts(_) => {
-                // todo(hl): find a way to support bulk inserts in metric engine.
-                UnsupportedRegionRequestSnafu { request }.fail()
+            RegionRequest::BulkInserts(bulk) => {
+                self.inner.bulk_insert_region(region_id, bulk).await
             }
         };
 
@@ -575,7 +575,7 @@ struct MetricEngineInner {
 
 #[cfg(test)]
 mod test {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
 
     use common_telemetry::info;
@@ -688,9 +688,7 @@ mod test {
         metric_engine
             .handle_request(
                 physical_region_id,
-                RegionRequest::Flush(RegionFlushRequest {
-                    row_group_size: None,
-                }),
+                RegionRequest::Flush(RegionFlushRequest::default()),
             )
             .await
             .unwrap();
diff --git a/src/metric-engine/src/engine/alter.rs b/src/metric-engine/src/engine/alter.rs
index 4b6b67f31b..892e0c91c2 100644
--- a/src/metric-engine/src/engine/alter.rs
+++ b/src/metric-engine/src/engine/alter.rs
@@ -189,7 +189,7 @@ impl MetricEngineInner {
         let new_add_columns = new_column_names.iter().map(|name| {
             // Safety: previous steps ensure the physical region exist
             let column_metadata = *physical_schema_map.get(name).unwrap();
-            (name.to_string(), column_metadata.column_id)
+            (name.to_string(), column_metadata.clone())
         });
 
         // Writes logical regions metadata to metadata region
diff --git a/src/metric-engine/src/engine/alter/extract_new_columns.rs b/src/metric-engine/src/engine/alter/extract_new_columns.rs
index fdb1ef6126..c3bef1ba4f 100644
--- a/src/metric-engine/src/engine/alter/extract_new_columns.rs
+++ b/src/metric-engine/src/engine/alter/extract_new_columns.rs
@@ -14,11 +14,13 @@
 
 use std::collections::{HashMap, HashSet};
 
+use api::v1::SemanticType;
+use snafu::ensure;
 use store_api::metadata::ColumnMetadata;
 use store_api::region_request::{AlterKind, RegionAlterRequest};
-use store_api::storage::{ColumnId, RegionId};
+use store_api::storage::RegionId;
 
-use crate::error::Result;
+use crate::error::{AddingFieldColumnSnafu, Result};
 
 /// Extract new columns from the create requests.
 ///
@@ -27,7 +29,7 @@ use crate::error::Result;
 /// This function will panic if the alter kind is not `AddColumns`.
 pub fn extract_new_columns<'a>(
     requests: &'a [(RegionId, RegionAlterRequest)],
-    physical_columns: &HashMap<String, ColumnId>,
+    physical_columns: &HashMap<String, ColumnMetadata>,
     new_column_names: &mut HashSet<&'a str>,
     new_columns: &mut Vec<ColumnMetadata>,
 ) -> Result<()> {
@@ -40,6 +42,12 @@ pub fn extract_new_columns<'a>(
             if !physical_columns.contains_key(column_name)
                 && !new_column_names.contains(column_name)
             {
+                ensure!(
+                    col.column_metadata.semantic_type != SemanticType::Field,
+                    AddingFieldColumnSnafu {
+                        name: column_name.to_string(),
+                    }
+                );
                 new_column_names.insert(column_name);
                 // TODO(weny): avoid clone
                 new_columns.push(col.column_metadata.clone());
@@ -49,3 +57,55 @@ pub fn extract_new_columns<'a>(
 
     Ok(())
 }
+
+#[cfg(test)]
+mod tests {
+    use std::collections::{HashMap, HashSet};
+
+    use api::v1::SemanticType;
+    use datatypes::prelude::ConcreteDataType;
+    use datatypes::schema::ColumnSchema;
+    use store_api::metadata::ColumnMetadata;
+    use store_api::region_request::{AddColumn, AlterKind, RegionAlterRequest};
+    use store_api::storage::RegionId;
+
+    use super::*;
+    use crate::error::Error;
+
+    #[test]
+    fn test_extract_new_columns_with_field_type() {
+        let requests = vec![(
+            RegionId::new(1, 1),
+            RegionAlterRequest {
+                kind: AlterKind::AddColumns {
+                    columns: vec![AddColumn {
+                        column_metadata: ColumnMetadata {
+                            column_schema: ColumnSchema::new(
+                                "new_column".to_string(),
+                                ConcreteDataType::string_datatype(),
+                                false,
+                            ),
+                            semantic_type: SemanticType::Field,
+                            column_id: 0,
+                        },
+                        location: None,
+                    }],
+                },
+            },
+        )];
+
+        let physical_columns = HashMap::new();
+        let mut new_column_names = HashSet::new();
+        let mut new_columns = Vec::new();
+
+        let err = extract_new_columns(
+            &requests,
+            &physical_columns,
+            &mut new_column_names,
+            &mut new_columns,
+        )
+        .unwrap_err();
+
+        assert!(matches!(err, Error::AddingFieldColumn { .. }));
+    }
+}
diff --git a/src/metric-engine/src/engine/bulk_insert.rs b/src/metric-engine/src/engine/bulk_insert.rs
new file mode 100644
index 0000000000..24c9e7934c
--- /dev/null
+++ b/src/metric-engine/src/engine/bulk_insert.rs
@@ -0,0 +1,874 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::HashSet;
+
+use api::v1::{ArrowIpc, ColumnDataType, SemanticType};
+use bytes::Bytes;
+use common_error::ext::ErrorExt;
+use common_error::status_code::StatusCode;
+use common_grpc::flight::{FlightEncoder, FlightMessage};
+use common_query::prelude::{greptime_timestamp, greptime_value};
+use datatypes::arrow::array::{Array, Float64Array, StringArray, TimestampMillisecondArray};
+use datatypes::arrow::record_batch::RecordBatch;
+use snafu::{OptionExt, ensure};
+use store_api::codec::PrimaryKeyEncoding;
+use store_api::metadata::RegionMetadataRef;
+use store_api::region_request::{
+    AffectedRows, RegionBulkInsertsRequest, RegionPutRequest, RegionRequest,
+};
+use store_api::storage::RegionId;
+
+use crate::batch_modifier::{TagColumnInfo, modify_batch_sparse};
+use crate::engine::MetricEngineInner;
+use crate::error;
+use crate::error::Result;
+use crate::metrics::MITO_OPERATION_ELAPSED;
+
+impl MetricEngineInner {
+    /// Bulk-inserts rows into a metric region.
+    ///
+    /// **Logical region path:** The request payload is a logical `RecordBatch`
+    /// (timestamp, value and tag columns). It is transformed to physical format
+    /// via `modify_batch_sparse`, encoded to Arrow IPC, and forwarded as a
+    /// `BulkInserts` request to the data region. If mito reports
+    /// `StatusCode::Unsupported`, the request is transparently retried as a `Put`.
+    ///
+    /// **Physical region path:** The request payload is already in physical format
+    /// (produced by the batcher's `flush_batch_physical`). It is forwarded directly
+    /// to the data region with no transformation.
+    ///
+    /// Returns the number of affected rows, or `0` if the input batch is empty.
+    pub async fn bulk_insert_region(
+        &self,
+        region_id: RegionId,
+        request: RegionBulkInsertsRequest,
+    ) -> Result<AffectedRows> {
+        if request.payload.num_rows() == 0 {
+            return Ok(0);
+        }
+        if self.is_physical_region(region_id) {
+            let _timer = MITO_OPERATION_ELAPSED
+                .with_label_values(&["bulk_insert_physical"])
+                .start_timer();
+            return self.bulk_insert_physical_region(region_id, request).await;
+        }
+
+        let _timer = MITO_OPERATION_ELAPSED
+            .with_label_values(&["bulk_insert_logical"])
+            .start_timer();
+        self.bulk_insert_logical_region(region_id, request).await
+    }
+
+    /// Passthrough for bulk inserts targeting a physical data region.
+    ///
+    /// The batch is already in physical format (with `__primary_key`, timestamp,
+    /// value columns), so no logical-to-physical transformation is needed.
+    async fn bulk_insert_physical_region(
+        &self,
+        region_id: RegionId,
+        request: RegionBulkInsertsRequest,
+    ) -> Result<AffectedRows> {
+        self.data_region
+            .write_data(region_id, RegionRequest::BulkInserts(request))
+            .await
+    }
+
+    /// Bulk-inserts logical rows, transforming them to physical format first.
+    async fn bulk_insert_logical_region(
+        &self,
+        region_id: RegionId,
+        request: RegionBulkInsertsRequest,
+    ) -> Result<AffectedRows> {
+        let (physical_region_id, data_region_id, primary_key_encoding) =
+            self.find_data_region_meta(region_id)?;
+
+        if primary_key_encoding != PrimaryKeyEncoding::Sparse {
+            return error::UnsupportedRegionRequestSnafu {
+                request: RegionRequest::BulkInserts(request),
+            }
+            .fail();
+        }
+
+        let batch = request.payload;
+        if batch.num_rows() == 0 {
+            return Ok(0);
+        }
+
+        let logical_metadata = self
+            .logical_region_metadata(physical_region_id, region_id)
+            .await?;
+        let (tag_columns, non_tag_indices) = self.resolve_tag_columns_from_metadata(
+            region_id,
+            data_region_id,
+            &batch,
+            &logical_metadata,
+        )?;
+        let modified_batch = modify_batch_sparse(
+            batch.clone(),
+            region_id.table_id(),
+            &tag_columns,
+            &non_tag_indices,
+        )?;
+        let (schema, data_header, payload) = record_batch_to_ipc(&modified_batch)?;
+
+        let partition_expr_version = request.partition_expr_version;
+        let request = RegionBulkInsertsRequest {
+            region_id: data_region_id,
+            payload: modified_batch,
+            raw_data: ArrowIpc {
+                schema,
+                data_header,
+                payload,
+            },
+            partition_expr_version,
+        };
+        match self
+            .data_region
+            .write_data(data_region_id, RegionRequest::BulkInserts(request))
+            .await
+        {
+            Ok(affected_rows) => Ok(affected_rows),
+            Err(err) if err.status_code() == StatusCode::Unsupported => {
+                // todo(hl): fallback path for PartitionTreeMemtable, remove this once we remove it
+                let rows = record_batch_to_rows(&batch, region_id)?;
+                self.put_region(
+                    region_id,
+                    RegionPutRequest {
+                        rows,
+                        hint: None,
+                        partition_expr_version,
+                    },
+                )
+                .await
+            }
+            Err(err) => Err(err),
+        }
+    }
+
+    fn resolve_tag_columns_from_metadata(
+        &self,
+        logical_region_id: RegionId,
+        data_region_id: RegionId,
+        batch: &RecordBatch,
+        logical_metadata: &RegionMetadataRef,
+    ) -> Result<(Vec<TagColumnInfo>, Vec<usize>)> {
+        let tag_names: HashSet<&str> = logical_metadata
+            .column_metadatas
+            .iter()
+            .filter_map(|column| {
+                if column.semantic_type == SemanticType::Tag {
+                    Some(column.column_schema.name.as_str())
+                } else {
+                    None
+                }
+            })
+            .collect();
+
+        let mut tag_columns = Vec::new();
+        let mut non_tag_indices = Vec::new();
+        {
+            let state = self.state.read().unwrap();
+            let physical_columns = state
+                .physical_region_states()
+                .get(&data_region_id)
+                .context(error::PhysicalRegionNotFoundSnafu {
+                    region_id: data_region_id,
+                })?
+                .physical_columns();
+
+            for (index, field) in batch.schema().fields().iter().enumerate() {
+                let name = field.name();
+                let column_id = physical_columns
+                    .get(name)
+                    .map(|info| info.column_id)
+                    .with_context(|| error::ColumnNotFoundSnafu {
+                        name: name.clone(),
+                        region_id: logical_region_id,
+                    })?;
+                if tag_names.contains(name.as_str()) {
+                    tag_columns.push(TagColumnInfo {
+                        name: name.clone(),
+                        index,
+                        column_id,
+                    });
+                } else {
+                    non_tag_indices.push(index);
+                }
+            }
+        }
+
+        tag_columns.sort_by(|a, b| a.name.cmp(&b.name));
+        Ok((tag_columns, non_tag_indices))
+    }
+}
+
+fn record_batch_to_rows(batch: &RecordBatch, logical_region_id: RegionId) -> Result<api::v1::Rows> {
+    let schema_ref = batch.schema();
+    let fields = schema_ref.fields();
+
+    let mut ts_idx = None;
+    let mut val_idx = None;
+    let mut tag_indices = Vec::new();
+
+    for (idx, field) in fields.iter().enumerate() {
+        if field.name() == greptime_timestamp() {
+            ts_idx = Some(idx);
+            if !matches!(
+                field.data_type(),
+                datatypes::arrow::datatypes::DataType::Timestamp(
+                    datatypes::arrow::datatypes::TimeUnit::Millisecond,
+                    _
+                )
+            ) {
+                return error::UnexpectedRequestSnafu {
+                    reason: format!(
+                        "Timestamp column '{}' in region {:?} has incompatible type: {:?}",
+                        field.name(),
+                        logical_region_id,
+                        field.data_type()
+                    ),
+                }
+                .fail();
+            }
+        } else if field.name() == greptime_value() {
+            val_idx = Some(idx);
+            if !matches!(
+                field.data_type(),
+                datatypes::arrow::datatypes::DataType::Float64
+            ) {
+                return error::UnexpectedRequestSnafu {
+                    reason: format!(
+                        "Value column '{}' in region {:?} has incompatible type: {:?}",
+                        field.name(),
+                        logical_region_id,
+                        field.data_type()
+                    ),
+                }
+                .fail();
+            }
+        } else {
+            if !matches!(
+                field.data_type(),
+                datatypes::arrow::datatypes::DataType::Utf8
+            ) {
+                return error::UnexpectedRequestSnafu {
+                    reason: format!(
+                        "Tag column '{}' in region {:?} must be Utf8, found: {:?}",
+                        field.name(),
+                        logical_region_id,
+                        field.data_type()
+                    ),
+                }
+                .fail();
+            }
+            tag_indices.push(idx);
+        }
+    }
+
+    let ts_idx = ts_idx.with_context(|| error::UnexpectedRequestSnafu {
+        reason: format!(
+            "Timestamp column '{}' not found in RecordBatch for region {:?}",
+            greptime_timestamp(),
+            logical_region_id
+        ),
+    })?;
+    let val_idx = val_idx.with_context(|| error::UnexpectedRequestSnafu {
+        reason: format!(
+            "Value column '{}' not found in RecordBatch for region {:?}",
+            greptime_value(),
+            logical_region_id
+        ),
+    })?;
+
+    let mut schema = Vec::with_capacity(2 + tag_indices.len());
+    schema.push(api::v1::ColumnSchema {
+        column_name: greptime_timestamp().to_string(),
+        datatype: ColumnDataType::TimestampMillisecond as i32,
+        semantic_type: SemanticType::Timestamp as i32,
+        datatype_extension: None,
+        options: None,
+    });
+    schema.push(api::v1::ColumnSchema {
+        column_name: greptime_value().to_string(),
+        datatype: ColumnDataType::Float64 as i32,
+        semantic_type: SemanticType::Field as i32,
+        datatype_extension: None,
+        options: None,
+    });
+    for &idx in &tag_indices {
+        let field = &fields[idx];
+        schema.push(api::v1::ColumnSchema {
+            column_name: field.name().clone(),
+            datatype: ColumnDataType::String as i32,
+            semantic_type: SemanticType::Tag as i32,
+            datatype_extension: None,
+            options: None,
+        });
+    }
+
+    let ts_array = batch
+        .column(ts_idx)
+        .as_any()
+        .downcast_ref::<TimestampMillisecondArray>()
+        .expect("validated as TimestampMillisecond");
+    let val_array = batch
+        .column(val_idx)
+        .as_any()
+        .downcast_ref::<Float64Array>()
+        .expect("validated as Float64");
+    let tag_arrays: Vec<&StringArray> = tag_indices
+        .iter()
+        .map(|&idx| {
+            batch
+                .column(idx)
+                .as_any()
+                .downcast_ref::<StringArray>()
+                .expect("validated as Utf8")
+        })
+        .collect();
+
+    let num_rows = batch.num_rows();
+    let mut rows = Vec::with_capacity(num_rows);
+    for row_idx in 0..num_rows {
+        let mut values = Vec::with_capacity(2 + tag_arrays.len());
+
+        if ts_array.is_null(row_idx) {
+            values.push(api::v1::Value { value_data: None });
+        } else {
+            values.push(api::v1::Value {
+                value_data: Some(api::v1::value::ValueData::TimestampMillisecondValue(
+                    ts_array.value(row_idx),
+                )),
+            });
+        }
+
+        if val_array.is_null(row_idx) {
+            values.push(api::v1::Value { value_data: None });
+        } else {
+            values.push(api::v1::Value {
+                value_data: Some(api::v1::value::ValueData::F64Value(
+                    val_array.value(row_idx),
+                )),
+            });
+        }
+
+        for arr in &tag_arrays {
+            if arr.is_null(row_idx) {
+                values.push(api::v1::Value { value_data: None });
+            } else {
+                values.push(api::v1::Value {
+                    value_data: Some(api::v1::value::ValueData::StringValue(
+                        arr.value(row_idx).to_string(),
+                    )),
+                });
+            }
+        }
+
+        rows.push(api::v1::Row { values });
+    }
+
+    Ok(api::v1::Rows { schema, rows })
+}
+
+fn record_batch_to_ipc(record_batch: &RecordBatch) -> Result<(Bytes, Bytes, Bytes)> {
+    let mut encoder = FlightEncoder::default();
+    let schema = encoder.encode_schema(record_batch.schema().as_ref());
+    let mut iter = encoder
+        .encode(FlightMessage::RecordBatch(record_batch.clone()))
+        .into_iter();
+
+    let Some(flight_data) = iter.next() else {
+        return error::UnexpectedRequestSnafu {
+            reason: "Failed to encode empty flight data",
+        }
+        .fail();
+    };
+    ensure!(
+        iter.next().is_none(),
+        error::UnexpectedRequestSnafu {
+            reason: "Bulk insert RecordBatch with dictionary arrays is unsupported".to_string(),
+        }
+    );
+
+    Ok((
+        schema.data_header,
+        flight_data.data_header,
+        flight_data.data_body,
+    ))
+}
+
+#[cfg(test)]
+mod tests {
+    use std::assert_matches;
+    use std::sync::Arc;
+
+    use api::v1::ArrowIpc;
+    use common_error::ext::ErrorExt;
+    use common_query::prelude::{greptime_timestamp, greptime_value};
+    use common_recordbatch::RecordBatches;
+    use datatypes::arrow::array::{Float64Array, StringArray, TimestampMillisecondArray};
+    use datatypes::arrow::datatypes::{DataType, Field, Schema as ArrowSchema, TimeUnit};
+    use datatypes::arrow::record_batch::RecordBatch;
+    use mito2::config::MitoConfig;
+    use store_api::metric_engine_consts::MEMTABLE_PARTITION_TREE_PRIMARY_KEY_ENCODING;
+    use store_api::path_utils::table_dir;
+    use store_api::region_engine::RegionEngine;
+    use store_api::region_request::{RegionBulkInsertsRequest, RegionPutRequest, RegionRequest};
+    use store_api::storage::{RegionId, ScanRequest};
+
+    use super::record_batch_to_ipc;
+    use crate::batch_modifier::{TagColumnInfo, modify_batch_sparse};
+    use crate::error::Error;
+    use crate::test_util::{self, TestEnv};
+
+    fn build_logical_batch(start: usize, rows: usize) -> RecordBatch {
+        let schema = Arc::new(ArrowSchema::new(vec![
+            Field::new(
+                greptime_timestamp(),
+                DataType::Timestamp(TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new(greptime_value(), DataType::Float64, true),
+            Field::new("job", DataType::Utf8, true),
+        ]));
+
+        let mut ts = Vec::with_capacity(rows);
+        let mut values = Vec::with_capacity(rows);
+        let mut tags = Vec::with_capacity(rows);
+        for i in start..start + rows {
+            ts.push(i as i64);
+            values.push(i as f64);
+            tags.push("tag_0".to_string());
+        }
+
+        RecordBatch::try_new(
+            schema,
+            vec![
+                Arc::new(TimestampMillisecondArray::from(ts)),
+                Arc::new(Float64Array::from(values)),
+                Arc::new(StringArray::from(tags)),
+            ],
+        )
+        .unwrap()
+    }
+
+    fn build_bulk_request(logical_region_id: RegionId, batch: RecordBatch) -> RegionRequest {
+        let (schema, data_header, payload) = record_batch_to_ipc(&batch).unwrap();
+        RegionRequest::BulkInserts(RegionBulkInsertsRequest {
+            region_id: logical_region_id,
+            payload: batch,
+            raw_data: ArrowIpc {
+                schema,
+                data_header,
+                payload,
+            },
+            partition_expr_version: None,
+        })
+    }
+
+    async fn init_dense_metric_region(env: &TestEnv) -> RegionId {
+        let physical_region_id = env.default_physical_region_id();
+        env.create_physical_region(
+            physical_region_id,
+            &TestEnv::default_table_dir(),
+            vec![(
+                MEMTABLE_PARTITION_TREE_PRIMARY_KEY_ENCODING.to_string(),
+                "dense".to_string(),
+            )],
+        )
+        .await;
+
+        let logical_region_id = env.default_logical_region_id();
+        let request = test_util::create_logical_region_request(
+            &["job"],
+            physical_region_id,
+            &table_dir("test", logical_region_id.table_id()),
+        );
+        env.metric()
+            .handle_request(logical_region_id, RegionRequest::Create(request))
+            .await
+            .unwrap();
+        logical_region_id
+    }
+
+    #[tokio::test]
+    async fn test_bulk_insert_empty_batch_returns_zero() {
+        let env = TestEnv::new().await;
+        env.init_metric_region().await;
+        let logical_region_id = env.default_logical_region_id();
+
+        let batch = build_logical_batch(0, 0);
+        let request = RegionRequest::BulkInserts(RegionBulkInsertsRequest {
+            region_id: logical_region_id,
+            payload: batch,
+            raw_data: ArrowIpc::default(),
+            partition_expr_version: None,
+        });
+        let response = env
+            .metric()
+            .handle_request(logical_region_id, request)
+            .await
+            .unwrap();
+        assert_eq!(response.affected_rows, 0);
+    }
+
+    #[tokio::test]
+    async fn test_bulk_insert_physical_region_passthrough() {
+        // Use flat format so that BulkMemtable is used (supports write_bulk).
+        let mito_config = MitoConfig {
+            default_flat_format: true,
+            ..Default::default()
+        };
+        let env = TestEnv::with_mito_config("", mito_config, Default::default()).await;
+        env.init_metric_region().await;
+        let physical_region_id = env.default_physical_region_id();
+        let logical_region_id = env.default_logical_region_id();
+
+        // First, do a normal logical bulk insert so we can compare results.
+        let logical_batch = build_logical_batch(0, 3);
+        let logical_request = build_bulk_request(logical_region_id, logical_batch.clone());
+        let response = env
+            .metric()
+            .handle_request(logical_region_id, logical_request)
+            .await
+            .unwrap();
+        assert_eq!(response.affected_rows, 3);
+
+        // Now build a physical-format batch using modify_batch_sparse (simulating
+        // what the batcher's flush_batch_physical does) and send it directly to
+        // the physical region.
+        let tag_columns = vec![TagColumnInfo {
+            name: "job".to_string(),
+            index: 2,
+            column_id: 2, // column_id for "job" in the physical table
+        }];
+        let non_tag_indices = vec![0, 1]; // timestamp, value
+        let second_batch = build_logical_batch(3, 3);
+        let physical_batch = modify_batch_sparse(
+            second_batch,
+            logical_region_id.table_id(),
+            &tag_columns,
+            &non_tag_indices,
+        )
+        .unwrap();
+        let request = build_bulk_request(physical_region_id, physical_batch);
+        let response = env
+            .metric()
+            .handle_request(physical_region_id, request)
+            .await
+            .unwrap();
+        assert_eq!(response.affected_rows, 3);
+
+        // Verify all 6 rows are readable from the logical region.
+        let stream = env
+            .metric()
+            .scan_to_stream(logical_region_id, ScanRequest::default())
+            .await
+            .unwrap();
+        let batches = RecordBatches::try_collect(stream).await.unwrap();
+        assert_eq!(batches.iter().map(|b| b.num_rows()).sum::<usize>(), 6);
+    }
+
+    #[tokio::test]
+    async fn test_bulk_insert_physical_region_empty_batch() {
+        // Use flat format so that BulkMemtable is used (supports write_bulk).
+        let mito_config = MitoConfig {
+            default_flat_format: true,
+            ..Default::default()
+        };
+        let env = TestEnv::with_mito_config("", mito_config, Default::default()).await;
+        env.init_metric_region().await;
+        let physical_region_id = env.default_physical_region_id();
+
+        let batch = build_logical_batch(0, 0);
+        let request = build_bulk_request(physical_region_id, batch);
+        let response = env
+            .metric()
+            .handle_request(physical_region_id, request)
+            .await
+            .unwrap();
+        assert_eq!(response.affected_rows, 0);
+    }
+
+    #[tokio::test]
+    async fn test_bulk_insert_unknown_column_errors() {
+        let env = TestEnv::new().await;
+        env.init_metric_region().await;
+        let logical_region_id = env.default_logical_region_id();
+
+        let schema = Arc::new(ArrowSchema::new(vec![
+            Field::new(
+                greptime_timestamp(),
+                DataType::Timestamp(TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new(greptime_value(), DataType::Float64, true),
+            Field::new("nonexistent_column", DataType::Utf8, true),
+        ]));
+        let batch = RecordBatch::try_new(
+            schema,
+            vec![
+                Arc::new(TimestampMillisecondArray::from(vec![0i64])),
+                Arc::new(Float64Array::from(vec![1.0])),
+                Arc::new(StringArray::from(vec!["val"])),
+            ],
+        )
+        .unwrap();
+
+        let request = build_bulk_request(logical_region_id, batch);
+        let err = env
+            .metric()
+            .handle_request(logical_region_id, request)
+            .await
+            .unwrap_err();
+        let Some(err) = err.as_any().downcast_ref::<Error>() else {
+            panic!("unexpected error type");
+        };
+        assert_matches!(err, Error::ColumnNotFound { .. });
+    }
+
+    #[tokio::test]
+    async fn test_bulk_insert_multiple_tag_columns() {
+        let env = TestEnv::new().await;
+        let physical_region_id = env.default_physical_region_id();
+        env.create_physical_region(physical_region_id, &TestEnv::default_table_dir(), vec![])
+            .await;
+        let logical_region_id = env.default_logical_region_id();
+        let request = test_util::create_logical_region_request(
+            &["host", "region"],
+            physical_region_id,
+            &table_dir("test", logical_region_id.table_id()),
+        );
+        env.metric()
+            .handle_request(logical_region_id, RegionRequest::Create(request))
+            .await
+            .unwrap();
+
+        let schema = Arc::new(ArrowSchema::new(vec![
+            Field::new(
+                greptime_timestamp(),
+                DataType::Timestamp(TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new(greptime_value(), DataType::Float64, true),
+            Field::new("host", DataType::Utf8, true),
+            Field::new("region", DataType::Utf8, true),
+        ]));
+        let batch = RecordBatch::try_new(
+            schema,
+            vec![
+                Arc::new(TimestampMillisecondArray::from(vec![0i64, 1, 2])),
+                Arc::new(Float64Array::from(vec![10.0, 20.0, 30.0])),
+                Arc::new(StringArray::from(vec!["h1", "h2", "h1"])),
+                Arc::new(StringArray::from(vec!["us-east", "us-west", "eu-west"])),
+            ],
+        )
+        .unwrap();
+
+        let request = build_bulk_request(logical_region_id, batch);
+        let response = env
+            .metric()
+            .handle_request(logical_region_id, request)
+            .await
+            .unwrap();
+        assert_eq!(response.affected_rows, 3);
+
+        let stream = env
+            .metric()
+            .scan_to_stream(logical_region_id, ScanRequest::default())
+            .await
+            .unwrap();
+        let batches = RecordBatches::try_collect(stream).await.unwrap();
+        assert_eq!(batches.iter().map(|b| b.num_rows()).sum::<usize>(), 3);
+    }
+
+    #[tokio::test]
+    async fn test_bulk_insert_accumulates_rows() {
+        let env = TestEnv::new().await;
+        env.init_metric_region().await;
+        let logical_region_id = env.default_logical_region_id();
+
+        let request = build_bulk_request(logical_region_id, build_logical_batch(0, 3));
+        let response = env
+            .metric()
+            .handle_request(logical_region_id, request)
+            .await
+            .unwrap();
+        assert_eq!(response.affected_rows, 3);
+
+        let request = build_bulk_request(logical_region_id, build_logical_batch(3, 5));
+        let response = env
+            .metric()
+            .handle_request(logical_region_id, request)
+            .await
+            .unwrap();
+        assert_eq!(response.affected_rows, 5);
+
+        let stream = env
+            .metric()
+            .scan_to_stream(logical_region_id, ScanRequest::default())
+            .await
+            .unwrap();
+        let batches = RecordBatches::try_collect(stream).await.unwrap();
+        assert_eq!(batches.iter().map(|b| b.num_rows()).sum::<usize>(), 8);
+    }
+
+    #[tokio::test]
+    async fn test_bulk_insert_sparse_encoding() {
+        let env = TestEnv::new().await;
+        env.init_metric_region().await;
+        let logical_region_id = env.default_logical_region_id();
+
+        let request = build_bulk_request(logical_region_id, build_logical_batch(0, 4));
+        let response = env
+            .metric()
+            .handle_request(logical_region_id, request)
+            .await
+            .unwrap();
+        assert_eq!(response.affected_rows, 4);
+
+        let stream = env
+            .metric()
+            .scan_to_stream(logical_region_id, ScanRequest::default())
+            .await
+            .unwrap();
+        let batches = RecordBatches::try_collect(stream).await.unwrap();
+        assert_eq!(batches.iter().map(|b| b.num_rows()).sum::<usize>(), 4);
+    }
+
+    #[tokio::test]
+    async fn test_bulk_insert_dense_encoding_rejected() {
+        let env = TestEnv::new().await;
+        let logical_region_id = init_dense_metric_region(&env).await;
+
+        let request = build_bulk_request(logical_region_id, build_logical_batch(0, 2));
+        let err = env
+            .metric()
+            .handle_request(logical_region_id, request)
+            .await
+            .unwrap_err();
+        let Some(err) = err.as_any().downcast_ref::<Error>() else {
+            panic!("unexpected error type");
+        };
+        assert_matches!(err, Error::UnsupportedRegionRequest { .. });
+    }
+
+    #[tokio::test]
+    async fn test_bulk_insert_matches_put() {
+        let env_put = TestEnv::new().await;
+        env_put.init_metric_region().await;
+        let logical_region_id = env_put.default_logical_region_id();
+        let schema = test_util::row_schema_with_tags(&["job"]);
+        let rows = test_util::build_rows(1, 5);
+        env_put
+            .metric()
+            .handle_request(
+                logical_region_id,
+                RegionRequest::Put(RegionPutRequest {
+                    rows: api::v1::Rows { schema, rows },
+                    hint: None,
+                    partition_expr_version: None,
+                }),
+            )
+            .await
+            .unwrap();
+        let put_stream = env_put
+            .metric()
+            .scan_to_stream(logical_region_id, ScanRequest::default())
+            .await
+            .unwrap();
+        let put_batches = RecordBatches::try_collect(put_stream).await.unwrap();
+        let put_output = put_batches.pretty_print().unwrap();
+
+        let env_bulk = TestEnv::new().await;
+        env_bulk.init_metric_region().await;
+        let request = build_bulk_request(logical_region_id, build_logical_batch(0, 5));
+        env_bulk
+            .metric()
+            .handle_request(logical_region_id, request)
+            .await
+            .unwrap();
+        let bulk_stream = env_bulk
+            .metric()
+            .scan_to_stream(logical_region_id, ScanRequest::default())
+            .await
+            .unwrap();
+        let bulk_batches = RecordBatches::try_collect(bulk_stream).await.unwrap();
+        let bulk_output = bulk_batches.pretty_print().unwrap();
+
+        assert_eq!(put_output, bulk_output);
+    }
+
+    #[test]
+    fn test_record_batch_to_rows_with_null_values() {
+        use datatypes::arrow::array::{Float64Array, StringArray, TimestampMillisecondArray};
+        use datatypes::arrow::datatypes::{DataType, Field, Schema as ArrowSchema, TimeUnit};
+        use datatypes::arrow::record_batch::RecordBatch;
+        use store_api::storage::RegionId;
+
+        use crate::engine::bulk_insert::record_batch_to_rows;
+
+        let schema = Arc::new(ArrowSchema::new(vec![
+            Field::new(
+                greptime_timestamp(),
+                DataType::Timestamp(TimeUnit::Millisecond, None),
+                true,
+            ),
+            Field::new(greptime_value(), DataType::Float64, true),
+            Field::new("job", DataType::Utf8, true),
+            Field::new("host", DataType::Utf8, true),
+        ]));
+
+        let ts_array = TimestampMillisecondArray::from(vec![Some(1000), None, Some(3000)]);
+        let val_array = Float64Array::from(vec![Some(1.0), Some(2.0), None]);
+        let job_array = StringArray::from(vec![Some("job1"), None, Some("job3")]);
+        let host_array = StringArray::from(vec![None, Some("host2"), Some("host3")]);
+
+        let batch = RecordBatch::try_new(
+            schema,
+            vec![
+                Arc::new(ts_array),
+                Arc::new(val_array),
+                Arc::new(job_array),
+                Arc::new(host_array),
+            ],
+        )
+        .unwrap();
+
+        let region_id = RegionId::new(1, 1);
+        let rows = record_batch_to_rows(&batch, region_id).unwrap();
+
+        assert_eq!(rows.rows.len(), 3);
+        assert_eq!(rows.schema.len(), 4);
+
+        // Row 0: all non-null except host
+        assert!(rows.rows[0].values[0].value_data.is_some());
+        assert!(rows.rows[0].values[1].value_data.is_some());
+        assert!(rows.rows[0].values[2].value_data.is_some());
+        assert!(rows.rows[0].values[3].value_data.is_none());
+
+        // Row 1: null timestamp, null job
+        assert!(rows.rows[1].values[0].value_data.is_none());
+        assert!(rows.rows[1].values[1].value_data.is_some());
+        assert!(rows.rows[1].values[2].value_data.is_none());
+        assert!(rows.rows[1].values[3].value_data.is_some());
+
+        // Row 2: null value
+        assert!(rows.rows[2].values[0].value_data.is_some());
+        assert!(rows.rows[2].values[1].value_data.is_none());
+        assert!(rows.rows[2].values[2].value_data.is_some());
+        assert!(rows.rows[2].values[3].value_data.is_some());
+    }
+}
diff --git a/src/metric-engine/src/engine/create.rs b/src/metric-engine/src/engine/create.rs
index 131ec1dbd4..34aeb5577c 100644
--- a/src/metric-engine/src/engine/create.rs
+++ b/src/metric-engine/src/engine/create.rs
@@ -145,7 +145,7 @@ impl MetricEngineInner {
         let physical_columns = create_data_region_request
             .column_metadatas
             .iter()
-            .map(|metadata| (metadata.column_schema.name.clone(), metadata.column_id))
+            .map(|metadata| (metadata.column_schema.name.clone(), metadata.clone()))
             .collect::<HashMap<_, _>>();
         let time_index_unit = create_data_region_request
             .column_metadatas
@@ -321,7 +321,7 @@ impl MetricEngineInner {
         let new_add_columns = new_column_names.iter().map(|name| {
             // Safety: previous steps ensure the physical region exist
             let column_metadata = *physical_schema_map.get(name).unwrap();
-            (name.to_string(), column_metadata.column_id)
+            (name.to_string(), column_metadata.clone())
         });
 
         extension_return_value.insert(
diff --git a/src/metric-engine/src/engine/create/extract_new_columns.rs b/src/metric-engine/src/engine/create/extract_new_columns.rs
index b3eabd5706..82cc53ce3d 100644
--- a/src/metric-engine/src/engine/create/extract_new_columns.rs
+++ b/src/metric-engine/src/engine/create/extract_new_columns.rs
@@ -18,14 +18,14 @@ use api::v1::SemanticType;
 use snafu::ensure;
 use store_api::metadata::ColumnMetadata;
 use store_api::region_request::RegionCreateRequest;
-use store_api::storage::{ColumnId, RegionId};
+use store_api::storage::RegionId;
 
 use crate::error::{AddingFieldColumnSnafu, Result};
 
 /// Extract new columns from the create requests.
 pub fn extract_new_columns<'a>(
     requests: &'a [(RegionId, RegionCreateRequest)],
-    physical_columns: &HashMap<String, ColumnId>,
+    physical_columns: &HashMap<String, ColumnMetadata>,
     new_column_names: &mut HashSet<&'a str>,
     new_columns: &mut Vec<ColumnMetadata>,
 ) -> Result<()> {
@@ -52,7 +52,7 @@ pub fn extract_new_columns<'a>(
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::{HashMap, HashSet};
 
     use api::v1::SemanticType;
@@ -123,7 +123,18 @@ mod tests {
         ];
 
         let mut physical_columns = HashMap::new();
-        physical_columns.insert("existing_column".to_string(), 0);
+        physical_columns.insert(
+            "existing_column".to_string(),
+            ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "existing_column".to_string(),
+                    ConcreteDataType::string_datatype(),
+                    false,
+                ),
+                semantic_type: SemanticType::Tag,
+                column_id: 0,
+            },
+        );
         let mut new_column_names = HashSet::new();
         let mut new_columns = Vec::new();
 
diff --git a/src/metric-engine/src/engine/flush.rs b/src/metric-engine/src/engine/flush.rs
index 5d7479c5d0..5953859cf0 100644
--- a/src/metric-engine/src/engine/flush.rs
+++ b/src/metric-engine/src/engine/flush.rs
@@ -121,6 +121,10 @@ mod tests {
                     .map(|path| path.replace(&e.file_id, "<file_id>"));
                 e.file_id = "<file_id>".to_string();
                 e.index_version = 0;
+                // Round down sizes to nearest 1000 to avoid exact size
+                // comparisons that break when the SST format changes.
+                e.file_size = e.file_size / 1000 * 1000;
+                e.index_file_size = e.index_file_size.map(|s| s / 1000 * 1000);
                 format!("\n{:?}", e)
             })
             .sorted()
@@ -129,12 +133,12 @@ mod tests {
         assert_eq!(
             debug_format,
             r#"
-ManifestSstEntry { table_dir: "test_metric_region/", region_id: 47244640257(11, 1), table_id: 11, region_number: 1, region_group: 0, region_sequence: 1, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test_metric_region/11_0000000001/data/<file_id>.parquet", file_size: 3217, index_file_path: Some("test_metric_region/11_0000000001/data/index/<file_id>.puffin"), index_file_size: Some(235), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9::Millisecond, sequence: Some(20), origin_region_id: 47244640257(11, 1), node_id: None, visible: true }
-ManifestSstEntry { table_dir: "test_metric_region/", region_id: 47244640258(11, 2), table_id: 11, region_number: 2, region_group: 0, region_sequence: 2, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test_metric_region/11_0000000002/data/<file_id>.parquet", file_size: 3217, index_file_path: Some("test_metric_region/11_0000000002/data/index/<file_id>.puffin"), index_file_size: Some(235), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9::Millisecond, sequence: Some(10), origin_region_id: 47244640258(11, 2), node_id: None, visible: true }
-ManifestSstEntry { table_dir: "test_metric_region/", region_id: 47261417473(11, 16777217), table_id: 11, region_number: 16777217, region_group: 1, region_sequence: 1, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test_metric_region/11_0000000001/metadata/<file_id>.parquet", file_size: 3487, index_file_path: None, index_file_size: None, num_rows: 8, num_row_groups: 1, num_series: Some(8), min_ts: 0::Millisecond, max_ts: 0::Millisecond, sequence: Some(8), origin_region_id: 47261417473(11, 16777217), node_id: None, visible: true }
-ManifestSstEntry { table_dir: "test_metric_region/", region_id: 47261417474(11, 16777218), table_id: 11, region_number: 16777218, region_group: 1, region_sequence: 2, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test_metric_region/11_0000000002/metadata/<file_id>.parquet", file_size: 3471, index_file_path: None, index_file_size: None, num_rows: 4, num_row_groups: 1, num_series: Some(4), min_ts: 0::Millisecond, max_ts: 0::Millisecond, sequence: Some(4), origin_region_id: 47261417474(11, 16777218), node_id: None, visible: true }
-ManifestSstEntry { table_dir: "test_metric_region/", region_id: 94489280554(22, 42), table_id: 22, region_number: 42, region_group: 0, region_sequence: 42, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test_metric_region/22_0000000042/data/<file_id>.parquet", file_size: 3217, index_file_path: Some("test_metric_region/22_0000000042/data/index/<file_id>.puffin"), index_file_size: Some(235), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9::Millisecond, sequence: Some(10), origin_region_id: 94489280554(22, 42), node_id: None, visible: true }
-ManifestSstEntry { table_dir: "test_metric_region/", region_id: 94506057770(22, 16777258), table_id: 22, region_number: 16777258, region_group: 1, region_sequence: 42, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test_metric_region/22_0000000042/metadata/<file_id>.parquet", file_size: 3471, index_file_path: None, index_file_size: None, num_rows: 4, num_row_groups: 1, num_series: Some(4), min_ts: 0::Millisecond, max_ts: 0::Millisecond, sequence: Some(4), origin_region_id: 94506057770(22, 16777258), node_id: None, visible: true }"#,
+ManifestSstEntry { table_dir: "test_metric_region/", region_id: 47244640257(11, 1), table_id: 11, region_number: 1, region_group: 0, region_sequence: 1, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test_metric_region/11_0000000001/data/<file_id>.parquet", file_size: 3000, index_file_path: Some("test_metric_region/11_0000000001/data/index/<file_id>.puffin"), index_file_size: Some(0), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9::Millisecond, sequence: Some(20), origin_region_id: 47244640257(11, 1), node_id: None, visible: true }
+ManifestSstEntry { table_dir: "test_metric_region/", region_id: 47244640258(11, 2), table_id: 11, region_number: 2, region_group: 0, region_sequence: 2, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test_metric_region/11_0000000002/data/<file_id>.parquet", file_size: 3000, index_file_path: Some("test_metric_region/11_0000000002/data/index/<file_id>.puffin"), index_file_size: Some(0), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9::Millisecond, sequence: Some(10), origin_region_id: 47244640258(11, 2), node_id: None, visible: true }
+ManifestSstEntry { table_dir: "test_metric_region/", region_id: 47261417473(11, 16777217), table_id: 11, region_number: 16777217, region_group: 1, region_sequence: 1, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test_metric_region/11_0000000001/metadata/<file_id>.parquet", file_size: 4000, index_file_path: None, index_file_size: None, num_rows: 8, num_row_groups: 1, num_series: Some(8), min_ts: 0::Millisecond, max_ts: 0::Millisecond, sequence: Some(8), origin_region_id: 47261417473(11, 16777217), node_id: None, visible: true }
+ManifestSstEntry { table_dir: "test_metric_region/", region_id: 47261417474(11, 16777218), table_id: 11, region_number: 16777218, region_group: 1, region_sequence: 2, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test_metric_region/11_0000000002/metadata/<file_id>.parquet", file_size: 4000, index_file_path: None, index_file_size: None, num_rows: 4, num_row_groups: 1, num_series: Some(4), min_ts: 0::Millisecond, max_ts: 0::Millisecond, sequence: Some(4), origin_region_id: 47261417474(11, 16777218), node_id: None, visible: true }
+ManifestSstEntry { table_dir: "test_metric_region/", region_id: 94489280554(22, 42), table_id: 22, region_number: 42, region_group: 0, region_sequence: 42, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test_metric_region/22_0000000042/data/<file_id>.parquet", file_size: 3000, index_file_path: Some("test_metric_region/22_0000000042/data/index/<file_id>.puffin"), index_file_size: Some(0), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9::Millisecond, sequence: Some(10), origin_region_id: 94489280554(22, 42), node_id: None, visible: true }
+ManifestSstEntry { table_dir: "test_metric_region/", region_id: 94506057770(22, 16777258), table_id: 22, region_number: 16777258, region_group: 1, region_sequence: 42, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test_metric_region/22_0000000042/metadata/<file_id>.parquet", file_size: 4000, index_file_path: None, index_file_size: None, num_rows: 4, num_row_groups: 1, num_series: Some(4), min_ts: 0::Millisecond, max_ts: 0::Millisecond, sequence: Some(4), origin_region_id: 94506057770(22, 16777258), node_id: None, visible: true }"#,
         );
         // list from storage
         let storage_entries = mito
@@ -147,6 +151,8 @@ ManifestSstEntry { table_dir: "test_metric_region/", region_id: 94506057770(22,
             .map(|mut e| {
                 let i = e.file_path.rfind('/').unwrap();
                 e.file_path.replace_range(i..(i + 37), "/<file_id>");
+                e.file_size = None;
+                e.last_modified_ms = None;
                 format!("\n{:?}", e)
             })
             .sorted()
diff --git a/src/metric-engine/src/engine/open.rs b/src/metric-engine/src/engine/open.rs
index 70e3280183..59b1cfd928 100644
--- a/src/metric-engine/src/engine/open.rs
+++ b/src/metric-engine/src/engine/open.rs
@@ -326,7 +326,7 @@ impl MetricEngineInner {
                 .unwrap();
             let physical_columns = physical_columns
                 .into_iter()
-                .map(|col| (col.column_schema.name, col.column_id))
+                .map(|col| (col.column_schema.name.clone(), col))
                 .collect();
             state.add_physical_region(
                 physical_region_id,
diff --git a/src/metric-engine/src/engine/put.rs b/src/metric-engine/src/engine/put.rs
index 9251605aea..07adfae120 100644
--- a/src/metric-engine/src/engine/put.rs
+++ b/src/metric-engine/src/engine/put.rs
@@ -14,13 +14,16 @@
 
 use std::collections::HashMap;
 
+use api::helper::ColumnDataTypeWrapper;
 use api::v1::{
-    ColumnSchema, PrimaryKeyEncoding as PrimaryKeyEncodingProto, Row, Rows, Value, WriteHint,
+    ColumnSchema, PrimaryKeyEncoding as PrimaryKeyEncodingProto, Row, Rows, SemanticType, Value,
+    WriteHint,
 };
 use common_telemetry::{error, info};
 use fxhash::FxHashMap;
-use snafu::{OptionExt, ensure};
+use snafu::{OptionExt, ResultExt, ensure};
 use store_api::codec::PrimaryKeyEncoding;
+use store_api::metadata::ColumnMetadata;
 use store_api::region_request::{
     AffectedRows, RegionDeleteRequest, RegionPutRequest, RegionRequest,
 };
@@ -28,8 +31,9 @@ use store_api::storage::{RegionId, TableId};
 
 use crate::engine::MetricEngineInner;
 use crate::error::{
-    ColumnNotFoundSnafu, ForbiddenPhysicalAlterSnafu, InvalidRequestSnafu,
-    LogicalRegionNotFoundSnafu, PhysicalRegionNotFoundSnafu, Result, UnsupportedRegionRequestSnafu,
+    ColumnNotFoundSnafu, CreateDefaultSnafu, ForbiddenPhysicalAlterSnafu, InvalidRequestSnafu,
+    LogicalRegionNotFoundSnafu, PhysicalRegionNotFoundSnafu, Result, UnexpectedRequestSnafu,
+    UnsupportedRegionRequestSnafu,
 };
 use crate::metrics::{FORBIDDEN_OPERATION_COUNT, MITO_OPERATION_ELAPSED};
 use crate::row_modifier::{RowsIter, TableIdInput};
@@ -116,7 +120,7 @@ impl MetricEngineInner {
     async fn put_regions_batch_single_physical(
         &self,
         physical_region_id: RegionId,
-        requests: Vec<(RegionId, RegionPutRequest)>,
+        mut requests: Vec<(RegionId, RegionPutRequest)>,
     ) -> Result<AffectedRows> {
         if requests.is_empty() {
             return Ok(0);
@@ -126,7 +130,7 @@ impl MetricEngineInner {
         let primary_key_encoding = self.get_primary_key_encoding(data_region_id)?;
 
         // Validate all requests
-        self.validate_batch_requests(physical_region_id, &requests)
+        self.validate_batch_requests(physical_region_id, &mut requests)
             .await?;
 
         // Merge requests according to encoding strategy
@@ -157,11 +161,16 @@ impl MetricEngineInner {
     async fn validate_batch_requests(
         &self,
         physical_region_id: RegionId,
-        requests: &[(RegionId, RegionPutRequest)],
+        requests: &mut [(RegionId, RegionPutRequest)],
     ) -> Result<()> {
         for (logical_region_id, request) in requests {
-            self.verify_rows(*logical_region_id, physical_region_id, &request.rows)
-                .await?;
+            self.verify_rows(
+                *logical_region_id,
+                physical_region_id,
+                &mut request.rows,
+                true,
+            )
+            .await?;
         }
         Ok(())
     }
@@ -248,7 +257,7 @@ impl MetricEngineInner {
         // Batch-modify all rows (add __table_id and __tsid columns)
         let final_rows = {
             let state = self.state.read().unwrap();
-            let name_to_id = state
+            let physical_columns = state
                 .physical_region_states()
                 .get(&data_region_id)
                 .with_context(|| PhysicalRegionNotFoundSnafu {
@@ -261,7 +270,7 @@ impl MetricEngineInner {
                     schema: merged_schema,
                     rows: merged_rows,
                 },
-                name_to_id,
+                physical_columns,
             );
 
             self.row_modifier.modify_rows(
@@ -406,8 +415,13 @@ impl MetricEngineInner {
         let (physical_region_id, data_region_id, primary_key_encoding) =
             self.find_data_region_meta(logical_region_id)?;
 
-        self.verify_rows(logical_region_id, physical_region_id, &request.rows)
-            .await?;
+        self.verify_rows(
+            logical_region_id,
+            physical_region_id,
+            &mut request.rows,
+            true,
+        )
+        .await?;
 
         // write to data region
         // TODO: retrieve table name
@@ -439,8 +453,13 @@ impl MetricEngineInner {
         let (physical_region_id, data_region_id, primary_key_encoding) =
             self.find_data_region_meta(logical_region_id)?;
 
-        self.verify_rows(logical_region_id, physical_region_id, &request.rows)
-            .await?;
+        self.verify_rows(
+            logical_region_id,
+            physical_region_id,
+            &mut request.rows,
+            false,
+        )
+        .await?;
 
         // write to data region
         // TODO: retrieve table name
@@ -460,7 +479,7 @@ impl MetricEngineInner {
             .await
     }
 
-    fn find_data_region_meta(
+    pub(crate) fn find_data_region_meta(
         &self,
         logical_region_id: RegionId,
     ) -> Result<(RegionId, RegionId, PrimaryKeyEncoding)> {
@@ -484,12 +503,18 @@ impl MetricEngineInner {
     ///
     /// Includes:
     /// - Check if the logical region exists
-    /// - Check if the columns exist
+    /// - Check if every column in the request exists in the physical region
+    /// - Check each column's datatype and semantic type match the physical region's schema
+    /// - Check the time index column is present
+    /// - When `check_fields` is true, check every physical field column is present.
+    ///   Set this to `false` for delete requests, which legitimately carry only
+    ///   the primary key + timestamp.
     async fn verify_rows(
         &self,
         logical_region_id: RegionId,
         physical_region_id: RegionId,
-        rows: &Rows,
+        rows: &mut Rows,
+        check_fields: bool,
     ) -> Result<()> {
         // Check if the region exists
         let data_region_id = to_data_region_id(physical_region_id);
@@ -502,22 +527,137 @@ impl MetricEngineInner {
             .fail();
         }
 
-        // Check if a physical column exists
-        let physical_columns = state
+        // Type + semantic check on every column in the request schema.
+        let physical_state = state
             .physical_region_states()
             .get(&data_region_id)
             .context(PhysicalRegionNotFoundSnafu {
                 region_id: data_region_id,
-            })?
-            .physical_columns();
+            })?;
+        let physical_columns = physical_state.physical_columns();
         for col in &rows.schema {
-            ensure!(
-                physical_columns.contains_key(&col.column_name),
-                ColumnNotFoundSnafu {
-                    name: col.column_name.clone(),
+            let info = physical_columns
+                .get(&col.column_name)
+                .context(ColumnNotFoundSnafu {
+                    name: &col.column_name,
                     region_id: logical_region_id,
+                })?;
+
+            ensure!(
+                api::helper::is_column_type_value_eq(
+                    col.datatype,
+                    col.datatype_extension.clone(),
+                    &info.column_schema.data_type
+                ),
+                InvalidRequestSnafu {
+                    region_id: logical_region_id,
+                    reason: format!(
+                        "column {} expect type {:?}, given: {}({})",
+                        col.column_name,
+                        info.column_schema.data_type,
+                        api::v1::ColumnDataType::try_from(col.datatype)
+                            .map(|v| v.as_str_name())
+                            .unwrap_or("Unknown"),
+                        col.datatype,
+                    ),
                 }
             );
+
+            ensure!(
+                api::helper::is_semantic_type_eq(col.semantic_type, info.semantic_type),
+                InvalidRequestSnafu {
+                    region_id: logical_region_id,
+                    reason: format!(
+                        "column {} expect semantic type {:?}, given: {}({})",
+                        col.column_name,
+                        info.semantic_type,
+                        api::v1::SemanticType::try_from(col.semantic_type)
+                            .map(|v| v.as_str_name())
+                            .unwrap_or("Unknown"),
+                        col.semantic_type,
+                    ),
+                }
+            );
+        }
+
+        let ts_name = physical_state.time_index_column_name();
+        ensure!(
+            rows.schema.iter().any(|col| col.column_name == ts_name),
+            InvalidRequestSnafu {
+                region_id: logical_region_id,
+                reason: format!("missing required time index column {ts_name}"),
+            }
+        );
+
+        if check_fields {
+            let field_name = physical_state.field_column_name();
+            if !rows.schema.iter().any(|col| col.column_name == field_name) {
+                let field_meta =
+                    physical_columns
+                        .get(field_name)
+                        .with_context(|| ColumnNotFoundSnafu {
+                            name: field_name,
+                            region_id: logical_region_id,
+                        })?;
+                Self::fill_missing_field_column(logical_region_id, field_name, field_meta, rows)?;
+            }
+        }
+
+        Ok(())
+    }
+
+    fn fill_missing_field_column(
+        logical_region_id: RegionId,
+        field_name: &str,
+        field_meta: &ColumnMetadata,
+        rows: &mut Rows,
+    ) -> Result<()> {
+        ensure!(
+            !field_meta.column_schema.is_default_impure(),
+            UnexpectedRequestSnafu {
+                reason: format!(
+                    "unexpected impure default value with region_id: {logical_region_id}, column: {field_name}, default_value: {:?}",
+                    field_meta.column_schema.default_constraint(),
+                ),
+            }
+        );
+
+        let default_value = field_meta
+            .column_schema
+            .create_default()
+            .context(CreateDefaultSnafu {
+                region_id: logical_region_id,
+                column: field_name,
+            })?
+            .with_context(|| InvalidRequestSnafu {
+                region_id: logical_region_id,
+                reason: format!("missing required field column {field_name}"),
+            })?;
+        let default_value = api::helper::to_grpc_value(default_value);
+        let (datatype, datatype_extension) =
+            ColumnDataTypeWrapper::try_from(field_meta.column_schema.data_type.clone())
+                .map_err(|e| {
+                    InvalidRequestSnafu {
+                        region_id: logical_region_id,
+                        reason: format!(
+                            "no protobuf type for field column {field_name} ({:?}): {e}",
+                            field_meta.column_schema.data_type
+                        ),
+                    }
+                    .build()
+                })?
+                .to_parts();
+
+        rows.schema.push(ColumnSchema {
+            column_name: field_name.to_string(),
+            datatype: datatype as i32,
+            semantic_type: SemanticType::Field as i32,
+            datatype_extension,
+            options: None,
+        });
+
+        for row in &mut rows.rows {
+            row.values.push(default_value.clone());
         }
 
         Ok(())
@@ -536,14 +676,14 @@ impl MetricEngineInner {
         let input = std::mem::take(rows);
         let iter = {
             let state = self.state.read().unwrap();
-            let name_to_id = state
+            let physical_columns = state
                 .physical_region_states()
                 .get(&physical_region_id)
                 .with_context(|| PhysicalRegionNotFoundSnafu {
                     region_id: physical_region_id,
                 })?
                 .physical_columns();
-            RowsIter::new(input, name_to_id)
+            RowsIter::new(input, physical_columns)
         };
         let output =
             self.row_modifier
@@ -557,12 +697,17 @@ impl MetricEngineInner {
 mod tests {
     use std::collections::HashSet;
 
+    use api::v1::value::ValueData;
+    use api::v1::{ColumnDataType, ColumnSchema as PbColumnSchema};
     use common_error::ext::ErrorExt;
     use common_error::status_code::StatusCode;
     use common_function::utils::partition_expr_version;
     use common_recordbatch::RecordBatches;
+    use datatypes::prelude::ConcreteDataType;
+    use datatypes::schema::{ColumnDefaultConstraint, ColumnSchema};
     use datatypes::value::Value as PartitionValue;
     use partition::expr::col;
+    use store_api::metadata::ColumnMetadata;
     use store_api::metric_engine_consts::{
         DATA_SCHEMA_TABLE_ID_COLUMN_NAME, DATA_SCHEMA_TSID_COLUMN_NAME,
         MEMTABLE_PARTITION_TREE_PRIMARY_KEY_ENCODING,
@@ -1258,4 +1403,271 @@ mod tests {
             .unwrap();
         assert_eq!(response.affected_rows, 3);
     }
+
+    /// Regression test for issue #7990: the metric engine must reject a row
+    /// whose timestamp column carries a non-timestamp datatype, rather than
+    /// letting it panic inside mito's `ValueBuilder::push`.
+    #[tokio::test]
+    async fn test_verify_rows_rejects_wrong_type() {
+        use api::v1::value::ValueData;
+        use api::v1::{ColumnDataType, ColumnSchema as PbColumnSchema, SemanticType};
+        use common_query::prelude::{greptime_timestamp, greptime_value};
+
+        let env = TestEnv::new().await;
+        env.init_metric_region().await;
+
+        let logical_region_id = env.default_logical_region_id();
+
+        // Timestamp column is declared as String — the very payload that
+        // caused #7990. It should surface a typed error rather than panic.
+        let schema = vec![
+            PbColumnSchema {
+                column_name: greptime_timestamp().to_string(),
+                datatype: ColumnDataType::String as i32,
+                semantic_type: SemanticType::Timestamp as _,
+                datatype_extension: None,
+                options: None,
+            },
+            PbColumnSchema {
+                column_name: greptime_value().to_string(),
+                datatype: ColumnDataType::Float64 as i32,
+                semantic_type: SemanticType::Field as _,
+                datatype_extension: None,
+                options: None,
+            },
+            PbColumnSchema {
+                column_name: "job".to_string(),
+                datatype: ColumnDataType::String as i32,
+                semantic_type: SemanticType::Tag as _,
+                datatype_extension: None,
+                options: None,
+            },
+        ];
+        let rows = vec![Row {
+            values: vec![
+                Value {
+                    value_data: Some(ValueData::StringValue("not-a-timestamp".to_string())),
+                },
+                Value {
+                    value_data: Some(ValueData::F64Value(1.0)),
+                },
+                Value {
+                    value_data: Some(ValueData::StringValue("tag_0".to_string())),
+                },
+            ],
+        }];
+
+        let err = env
+            .metric()
+            .handle_request(
+                logical_region_id,
+                RegionRequest::Put(RegionPutRequest {
+                    rows: Rows { schema, rows },
+                    hint: None,
+                    partition_expr_version: None,
+                }),
+            )
+            .await
+            .unwrap_err();
+        assert_eq!(err.status_code(), StatusCode::InvalidArguments);
+    }
+
+    /// The completeness check must reject requests that omit the time index
+    /// column, since mito cannot default-fill a `TimeIndex` column and would
+    /// previously panic on the empty builder.
+    #[tokio::test]
+    async fn test_verify_rows_rejects_missing_time_index() {
+        use api::v1::{ColumnDataType, ColumnSchema as PbColumnSchema, SemanticType};
+        use common_query::prelude::greptime_value;
+
+        let env = TestEnv::new().await;
+        env.init_metric_region().await;
+
+        let logical_region_id = env.default_logical_region_id();
+
+        // Payload only carries the field and a tag — no timestamp column.
+        let schema = vec![
+            PbColumnSchema {
+                column_name: greptime_value().to_string(),
+                datatype: ColumnDataType::Float64 as i32,
+                semantic_type: SemanticType::Field as _,
+                datatype_extension: None,
+                options: None,
+            },
+            PbColumnSchema {
+                column_name: "job".to_string(),
+                datatype: ColumnDataType::String as i32,
+                semantic_type: SemanticType::Tag as _,
+                datatype_extension: None,
+                options: None,
+            },
+        ];
+        let rows = vec![Row {
+            values: vec![
+                Value {
+                    value_data: Some(api::v1::value::ValueData::F64Value(1.0)),
+                },
+                Value {
+                    value_data: Some(api::v1::value::ValueData::StringValue("tag_0".to_string())),
+                },
+            ],
+        }];
+
+        let err = env
+            .metric()
+            .handle_request(
+                logical_region_id,
+                RegionRequest::Put(RegionPutRequest {
+                    rows: Rows { schema, rows },
+                    hint: None,
+                    partition_expr_version: None,
+                }),
+            )
+            .await
+            .unwrap_err();
+        assert_eq!(err.status_code(), StatusCode::InvalidArguments);
+    }
+
+    #[tokio::test]
+    async fn test_verify_rows_rejects_missing_field() {
+        use api::v1::value::ValueData;
+        use api::v1::{ColumnDataType, ColumnSchema as PbColumnSchema, SemanticType};
+        use common_query::prelude::greptime_timestamp;
+
+        let env = TestEnv::new().await;
+        env.init_metric_region().await;
+
+        let logical_region_id = env.default_logical_region_id();
+
+        // Schema has timestamp + tag but no field column.
+        let schema = vec![
+            PbColumnSchema {
+                column_name: greptime_timestamp().to_string(),
+                datatype: ColumnDataType::TimestampMillisecond as i32,
+                semantic_type: SemanticType::Timestamp as _,
+                datatype_extension: None,
+                options: None,
+            },
+            PbColumnSchema {
+                column_name: "job".to_string(),
+                datatype: ColumnDataType::String as i32,
+                semantic_type: SemanticType::Tag as _,
+                datatype_extension: None,
+                options: None,
+            },
+        ];
+        let rows = vec![Row {
+            values: vec![
+                Value {
+                    value_data: Some(ValueData::TimestampMillisecondValue(0)),
+                },
+                Value {
+                    value_data: Some(ValueData::StringValue("tag_0".to_string())),
+                },
+            ],
+        }];
+
+        let err = env
+            .metric()
+            .handle_request(
+                logical_region_id,
+                RegionRequest::Put(RegionPutRequest {
+                    rows: Rows { schema, rows },
+                    hint: None,
+                    partition_expr_version: None,
+                }),
+            )
+            .await
+            .unwrap_err();
+        let message = err.to_string();
+        assert!(
+            message.contains("missing required field column"),
+            "expected field-completeness rejection, got: {message}"
+        );
+        assert_eq!(err.status_code(), StatusCode::InvalidArguments);
+    }
+
+    #[test]
+    fn test_fill_missing_field_column_nullable_no_default() {
+        let field_meta = ColumnMetadata {
+            column_id: 1,
+            semantic_type: SemanticType::Field,
+            column_schema: ColumnSchema::new(
+                "greptime_value".to_string(),
+                ConcreteDataType::float64_datatype(),
+                true, // nullable, no default
+            ),
+        };
+        let mut rows = Rows {
+            schema: vec![PbColumnSchema {
+                column_name: "ts".to_string(),
+                datatype: ColumnDataType::TimestampMillisecond as i32,
+                semantic_type: SemanticType::Timestamp as _,
+                datatype_extension: None,
+                options: None,
+            }],
+            rows: vec![Row {
+                values: vec![Value {
+                    value_data: Some(ValueData::TimestampMillisecondValue(0)),
+                }],
+            }],
+        };
+
+        MetricEngineInner::fill_missing_field_column(
+            RegionId::new(1, 1),
+            "greptime_value",
+            &field_meta,
+            &mut rows,
+        )
+        .unwrap();
+
+        assert_eq!(rows.schema.len(), 2);
+        assert_eq!(rows.schema[1].column_name, "greptime_value");
+        assert_eq!(rows.rows[0].values.len(), 2);
+        assert!(
+            rows.rows[0].values[1].value_data.is_none(),
+            "missing nullable field should be filled with null"
+        );
+    }
+
+    #[test]
+    fn test_fill_missing_field_column_rejects_impure_default() {
+        let field_meta = ColumnMetadata {
+            column_id: 1,
+            semantic_type: SemanticType::Field,
+            column_schema: ColumnSchema::new(
+                "greptime_value".to_string(),
+                ConcreteDataType::timestamp_millisecond_datatype(),
+                false,
+            )
+            .with_default_constraint(Some(ColumnDefaultConstraint::Function("now()".to_string())))
+            .unwrap(),
+        };
+        let mut rows = Rows {
+            schema: vec![PbColumnSchema {
+                column_name: "ts".to_string(),
+                datatype: api::v1::ColumnDataType::TimestampMillisecond as i32,
+                semantic_type: SemanticType::Timestamp as _,
+                datatype_extension: None,
+                options: None,
+            }],
+            rows: vec![Row {
+                values: vec![Value {
+                    value_data: Some(ValueData::TimestampMillisecondValue(0)),
+                }],
+            }],
+        };
+
+        let err = MetricEngineInner::fill_missing_field_column(
+            RegionId::new(1, 1),
+            "greptime_value",
+            &field_meta,
+            &mut rows,
+        )
+        .unwrap_err();
+        assert!(
+            err.to_string().contains("impure default value"),
+            "expected impure-default rejection, got: {err}"
+        );
+    }
 }
diff --git a/src/metric-engine/src/engine/read.rs b/src/metric-engine/src/engine/read.rs
index 13ae461db1..178c2e5b92 100644
--- a/src/metric-engine/src/engine/read.rs
+++ b/src/metric-engine/src/engine/read.rs
@@ -143,17 +143,26 @@ impl MetricEngineInner {
         mut request: ScanRequest,
     ) -> Result<ScanRequest> {
         // transform projection
-        let physical_projection = if let Some(projection) = &request.projection {
-            self.transform_projection(physical_region_id, logical_region_id, projection)
-                .await?
-        } else {
-            self.default_projection(physical_region_id, logical_region_id)
+        let physical_projection = match request.projection_input.as_ref() {
+            Some(projection_input) => {
+                self.transform_projection(
+                    physical_region_id,
+                    logical_region_id,
+                    &projection_input.projection,
+                )
                 .await?
+            }
+            None => {
+                self.default_projection(physical_region_id, logical_region_id)
+                    .await?
+            }
         };
 
-        request.projection = Some(physical_projection);
+        // Rewrite the top-level projection from logical-region schema indices to
+        // physical-region schema indices. `nested_paths` are left unchanged because
+        // they are expressed by column name rather than schema index.
+        request.projection_input.get_or_insert_default().projection = physical_projection;
 
-        // add table filter
         request
             .filters
             .push(self.table_id_filter(logical_region_id));
@@ -325,8 +334,9 @@ mod test {
             .unwrap();
 
         // check explicit projection
+        let projection_input = Some(vec![0, 1, 2, 3, 4, 5, 6].into());
         let scan_req = ScanRequest {
-            projection: Some(vec![0, 1, 2, 3, 4, 5, 6]),
+            projection_input,
             filters: vec![],
             ..Default::default()
         };
@@ -338,7 +348,10 @@ mod test {
             .await
             .unwrap();
 
-        assert_eq!(scan_req.projection.unwrap(), vec![11, 10, 9, 8, 0, 1, 4]);
+        assert_eq!(
+            scan_req.projection_indices().unwrap(),
+            &[11, 10, 9, 8, 0, 1, 4]
+        );
         assert_eq!(scan_req.filters.len(), 1);
         assert_eq!(
             scan_req.filters[0],
@@ -354,6 +367,9 @@ mod test {
             .transform_request(physical_region_id, logical_region_id, scan_req)
             .await
             .unwrap();
-        assert_eq!(scan_req.projection.unwrap(), vec![11, 10, 9, 8, 0, 1, 4]);
+        assert_eq!(
+            scan_req.projection_indices().unwrap(),
+            &[11, 10, 9, 8, 0, 1, 4]
+        );
     }
 }
diff --git a/src/metric-engine/src/engine/state.rs b/src/metric-engine/src/engine/state.rs
index 7c64758f48..25416380f1 100644
--- a/src/metric-engine/src/engine/state.rs
+++ b/src/metric-engine/src/engine/state.rs
@@ -16,11 +16,13 @@
 
 use std::collections::{HashMap, HashSet};
 
+use api::v1::SemanticType;
+use common_telemetry::warn;
 use common_time::timestamp::TimeUnit;
 use snafu::OptionExt;
 use store_api::codec::PrimaryKeyEncoding;
 use store_api::metadata::ColumnMetadata;
-use store_api::storage::{ColumnId, RegionId};
+use store_api::storage::RegionId;
 
 use crate::engine::options::PhysicalRegionOptions;
 use crate::error::{PhysicalRegionNotFoundSnafu, Result};
@@ -29,7 +31,16 @@ use crate::utils::to_data_region_id;
 
 pub struct PhysicalRegionState {
     logical_regions: HashSet<RegionId>,
-    physical_columns: HashMap<String, ColumnId>,
+    physical_columns: HashMap<String, ColumnMetadata>,
+    /// Name of the time index column, cached at region load so that the write
+    /// path doesn't have to scan `physical_columns` for the timestamp on every
+    /// row batch. The time index is fixed at region creation and never
+    /// changes, so this stays in sync with `physical_columns`.
+    time_index_column_name: String,
+    /// Name of the field column. Metric regions have exactly one field column
+    /// verified at creation time, so the write path can validate completeness
+    /// without consulting per-logical-region metadata.
+    field_column_name: String,
     primary_key_encoding: PrimaryKeyEncoding,
     options: PhysicalRegionOptions,
     time_index_unit: TimeUnit,
@@ -37,14 +48,29 @@ pub struct PhysicalRegionState {
 
 impl PhysicalRegionState {
     pub fn new(
-        physical_columns: HashMap<String, ColumnId>,
+        physical_columns: HashMap<String, ColumnMetadata>,
         primary_key_encoding: PrimaryKeyEncoding,
         options: PhysicalRegionOptions,
         time_index_unit: TimeUnit,
     ) -> Self {
+        // Safety: a valid physical region always has exactly one time index
+        // column; callers validate this before reaching here (see
+        // `create_data_region_request` and the open path).
+        let time_index_column_name = physical_columns
+            .iter()
+            .find(|(_, meta)| meta.semantic_type == SemanticType::Timestamp)
+            .map(|(name, _)| name.clone())
+            .unwrap_or_default();
+        let field_column_name = physical_columns
+            .iter()
+            .find(|(_, meta)| meta.semantic_type == SemanticType::Field)
+            .map(|(name, _)| name.clone())
+            .unwrap_or_default();
         Self {
             logical_regions: HashSet::new(),
             physical_columns,
+            time_index_column_name,
+            field_column_name,
             primary_key_encoding,
             options,
             time_index_unit,
@@ -57,10 +83,20 @@ impl PhysicalRegionState {
     }
 
     /// Returns a reference to the physical columns.
-    pub fn physical_columns(&self) -> &HashMap<String, ColumnId> {
+    pub fn physical_columns(&self) -> &HashMap<String, ColumnMetadata> {
         &self.physical_columns
     }
 
+    /// Returns the cached name of the time index column.
+    pub fn time_index_column_name(&self) -> &str {
+        &self.time_index_column_name
+    }
+
+    /// Returns the cached name of the field column.
+    pub fn field_column_name(&self) -> &str {
+        &self.field_column_name
+    }
+
     /// Returns a reference to the physical region options.
     pub fn options(&self) -> &PhysicalRegionOptions {
         &self.options
@@ -90,7 +126,7 @@ impl MetricEngineState {
     pub fn add_physical_region(
         &mut self,
         physical_region_id: RegionId,
-        physical_columns: HashMap<String, ColumnId>,
+        physical_columns: HashMap<String, ColumnMetadata>,
         primary_key_encoding: PrimaryKeyEncoding,
         options: PhysicalRegionOptions,
         time_index_unit: TimeUnit,
@@ -112,12 +148,25 @@ impl MetricEngineState {
     pub fn add_physical_columns(
         &mut self,
         physical_region_id: RegionId,
-        physical_columns: impl IntoIterator<Item = (String, ColumnId)>,
+        physical_columns: impl IntoIterator<Item = (String, ColumnMetadata)>,
     ) {
         let physical_region_id = to_data_region_id(physical_region_id);
         let state = self.physical_regions.get_mut(&physical_region_id).unwrap();
-        for (col, id) in physical_columns {
-            state.physical_columns.insert(col, id);
+        for (col, meta) in physical_columns {
+            // The time index is fixed at region creation and alter cannot add
+            // a new one; keep the cached name in sync defensively.
+            debug_assert_ne!(
+                meta.semantic_type,
+                SemanticType::Timestamp,
+                "unexpected time index column {col} added to an existing physical region"
+            );
+            if meta.semantic_type == SemanticType::Field {
+                warn!(
+                    "Unexpected field column {col} added to physical region {physical_region_id}; cached field column remains {}",
+                    state.field_column_name
+                );
+            }
+            state.physical_columns.insert(col, meta);
         }
     }
 
diff --git a/src/metric-engine/src/engine/sync/region.rs b/src/metric-engine/src/engine/sync/region.rs
index 3fe51d1b68..cbe6515a19 100644
--- a/src/metric-engine/src/engine/sync/region.rs
+++ b/src/metric-engine/src/engine/sync/region.rs
@@ -231,9 +231,7 @@ mod tests {
         metric_engine
             .handle_request(
                 source_physical_region_id,
-                RegionRequest::Flush(RegionFlushRequest {
-                    row_group_size: None,
-                }),
+                RegionRequest::Flush(RegionFlushRequest::default()),
             )
             .await
             .unwrap();
diff --git a/src/metric-engine/src/error.rs b/src/metric-engine/src/error.rs
index 6fb3406826..284b1b0298 100644
--- a/src/metric-engine/src/error.rs
+++ b/src/metric-engine/src/error.rs
@@ -343,6 +343,19 @@ pub enum Error {
         location: Location,
     },
 
+    #[snafu(display(
+        "Failed to create default value for column {} of region {}",
+        column,
+        region_id
+    ))]
+    CreateDefault {
+        region_id: RegionId,
+        column: String,
+        source: datatypes::error::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
     #[snafu(display("Unexpected request: {}", reason))]
     UnexpectedRequest {
         reason: String,
@@ -394,7 +407,8 @@ impl ErrorExt for Error {
             | UnsupportedAlterKind { .. }
             | UnsupportedRemapManifestsRequest { .. }
             | UnsupportedSyncRegionFromRequest { .. }
-            | InvalidRequest { .. } => StatusCode::InvalidArguments,
+            | InvalidRequest { .. }
+            | CreateDefault { .. } => StatusCode::InvalidArguments,
 
             ForbiddenPhysicalAlter { .. }
             | UnsupportedRegionRequest { .. }
diff --git a/src/metric-engine/src/lib.rs b/src/metric-engine/src/lib.rs
index 30daa80b91..d209eb7588 100644
--- a/src/metric-engine/src/lib.rs
+++ b/src/metric-engine/src/lib.rs
@@ -50,8 +50,9 @@
 //!              └─────────────────────┘
 //! ```
 
-#![feature(assert_matches)]
+#![recursion_limit = "256"]
 
+pub mod batch_modifier;
 pub mod config;
 mod data_region;
 pub mod engine;
diff --git a/src/metric-engine/src/metadata_region.rs b/src/metric-engine/src/metadata_region.rs
index d84cbad946..76abaacf10 100644
--- a/src/metric-engine/src/metadata_region.rs
+++ b/src/metric-engine/src/metadata_region.rs
@@ -349,8 +349,9 @@ impl MetadataRegion {
                 METADATA_SCHEMA_VALUE_COLUMN_INDEX,
             ]
         };
+        let projection_input = Some(projection.into());
         ScanRequest {
-            projection: Some(projection),
+            projection_input,
             filters: vec![filter_expr],
             ..Default::default()
         }
@@ -361,8 +362,9 @@ impl MetadataRegion {
             METADATA_SCHEMA_KEY_COLUMN_INDEX,
             METADATA_SCHEMA_VALUE_COLUMN_INDEX,
         ];
+        let projection_input = Some(projection.into());
         ScanRequest {
-            projection: Some(projection),
+            projection_input,
             ..Default::default()
         }
     }
@@ -679,8 +681,9 @@ impl MetadataRegion {
         let filter_expr = datafusion::prelude::col(METADATA_SCHEMA_KEY_COLUMN_NAME)
             .eq(datafusion::prelude::lit(key));
 
+        let projection_input = Some(vec![METADATA_SCHEMA_VALUE_COLUMN_INDEX].into());
         let scan_req = ScanRequest {
-            projection: Some(vec![METADATA_SCHEMA_VALUE_COLUMN_INDEX]),
+            projection_input,
             filters: vec![filter_expr],
             ..Default::default()
         };
diff --git a/src/metric-engine/src/repeated_task.rs b/src/metric-engine/src/repeated_task.rs
index fa382d7844..d0299035d9 100644
--- a/src/metric-engine/src/repeated_task.rs
+++ b/src/metric-engine/src/repeated_task.rs
@@ -64,9 +64,7 @@ impl TaskFunction<Error> for FlushMetadataRegionTask {
                 .mito
                 .handle_request(
                     metadata_region_id,
-                    RegionRequest::Flush(RegionFlushRequest {
-                        row_group_size: None,
-                    }),
+                    RegionRequest::Flush(RegionFlushRequest::default()),
                 )
                 .await
             {
@@ -85,7 +83,7 @@ impl TaskFunction<Error> for FlushMetadataRegionTask {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::time::Duration;
 
     use store_api::region_engine::{RegionEngine, RegionManifestInfo};
diff --git a/src/metric-engine/src/row_modifier.rs b/src/metric-engine/src/row_modifier.rs
index d5ed1cc9b0..7d68f94aa0 100644
--- a/src/metric-engine/src/row_modifier.rs
+++ b/src/metric-engine/src/row_modifier.rs
@@ -23,6 +23,7 @@ use mito_codec::row_converter::SparsePrimaryKeyCodec;
 use smallvec::SmallVec;
 use snafu::ResultExt;
 use store_api::codec::PrimaryKeyEncoding;
+use store_api::metadata::ColumnMetadata;
 use store_api::metric_engine_consts::{
     DATA_SCHEMA_TABLE_ID_COLUMN_NAME, DATA_SCHEMA_TSID_COLUMN_NAME,
 };
@@ -264,7 +265,10 @@ struct IterIndex {
 }
 
 impl IterIndex {
-    fn new(row_schema: &[ColumnSchema], name_to_column_id: &HashMap<String, ColumnId>) -> Self {
+    fn new(
+        row_schema: &[ColumnSchema],
+        physical_columns: &HashMap<String, ColumnMetadata>,
+    ) -> Self {
         let mut reserved_indices = SmallVec::<[ValueIndex; 2]>::new();
         // Uses BTreeMap to keep the primary key column name order (lexicographical)
         let mut primary_key_indices = BTreeMap::new();
@@ -290,7 +294,10 @@ impl IterIndex {
                         primary_key_indices.insert(
                             col.column_name.as_str(),
                             ValueIndex {
-                                column_id: *name_to_column_id.get(&col.column_name).unwrap(),
+                                column_id: physical_columns
+                                    .get(&col.column_name)
+                                    .unwrap()
+                                    .column_id,
                                 index: idx,
                             },
                         );
@@ -298,13 +305,13 @@ impl IterIndex {
                 },
                 SemanticType::Field => {
                     field_indices.push(ValueIndex {
-                        column_id: *name_to_column_id.get(&col.column_name).unwrap(),
+                        column_id: physical_columns.get(&col.column_name).unwrap().column_id,
                         index: idx,
                     });
                 }
                 SemanticType::Timestamp => {
                     ts_index = Some(ValueIndex {
-                        column_id: *name_to_column_id.get(&col.column_name).unwrap(),
+                        column_id: physical_columns.get(&col.column_name).unwrap().column_id,
                         index: idx,
                     });
                 }
@@ -338,8 +345,8 @@ pub struct RowsIter {
 }
 
 impl RowsIter {
-    pub fn new(rows: Rows, name_to_column_id: &HashMap<String, ColumnId>) -> Self {
-        let index: IterIndex = IterIndex::new(&rows.schema, name_to_column_id);
+    pub fn new(rows: Rows, physical_columns: &HashMap<String, ColumnMetadata>) -> Self {
+        let index: IterIndex = IterIndex::new(&rows.schema, physical_columns);
         Self { rows, index }
     }
 
@@ -455,8 +462,23 @@ mod tests {
         }
     }
 
-    fn test_name_to_column_id() -> HashMap<String, ColumnId> {
-        HashMap::from([("namespace".to_string(), 1), ("host".to_string(), 2)])
+    fn make_info(name: &str, column_id: ColumnId) -> ColumnMetadata {
+        ColumnMetadata {
+            column_schema: datatypes::schema::ColumnSchema::new(
+                name.to_string(),
+                datatypes::prelude::ConcreteDataType::string_datatype(),
+                false,
+            ),
+            semantic_type: SemanticType::Tag,
+            column_id,
+        }
+    }
+
+    fn test_name_to_column_id() -> HashMap<String, ColumnMetadata> {
+        HashMap::from([
+            ("namespace".to_string(), make_info("namespace", 1)),
+            ("host".to_string(), make_info("host", 2)),
+        ])
     }
 
     #[test]
@@ -657,11 +679,11 @@ mod tests {
     }
 
     /// Helper function to create a name_to_column_id map
-    fn create_name_to_column_id(labels: &[&str]) -> HashMap<String, ColumnId> {
+    fn create_name_to_column_id(labels: &[&str]) -> HashMap<String, ColumnMetadata> {
         labels
             .iter()
             .enumerate()
-            .map(|(idx, name)| (name.to_string(), idx as ColumnId + 1))
+            .map(|(idx, name)| (name.to_string(), make_info(name, idx as ColumnId + 1)))
             .collect()
     }
 
@@ -692,7 +714,7 @@ mod tests {
     fn extract_tsid(
         schema: Vec<ColumnSchema>,
         row: Row,
-        name_to_column_id: &HashMap<String, ColumnId>,
+        name_to_column_id: &HashMap<String, ColumnMetadata>,
         table_id: TableId,
     ) -> u64 {
         let rows = Rows {
diff --git a/src/metric-engine/src/test_util.rs b/src/metric-engine/src/test_util.rs
index d81240d47f..ec55a01903 100644
--- a/src/metric-engine/src/test_util.rs
+++ b/src/metric-engine/src/test_util.rs
@@ -76,6 +76,22 @@ impl TestEnv {
         }
     }
 
+    /// Returns a new env with specific `prefix`, `mito_config`, and `config` for test.
+    pub async fn with_mito_config(
+        prefix: &str,
+        mito_config: MitoConfig,
+        config: EngineConfig,
+    ) -> Self {
+        let mut mito_env = MitoTestEnv::with_prefix(prefix).await;
+        let mito = mito_env.create_engine(mito_config).await;
+        let metric = MetricEngine::try_new(mito.clone(), config).unwrap();
+        Self {
+            mito_env,
+            mito,
+            metric,
+        }
+    }
+
     /// Returns a new env with specific `prefix` and `mito_env` for test.
     pub async fn with_mito_env(mut mito_env: MitoTestEnv) -> Self {
         let mito = mito_env.create_engine(MitoConfig::default()).await;
@@ -167,7 +183,7 @@ impl TestEnv {
             primary_key: vec![],
             options: [(PHYSICAL_TABLE_METADATA_KEY.to_string(), String::new())]
                 .into_iter()
-                .chain(options.into_iter())
+                .chain(options)
                 .collect(),
             table_dir: table_dir.to_string(),
             path_type: PathType::Bare, // Use Bare path type for engine regions
diff --git a/src/mito-codec/Cargo.toml b/src/mito-codec/Cargo.toml
index 07d64482e0..c5a6625cc2 100644
--- a/src/mito-codec/Cargo.toml
+++ b/src/mito-codec/Cargo.toml
@@ -15,7 +15,6 @@ common-base.workspace = true
 common-decimal.workspace = true
 common-error.workspace = true
 common-macro.workspace = true
-common-query.workspace = true
 common-recordbatch.workspace = true
 common-telemetry.workspace = true
 common-time.workspace = true
@@ -27,6 +26,7 @@ snafu.workspace = true
 store-api.workspace = true
 
 [dev-dependencies]
+common-query.workspace = true
 criterion = "0.7"
 datafusion-common.workspace = true
 datafusion-expr.workspace = true
diff --git a/src/mito-codec/benches/bench_primary_key_filter.rs b/src/mito-codec/benches/bench_primary_key_filter.rs
index 47158f87f8..528c374761 100644
--- a/src/mito-codec/benches/bench_primary_key_filter.rs
+++ b/src/mito-codec/benches/bench_primary_key_filter.rs
@@ -25,7 +25,8 @@ use datatypes::prelude::ConcreteDataType;
 use datatypes::schema::ColumnSchema;
 use datatypes::value::{Value, ValueRef};
 use mito_codec::row_converter::{
-    DensePrimaryKeyCodec, PrimaryKeyCodec, PrimaryKeyCodecExt, SparsePrimaryKeyCodec,
+    DensePrimaryKeyCodec, PrimaryKeyCodec, PrimaryKeyCodecExt, SparseOffsetsCache,
+    SparsePrimaryKeyCodec,
 };
 use store_api::metadata::{ColumnMetadata, RegionMetadataBuilder, RegionMetadataRef};
 use store_api::storage::{ColumnId, RegionId};
@@ -182,7 +183,7 @@ fn matches_sparse_scalar(
     codec: &SparsePrimaryKeyCodec,
     filters: &[SimpleFilterEvaluator],
     pk: &[u8],
-    offsets_map: &mut std::collections::HashMap<ColumnId, usize>,
+    offsets_map: &mut SparseOffsetsCache,
 ) -> bool {
     offsets_map.clear();
     if filters.is_empty() || metadata.primary_key.is_empty() {
@@ -246,18 +247,18 @@ fn bench_primary_key_filter(c: &mut Criterion) {
 
         let dense_pk = encode_dense_pk(&metadata, &row);
         let dense_codec = DensePrimaryKeyCodec::new(&metadata);
-        let mut dense_fast = dense_codec.primary_key_filter(&metadata, filters.clone());
+        let mut dense_fast = dense_codec.primary_key_filter(&metadata, filters.clone(), false);
         let mut dense_offsets = Vec::new();
 
         let sparse_pk = encode_sparse_pk(&metadata, &row);
         let sparse_codec = SparsePrimaryKeyCodec::new(&metadata);
-        let mut sparse_fast = sparse_codec.primary_key_filter(&metadata, filters.clone());
-        let mut sparse_offsets = std::collections::HashMap::new();
+        let mut sparse_fast = sparse_codec.primary_key_filter(&metadata, filters.clone(), false);
+        let mut sparse_offsets = SparseOffsetsCache::new();
 
         let mut group = c.benchmark_group(format!("primary_key_filter/{case_name}"));
 
         group.bench_function("dense/fast", |b| {
-            b.iter(|| black_box(dense_fast.matches(black_box(&dense_pk))))
+            b.iter(|| black_box(dense_fast.matches(black_box(&dense_pk)).unwrap()))
         });
         group.bench_function("dense/scalar", |b| {
             b.iter(|| {
@@ -272,7 +273,7 @@ fn bench_primary_key_filter(c: &mut Criterion) {
         });
 
         group.bench_function("sparse/fast", |b| {
-            b.iter(|| black_box(sparse_fast.matches(black_box(&sparse_pk))))
+            b.iter(|| black_box(sparse_fast.matches(black_box(&sparse_pk)).unwrap()))
         });
         group.bench_function("sparse/scalar", |b| {
             b.iter(|| {
diff --git a/src/mito-codec/benches/bench_sparse_encoding.rs b/src/mito-codec/benches/bench_sparse_encoding.rs
index 5ad8964c0c..a0f978fd9f 100644
--- a/src/mito-codec/benches/bench_sparse_encoding.rs
+++ b/src/mito-codec/benches/bench_sparse_encoding.rs
@@ -12,13 +12,17 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+use std::collections::HashMap;
 use std::hint::black_box;
 
 use bytes::Bytes;
-use criterion::{Criterion, criterion_group, criterion_main};
+use criterion::measurement::WallTime;
+use criterion::{BenchmarkGroup, Criterion, criterion_group, criterion_main};
 use datatypes::prelude::ValueRef;
-use mito_codec::row_converter::SparsePrimaryKeyCodec;
+use datatypes::value::Value;
 use mito_codec::row_converter::sparse::{RESERVED_COLUMN_ID_TABLE_ID, RESERVED_COLUMN_ID_TSID};
+use mito_codec::row_converter::{SparseOffsetsCache, SparsePrimaryKeyCodec};
+use store_api::storage::ColumnId;
 
 fn encode_sparse(c: &mut Criterion) {
     let num_tags = 10;
@@ -82,5 +86,424 @@ fn encode_sparse(c: &mut Criterion) {
     group.finish();
 }
 
-criterion_group!(benches, encode_sparse);
+/// Encodes a primary key with the given number of tags.
+fn encode_pk(num_tags: u32) -> (SparsePrimaryKeyCodec, Vec<u8>) {
+    // Use schemaless() so all columns (including reserved ones) are recognized during parsing.
+    let codec = SparsePrimaryKeyCodec::schemaless();
+    let dummy_table_id = 1024u32;
+    let dummy_tsid = 42u64;
+
+    let tags: Vec<_> = (0..num_tags)
+        .map(|idx| {
+            let tag_value = idx.to_string().repeat(10);
+            (idx, Bytes::copy_from_slice(tag_value.as_bytes()))
+        })
+        .collect();
+
+    let mut buffer = Vec::new();
+    codec
+        .encode_internal(dummy_table_id, dummy_tsid, &mut buffer)
+        .unwrap();
+    codec
+        .encode_raw_tag_value(tags.iter().map(|(c, b)| (*c, &b[..])), &mut buffer)
+        .unwrap();
+
+    (codec, buffer)
+}
+
+fn bench_has_column(c: &mut Criterion) {
+    for num_tags in [5, 10, 50, 100] {
+        let (codec, pk) = encode_pk(num_tags);
+        let mut group = c.benchmark_group(format!("has_column/{num_tags}_tags"));
+
+        group.bench_function("table_id", |b| {
+            b.iter(|| {
+                let mut offsets_map = SparseOffsetsCache::new();
+                black_box(codec.has_column(&pk, &mut offsets_map, RESERVED_COLUMN_ID_TABLE_ID));
+            });
+        });
+
+        group.bench_function("tsid", |b| {
+            b.iter(|| {
+                let mut offsets_map = SparseOffsetsCache::new();
+                black_box(codec.has_column(&pk, &mut offsets_map, RESERVED_COLUMN_ID_TSID));
+            });
+        });
+
+        group.bench_function("first_tag", |b| {
+            b.iter(|| {
+                let mut offsets_map = SparseOffsetsCache::new();
+                black_box(codec.has_column(&pk, &mut offsets_map, 0));
+            });
+        });
+
+        group.bench_function("middle_tag", |b| {
+            b.iter(|| {
+                let mut offsets_map = SparseOffsetsCache::new();
+                black_box(codec.has_column(&pk, &mut offsets_map, num_tags / 2));
+            });
+        });
+
+        group.bench_function("last_tag", |b| {
+            b.iter(|| {
+                let mut offsets_map = SparseOffsetsCache::new();
+                black_box(codec.has_column(&pk, &mut offsets_map, num_tags - 1));
+            });
+        });
+
+        group.finish();
+    }
+}
+
+/// Benchmarks Vec linear scan vs HashMap lookup at various collection sizes
+/// to find the optimal `SPARSE_OFFSETS_INLINE_CAP` threshold.
+fn bench_inline_threshold(c: &mut Criterion) {
+    for size in [4, 8, 12, 16, 20, 24, 32, 48, 64] {
+        let vec: Vec<(u32, usize)> = (0..size).map(|i| (i as u32, i * 8)).collect();
+        let map: HashMap<u32, usize> = vec.iter().copied().collect();
+
+        let last_id = (size - 1) as u32;
+        let missing_id = size as u32;
+
+        let mut group = c.benchmark_group(format!("inline_threshold/{size}"));
+
+        // Vec: best case (first element)
+        group.bench_function("vec_first", |b| {
+            b.iter(|| {
+                let target = black_box(0u32);
+                for entry in &vec {
+                    if entry.0 == target {
+                        return black_box(Some(entry.1));
+                    }
+                }
+                black_box(None)
+            });
+        });
+
+        // Vec: worst case (last element)
+        group.bench_function("vec_last", |b| {
+            b.iter(|| {
+                let target = black_box(last_id);
+                for entry in &vec {
+                    if entry.0 == target {
+                        return black_box(Some(entry.1));
+                    }
+                }
+                black_box(None)
+            });
+        });
+
+        // Vec: miss
+        group.bench_function("vec_miss", |b| {
+            b.iter(|| {
+                let target = black_box(missing_id);
+                for entry in &vec {
+                    if entry.0 == target {
+                        return black_box(Some(entry.1));
+                    }
+                }
+                black_box(None)
+            });
+        });
+
+        // HashMap: hit (last element)
+        group.bench_function("map_hit", |b| {
+            b.iter(|| black_box(map.get(&black_box(last_id)).copied()));
+        });
+
+        // HashMap: miss
+        group.bench_function("map_miss", |b| {
+            b.iter(|| black_box(map.get(&black_box(missing_id)).copied()));
+        });
+
+        group.finish();
+    }
+}
+
+/// Lookup-strategy comparison for `SparseValues`-shaped collections.
+///
+/// `SparseValues` is now a `Vec<(ColumnId, Value)>` with linear-scan lookups;
+/// these benches keep the `HashMap` and `HybridValues<CAP>` baselines around
+/// to track the crossover and to make it easy to revisit the choice later.
+/// The companion `SparseOffsetsCache` uses a small-vec fast path with
+/// `SPARSE_OFFSETS_INLINE_CAP = 32` entries against a `usize` payload;
+/// `Value` is a much fatter enum, so its crossover may land elsewhere.
+/// The benches measure both the crossover (`bench_sparse_values_threshold`)
+/// and the workload impact at realistic primary-key widths
+/// (`bench_sparse_values_lookup`).
+fn sample_value(idx: u32) -> Value {
+    Value::String(idx.to_string().repeat(10).into())
+}
+
+/// Common lookup surface for the variants under test.
+trait ValueLookup {
+    fn lookup_get(&self, col: ColumnId) -> Option<&Value>;
+    fn lookup_or_null(&self, col: ColumnId) -> &Value;
+}
+
+impl ValueLookup for HashMap<ColumnId, Value> {
+    fn lookup_get(&self, col: ColumnId) -> Option<&Value> {
+        self.get(&col)
+    }
+    fn lookup_or_null(&self, col: ColumnId) -> &Value {
+        self.get(&col).unwrap_or(&Value::Null)
+    }
+}
+
+impl ValueLookup for Vec<(ColumnId, Value)> {
+    fn lookup_get(&self, col: ColumnId) -> Option<&Value> {
+        for entry in self {
+            if entry.0 == col {
+                return Some(&entry.1);
+            }
+        }
+        None
+    }
+    fn lookup_or_null(&self, col: ColumnId) -> &Value {
+        for entry in self {
+            if entry.0 == col {
+                return &entry.1;
+            }
+        }
+        &Value::Null
+    }
+}
+
+/// Inline-Vec + HashMap-overflow mirror of `SparseOffsetsCache`'s layout,
+/// parameterized so the bench can sweep candidate inline caps.
+struct HybridValues<const CAP: usize> {
+    inline: Vec<(ColumnId, Value)>,
+    overflow: HashMap<ColumnId, Value>,
+}
+
+impl<const CAP: usize> HybridValues<CAP> {
+    fn new() -> Self {
+        Self {
+            inline: Vec::with_capacity(CAP),
+            overflow: HashMap::new(),
+        }
+    }
+
+    fn insert(&mut self, col: ColumnId, value: Value) {
+        if self.inline.len() < CAP {
+            self.inline.push((col, value));
+        } else {
+            self.overflow.insert(col, value);
+        }
+    }
+}
+
+impl<const CAP: usize> ValueLookup for HybridValues<CAP> {
+    fn lookup_get(&self, col: ColumnId) -> Option<&Value> {
+        for entry in &self.inline {
+            if entry.0 == col {
+                return Some(&entry.1);
+            }
+        }
+        if self.overflow.is_empty() {
+            return None;
+        }
+        self.overflow.get(&col)
+    }
+    fn lookup_or_null(&self, col: ColumnId) -> &Value {
+        self.lookup_get(col).unwrap_or(&Value::Null)
+    }
+}
+
+fn build_hashmap_values(pairs: &[(ColumnId, Value)]) -> HashMap<ColumnId, Value> {
+    let mut m = HashMap::with_capacity(pairs.len());
+    for (c, v) in pairs {
+        m.insert(*c, v.clone());
+    }
+    m
+}
+
+fn build_vec_values(pairs: &[(ColumnId, Value)]) -> Vec<(ColumnId, Value)> {
+    pairs.to_vec()
+}
+
+fn build_hybrid_values<const CAP: usize>(pairs: &[(ColumnId, Value)]) -> HybridValues<CAP> {
+    let mut h = HybridValues::<CAP>::new();
+    for (c, v) in pairs {
+        h.insert(*c, v.clone());
+    }
+    h
+}
+
+/// Runs the five workloads (lookup_first/last/miss, iter_all, build_then_iter)
+/// for one variant inside a `sparse_values/{size}` group.
+fn run_lookup_workloads<V, B>(
+    group: &mut BenchmarkGroup<'_, WallTime>,
+    name: &str,
+    column_ids: &[ColumnId],
+    last_id: ColumnId,
+    missing_id: ColumnId,
+    build: B,
+) where
+    V: ValueLookup,
+    B: Fn() -> V,
+{
+    let built = build();
+
+    group.bench_function(format!("{name}/lookup_first"), |b| {
+        b.iter(|| {
+            let target = black_box(0 as ColumnId);
+            black_box(built.lookup_or_null(target))
+        });
+    });
+
+    group.bench_function(format!("{name}/lookup_last"), |b| {
+        b.iter(|| {
+            let target = black_box(last_id);
+            black_box(built.lookup_or_null(target))
+        });
+    });
+
+    group.bench_function(format!("{name}/lookup_miss"), |b| {
+        b.iter(|| {
+            let target = black_box(missing_id);
+            black_box(built.lookup_get(target))
+        });
+    });
+
+    group.bench_function(format!("{name}/iter_all"), |b| {
+        b.iter(|| {
+            for col in column_ids {
+                black_box(built.lookup_or_null(*col));
+            }
+        });
+    });
+
+    group.bench_function(format!("{name}/build_then_iter"), |b| {
+        b.iter(|| {
+            let c = build();
+            for col in column_ids {
+                black_box(c.lookup_or_null(*col));
+            }
+            c
+        });
+    });
+}
+
+/// Sweeps collection sizes to find the linear-scan-vs-HashMap crossover for
+/// `(ColumnId, Value)` entries — i.e. the natural inline cap for a
+/// `SparseValues`-shaped hybrid.
+fn bench_sparse_values_threshold(c: &mut Criterion) {
+    for size in [2usize, 4, 8, 12, 16, 20, 24, 32, 48, 64] {
+        let vec: Vec<(ColumnId, Value)> = (0..size as u32)
+            .map(|i| (i as ColumnId, sample_value(i)))
+            .collect();
+        let map: HashMap<ColumnId, Value> = vec.iter().cloned().collect();
+        let last_id = (size - 1) as ColumnId;
+        let missing_id = size as ColumnId;
+
+        let mut group = c.benchmark_group(format!("sparse_values_threshold/{size}"));
+
+        group.bench_function("vec_first", |b| {
+            b.iter(|| {
+                let target = black_box(0 as ColumnId);
+                for entry in &vec {
+                    if entry.0 == target {
+                        return black_box(Some(&entry.1));
+                    }
+                }
+                black_box(None)
+            });
+        });
+
+        group.bench_function("vec_last", |b| {
+            b.iter(|| {
+                let target = black_box(last_id);
+                for entry in &vec {
+                    if entry.0 == target {
+                        return black_box(Some(&entry.1));
+                    }
+                }
+                black_box(None)
+            });
+        });
+
+        group.bench_function("vec_miss", |b| {
+            b.iter(|| {
+                let target = black_box(missing_id);
+                for entry in &vec {
+                    if entry.0 == target {
+                        return black_box(Some(&entry.1));
+                    }
+                }
+                black_box(None)
+            });
+        });
+
+        group.bench_function("map_hit_last", |b| {
+            b.iter(|| black_box(map.get(&black_box(last_id))));
+        });
+
+        group.bench_function("map_miss", |b| {
+            b.iter(|| black_box(map.get(&black_box(missing_id))));
+        });
+
+        group.finish();
+    }
+}
+
+/// Workload comparison at realistic primary-key widths. Compares the current
+/// `Vec`-backed `SparseValues` against a `HashMap` and two hybrid candidates.
+/// The hybrid caps are picked to bracket the likely crossover (smaller than
+/// the offsets-cache's 32, since `Value` is fatter than `usize`); refine
+/// after reading `bench_sparse_values_threshold`.
+fn bench_sparse_values_lookup(c: &mut Criterion) {
+    for size in [2usize, 4, 8, 16, 32, 50] {
+        let pairs: Vec<(ColumnId, Value)> = (0..size as u32)
+            .map(|i| (i as ColumnId, sample_value(i)))
+            .collect();
+        let column_ids: Vec<ColumnId> = pairs.iter().map(|(c, _)| *c).collect();
+        let last_id = (size - 1) as ColumnId;
+        let missing_id = size as ColumnId;
+
+        let mut group = c.benchmark_group(format!("sparse_values/{size}"));
+
+        run_lookup_workloads(
+            &mut group,
+            "hashmap",
+            &column_ids,
+            last_id,
+            missing_id,
+            || build_hashmap_values(&pairs),
+        );
+
+        run_lookup_workloads(&mut group, "vec", &column_ids, last_id, missing_id, || {
+            build_vec_values(&pairs)
+        });
+
+        run_lookup_workloads(
+            &mut group,
+            "hybrid_8",
+            &column_ids,
+            last_id,
+            missing_id,
+            || build_hybrid_values::<8>(&pairs),
+        );
+
+        run_lookup_workloads(
+            &mut group,
+            "hybrid_32",
+            &column_ids,
+            last_id,
+            missing_id,
+            || build_hybrid_values::<32>(&pairs),
+        );
+
+        group.finish();
+    }
+}
+
+criterion_group!(
+    benches,
+    encode_sparse,
+    bench_has_column,
+    bench_inline_threshold,
+    bench_sparse_values_threshold,
+    bench_sparse_values_lookup
+);
 criterion_main!(benches);
diff --git a/src/mito-codec/src/error.rs b/src/mito-codec/src/error.rs
index 1be0074b1e..78a656415a 100644
--- a/src/mito-codec/src/error.rs
+++ b/src/mito-codec/src/error.rs
@@ -72,6 +72,14 @@ pub enum Error {
         #[snafu(implicit)]
         location: Location,
     },
+
+    #[snafu(display("Failed to evaluate filter"))]
+    EvaluateFilter {
+        #[snafu(source(from(common_recordbatch::error::Error, Box::new)))]
+        source: Box<common_recordbatch::error::Error>,
+        #[snafu(implicit)]
+        location: Location,
+    },
 }
 
 pub type Result<T, E = Error> = std::result::Result<T, E>;
@@ -86,6 +94,7 @@ impl ErrorExt for Error {
                 StatusCode::InvalidArguments
             }
             NotSupportedField { .. } | UnsupportedOperation { .. } => StatusCode::Unsupported,
+            EvaluateFilter { source, .. } => source.status_code(),
         }
     }
 
diff --git a/src/mito-codec/src/key_values.rs b/src/mito-codec/src/key_values.rs
index 5afacc3718..d66110bacf 100644
--- a/src/mito-codec/src/key_values.rs
+++ b/src/mito-codec/src/key_values.rs
@@ -431,11 +431,9 @@ mod tests {
         values: &[Option<i64>],
     ) {
         assert_eq!(num_rows, kvs.num_rows());
-        let mut expect_seq = START_SEQ;
         let expect_ts = ValueRef::Int64(ts);
-        for kv in kvs.iter() {
+        for (expect_seq, kv) in (START_SEQ..).zip(kvs.iter()) {
             assert_eq!(expect_seq, kv.sequence());
-            expect_seq += 1;
             assert_eq!(OpType::Put, kv.op_type);
             assert_eq!(keys.len(), kv.num_primary_keys());
             assert_eq!(values.len(), kv.num_fields());
diff --git a/src/mito-codec/src/primary_key_filter.rs b/src/mito-codec/src/primary_key_filter.rs
index 189c7a08cd..2450a6e44a 100644
--- a/src/mito-codec/src/primary_key_filter.rs
+++ b/src/mito-codec/src/primary_key_filter.rs
@@ -12,7 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::collections::HashMap;
 use std::sync::Arc;
 
 use api::v1::SemanticType;
@@ -20,13 +19,14 @@ use common_recordbatch::filter::SimpleFilterEvaluator;
 use datatypes::data_type::ConcreteDataType;
 use datatypes::value::Value;
 use memcomparable::Serializer;
+use snafu::ResultExt;
 use store_api::metadata::RegionMetadataRef;
 use store_api::metric_engine_consts::DATA_SCHEMA_TABLE_ID_COLUMN_NAME;
 use store_api::storage::ColumnId;
 
-use crate::error::Result;
+use crate::error::{EvaluateFilterSnafu, Result};
 use crate::row_converter::{
-    DensePrimaryKeyCodec, PrimaryKeyFilter, SortField, SparsePrimaryKeyCodec,
+    DensePrimaryKeyCodec, PrimaryKeyFilter, SortField, SparseOffsetsCache, SparsePrimaryKeyCodec,
 };
 
 /// Returns true if this is a partition column for metrics in the memtable.
@@ -41,8 +41,12 @@ struct PrimaryKeyFilterInner {
 }
 
 impl PrimaryKeyFilterInner {
-    fn new(metadata: RegionMetadataRef, filters: Arc<Vec<SimpleFilterEvaluator>>) -> Self {
-        let compiled_filters = Self::compile_filters(&metadata, &filters);
+    fn new(
+        metadata: RegionMetadataRef,
+        filters: Arc<Vec<SimpleFilterEvaluator>>,
+        skip_partition_column: bool,
+    ) -> Self {
+        let compiled_filters = Self::compile_filters(&metadata, &filters, skip_partition_column);
         Self {
             filters,
             compiled_filters,
@@ -52,6 +56,7 @@ impl PrimaryKeyFilterInner {
     fn compile_filters(
         metadata: &RegionMetadataRef,
         filters: &[SimpleFilterEvaluator],
+        skip_partition_column: bool,
     ) -> Vec<CompiledPrimaryKeyFilter> {
         if filters.is_empty() || metadata.primary_key.is_empty() {
             return Vec::new();
@@ -59,7 +64,7 @@ impl PrimaryKeyFilterInner {
 
         let mut compiled_filters = Vec::with_capacity(filters.len());
         for (filter_idx, filter) in filters.iter().enumerate() {
-            if is_partition_column(filter.column_name()) {
+            if skip_partition_column && is_partition_column(filter.column_name()) {
                 continue;
             }
 
@@ -91,43 +96,36 @@ impl PrimaryKeyFilterInner {
         compiled_filters
     }
 
-    fn evaluate_filters<'a>(&self, accessor: &mut impl PrimaryKeyValueAccessor<'a>) -> bool {
+    fn evaluate_filters<'a>(
+        &self,
+        accessor: &mut impl PrimaryKeyValueAccessor<'a>,
+    ) -> Result<bool> {
         if self.compiled_filters.is_empty() {
-            return true;
+            return Ok(true);
         }
 
         for compiled in &self.compiled_filters {
             let filter = &self.filters[compiled.filter_idx];
 
             let passed = if let Some(fast_path) = &compiled.fast_path {
-                let encoded_value = match accessor.encoded_value(compiled) {
-                    Ok(v) => v,
-                    Err(e) => {
-                        common_telemetry::error!(e; "Failed to decode primary key");
-                        return true;
-                    }
-                };
+                let encoded_value = accessor.encoded_value(compiled)?;
                 fast_path.matches(encoded_value)
             } else {
-                let value = match accessor.decode_value(compiled) {
-                    Ok(v) => v,
-                    Err(e) => {
-                        common_telemetry::error!(e; "Failed to decode primary key");
-                        return true;
-                    }
-                };
+                let value = accessor.decode_value(compiled)?;
 
                 // Safety: arrow schema and datatypes are constructed from the same source.
                 let scalar_value = value.try_to_scalar_value(&compiled.data_type).unwrap();
-                filter.evaluate_scalar(&scalar_value).unwrap_or(true)
+                filter
+                    .evaluate_scalar(&scalar_value)
+                    .context(EvaluateFilterSnafu)?
             };
 
             if !passed {
-                return false;
+                return Ok(false);
             }
         }
 
-        true
+        Ok(true)
     }
 }
 
@@ -258,9 +256,10 @@ impl DensePrimaryKeyFilter {
         metadata: RegionMetadataRef,
         filters: Arc<Vec<SimpleFilterEvaluator>>,
         codec: DensePrimaryKeyCodec,
+        skip_partition_column: bool,
     ) -> Self {
         Self {
-            inner: PrimaryKeyFilterInner::new(metadata, filters),
+            inner: PrimaryKeyFilterInner::new(metadata, filters, skip_partition_column),
             codec,
             offsets_buf: Vec::new(),
         }
@@ -268,7 +267,7 @@ impl DensePrimaryKeyFilter {
 }
 
 impl PrimaryKeyFilter for DensePrimaryKeyFilter {
-    fn matches(&mut self, pk: &[u8]) -> bool {
+    fn matches(&mut self, pk: &[u8]) -> Result<bool> {
         self.offsets_buf.clear();
         let mut accessor = DensePrimaryKeyValueAccessor {
             pk,
@@ -303,7 +302,7 @@ impl<'a> PrimaryKeyValueAccessor<'a> for DensePrimaryKeyValueAccessor<'a, '_> {
 pub struct SparsePrimaryKeyFilter {
     inner: PrimaryKeyFilterInner,
     codec: SparsePrimaryKeyCodec,
-    offsets_map: HashMap<ColumnId, usize>,
+    offsets_cache: SparseOffsetsCache,
 }
 
 impl SparsePrimaryKeyFilter {
@@ -311,22 +310,23 @@ impl SparsePrimaryKeyFilter {
         metadata: RegionMetadataRef,
         filters: Arc<Vec<SimpleFilterEvaluator>>,
         codec: SparsePrimaryKeyCodec,
+        skip_partition_column: bool,
     ) -> Self {
         Self {
-            inner: PrimaryKeyFilterInner::new(metadata, filters),
+            inner: PrimaryKeyFilterInner::new(metadata, filters, skip_partition_column),
             codec,
-            offsets_map: HashMap::new(),
+            offsets_cache: SparseOffsetsCache::new(),
         }
     }
 }
 
 impl PrimaryKeyFilter for SparsePrimaryKeyFilter {
-    fn matches(&mut self, pk: &[u8]) -> bool {
-        self.offsets_map.clear();
+    fn matches(&mut self, pk: &[u8]) -> Result<bool> {
+        self.offsets_cache.clear();
         let mut accessor = SparsePrimaryKeyValueAccessor {
             pk,
             codec: &self.codec,
-            offsets_map: &mut self.offsets_map,
+            offsets_cache: &mut self.offsets_cache,
         };
         self.inner.evaluate_filters(&mut accessor)
     }
@@ -335,19 +335,19 @@ impl PrimaryKeyFilter for SparsePrimaryKeyFilter {
 struct SparsePrimaryKeyValueAccessor<'a, 'b> {
     pk: &'a [u8],
     codec: &'b SparsePrimaryKeyCodec,
-    offsets_map: &'b mut HashMap<ColumnId, usize>,
+    offsets_cache: &'b mut SparseOffsetsCache,
 }
 
 impl<'a> PrimaryKeyValueAccessor<'a> for SparsePrimaryKeyValueAccessor<'a, '_> {
     fn encoded_value(&mut self, filter: &CompiledPrimaryKeyFilter) -> Result<Option<&'a [u8]>> {
         self.codec
-            .encoded_value_for_column(self.pk, self.offsets_map, filter.column_id)
+            .encoded_value_for_column(self.pk, self.offsets_cache, filter.column_id)
     }
 
     fn decode_value(&mut self, filter: &CompiledPrimaryKeyFilter) -> Result<Value> {
         if let Some(offset) = self
             .codec
-            .has_column(self.pk, self.offsets_map, filter.column_id)
+            .has_column(self.pk, self.offsets_cache, filter.column_id)
         {
             self.codec
                 .decode_value_at(self.pk, offset, filter.column_id)
@@ -369,6 +369,7 @@ mod tests {
     use datatypes::schema::ColumnSchema;
     use datatypes::value::{OrderedFloat, ValueRef};
     use store_api::metadata::{ColumnMetadata, RegionMetadataBuilder};
+    use store_api::metric_engine_consts::DATA_SCHEMA_TABLE_ID_COLUMN_NAME;
     use store_api::storage::{ColumnId, RegionId};
 
     use super::*;
@@ -423,6 +424,36 @@ mod tests {
         Arc::new(metadata)
     }
 
+    fn setup_partitioned_metadata() -> RegionMetadataRef {
+        let mut builder = RegionMetadataBuilder::new(RegionId::new(1, 1));
+        builder
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    DATA_SCHEMA_TABLE_ID_COLUMN_NAME,
+                    ConcreteDataType::uint32_datatype(),
+                    false,
+                ),
+                semantic_type: SemanticType::Tag,
+                column_id: 10,
+            })
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new("tag", ConcreteDataType::string_datatype(), true),
+                semantic_type: SemanticType::Tag,
+                column_id: 1,
+            })
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    greptime_timestamp(),
+                    ConcreteDataType::timestamp_nanosecond_datatype(),
+                    false,
+                ),
+                semantic_type: SemanticType::Timestamp,
+                column_id: 2,
+            })
+            .primary_key(vec![10, 1]);
+        Arc::new(builder.build().unwrap())
+    }
+
     fn create_test_row() -> Vec<(ColumnId, ValueRef<'static>)> {
         vec![
             (1, ValueRef::String("greptime-frontend-6989d9899-22222")),
@@ -450,10 +481,13 @@ mod tests {
 
     fn encode_sparse_pk(
         metadata: &RegionMetadataRef,
+        table_id: u32,
+        tsid: u64,
         row: Vec<(ColumnId, ValueRef<'static>)>,
     ) -> Vec<u8> {
         let codec = SparsePrimaryKeyCodec::new(metadata);
         let mut pk = Vec::new();
+        codec.encode_internal(table_id, tsid, &mut pk).unwrap();
         codec.encode_to_vec(row.into_iter(), &mut pk).unwrap();
         pk
     }
@@ -477,10 +511,10 @@ mod tests {
             "pod",
             "greptime-frontend-6989d9899-22222",
         )]);
-        let pk = encode_sparse_pk(&metadata, create_test_row());
+        let pk = encode_sparse_pk(&metadata, 1, 0, create_test_row());
         let codec = SparsePrimaryKeyCodec::new(&metadata);
-        let mut filter = SparsePrimaryKeyFilter::new(metadata, filters, codec);
-        assert!(filter.matches(&pk));
+        let mut filter = SparsePrimaryKeyFilter::new(metadata, filters, codec, false);
+        assert!(filter.matches(&pk).unwrap());
     }
 
     #[test]
@@ -490,10 +524,10 @@ mod tests {
             "pod",
             "greptime-frontend-6989d9899-22223",
         )]);
-        let pk = encode_sparse_pk(&metadata, create_test_row());
+        let pk = encode_sparse_pk(&metadata, 1, 0, create_test_row());
         let codec = SparsePrimaryKeyCodec::new(&metadata);
-        let mut filter = SparsePrimaryKeyFilter::new(metadata, filters, codec);
-        assert!(!filter.matches(&pk));
+        let mut filter = SparsePrimaryKeyFilter::new(metadata, filters, codec, false);
+        assert!(!filter.matches(&pk).unwrap());
     }
 
     #[test]
@@ -503,10 +537,10 @@ mod tests {
             "non-exist-label",
             "greptime-frontend-6989d9899-22222",
         )]);
-        let pk = encode_sparse_pk(&metadata, create_test_row());
+        let pk = encode_sparse_pk(&metadata, 1, 0, create_test_row());
         let codec = SparsePrimaryKeyCodec::new(&metadata);
-        let mut filter = SparsePrimaryKeyFilter::new(metadata, filters, codec);
-        assert!(filter.matches(&pk));
+        let mut filter = SparsePrimaryKeyFilter::new(metadata, filters, codec, false);
+        assert!(filter.matches(&pk).unwrap());
     }
 
     #[test]
@@ -518,8 +552,8 @@ mod tests {
         )]);
         let pk = encode_dense_pk(&metadata, create_test_row());
         let codec = DensePrimaryKeyCodec::new(&metadata);
-        let mut filter = DensePrimaryKeyFilter::new(metadata, filters, codec);
-        assert!(filter.matches(&pk));
+        let mut filter = DensePrimaryKeyFilter::new(metadata, filters, codec, false);
+        assert!(filter.matches(&pk).unwrap());
     }
 
     #[test]
@@ -531,8 +565,8 @@ mod tests {
         )]);
         let pk = encode_dense_pk(&metadata, create_test_row());
         let codec = DensePrimaryKeyCodec::new(&metadata);
-        let mut filter = DensePrimaryKeyFilter::new(metadata, filters, codec);
-        assert!(!filter.matches(&pk));
+        let mut filter = DensePrimaryKeyFilter::new(metadata, filters, codec, false);
+        assert!(!filter.matches(&pk).unwrap());
     }
 
     #[test]
@@ -544,8 +578,8 @@ mod tests {
         )]);
         let pk = encode_dense_pk(&metadata, create_test_row());
         let codec = DensePrimaryKeyCodec::new(&metadata);
-        let mut filter = DensePrimaryKeyFilter::new(metadata, filters, codec);
-        assert!(filter.matches(&pk));
+        let mut filter = DensePrimaryKeyFilter::new(metadata, filters, codec, false);
+        assert!(filter.matches(&pk).unwrap());
     }
 
     #[test]
@@ -563,15 +597,16 @@ mod tests {
 
         for (op, value, expected) in cases {
             let filters = Arc::new(vec![create_filter_with_op("pod", op, value)]);
-            let mut filter = DensePrimaryKeyFilter::new(metadata.clone(), filters, codec.clone());
-            assert_eq!(expected, filter.matches(&pk));
+            let mut filter =
+                DensePrimaryKeyFilter::new(metadata.clone(), filters, codec.clone(), false);
+            assert_eq!(expected, filter.matches(&pk).unwrap());
         }
     }
 
     #[test]
     fn test_sparse_primary_key_filter_order_ops() {
         let metadata = setup_metadata();
-        let pk = encode_sparse_pk(&metadata, create_test_row());
+        let pk = encode_sparse_pk(&metadata, 1, 0, create_test_row());
         let codec = SparsePrimaryKeyCodec::new(&metadata);
 
         let cases = [
@@ -583,8 +618,9 @@ mod tests {
 
         for (op, value, expected) in cases {
             let filters = Arc::new(vec![create_filter_with_op("pod", op, value)]);
-            let mut filter = SparsePrimaryKeyFilter::new(metadata.clone(), filters, codec.clone());
-            assert_eq!(expected, filter.matches(&pk));
+            let mut filter =
+                SparsePrimaryKeyFilter::new(metadata.clone(), filters, codec.clone(), false);
+            assert_eq!(expected, filter.matches(&pk).unwrap());
         }
     }
 
@@ -616,8 +652,52 @@ mod tests {
             .unwrap();
 
         let filters = Arc::new(vec![create_filter_with_op("f", Operator::Eq, 0.0_f64)]);
-        let mut filter = DensePrimaryKeyFilter::new(metadata, filters, codec);
+        let mut filter = DensePrimaryKeyFilter::new(metadata, filters, codec, false);
 
-        assert!(filter.matches(&pk));
+        assert!(filter.matches(&pk).unwrap());
+    }
+
+    #[test]
+    fn test_dense_primary_key_filter_matches_partition_column_by_default() {
+        let metadata = setup_partitioned_metadata();
+        let codec = DensePrimaryKeyCodec::new(&metadata);
+        let mut pk = Vec::new();
+        codec
+            .encode_to_vec(
+                [ValueRef::UInt32(42), ValueRef::String("host-a")].into_iter(),
+                &mut pk,
+            )
+            .unwrap();
+
+        let filters = Arc::new(vec![create_filter_with_op(
+            DATA_SCHEMA_TABLE_ID_COLUMN_NAME,
+            Operator::Eq,
+            42_u32,
+        )]);
+        let mut filter = DensePrimaryKeyFilter::new(metadata, filters, codec, false);
+
+        assert!(filter.matches(&pk).unwrap());
+    }
+
+    #[test]
+    fn test_dense_primary_key_filter_can_skip_partition_column() {
+        let metadata = setup_partitioned_metadata();
+        let codec = DensePrimaryKeyCodec::new(&metadata);
+        let mut pk = Vec::new();
+        codec
+            .encode_to_vec(
+                [ValueRef::UInt32(42), ValueRef::String("host-a")].into_iter(),
+                &mut pk,
+            )
+            .unwrap();
+
+        let filters = Arc::new(vec![create_filter_with_op(
+            DATA_SCHEMA_TABLE_ID_COLUMN_NAME,
+            Operator::Eq,
+            7_u32,
+        )]);
+        let mut filter = DensePrimaryKeyFilter::new(metadata, filters, codec, true);
+
+        assert!(filter.matches(&pk).unwrap());
     }
 }
diff --git a/src/mito-codec/src/row_converter.rs b/src/mito-codec/src/row_converter.rs
index 6fe33a9ee7..0a3205ce9e 100644
--- a/src/mito-codec/src/row_converter.rs
+++ b/src/mito-codec/src/row_converter.rs
@@ -21,7 +21,7 @@ use std::sync::Arc;
 use common_recordbatch::filter::SimpleFilterEvaluator;
 use datatypes::value::{Value, ValueRef};
 pub use dense::{DensePrimaryKeyCodec, SortField};
-pub use sparse::{COLUMN_ID_ENCODE_SIZE, SparsePrimaryKeyCodec, SparseValues};
+pub use sparse::{COLUMN_ID_ENCODE_SIZE, SparseOffsetsCache, SparsePrimaryKeyCodec, SparseValues};
 use store_api::codec::PrimaryKeyEncoding;
 use store_api::metadata::{RegionMetadata, RegionMetadataRef};
 use store_api::storage::ColumnId;
@@ -53,7 +53,7 @@ pub trait PrimaryKeyCodecExt {
 
 pub trait PrimaryKeyFilter: Send + Sync {
     /// Returns true if the primary key matches the filter.
-    fn matches(&mut self, pk: &[u8]) -> bool;
+    fn matches(&mut self, pk: &[u8]) -> Result<bool>;
 }
 
 /// Composite values decoded from primary key bytes.
@@ -65,6 +65,10 @@ pub enum CompositeValues {
 
 impl CompositeValues {
     /// Extends the composite values with the given values.
+    ///
+    /// Append-only: `values` must not contain a column id already present in
+    /// the composite; otherwise the existing entry would shadow the new one on
+    /// `SparseValues` lookup.
     pub fn extend(&mut self, values: &[(ColumnId, Value)]) {
         match self {
             CompositeValues::Dense(dense_values) => {
@@ -120,6 +124,7 @@ pub trait PrimaryKeyCodec: Send + Sync + Debug {
         &self,
         metadata: &RegionMetadataRef,
         filters: Arc<Vec<SimpleFilterEvaluator>>,
+        skip_partition_column: bool,
     ) -> Box<dyn PrimaryKeyFilter>;
 
     /// Returns the estimated size of the primary key.
diff --git a/src/mito-codec/src/row_converter/dense.rs b/src/mito-codec/src/row_converter/dense.rs
index 6cc70feaea..4bc774c941 100644
--- a/src/mito-codec/src/row_converter/dense.rs
+++ b/src/mito-codec/src/row_converter/dense.rs
@@ -556,11 +556,13 @@ impl PrimaryKeyCodec for DensePrimaryKeyCodec {
         &self,
         metadata: &RegionMetadataRef,
         filters: Arc<Vec<SimpleFilterEvaluator>>,
+        skip_partition_column: bool,
     ) -> Box<dyn PrimaryKeyFilter> {
         Box::new(DensePrimaryKeyFilter::new(
             metadata.clone(),
             filters,
             self.clone(),
+            skip_partition_column,
         ))
     }
 
diff --git a/src/mito-codec/src/row_converter/sparse.rs b/src/mito-codec/src/row_converter/sparse.rs
index 4638ddcefb..9c0e488576 100644
--- a/src/mito-codec/src/row_converter/sparse.rs
+++ b/src/mito-codec/src/row_converter/sparse.rs
@@ -34,6 +34,20 @@ use crate::row_converter::dense::SortField;
 use crate::row_converter::{CompositeValues, PrimaryKeyCodec, PrimaryKeyFilter};
 
 /// A codec for sparse key of metrics.
+///
+/// ## Encoding format
+/// Each primary key is encoded as a sequence of `(column_id, value)` pairs:
+/// - The first two fields are always the reserved `table_id` (uint32) and `tsid` (uint64).
+/// - User-defined labels follow, sorted by **column name** in lexicographical order.
+/// - Null values are omitted (not encoded).
+///
+/// The `column_id` is encoded as a 4-byte big-endian integer, and the value is encoded
+/// using memcomparable serialization.
+///
+/// `decode_leftmost` always decodes the first value from the encoded bytes (i.e., the
+/// `table_id` field).
+///
+/// ## Requirements
 /// It requires the input primary key columns are sorted by the column name in lexicographical order.
 /// It encodes the column id of the physical region.
 #[derive(Clone, Debug)]
@@ -57,36 +71,56 @@ struct SparsePrimaryKeyCodecInner {
 
 /// Sparse values representation.
 ///
-/// A map of [`ColumnId`] to [`Value`].
-#[derive(Debug, Clone, PartialEq, Eq)]
+/// Callers must not insert a column id that is already present; otherwise
+/// the existing entry will shadow the newly inserted value on lookup.
+#[derive(Debug, Clone, PartialEq, Eq, Default)]
 pub struct SparseValues {
-    values: HashMap<ColumnId, Value>,
+    values: Vec<(ColumnId, Value)>,
 }
 
 impl SparseValues {
-    /// Creates a new [`SparseValues`] instance.
-    pub fn new(values: HashMap<ColumnId, Value>) -> Self {
-        Self { values }
+    /// Creates an empty [`SparseValues`].
+    pub fn new() -> Self {
+        Self { values: Vec::new() }
+    }
+
+    /// Creates an empty [`SparseValues`] with space reserved for `cap` entries.
+    pub fn with_capacity(cap: usize) -> Self {
+        Self {
+            values: Vec::with_capacity(cap),
+        }
     }
 
     /// Returns the value of the given column, or [`Value::Null`] if the column is not present.
     pub fn get_or_null(&self, column_id: ColumnId) -> &Value {
-        self.values.get(&column_id).unwrap_or(&Value::Null)
+        for (id, value) in &self.values {
+            if *id == column_id {
+                return value;
+            }
+        }
+        &Value::Null
     }
 
     /// Returns the value of the given column, or [`None`] if the column is not present.
     pub fn get(&self, column_id: &ColumnId) -> Option<&Value> {
-        self.values.get(column_id)
+        for (id, value) in &self.values {
+            if id == column_id {
+                return Some(value);
+            }
+        }
+        None
     }
 
-    /// Inserts a new value into the [`SparseValues`].
+    /// Appends a new `(column_id, value)` pair.
+    ///
+    /// Append-only: the caller must ensure `column_id` is not already present.
     pub fn insert(&mut self, column_id: ColumnId, value: Value) {
-        self.values.insert(column_id, value);
+        self.values.push((column_id, value));
     }
 
     /// Returns an iterator over all stored column id/value pairs.
     pub fn iter(&self) -> impl Iterator<Item = (&ColumnId, &Value)> {
-        self.values.iter()
+        self.values.iter().map(|(id, value)| (id, value))
     }
 }
 
@@ -97,6 +131,88 @@ pub const RESERVED_COLUMN_ID_TABLE_ID: ColumnId = ReservedColumnId::table_id();
 /// The size of the column id in the encoded sparse row.
 pub const COLUMN_ID_ENCODE_SIZE: usize = 4;
 
+// Fixed byte offsets for reserved columns in the sparse encoding.
+// Layout: [table_id_col_id: 4B][marker: 1B][table_id: 4B][tsid_col_id: 4B][marker: 1B][tsid: 8B]
+/// Byte offset to the table_id value (after its 4-byte column id).
+const TABLE_ID_VALUE_OFFSET: usize = COLUMN_ID_ENCODE_SIZE;
+/// Byte offset to the tsid value (after 9-byte table_id entry + 4-byte tsid column id).
+const TSID_VALUE_OFFSET: usize = COLUMN_ID_ENCODE_SIZE + 5 + COLUMN_ID_ENCODE_SIZE;
+/// Byte offset where tag columns start (after 9-byte table_id + 13-byte tsid entries).
+const TAGS_START_OFFSET: usize = COLUMN_ID_ENCODE_SIZE + 5 + COLUMN_ID_ENCODE_SIZE + 9;
+
+/// Inline capacity for the small-vec fast path of [`SparseOffsetsCache`].
+///
+/// Most sparse primary keys carry only a handful of tags, so a linear scan
+/// over a short `Vec` beats a `HashMap` lookup. Tags beyond this capacity
+/// spill into the overflow `HashMap`.
+const SPARSE_OFFSETS_INLINE_CAP: usize = 32;
+
+/// A lazily populated cache of tag column offsets inside a sparse primary key.
+#[derive(Debug, Clone)]
+pub struct SparseOffsetsCache {
+    /// Small-vec fast path. Reserves [`SPARSE_OFFSETS_INLINE_CAP`] slots on
+    /// the first insert.
+    inline: Vec<(ColumnId, usize)>,
+    /// Overflow for columns beyond the inline capacity. Lazily allocated.
+    overflow: HashMap<ColumnId, usize>,
+    /// Next byte position in the pk to resume parsing from.
+    cursor: usize,
+    /// True once the decoder has walked past the last tag column (or stopped
+    /// on an unknown column id); no further offsets can be discovered.
+    finished: bool,
+}
+
+impl Default for SparseOffsetsCache {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl SparseOffsetsCache {
+    pub fn new() -> Self {
+        Self {
+            inline: Vec::new(),
+            overflow: HashMap::new(),
+            cursor: TAGS_START_OFFSET,
+            finished: false,
+        }
+    }
+
+    pub fn clear(&mut self) {
+        self.inline.clear();
+        self.overflow.clear();
+        self.cursor = TAGS_START_OFFSET;
+        self.finished = false;
+    }
+
+    /// Returns the cached offset for `column_id`, if any.
+    fn get(&self, column_id: ColumnId) -> Option<usize> {
+        for entry in &self.inline {
+            if entry.0 == column_id {
+                return Some(entry.1);
+            }
+        }
+        self.overflow.get(&column_id).copied()
+    }
+
+    /// Records a new `(column_id, offset)` entry.
+    fn insert(&mut self, column_id: ColumnId, offset: usize) {
+        if self.inline.len() < SPARSE_OFFSETS_INLINE_CAP {
+            if self.inline.capacity() == 0 {
+                self.inline.reserve_exact(SPARSE_OFFSETS_INLINE_CAP);
+            }
+            self.inline.push((column_id, offset));
+        } else {
+            self.overflow.insert(column_id, offset);
+        }
+    }
+
+    #[cfg(test)]
+    fn contains(&self, column_id: ColumnId) -> bool {
+        self.get(column_id).is_some()
+    }
+}
+
 impl SparsePrimaryKeyCodec {
     /// Creates a new [`SparsePrimaryKeyCodec`] instance.
     pub fn from_columns(columns_ids: impl Iterator<Item = ColumnId>) -> Self {
@@ -233,7 +349,7 @@ impl SparsePrimaryKeyCodec {
     /// Decodes the given bytes into a [`SparseValues`].
     fn decode_sparse(&self, bytes: &[u8]) -> Result<SparseValues> {
         let mut deserializer = Deserializer::new(bytes);
-        let mut values = SparseValues::new(HashMap::new());
+        let mut values = SparseValues::with_capacity(16);
 
         let column_id = u32::deserialize(&mut deserializer).context(DeserializeFieldSnafu)?;
         let value = self.inner.table_id_field.deserialize(&mut deserializer)?;
@@ -261,31 +377,65 @@ impl SparsePrimaryKeyCodec {
     }
 
     /// Returns the offset of the given column id in the given primary key.
+    ///
+    /// The pk must start with the table_id + tsid prefix written by
+    /// `encode_internal`.
+    ///
+    /// # Panics
+    ///
+    /// Panics if `pk` is not a well-formed sparse primary key produced by
+    /// this codec (e.g. truncated or otherwise malformed bytes).
     pub fn has_column(
         &self,
         pk: &[u8],
-        offsets_map: &mut HashMap<u32, usize>,
+        cache: &mut SparseOffsetsCache,
         column_id: ColumnId,
     ) -> Option<usize> {
-        if offsets_map.is_empty() {
-            let mut deserializer = Deserializer::new(pk);
-            let mut offset = 0;
-            while deserializer.has_remaining() {
-                let column_id = u32::deserialize(&mut deserializer).unwrap();
-                offset += 4;
-                offsets_map.insert(column_id, offset);
-                let Some(field) = self.get_field(column_id) else {
-                    break;
-                };
-
-                let skip = field.skip_deserialize(pk, &mut deserializer).unwrap();
-                offset += skip;
-            }
-
-            offsets_map.get(&column_id).copied()
-        } else {
-            offsets_map.get(&column_id).copied()
+        // Decoding is lazy: on each call we only advance the cache's cursor as
+        // far as needed to answer the query. A column that has already been
+        // seen returns immediately; a column we haven't reached yet causes the
+        // parser to resume from `cache.cursor` and stop as soon as the column
+        // is located. Once the cursor walks off the end (or hits an unknown
+        // column id) the cache is marked finished, so subsequent misses are
+        // O(1).
+        // table_id and tsid are at fixed offsets.
+        match column_id {
+            RESERVED_COLUMN_ID_TABLE_ID => return Some(TABLE_ID_VALUE_OFFSET),
+            RESERVED_COLUMN_ID_TSID => return Some(TSID_VALUE_OFFSET),
+            _ => {}
         }
+
+        if let Some(offset) = cache.get(column_id) {
+            return Some(offset);
+        }
+        if cache.finished {
+            return None;
+        }
+
+        let mut deserializer = Deserializer::new(pk);
+        deserializer.advance(cache.cursor);
+        let mut offset = cache.cursor;
+        while deserializer.has_remaining() {
+            let col = u32::deserialize(&mut deserializer).unwrap();
+            offset += COLUMN_ID_ENCODE_SIZE;
+            let value_offset = offset;
+            cache.insert(col, value_offset);
+            let Some(field) = self.get_field(col) else {
+                cache.finished = true;
+                cache.cursor = offset;
+                return None;
+            };
+
+            let skip = field.skip_deserialize(pk, &mut deserializer).unwrap();
+            offset += skip;
+            cache.cursor = offset;
+            if col == column_id {
+                return Some(value_offset);
+            }
+        }
+
+        cache.finished = true;
+        None
     }
 
     /// Decode value at `offset` in `pk`.
@@ -303,10 +453,10 @@ impl SparsePrimaryKeyCodec {
     pub fn encoded_value_for_column<'a>(
         &self,
         pk: &'a [u8],
-        offsets_map: &mut HashMap<u32, usize>,
+        cache: &mut SparseOffsetsCache,
         column_id: ColumnId,
     ) -> Result<Option<&'a [u8]>> {
-        let Some(offset) = self.has_column(pk, offsets_map, column_id) else {
+        let Some(offset) = self.has_column(pk, cache, column_id) else {
             return Ok(None);
         };
 
@@ -357,11 +507,13 @@ impl PrimaryKeyCodec for SparsePrimaryKeyCodec {
         &self,
         metadata: &RegionMetadataRef,
         filters: Arc<Vec<SimpleFilterEvaluator>>,
+        skip_partition_column: bool,
     ) -> Box<dyn PrimaryKeyFilter> {
         Box::new(SparsePrimaryKeyFilter::new(
             metadata.clone(),
             filters,
             self.clone(),
+            skip_partition_column,
         ))
     }
 
@@ -521,9 +673,8 @@ mod tests {
 
     #[test]
     fn test_sparse_value_new_and_get_or_null() {
-        let mut values = HashMap::new();
-        values.insert(1, Value::Int32(42));
-        let sparse_value = SparseValues::new(values);
+        let mut sparse_value = SparseValues::new();
+        sparse_value.insert(1, Value::Int32(42));
 
         assert_eq!(sparse_value.get_or_null(1), &Value::Int32(42));
         assert_eq!(sparse_value.get_or_null(2), &Value::Null);
@@ -531,7 +682,7 @@ mod tests {
 
     #[test]
     fn test_sparse_value_insert() {
-        let mut sparse_value = SparseValues::new(HashMap::new());
+        let mut sparse_value = SparseValues::new();
         sparse_value.insert(1, Value::Int32(42));
 
         assert_eq!(sparse_value.get_or_null(1), &Value::Int32(42));
@@ -665,7 +816,7 @@ mod tests {
         codec.encode_to_vec(row.into_iter(), &mut buffer).unwrap();
         assert!(!buffer.is_empty());
 
-        let mut offsets_map = HashMap::new();
+        let mut offsets_map = SparseOffsetsCache::new();
         for column_id in [
             RESERVED_COLUMN_ID_TABLE_ID,
             RESERVED_COLUMN_ID_TSID,
@@ -683,6 +834,36 @@ mod tests {
         assert!(offset.is_none());
     }
 
+    #[test]
+    fn test_has_column_lazy_resume() {
+        let region_metadata = test_region_metadata();
+        let codec = SparsePrimaryKeyCodec::new(&region_metadata);
+        let mut buffer = Vec::new();
+        codec
+            .encode_to_vec(test_row().into_iter(), &mut buffer)
+            .unwrap();
+
+        let mut cache = SparseOffsetsCache::new();
+        // Look up an early column: only a prefix of tags is decoded.
+        assert!(codec.has_column(&buffer, &mut cache, 1).is_some());
+        assert!(!cache.finished);
+        assert!(cache.contains(1));
+        assert!(!cache.contains(5));
+
+        // A later column resumes from the cursor.
+        assert!(codec.has_column(&buffer, &mut cache, 5).is_some());
+        assert!(cache.contains(5));
+
+        // An earlier column that was already cached still resolves.
+        assert!(codec.has_column(&buffer, &mut cache, 2).is_some());
+
+        // A non-existent column walks off the end and marks the cache finished.
+        assert!(codec.has_column(&buffer, &mut cache, 999).is_none());
+        assert!(cache.finished);
+        // Further misses are O(1).
+        assert!(codec.has_column(&buffer, &mut cache, 998).is_none());
+    }
+
     #[test]
     fn test_decode_value_at() {
         let region_metadata = test_region_metadata();
@@ -693,7 +874,7 @@ mod tests {
         assert!(!buffer.is_empty());
 
         let row = test_row();
-        let mut offsets_map = HashMap::new();
+        let mut offsets_map = SparseOffsetsCache::new();
         for column_id in [
             RESERVED_COLUMN_ID_TABLE_ID,
             RESERVED_COLUMN_ID_TSID,
@@ -728,7 +909,7 @@ mod tests {
             .unwrap();
         assert!(!buffer.is_empty());
 
-        let mut offsets_map = HashMap::new();
+        let mut offsets_map = SparseOffsetsCache::new();
         for column_id in [
             RESERVED_COLUMN_ID_TABLE_ID,
             RESERVED_COLUMN_ID_TSID,
diff --git a/src/mito2/Cargo.toml b/src/mito2/Cargo.toml
index 1d7cf7b6d7..964f6eedea 100644
--- a/src/mito2/Cargo.toml
+++ b/src/mito2/Cargo.toml
@@ -16,10 +16,12 @@ workspace = true
 [dependencies]
 api.workspace = true
 aquamarine.workspace = true
+arrow-schema.workspace = true
 async-channel = "1.9"
 common-stat.workspace = true
 async-stream.workspace = true
 async-trait.workspace = true
+base64.workspace = true
 bytemuck.workspace = true
 bytes.workspace = true
 chrono.workspace = true
@@ -108,6 +110,11 @@ name = "memtable_bench"
 harness = false
 required-features = ["test"]
 
+[[bench]]
+name = "bench_cache_stream"
+harness = false
+required-features = ["test"]
+
 [[bench]]
 name = "bench_filter_time_partition"
 harness = false
diff --git a/src/mito2/benches/bench_cache_stream.rs b/src/mito2/benches/bench_cache_stream.rs
new file mode 100644
index 0000000000..f2314f2ccb
--- /dev/null
+++ b/src/mito2/benches/bench_cache_stream.rs
@@ -0,0 +1,126 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Benchmarks for `cache_flat_range_stream` overhead.
+//!
+//! Compares consuming batches from a plain stream vs through the caching wrapper
+//! that clones batches for the range cache.
+//!
+//! Run with:
+//! ```sh
+//! cargo bench -p mito2 --features test --bench bench_cache_stream
+//! ```
+
+use std::collections::VecDeque;
+use std::sync::Arc;
+
+use criterion::{Criterion, criterion_group, criterion_main};
+use futures::TryStreamExt;
+use mito_codec::row_converter::DensePrimaryKeyCodec;
+use mito2::memtable::bulk::context::BulkIterContext;
+use mito2::memtable::bulk::part::{BulkPartConverter, BulkPartEncoder};
+use mito2::memtable::bulk::part_reader::EncodedBulkPartIter;
+use mito2::read::range_cache::bench_cache_flat_range_stream;
+use mito2::sst::parquet::DEFAULT_ROW_GROUP_SIZE;
+use mito2::sst::{FlatSchemaOptions, to_flat_sst_arrow_schema};
+use mito2::test_util::bench_util::{CpuDataGenerator, cpu_metadata};
+
+fn cache_flat_range_stream_bench(c: &mut Criterion) {
+    let metadata = Arc::new(cpu_metadata());
+    let region_id = metadata.region_id;
+    let start_sec = 1710043200;
+    // 2000 hosts × 51 steps = 102,000 rows ≈ DEFAULT_ROW_GROUP_SIZE
+    let num_hosts = 2000;
+    let end_sec = start_sec + 510;
+    let generator = CpuDataGenerator::new(metadata.clone(), num_hosts, start_sec, end_sec);
+
+    // Build a BulkPart from all the generated data
+    let schema = to_flat_sst_arrow_schema(&metadata, &FlatSchemaOptions::default());
+    let codec = Arc::new(DensePrimaryKeyCodec::new(&metadata));
+
+    let mut converter = BulkPartConverter::new(
+        &metadata,
+        schema,
+        DEFAULT_ROW_GROUP_SIZE,
+        codec,
+        true, // store_pk_columns
+    );
+    for kvs in generator.iter() {
+        converter.append_key_values(&kvs).unwrap();
+    }
+    let bulk_part = converter.convert().unwrap();
+
+    // Encode to parquet
+    let encoder = BulkPartEncoder::new(metadata.clone(), DEFAULT_ROW_GROUP_SIZE).unwrap();
+    let encoded_part = encoder.encode_part(&bulk_part).unwrap().unwrap();
+
+    // Decode all record batches
+    let num_row_groups = encoded_part.metadata().parquet_metadata.num_row_groups();
+    let context = Arc::new(
+        BulkIterContext::new(
+            metadata.clone(),
+            None, // No projection
+            None, // No predicate
+            false,
+        )
+        .unwrap(),
+    );
+    let row_groups: VecDeque<usize> = (0..num_row_groups).collect();
+
+    let rt = tokio::runtime::Runtime::new().unwrap();
+
+    let mut group = c.benchmark_group("cache_flat_range_stream");
+    group.sample_size(10);
+
+    group.bench_function("baseline_iter_stream", |b| {
+        b.iter(|| {
+            rt.block_on(async {
+                let iter = EncodedBulkPartIter::try_new(
+                    &encoded_part,
+                    context.clone(),
+                    row_groups.clone(),
+                    None,
+                    None,
+                )
+                .unwrap();
+                let stream: mito2::read::BoxedRecordBatchStream =
+                    Box::pin(futures::stream::iter(iter));
+                let mut stream = stream;
+                while let Some(_batch) = stream.try_next().await.unwrap() {}
+            });
+        });
+    });
+
+    group.bench_function("cache_flat_range_stream", |b| {
+        b.iter(|| {
+            rt.block_on(async {
+                let iter = EncodedBulkPartIter::try_new(
+                    &encoded_part,
+                    context.clone(),
+                    row_groups.clone(),
+                    None,
+                    None,
+                )
+                .unwrap();
+                let stream: mito2::read::BoxedRecordBatchStream =
+                    Box::pin(futures::stream::iter(iter));
+                let mut stream = bench_cache_flat_range_stream(stream, 64 * 1024 * 1024, region_id);
+                while let Some(_batch) = stream.try_next().await.unwrap() {}
+            });
+        });
+    });
+}
+
+criterion_group!(benches, cache_flat_range_stream_bench);
+criterion_main!(benches);
diff --git a/src/mito2/benches/memtable_bench.rs b/src/mito2/benches/memtable_bench.rs
index ebe994f861..8336625e3c 100644
--- a/src/mito2/benches/memtable_bench.rs
+++ b/src/mito2/benches/memtable_bench.rs
@@ -12,15 +12,17 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+//! Benchmarks for memtable operations: writes, full scans, filtered scans,
+//! bulk part conversion, record batch iteration with filters, and flat merge.
+//!
+//! Run with:
+//! ```sh
+//! cargo bench -p mito2 --features test --bench memtable_bench
+//! ```
+
 use std::sync::Arc;
 
-use api::v1::value::ValueData;
-use api::v1::{Row, Rows, SemanticType};
 use criterion::{Criterion, criterion_group, criterion_main};
-use datafusion_common::Column;
-use datafusion_expr::{Expr, lit};
-use datatypes::data_type::ConcreteDataType;
-use datatypes::schema::ColumnSchema;
 use mito_codec::row_converter::DensePrimaryKeyCodec;
 use mito2::memtable::bulk::context::BulkIterContext;
 use mito2::memtable::bulk::part::BulkPartConverter;
@@ -28,20 +30,13 @@ use mito2::memtable::bulk::part_reader::BulkPartBatchIter;
 use mito2::memtable::bulk::{BulkMemtable, BulkMemtableConfig};
 use mito2::memtable::partition_tree::{PartitionTreeConfig, PartitionTreeMemtable};
 use mito2::memtable::time_series::TimeSeriesMemtable;
-use mito2::memtable::{KeyValues, Memtable, RangesOptions};
+use mito2::memtable::{IterBuilder, Memtable, RangesOptions};
 use mito2::read::flat_merge::FlatMergeIterator;
 use mito2::read::scan_region::PredicateGroup;
 use mito2::region::options::MergeMode;
 use mito2::sst::{FlatSchemaOptions, to_flat_sst_arrow_schema};
-use mito2::test_util::memtable_util::{self, region_metadata_to_row_schema};
-use rand::Rng;
-use rand::rngs::ThreadRng;
-use rand::seq::IndexedRandom;
-use store_api::metadata::{
-    ColumnMetadata, RegionMetadata, RegionMetadataBuilder, RegionMetadataRef,
-};
-use store_api::storage::RegionId;
-use table::predicate::Predicate;
+use mito2::test_util::bench_util::{CpuDataGenerator, cpu_metadata};
+use mito2::test_util::memtable_util;
 
 /// Writes rows.
 fn write_rows(c: &mut Criterion) {
@@ -105,7 +100,11 @@ fn full_scan(c: &mut Criterion) {
         }
 
         b.iter(|| {
-            let iter = memtable.iter(None, None, None).unwrap();
+            let iter = memtable
+                .ranges(None, RangesOptions::default())
+                .unwrap()
+                .build(None)
+                .unwrap();
             for batch in iter {
                 let _batch = batch.unwrap();
             }
@@ -145,7 +144,17 @@ fn filter_1_host(c: &mut Criterion) {
         let predicate = generator.random_host_filter();
 
         b.iter(|| {
-            let iter = memtable.iter(None, Some(predicate.clone()), None).unwrap();
+            let iter = memtable
+                .ranges(
+                    None,
+                    RangesOptions {
+                        predicate: PredicateGroup::new(&metadata, predicate.exprs()).unwrap(),
+                        ..Default::default()
+                    },
+                )
+                .unwrap()
+                .build(None)
+                .unwrap();
             for batch in iter {
                 let _batch = batch.unwrap();
             }
@@ -202,224 +211,6 @@ fn filter_1_host(c: &mut Criterion) {
     });
 }
 
-struct Host {
-    hostname: String,
-    region: String,
-    datacenter: String,
-    rack: String,
-    os: String,
-    arch: String,
-    team: String,
-    service: String,
-    service_version: String,
-    service_environment: String,
-}
-
-impl Host {
-    fn random_with_id(id: usize) -> Host {
-        let mut rng = rand::rng();
-        let region = format!("ap-southeast-{}", rng.random_range(0..10));
-        let datacenter = format!(
-            "{}{}",
-            region,
-            ['a', 'b', 'c', 'd', 'e'].choose(&mut rng).unwrap()
-        );
-        Host {
-            hostname: format!("host_{id}"),
-            region,
-            datacenter,
-            rack: rng.random_range(0..100).to_string(),
-            os: "Ubuntu16.04LTS".to_string(),
-            arch: "x86".to_string(),
-            team: "CHI".to_string(),
-            service: rng.random_range(0..100).to_string(),
-            service_version: rng.random_range(0..10).to_string(),
-            service_environment: "test".to_string(),
-        }
-    }
-
-    fn fill_values(&self, values: &mut Vec<api::v1::Value>) {
-        let tags = [
-            api::v1::Value {
-                value_data: Some(ValueData::StringValue(self.hostname.clone())),
-            },
-            api::v1::Value {
-                value_data: Some(ValueData::StringValue(self.region.clone())),
-            },
-            api::v1::Value {
-                value_data: Some(ValueData::StringValue(self.datacenter.clone())),
-            },
-            api::v1::Value {
-                value_data: Some(ValueData::StringValue(self.rack.clone())),
-            },
-            api::v1::Value {
-                value_data: Some(ValueData::StringValue(self.os.clone())),
-            },
-            api::v1::Value {
-                value_data: Some(ValueData::StringValue(self.arch.clone())),
-            },
-            api::v1::Value {
-                value_data: Some(ValueData::StringValue(self.team.clone())),
-            },
-            api::v1::Value {
-                value_data: Some(ValueData::StringValue(self.service.clone())),
-            },
-            api::v1::Value {
-                value_data: Some(ValueData::StringValue(self.service_version.clone())),
-            },
-            api::v1::Value {
-                value_data: Some(ValueData::StringValue(self.service_environment.clone())),
-            },
-        ];
-        for tag in tags {
-            values.push(tag);
-        }
-    }
-}
-
-struct CpuDataGenerator {
-    metadata: RegionMetadataRef,
-    column_schemas: Vec<api::v1::ColumnSchema>,
-    hosts: Vec<Host>,
-    start_sec: i64,
-    end_sec: i64,
-}
-
-impl CpuDataGenerator {
-    fn new(metadata: RegionMetadataRef, num_hosts: usize, start_sec: i64, end_sec: i64) -> Self {
-        let column_schemas = region_metadata_to_row_schema(&metadata);
-        Self {
-            metadata,
-            column_schemas,
-            hosts: Self::generate_hosts(num_hosts),
-            start_sec,
-            end_sec,
-        }
-    }
-
-    fn iter(&self) -> impl Iterator<Item = KeyValues> + '_ {
-        // point per 10s.
-        (self.start_sec..self.end_sec)
-            .step_by(10)
-            .enumerate()
-            .map(|(seq, ts)| self.build_key_values(seq, ts))
-    }
-
-    fn build_key_values(&self, seq: usize, current_sec: i64) -> KeyValues {
-        let rows = self
-            .hosts
-            .iter()
-            .map(|host| {
-                let mut rng = rand::rng();
-                let mut values = Vec::with_capacity(21);
-                values.push(api::v1::Value {
-                    value_data: Some(ValueData::TimestampMillisecondValue(current_sec * 1000)),
-                });
-                host.fill_values(&mut values);
-                for _ in 0..10 {
-                    values.push(api::v1::Value {
-                        value_data: Some(ValueData::F64Value(Self::random_f64(&mut rng))),
-                    });
-                }
-                Row { values }
-            })
-            .collect();
-        let mutation = api::v1::Mutation {
-            op_type: api::v1::OpType::Put as i32,
-            sequence: seq as u64,
-            rows: Some(Rows {
-                schema: self.column_schemas.clone(),
-                rows,
-            }),
-            write_hint: None,
-        };
-
-        KeyValues::new(&self.metadata, mutation).unwrap()
-    }
-
-    fn random_host_filter(&self) -> Predicate {
-        let host = self.random_hostname();
-        let expr = Expr::Column(Column::from_name("hostname")).eq(lit(host));
-        Predicate::new(vec![expr])
-    }
-
-    fn random_host_filter_exprs(&self) -> Vec<Expr> {
-        let host = self.random_hostname();
-        vec![Expr::Column(Column::from_name("hostname")).eq(lit(host))]
-    }
-
-    fn random_hostname(&self) -> String {
-        let mut rng = rand::rng();
-        self.hosts.choose(&mut rng).unwrap().hostname.clone()
-    }
-
-    fn random_f64(rng: &mut ThreadRng) -> f64 {
-        let base: u32 = rng.random_range(30..95);
-        base as f64
-    }
-
-    fn generate_hosts(num_hosts: usize) -> Vec<Host> {
-        (0..num_hosts).map(Host::random_with_id).collect()
-    }
-}
-
-/// Creates a metadata for TSBS cpu-like table.
-fn cpu_metadata() -> RegionMetadata {
-    let mut builder = RegionMetadataBuilder::new(RegionId::new(1, 1));
-    builder.push_column_metadata(ColumnMetadata {
-        column_schema: ColumnSchema::new(
-            "ts",
-            ConcreteDataType::timestamp_millisecond_datatype(),
-            false,
-        ),
-        semantic_type: SemanticType::Timestamp,
-        column_id: 0,
-    });
-    let mut column_id = 1;
-    let tags = [
-        "hostname",
-        "region",
-        "datacenter",
-        "rack",
-        "os",
-        "arch",
-        "team",
-        "service",
-        "service_version",
-        "service_environment",
-    ];
-    for tag in tags {
-        builder.push_column_metadata(ColumnMetadata {
-            column_schema: ColumnSchema::new(tag, ConcreteDataType::string_datatype(), true),
-            semantic_type: SemanticType::Tag,
-            column_id,
-        });
-        column_id += 1;
-    }
-    let fields = [
-        "usage_user",
-        "usage_system",
-        "usage_idle",
-        "usage_nice",
-        "usage_iowait",
-        "usage_irq",
-        "usage_softirq",
-        "usage_steal",
-        "usage_guest",
-        "usage_guest_nice",
-    ];
-    for field in fields {
-        builder.push_column_metadata(ColumnMetadata {
-            column_schema: ColumnSchema::new(field, ConcreteDataType::float64_datatype(), true),
-            semantic_type: SemanticType::Field,
-            column_id,
-        });
-        column_id += 1;
-    }
-    builder.primary_key(vec![1, 2, 3, 4, 5, 6, 7, 8, 9, 10]);
-    builder.build().unwrap()
-}
-
 fn bulk_part_converter(c: &mut Criterion) {
     let metadata = Arc::new(cpu_metadata());
     let start_sec = 1710043200;
diff --git a/src/mito2/benches/simple_bulk_memtable.rs b/src/mito2/benches/simple_bulk_memtable.rs
index 0277397768..8a199f46f1 100644
--- a/src/mito2/benches/simple_bulk_memtable.rs
+++ b/src/mito2/benches/simple_bulk_memtable.rs
@@ -21,11 +21,7 @@ use criterion::{Criterion, criterion_group, criterion_main};
 use datatypes::data_type::ConcreteDataType;
 use datatypes::schema::ColumnSchema;
 use mito2::memtable::simple_bulk_memtable::SimpleBulkMemtable;
-use mito2::memtable::{KeyValues, Memtable, MemtableRanges, RangesOptions};
-use mito2::read;
-use mito2::read::Source;
-use mito2::read::dedup::DedupReader;
-use mito2::read::merge::MergeReaderBuilder;
+use mito2::memtable::{IterBuilder, KeyValues, Memtable, RangesOptions};
 use mito2::region::options::MergeMode;
 use mito2::test_util::column_metadata_to_column_schema;
 use store_api::metadata::{ColumnMetadata, RegionMetadataBuilder};
@@ -126,37 +122,11 @@ fn create_memtable_with_rows(num_batches: usize) -> SimpleBulkMemtable {
 }
 
 async fn flush(mem: &SimpleBulkMemtable) {
-    let MemtableRanges { ranges, .. } = mem.ranges(None, RangesOptions::for_flush()).unwrap();
-
-    let mut source = if ranges.len() == 1 {
-        let only_range = ranges.into_values().next().unwrap();
-        let iter = only_range.build_iter().unwrap();
-        Source::Iter(iter)
-    } else {
-        let sources = ranges
-            .into_values()
-            .map(|r| r.build_iter().map(Source::Iter))
-            .collect::<mito2::error::Result<Vec<_>>>()
-            .unwrap();
-        let merge_reader = MergeReaderBuilder::from_sources(sources)
-            .build()
-            .await
-            .unwrap();
-        let reader = Box::new(DedupReader::new(
-            merge_reader,
-            read::dedup::LastRow::new(true),
-            None,
-        ));
-        Source::Reader(reader)
-    };
-
-    while let Some(b) = source.next_batch().await.unwrap() {
-        black_box(b);
-    }
-}
-
-async fn flush_original(mem: &SimpleBulkMemtable) {
-    let iter = mem.iter(None, None, None).unwrap();
+    let iter = mem
+        .ranges(None, RangesOptions::default())
+        .unwrap()
+        .build(None)
+        .unwrap();
     for b in iter {
         black_box(b.unwrap());
     }
@@ -175,19 +145,10 @@ fn bench_ranges_parallel_vs_sequential(c: &mut Criterion) {
         let total_rows_k = num_batch * 10;
         let memtable = create_memtable_with_rows(num_batch);
 
-        group.bench_with_input(
-            BenchmarkId::new("flush_by_merge_reader", format!("{}k_rows", total_rows_k)),
-            &memtable,
-            |b, memtable| b.to_async(&rt).iter(|| async { flush(memtable).await }),
-        );
-
         group.bench_with_input(
             BenchmarkId::new("flush_by_iter", format!("{}k_rows", total_rows_k)),
             &memtable,
-            |b, memtable| {
-                b.to_async(&rt)
-                    .iter(|| async { flush_original(memtable).await })
-            },
+            |b, memtable| b.to_async(&rt).iter(|| async { flush(memtable).await }),
         );
     }
 
diff --git a/src/mito2/src/access_layer.rs b/src/mito2/src/access_layer.rs
index 92c8a3bc36..b6adb6eb59 100644
--- a/src/mito2/src/access_layer.rs
+++ b/src/mito2/src/access_layer.rs
@@ -17,7 +17,6 @@ use std::time::{Duration, Instant};
 
 use async_stream::try_stream;
 use common_time::Timestamp;
-use either::Either;
 use futures::{Stream, TryStreamExt};
 use object_store::services::Fs;
 use object_store::util::{join_dir, with_instrument_layers};
@@ -37,7 +36,7 @@ use crate::error::{
     CleanDirSnafu, DeleteIndexSnafu, DeleteIndexesSnafu, DeleteSstsSnafu, OpenDalSnafu, Result,
 };
 use crate::metrics::{COMPACTION_STAGE_ELAPSED, FLUSH_ELAPSED};
-use crate::read::{FlatSource, Source};
+use crate::read::FlatSource;
 use crate::region::options::IndexOptions;
 use crate::sst::file::{FileHandle, RegionFileId, RegionIndexId};
 use crate::sst::index::IndexerBuilderImpl;
@@ -47,7 +46,7 @@ use crate::sst::location::{self, region_dir_from_table_dir};
 use crate::sst::parquet::reader::ParquetReaderBuilder;
 use crate::sst::parquet::writer::ParquetWriter;
 use crate::sst::parquet::{SstInfo, WriteOptions};
-use crate::sst::{DEFAULT_WRITE_BUFFER_SIZE, DEFAULT_WRITE_CONCURRENCY};
+use crate::sst::{DEFAULT_WRITE_BUFFER_SIZE, DEFAULT_WRITE_CONCURRENCY, FormatType};
 
 pub type AccessLayerRef = Arc<AccessLayer>;
 /// SST write results.
@@ -339,6 +338,7 @@ impl AccessLayer {
         metrics: &mut Metrics,
     ) -> Result<SstInfoArray> {
         let region_id = request.metadata.region_id;
+        let region_metadata = request.metadata.clone();
         let cache_manager = request.cache_manager.clone();
 
         let sst_info = if let Some(write_cache) = cache_manager.write_cache() {
@@ -391,15 +391,19 @@ impl AccessLayer {
             )
             .await
             .with_file_cleaner(cleaner);
-            match request.source {
-                Either::Left(source) => {
+            match request.sst_write_format {
+                FormatType::PrimaryKey => {
                     writer
-                        .write_all(source, request.max_sequence, write_opts)
+                        .write_all_flat_as_primary_key(
+                            request.source,
+                            request.max_sequence,
+                            write_opts,
+                        )
                         .await?
                 }
-                Either::Right(flat_source) => {
+                FormatType::Flat => {
                     writer
-                        .write_all_flat(flat_source, request.max_sequence, write_opts)
+                        .write_all_flat(request.source, request.max_sequence, write_opts)
                         .await?
                 }
             }
@@ -412,6 +416,7 @@ impl AccessLayer {
                     cache_manager.put_parquet_meta_data(
                         RegionFileId::new(region_id, sst.file_id),
                         parquet_metadata.clone(),
+                        Some(region_metadata.clone()),
                     )
                 }
             }
@@ -494,7 +499,7 @@ impl AccessLayer {
                 let file_size = if file_size == 0 { None } else { Some(file_size) };
                 let last_modified_ms = metadata
                     .last_modified()
-                    .map(|ts| Timestamp::new_millisecond(ts.timestamp_millis()));
+                    .map(|ts| Timestamp::new_millisecond(ts.into_inner().as_millisecond()));
 
                 let entry = StorageSstEntry {
                     file_path: path.to_string(),
@@ -520,11 +525,12 @@ pub enum OperationType {
 pub struct SstWriteRequest {
     pub op_type: OperationType,
     pub metadata: RegionMetadataRef,
-    pub source: Either<Source, FlatSource>,
+    pub source: FlatSource,
     pub cache_manager: CacheManagerRef,
     #[allow(dead_code)]
     pub storage: Option<String>,
     pub max_sequence: Option<SequenceNumber>,
+    pub sst_write_format: FormatType,
 
     /// Configs for index
     pub index_options: IndexOptions,
diff --git a/src/mito2/src/cache.rs b/src/mito2/src/cache.rs
index 3ad71d2a61..48881e5f05 100644
--- a/src/mito2/src/cache.rs
+++ b/src/mito2/src/cache.rs
@@ -28,6 +28,8 @@ use std::ops::Range;
 use std::sync::Arc;
 
 use bytes::Bytes;
+use common_base::readable_size::ReadableSize;
+use common_telemetry::warn;
 use datatypes::arrow::record_batch::RecordBatch;
 use datatypes::value::Value;
 use datatypes::vectors::VectorRef;
@@ -36,8 +38,10 @@ use index::result_cache::IndexResultCache;
 use moka::notification::RemovalCause;
 use moka::sync::Cache;
 use object_store::ObjectStore;
-use parquet::file::metadata::{PageIndexPolicy, ParquetMetaData};
+use parquet::file::metadata::{FileMetaData, PageIndexPolicy, ParquetMetaData};
 use puffin::puffin_manager::cache::{PuffinMetadataCache, PuffinMetadataCacheRef};
+use snafu::{OptionExt, ResultExt};
+use store_api::metadata::RegionMetadataRef;
 use store_api::storage::{ConcreteDataType, FileId, RegionId, TimeSeriesRowSelector};
 
 use crate::cache::cache_size::parquet_meta_size;
@@ -46,10 +50,14 @@ use crate::cache::index::inverted_index::{InvertedIndexCache, InvertedIndexCache
 #[cfg(feature = "vector_index")]
 use crate::cache::index::vector_index::{VectorIndexCache, VectorIndexCacheRef};
 use crate::cache::write_cache::WriteCacheRef;
+use crate::error::{InvalidMetadataSnafu, InvalidParquetSnafu, Result, UnexpectedSnafu};
 use crate::memtable::record_batch_estimated_size;
 use crate::metrics::{CACHE_BYTES, CACHE_EVICTION, CACHE_HIT, CACHE_MISS};
 use crate::read::Batch;
+use crate::read::range_cache::{RangeScanCacheKey, RangeScanCacheValue};
 use crate::sst::file::{RegionFileId, RegionIndexId};
+use crate::sst::parquet::PARQUET_METADATA_KEY;
+use crate::sst::parquet::read_columns::ParquetReadColumns;
 use crate::sst::parquet::reader::MetadataCacheMetrics;
 
 /// Metrics type key for sst meta.
@@ -64,6 +72,172 @@ const FILE_TYPE: &str = "file";
 const INDEX_TYPE: &str = "index";
 /// Metrics type key for selector result cache.
 const SELECTOR_RESULT_TYPE: &str = "selector_result";
+/// Metrics type key for range scan result cache.
+const RANGE_RESULT_TYPE: &str = "range_result";
+const RANGE_RESULT_CONCAT_MEMORY_LIMIT: ReadableSize = ReadableSize::mb(512);
+const RANGE_RESULT_CONCAT_MEMORY_PERMIT: ReadableSize = ReadableSize::kb(1);
+
+#[derive(Debug)]
+pub(crate) struct RangeResultMemoryLimiter {
+    semaphore: Arc<tokio::sync::Semaphore>,
+    permit_bytes: usize,
+    total_permits: usize,
+}
+
+impl Default for RangeResultMemoryLimiter {
+    fn default() -> Self {
+        Self::new(
+            RANGE_RESULT_CONCAT_MEMORY_LIMIT.as_bytes() as usize,
+            RANGE_RESULT_CONCAT_MEMORY_PERMIT.as_bytes() as usize,
+        )
+    }
+}
+
+impl RangeResultMemoryLimiter {
+    pub(crate) fn new(limit_bytes: usize, permit_bytes: usize) -> Self {
+        let permit_bytes = permit_bytes.max(1);
+        let total_permits = limit_bytes
+            .div_ceil(permit_bytes)
+            .clamp(1, tokio::sync::Semaphore::MAX_PERMITS);
+        Self {
+            semaphore: Arc::new(tokio::sync::Semaphore::new(total_permits)),
+            permit_bytes,
+            total_permits,
+        }
+    }
+
+    #[cfg(test)]
+    pub(crate) fn permit_bytes(&self) -> usize {
+        self.permit_bytes
+    }
+
+    #[cfg(test)]
+    pub(crate) fn available_permits(&self) -> usize {
+        self.semaphore.available_permits()
+    }
+
+    pub(crate) async fn acquire(&self, bytes: usize) -> Result<tokio::sync::SemaphorePermit<'_>> {
+        let permits = bytes.div_ceil(self.permit_bytes).max(1);
+        if permits > self.total_permits {
+            return UnexpectedSnafu {
+                reason: format!(
+                    "range result memory request of {bytes} bytes exceeds limiter capacity of {} bytes",
+                    self.total_permits.saturating_mul(self.permit_bytes)
+                ),
+            }
+            .fail();
+        }
+        self.semaphore
+            .acquire_many(permits as u32)
+            .await
+            .map_err(|_| {
+                UnexpectedSnafu {
+                    reason: "range result memory limiter is unexpectedly closed",
+                }
+                .build()
+            })
+    }
+}
+
+/// Cached SST metadata combines the parquet footer with the decoded region metadata.
+///
+/// The cached parquet footer strips the `greptime:metadata` JSON payload and stores the decoded
+/// [RegionMetadata] separately so readers can skip repeated deserialization work.
+#[derive(Debug)]
+pub(crate) struct CachedSstMeta {
+    parquet_metadata: Arc<ParquetMetaData>,
+    region_metadata: RegionMetadataRef,
+    region_metadata_weight: usize,
+}
+
+impl CachedSstMeta {
+    pub(crate) fn try_new(file_path: &str, parquet_metadata: ParquetMetaData) -> Result<Self> {
+        Self::try_new_with_region_metadata(file_path, parquet_metadata, None)
+    }
+
+    pub(crate) fn try_new_with_region_metadata(
+        file_path: &str,
+        parquet_metadata: ParquetMetaData,
+        region_metadata: Option<RegionMetadataRef>,
+    ) -> Result<Self> {
+        let file_metadata = parquet_metadata.file_metadata();
+        let key_values = file_metadata
+            .key_value_metadata()
+            .context(InvalidParquetSnafu {
+                file: file_path,
+                reason: "missing key value meta",
+            })?;
+        let meta_value = key_values
+            .iter()
+            .find(|kv| kv.key == PARQUET_METADATA_KEY)
+            .with_context(|| InvalidParquetSnafu {
+                file: file_path,
+                reason: format!("key {} not found", PARQUET_METADATA_KEY),
+            })?;
+        let json = meta_value
+            .value
+            .as_ref()
+            .with_context(|| InvalidParquetSnafu {
+                file: file_path,
+                reason: format!("No value for key {}", PARQUET_METADATA_KEY),
+            })?;
+        let region_metadata = match region_metadata {
+            Some(region_metadata) => region_metadata,
+            None => Arc::new(
+                store_api::metadata::RegionMetadata::from_json(json)
+                    .context(InvalidMetadataSnafu)?,
+            ),
+        };
+        // Keep the previous JSON-byte floor and charge the decoded structures as well.
+        let region_metadata_weight = region_metadata.estimated_size().max(json.len());
+        let parquet_metadata = Arc::new(strip_region_metadata_from_parquet(parquet_metadata));
+
+        Ok(Self {
+            parquet_metadata,
+            region_metadata,
+            region_metadata_weight,
+        })
+    }
+
+    pub(crate) fn parquet_metadata(&self) -> Arc<ParquetMetaData> {
+        self.parquet_metadata.clone()
+    }
+
+    pub(crate) fn region_metadata(&self) -> RegionMetadataRef {
+        self.region_metadata.clone()
+    }
+}
+
+fn strip_region_metadata_from_parquet(parquet_metadata: ParquetMetaData) -> ParquetMetaData {
+    let file_metadata = parquet_metadata.file_metadata();
+    let filtered_key_values = file_metadata.key_value_metadata().and_then(|key_values| {
+        let filtered = key_values
+            .iter()
+            .filter(|kv| kv.key != PARQUET_METADATA_KEY)
+            .cloned()
+            .collect::<Vec<_>>();
+        (!filtered.is_empty()).then_some(filtered)
+    });
+    let stripped_file_metadata = FileMetaData::new(
+        file_metadata.version(),
+        file_metadata.num_rows(),
+        file_metadata.created_by().map(ToString::to_string),
+        filtered_key_values,
+        file_metadata.schema_descr_ptr(),
+        file_metadata.column_orders().cloned(),
+    );
+
+    let mut builder = parquet_metadata.into_builder();
+    let row_groups = builder.take_row_groups();
+    let column_index = builder.take_column_index();
+    let offset_index = builder.take_offset_index();
+
+    parquet::file::metadata::ParquetMetaDataBuilder::new(stripped_file_metadata)
+        .set_row_groups(row_groups)
+        .set_column_index(column_index)
+        .set_offset_index(offset_index)
+        .build()
+}
 
 /// Cache strategies that may only enable a subset of caches.
 #[derive(Clone)]
@@ -81,18 +255,17 @@ pub enum CacheStrategy {
 }
 
 impl CacheStrategy {
-    /// Gets parquet metadata with cache metrics tracking.
-    /// Returns the metadata and updates the provided metrics.
-    pub(crate) async fn get_parquet_meta_data(
+    /// Gets fused SST metadata with cache metrics tracking.
+    pub(crate) async fn get_sst_meta_data(
         &self,
         file_id: RegionFileId,
         metrics: &mut MetadataCacheMetrics,
         page_index_policy: PageIndexPolicy,
-    ) -> Option<Arc<ParquetMetaData>> {
+    ) -> Option<Arc<CachedSstMeta>> {
         match self {
             CacheStrategy::EnableAll(cache_manager) | CacheStrategy::Compaction(cache_manager) => {
                 cache_manager
-                    .get_parquet_meta_data(file_id, metrics, page_index_policy)
+                    .get_sst_meta_data(file_id, metrics, page_index_policy)
                     .await
             }
             CacheStrategy::Disabled => {
@@ -102,30 +275,48 @@ impl CacheStrategy {
         }
     }
 
-    /// Calls [CacheManager::get_parquet_meta_data_from_mem_cache()].
-    pub fn get_parquet_meta_data_from_mem_cache(
+    /// Calls [CacheManager::get_sst_meta_data_from_mem_cache()].
+    pub(crate) fn get_sst_meta_data_from_mem_cache(
         &self,
         file_id: RegionFileId,
-    ) -> Option<Arc<ParquetMetaData>> {
+    ) -> Option<Arc<CachedSstMeta>> {
         match self {
-            CacheStrategy::EnableAll(cache_manager) => {
-                cache_manager.get_parquet_meta_data_from_mem_cache(file_id)
-            }
-            CacheStrategy::Compaction(cache_manager) => {
-                cache_manager.get_parquet_meta_data_from_mem_cache(file_id)
+            CacheStrategy::EnableAll(cache_manager) | CacheStrategy::Compaction(cache_manager) => {
+                cache_manager.get_sst_meta_data_from_mem_cache(file_id)
             }
             CacheStrategy::Disabled => None,
         }
     }
 
-    /// Calls [CacheManager::put_parquet_meta_data()].
-    pub fn put_parquet_meta_data(&self, file_id: RegionFileId, metadata: Arc<ParquetMetaData>) {
+    /// Calls [CacheManager::get_parquet_meta_data_from_mem_cache()].
+    pub fn get_parquet_meta_data_from_mem_cache(
+        &self,
+        file_id: RegionFileId,
+    ) -> Option<Arc<ParquetMetaData>> {
+        self.get_sst_meta_data_from_mem_cache(file_id)
+            .map(|metadata| metadata.parquet_metadata())
+    }
+
+    /// Calls [CacheManager::put_sst_meta_data()].
+    pub(crate) fn put_sst_meta_data(&self, file_id: RegionFileId, metadata: Arc<CachedSstMeta>) {
         match self {
-            CacheStrategy::EnableAll(cache_manager) => {
-                cache_manager.put_parquet_meta_data(file_id, metadata);
+            CacheStrategy::EnableAll(cache_manager) | CacheStrategy::Compaction(cache_manager) => {
+                cache_manager.put_sst_meta_data(file_id, metadata);
             }
-            CacheStrategy::Compaction(cache_manager) => {
-                cache_manager.put_parquet_meta_data(file_id, metadata);
+            CacheStrategy::Disabled => {}
+        }
+    }
+
+    /// Calls [CacheManager::put_parquet_meta_data()].
+    pub fn put_parquet_meta_data(
+        &self,
+        file_id: RegionFileId,
+        metadata: Arc<ParquetMetaData>,
+        region_metadata: Option<RegionMetadataRef>,
+    ) {
+        match self {
+            CacheStrategy::EnableAll(cache_manager) | CacheStrategy::Compaction(cache_manager) => {
+                cache_manager.put_parquet_meta_data(file_id, metadata, region_metadata);
             }
             CacheStrategy::Disabled => {}
         }
@@ -223,6 +414,57 @@ impl CacheStrategy {
         }
     }
 
+    /// Calls [CacheManager::get_range_result()].
+    /// It returns None if the strategy is [CacheStrategy::Compaction] or [CacheStrategy::Disabled].
+    #[allow(dead_code)]
+    pub(crate) fn get_range_result(
+        &self,
+        key: &RangeScanCacheKey,
+    ) -> Option<Arc<RangeScanCacheValue>> {
+        match self {
+            CacheStrategy::EnableAll(cache_manager) => cache_manager.get_range_result(key),
+            CacheStrategy::Compaction(_) | CacheStrategy::Disabled => None,
+        }
+    }
+
+    /// Calls [CacheManager::put_range_result()].
+    /// It does nothing if the strategy isn't [CacheStrategy::EnableAll].
+    pub(crate) fn put_range_result(
+        &self,
+        key: RangeScanCacheKey,
+        result: Arc<RangeScanCacheValue>,
+    ) {
+        if let CacheStrategy::EnableAll(cache_manager) = self {
+            cache_manager.put_range_result(key, result);
+        }
+    }
+
+    /// Returns true if the range result cache is enabled.
+    pub(crate) fn has_range_result_cache(&self) -> bool {
+        match self {
+            CacheStrategy::EnableAll(cache_manager) => cache_manager.has_range_result_cache(),
+            CacheStrategy::Compaction(_) | CacheStrategy::Disabled => false,
+        }
+    }
+
+    pub(crate) fn range_result_memory_limiter(&self) -> Option<&Arc<RangeResultMemoryLimiter>> {
+        match self {
+            CacheStrategy::EnableAll(cache_manager) => {
+                Some(cache_manager.range_result_memory_limiter())
+            }
+            CacheStrategy::Compaction(_) | CacheStrategy::Disabled => None,
+        }
+    }
+
+    pub(crate) fn range_result_cache_size(&self) -> Option<usize> {
+        match self {
+            CacheStrategy::EnableAll(cache_manager) => {
+                Some(cache_manager.range_result_cache_size())
+            }
+            CacheStrategy::Compaction(_) | CacheStrategy::Disabled => None,
+        }
+    }
+
     /// Calls [CacheManager::write_cache()].
     /// It returns None if the strategy is [CacheStrategy::Disabled].
     pub fn write_cache(&self) -> Option<&WriteCacheRef> {
@@ -324,6 +566,12 @@ pub struct CacheManager {
     puffin_metadata_cache: Option<PuffinMetadataCacheRef>,
     /// Cache for time series selectors.
     selector_result_cache: Option<SelectorResultCache>,
+    /// Cache for range scan outputs in flat format.
+    range_result_cache: Option<RangeResultCache>,
+    /// Configured capacity for range scan outputs in flat format.
+    range_result_cache_size: u64,
+    /// Shared memory limiter for async range-result cache tasks.
+    range_result_memory_limiter: Arc<RangeResultMemoryLimiter>,
     /// Cache for index result.
     index_result_cache: Option<IndexResultCache>,
 }
@@ -336,6 +584,35 @@ impl CacheManager {
         CacheManagerBuilder::default()
     }
 
+    /// Gets fused SST metadata with metrics tracking.
+    /// Tries in-memory cache first, then file cache, updating metrics accordingly.
+    pub(crate) async fn get_sst_meta_data(
+        &self,
+        file_id: RegionFileId,
+        metrics: &mut MetadataCacheMetrics,
+        page_index_policy: PageIndexPolicy,
+    ) -> Option<Arc<CachedSstMeta>> {
+        if let Some(metadata) = self.get_sst_meta_data_from_mem_cache(file_id) {
+            metrics.mem_cache_hit += 1;
+            return Some(metadata);
+        }
+
+        let key = IndexKey::new(file_id.region_id(), file_id.file_id(), FileType::Parquet);
+        if let Some(write_cache) = &self.write_cache
+            && let Some(metadata) = write_cache
+                .file_cache()
+                .get_sst_meta_data(key, metrics, page_index_policy)
+                .await
+        {
+            metrics.file_cache_hit += 1;
+            self.put_sst_meta_data(file_id, metadata.clone());
+            return Some(metadata);
+        }
+
+        metrics.cache_miss += 1;
+        None
+    }
+
     /// Gets cached [ParquetMetaData] with metrics tracking.
     /// Tries in-memory cache first, then file cache, updating metrics accordingly.
     pub(crate) async fn get_parquet_meta_data(
@@ -344,29 +621,21 @@ impl CacheManager {
         metrics: &mut MetadataCacheMetrics,
         page_index_policy: PageIndexPolicy,
     ) -> Option<Arc<ParquetMetaData>> {
-        // Try to get metadata from sst meta cache
-        if let Some(metadata) = self.get_parquet_meta_data_from_mem_cache(file_id) {
-            metrics.mem_cache_hit += 1;
-            return Some(metadata);
-        }
+        self.get_sst_meta_data(file_id, metrics, page_index_policy)
+            .await
+            .map(|metadata| metadata.parquet_metadata())
+    }
 
-        // Try to get metadata from write cache
-        let key = IndexKey::new(file_id.region_id(), file_id.file_id(), FileType::Parquet);
-        if let Some(write_cache) = &self.write_cache
-            && let Some(metadata) = write_cache
-                .file_cache()
-                .get_parquet_meta_data(key, metrics, page_index_policy)
-                .await
-        {
-            metrics.file_cache_hit += 1;
-            let metadata = Arc::new(metadata);
-            // Put metadata into sst meta cache
-            self.put_parquet_meta_data(file_id, metadata.clone());
-            return Some(metadata);
-        };
-        metrics.cache_miss += 1;
-
-        None
+    /// Gets cached fused SST metadata from in-memory cache.
+    /// This method does not perform I/O.
+    pub(crate) fn get_sst_meta_data_from_mem_cache(
+        &self,
+        file_id: RegionFileId,
+    ) -> Option<Arc<CachedSstMeta>> {
+        self.sst_meta_cache.as_ref().and_then(|sst_meta_cache| {
+            let value = sst_meta_cache.get(&SstMetaKey(file_id.region_id(), file_id.file_id()));
+            update_hit_miss(value, SST_META_TYPE)
+        })
     }
 
     /// Gets cached [ParquetMetaData] from in-memory cache.
@@ -375,15 +644,12 @@ impl CacheManager {
         &self,
         file_id: RegionFileId,
     ) -> Option<Arc<ParquetMetaData>> {
-        // Try to get metadata from sst meta cache
-        self.sst_meta_cache.as_ref().and_then(|sst_meta_cache| {
-            let value = sst_meta_cache.get(&SstMetaKey(file_id.region_id(), file_id.file_id()));
-            update_hit_miss(value, SST_META_TYPE)
-        })
+        self.get_sst_meta_data_from_mem_cache(file_id)
+            .map(|metadata| metadata.parquet_metadata())
     }
 
-    /// Puts [ParquetMetaData] into the cache.
-    pub fn put_parquet_meta_data(&self, file_id: RegionFileId, metadata: Arc<ParquetMetaData>) {
+    /// Puts fused SST metadata into the cache.
+    pub(crate) fn put_sst_meta_data(&self, file_id: RegionFileId, metadata: Arc<CachedSstMeta>) {
         if let Some(cache) = &self.sst_meta_cache {
             let key = SstMetaKey(file_id.region_id(), file_id.file_id());
             CACHE_BYTES
@@ -393,6 +659,34 @@ impl CacheManager {
         }
     }
 
+    /// Puts [ParquetMetaData] into the cache.
+    pub fn put_parquet_meta_data(
+        &self,
+        file_id: RegionFileId,
+        metadata: Arc<ParquetMetaData>,
+        region_metadata: Option<RegionMetadataRef>,
+    ) {
+        if self.sst_meta_cache.is_some() {
+            let file_path = format!(
+                "region_id={}, file_id={}",
+                file_id.region_id(),
+                file_id.file_id()
+            );
+            match CachedSstMeta::try_new_with_region_metadata(
+                &file_path,
+                Arc::unwrap_or_clone(metadata),
+                region_metadata,
+            ) {
+                Ok(metadata) => self.put_sst_meta_data(file_id, Arc::new(metadata)),
+                Err(err) => warn!(
+                    err; "Failed to decode region metadata while caching parquet metadata, region_id: {}, file_id: {}",
+                    file_id.region_id(),
+                    file_id.file_id()
+                ),
+            }
+        }
+    }
+
     /// Removes [ParquetMetaData] from the cache.
     pub fn remove_parquet_meta_data(&self, file_id: RegionFileId) {
         if let Some(cache) = &self.sst_meta_cache {
@@ -512,6 +806,44 @@ impl CacheManager {
         }
     }
 
+    /// Gets cached result for range scan.
+    #[allow(dead_code)]
+    pub(crate) fn get_range_result(
+        &self,
+        key: &RangeScanCacheKey,
+    ) -> Option<Arc<RangeScanCacheValue>> {
+        self.range_result_cache
+            .as_ref()
+            .and_then(|cache| update_hit_miss(cache.get(key), RANGE_RESULT_TYPE))
+    }
+
+    /// Puts range scan result into cache.
+    pub(crate) fn put_range_result(
+        &self,
+        key: RangeScanCacheKey,
+        result: Arc<RangeScanCacheValue>,
+    ) {
+        if let Some(cache) = &self.range_result_cache {
+            CACHE_BYTES
+                .with_label_values(&[RANGE_RESULT_TYPE])
+                .add(range_result_cache_weight(&key, &result).into());
+            cache.insert(key, result);
+        }
+    }
+
+    /// Returns true if the range result cache is enabled.
+    pub(crate) fn has_range_result_cache(&self) -> bool {
+        self.range_result_cache.is_some()
+    }
+
+    pub(crate) fn range_result_memory_limiter(&self) -> &Arc<RangeResultMemoryLimiter> {
+        &self.range_result_memory_limiter
+    }
+
+    pub(crate) fn range_result_cache_size(&self) -> usize {
+        self.range_result_cache_size as usize
+    }
+
     /// Gets the write cache.
     pub(crate) fn write_cache(&self) -> Option<&WriteCacheRef> {
         self.write_cache.as_ref()
@@ -562,6 +894,7 @@ pub struct CacheManagerBuilder {
     puffin_metadata_size: u64,
     write_cache: Option<WriteCacheRef>,
     selector_result_cache_size: u64,
+    range_result_cache_size: u64,
 }
 
 impl CacheManagerBuilder {
@@ -625,6 +958,12 @@ impl CacheManagerBuilder {
         self
     }
 
+    /// Sets range result cache size.
+    pub fn range_result_cache_size(mut self, bytes: u64) -> Self {
+        self.range_result_cache_size = bytes;
+        self
+    }
+
     /// Builds the [CacheManager].
     pub fn build(self) -> CacheManager {
         fn to_str(cause: RemovalCause) -> &'static str {
@@ -712,6 +1051,21 @@ impl CacheManagerBuilder {
                 })
                 .build()
         });
+        let range_result_cache = (self.range_result_cache_size != 0).then(|| {
+            Cache::builder()
+                .max_capacity(self.range_result_cache_size)
+                .weigher(range_result_cache_weight)
+                .eviction_listener(move |k, v, cause| {
+                    let size = range_result_cache_weight(&k, &v);
+                    CACHE_BYTES
+                        .with_label_values(&[RANGE_RESULT_TYPE])
+                        .sub(size.into());
+                    CACHE_EVICTION
+                        .with_label_values(&[RANGE_RESULT_TYPE, to_str(cause)])
+                        .inc();
+                })
+                .build()
+        });
         CacheManager {
             sst_meta_cache,
             vector_cache,
@@ -723,14 +1077,20 @@ impl CacheManagerBuilder {
             vector_index_cache,
             puffin_metadata_cache: Some(Arc::new(puffin_metadata_cache)),
             selector_result_cache,
+            range_result_cache,
+            range_result_cache_size: self.range_result_cache_size,
+            range_result_memory_limiter: Arc::new(RangeResultMemoryLimiter::new(
+                self.range_result_cache_size as usize,
+                RANGE_RESULT_CONCAT_MEMORY_PERMIT.as_bytes() as usize,
+            )),
             index_result_cache,
         }
     }
 }
 
-fn meta_cache_weight(k: &SstMetaKey, v: &Arc<ParquetMetaData>) -> u32 {
+fn meta_cache_weight(k: &SstMetaKey, v: &Arc<CachedSstMeta>) -> u32 {
     // We ignore the size of `Arc`.
-    (k.estimated_size() + parquet_meta_size(v)) as u32
+    (k.estimated_size() + parquet_meta_size(&v.parquet_metadata) + v.region_metadata_weight) as u32
 }
 
 fn vector_cache_weight(_k: &(ConcreteDataType, Value), v: &VectorRef) -> u32 {
@@ -746,6 +1106,10 @@ fn selector_result_cache_weight(k: &SelectorResultKey, v: &Arc<SelectorResultVal
     (mem::size_of_val(k) + v.estimated_size()) as u32
 }
 
+fn range_result_cache_weight(k: &RangeScanCacheKey, v: &Arc<RangeScanCacheValue>) -> u32 {
+    (k.estimated_size() + v.estimated_size()) as u32
+}
+
 /// Updates cache hit/miss metrics.
 fn update_hit_miss<T>(value: Option<T>, cache_type: &str) -> Option<T> {
     if value.is_some() {
@@ -860,24 +1224,27 @@ pub enum SelectorResult {
 pub struct SelectorResultValue {
     /// Batches of rows selected by the selector.
     pub result: SelectorResult,
-    /// Projection of rows.
-    pub projection: Vec<usize>,
+    /// The read columns of rows.
+    pub read_cols: ParquetReadColumns,
 }
 
 impl SelectorResultValue {
     /// Creates a new selector result value with primary key format.
-    pub fn new(result: Vec<Batch>, projection: Vec<usize>) -> SelectorResultValue {
+    pub fn new(result: Vec<Batch>, read_cols: ParquetReadColumns) -> SelectorResultValue {
         SelectorResultValue {
             result: SelectorResult::PrimaryKey(result),
-            projection,
+            read_cols,
         }
     }
 
     /// Creates a new selector result value with flat format.
-    pub fn new_flat(result: Vec<RecordBatch>, projection: Vec<usize>) -> SelectorResultValue {
+    pub fn new_flat(
+        result: Vec<RecordBatch>,
+        read_cols: ParquetReadColumns,
+    ) -> SelectorResultValue {
         SelectorResultValue {
             result: SelectorResult::Flat(result),
-            projection,
+            read_cols,
         }
     }
 
@@ -892,8 +1259,8 @@ impl SelectorResultValue {
     }
 }
 
-/// Maps (region id, file id) to [ParquetMetaData].
-type SstMetaCache = Cache<SstMetaKey, Arc<ParquetMetaData>>;
+/// Maps (region id, file id) to fused SST metadata.
+type SstMetaCache = Cache<SstMetaKey, Arc<CachedSstMeta>>;
 /// Maps [Value] to a vector that holds this value repeatedly.
 ///
 /// e.g. `"hello" => ["hello", "hello", "hello"]`
@@ -902,20 +1269,31 @@ type VectorCache = Cache<(ConcreteDataType, Value), VectorRef>;
 type PageCache = Cache<PageKey, Arc<PageValue>>;
 /// Maps (file id, row group id, time series row selector) to [SelectorResultValue].
 type SelectorResultCache = Cache<SelectorResultKey, Arc<SelectorResultValue>>;
+/// Maps partition-range scan key to cached flat batches.
+type RangeResultCache = Cache<RangeScanCacheKey, Arc<RangeScanCacheValue>>;
 
 #[cfg(test)]
 mod tests {
     use std::sync::Arc;
 
+    use api::v1::SemanticType;
     use api::v1::index::{BloomFilterMeta, InvertedIndexMetas};
+    use datatypes::schema::ColumnSchema;
     use datatypes::vectors::Int64Vector;
     use puffin::file_metadata::FileMetadata;
+    use store_api::metadata::{ColumnMetadata, RegionMetadata, RegionMetadataBuilder};
     use store_api::storage::ColumnId;
 
     use super::*;
     use crate::cache::index::bloom_filter_index::Tag;
     use crate::cache::index::result_cache::PredicateKey;
-    use crate::cache::test_util::parquet_meta;
+    use crate::cache::test_util::{
+        parquet_meta, sst_parquet_meta, sst_parquet_meta_with_region_metadata,
+    };
+    use crate::read::range_cache::{
+        RangeScanCacheKey, RangeScanCacheValue, ScanRequestFingerprintBuilder,
+    };
+    use crate::read::read_columns::ReadColumns;
     use crate::sst::parquet::row_selection::RowGroupSelection;
 
     #[tokio::test]
@@ -929,7 +1307,7 @@ mod tests {
         let file_id = RegionFileId::new(region_id, FileId::random());
         let metadata = parquet_meta();
         let mut metrics = MetadataCacheMetrics::default();
-        cache.put_parquet_meta_data(file_id, metadata);
+        cache.put_parquet_meta_data(file_id, metadata, None);
         assert!(
             cache
                 .get_parquet_meta_data(file_id, &mut metrics, Default::default())
@@ -966,13 +1344,23 @@ mod tests {
                 .await
                 .is_none()
         );
-        let metadata = parquet_meta();
-        cache.put_parquet_meta_data(file_id, metadata);
+        let (metadata, region_metadata) = sst_parquet_meta();
+        cache.put_parquet_meta_data(file_id, metadata, None);
+        let cached = cache
+            .get_sst_meta_data(file_id, &mut metrics, Default::default())
+            .await
+            .unwrap();
+        assert_eq!(region_metadata, cached.region_metadata());
         assert!(
-            cache
-                .get_parquet_meta_data(file_id, &mut metrics, Default::default())
-                .await
-                .is_some()
+            cached
+                .parquet_metadata()
+                .file_metadata()
+                .key_value_metadata()
+                .is_none_or(|key_values| {
+                    key_values
+                        .iter()
+                        .all(|key_value| key_value.key != PARQUET_METADATA_KEY)
+                })
         );
         cache.remove_parquet_meta_data(file_id);
         assert!(
@@ -983,6 +1371,42 @@ mod tests {
         );
     }
 
+    #[tokio::test]
+    async fn test_parquet_meta_cache_with_provided_region_metadata() {
+        let cache = CacheManager::builder().sst_meta_cache_size(2000).build();
+        let mut metrics = MetadataCacheMetrics::default();
+        let region_id = RegionId::new(1, 1);
+        let file_id = RegionFileId::new(region_id, FileId::random());
+        let (metadata, region_metadata) = sst_parquet_meta();
+
+        cache.put_parquet_meta_data(file_id, metadata, Some(region_metadata.clone()));
+
+        let cached = cache
+            .get_sst_meta_data(file_id, &mut metrics, Default::default())
+            .await
+            .unwrap();
+        assert!(Arc::ptr_eq(&region_metadata, &cached.region_metadata()));
+    }
+
+    #[test]
+    fn test_meta_cache_weight_accounts_for_decoded_region_metadata() {
+        let region_metadata = Arc::new(wide_region_metadata(128));
+        let json_len = region_metadata.to_json().unwrap().len();
+        let metadata = sst_parquet_meta_with_region_metadata(region_metadata.clone());
+        let cached = Arc::new(
+            CachedSstMeta::try_new("test.parquet", Arc::unwrap_or_clone(metadata)).unwrap(),
+        );
+        let key = SstMetaKey(region_metadata.region_id, FileId::random());
+
+        assert!(cached.region_metadata_weight > json_len);
+        assert_eq!(
+            meta_cache_weight(&key, &cached) as usize,
+            key.estimated_size()
+                + parquet_meta_size(&cached.parquet_metadata)
+                + cached.region_metadata_weight
+        );
+    }
+
     #[test]
     fn test_repeated_vector_cache() {
         let cache = CacheManager::builder().vector_cache_size(4096).build();
@@ -1023,11 +1447,98 @@ mod tests {
             selector: TimeSeriesRowSelector::LastRow,
         };
         assert!(cache.get_selector_result(&key).is_none());
-        let result = Arc::new(SelectorResultValue::new(Vec::new(), Vec::new()));
+        let result = Arc::new(SelectorResultValue::new(
+            Vec::new(),
+            ParquetReadColumns::from_deduped(Vec::new()),
+        ));
         cache.put_selector_result(key, result);
         assert!(cache.get_selector_result(&key).is_some());
     }
 
+    #[test]
+    fn test_range_result_cache() {
+        let cache = Arc::new(
+            CacheManager::builder()
+                .range_result_cache_size(1024 * 1024)
+                .build(),
+        );
+
+        let key = RangeScanCacheKey {
+            region_id: RegionId::new(1, 1),
+            row_groups: vec![(FileId::random(), 0)],
+            scan: ScanRequestFingerprintBuilder {
+                read_columns: ReadColumns::from_deduped_column_ids(std::iter::empty()),
+                read_column_types: vec![],
+                filters: vec!["tag_0 = 1".to_string()],
+                time_filters: vec![],
+                series_row_selector: None,
+                append_mode: false,
+                filter_deleted: true,
+                merge_mode: crate::region::options::MergeMode::LastRow,
+                partition_expr_version: 0,
+            }
+            .build(),
+        };
+        let value = Arc::new(RangeScanCacheValue::new(Vec::new(), 0));
+
+        assert!(cache.get_range_result(&key).is_none());
+        cache.put_range_result(key.clone(), value.clone());
+        assert!(cache.get_range_result(&key).is_some());
+
+        let enable_all = CacheStrategy::EnableAll(cache.clone());
+        assert!(enable_all.get_range_result(&key).is_some());
+
+        let compaction = CacheStrategy::Compaction(cache.clone());
+        assert!(compaction.get_range_result(&key).is_none());
+        compaction.put_range_result(key.clone(), value.clone());
+        assert!(cache.get_range_result(&key).is_some());
+
+        let disabled = CacheStrategy::Disabled;
+        assert!(disabled.get_range_result(&key).is_none());
+        disabled.put_range_result(key.clone(), value);
+        assert!(cache.get_range_result(&key).is_some());
+    }
+
+    #[test]
+    fn test_range_result_cache_size_configures_limiter() {
+        let cache_size = 3 * 1024_u64;
+        let cache = CacheManager::builder()
+            .range_result_cache_size(cache_size)
+            .build();
+
+        assert_eq!(cache.range_result_cache_size(), cache_size as usize);
+        assert_eq!(
+            cache.range_result_memory_limiter().permit_bytes(),
+            RANGE_RESULT_CONCAT_MEMORY_PERMIT.as_bytes() as usize
+        );
+        assert_eq!(
+            cache.range_result_memory_limiter().available_permits(),
+            (cache_size as usize).div_ceil(RANGE_RESULT_CONCAT_MEMORY_PERMIT.as_bytes() as usize)
+        );
+    }
+
+    #[tokio::test]
+    async fn range_result_memory_limiter_rejects_oversized_request() {
+        let limiter = RangeResultMemoryLimiter::new(2 * 1024, 1024);
+        assert_eq!(limiter.available_permits(), 2);
+
+        let err = limiter.acquire(10 * 1024).await.unwrap_err();
+        assert!(
+            err.to_string().contains("exceeds limiter capacity"),
+            "unexpected error: {err}"
+        );
+        assert_eq!(limiter.available_permits(), 2);
+    }
+
+    #[tokio::test]
+    async fn range_result_memory_limiter_allows_request_up_to_capacity() {
+        let limiter = RangeResultMemoryLimiter::new(2 * 1024, 1024);
+        let permit = limiter.acquire(2 * 1024).await.unwrap();
+        assert_eq!(limiter.available_permits(), 0);
+        drop(permit);
+        assert_eq!(limiter.available_permits(), 2);
+    }
+
     #[tokio::test]
     async fn test_evict_puffin_cache_clears_all_entries() {
         use std::collections::{BTreeMap, HashMap};
@@ -1122,4 +1633,45 @@ mod tests {
         assert!(result_cache.get(&predicate, index_id.file_id()).is_none());
         assert!(puffin_metadata_cache.get_metadata(&file_id_str).is_none());
     }
+
+    fn wide_region_metadata(column_count: u32) -> RegionMetadata {
+        let region_id = RegionId::new(1024, 7);
+        let mut builder = RegionMetadataBuilder::new(region_id);
+        let mut primary_key = Vec::new();
+
+        for column_id in 0..column_count {
+            let semantic_type = if column_id < 32 {
+                primary_key.push(column_id);
+                SemanticType::Tag
+            } else {
+                SemanticType::Field
+            };
+            let mut column_schema = ColumnSchema::new(
+                format!("wide_column_{column_id}"),
+                ConcreteDataType::string_datatype(),
+                true,
+            );
+            column_schema
+                .mut_metadata()
+                .insert(format!("cache_key_{column_id}"), "cache_value".repeat(4));
+            builder.push_column_metadata(ColumnMetadata {
+                column_schema,
+                semantic_type,
+                column_id,
+            });
+        }
+
+        builder.push_column_metadata(ColumnMetadata {
+            column_schema: ColumnSchema::new(
+                "ts",
+                ConcreteDataType::timestamp_millisecond_datatype(),
+                false,
+            ),
+            semantic_type: SemanticType::Timestamp,
+            column_id: column_count,
+        });
+        builder.primary_key(primary_key);
+
+        builder.build().unwrap()
+    }
 }
diff --git a/src/mito2/src/cache/file_cache.rs b/src/mito2/src/cache/file_cache.rs
index 32a276d0e4..9b987c810b 100644
--- a/src/mito2/src/cache/file_cache.rs
+++ b/src/mito2/src/cache/file_cache.rs
@@ -34,7 +34,7 @@ use store_api::storage::{FileId, RegionId};
 use tokio::sync::mpsc::{Sender, UnboundedReceiver};
 
 use crate::access_layer::TempFileCleaner;
-use crate::cache::{FILE_TYPE, INDEX_TYPE};
+use crate::cache::{CachedSstMeta, FILE_TYPE, INDEX_TYPE};
 use crate::error::{self, OpenDalSnafu, Result};
 use crate::metrics::{
     CACHE_BYTES, CACHE_HIT, CACHE_MISS, WRITE_CACHE_DOWNLOAD_BYTES_TOTAL,
@@ -288,6 +288,17 @@ pub(crate) struct FileCache {
 pub(crate) type FileCacheRef = Arc<FileCache>;
 
 impl FileCache {
+    /// Splits the configured total capacity between parquet and puffin caches
+    /// without exceeding the requested overall budget.
+    fn split_cache_capacities(total_capacity: u64, index_percent: u8) -> (u64, u64) {
+        let desired_puffin_capacity = total_capacity * u64::from(index_percent) / 100;
+        let min_cache_capacity = MIN_CACHE_CAPACITY.min(total_capacity / 2);
+        let puffin_capacity =
+            desired_puffin_capacity.clamp(min_cache_capacity, total_capacity - min_cache_capacity);
+        let parquet_capacity = total_capacity - puffin_capacity;
+        (parquet_capacity, puffin_capacity)
+    }
+
     /// Creates a new file cache.
     pub(crate) fn new(
         local_store: ObjectStore,
@@ -302,14 +313,8 @@ impl FileCache {
             .unwrap_or(DEFAULT_INDEX_CACHE_PERCENT);
         let total_capacity = capacity.as_bytes();
 
-        // Convert percent to ratio and calculate capacity for each cache
-        let index_ratio = index_percent as f64 / 100.0;
-        let puffin_capacity = (total_capacity as f64 * index_ratio) as u64;
-        let parquet_capacity = total_capacity - puffin_capacity;
-
-        // Ensure both capacities are at least 512MB
-        let puffin_capacity = puffin_capacity.max(MIN_CACHE_CAPACITY);
-        let parquet_capacity = parquet_capacity.max(MIN_CACHE_CAPACITY);
+        let (parquet_capacity, puffin_capacity) =
+            Self::split_cache_capacities(total_capacity, index_percent);
 
         info!(
             "Initializing file cache with index_percent: {}%, total_capacity: {}, parquet_capacity: {}, puffin_capacity: {}",
@@ -612,6 +617,34 @@ impl FileCache {
         }
     }
 
+    /// Get fused SST metadata from the file cache.
+    /// If the file is not in the cache, or metadata loading/decoding fails, return None.
+    pub(crate) async fn get_sst_meta_data(
+        &self,
+        key: IndexKey,
+        cache_metrics: &mut MetadataCacheMetrics,
+        page_index_policy: PageIndexPolicy,
+    ) -> Option<Arc<CachedSstMeta>> {
+        let file_path = self.inner.cache_file_path(key);
+        self.get_parquet_meta_data(key, cache_metrics, page_index_policy)
+            .await
+            .and_then(
+                |metadata| match CachedSstMeta::try_new(&file_path, metadata) {
+                    Ok(metadata) => Some(Arc::new(metadata)),
+                    Err(err) => {
+                        CACHE_MISS
+                            .with_label_values(&[key.file_type.metric_label()])
+                            .inc();
+                        warn!(
+                            err; "Failed to decode cached parquet metadata for key {:?}",
+                            key
+                        );
+                        None
+                    }
+                },
+            )
+    }
+
     async fn get_reader(&self, file_path: &str) -> object_store::Result<Option<Reader>> {
         if self.inner.local_store.exists(file_path).await? {
             Ok(Some(self.inner.local_store.reader(file_path).await?))
@@ -1036,6 +1069,28 @@ mod tests {
         assert_eq!(data, bytes[3].as_ref());
     }
 
+    #[test]
+    fn test_file_cache_capacity_respects_total_budget() {
+        let total_capacity = ReadableSize::mb(256).as_bytes();
+        let (parquet_capacity, puffin_capacity) =
+            FileCache::split_cache_capacities(total_capacity, 20);
+
+        assert_eq!(total_capacity, parquet_capacity + puffin_capacity);
+        assert_eq!(ReadableSize::mb(128).as_bytes(), parquet_capacity);
+        assert_eq!(ReadableSize::mb(128).as_bytes(), puffin_capacity);
+    }
+
+    #[test]
+    fn test_file_cache_capacity_keeps_split_when_total_allows_it() {
+        let total_capacity = ReadableSize::gb(5).as_bytes();
+        let (parquet_capacity, puffin_capacity) =
+            FileCache::split_cache_capacities(total_capacity, 20);
+
+        assert_eq!(total_capacity, parquet_capacity + puffin_capacity);
+        assert_eq!(ReadableSize::gb(4).as_bytes(), parquet_capacity);
+        assert_eq!(ReadableSize::gb(1).as_bytes(), puffin_capacity);
+    }
+
     #[test]
     fn test_cache_file_path() {
         let file_id = FileId::parse_str("3368731b-a556-42b8-a5df-9c31ce155095").unwrap();
diff --git a/src/mito2/src/cache/index.rs b/src/mito2/src/cache/index.rs
index 74f16fdb71..a404978714 100644
--- a/src/mito2/src/cache/index.rs
+++ b/src/mito2/src/cache/index.rs
@@ -229,14 +229,14 @@ where
         }
         if !cache_miss_range.is_empty() {
             let pages = load(cache_miss_range).await?;
-            for (i, page) in cache_miss_idx.into_iter().zip(pages.into_iter()) {
+            for (i, page) in cache_miss_idx.into_iter().zip(pages) {
                 let page_key = page_keys[i];
                 metrics.page_bytes += page.len() as u64;
                 data[i] = page.clone();
                 self.put_page(key, page_key, page.clone());
             }
         }
-        let buffer = Buffer::from_iter(data.into_iter());
+        let buffer = Buffer::from_iter(data);
         Ok((
             buffer
                 .slice(PageKey::calculate_range(offset, size, self.page_size))
diff --git a/src/mito2/src/cache/manifest_cache.rs b/src/mito2/src/cache/manifest_cache.rs
index 458b199386..08ab0fb14a 100644
--- a/src/mito2/src/cache/manifest_cache.rs
+++ b/src/mito2/src/cache/manifest_cache.rs
@@ -378,7 +378,7 @@ impl ManifestCache {
                     warn!(e; "Failed to remove empty root dir {}", dir.display());
                     return Err(e);
                 } else {
-                    warn!("Empty root dir not found before removal {}", dir.display());
+                    info!("Empty root dir not found before removal {}", dir.display());
                 }
             } else {
                 info!(
diff --git a/src/mito2/src/cache/test_util.rs b/src/mito2/src/cache/test_util.rs
index 65ad9d87eb..ef3d8e9315 100644
--- a/src/mito2/src/cache/test_util.rs
+++ b/src/mito2/src/cache/test_util.rs
@@ -23,8 +23,13 @@ use object_store::ObjectStore;
 use object_store::services::Fs;
 use parquet::arrow::ArrowWriter;
 use parquet::arrow::arrow_reader::ParquetRecordBatchReaderBuilder;
-use parquet::file::metadata::ParquetMetaData;
+use parquet::file::metadata::{KeyValue, ParquetMetaData};
+use parquet::file::properties::WriterProperties;
 use parquet::file::statistics::Statistics;
+use store_api::metadata::RegionMetadataRef;
+
+use crate::sst::parquet::PARQUET_METADATA_KEY;
+use crate::test_util::sst_util::sst_region_metadata;
 
 /// Returns a parquet meta data.
 pub(crate) fn parquet_meta() -> Arc<ParquetMetaData> {
@@ -33,13 +38,43 @@ pub(crate) fn parquet_meta() -> Arc<ParquetMetaData> {
     builder.metadata().clone()
 }
 
+/// Returns parquet metadata for an SST parquet file and its decoded region metadata.
+pub(crate) fn sst_parquet_meta() -> (Arc<ParquetMetaData>, RegionMetadataRef) {
+    let region_metadata = Arc::new(sst_region_metadata());
+    let file_data = parquet_file_data_with_region_metadata(&region_metadata);
+    let builder = ParquetRecordBatchReaderBuilder::try_new(Bytes::from(file_data)).unwrap();
+    (builder.metadata().clone(), region_metadata)
+}
+
+/// Returns parquet metadata for an SST parquet file with custom region metadata.
+pub(crate) fn sst_parquet_meta_with_region_metadata(
+    region_metadata: RegionMetadataRef,
+) -> Arc<ParquetMetaData> {
+    let file_data = parquet_file_data_with_region_metadata(&region_metadata);
+    let builder = ParquetRecordBatchReaderBuilder::try_new(Bytes::from(file_data)).unwrap();
+    builder.metadata().clone()
+}
+
 /// Write a test parquet file to a buffer
 fn parquet_file_data() -> Vec<u8> {
+    parquet_file_data_inner(None)
+}
+
+fn parquet_file_data_with_region_metadata(region_metadata: &RegionMetadataRef) -> Vec<u8> {
+    let json = region_metadata.to_json().unwrap();
+    let key_value = KeyValue::new(PARQUET_METADATA_KEY.to_string(), json);
+    parquet_file_data_inner(Some(vec![key_value]))
+}
+
+fn parquet_file_data_inner(key_value_metadata: Option<Vec<KeyValue>>) -> Vec<u8> {
     let col = Arc::new(Int64Array::from_iter_values([1, 2, 3])) as ArrayRef;
     let to_write = RecordBatch::try_from_iter([("col", col)]).unwrap();
 
     let mut buffer = Vec::new();
-    let mut writer = ArrowWriter::try_new(&mut buffer, to_write.schema(), None).unwrap();
+    let props = WriterProperties::builder()
+        .set_key_value_metadata(key_value_metadata)
+        .build();
+    let mut writer = ArrowWriter::try_new(&mut buffer, to_write.schema(), Some(props)).unwrap();
     writer.write(&to_write).unwrap();
     writer.close().unwrap();
 
diff --git a/src/mito2/src/cache/write_cache.rs b/src/mito2/src/cache/write_cache.rs
index a28df3f54c..6b02a92811 100644
--- a/src/mito2/src/cache/write_cache.rs
+++ b/src/mito2/src/cache/write_cache.rs
@@ -244,15 +244,19 @@ impl WriteCache {
         .await
         .with_file_cleaner(cleaner);
 
-        let sst_info = match write_request.source {
-            either::Left(source) => {
+        let sst_info = match write_request.sst_write_format {
+            crate::sst::FormatType::PrimaryKey => {
                 writer
-                    .write_all(source, write_request.max_sequence, write_opts)
+                    .write_all_flat_as_primary_key(
+                        write_request.source,
+                        write_request.max_sequence,
+                        write_opts,
+                    )
                     .await?
             }
-            either::Right(flat_source) => {
+            crate::sst::FormatType::Flat => {
                 writer
-                    .write_all_flat(flat_source, write_request.max_sequence, write_opts)
+                    .write_all_flat(write_request.source, write_request.max_sequence, write_opts)
                     .await?
             }
         };
@@ -509,12 +513,13 @@ mod tests {
     use crate::cache::test_util::{assert_parquet_metadata_equal, new_fs_store};
     use crate::cache::{CacheManager, CacheStrategy};
     use crate::error::InvalidBatchSnafu;
-    use crate::read::Source;
+    use crate::read::FlatSource;
     use crate::region::options::IndexOptions;
     use crate::sst::parquet::reader::ParquetReaderBuilder;
     use crate::test_util::TestEnv;
     use crate::test_util::sst_util::{
-        new_batch_by_range, new_source, sst_file_handle_with_file_id, sst_region_metadata,
+        new_flat_source_from_record_batches, new_record_batch_by_range,
+        sst_file_handle_with_file_id, sst_region_metadata,
     };
 
     #[tokio::test]
@@ -532,21 +537,22 @@ mod tests {
             .create_write_cache(local_store.clone(), ReadableSize::mb(10))
             .await;
 
-        // Create Source
+        // Create source.
         let metadata = Arc::new(sst_region_metadata());
         let region_id = metadata.region_id;
-        let source = new_source(&[
-            new_batch_by_range(&["a", "d"], 0, 60),
-            new_batch_by_range(&["b", "f"], 0, 40),
-            new_batch_by_range(&["b", "h"], 100, 200),
+        let source = new_flat_source_from_record_batches(vec![
+            new_record_batch_by_range(&["a", "d"], 0, 60),
+            new_record_batch_by_range(&["b", "f"], 0, 40),
+            new_record_batch_by_range(&["b", "h"], 100, 200),
         ]);
 
         let write_request = SstWriteRequest {
             op_type: OperationType::Flush,
             metadata,
-            source: either::Left(source),
+            source,
             storage: None,
             max_sequence: None,
+            sst_write_format: Default::default(),
             cache_manager: Default::default(),
             index_options: IndexOptions::default(),
             index_config: Default::default(),
@@ -636,19 +642,20 @@ mod tests {
         // Create source
         let metadata = Arc::new(sst_region_metadata());
 
-        let source = new_source(&[
-            new_batch_by_range(&["a", "d"], 0, 60),
-            new_batch_by_range(&["b", "f"], 0, 40),
-            new_batch_by_range(&["b", "h"], 100, 200),
+        let source = new_flat_source_from_record_batches(vec![
+            new_record_batch_by_range(&["a", "d"], 0, 60),
+            new_record_batch_by_range(&["b", "f"], 0, 40),
+            new_record_batch_by_range(&["b", "h"], 100, 200),
         ]);
 
         // Write to local cache and upload sst to mock remote store
         let write_request = SstWriteRequest {
             op_type: OperationType::Flush,
             metadata,
-            source: either::Left(source),
+            source,
             storage: None,
             max_sequence: None,
+            sst_write_format: Default::default(),
             cache_manager: cache_manager.clone(),
             index_options: IndexOptions::default(),
             index_config: Default::default(),
@@ -686,9 +693,15 @@ mod tests {
         .cache(CacheStrategy::EnableAll(cache_manager.clone()))
         .page_index_policy(PageIndexPolicy::Optional);
         let reader = builder.build().await.unwrap().unwrap();
+        let cached_write_parquet_metadata = crate::cache::CachedSstMeta::try_new(
+            "test.sst",
+            Arc::unwrap_or_clone(write_parquet_metadata),
+        )
+        .unwrap()
+        .parquet_metadata();
 
         // Check parquet metadata
-        assert_parquet_metadata_equal(write_parquet_metadata, reader.parquet_metadata());
+        assert_parquet_metadata_equal(cached_write_parquet_metadata, reader.parquet_metadata());
     }
 
     #[tokio::test]
@@ -715,24 +728,28 @@ mod tests {
         let metadata = Arc::new(sst_region_metadata());
 
         // Creates a source that can return an error to abort the writer.
-        let source = Source::Iter(Box::new(
+        let record_batch = new_record_batch_by_range(&["a", "d"], 0, 60);
+        let schema = record_batch.schema();
+        let iter = Box::new(
             [
-                Ok(new_batch_by_range(&["a", "d"], 0, 60)),
+                Ok(record_batch),
                 InvalidBatchSnafu {
                     reason: "Abort the writer",
                 }
                 .fail(),
             ]
             .into_iter(),
-        ));
+        );
+        let source = FlatSource::new_iter(schema, iter);
 
         // Write to local cache and upload sst to mock remote store
         let write_request = SstWriteRequest {
             op_type: OperationType::Flush,
             metadata,
-            source: either::Left(source),
+            source,
             storage: None,
             max_sequence: None,
+            sst_write_format: Default::default(),
             cache_manager: cache_manager.clone(),
             index_options: IndexOptions::default(),
             index_config: Default::default(),
diff --git a/src/mito2/src/compaction.rs b/src/mito2/src/compaction.rs
index 6d51d1dd59..1d1151177d 100644
--- a/src/mito2/src/compaction.rs
+++ b/src/mito2/src/compaction.rs
@@ -24,12 +24,13 @@ mod twcs;
 mod window;
 
 use std::collections::HashMap;
-use std::sync::Arc;
+use std::sync::{Arc, Mutex};
 use std::time::Instant;
 
 use api::v1::region::compact_request;
 use api::v1::region::compact_request::Options;
 use common_base::Plugins;
+use common_base::cancellation::CancellationHandle;
 use common_memory_manager::OnExhaustedPolicy;
 use common_meta::key::SchemaMetadataManagerRef;
 use common_telemetry::{debug, error, info, warn};
@@ -53,15 +54,15 @@ use crate::compaction::picker::{CompactionTask, PickerOutput, new_picker};
 use crate::compaction::task::CompactionTaskImpl;
 use crate::config::MitoConfig;
 use crate::error::{
-    CompactRegionSnafu, Error, GetSchemaMetadataSnafu, ManualCompactionOverrideSnafu,
-    RegionClosedSnafu, RegionDroppedSnafu, RegionTruncatedSnafu, RemoteCompactionSnafu, Result,
-    TimeRangePredicateOverflowSnafu, TimeoutSnafu,
+    CompactRegionSnafu, CompactionCancelledSnafu, Error, GetSchemaMetadataSnafu,
+    ManualCompactionOverrideSnafu, RegionClosedSnafu, RegionDroppedSnafu, RegionTruncatedSnafu,
+    RemoteCompactionSnafu, Result, TimeRangePredicateOverflowSnafu, TimeoutSnafu,
 };
 use crate::metrics::{COMPACTION_STAGE_ELAPSED, INFLIGHT_COMPACTION_COUNT};
-use crate::read::projection::ProjectionMapper;
+use crate::read::FlatSource;
+use crate::read::flat_projection::FlatProjectionMapper;
 use crate::read::scan_region::{PredicateGroup, ScanInput};
 use crate::read::seq_scan::SeqScan;
-use crate::read::{BoxedBatchReader, BoxedRecordBatchStream};
 use crate::region::options::{MergeMode, RegionOptions};
 use crate::region::version::VersionControlRef;
 use crate::region::{ManifestContextRef, RegionLeaderState, RegionRoleState};
@@ -186,7 +187,7 @@ impl CompactionScheduler {
         }
 
         // The region can compact directly.
-        let mut status: CompactionStatus =
+        let mut status =
             CompactionStatus::new(region_id, version_control.clone(), access_layer.clone());
         let request = status.new_compaction_request(
             self.request_sender.clone(),
@@ -199,17 +200,25 @@ impl CompactionScheduler {
             max_parallelism,
         );
 
-        let result = self
+        let result = match self
             .schedule_compaction_request(request, compact_options)
-            .await;
-        if matches!(result, Ok(true)) {
-            // Only if the compaction request is scheduled successfully,
-            // we insert the region into the status map.
-            self.region_status.insert(region_id, status);
-        }
+            .await
+        {
+            Ok(Some(active_compaction)) => {
+                // Publish CompactionStatus only after a task has been accepted by the scheduler.
+                // This avoids exposing a half-initialized region status that could collect pending
+                // DDL/compaction state even though no compaction is actually running.
+                status.active_compaction = Some(active_compaction);
+                self.region_status.insert(region_id, status);
+
+                Ok(())
+            }
+            Ok(None) => Ok(()),
+            Err(e) => Err(e),
+        };
 
         self.listener.on_compaction_scheduled(region_id);
-        result.map(|_| ())
+        result
     }
 
     // Handle pending manual compaction request for the region.
@@ -251,14 +260,16 @@ impl CompactionScheduler {
         };
 
         match self.schedule_compaction_request(request, options).await {
-            Ok(true) => {
+            Ok(Some(active_compaction)) => {
+                let status = self.region_status.get_mut(&region_id).unwrap();
+                status.active_compaction = Some(active_compaction);
                 debug!(
                     "Successfully scheduled manual compaction for region id: {}",
                     region_id
                 );
                 true
             }
-            Ok(false) => {
+            Ok(None) => {
                 // We still need to handle the pending DDL requests.
                 // So we can't return early here.
                 false
@@ -278,6 +289,11 @@ impl CompactionScheduler {
         manifest_ctx: &ManifestContextRef,
         schema_metadata_manager: SchemaMetadataManagerRef,
     ) -> Vec<SenderDdlRequest> {
+        let Some(status) = self.region_status.get_mut(&region_id) else {
+            return Vec::new();
+        };
+        status.clear_running_task();
+
         // If there a pending compaction request, handle it first
         // and defer returning the pending DDL requests to the caller.
         if self
@@ -297,7 +313,6 @@ impl CompactionScheduler {
             return Vec::new();
         };
 
-        // Notify all waiters that compaction is finished.
         for waiter in std::mem::take(&mut status.waiters) {
             waiter.send(Ok(0));
         }
@@ -331,13 +346,17 @@ impl CompactionScheduler {
             )
             .await
         {
-            Ok(true) => {
+            Ok(Some(active_compaction)) => {
+                self.region_status
+                    .get_mut(&region_id)
+                    .unwrap()
+                    .active_compaction = Some(active_compaction);
                 debug!(
                     "Successfully scheduled next compaction for region id: {}",
                     region_id
                 );
             }
-            Ok(false) => {
+            Ok(None) => {
                 // No further compaction tasks can be scheduled; cleanup the `CompactionStatus` for this region.
                 // All DDL requests and pending compaction requests have already been processed.
                 // Safe to remove the region from status tracking.
@@ -352,6 +371,14 @@ impl CompactionScheduler {
         Vec::new()
     }
 
+    /// Notifies the scheduler that the compaction job is cancelled cooperatively.
+    pub(crate) async fn on_compaction_cancelled(
+        &mut self,
+        region_id: RegionId,
+    ) -> Vec<SenderDdlRequest> {
+        self.remove_region_on_cancel(region_id)
+    }
+
     /// Notifies the scheduler that the compaction job is failed.
     pub(crate) fn on_compaction_failed(&mut self, region_id: RegionId, err: Arc<Error>) {
         error!(err; "Region {} failed to compact, cancel all pending tasks", region_id);
@@ -406,20 +433,23 @@ impl CompactionScheduler {
         has_pending
     }
 
-    /// Returns true if the region is compacting.
-    pub(crate) fn is_compacting(&self, region_id: RegionId) -> bool {
-        self.region_status.contains_key(&region_id)
+    pub(crate) fn request_cancel(&mut self, region_id: RegionId) -> RequestCancelResult {
+        let Some(status) = self.region_status.get_mut(&region_id) else {
+            return RequestCancelResult::NotRunning;
+        };
+
+        status.request_cancel()
     }
 
     /// Schedules a compaction request.
     ///
-    /// Returns true if the compaction request is scheduled successfully.
-    /// Returns false if no compaction task can be scheduled for this region.
+    /// Returns the active compaction state if the request is scheduled successfully.
+    /// Returns `None` if no compaction task can be scheduled for this region.
     async fn schedule_compaction_request(
         &mut self,
         request: CompactionRequest,
         options: compact_request::Options,
-    ) -> Result<bool> {
+    ) -> Result<Option<ActiveCompaction>> {
         let region_id = request.region_id();
         let (dynamic_compaction_opts, ttl) = find_dynamic_options(
             region_id.table_id(),
@@ -492,7 +522,7 @@ impl CompactionScheduler {
             for waiter in waiters {
                 waiter.send(Ok(0));
             }
-            return Ok(false);
+            return Ok(None);
         };
 
         // If specified to run compaction remotely, we schedule the compaction job remotely.
@@ -523,7 +553,7 @@ impl CompactionScheduler {
                             job_id, region_id
                         );
                         INFLIGHT_COMPACTION_COUNT.inc();
-                        return Ok(true);
+                        return Ok(Some(ActiveCompaction::Remote));
                     }
                     Err(e) => {
                         if !dynamic_compaction_opts.fallback_to_local() {
@@ -555,21 +585,25 @@ impl CompactionScheduler {
 
         // Create a local compaction task.
         let estimated_bytes = estimate_compaction_bytes(&picker_output);
+
+        let cancel_handle = Arc::new(CancellationHandle::default());
+        let state = LocalCompactionState::new(cancel_handle.clone());
         let local_compaction_task = Box::new(CompactionTaskImpl {
+            state: state.clone(),
             request_sender,
             waiters,
             start_time,
             listener,
             picker_output,
             compaction_region,
-            compactor: Arc::new(DefaultCompactor {}),
+            compactor: Arc::new(DefaultCompactor::with_cancel_handle(cancel_handle.clone())),
             memory_manager: self.memory_manager.clone(),
             memory_policy: self.memory_policy,
             estimated_memory_bytes: estimated_bytes,
         });
 
         self.submit_compaction_task(local_compaction_task, region_id)
-            .map(|_| true)
+            .map(|_| Some(ActiveCompaction::Local { state }))
     }
 
     fn submit_compaction_task(
@@ -597,6 +631,77 @@ impl CompactionScheduler {
         // Notifies all pending tasks.
         status.on_failure(err);
     }
+
+    fn remove_region_on_cancel(&mut self, region_id: RegionId) -> Vec<SenderDdlRequest> {
+        let Some(status) = self.region_status.remove(&region_id) else {
+            return Vec::new();
+        };
+
+        status.on_cancel()
+    }
+}
+
+#[derive(Debug, Clone)]
+pub(crate) struct LocalCompactionState {
+    cancel_handle: Arc<CancellationHandle>,
+    commit_started: Arc<Mutex<bool>>,
+}
+
+#[derive(Debug)]
+enum ActiveCompaction {
+    Local { state: LocalCompactionState },
+    Remote,
+}
+
+impl LocalCompactionState {
+    fn new(cancel_handle: Arc<CancellationHandle>) -> Self {
+        Self {
+            cancel_handle,
+            commit_started: Arc::new(Mutex::new(false)),
+        }
+    }
+
+    /// Returns the cancellation handle for this compaction task.
+    pub(crate) fn cancel_handle(&self) -> Arc<CancellationHandle> {
+        self.cancel_handle.clone()
+    }
+
+    /// Marks the compaction task as started to commit,
+    /// which means the compaction task is in the final stage and is about to update region version and manifest.
+    /// It will reject cancellation request after this method is called.
+    ///
+    /// Returns true if this is the first time to mark commit started, false otherwise.
+    pub(crate) fn mark_commit_started(&self) -> bool {
+        let mut commit_started = self.commit_started.lock().unwrap();
+        if self.cancel_handle.is_cancelled() {
+            return false;
+        }
+        *commit_started = true;
+        true
+    }
+
+    /// Request cancellation for this compaction task.
+    pub(crate) fn request_cancel(&self) -> RequestCancelResult {
+        // The cancel handle must under the lock of `commit_started` to avoid racing between cancellation and commit.
+        let commit_started = self.commit_started.lock().unwrap();
+        if *commit_started {
+            return RequestCancelResult::TooLateToCancel;
+        }
+        if self.cancel_handle.is_cancelled() {
+            return RequestCancelResult::AlreadyCancelling;
+        }
+
+        self.cancel_handle.cancel();
+        RequestCancelResult::CancelIssued
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(crate) enum RequestCancelResult {
+    CancelIssued,
+    AlreadyCancelling,
+    TooLateToCancel,
+    NotRunning,
 }
 
 impl Drop for CompactionScheduler {
@@ -614,15 +719,12 @@ async fn find_dynamic_options(
     region_options: &crate::region::options::RegionOptions,
     schema_metadata_manager: &SchemaMetadataManagerRef,
 ) -> Result<(crate::region::options::CompactionOptions, TimeToLive)> {
-    if region_options.compaction_override && region_options.ttl.is_some() {
+    if let (true, Some(ttl)) = (region_options.compaction_override, region_options.ttl) {
         debug!(
             "Use region options directly for table {}: compaction={:?}, ttl={:?}",
             table_id, region_options.compaction, region_options.ttl
         );
-        return Ok((
-            region_options.compaction.clone(),
-            region_options.ttl.unwrap(),
-        ));
+        return Ok((region_options.compaction.clone(), ttl));
     }
 
     let db_options = tokio::time::timeout(
@@ -633,12 +735,12 @@ async fn find_dynamic_options(
     .context(TimeoutSnafu)?
     .context(GetSchemaMetadataSnafu)?;
 
-    let ttl = if region_options.ttl.is_some() {
+    let ttl = if let Some(ttl) = region_options.ttl {
         debug!(
             "Use region TTL directly for table {}: ttl={:?}",
             table_id, region_options.ttl
         );
-        region_options.ttl.unwrap()
+        ttl
     } else {
         db_options
             .as_ref()
@@ -706,6 +808,8 @@ struct CompactionStatus {
     pending_request: Option<PendingCompaction>,
     /// Pending DDL requests that should run when compaction is done.
     pending_ddl_requests: Vec<SenderDdlRequest>,
+    /// Active compaction state.
+    active_compaction: Option<ActiveCompaction>,
 }
 
 impl CompactionStatus {
@@ -722,9 +826,39 @@ impl CompactionStatus {
             waiters: Vec::new(),
             pending_request: None,
             pending_ddl_requests: Vec::new(),
+            active_compaction: None,
         }
     }
 
+    #[cfg(test)]
+    fn start_local_task(&mut self) -> LocalCompactionState {
+        let state = LocalCompactionState::new(Arc::new(CancellationHandle::default()));
+        self.active_compaction = Some(ActiveCompaction::Local {
+            state: state.clone(),
+        });
+        state
+    }
+
+    #[cfg(test)]
+    fn start_remote_task(&mut self) {
+        self.active_compaction = Some(ActiveCompaction::Remote);
+    }
+
+    fn request_cancel(&mut self) -> RequestCancelResult {
+        let Some(active_compaction) = &self.active_compaction else {
+            return RequestCancelResult::NotRunning;
+        };
+
+        match active_compaction {
+            ActiveCompaction::Local { state, .. } => state.request_cancel(),
+            ActiveCompaction::Remote => RequestCancelResult::TooLateToCancel,
+        }
+    }
+
+    fn clear_running_task(&mut self) -> bool {
+        self.active_compaction.take().is_some()
+    }
+
     /// Merge the waiter to the pending compaction.
     fn merge_waiter(&mut self, mut waiter: OptionOutputTx) {
         if let Some(waiter) = waiter.take_inner() {
@@ -767,6 +901,23 @@ impl CompactionStatus {
         }
     }
 
+    #[must_use]
+    fn on_cancel(mut self) -> Vec<SenderDdlRequest> {
+        for waiter in self.waiters.drain(..) {
+            waiter.send(CompactionCancelledSnafu.fail());
+        }
+
+        if let Some(pending_compaction) = self.pending_request {
+            pending_compaction.waiter.send(
+                Err(Arc::new(CompactionCancelledSnafu.build())).context(CompactRegionSnafu {
+                    region_id: self.region_id,
+                }),
+            );
+        }
+
+        std::mem::take(&mut self.pending_ddl_requests)
+    }
+
     /// Creates a new compaction request for compaction picker.
     ///
     /// It consumes all pending compaction waiters.
@@ -828,7 +979,7 @@ pub struct SerializedCompactionOutput {
     output_time_range: Option<TimestampRange>,
 }
 
-/// Builders to create [BoxedBatchReader] for compaction.
+/// Builders to create [BoxedRecordBatchStream] for compaction.
 struct CompactionSstReaderBuilder<'a> {
     metadata: RegionMetadataRef,
     sst_layer: AccessLayerRef,
@@ -841,24 +992,22 @@ struct CompactionSstReaderBuilder<'a> {
 }
 
 impl CompactionSstReaderBuilder<'_> {
-    /// Builds [BoxedBatchReader] that reads all SST files and yields batches in primary key order.
-    async fn build_sst_reader(self) -> Result<BoxedBatchReader> {
-        let scan_input = self.build_scan_input(false)?.with_compaction(true);
+    /// Build a [FlatSource] that yields Arrow `RecordBatch`s from reading all the input SST files,
+    /// for compaction. The schema of the [FlatSource] is unified.
+    async fn build_flat_sst_reader(self) -> Result<FlatSource> {
+        let scan_input = self.build_scan_input()?.with_compaction(true);
 
-        SeqScan::new(scan_input).build_reader_for_compaction().await
-    }
-
-    /// Builds [BoxedRecordBatchStream] that reads all SST files and yields batches in flat format for compaction.
-    async fn build_flat_sst_reader(self) -> Result<BoxedRecordBatchStream> {
-        let scan_input = self.build_scan_input(true)?.with_compaction(true);
+        let schema = scan_input.mapper.output_schema();
+        let schema = schema.arrow_schema();
 
         SeqScan::new(scan_input)
             .build_flat_reader_for_compaction()
             .await
+            .map(|stream| FlatSource::new_stream(schema.clone(), stream))
     }
 
-    fn build_scan_input(self, flat_format: bool) -> Result<ScanInput> {
-        let mapper = ProjectionMapper::all(&self.metadata, flat_format)?;
+    fn build_scan_input(self) -> Result<ScanInput> {
+        let mapper = FlatProjectionMapper::all(&self.metadata)?;
         let mut scan_input = ScanInput::new(self.sst_layer, mapper)
             .with_files(self.inputs.to_vec())
             .with_append_mode(self.append_mode)
@@ -867,8 +1016,7 @@ impl CompactionSstReaderBuilder<'_> {
             .with_filter_deleted(self.filter_deleted)
             // We ignore file not found error during compaction.
             .with_ignore_file_not_found(true)
-            .with_merge_mode(self.merge_mode)
-            .with_flat_format(flat_format);
+            .with_merge_mode(self.merge_mode);
 
         // This serves as a workaround of https://github.com/GreptimeTeam/greptimedb/issues/3944
         // by converting time ranges into predicate.
@@ -987,7 +1135,7 @@ struct PendingCompaction {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::time::Duration;
 
     use api::v1::region::StrictWindow;
@@ -1373,6 +1521,58 @@ mod tests {
         );
     }
 
+    #[tokio::test]
+    async fn test_schedule_compaction_does_not_publish_status_when_schedule_fails() {
+        common_telemetry::init_default_ut_logging();
+        let env = SchedulerEnv::new()
+            .await
+            .scheduler(Arc::new(FailingScheduler));
+        let (tx, _rx) = mpsc::channel(4);
+        let mut scheduler = env.mock_compaction_scheduler(tx);
+        let mut builder = VersionControlBuilder::new();
+        let end = 1000 * 1000;
+        let version_control = Arc::new(
+            builder
+                .push_l0_file(0, end)
+                .push_l0_file(10, end)
+                .push_l0_file(50, end)
+                .push_l0_file(80, end)
+                .push_l0_file(90, end)
+                .build(),
+        );
+        let region_id = builder.region_id();
+        let manifest_ctx = env
+            .mock_manifest_context(version_control.current().version.metadata.clone())
+            .await;
+        let (schema_metadata_manager, kv_backend) = mock_schema_metadata_manager();
+        schema_metadata_manager
+            .register_region_table_info(
+                builder.region_id().table_id(),
+                "test_table",
+                "test_catalog",
+                "test_schema",
+                None,
+                kv_backend,
+            )
+            .await;
+
+        let result = scheduler
+            .schedule_compaction(
+                region_id,
+                compact_request::Options::Regular(Default::default()),
+                &version_control,
+                &env.access_layer,
+                OptionOutputTx::none(),
+                &manifest_ctx,
+                schema_metadata_manager,
+                1,
+            )
+            .await;
+
+        assert!(result.is_err());
+        assert!(!scheduler.region_status.contains_key(&region_id));
+    }
+
     #[tokio::test]
     async fn test_manual_compaction_when_compaction_in_progress() {
         common_telemetry::init_default_ut_logging();
@@ -1553,6 +1753,11 @@ mod tests {
             region_id,
             CompactionStatus::new(region_id, version_control, env.access_layer.clone()),
         );
+        scheduler
+            .region_status
+            .get_mut(&region_id)
+            .unwrap()
+            .start_local_task();
 
         let (output_tx, _output_rx) = oneshot::channel();
         scheduler.add_ddl_request_to_pending(SenderDdlRequest {
@@ -1569,6 +1774,142 @@ mod tests {
         assert!(scheduler.has_pending_ddls(region_id));
     }
 
+    #[tokio::test]
+    async fn test_request_cancel_state_transitions() {
+        let env = SchedulerEnv::new().await;
+        let builder = VersionControlBuilder::new();
+        let region_id = builder.region_id();
+        let version_control = Arc::new(builder.build());
+        let mut status =
+            CompactionStatus::new(region_id, version_control, env.access_layer.clone());
+        let state = status.start_local_task();
+
+        assert_eq!(status.request_cancel(), RequestCancelResult::CancelIssued);
+        assert!(state.cancel_handle().is_cancelled());
+        assert_eq!(
+            status.request_cancel(),
+            RequestCancelResult::AlreadyCancelling
+        );
+
+        assert!(!state.mark_commit_started());
+        assert_eq!(
+            status.request_cancel(),
+            RequestCancelResult::AlreadyCancelling
+        );
+
+        assert!(status.clear_running_task());
+        assert_eq!(status.request_cancel(), RequestCancelResult::NotRunning);
+    }
+
+    #[tokio::test]
+    async fn test_request_cancel_remote_compaction_is_too_late() {
+        let env = SchedulerEnv::new().await;
+        let builder = VersionControlBuilder::new();
+        let region_id = builder.region_id();
+        let version_control = Arc::new(builder.build());
+        let mut status =
+            CompactionStatus::new(region_id, version_control, env.access_layer.clone());
+
+        status.start_remote_task();
+
+        assert_eq!(
+            status.request_cancel(),
+            RequestCancelResult::TooLateToCancel
+        );
+        assert!(status.active_compaction.is_some());
+    }
+
+    #[tokio::test]
+    async fn test_on_compaction_cancelled_returns_pending_ddl_requests() {
+        let job_scheduler = Arc::new(VecScheduler::default());
+        let env = SchedulerEnv::new().await.scheduler(job_scheduler.clone());
+        let (tx, _rx) = mpsc::channel(4);
+        let mut scheduler = env.mock_compaction_scheduler(tx);
+        let builder = VersionControlBuilder::new();
+        let version_control = Arc::new(builder.build());
+        let region_id = builder.region_id();
+        let _manifest_ctx = env
+            .mock_manifest_context(version_control.current().version.metadata.clone())
+            .await;
+        let (_schema_metadata_manager, _kv_backend) = mock_schema_metadata_manager();
+
+        scheduler.region_status.insert(
+            region_id,
+            CompactionStatus::new(region_id, version_control, env.access_layer.clone()),
+        );
+        scheduler
+            .region_status
+            .get_mut(&region_id)
+            .unwrap()
+            .start_local_task();
+
+        let (output_tx, _output_rx) = oneshot::channel();
+        scheduler.add_ddl_request_to_pending(SenderDdlRequest {
+            region_id,
+            sender: OptionOutputTx::from(output_tx),
+            request: crate::request::DdlRequest::EnterStaging(
+                store_api::region_request::EnterStagingRequest {
+                    partition_directive:
+                        store_api::region_request::StagingPartitionDirective::RejectAllWrites,
+                },
+            ),
+        });
+
+        let pending_ddls = scheduler.on_compaction_cancelled(region_id).await;
+
+        assert_eq!(pending_ddls.len(), 1);
+        assert!(!scheduler.has_pending_ddls(region_id));
+        assert!(!scheduler.region_status.contains_key(&region_id));
+        assert_eq!(job_scheduler.num_jobs(), 0);
+    }
+
+    #[tokio::test]
+    async fn test_on_compaction_cancelled_prioritizes_pending_ddls_over_pending_compaction() {
+        let job_scheduler = Arc::new(VecScheduler::default());
+        let env = SchedulerEnv::new().await.scheduler(job_scheduler.clone());
+        let (tx, _rx) = mpsc::channel(4);
+        let mut scheduler = env.mock_compaction_scheduler(tx);
+        let builder = VersionControlBuilder::new();
+        let version_control = Arc::new(builder.build());
+        let region_id = builder.region_id();
+        let _manifest_ctx = env
+            .mock_manifest_context(version_control.current().version.metadata.clone())
+            .await;
+        let (_schema_metadata_manager, _kv_backend) = mock_schema_metadata_manager();
+
+        scheduler.region_status.insert(
+            region_id,
+            CompactionStatus::new(region_id, version_control, env.access_layer.clone()),
+        );
+        let status = scheduler.region_status.get_mut(&region_id).unwrap();
+        status.start_local_task();
+        let (manual_tx, manual_rx) = oneshot::channel();
+        status.set_pending_request(PendingCompaction {
+            options: compact_request::Options::StrictWindow(StrictWindow { window_seconds: 60 }),
+            waiter: OptionOutputTx::from(manual_tx),
+            max_parallelism: 1,
+        });
+
+        let (output_tx, _output_rx) = oneshot::channel();
+        scheduler.add_ddl_request_to_pending(SenderDdlRequest {
+            region_id,
+            sender: OptionOutputTx::from(output_tx),
+            request: crate::request::DdlRequest::EnterStaging(
+                store_api::region_request::EnterStagingRequest {
+                    partition_directive:
+                        store_api::region_request::StagingPartitionDirective::RejectAllWrites,
+                },
+            ),
+        });
+
+        let pending_ddls = scheduler.on_compaction_cancelled(region_id).await;
+
+        assert_eq!(pending_ddls.len(), 1);
+        assert!(!scheduler.region_status.contains_key(&region_id));
+        assert_eq!(job_scheduler.num_jobs(), 0);
+        assert_matches!(manual_rx.await.unwrap(), Err(_));
+    }
+
     #[tokio::test]
     async fn test_pending_ddl_request_failed_on_compaction_failed() {
         let env = SchedulerEnv::new().await;
@@ -1724,6 +2065,11 @@ mod tests {
             region_id,
             CompactionStatus::new(region_id, version_control, env.access_layer.clone()),
         );
+        scheduler
+            .region_status
+            .get_mut(&region_id)
+            .unwrap()
+            .start_local_task();
 
         let (output_tx, _output_rx) = oneshot::channel();
         scheduler.add_ddl_request_to_pending(SenderDdlRequest {
@@ -1763,6 +2109,7 @@ mod tests {
         let (manual_tx, manual_rx) = oneshot::channel();
         let mut status =
             CompactionStatus::new(region_id, version_control.clone(), env.access_layer.clone());
+        status.start_local_task();
         status.set_pending_request(PendingCompaction {
             options: compact_request::Options::Regular(Default::default()),
             waiter: OptionOutputTx::from(manual_tx),
@@ -1838,6 +2185,7 @@ mod tests {
         let (manual_tx, manual_rx) = oneshot::channel();
         let mut status =
             CompactionStatus::new(region_id, version_control.clone(), env.access_layer.clone());
+        status.start_local_task();
         status.set_pending_request(PendingCompaction {
             options: compact_request::Options::Regular(Default::default()),
             waiter: OptionOutputTx::from(manual_tx),
@@ -1884,6 +2232,11 @@ mod tests {
             region_id,
             CompactionStatus::new(region_id, version_control, env.access_layer.clone()),
         );
+        scheduler
+            .region_status
+            .get_mut(&region_id)
+            .unwrap()
+            .start_local_task();
 
         let pending_ddls = scheduler
             .on_compaction_finished(region_id, &manifest_ctx, schema_metadata_manager)
@@ -1921,6 +2274,11 @@ mod tests {
             region_id,
             CompactionStatus::new(region_id, version_control, env.access_layer.clone()),
         );
+        scheduler
+            .region_status
+            .get_mut(&region_id)
+            .unwrap()
+            .start_local_task();
 
         let pending_ddls = scheduler
             .on_compaction_finished(region_id, &manifest_ctx, schema_metadata_manager)
diff --git a/src/mito2/src/compaction/compactor.rs b/src/mito2/src/compaction/compactor.rs
index 1876972b0d..e5dae0af0c 100644
--- a/src/mito2/src/compaction/compactor.rs
+++ b/src/mito2/src/compaction/compactor.rs
@@ -16,9 +16,9 @@ use std::num::NonZero;
 use std::sync::Arc;
 use std::time::Duration;
 
-use api::v1::region::compact_request;
+use common_base::cancellation::{CancellableFuture, CancellationHandle};
 use common_meta::key::SchemaMetadataManagerRef;
-use common_telemetry::{info, warn};
+use common_telemetry::{debug, info, warn};
 use common_time::TimeToLive;
 use either::Either;
 use itertools::Itertools;
@@ -34,16 +34,15 @@ use crate::access_layer::{
     AccessLayer, AccessLayerRef, Metrics, OperationType, SstWriteRequest, WriteType,
 };
 use crate::cache::{CacheManager, CacheManagerRef};
-use crate::compaction::picker::{PickerOutput, new_picker};
+use crate::compaction::picker::PickerOutput;
 use crate::compaction::{CompactionOutput, CompactionSstReaderBuilder, find_dynamic_options};
 use crate::config::MitoConfig;
+use crate::error;
 use crate::error::{
-    EmptyRegionDirSnafu, InvalidPartitionExprSnafu, JoinSnafu, ObjectStoreNotFoundSnafu, Result,
+    EmptyRegionDirSnafu, InvalidPartitionExprSnafu, ObjectStoreNotFoundSnafu, Result,
 };
 use crate::manifest::action::{RegionEdit, RegionMetaAction, RegionMetaActionList};
 use crate::manifest::manager::{RegionManifestManager, RegionManifestOptions};
-use crate::metrics;
-use crate::read::{FlatSource, Source};
 use crate::region::options::RegionOptions;
 use crate::region::version::VersionRef;
 use crate::region::{ManifestContext, RegionLeaderState, RegionRoleState};
@@ -55,6 +54,7 @@ use crate::sst::index::intermediate::IntermediateManager;
 use crate::sst::index::puffin_manager::PuffinManagerFactory;
 use crate::sst::location::region_dir_from_table_dir;
 use crate::sst::parquet::WriteOptions;
+use crate::sst::parquet::metadata::extract_primary_key_range;
 use crate::sst::version::{SstVersion, SstVersionRef};
 
 /// Region version for compaction that does not hold memtables.
@@ -290,21 +290,30 @@ pub trait Compactor: Send + Sync + 'static {
         compaction_region: &CompactionRegion,
         merge_output: MergeOutput,
     ) -> Result<RegionEdit>;
-
-    /// Execute compaction for a region.
-    async fn compact(
-        &self,
-        compaction_region: &CompactionRegion,
-        compact_request_options: compact_request::Options,
-    ) -> Result<()>;
 }
 
-/// DefaultCompactor is the default implementation of Compactor.
-pub struct DefaultCompactor;
-
-impl DefaultCompactor {
-    /// Merge a single compaction output into SST files.
+/// Trait for merging a single compaction output into SST files.
+///
+/// This is extracted from `DefaultCompactor` to allow injecting mock
+/// implementations in tests.
+#[async_trait::async_trait]
+pub trait SstMerger: Send + Sync + 'static {
     async fn merge_single_output(
+        &self,
+        compaction_region: CompactionRegion,
+        output: CompactionOutput,
+        write_opts: WriteOptions,
+    ) -> Result<Vec<FileMeta>>;
+}
+
+/// The production [`SstMerger`] that reads, merges, and writes SST files.
+#[derive(Clone)]
+pub struct DefaultSstMerger;
+
+#[async_trait::async_trait]
+impl SstMerger for DefaultSstMerger {
+    async fn merge_single_output(
+        &self,
         compaction_region: CompactionRegion,
         output: CompactionOutput,
         write_opts: WriteOptions,
@@ -322,11 +331,7 @@ impl DefaultCompactor {
             .region_options
             .sst_format
             .map(|format| format == FormatType::Flat)
-            .unwrap_or(
-                compaction_region
-                    .engine_config
-                    .default_experimental_flat_format,
-            );
+            .unwrap_or(compaction_region.engine_config.default_flat_format);
 
         let index_config = compaction_region.engine_config.index.clone();
         let inverted_index_config = compaction_region.engine_config.inverted_index.clone();
@@ -356,13 +361,7 @@ impl DefaultCompactor {
             time_range: output.output_time_range,
             merge_mode,
         };
-        let source = if flat_format {
-            let reader = builder.build_flat_sst_reader().await?;
-            Either::Right(FlatSource::Stream(reader))
-        } else {
-            let reader = builder.build_sst_reader().await?;
-            Either::Left(Source::Reader(reader))
-        };
+        let source = builder.build_flat_sst_reader().await?;
         let mut metrics = Metrics::new(WriteType::Compaction);
         let region_metadata = compaction_region.region_metadata.clone();
         let sst_infos = compaction_region
@@ -375,6 +374,11 @@ impl DefaultCompactor {
                     cache_manager: compaction_region.cache_manager.clone(),
                     storage,
                     max_sequence: max_sequence.map(NonZero::get),
+                    sst_write_format: if flat_format {
+                        FormatType::Flat
+                    } else {
+                        FormatType::PrimaryKey
+                    },
                     index_options,
                     index_config,
                     inverted_index_config,
@@ -400,22 +404,35 @@ impl DefaultCompactor {
 
         let output_files = sst_infos
             .into_iter()
-            .map(|sst_info| FileMeta {
-                region_id,
-                file_id: sst_info.file_id,
-                time_range: sst_info.time_range,
-                level: output.output_level,
-                file_size: sst_info.file_size,
-                max_row_group_uncompressed_size: sst_info.max_row_group_uncompressed_size,
-                available_indexes: sst_info.index_metadata.build_available_indexes(),
-                indexes: sst_info.index_metadata.build_indexes(),
-                index_file_size: sst_info.index_metadata.file_size,
-                index_version: 0,
-                num_rows: sst_info.num_rows as u64,
-                num_row_groups: sst_info.num_row_groups,
-                sequence: max_sequence,
-                partition_expr: partition_expr.clone(),
-                num_series: sst_info.num_series,
+            .map(|sst_info| {
+                let pk_range = sst_info
+                    .file_metadata
+                    .as_ref()
+                    .and_then(|meta| extract_primary_key_range(meta, &region_metadata));
+                let (primary_key_min, primary_key_max) = match pk_range {
+                    Some((min, max)) => (Some(min), Some(max)),
+                    None => (None, None),
+                };
+
+                FileMeta {
+                    region_id,
+                    file_id: sst_info.file_id,
+                    time_range: sst_info.time_range,
+                    level: output.output_level,
+                    file_size: sst_info.file_size,
+                    max_row_group_uncompressed_size: sst_info.max_row_group_uncompressed_size,
+                    available_indexes: sst_info.index_metadata.build_available_indexes(),
+                    indexes: sst_info.index_metadata.build_indexes(),
+                    index_file_size: sst_info.index_metadata.file_size,
+                    index_version: 0,
+                    num_rows: sst_info.num_rows as u64,
+                    num_row_groups: sst_info.num_row_groups,
+                    sequence: max_sequence,
+                    partition_expr: partition_expr.clone(),
+                    num_series: sst_info.num_series,
+                    primary_key_min,
+                    primary_key_max,
+                }
             })
             .collect::<Vec<_>>();
         let output_file_names = output_files.iter().map(|f| f.file_id.to_string()).join(",");
@@ -428,54 +445,145 @@ impl DefaultCompactor {
     }
 }
 
+/// DefaultCompactor is the default implementation of Compactor.
+///
+/// It is parameterized by an [`SstMerger`] to allow injecting mock
+/// implementations in tests.
+pub struct DefaultCompactor<M = DefaultSstMerger> {
+    merger: M,
+    cancel_handle: Arc<CancellationHandle>,
+}
+
+#[cfg(test)]
+impl<M: SstMerger> DefaultCompactor<M> {
+    pub fn with_merger(merger: M) -> Self {
+        Self {
+            merger,
+            cancel_handle: Arc::new(CancellationHandle::default()),
+        }
+    }
+}
+
+impl DefaultCompactor {
+    pub fn with_cancel_handle(cancel_handle: Arc<CancellationHandle>) -> Self {
+        Self {
+            merger: DefaultSstMerger,
+            cancel_handle,
+        }
+    }
+}
+
 #[async_trait::async_trait]
-impl Compactor for DefaultCompactor {
+impl<M: SstMerger> Compactor for DefaultCompactor<M>
+where
+    M: Clone,
+{
     async fn merge_ssts(
         &self,
         compaction_region: &CompactionRegion,
         mut picker_output: PickerOutput,
     ) -> Result<MergeOutput> {
-        let mut futs = Vec::with_capacity(picker_output.outputs.len());
-        let mut compacted_inputs =
-            Vec::with_capacity(picker_output.outputs.iter().map(|o| o.inputs.len()).sum());
         let internal_parallelism = compaction_region.max_parallelism.max(1);
         let compaction_time_window = picker_output.time_window_size;
+        let region_id = compaction_region.region_id;
+
+        // Build tasks along with their input file metas so we can track which
+        // inputs correspond to each task.
+        let mut tasks: Vec<(Vec<FileMeta>, _)> = Vec::with_capacity(picker_output.outputs.len());
 
         for output in picker_output.outputs.drain(..) {
             let inputs_to_remove: Vec<_> =
                 output.inputs.iter().map(|f| f.meta_ref().clone()).collect();
-            compacted_inputs.extend(inputs_to_remove.iter().cloned());
             let write_opts = WriteOptions {
                 write_buffer_size: compaction_region.engine_config.sst_write_buffer_size,
                 max_file_size: picker_output.max_file_size,
                 ..Default::default()
             };
-            futs.push(Self::merge_single_output(
-                compaction_region.clone(),
-                output,
-                write_opts,
-            ));
-        }
-        let mut output_files = Vec::with_capacity(futs.len());
-        while !futs.is_empty() {
-            let mut task_chunk = Vec::with_capacity(internal_parallelism);
-            for _ in 0..internal_parallelism {
-                if let Some(task) = futs.pop() {
-                    task_chunk.push(common_runtime::spawn_compact(task));
-                }
-            }
-            let metas = futures::future::try_join_all(task_chunk)
-                .await
-                .context(JoinSnafu)?
-                .into_iter()
-                .collect::<Result<Vec<Vec<_>>>>()?;
-            output_files.extend(metas.into_iter().flatten());
+            let merger = self.merger.clone();
+            let compaction_region = compaction_region.clone();
+            let fut = async move {
+                merger
+                    .merge_single_output(compaction_region, output, write_opts)
+                    .await
+            };
+            tasks.push((inputs_to_remove, fut));
         }
 
-        // In case of remote compaction, we still allow the region edit after merge to
-        // clean expired ssts.
-        let mut inputs: Vec<_> = compacted_inputs.into_iter().collect();
-        inputs.extend(
+        let mut output_files = Vec::with_capacity(tasks.len());
+        let mut compacted_inputs = Vec::with_capacity(
+            tasks.iter().map(|(inputs, _)| inputs.len()).sum::<usize>()
+                + picker_output.expired_ssts.len(),
+        );
+
+        while !tasks.is_empty() {
+            let mut chunk: Vec<(Vec<FileMeta>, _)> = Vec::with_capacity(internal_parallelism);
+            for _ in 0..internal_parallelism {
+                if let Some(task) = tasks.pop() {
+                    chunk.push(task);
+                }
+            }
+            let mut spawned: Vec<_> = chunk
+                .into_iter()
+                .map(|(inputs, fut)| {
+                    let handle = common_runtime::spawn_compact(fut);
+                    (inputs, handle)
+                })
+                .collect();
+
+            while let Some((inputs, handle)) = spawned.pop() {
+                let abort_handle = handle.abort_handle();
+                match CancellableFuture::new(handle, self.cancel_handle.clone()).await {
+                    Ok(Ok(Ok(files))) => {
+                        output_files.extend(files);
+                        compacted_inputs.extend(inputs);
+                    }
+                    Ok(Ok(Err(e))) => {
+                        warn!(
+                            e; "Failed to merge compaction output for region: {}, inputs: [{}]",
+                            region_id,
+                            inputs.iter().map(|f| f.file_id.to_string()).join(",")
+                        );
+                    }
+                    Ok(Err(e)) => {
+                        warn!(
+                            "Region {} compaction task join error for inputs: [{}], skipping: {}",
+                            region_id,
+                            inputs.iter().map(|f| f.file_id.to_string()).join(","),
+                            e
+                        );
+                        // If the cancel handle is cancelled,
+                        // cancel the remaining tasks before returns the error.
+                        if self.cancel_handle.is_cancelled() {
+                            abort_handle.abort();
+                            for (_, handle) in spawned {
+                                handle.abort();
+                            }
+                        }
+                        return Err(e).context(error::JoinSnafu);
+                    }
+                    Err(_) => {
+                        debug!(
+                            "Compaction merge cancelled for region: {}, aborting remaining {} spawned tasks",
+                            region_id,
+                            spawned.len(),
+                        );
+                        abort_handle.abort();
+                        for (_, handle) in spawned {
+                            handle.abort();
+                        }
+                        break;
+                    }
+                }
+            }
+
+            if self.cancel_handle.is_cancelled() {
+                info!("Compaction merge cancelled for region: {}", region_id);
+                break;
+            }
+        }
+
+        // Include expired SSTs in removals — these don't depend on merge success.
+        compacted_inputs.extend(
             picker_output
                 .expired_ssts
                 .iter()
@@ -484,7 +592,7 @@ impl Compactor for DefaultCompactor {
 
         Ok(MergeOutput {
             files_to_add: output_files,
-            files_to_remove: inputs,
+            files_to_remove: compacted_inputs,
             compaction_time_window: Some(compaction_time_window),
         })
     }
@@ -517,48 +625,329 @@ impl Compactor for DefaultCompactor {
 
         Ok(edit)
     }
+}
 
-    // The default implementation of compact combines the merge_ssts and update_manifest functions.
-    // Note: It's local compaction and only used for testing purpose.
-    async fn compact(
-        &self,
-        compaction_region: &CompactionRegion,
-        compact_request_options: compact_request::Options,
-    ) -> Result<()> {
-        let picker_output = {
-            let picker_output = new_picker(
-                &compact_request_options,
-                &compaction_region.region_options.compaction,
-                compaction_region.region_options.append_mode,
-                None,
-            )
-            .pick(compaction_region);
+#[cfg(test)]
+mod tests {
+    use std::sync::atomic::{AtomicUsize, Ordering};
+    use std::sync::{Arc, Mutex};
+    use std::time::Duration;
 
-            if let Some(picker_output) = picker_output {
-                picker_output
-            } else {
-                info!(
-                    "No files to compact for region_id: {}",
-                    compaction_region.region_id
-                );
-                return Ok(());
+    use store_api::storage::{FileId, RegionId};
+    use tokio::time::sleep;
+
+    use super::{DefaultCompactor, *};
+    use crate::cache::CacheManager;
+    use crate::compaction::picker::PickerOutput;
+    use crate::error::Result;
+    use crate::sst::file::FileHandle;
+    use crate::sst::file_purger::NoopFilePurger;
+    use crate::sst::version::SstVersion;
+    use crate::test_util::memtable_util::metadata_for_test;
+    use crate::test_util::scheduler_util::SchedulerEnv;
+
+    fn dummy_file_meta() -> FileMeta {
+        FileMeta {
+            region_id: RegionId::new(1, 1),
+            file_id: FileId::random(),
+            file_size: 100,
+            ..Default::default()
+        }
+    }
+
+    fn new_file_handle(meta: FileMeta) -> FileHandle {
+        FileHandle::new(meta, Arc::new(NoopFilePurger))
+    }
+
+    /// Build a minimal [`CompactionRegion`] suitable for tests where the
+    /// [`SstMerger`] is mocked and never touches the access layer.
+    async fn new_test_compaction_region() -> CompactionRegion {
+        let env = SchedulerEnv::new().await;
+        let metadata = metadata_for_test();
+        let manifest_ctx = env.mock_manifest_context(metadata.clone()).await;
+        CompactionRegion {
+            region_id: RegionId::new(1, 1),
+            region_options: RegionOptions::default(),
+            engine_config: Arc::new(MitoConfig::default()),
+            region_metadata: metadata.clone(),
+            cache_manager: Arc::new(CacheManager::default()),
+            access_layer: env.access_layer.clone(),
+            manifest_ctx,
+            current_version: CompactionVersion {
+                metadata,
+                options: RegionOptions::default(),
+                ssts: Arc::new(SstVersion::new()),
+                compaction_time_window: None,
+            },
+            file_purger: None,
+            ttl: None,
+            max_parallelism: 1,
+        }
+    }
+
+    /// An [`SstMerger`] that returns pre-configured results per call index.
+    ///
+    /// Call 0 gets `results[0]`, call 1 gets `results[1]`, etc.
+    #[derive(Clone)]
+    struct MockMerger {
+        results: Arc<Mutex<Vec<Result<Vec<FileMeta>>>>>,
+        call_idx: Arc<AtomicUsize>,
+    }
+
+    impl MockMerger {
+        fn new(results: Vec<Result<Vec<FileMeta>>>) -> Self {
+            Self {
+                results: Arc::new(Mutex::new(results)),
+                call_idx: Arc::new(AtomicUsize::new(0)),
             }
+        }
+    }
+
+    #[async_trait::async_trait]
+    impl SstMerger for MockMerger {
+        async fn merge_single_output(
+            &self,
+            _compaction_region: CompactionRegion,
+            _output: CompactionOutput,
+            _write_opts: WriteOptions,
+        ) -> Result<Vec<FileMeta>> {
+            let idx = self.call_idx.fetch_add(1, Ordering::SeqCst);
+            match self.results.lock().unwrap().get(idx) {
+                Some(Ok(files)) => Ok(files.clone()),
+                Some(Err(_)) => error::InvalidMetaSnafu {
+                    reason: format!("simulated failure at index {idx}"),
+                }
+                .fail(),
+                None => panic!("MockMerger: no result configured for call index {idx}"),
+            }
+        }
+    }
+
+    #[tokio::test]
+    async fn test_partial_merge_failure_collects_only_successful_outputs() {
+        common_telemetry::init_default_ut_logging();
+
+        let compaction_region = new_test_compaction_region().await;
+
+        // Prepare 3 compaction outputs: output 0 and 2 succeed, output 1 fails.
+        let input_meta_0 = dummy_file_meta();
+        let input_meta_1 = dummy_file_meta();
+        let input_meta_2 = dummy_file_meta();
+
+        let output_meta_0 = vec![dummy_file_meta()];
+        let output_meta_2 = vec![dummy_file_meta(), dummy_file_meta()];
+
+        let merger = MockMerger::new(vec![
+            Ok(output_meta_0.clone()),
+            Err(error::InvalidMetaSnafu {
+                reason: "boom".to_string(),
+            }
+            .build()),
+            Ok(output_meta_2.clone()),
+        ]);
+        let compactor = DefaultCompactor::with_merger(merger);
+
+        let picker_output = PickerOutput {
+            outputs: vec![
+                CompactionOutput {
+                    output_level: 1,
+                    inputs: vec![new_file_handle(input_meta_0.clone())],
+                    filter_deleted: false,
+                    output_time_range: None,
+                },
+                CompactionOutput {
+                    output_level: 1,
+                    inputs: vec![new_file_handle(input_meta_1.clone())],
+                    filter_deleted: false,
+                    output_time_range: None,
+                },
+                CompactionOutput {
+                    output_level: 1,
+                    inputs: vec![new_file_handle(input_meta_2.clone())],
+                    filter_deleted: false,
+                    output_time_range: None,
+                },
+            ],
+            expired_ssts: vec![],
+            time_window_size: 3600,
+            max_file_size: None,
         };
 
-        let merge_output = self.merge_ssts(compaction_region, picker_output).await?;
-        if merge_output.is_empty() {
-            info!(
-                "No files to compact for region_id: {}",
-                compaction_region.region_id
-            );
-            return Ok(());
+        let merge_output = compactor
+            .merge_ssts(&compaction_region, picker_output)
+            .await
+            .unwrap();
+
+        // Outputs 0 and 2 succeeded (1 + 2 = 3 files added).
+        assert_eq!(merge_output.files_to_add.len(), 3);
+        // Only inputs from successful merges should be removed.
+        assert_eq!(merge_output.files_to_remove.len(), 2);
+
+        let removed_ids: Vec<_> = merge_output
+            .files_to_remove
+            .iter()
+            .map(|f| f.file_id)
+            .collect();
+        assert!(removed_ids.contains(&input_meta_0.file_id));
+        assert!(removed_ids.contains(&input_meta_2.file_id));
+        // The failed output's input must NOT be removed.
+        assert!(!removed_ids.contains(&input_meta_1.file_id));
+    }
+
+    #[tokio::test]
+    async fn test_all_outputs_succeed() {
+        common_telemetry::init_default_ut_logging();
+
+        let compaction_region = new_test_compaction_region().await;
+        let input_meta = dummy_file_meta();
+        let output_meta = vec![dummy_file_meta()];
+
+        let merger = MockMerger::new(vec![Ok(output_meta.clone())]);
+        let compactor = DefaultCompactor::with_merger(merger);
+
+        let picker_output = PickerOutput {
+            outputs: vec![CompactionOutput {
+                output_level: 1,
+                inputs: vec![new_file_handle(input_meta.clone())],
+                filter_deleted: false,
+                output_time_range: None,
+            }],
+            expired_ssts: vec![],
+            time_window_size: 3600,
+            max_file_size: None,
+        };
+
+        let merge_output = compactor
+            .merge_ssts(&compaction_region, picker_output)
+            .await
+            .unwrap();
+
+        assert_eq!(merge_output.files_to_add.len(), 1);
+        assert_eq!(merge_output.files_to_add[0].file_id, output_meta[0].file_id);
+        assert_eq!(merge_output.files_to_remove.len(), 1);
+        assert_eq!(merge_output.files_to_remove[0].file_id, input_meta.file_id);
+    }
+
+    #[tokio::test]
+    async fn test_expired_ssts_always_removed() {
+        common_telemetry::init_default_ut_logging();
+
+        let compaction_region = new_test_compaction_region().await;
+        let input_meta = dummy_file_meta();
+        let expired_meta = dummy_file_meta();
+
+        // The single merge output fails, but expired SSTs should still be removed.
+        let merger = MockMerger::new(vec![Err(error::InvalidMetaSnafu {
+            reason: "fail".to_string(),
         }
+        .build())]);
+        let compactor = DefaultCompactor::with_merger(merger);
 
-        metrics::COMPACTION_INPUT_BYTES.inc_by(merge_output.input_file_size() as f64);
-        metrics::COMPACTION_OUTPUT_BYTES.inc_by(merge_output.output_file_size() as f64);
-        self.update_manifest(compaction_region, merge_output)
-            .await?;
+        let picker_output = PickerOutput {
+            outputs: vec![CompactionOutput {
+                output_level: 1,
+                inputs: vec![new_file_handle(input_meta.clone())],
+                filter_deleted: false,
+                output_time_range: None,
+            }],
+            expired_ssts: vec![new_file_handle(expired_meta.clone())],
+            time_window_size: 3600,
+            max_file_size: None,
+        };
 
-        Ok(())
+        let merge_output = compactor
+            .merge_ssts(&compaction_region, picker_output)
+            .await
+            .unwrap();
+
+        // No files added (merge failed).
+        assert!(merge_output.files_to_add.is_empty());
+        // Only the expired SST should be in files_to_remove (not the failed merge's input).
+        assert_eq!(merge_output.files_to_remove.len(), 1);
+        assert_eq!(
+            merge_output.files_to_remove[0].file_id,
+            expired_meta.file_id
+        );
+    }
+
+    #[derive(Clone)]
+    struct BlockingMerger {
+        call_idx: Arc<AtomicUsize>,
+    }
+
+    #[async_trait::async_trait]
+    impl SstMerger for BlockingMerger {
+        async fn merge_single_output(
+            &self,
+            _compaction_region: CompactionRegion,
+            _output: CompactionOutput,
+            _write_opts: WriteOptions,
+        ) -> Result<Vec<FileMeta>> {
+            self.call_idx.fetch_add(1, Ordering::SeqCst);
+            std::future::pending().await
+        }
+    }
+
+    #[tokio::test(flavor = "multi_thread")]
+    async fn test_merge_ssts_cancels_spawned_tasks() {
+        common_telemetry::init_default_ut_logging();
+
+        let mut compaction_region = new_test_compaction_region().await;
+        compaction_region.max_parallelism = 2;
+
+        let cancel_handle = Arc::new(CancellationHandle::default());
+        let call_idx = Arc::new(AtomicUsize::new(0));
+        let compactor = DefaultCompactor {
+            merger: BlockingMerger {
+                call_idx: call_idx.clone(),
+            },
+            cancel_handle: cancel_handle.clone(),
+        };
+
+        let picker_output = PickerOutput {
+            outputs: vec![
+                CompactionOutput {
+                    output_level: 1,
+                    inputs: vec![new_file_handle(dummy_file_meta())],
+                    filter_deleted: false,
+                    output_time_range: None,
+                },
+                CompactionOutput {
+                    output_level: 1,
+                    inputs: vec![new_file_handle(dummy_file_meta())],
+                    filter_deleted: false,
+                    output_time_range: None,
+                },
+                CompactionOutput {
+                    output_level: 1,
+                    inputs: vec![new_file_handle(dummy_file_meta())],
+                    filter_deleted: false,
+                    output_time_range: None,
+                },
+            ],
+            expired_ssts: vec![],
+            time_window_size: 3600,
+            max_file_size: None,
+        };
+
+        let task = tokio::spawn(async move {
+            compactor
+                .merge_ssts(&compaction_region, picker_output)
+                .await
+        });
+
+        sleep(Duration::from_millis(100)).await;
+        cancel_handle.cancel();
+
+        let merge_output = task
+            .await
+            .expect("merge_ssts should stop after cancellation")
+            .unwrap();
+
+        let started = call_idx.load(Ordering::SeqCst);
+
+        assert!(merge_output.files_to_add.is_empty());
+        assert!(merge_output.files_to_remove.is_empty());
+        assert_eq!(started, 2);
     }
 }
diff --git a/src/mito2/src/compaction/memory_manager.rs b/src/mito2/src/compaction/memory_manager.rs
index 8cbb5d293a..94b0779254 100644
--- a/src/mito2/src/compaction/memory_manager.rs
+++ b/src/mito2/src/compaction/memory_manager.rs
@@ -31,7 +31,7 @@ impl MemoryMetrics for CompactionMemoryMetrics {
         COMPACTION_MEMORY_IN_USE.set(bytes);
     }
 
-    fn inc_rejected(&self, reason: &str) {
+    fn inc_exhausted(&self, reason: &str) {
         COMPACTION_MEMORY_REJECTED
             .with_label_values(&[reason])
             .inc();
diff --git a/src/mito2/src/compaction/run.rs b/src/mito2/src/compaction/run.rs
index a7e5ca490c..27d4f5779e 100644
--- a/src/mito2/src/compaction/run.rs
+++ b/src/mito2/src/compaction/run.rs
@@ -15,6 +15,7 @@
 //! This file contains code to find sorted runs in a set if ranged items and
 //! along with the best way to merge these items to satisfy the desired run count.
 
+use bytes::{Buf, Bytes};
 use common_base::BitVec;
 use common_base::readable_size::ReadableSize;
 use common_time::Timestamp;
@@ -32,15 +33,37 @@ pub trait Ranged {
     /// Returns the inclusive range of item.
     fn range(&self) -> (Self::BoundType, Self::BoundType);
 
-    fn overlap<T>(&self, other: &T) -> bool
-    where
-        T: Ranged<BoundType = Self::BoundType>,
-    {
+    fn overlap(&self, other: &Self) -> bool {
         let (lhs_start, lhs_end) = self.range();
         let (rhs_start, rhs_end) = other.range();
 
         lhs_start.max(rhs_start) < lhs_end.min(rhs_end)
     }
+
+    /// Like `overlap`, but treats touching boundaries as overlapping (inclusive).
+    /// Used by `find_overlapping_items` where shared boundaries count as overlap.
+    fn overlap_inclusive(&self, other: &Self) -> bool {
+        let (lhs_start, lhs_end) = self.range();
+        let (rhs_start, rhs_end) = other.range();
+
+        lhs_start.max(rhs_start) <= lhs_end.min(rhs_end)
+    }
+}
+
+pub(crate) fn primary_key_ranges_overlap(lhs: &(Bytes, Bytes), rhs: &(Bytes, Bytes)) -> bool {
+    lhs.0.chunk().max(rhs.0.chunk()) <= lhs.1.chunk().min(rhs.1.chunk())
+}
+
+pub(crate) fn merge_primary_key_ranges(
+    lhs: Option<(Bytes, Bytes)>,
+    rhs: Option<(Bytes, Bytes)>,
+) -> Option<(Bytes, Bytes)> {
+    match (lhs, rhs) {
+        (Some((lhs_min, lhs_max)), Some((rhs_min, rhs_max))) => {
+            Some((lhs_min.min(rhs_min), lhs_max.max(rhs_max)))
+        }
+        _ => None,
+    }
 }
 
 pub fn find_overlapping_items<T: Item + Clone>(
@@ -85,15 +108,15 @@ pub fn find_overlapping_items<T: Item + Clone>(
         // Check for overlaps with remaining right elements
         let mut j = r_idx;
         while j < r.items.len() {
-            let (rhs_start, rhs_end) = r.items[j].range();
+            let (rhs_start, _rhs_end) = r.items[j].range();
 
             // If right element starts after left element ends, no more overlaps possible
             if rhs_start > lhs_end {
                 break;
             }
 
-            // We have an overlap
-            if lhs_start.max(rhs_start) <= lhs_end.min(rhs_end) {
+            // We have an overlap (inclusive: touching boundaries count)
+            if lhs.overlap_inclusive(&r.items[j]) {
                 if !selected[lhs_idx] {
                     result.push(lhs.clone());
                     selected.set(lhs_idx, true);
@@ -134,6 +157,7 @@ pub struct FileGroup {
     num_rows: usize,
     min_timestamp: Timestamp,
     max_timestamp: Timestamp,
+    primary_key_range: Option<(Bytes, Bytes)>,
 }
 
 impl FileGroup {
@@ -141,12 +165,14 @@ impl FileGroup {
         let size = file.size() as usize;
         let (min_timestamp, max_timestamp) = file.time_range();
         let num_rows = file.num_rows();
+        let primary_key_range = file.primary_key_range();
         Self {
             files: smallvec![file],
             size,
             num_rows,
             min_timestamp,
             max_timestamp,
+            primary_key_range,
         }
     }
 
@@ -160,6 +186,8 @@ impl FileGroup {
         let (min_timestamp, max_timestamp) = file.time_range();
         self.min_timestamp = self.min_timestamp.min(min_timestamp);
         self.max_timestamp = self.max_timestamp.max(max_timestamp);
+        self.primary_key_range =
+            merge_primary_key_ranges(self.primary_key_range.take(), file.primary_key_range());
         self.files.push(file);
     }
 
@@ -187,6 +215,32 @@ impl Ranged for FileGroup {
     fn range(&self) -> (Self::BoundType, Self::BoundType) {
         (self.min_timestamp, self.max_timestamp)
     }
+
+    fn overlap(&self, other: &Self) -> bool {
+        let (lhs_start, lhs_end) = self.range();
+        let (rhs_start, rhs_end) = other.range();
+        if lhs_start.max(rhs_start) >= lhs_end.min(rhs_end) {
+            return false;
+        }
+
+        match (&self.primary_key_range, &other.primary_key_range) {
+            (Some(lhs), Some(rhs)) => primary_key_ranges_overlap(lhs, rhs),
+            _ => true,
+        }
+    }
+
+    fn overlap_inclusive(&self, other: &Self) -> bool {
+        let (lhs_start, lhs_end) = self.range();
+        let (rhs_start, rhs_end) = other.range();
+        if lhs_start.max(rhs_start) > lhs_end.min(rhs_end) {
+            return false;
+        }
+
+        match (&self.primary_key_range, &other.primary_key_range) {
+            (Some(lhs), Some(rhs)) => primary_key_ranges_overlap(lhs, rhs),
+            _ => true,
+        }
+    }
 }
 
 impl Item for FileGroup {
@@ -281,21 +335,20 @@ where
                 // item is already assigned.
                 continue;
             }
-            match current_run.items.last() {
-                None => {
-                    // current run is empty, just add current_item
+            if current_run.items.is_empty() {
+                // current run is empty, just add current_item
+                selected.set(true);
+                current_run.push_item(item.clone());
+            } else {
+                // the current item does not overlap with any item in current run,
+                // then it belongs to current run. Because now we introduced primary
+                // key range, we cannot simply use timestamps to check overlapping.
+                let overlaps_any = current_run.items.iter().any(|i| i.overlap(item));
+                if !overlaps_any {
+                    // does not overlap, push to current run
                     selected.set(true);
                     current_run.push_item(item.clone());
                 }
-                Some(last) => {
-                    // the current item does not overlap with the last item in current run,
-                    // then it belongs to current run.
-                    if !last.overlap(item) {
-                        // does not overlap, push to current run
-                        selected.set(true);
-                        current_run.push_item(item.clone());
-                    }
-                }
             }
         }
         // finished an iteration, we've found a new run.
@@ -309,7 +362,7 @@ where
 pub fn reduce_runs<T: Item>(mut runs: Vec<SortedRun<T>>) -> Vec<T> {
     assert!(runs.len() > 1);
     // sort runs by size
-    runs.sort_unstable_by(|a, b| a.size.cmp(&b.size));
+    runs.sort_unstable_by_key(|a| a.size);
     // limit max probe runs to 100
     let probe_end = runs.len().min(100);
     let mut min_penalty = usize::MAX;
@@ -422,7 +475,11 @@ pub fn merge_seq_files<T: Item>(input_files: &[T], max_file_size: Option<u64>) -
 mod tests {
     use std::collections::HashSet;
 
+    use bytes::Bytes;
+    use store_api::storage::FileId;
+
     use super::*;
+    use crate::compaction::test_util::new_file_handle_with_size_sequence_and_primary_key_range;
 
     #[derive(Clone, Debug, PartialEq)]
     struct MockFile {
@@ -467,6 +524,10 @@ mod tests {
             .collect()
     }
 
+    fn pk_range(min: &'static [u8], max: &'static [u8]) -> Option<(Bytes, Bytes)> {
+        Some((Bytes::from_static(min), Bytes::from_static(max)))
+    }
+
     fn check_sorted_runs(
         ranges: &[(i64, i64)],
         expected_runs: &[Vec<(i64, i64)>],
@@ -726,6 +787,157 @@ mod tests {
         assert_eq!(result.len(), 4); // Should find both overlaps
     }
 
+    #[test]
+    fn test_file_group_overlap_time_overlap_pk_disjoint() {
+        let lhs =
+            FileGroup::new_with_file(new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                0,
+                100,
+                0,
+                1,
+                10,
+                pk_range(b"a", b"f"),
+            ));
+        let rhs =
+            FileGroup::new_with_file(new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                50,
+                150,
+                0,
+                2,
+                10,
+                pk_range(b"x", b"z"),
+            ));
+
+        assert!(!lhs.overlap(&rhs));
+    }
+
+    #[test]
+    fn test_find_sorted_runs_collapses_pk_disjoint_files_into_one_run() {
+        let mut files = vec![
+            FileGroup::new_with_file(new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                0,
+                100,
+                0,
+                1,
+                10,
+                pk_range(b"a", b"f"),
+            )),
+            FileGroup::new_with_file(new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                50,
+                150,
+                0,
+                2,
+                10,
+                pk_range(b"x", b"z"),
+            )),
+        ];
+
+        let runs = find_sorted_runs(&mut files);
+
+        assert_eq!(1, runs.len());
+        assert_eq!(2, runs[0].items().len());
+    }
+
+    #[test]
+    fn test_find_sorted_runs_handles_2d_transitivity_break() {
+        let mut files = vec![
+            FileGroup::new_with_file(new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                0,
+                100,
+                0,
+                1,
+                10,
+                pk_range(b"a", b"f"),
+            )),
+            FileGroup::new_with_file(new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                50,
+                150,
+                0,
+                2,
+                10,
+                pk_range(b"x", b"z"),
+            )),
+            FileGroup::new_with_file(new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                50,
+                150,
+                0,
+                3,
+                10,
+                pk_range(b"a", b"f"),
+            )),
+        ];
+
+        let runs = find_sorted_runs(&mut files);
+
+        assert_eq!(2, runs.len());
+        assert_eq!(2, runs[0].items().len());
+        assert_eq!(1, runs[1].items().len());
+    }
+
+    #[test]
+    fn test_find_overlapping_items_skips_pk_disjoint_pairs() {
+        let mut left = SortedRun::from(vec![FileGroup::new_with_file(
+            new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                0,
+                100,
+                0,
+                1,
+                10,
+                pk_range(b"a", b"f"),
+            ),
+        )]);
+        let mut right = SortedRun::from(vec![FileGroup::new_with_file(
+            new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                50,
+                150,
+                0,
+                2,
+                10,
+                pk_range(b"x", b"z"),
+            ),
+        )]);
+        let mut result = Vec::new();
+
+        find_overlapping_items(&mut left, &mut right, &mut result);
+
+        assert!(result.is_empty());
+    }
+
+    #[test]
+    fn test_file_group_touching_time_boundary_with_same_pk_is_not_overlap() {
+        let lhs =
+            FileGroup::new_with_file(new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                0,
+                100,
+                0,
+                1,
+                10,
+                pk_range(b"a", b"f"),
+            ));
+        let rhs =
+            FileGroup::new_with_file(new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                100,
+                150,
+                0,
+                2,
+                10,
+                pk_range(b"a", b"f"),
+            ));
+
+        assert!(!lhs.overlap(&rhs));
+    }
+
     #[test]
     fn test_merge_seq_files() {
         // Test empty input
diff --git a/src/mito2/src/compaction/task.rs b/src/mito2/src/compaction/task.rs
index e21db2bab8..a8182353aa 100644
--- a/src/mito2/src/compaction/task.rs
+++ b/src/mito2/src/compaction/task.rs
@@ -16,13 +16,15 @@ use std::fmt::{Debug, Formatter};
 use std::sync::Arc;
 use std::time::Instant;
 
+use common_base::cancellation::CancellableFuture;
 use common_memory_manager::OnExhaustedPolicy;
 use common_telemetry::{error, info, warn};
 use itertools::Itertools;
 use snafu::ResultExt;
 use tokio::sync::mpsc;
 
-use crate::compaction::compactor::{CompactionRegion, Compactor};
+use crate::compaction::LocalCompactionState;
+use crate::compaction::compactor::{CompactionRegion, Compactor, MergeOutput};
 use crate::compaction::memory_manager::{CompactionMemoryGuard, CompactionMemoryManager};
 use crate::compaction::picker::{CompactionTask, PickerOutput};
 use crate::error::{CompactRegionSnafu, CompactionMemoryExhaustedSnafu};
@@ -30,8 +32,8 @@ use crate::manifest::action::{RegionEdit, RegionMetaAction, RegionMetaActionList
 use crate::metrics::{COMPACTION_FAILURE_COUNT, COMPACTION_MEMORY_WAIT, COMPACTION_STAGE_ELAPSED};
 use crate::region::RegionRoleState;
 use crate::request::{
-    BackgroundNotify, CompactionFailed, CompactionFinished, OutputTx, RegionEditResult,
-    WorkerRequest, WorkerRequestWithTime,
+    BackgroundNotify, CompactionCancelled, CompactionFailed, CompactionFinished, OutputTx,
+    RegionEditResult, WorkerRequest, WorkerRequestWithTime,
 };
 use crate::sst::file::FileMeta;
 use crate::worker::WorkerListener;
@@ -41,6 +43,8 @@ use crate::{error, metrics};
 pub const MAX_PARALLEL_COMPACTION: usize = 1;
 
 pub(crate) struct CompactionTaskImpl {
+    /// Shared local-compaction state for cooperative cancellation.
+    pub(crate) state: LocalCompactionState,
     pub compaction_region: CompactionRegion,
     /// Request sender to notify the worker.
     pub(crate) request_sender: mpsc::Sender<WorkerRequestWithTime>,
@@ -184,9 +188,7 @@ impl CompactionTaskImpl {
         );
     }
 
-    async fn handle_expiration_and_compaction(&mut self) -> error::Result<RegionEdit> {
-        self.mark_files_compacting(true);
-
+    async fn handle_expiration(&mut self) {
         // 1. In case of local compaction, we can delete expired ssts in advance.
         if !self.picker_output.expired_ssts.is_empty() {
             let remove_timer = COMPACTION_STAGE_ELAPSED
@@ -203,7 +205,9 @@ impl CompactionTaskImpl {
                 .await;
             remove_timer.observe_duration();
         }
+    }
 
+    async fn handle_compaction(&mut self) -> error::Result<MergeOutput> {
         // 2. Merge inputs
         let merge_timer = COMPACTION_STAGE_ELAPSED
             .with_label_values(&["merge"])
@@ -239,6 +243,13 @@ impl CompactionTaskImpl {
             .on_merge_ssts_finished(self.compaction_region.region_id)
             .await;
 
+        Ok(compaction_result)
+    }
+
+    async fn update_manifest(
+        &self,
+        compaction_result: crate::compaction::compactor::MergeOutput,
+    ) -> error::Result<RegionEdit> {
         let _manifest_timer = COMPACTION_STAGE_ELAPSED
             .with_label_values(&["write_manifest"])
             .start_timer();
@@ -296,14 +307,61 @@ impl CompactionTask for CompactionTaskImpl {
             }
         };
 
-        let notify = match self.handle_expiration_and_compaction().await {
-            Ok(edit) => BackgroundNotify::CompactionFinished(CompactionFinished {
-                region_id: self.compaction_region.region_id,
-                senders: std::mem::take(&mut self.waiters),
-                start_time: self.start_time,
-                edit,
-            }),
-            Err(e) => {
+        // Marks files compacting before compaction and unmark after compaction (even if compaction is cancelled or failed), so that they won't be picked by other compaction tasks.
+        self.mark_files_compacting(true);
+        self.handle_expiration().await;
+
+        let cancel_handle = self.state.cancel_handle();
+        // Run compaction with cooperative cancellation.
+        let notify = match CancellableFuture::new(
+            async { self.handle_compaction().await },
+            cancel_handle,
+        )
+        .await
+        {
+            Ok(Ok(merge_output)) => {
+                // Stop accepting cancellation once we are about to publish the compaction edit.
+                if !self.state.mark_commit_started() {
+                    let senders = std::mem::take(&mut self.waiters);
+                    BackgroundNotify::CompactionCancelled(CompactionCancelled {
+                        region_id: self.compaction_region.region_id,
+                        senders,
+                    })
+                } else {
+                    match self.update_manifest(merge_output).await {
+                        Ok(edit) => {
+                            let senders = std::mem::take(&mut self.waiters);
+                            BackgroundNotify::CompactionFinished(CompactionFinished {
+                                region_id: self.compaction_region.region_id,
+                                senders,
+                                start_time: self.start_time,
+                                edit,
+                            })
+                        }
+                        Err(e) => {
+                            error!(e; "Failed to compact region, region id: {}", self.compaction_region.region_id);
+                            let err = Arc::new(e);
+                            self.on_failure(err.clone());
+                            BackgroundNotify::CompactionFailed(CompactionFailed {
+                                region_id: self.compaction_region.region_id,
+                                err,
+                            })
+                        }
+                    }
+                }
+            }
+            Err(_) => {
+                info!(
+                    "Compaction cancelled, region id: {}",
+                    self.compaction_region.region_id
+                );
+                let senders = std::mem::take(&mut self.waiters);
+                BackgroundNotify::CompactionCancelled(CompactionCancelled {
+                    region_id: self.compaction_region.region_id,
+                    senders,
+                })
+            }
+            Ok(Err(e)) => {
                 error!(e; "Failed to compact region, region id: {}", self.compaction_region.region_id);
                 let err = Arc::new(e);
                 // notify compaction waiters
@@ -334,7 +392,7 @@ mod tests {
     fn test_picker_output_with_expired_ssts() {
         // Test that PickerOutput correctly includes expired_ssts
         // This verifies that expired SSTs are properly identified and included
-        // in the picker output, which is then handled by handle_expiration_and_compaction
+        // in the picker output, which is then handled by handle_expiration()
 
         let file_ids = (0..3).map(|_| FileId::random()).collect::<Vec<_>>();
         let expired_ssts = vec![
@@ -382,6 +440,6 @@ mod tests {
     //
     // The behavior is tested indirectly through integration tests:
     // - remove_expired() logs errors but doesn't stop compaction
-    // - handle_expiration_and_compaction() continues even if remove_expired() encounters errors
-    // - The function is designed to be non-blocking for compaction
+    // - handle_expiration() continues even if remove_expired() encounters errors
+    // - The expiration stage is designed to be non-blocking for compaction
 }
diff --git a/src/mito2/src/compaction/test_util.rs b/src/mito2/src/compaction/test_util.rs
index 6061e294bd..f6b39fdc8c 100644
--- a/src/mito2/src/compaction/test_util.rs
+++ b/src/mito2/src/compaction/test_util.rs
@@ -14,6 +14,7 @@
 
 use std::num::NonZeroU64;
 
+use bytes::Bytes;
 use common_time::Timestamp;
 use store_api::storage::FileId;
 
@@ -84,7 +85,46 @@ pub fn new_file_handle_with_size_and_sequence(
             num_series: 0,
             sequence: NonZeroU64::new(sequence),
             partition_expr: None,
+            ..Default::default()
         },
         file_purger,
     )
 }
+
+/// Test util to create file handles with custom size and primary-key range.
+pub fn new_file_handle_with_size_sequence_and_primary_key_range(
+    file_id: FileId,
+    start_ts_millis: i64,
+    end_ts_millis: i64,
+    level: Level,
+    sequence: u64,
+    file_size: u64,
+    primary_key_range: Option<(Bytes, Bytes)>,
+) -> FileHandle {
+    let file_purger = new_noop_file_purger();
+    FileHandle::new_with_primary_key_range(
+        FileMeta {
+            region_id: 0.into(),
+            file_id,
+            time_range: (
+                Timestamp::new_millisecond(start_ts_millis),
+                Timestamp::new_millisecond(end_ts_millis),
+            ),
+            level,
+            file_size,
+            max_row_group_uncompressed_size: file_size,
+            available_indexes: Default::default(),
+            indexes: Default::default(),
+            index_file_size: 0,
+            index_version: 0,
+            num_rows: 0,
+            num_row_groups: 0,
+            num_series: 0,
+            sequence: NonZeroU64::new(sequence),
+            partition_expr: None,
+            ..Default::default()
+        },
+        file_purger,
+        primary_key_range,
+    )
+}
diff --git a/src/mito2/src/compaction/twcs.rs b/src/mito2/src/compaction/twcs.rs
index 9012457f75..87dcac7279 100644
--- a/src/mito2/src/compaction/twcs.rs
+++ b/src/mito2/src/compaction/twcs.rs
@@ -28,7 +28,8 @@ use crate::compaction::buckets::infer_time_bucket;
 use crate::compaction::compactor::CompactionRegion;
 use crate::compaction::picker::{Picker, PickerOutput};
 use crate::compaction::run::{
-    FileGroup, Item, Ranged, find_sorted_runs, merge_seq_files, reduce_runs,
+    FileGroup, Item, Ranged, find_sorted_runs, merge_primary_key_ranges, merge_seq_files,
+    primary_key_ranges_overlap, reduce_runs,
 };
 use crate::compaction::{CompactionOutput, get_expired_ssts};
 use crate::sst::file::{FileHandle, Level, overlaps};
@@ -268,12 +269,14 @@ struct Window {
     files: HashMap<Option<NonZeroU64>, FileGroup>,
     time_window: i64,
     overlapping: bool,
+    primary_key_range: Option<(bytes::Bytes, bytes::Bytes)>,
 }
 
 impl Window {
     /// Creates a new [Window] with given file.
     fn new_with_file(file: FileHandle) -> Self {
         let (start, end) = file.time_range();
+        let primary_key_range = file.primary_key_range();
         let files = HashMap::from([(file.meta_ref().sequence, FileGroup::new_with_file(file))]);
         Self {
             start,
@@ -281,6 +284,7 @@ impl Window {
             files,
             time_window: 0,
             overlapping: false,
+            primary_key_range,
         }
     }
 
@@ -294,6 +298,8 @@ impl Window {
         let (start, end) = file.time_range();
         self.start = self.start.min(start);
         self.end = self.end.max(end);
+        self.primary_key_range =
+            merge_primary_key_ranges(self.primary_key_range.take(), file.primary_key_range());
 
         match self.files.entry(file.meta_ref().sequence) {
             Entry::Occupied(mut o) => {
@@ -347,18 +353,27 @@ fn assign_to_windows<'a>(
     let mut windows = windows.into_values().collect::<Vec<_>>();
     windows.sort_unstable_by(|l, r| l.start.cmp(&r.start).then(l.end.cmp(&r.end).reverse()));
 
-    let mut current_range: (Timestamp, Timestamp) = windows[0].range(); // windows cannot be empty.
+    for idx in 0..windows.len() {
+        let lhs_range = windows[idx].range();
+        for next_idx in idx + 1..windows.len() {
+            let rhs_range = windows[next_idx].range();
+            if rhs_range.0 > lhs_range.1 {
+                break;
+            }
 
-    for idx in 1..windows.len() {
-        let next_range = windows[idx].range();
-        if overlaps(&current_range, &next_range) {
-            windows[idx - 1].overlapping = true;
-            windows[idx].overlapping = true;
+            let windows_overlap = overlaps(&lhs_range, &rhs_range)
+                && match (
+                    &windows[idx].primary_key_range,
+                    &windows[next_idx].primary_key_range,
+                ) {
+                    (Some(lhs), Some(rhs)) => primary_key_ranges_overlap(lhs, rhs),
+                    _ => true,
+                };
+            if windows_overlap {
+                windows[idx].overlapping = true;
+                windows[next_idx].overlapping = true;
+            }
         }
-        current_range = (
-            current_range.0.min(next_range.0),
-            current_range.1.max(next_range.1),
-        );
     }
 
     windows.into_iter().map(|w| (w.time_window, w)).collect()
@@ -390,11 +405,13 @@ fn find_latest_window_in_seconds<'a>(
 mod tests {
     use std::collections::HashSet;
 
+    use bytes::Bytes;
     use store_api::storage::FileId;
 
     use super::*;
     use crate::compaction::test_util::{
         new_file_handle, new_file_handle_with_sequence, new_file_handle_with_size_and_sequence,
+        new_file_handle_with_size_sequence_and_primary_key_range,
     };
     use crate::sst::file::Level;
 
@@ -566,6 +583,10 @@ mod tests {
     /// (Window value, overlapping, files' time ranges in window)
     type ExpectedWindowSpec = (i64, bool, Vec<(i64, i64)>);
 
+    fn pk_range(min: &'static [u8], max: &'static [u8]) -> Option<(Bytes, Bytes)> {
+        Some((Bytes::from_static(min), Bytes::from_static(max)))
+    }
+
     fn check_assign_to_windows_with_overlapping(
         file_time_ranges: &[(i64, i64)],
         time_window: i64,
@@ -588,8 +609,8 @@ mod tests {
             assert_eq!(*overlapping, actual_window.overlapping);
             let mut file_ranges = actual_window
                 .files
-                .iter()
-                .flat_map(|(_, f)| {
+                .values()
+                .flat_map(|f| {
                     f.files().iter().map(|f| {
                         let (s, e) = f.time_range();
                         (s.value(), e.value())
@@ -698,6 +719,63 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_assign_to_windows_not_overlapping_when_pk_disjoint() {
+        let files = [
+            new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                0,
+                1000,
+                0,
+                1,
+                10,
+                pk_range(b"a", b"f"),
+            ),
+            new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                500,
+                1999,
+                0,
+                2,
+                10,
+                pk_range(b"x", b"z"),
+            ),
+        ];
+
+        let windows = assign_to_windows(files.iter(), 2);
+
+        assert!(!windows.get(&2).unwrap().overlapping);
+    }
+
+    #[test]
+    fn test_assign_to_windows_pk_unknown_in_earlier_window_does_not_poison_later_windows() {
+        let files = [
+            new_file_handle(FileId::random(), 0, 1999, 0),
+            new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                2000,
+                3999,
+                0,
+                1,
+                10,
+                pk_range(b"a", b"f"),
+            ),
+            new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                3000,
+                4999,
+                0,
+                2,
+                10,
+                pk_range(b"x", b"z"),
+            ),
+        ];
+
+        let windows = assign_to_windows(files.iter(), 2);
+
+        assert!(!windows.get(&4).unwrap().overlapping);
+    }
+
     struct CompactionPickerTestCase {
         window_size: i64,
         input_files: Vec<FileHandle>,
@@ -832,6 +910,42 @@ mod tests {
         .check();
     }
 
+    #[test]
+    fn test_build_output_skips_pk_disjoint_files() {
+        let files = [
+            new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                0,
+                2999,
+                0,
+                1,
+                10,
+                pk_range(b"a", b"f"),
+            ),
+            new_file_handle_with_size_sequence_and_primary_key_range(
+                FileId::random(),
+                50,
+                2998,
+                0,
+                2,
+                10,
+                pk_range(b"x", b"z"),
+            ),
+        ];
+        let mut windows = assign_to_windows(files.iter(), 3);
+        let active_window = find_latest_window_in_seconds(files.iter(), 3);
+        let output = TwcsPicker {
+            trigger_file_num: 4,
+            time_window_seconds: None,
+            max_output_file_size: None,
+            append_mode: false,
+            max_background_tasks: None,
+        }
+        .build_output(RegionId::from_u64(0), &mut windows, active_window);
+
+        assert!(output.is_empty());
+    }
+
     #[test]
     fn test_append_mode_filter_large_files() {
         let file_ids = (0..4).map(|_| FileId::random()).collect::<Vec<_>>();
diff --git a/src/mito2/src/config.rs b/src/mito2/src/config.rs
index 602f5508ba..120b5adbe3 100644
--- a/src/mito2/src/config.rs
+++ b/src/mito2/src/config.rs
@@ -33,8 +33,6 @@ use crate::memtable::MemtableConfig;
 use crate::sst::DEFAULT_WRITE_BUFFER_SIZE;
 
 const MULTIPART_UPLOAD_MINIMUM_SIZE: ReadableSize = ReadableSize::mb(5);
-/// Default channel size for parallel scan task.
-pub(crate) const DEFAULT_SCAN_CHANNEL_SIZE: usize = 32;
 /// Default maximum number of SST files to scan concurrently.
 pub(crate) const DEFAULT_MAX_CONCURRENT_SCAN_FILES: usize = 384;
 
@@ -93,7 +91,9 @@ pub struct MitoConfig {
     pub max_background_compactions: usize,
     /// Max number of running background purge jobs (default: number of cpu cores).
     pub max_background_purges: usize,
-    /// Memory budget for compaction tasks. Setting it to 0 or "unlimited" disables the limit.
+    /// Memory budget for compaction tasks.
+    /// Supports absolute size (e.g., "2GiB", "512MB") or percentage of system memory (e.g., "50%").
+    /// Setting it to 0 or "unlimited" disables the limit.
     pub experimental_compaction_memory_limit: MemoryLimit,
     /// Behavior when compaction cannot acquire memory from the budget.
     pub experimental_compaction_on_exhausted: OnExhaustedPolicy,
@@ -116,6 +116,8 @@ pub struct MitoConfig {
     pub page_cache_size: ReadableSize,
     /// Cache size for time series selector (e.g. `last_value()`). Setting it to 0 to disable the cache.
     pub selector_result_cache_size: ReadableSize,
+    /// Cache size for flat range scan results. Setting it to 0 to disable the cache.
+    pub range_result_cache_size: ReadableSize,
     /// Whether to enable the write cache.
     pub enable_write_cache: bool,
     /// File system path for write cache dir's root, defaults to `{data_home}`.
@@ -140,8 +142,6 @@ pub struct MitoConfig {
     // Other configs:
     /// Buffer size for SST writing.
     pub sst_write_buffer_size: ReadableSize,
-    /// Capacity of the channel to send data from parallel scan tasks to the main task (default 32).
-    pub parallel_scan_channel_size: usize,
     /// Maximum number of SST files to scan concurrently (default 384).
     pub max_concurrent_scan_files: usize,
     /// Whether to allow stale entries read during replay.
@@ -149,6 +149,11 @@ pub struct MitoConfig {
     /// Memory limit for table scans across all queries. Setting it to 0 disables the limit.
     /// Supports absolute size (e.g., "2GB") or percentage (e.g., "50%").
     pub scan_memory_limit: MemoryLimit,
+    /// Behavior when scan memory tracking cannot acquire memory from the budget.
+    /// `wait` means `wait(10s)`, not unlimited waiting.
+    /// Defaults to [`OnExhaustedPolicy::Fail`], which intentionally differs from
+    /// [`OnExhaustedPolicy::default()`].
+    pub scan_memory_on_exhausted: OnExhaustedPolicy,
 
     /// Index configs.
     pub index: IndexConfig,
@@ -170,9 +175,9 @@ pub struct MitoConfig {
     #[serde(with = "humantime_serde")]
     pub min_compaction_interval: Duration,
 
-    /// Whether to enable experimental flat format as the default format.
+    /// Whether to enable flat format as the default SST format.
     /// When enabled, forces using BulkMemtable and BulkMemtableBuilder.
-    pub default_experimental_flat_format: bool,
+    pub default_flat_format: bool,
 
     pub gc: GcConfig,
 }
@@ -200,6 +205,7 @@ impl Default for MitoConfig {
             vector_cache_size: ReadableSize::mb(512),
             page_cache_size: ReadableSize::mb(512),
             selector_result_cache_size: ReadableSize::mb(512),
+            range_result_cache_size: ReadableSize::mb(512),
             enable_write_cache: false,
             write_cache_path: String::new(),
             write_cache_size: ReadableSize::gb(5),
@@ -209,10 +215,10 @@ impl Default for MitoConfig {
             enable_refill_cache_on_read: true,
             manifest_cache_size: ReadableSize::mb(256),
             sst_write_buffer_size: DEFAULT_WRITE_BUFFER_SIZE,
-            parallel_scan_channel_size: DEFAULT_SCAN_CHANNEL_SIZE,
             max_concurrent_scan_files: DEFAULT_MAX_CONCURRENT_SCAN_FILES,
             allow_stale_entries: false,
             scan_memory_limit: MemoryLimit::default(),
+            scan_memory_on_exhausted: OnExhaustedPolicy::Fail,
             index: IndexConfig::default(),
             inverted_index: InvertedIndexConfig::default(),
             fulltext_index: FulltextIndexConfig::default(),
@@ -221,7 +227,7 @@ impl Default for MitoConfig {
             vector_index: VectorIndexConfig::default(),
             memtable: MemtableConfig::default(),
             min_compaction_interval: Duration::from_secs(0),
-            default_experimental_flat_format: false,
+            default_flat_format: true,
             gc: GcConfig::default(),
         };
 
@@ -286,14 +292,6 @@ impl MitoConfig {
             );
         }
 
-        if self.parallel_scan_channel_size < 1 {
-            self.parallel_scan_channel_size = DEFAULT_SCAN_CHANNEL_SIZE;
-            warn!(
-                "Sanitize scan channel size to {}",
-                self.parallel_scan_channel_size
-            );
-        }
-
         // Sets write cache path if it is empty.
         if self.write_cache_path.trim().is_empty() {
             self.write_cache_path = data_home.to_string();
@@ -336,6 +334,7 @@ impl MitoConfig {
         self.vector_cache_size = mem_cache_size;
         self.page_cache_size = page_cache_size;
         self.selector_result_cache_size = mem_cache_size;
+        self.range_result_cache_size = mem_cache_size;
 
         self.index.adjust_buffer_and_cache_size(sys_memory);
     }
diff --git a/src/mito2/src/engine.rs b/src/mito2/src/engine.rs
index 1af79daff6..b3fa6ce75b 100644
--- a/src/mito2/src/engine.rs
+++ b/src/mito2/src/engine.rs
@@ -95,7 +95,7 @@ use common_base::Plugins;
 use common_error::ext::BoxedError;
 use common_meta::error::UnexpectedSnafu;
 use common_meta::key::SchemaMetadataManagerRef;
-use common_recordbatch::{MemoryPermit, QueryMemoryTracker, SendableRecordBatchStream};
+use common_recordbatch::{QueryMemoryTracker, SendableRecordBatchStream};
 use common_stat::get_total_memory_bytes;
 use common_telemetry::{info, tracing, warn};
 use common_wal::options::WalOptions;
@@ -129,8 +129,8 @@ use crate::cache::{CacheManagerRef, CacheStrategy};
 use crate::config::MitoConfig;
 use crate::engine::puffin_index::{IndexEntryContext, collect_index_entries_from_puffin};
 use crate::error::{
-    InvalidRequestSnafu, JoinSnafu, MitoManifestInfoSnafu, RecvSnafu, RegionNotFoundSnafu, Result,
-    SerdeJsonSnafu, SerializeColumnMetadataSnafu,
+    IncrementalQueryStaleSnafu, InvalidRequestSnafu, JoinSnafu, MitoManifestInfoSnafu, RecvSnafu,
+    RegionNotFoundSnafu, Result, SerdeJsonSnafu, SerializeColumnMetadataSnafu,
 };
 #[cfg(feature = "enterprise")]
 use crate::extension::BoxedExtensionRangeProviderFactory;
@@ -138,7 +138,8 @@ use crate::gc::GcLimiterRef;
 use crate::manifest::action::RegionEdit;
 use crate::memtable::MemtableStats;
 use crate::metrics::{
-    HANDLE_REQUEST_ELAPSED, SCAN_MEMORY_USAGE_BYTES, SCAN_REQUESTS_REJECTED_TOTAL,
+    HANDLE_REQUEST_ELAPSED, SCAN_MEMORY_EXHAUSTED_TOTAL, SCAN_MEMORY_USAGE_BYTES,
+    SCAN_REQUESTS_REJECTED_TOTAL,
 };
 use crate::read::scan_region::{ScanRegion, Scanner};
 use crate::read::stream::ScanBatchStream;
@@ -167,7 +168,6 @@ pub struct MitoEngineBuilder<'a, S: LogStore> {
     file_ref_manager: FileReferenceManagerRef,
     partition_expr_fetcher: PartitionExprFetcherRef,
     plugins: Plugins,
-    max_concurrent_queries: usize,
     #[cfg(feature = "enterprise")]
     extension_range_provider_factory: Option<BoxedExtensionRangeProviderFactory>,
 }
@@ -183,7 +183,6 @@ impl<'a, S: LogStore> MitoEngineBuilder<'a, S> {
         file_ref_manager: FileReferenceManagerRef,
         partition_expr_fetcher: PartitionExprFetcherRef,
         plugins: Plugins,
-        max_concurrent_queries: usize,
     ) -> Self {
         Self {
             data_home,
@@ -194,7 +193,6 @@ impl<'a, S: LogStore> MitoEngineBuilder<'a, S> {
             file_ref_manager,
             plugins,
             partition_expr_fetcher,
-            max_concurrent_queries,
             #[cfg(feature = "enterprise")]
             extension_range_provider_factory: None,
         }
@@ -230,13 +228,17 @@ impl<'a, S: LogStore> MitoEngineBuilder<'a, S> {
         let total_memory = get_total_memory_bytes().max(0) as u64;
         let scan_memory_limit = config.scan_memory_limit.resolve(total_memory) as usize;
         let scan_memory_tracker =
-            QueryMemoryTracker::new(scan_memory_limit, self.max_concurrent_queries)
-                .with_on_update(|usage| {
+            QueryMemoryTracker::builder(scan_memory_limit, config.scan_memory_on_exhausted)
+                .on_update(|usage| {
                     SCAN_MEMORY_USAGE_BYTES.set(usage as i64);
                 })
-                .with_on_reject(|| {
+                .on_exhausted(|| {
+                    SCAN_MEMORY_EXHAUSTED_TOTAL.inc();
+                })
+                .on_reject(|| {
                     SCAN_REQUESTS_REJECTED_TOTAL.inc();
-                });
+                })
+                .build();
 
         let inner = EngineInner {
             workers,
@@ -285,7 +287,6 @@ impl MitoEngine {
             file_ref_manager,
             partition_expr_fetcher,
             plugins,
-            0, // Default: no limit on concurrent queries
         );
         builder.try_build().await
     }
@@ -1012,11 +1013,29 @@ impl EngineInner {
 
     /// Handles the scan `request` and returns a [ScanRegion].
     #[tracing::instrument(skip_all, fields(region_id = %region_id))]
-    fn scan_region(&self, region_id: RegionId, request: ScanRequest) -> Result<ScanRegion> {
+    fn scan_region(&self, region_id: RegionId, mut request: ScanRequest) -> Result<ScanRegion> {
         let query_start = Instant::now();
         // Reading a region doesn't need to go through the region worker thread.
         let region = self.find_region(region_id)?;
-        let version = region.version();
+        let version_data = region.version_control.current();
+        let version = version_data.version;
+
+        if request.snapshot_on_scan && request.memtable_max_sequence.is_none() {
+            request.memtable_max_sequence = Some(version_data.committed_sequence);
+        }
+
+        if let Some(given_seq) = request.memtable_min_sequence {
+            let min_readable_seq = version.flushed_sequence;
+            ensure!(
+                given_seq >= min_readable_seq,
+                IncrementalQueryStaleSnafu {
+                    region_id,
+                    given_seq,
+                    min_readable_seq,
+                }
+            );
+        }
+
         // Get cache.
         let cache_manager = self.workers.cache_manager();
 
@@ -1026,7 +1045,6 @@ impl EngineInner {
             request,
             CacheStrategy::EnableAll(cache_manager),
         )
-        .with_parallel_scan_channel_size(self.config.parallel_scan_channel_size)
         .with_max_concurrent_scan_files(self.config.max_concurrent_scan_files)
         .with_ignore_inverted_index(self.config.inverted_index.apply_on_query.disabled())
         .with_ignore_fulltext_index(self.config.fulltext_index.apply_on_query.disabled())
@@ -1114,13 +1132,9 @@ impl EngineInner {
     }
 
     fn role(&self, region_id: RegionId) -> Option<RegionRole> {
-        self.workers.get_region(region_id).map(|region| {
-            if region.is_follower() {
-                RegionRole::Follower
-            } else {
-                RegionRole::Leader
-            }
-        })
+        self.workers
+            .get_region(region_id)
+            .map(|region| region.region_role())
     }
 }
 
@@ -1212,8 +1226,8 @@ impl RegionEngine for MitoEngine {
             .map_err(BoxedError::new)
     }
 
-    fn register_query_memory_permit(&self) -> Option<Arc<MemoryPermit>> {
-        Some(Arc::new(self.inner.scan_memory_tracker.register_permit()))
+    fn query_memory_tracker(&self) -> Option<QueryMemoryTracker> {
+        Some(self.inner.scan_memory_tracker.clone())
     }
 
     async fn get_committed_sequence(
@@ -1378,13 +1392,18 @@ impl MitoEngine {
         let wal_raw_entry_reader = Arc::new(LogStoreRawEntryReader::new(log_store.clone()));
         let total_memory = get_total_memory_bytes().max(0) as u64;
         let scan_memory_limit = config.scan_memory_limit.resolve(total_memory) as usize;
-        let scan_memory_tracker = QueryMemoryTracker::new(scan_memory_limit, 0)
-            .with_on_update(|usage| {
-                SCAN_MEMORY_USAGE_BYTES.set(usage as i64);
-            })
-            .with_on_reject(|| {
-                SCAN_REQUESTS_REJECTED_TOTAL.inc();
-            });
+        let scan_memory_tracker =
+            QueryMemoryTracker::builder(scan_memory_limit, config.scan_memory_on_exhausted)
+                .on_update(|usage| {
+                    SCAN_MEMORY_USAGE_BYTES.set(usage as i64);
+                })
+                .on_exhausted(|| {
+                    SCAN_MEMORY_EXHAUSTED_TOTAL.inc();
+                })
+                .on_reject(|| {
+                    SCAN_REQUESTS_REJECTED_TOTAL.inc();
+                })
+                .build();
         Ok(MitoEngine {
             inner: Arc::new(EngineInner {
                 workers: WorkerGroup::start_for_test(
diff --git a/src/mito2/src/engine/alter_test.rs b/src/mito2/src/engine/alter_test.rs
index e710e08688..a7798a7678 100644
--- a/src/mito2/src/engine/alter_test.rs
+++ b/src/mito2/src/engine/alter_test.rs
@@ -12,17 +12,18 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 use std::collections::HashMap;
 use std::sync::Arc;
 use std::time::Duration;
 
 use api::v1::helper::{row, tag_column_schema};
 use api::v1::value::ValueData;
-use api::v1::{ColumnDataType, Row, Rows, SemanticType};
+use api::v1::{ColumnDataType, Row, Rows, SemanticType, Value};
 use common_error::ext::ErrorExt;
 use common_meta::ddl::utils::{parse_column_metadatas, parse_manifest_infos_from_extensions};
 use common_recordbatch::RecordBatches;
+use datafusion_expr::col;
 use datatypes::prelude::ConcreteDataType;
 use datatypes::schema::{ColumnSchema, FulltextAnalyzer, FulltextBackend, FulltextOptions};
 use store_api::metadata::ColumnMetadata;
@@ -41,8 +42,8 @@ use crate::error;
 use crate::sst::FormatType;
 use crate::test_util::batch_util::sort_batches_and_print;
 use crate::test_util::{
-    CreateRequestBuilder, TestEnv, build_rows, build_rows_for_key, flush_region, put_rows,
-    rows_schema,
+    CreateRequestBuilder, TestEnv, build_rows, build_rows_for_key,
+    column_metadata_to_column_schema, flush_region, put_rows, rows_schema,
 };
 
 async fn scan_check_after_alter(engine: &MitoEngine, region_id: RegionId, expected: &str) {
@@ -102,6 +103,54 @@ fn alter_column_fulltext_options() -> RegionAlterRequest {
     }
 }
 
+fn add_nullable_field1() -> RegionAlterRequest {
+    RegionAlterRequest {
+        kind: AlterKind::AddColumns {
+            columns: vec![AddColumn {
+                column_metadata: ColumnMetadata {
+                    column_schema: ColumnSchema::new(
+                        "field_1",
+                        ConcreteDataType::float64_datatype(),
+                        true,
+                    ),
+                    semantic_type: SemanticType::Field,
+                    column_id: 3,
+                },
+                location: None,
+            }],
+        },
+    }
+}
+
+fn build_row_with_added_field(
+    metadata: &[ColumnMetadata],
+    tag_0: &str,
+    field_0: f64,
+    field_1: Option<f64>,
+    ts_millis: i64,
+) -> Row {
+    let values = metadata
+        .iter()
+        .map(|column| match column.column_schema.name.as_str() {
+            "tag_0" => Value {
+                value_data: Some(ValueData::StringValue(tag_0.to_string())),
+            },
+            "field_0" => Value {
+                value_data: Some(ValueData::F64Value(field_0)),
+            },
+            "field_1" => Value {
+                value_data: field_1.map(ValueData::F64Value),
+            },
+            "ts" => Value {
+                value_data: Some(ValueData::TimestampMillisecondValue(ts_millis)),
+            },
+            name => panic!("unexpected column {name}"),
+        })
+        .collect();
+
+    Row { values }
+}
+
 fn check_region_version(
     engine: &MitoEngine,
     region_id: RegionId,
@@ -141,7 +190,7 @@ async fn test_alter_region_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -213,7 +262,7 @@ async fn test_alter_region_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
@@ -236,6 +285,105 @@ async fn test_alter_region_with_format(flat_format: bool) {
     check_region_version(&engine, region_id, 1, 3, 1, 3);
 }
 
+#[tokio::test]
+async fn test_filter_is_null_after_alter_add_field() {
+    test_filter_is_null_after_alter_add_field_with_format(false).await;
+    test_filter_is_null_after_alter_add_field_with_format(true).await;
+}
+
+async fn test_filter_is_null_after_alter_add_field_with_format(flat_format: bool) {
+    common_telemetry::init_default_ut_logging();
+
+    let mut env = TestEnv::new().await;
+    let engine = env
+        .create_engine(MitoConfig {
+            default_flat_format: flat_format,
+            ..Default::default()
+        })
+        .await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+
+    env.get_schema_metadata_manager()
+        .register_region_table_info(
+            region_id.table_id(),
+            "test_table",
+            "test_catalog",
+            "test_schema",
+            None,
+            env.get_kv_backend(),
+        )
+        .await;
+
+    let column_schemas = rows_schema(&request);
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    put_rows(
+        &engine,
+        region_id,
+        Rows {
+            schema: column_schemas,
+            rows: vec![build_rows_for_key("a", 0, 1, 1).into_iter().next().unwrap()],
+        },
+    )
+    .await;
+    flush_region(&engine, region_id, None).await;
+
+    engine
+        .handle_request(region_id, RegionRequest::Alter(add_nullable_field1()))
+        .await
+        .unwrap();
+
+    let region = engine.get_region(region_id).unwrap();
+    let metadata = region.metadata().column_metadatas.clone();
+    let schema = metadata
+        .iter()
+        .map(column_metadata_to_column_schema)
+        .collect();
+
+    put_rows(
+        &engine,
+        region_id,
+        Rows {
+            schema,
+            rows: vec![build_row_with_added_field(
+                &metadata,
+                "a",
+                1.0,
+                Some(10.0),
+                0,
+            )],
+        },
+    )
+    .await;
+    flush_region(&engine, region_id, None).await;
+
+    // We skip field filters under merge mode because the flushed field values may be stale before
+    // the row is merged with newer field data.
+    let stream = engine
+        .scan_to_stream(
+            region_id,
+            ScanRequest {
+                filters: vec![col("field_1").is_null()],
+                ..Default::default()
+            },
+        )
+        .await
+        .unwrap();
+    let batches = RecordBatches::try_collect(stream).await.unwrap();
+    let expected = "\
++-------+---------+---------------------+---------+
+| tag_0 | field_0 | ts                  | field_1 |
++-------+---------+---------------------+---------+
+| a     | 1.0     | 1970-01-01T00:00:00 | 10.0    |
++-------+---------+---------------------+---------+";
+    assert_eq!(expected, batches.pretty_print().unwrap());
+}
+
 /// Build rows with schema (string, f64, ts_millis, string).
 fn build_rows_for_tags(
     tag0: &str,
@@ -267,7 +415,7 @@ async fn test_put_after_alter_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -318,7 +466,7 @@ async fn test_put_after_alter_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
@@ -387,7 +535,7 @@ async fn test_alter_region_retry_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -457,7 +605,7 @@ async fn test_alter_on_flushing_with_format(flat_format: bool) {
     let engine = env
         .create_engine_with(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             None,
@@ -574,7 +722,7 @@ async fn test_alter_column_fulltext_options_with_format(flat_format: bool) {
     let engine = env
         .create_engine_with(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             None,
@@ -681,7 +829,7 @@ async fn test_alter_column_fulltext_options_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
@@ -718,7 +866,7 @@ async fn test_alter_column_set_inverted_index_with_format(flat_format: bool) {
     let engine = env
         .create_engine_with(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             None,
@@ -816,7 +964,7 @@ async fn test_alter_column_set_inverted_index_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
@@ -853,7 +1001,7 @@ async fn test_alter_region_ttl_options_with_format(flat_format: bool) {
     let engine = env
         .create_engine_with(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             None,
@@ -916,7 +1064,7 @@ async fn test_write_stall_on_altering_with_format(flat_format: bool) {
     let engine = env
         .create_engine_with(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             None,
@@ -994,7 +1142,7 @@ async fn test_alter_region_sst_format_with_flush() {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: false,
+            default_flat_format: false,
             ..Default::default()
         })
         .await;
@@ -1085,7 +1233,7 @@ async fn test_alter_region_sst_format_with_flush() {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: false,
+                default_flat_format: false,
                 ..Default::default()
             },
         )
@@ -1118,7 +1266,7 @@ async fn test_alter_region_sst_format_without_flush() {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: false,
+            default_flat_format: false,
             ..Default::default()
         })
         .await;
@@ -1203,7 +1351,7 @@ async fn test_alter_region_sst_format_without_flush() {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: false,
+                default_flat_format: false,
                 ..Default::default()
             },
         )
@@ -1231,6 +1379,250 @@ async fn test_alter_region_sst_format_without_flush() {
     assert_eq!(expected_all_data, batches.pretty_print().unwrap());
 }
 
+#[tokio::test]
+async fn test_alter_region_sst_format_flat_to_pk_with_flush() {
+    common_telemetry::init_default_ut_logging();
+
+    let mut env = TestEnv::new().await;
+    let engine = env
+        .create_engine(MitoConfig {
+            default_flat_format: true,
+            ..Default::default()
+        })
+        .await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+
+    env.get_schema_metadata_manager()
+        .register_region_table_info(
+            region_id.table_id(),
+            "test_table",
+            "test_catalog",
+            "test_schema",
+            None,
+            env.get_kv_backend(),
+        )
+        .await;
+
+    let column_schemas = rows_schema(&request);
+    let table_dir = request.table_dir.clone();
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    // Inserts some data with flat format
+    let rows = Rows {
+        schema: column_schemas.clone(),
+        rows: build_rows(0, 3),
+    };
+    put_rows(&engine, region_id, rows).await;
+
+    // Flushes to create SST files with flat format
+    flush_region(&engine, region_id, None).await;
+
+    let expected_data = "\
++-------+---------+---------------------+
+| tag_0 | field_0 | ts                  |
++-------+---------+---------------------+
+| 0     | 0.0     | 1970-01-01T00:00:00 |
+| 1     | 1.0     | 1970-01-01T00:00:01 |
+| 2     | 2.0     | 1970-01-01T00:00:02 |
++-------+---------+---------------------+";
+    let request = ScanRequest::default();
+    let stream = engine.scan_to_stream(region_id, request).await.unwrap();
+    let batches = RecordBatches::try_collect(stream).await.unwrap();
+    assert_eq!(expected_data, batches.pretty_print().unwrap());
+
+    // Alters sst_format from flat to primary_key
+    let alter_format_request = RegionAlterRequest {
+        kind: AlterKind::SetRegionOptions {
+            options: vec![SetRegionOption::Format("primary_key".to_string())],
+        },
+    };
+    engine
+        .handle_request(region_id, RegionRequest::Alter(alter_format_request))
+        .await
+        .unwrap();
+
+    // Inserts more data after alter
+    let rows = Rows {
+        schema: column_schemas.clone(),
+        rows: build_rows(3, 6),
+    };
+    put_rows(&engine, region_id, rows).await;
+
+    // Flushes to create SST files with primary_key format
+    flush_region(&engine, region_id, None).await;
+
+    let expected_all_data = "\
++-------+---------+---------------------+
+| tag_0 | field_0 | ts                  |
++-------+---------+---------------------+
+| 0     | 0.0     | 1970-01-01T00:00:00 |
+| 1     | 1.0     | 1970-01-01T00:00:01 |
+| 2     | 2.0     | 1970-01-01T00:00:02 |
+| 3     | 3.0     | 1970-01-01T00:00:03 |
+| 4     | 4.0     | 1970-01-01T00:00:04 |
+| 5     | 5.0     | 1970-01-01T00:00:05 |
++-------+---------+---------------------+";
+    let request = ScanRequest::default();
+    let stream = engine.scan_to_stream(region_id, request).await.unwrap();
+    let batches = RecordBatches::try_collect(stream).await.unwrap();
+    assert_eq!(expected_all_data, batches.pretty_print().unwrap());
+
+    // Reopens region to verify format persists
+    let engine = env
+        .reopen_engine(
+            engine,
+            MitoConfig {
+                default_flat_format: false,
+                ..Default::default()
+            },
+        )
+        .await;
+    engine
+        .handle_request(
+            region_id,
+            RegionRequest::Open(RegionOpenRequest {
+                engine: String::new(),
+                table_dir,
+                path_type: PathType::Bare,
+                options: HashMap::default(),
+                skip_wal_replay: false,
+                checkpoint: None,
+            }),
+        )
+        .await
+        .unwrap();
+
+    let request = ScanRequest::default();
+    let stream = engine.scan_to_stream(region_id, request).await.unwrap();
+    let batches = RecordBatches::try_collect(stream).await.unwrap();
+    assert_eq!(expected_all_data, batches.pretty_print().unwrap());
+}
+
+#[tokio::test]
+async fn test_alter_region_sst_format_flat_to_pk_without_flush() {
+    common_telemetry::init_default_ut_logging();
+
+    let mut env = TestEnv::new().await;
+    let engine = env
+        .create_engine(MitoConfig {
+            default_flat_format: true,
+            ..Default::default()
+        })
+        .await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+
+    env.get_schema_metadata_manager()
+        .register_region_table_info(
+            region_id.table_id(),
+            "test_table",
+            "test_catalog",
+            "test_schema",
+            None,
+            env.get_kv_backend(),
+        )
+        .await;
+
+    let column_schemas = rows_schema(&request);
+    let table_dir = request.table_dir.clone();
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    let check_format = |engine: &MitoEngine, expected: Option<FormatType>| {
+        let current_format = engine
+            .get_region(region_id)
+            .unwrap()
+            .version()
+            .options
+            .sst_format;
+        assert_eq!(current_format, expected);
+    };
+    check_format(&engine, Some(FormatType::Flat));
+
+    // Inserts some data with flat format
+    let rows = Rows {
+        schema: column_schemas.clone(),
+        rows: build_rows(0, 3),
+    };
+    put_rows(&engine, region_id, rows).await;
+
+    // Alters sst_format from flat to primary_key
+    let alter_format_request = RegionAlterRequest {
+        kind: AlterKind::SetRegionOptions {
+            options: vec![SetRegionOption::Format("primary_key".to_string())],
+        },
+    };
+    engine
+        .handle_request(region_id, RegionRequest::Alter(alter_format_request))
+        .await
+        .unwrap();
+
+    check_format(&engine, Some(FormatType::PrimaryKey));
+
+    // Inserts more data after alter
+    let rows = Rows {
+        schema: column_schemas.clone(),
+        rows: build_rows(3, 6),
+    };
+    put_rows(&engine, region_id, rows).await;
+
+    let expected_all_data = "\
++-------+---------+---------------------+
+| tag_0 | field_0 | ts                  |
++-------+---------+---------------------+
+| 0     | 0.0     | 1970-01-01T00:00:00 |
+| 1     | 1.0     | 1970-01-01T00:00:01 |
+| 2     | 2.0     | 1970-01-01T00:00:02 |
+| 3     | 3.0     | 1970-01-01T00:00:03 |
+| 4     | 4.0     | 1970-01-01T00:00:04 |
+| 5     | 5.0     | 1970-01-01T00:00:05 |
++-------+---------+---------------------+";
+    let request = ScanRequest::default();
+    let stream = engine.scan_to_stream(region_id, request).await.unwrap();
+    let batches = RecordBatches::try_collect(stream).await.unwrap();
+    assert_eq!(expected_all_data, batches.pretty_print().unwrap());
+
+    // Reopens region to verify format persists
+    let engine = env
+        .reopen_engine(
+            engine,
+            MitoConfig {
+                default_flat_format: false,
+                ..Default::default()
+            },
+        )
+        .await;
+    engine
+        .handle_request(
+            region_id,
+            RegionRequest::Open(RegionOpenRequest {
+                engine: String::new(),
+                table_dir,
+                path_type: PathType::Bare,
+                options: HashMap::default(),
+                skip_wal_replay: false,
+                checkpoint: None,
+            }),
+        )
+        .await
+        .unwrap();
+
+    check_format(&engine, Some(FormatType::PrimaryKey));
+
+    let request = ScanRequest::default();
+    let stream = engine.scan_to_stream(region_id, request).await.unwrap();
+    let batches = RecordBatches::try_collect(stream).await.unwrap();
+    assert_eq!(expected_all_data, batches.pretty_print().unwrap());
+}
+
 #[tokio::test]
 async fn test_alter_region_append_mode_with_flush() {
     common_telemetry::init_default_ut_logging();
diff --git a/src/mito2/src/engine/append_mode_test.rs b/src/mito2/src/engine/append_mode_test.rs
index 61488b6592..de58e04e46 100644
--- a/src/mito2/src/engine/append_mode_test.rs
+++ b/src/mito2/src/engine/append_mode_test.rs
@@ -44,7 +44,7 @@ async fn test_append_mode_write_query_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -112,7 +112,7 @@ async fn test_append_mode_compaction_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -131,6 +131,7 @@ async fn test_append_mode_compaction_with_format(flat_format: bool) {
 
     let request = CreateRequestBuilder::new()
         .insert_option("compaction.type", "twcs")
+        .insert_option("compaction.twcs.trigger_file_num", "2")
         .insert_option("append_mode", "true")
         .build();
     let table_dir = request.table_dir.clone();
@@ -199,7 +200,7 @@ async fn test_append_mode_compaction_with_format(flat_format: bool) {
         .scanner(region_id, ScanRequest::default())
         .await
         .unwrap();
-    assert_eq!(1, scanner.num_files());
+    assert_eq!(2, scanner.num_files());
     assert_eq!(1, scanner.num_memtables());
     scanner.set_target_partitions(2);
     let stream = scanner.scan().await.unwrap();
@@ -211,7 +212,7 @@ async fn test_append_mode_compaction_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
@@ -238,7 +239,7 @@ async fn test_alter_append_mode_clears_merge_mode_with_format(flat_format: bool)
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -329,7 +330,7 @@ async fn test_alter_append_mode_clears_merge_mode_with_format(flat_format: bool)
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
@@ -376,7 +377,7 @@ async fn test_put_single_range_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -474,7 +475,7 @@ async fn test_put_single_range_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
diff --git a/src/mito2/src/engine/apply_staging_manifest_test.rs b/src/mito2/src/engine/apply_staging_manifest_test.rs
index 6904fbd624..e10ed0cbf0 100644
--- a/src/mito2/src/engine/apply_staging_manifest_test.rs
+++ b/src/mito2/src/engine/apply_staging_manifest_test.rs
@@ -12,9 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
-use std::fs;
 use std::sync::Arc;
+use std::{assert_matches, fs};
 
 use api::v1::Rows;
 use common_function::utils::partition_expr_version;
@@ -63,7 +62,7 @@ async fn test_apply_staging_manifest_invalid_region_state_with_format(flat_forma
     let mut env = TestEnv::with_prefix("invalid-region-state").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -126,7 +125,7 @@ async fn test_apply_staging_manifest_mismatched_partition_expr_with_format(flat_
     let mut env = TestEnv::with_prefix("mismatched-partition-expr").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -206,7 +205,7 @@ async fn test_apply_staging_manifest_success_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("success").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -231,9 +230,7 @@ async fn test_apply_staging_manifest_success_with_format(flat_format: bool) {
         engine
             .handle_request(
                 region_id,
-                RegionRequest::Flush(RegionFlushRequest {
-                    row_group_size: None,
-                }),
+                RegionRequest::Flush(RegionFlushRequest::default()),
             )
             .await
             .unwrap();
@@ -334,7 +331,7 @@ async fn test_apply_staging_manifest_success_with_format(flat_format: bool) {
     let staging_manifest = region.manifest_ctx.staging_manifest().await;
     assert!(staging_manifest.is_none());
     // The staging partition expr should be cleared.
-    assert!(region.staging_partition_info.lock().unwrap().is_none());
+    assert!(region.manifest_ctx.staging_partition_info().is_none());
     // The staging manifest directory should be empty.
     let data_home = env.data_home();
     let region_dir = format!("{}/data/test/1_0000000001", data_home.display());
@@ -407,7 +404,7 @@ async fn test_apply_staging_manifest_invalid_files_to_add_with_format(flat_forma
     let mut env = TestEnv::with_prefix("invalid-files-to-add").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -484,7 +481,7 @@ async fn test_apply_staging_manifest_change_edit_different_columns_fails_with_fo
     let mut env = TestEnv::with_prefix("apply-change-edit-different-columns").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -504,9 +501,7 @@ async fn test_apply_staging_manifest_change_edit_different_columns_fails_with_fo
     engine
         .handle_request(
             region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -600,7 +595,7 @@ async fn test_apply_staging_manifest_preserves_unflushed_memtable_with_format(fl
     let mut env = TestEnv::with_prefix("apply-preserve-memtable").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -625,9 +620,7 @@ async fn test_apply_staging_manifest_preserves_unflushed_memtable_with_format(fl
     engine
         .handle_request(
             region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -734,9 +727,7 @@ async fn test_split_repartition_causes_duplicate_data() {
     engine
         .handle_request(
             source_region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -929,9 +920,7 @@ async fn test_merge_repartition_data_integrity() {
     engine
         .handle_request(
             source_region_id_1,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -945,9 +934,7 @@ async fn test_merge_repartition_data_integrity() {
     engine
         .handle_request(
             source_region_id_2,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
diff --git a/src/mito2/src/engine/basic_test.rs b/src/mito2/src/engine/basic_test.rs
index ed92c2b4ac..c85c540488 100644
--- a/src/mito2/src/engine/basic_test.rs
+++ b/src/mito2/src/engine/basic_test.rs
@@ -56,7 +56,7 @@ async fn test_engine_new_stop_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("engine-stop").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -93,7 +93,7 @@ async fn test_write_to_region_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("write-to-region").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -134,7 +134,7 @@ async fn test_region_replay_with_format(factory: Option<LogStoreFactory>, flat_f
         .with_log_store_factory(factory.clone());
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -169,7 +169,7 @@ async fn test_region_replay_with_format(factory: Option<LogStoreFactory>, flat_f
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
@@ -234,7 +234,7 @@ async fn test_write_query_region_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -278,7 +278,7 @@ async fn test_different_order_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -339,7 +339,7 @@ async fn test_different_order_and_type_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -403,7 +403,7 @@ async fn test_put_delete_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -465,7 +465,7 @@ async fn test_delete_not_null_fields_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -524,7 +524,7 @@ async fn test_put_overwrite_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -594,7 +594,7 @@ async fn test_absent_and_invalid_columns_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -650,7 +650,7 @@ async fn test_region_usage_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("region_usage").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -716,7 +716,7 @@ async fn test_engine_with_write_cache_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let path = env.data_home().to_str().unwrap().to_string();
     let mito_config = MitoConfig {
-        default_experimental_flat_format: flat_format,
+        default_flat_format: flat_format,
         ..Default::default()
     }
     .enable_write_cache(path, ReadableSize::mb(512), None);
@@ -765,7 +765,7 @@ async fn test_cache_null_primary_key_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             vector_cache_size: ReadableSize::mb(32),
             ..Default::default()
         })
@@ -865,9 +865,9 @@ async fn test_cache_null_primary_key_with_format(flat_format: bool) {
 #[tokio::test]
 async fn test_list_ssts() {
     test_list_ssts_with_format(false, r#"
-ManifestSstEntry { table_dir: "test/", region_id: 47244640257(11, 1), table_id: 11, region_number: 1, region_group: 0, region_sequence: 1, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test/11_0000000001/<file_id>.parquet", file_size: 2513, index_file_path: Some("test/11_0000000001/index/<file_id>.puffin"), index_file_size: Some(250), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9000::Millisecond, sequence: Some(10), origin_region_id: 47244640257(11, 1), node_id: None, visible: true }
-ManifestSstEntry { table_dir: "test/", region_id: 47244640258(11, 2), table_id: 11, region_number: 2, region_group: 0, region_sequence: 2, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test/11_0000000002/<file_id>.parquet", file_size: 2513, index_file_path: Some("test/11_0000000002/index/<file_id>.puffin"), index_file_size: Some(250), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9000::Millisecond, sequence: Some(10), origin_region_id: 47244640258(11, 2), node_id: None, visible: true }
-ManifestSstEntry { table_dir: "test/", region_id: 94489280554(22, 42), table_id: 22, region_number: 42, region_group: 0, region_sequence: 42, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test/22_0000000042/<file_id>.parquet", file_size: 2513, index_file_path: Some("test/22_0000000042/index/<file_id>.puffin"), index_file_size: Some(250), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9000::Millisecond, sequence: Some(10), origin_region_id: 94489280554(22, 42), node_id: None, visible: true }"# ,
+ManifestSstEntry { table_dir: "test/", region_id: 47244640257(11, 1), table_id: 11, region_number: 1, region_group: 0, region_sequence: 1, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test/11_0000000001/<file_id>.parquet", file_size: 2701, index_file_path: Some("test/11_0000000001/index/<file_id>.puffin"), index_file_size: Some(250), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9000::Millisecond, sequence: Some(10), origin_region_id: 47244640257(11, 1), node_id: None, visible: true }
+ManifestSstEntry { table_dir: "test/", region_id: 47244640258(11, 2), table_id: 11, region_number: 2, region_group: 0, region_sequence: 2, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test/11_0000000002/<file_id>.parquet", file_size: 2701, index_file_path: Some("test/11_0000000002/index/<file_id>.puffin"), index_file_size: Some(250), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9000::Millisecond, sequence: Some(10), origin_region_id: 47244640258(11, 2), node_id: None, visible: true }
+ManifestSstEntry { table_dir: "test/", region_id: 94489280554(22, 42), table_id: 22, region_number: 42, region_group: 0, region_sequence: 42, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test/22_0000000042/<file_id>.parquet", file_size: 2701, index_file_path: Some("test/22_0000000042/index/<file_id>.puffin"), index_file_size: Some(250), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9000::Millisecond, sequence: Some(10), origin_region_id: 94489280554(22, 42), node_id: None, visible: true }"# ,
 r#"
 StorageSstEntry { file_path: "test/11_0000000001/<file_id>.parquet", file_size: None, last_modified_ms: None, node_id: None }
 StorageSstEntry { file_path: "test/11_0000000001/index/<file_id>.puffin", file_size: None, last_modified_ms: None, node_id: None }
@@ -876,9 +876,9 @@ StorageSstEntry { file_path: "test/11_0000000002/index/<file_id>.puffin", file_s
 StorageSstEntry { file_path: "test/22_0000000042/<file_id>.parquet", file_size: None, last_modified_ms: None, node_id: None }
 StorageSstEntry { file_path: "test/22_0000000042/index/<file_id>.puffin", file_size: None, last_modified_ms: None, node_id: None }"#).await;
     test_list_ssts_with_format(true, r#"
-ManifestSstEntry { table_dir: "test/", region_id: 47244640257(11, 1), table_id: 11, region_number: 1, region_group: 0, region_sequence: 1, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test/11_0000000001/<file_id>.parquet", file_size: 2837, index_file_path: Some("test/11_0000000001/index/<file_id>.puffin"), index_file_size: Some(250), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9000::Millisecond, sequence: Some(10), origin_region_id: 47244640257(11, 1), node_id: None, visible: true }
-ManifestSstEntry { table_dir: "test/", region_id: 47244640258(11, 2), table_id: 11, region_number: 2, region_group: 0, region_sequence: 2, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test/11_0000000002/<file_id>.parquet", file_size: 2837, index_file_path: Some("test/11_0000000002/index/<file_id>.puffin"), index_file_size: Some(250), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9000::Millisecond, sequence: Some(10), origin_region_id: 47244640258(11, 2), node_id: None, visible: true }
-ManifestSstEntry { table_dir: "test/", region_id: 94489280554(22, 42), table_id: 22, region_number: 42, region_group: 0, region_sequence: 42, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test/22_0000000042/<file_id>.parquet", file_size: 2837, index_file_path: Some("test/22_0000000042/index/<file_id>.puffin"), index_file_size: Some(250), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9000::Millisecond, sequence: Some(10), origin_region_id: 94489280554(22, 42), node_id: None, visible: true }"#, 
+ManifestSstEntry { table_dir: "test/", region_id: 47244640257(11, 1), table_id: 11, region_number: 1, region_group: 0, region_sequence: 1, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test/11_0000000001/<file_id>.parquet", file_size: 3099, index_file_path: Some("test/11_0000000001/index/<file_id>.puffin"), index_file_size: Some(250), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9000::Millisecond, sequence: Some(10), origin_region_id: 47244640257(11, 1), node_id: None, visible: true }
+ManifestSstEntry { table_dir: "test/", region_id: 47244640258(11, 2), table_id: 11, region_number: 2, region_group: 0, region_sequence: 2, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test/11_0000000002/<file_id>.parquet", file_size: 3099, index_file_path: Some("test/11_0000000002/index/<file_id>.puffin"), index_file_size: Some(250), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9000::Millisecond, sequence: Some(10), origin_region_id: 47244640258(11, 2), node_id: None, visible: true }
+ManifestSstEntry { table_dir: "test/", region_id: 94489280554(22, 42), table_id: 22, region_number: 42, region_group: 0, region_sequence: 42, file_id: "<file_id>", index_version: 0, level: 0, file_path: "test/22_0000000042/<file_id>.parquet", file_size: 3099, index_file_path: Some("test/22_0000000042/index/<file_id>.puffin"), index_file_size: Some(250), num_rows: 10, num_row_groups: 1, num_series: Some(1), min_ts: 0::Millisecond, max_ts: 9000::Millisecond, sequence: Some(10), origin_region_id: 94489280554(22, 42), node_id: None, visible: true }"#, 
 r#"
 StorageSstEntry { file_path: "test/11_0000000001/<file_id>.parquet", file_size: None, last_modified_ms: None, node_id: None }
 StorageSstEntry { file_path: "test/11_0000000001/index/<file_id>.puffin", file_size: None, last_modified_ms: None, node_id: None }
@@ -896,7 +896,7 @@ async fn test_list_ssts_with_format(
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -932,9 +932,7 @@ async fn test_list_ssts_with_format(
         engine
             .handle_request(
                 *region_id,
-                RegionRequest::Flush(RegionFlushRequest {
-                    row_group_size: None,
-                }),
+                RegionRequest::Flush(RegionFlushRequest::default()),
             )
             .await
             .unwrap();
@@ -970,6 +968,8 @@ async fn test_list_ssts_with_format(
         .map(|mut e| {
             let i = e.file_path.rfind('/').unwrap();
             e.file_path.replace_range(i..(i + 37), "/<file_id>");
+            e.file_size = None;
+            e.last_modified_ms = None;
             format!("\n{:?}", e)
         })
         .sorted()
@@ -1002,7 +1002,7 @@ async fn test_all_index_metas_list_all_types_with_format(flat_format: bool, expe
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -1122,9 +1122,7 @@ async fn test_all_index_metas_list_all_types_with_format(flat_format: bool, expe
     engine
         .handle_request(
             region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
diff --git a/src/mito2/src/engine/batch_catchup_test.rs b/src/mito2/src/engine/batch_catchup_test.rs
index d8c744a733..dc0b552adc 100644
--- a/src/mito2/src/engine/batch_catchup_test.rs
+++ b/src/mito2/src/engine/batch_catchup_test.rs
@@ -49,7 +49,7 @@ async fn test_batch_catchup_with_format(factory: Option<LogStoreFactory>, flat_f
         .with_log_store_factory(factory.clone());
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -135,7 +135,7 @@ async fn test_batch_catchup_with_format(factory: Option<LogStoreFactory>, flat_f
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
@@ -216,7 +216,7 @@ async fn test_batch_catchup_err_with_format(factory: Option<LogStoreFactory>, fl
         .with_log_store_factory(factory.clone());
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
diff --git a/src/mito2/src/engine/batch_open_test.rs b/src/mito2/src/engine/batch_open_test.rs
index c718ef248c..6b16b3c120 100644
--- a/src/mito2/src/engine/batch_open_test.rs
+++ b/src/mito2/src/engine/batch_open_test.rs
@@ -49,7 +49,7 @@ async fn test_batch_open_with_format(factory: Option<LogStoreFactory>, flat_form
         .with_log_store_factory(factory.clone());
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -157,7 +157,7 @@ async fn test_batch_open_with_format(factory: Option<LogStoreFactory>, flat_form
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
@@ -193,7 +193,7 @@ async fn test_batch_open_err_with_format(factory: Option<LogStoreFactory>, flat_
         .with_log_store_factory(factory.clone());
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
diff --git a/src/mito2/src/engine/bump_committed_sequence_test.rs b/src/mito2/src/engine/bump_committed_sequence_test.rs
index 00d2c0f51c..12db0044c5 100644
--- a/src/mito2/src/engine/bump_committed_sequence_test.rs
+++ b/src/mito2/src/engine/bump_committed_sequence_test.rs
@@ -35,7 +35,7 @@ async fn test_bump_committed_sequence_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -97,7 +97,7 @@ async fn test_bump_committed_sequence_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
@@ -136,7 +136,7 @@ async fn test_bump_committed_sequence_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
diff --git a/src/mito2/src/engine/catchup_test.rs b/src/mito2/src/engine/catchup_test.rs
index 0c7d058e4d..e10e91b51b 100644
--- a/src/mito2/src/engine/catchup_test.rs
+++ b/src/mito2/src/engine/catchup_test.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 use std::collections::HashMap;
 
 use api::v1::Rows;
@@ -701,7 +701,7 @@ async fn test_catchup_not_exist_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
diff --git a/src/mito2/src/engine/close_test.rs b/src/mito2/src/engine/close_test.rs
index 965a4f6fff..4c06583b0b 100644
--- a/src/mito2/src/engine/close_test.rs
+++ b/src/mito2/src/engine/close_test.rs
@@ -29,7 +29,7 @@ async fn test_engine_close_region_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("close").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
diff --git a/src/mito2/src/engine/compaction_test.rs b/src/mito2/src/engine/compaction_test.rs
index df8521535f..f76e9f8bf9 100644
--- a/src/mito2/src/engine/compaction_test.rs
+++ b/src/mito2/src/engine/compaction_test.rs
@@ -18,6 +18,8 @@ use std::sync::Arc;
 use std::time::Duration;
 
 use api::v1::{ColumnSchema, Rows};
+use common_error::ext::ErrorExt;
+use common_error::status_code::StatusCode;
 use common_recordbatch::{RecordBatches, SendableRecordBatchStream};
 use datatypes::arrow::array::AsArray;
 use datatypes::arrow::datatypes::TimestampMillisecondType;
@@ -53,9 +55,7 @@ pub(crate) async fn put_and_flush(
     let result = engine
         .handle_request(
             region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -66,9 +66,7 @@ async fn flush(engine: &MitoEngine, region_id: RegionId) {
     let result = engine
         .handle_request(
             region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -114,9 +112,7 @@ pub(crate) async fn delete_and_flush(
     let result = engine
         .handle_request(
             region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -147,7 +143,7 @@ async fn test_compaction_region_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -223,7 +219,7 @@ async fn test_infer_compaction_time_window_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -374,7 +370,7 @@ async fn test_compaction_overlapping_files_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -445,7 +441,7 @@ async fn test_compaction_region_with_overlapping_with_format(flat_format: bool)
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -503,7 +499,7 @@ async fn test_compaction_region_with_overlapping_delete_all_with_format(flat_for
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -571,7 +567,7 @@ async fn test_readonly_during_compaction_with_format(flat_format: bool) {
     let engine = env
         .create_engine_with(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 // Ensure there is only one background worker for purge task.
                 max_background_purges: 1,
                 ..Default::default()
@@ -650,7 +646,7 @@ async fn test_readonly_during_compaction_with_format(flat_format: bool) {
 }
 
 #[tokio::test]
-async fn test_enter_staging_deferred_by_inflight_compaction() {
+async fn test_enter_staging_cancels_inflight_local_compaction_before_commit() {
     common_telemetry::init_default_ut_logging();
     let mut env = TestEnv::new().await;
     let listener = Arc::new(CompactionListener::default());
@@ -706,17 +702,91 @@ async fn test_enter_staging_deferred_by_inflight_compaction() {
                 }),
             )
             .await
-            .unwrap();
     });
 
     tokio::time::sleep(Duration::from_millis(100)).await;
-    assert!(!enter_staging.is_finished());
+    // The enter staging should finished, and the compaction should be cancelled.
+    assert!(enter_staging.is_finished());
+    let _ = enter_staging.await.unwrap().unwrap();
+}
 
-    listener.wake();
-    enter_staging.await.unwrap();
+#[tokio::test]
+async fn test_manual_compaction_returns_cancelled_when_enter_staging_cancels_it() {
+    common_telemetry::init_default_ut_logging();
+    let mut env = TestEnv::new().await;
+    let listener = Arc::new(CompactionListener::default());
+    let engine = env
+        .create_engine_with(
+            MitoConfig {
+                max_background_purges: 1,
+                ..Default::default()
+            },
+            None,
+            Some(listener.clone()),
+            None,
+        )
+        .await;
 
-    let region = engine.get_region(region_id).unwrap();
-    assert!(region.is_staging());
+    let region_id = RegionId::new(2050, 1);
+    env.get_schema_metadata_manager()
+        .register_region_table_info(
+            region_id.table_id(),
+            "test_table",
+            "test_catalog",
+            "test_schema",
+            None,
+            env.get_kv_backend(),
+        )
+        .await;
+
+    let request = CreateRequestBuilder::new()
+        .insert_option("compaction.type", "twcs")
+        .build();
+    let column_schemas = request
+        .column_metadatas
+        .iter()
+        .map(column_metadata_to_column_schema)
+        .collect::<Vec<_>>();
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    put_and_flush(&engine, region_id, &column_schemas, 0..10).await;
+    put_and_flush(&engine, region_id, &column_schemas, 5..20).await;
+
+    let engine_cloned = engine.clone();
+    let compact = tokio::spawn(async move {
+        engine_cloned
+            .handle_request(
+                region_id,
+                RegionRequest::Compact(RegionCompactRequest::default()),
+            )
+            .await
+    });
+
+    listener.wait_handle_finished().await;
+
+    let engine_cloned = engine.clone();
+    let enter_staging = tokio::spawn(async move {
+        engine_cloned
+            .handle_request(
+                region_id,
+                RegionRequest::EnterStaging(EnterStagingRequest {
+                    partition_directive: StagingPartitionDirective::RejectAllWrites,
+                }),
+            )
+            .await
+    });
+
+    tokio::time::sleep(Duration::from_millis(100)).await;
+    assert!(compact.is_finished());
+    assert!(enter_staging.is_finished());
+
+    let err = compact.await.unwrap().unwrap_err();
+    assert_eq!(err.status_code(), StatusCode::Cancelled);
+
+    let _ = enter_staging.await.unwrap();
 }
 
 #[tokio::test]
@@ -730,7 +800,7 @@ async fn test_compaction_update_time_window_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -836,7 +906,7 @@ async fn test_change_region_compaction_window_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -938,7 +1008,7 @@ async fn test_change_region_compaction_window_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
@@ -981,7 +1051,7 @@ async fn test_open_overwrite_compaction_window_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -1040,7 +1110,7 @@ async fn test_open_overwrite_compaction_window_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
diff --git a/src/mito2/src/engine/copy_region_from_test.rs b/src/mito2/src/engine/copy_region_from_test.rs
index 75580d5c0b..97966af670 100644
--- a/src/mito2/src/engine/copy_region_from_test.rs
+++ b/src/mito2/src/engine/copy_region_from_test.rs
@@ -12,9 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
-use std::fs;
 use std::sync::Arc;
+use std::{assert_matches, fs};
 
 use api::v1::Rows;
 use common_error::ext::ErrorExt;
@@ -42,7 +41,7 @@ async fn test_engine_copy_region_from_with_format(flat_format: bool, with_index:
     let mut env = TestEnv::with_prefix("copy-region-from").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -72,9 +71,7 @@ async fn test_engine_copy_region_from_with_format(flat_format: bool, with_index:
     engine
         .handle_request(
             source_region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -157,7 +154,7 @@ async fn test_engine_copy_region_failure_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await.with_mock_layer(mock_layer);
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -185,9 +182,7 @@ async fn test_engine_copy_region_failure_with_format(flat_format: bool) {
     engine
         .handle_request(
             source_region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -284,7 +279,7 @@ async fn test_engine_copy_region_invalid_args_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -329,7 +324,7 @@ async fn test_engine_copy_region_unexpected_state_with_format(flat_format: bool)
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
diff --git a/src/mito2/src/engine/create_test.rs b/src/mito2/src/engine/create_test.rs
index e5980d9442..6dff346539 100644
--- a/src/mito2/src/engine/create_test.rs
+++ b/src/mito2/src/engine/create_test.rs
@@ -36,7 +36,7 @@ async fn test_engine_create_new_region_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("new-region").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -61,7 +61,7 @@ async fn test_engine_create_existing_region_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("create-existing").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -91,7 +91,7 @@ async fn test_engine_create_close_create_region_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("create-close-create").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -131,7 +131,7 @@ async fn test_engine_create_with_different_id_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -160,7 +160,7 @@ async fn test_engine_create_with_different_schema_with_format(flat_format: bool)
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -190,7 +190,7 @@ async fn test_engine_create_with_different_primary_key_with_format(flat_format:
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -220,7 +220,7 @@ async fn test_engine_create_with_options_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -253,7 +253,7 @@ async fn test_engine_create_with_custom_store_with_format(flat_format: bool) {
     let engine = env
         .create_engine_with_multiple_object_stores(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             None,
@@ -301,7 +301,7 @@ async fn test_engine_create_with_memtable_opts_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -353,7 +353,7 @@ async fn create_with_partition_expr_persists_manifest_with_format(flat_format: b
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -401,7 +401,7 @@ async fn test_engine_create_with_format_one_case(create_format: &str, default_fl
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: default_flat_format,
+            default_flat_format,
             ..Default::default()
         })
         .await;
diff --git a/src/mito2/src/engine/drop_test.rs b/src/mito2/src/engine/drop_test.rs
index b3da775117..a34a5d1172 100644
--- a/src/mito2/src/engine/drop_test.rs
+++ b/src/mito2/src/engine/drop_test.rs
@@ -45,7 +45,7 @@ async fn test_engine_drop_region_with_format(flat_format: bool) {
     let engine = env
         .create_engine_with(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             None,
@@ -175,7 +175,7 @@ async fn test_engine_drop_region_for_custom_store_with_format(flat_format: bool)
     let engine = env
         .create_engine_with_multiple_object_stores(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             None,
diff --git a/src/mito2/src/engine/edit_region_test.rs b/src/mito2/src/engine/edit_region_test.rs
index 01bdf60070..84bed66e38 100644
--- a/src/mito2/src/engine/edit_region_test.rs
+++ b/src/mito2/src/engine/edit_region_test.rs
@@ -15,21 +15,30 @@
 use std::sync::{Arc, Mutex};
 use std::time::Duration;
 
+use api::v1::{ArrowIpc, Rows};
+use common_error::ext::ErrorExt;
+use common_error::status_code::StatusCode;
+use common_recordbatch::DfRecordBatch;
+use common_test_util::flight::encode_to_flight_data;
 use common_time::util::current_time_millis;
+use datatypes::arrow::array::{ArrayRef, Float64Array, StringArray, TimestampMillisecondArray};
+use datatypes::arrow::datatypes::{DataType, Field, Schema, TimeUnit};
 use object_store::ObjectStore;
 use store_api::region_engine::RegionEngine;
-use store_api::region_request::RegionRequest;
+use store_api::region_request::{
+    RegionBulkInsertsRequest, RegionCloseRequest, RegionPutRequest, RegionRequest,
+};
 use store_api::storage::{FileId, RegionId};
-use tokio::sync::{Barrier, oneshot};
+use tokio::sync::{Barrier, mpsc, oneshot};
 
 use crate::config::MitoConfig;
 use crate::engine::MitoEngine;
 use crate::engine::flush_test::MockTimeProvider;
 use crate::engine::listener::EventListener;
 use crate::manifest::action::RegionEdit;
-use crate::region::MitoRegionRef;
+use crate::region::{MitoRegionRef, RegionLeaderState, RegionRoleState};
 use crate::sst::file::FileMeta;
-use crate::test_util::{CreateRequestBuilder, TestEnv};
+use crate::test_util::{CreateRequestBuilder, TestEnv, build_rows, rows_schema};
 
 #[tokio::test]
 async fn test_edit_region_schedule_compaction() {
@@ -54,7 +63,7 @@ async fn test_edit_region_schedule_compaction_with_format(flat_format: bool) {
     let (tx, mut rx) = oneshot::channel();
     let config = MitoConfig {
         min_compaction_interval: Duration::from_secs(60 * 60),
-        default_experimental_flat_format: flat_format,
+        default_flat_format: flat_format,
         ..Default::default()
     };
     let time_provider = Arc::new(MockTimeProvider::new(current_time_millis()));
@@ -154,7 +163,7 @@ async fn test_edit_region_fill_cache_with_format(flat_format: bool) {
             MitoConfig {
                 // Write cache must be enabled to download the ingested SST file.
                 enable_write_cache: true,
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             None,
@@ -210,6 +219,314 @@ async fn test_edit_region_fill_cache_with_format(flat_format: bool) {
     assert_eq!(file_id, actual);
 }
 
+#[tokio::test]
+async fn test_write_during_region_editing_is_queued() {
+    let mut env = TestEnv::new().await;
+    let (engine, mut request_rx) = create_engine_with_request_listener(&mut env).await;
+
+    let region_id = RegionId::new(1, 1);
+    let create_request = CreateRequestBuilder::new().build();
+    let column_schemas = rows_schema(&create_request);
+    engine
+        .handle_request(region_id, RegionRequest::Create(create_request))
+        .await
+        .unwrap();
+    let region = engine.get_region(region_id).unwrap();
+
+    let manifest_guard = region.manifest_ctx.manifest_manager.write().await;
+    let edit = test_region_edit(region.region_id, FileId::random());
+
+    let edit_engine = engine.clone();
+    let edit_task = tokio::spawn(async move { edit_engine.edit_region(region_id, edit).await });
+
+    tokio::time::timeout(Duration::from_secs(3), async {
+        while region.state() != RegionRoleState::Leader(RegionLeaderState::Editing) {
+            tokio::time::sleep(Duration::from_millis(10)).await;
+        }
+    })
+    .await
+    .unwrap();
+    drain_worker_recv_events(&mut request_rx);
+
+    let write_engine = engine.clone();
+    let rows = Rows {
+        schema: column_schemas,
+        rows: build_rows(0, 1),
+    };
+    let write_task = tokio::spawn(async move {
+        write_engine
+            .handle_request(
+                region_id,
+                RegionRequest::Put(RegionPutRequest {
+                    rows,
+                    hint: None,
+                    partition_expr_version: None,
+                }),
+            )
+            .await
+    });
+    wait_worker_recv_event(&mut request_rx).await;
+
+    let second_file_id = FileId::random();
+    let second_edit = test_region_edit(region.region_id, second_file_id);
+    let second_edit_engine = engine.clone();
+    let second_edit_task =
+        tokio::spawn(async move { second_edit_engine.edit_region(region_id, second_edit).await });
+    wait_worker_recv_event(&mut request_rx).await;
+
+    drop(manifest_guard);
+    edit_task.await.unwrap().unwrap();
+    let output = write_task.await.unwrap().unwrap();
+    assert_eq!(1, output.affected_rows);
+    second_edit_task.await.unwrap().unwrap();
+
+    let second_file_sequence = region.version().ssts.levels()[0]
+        .files
+        .iter()
+        .find(|(file_id, _)| **file_id == second_file_id)
+        .and_then(|(_, file)| file.meta_ref().sequence)
+        .map(|sequence| sequence.get());
+    assert_eq!(Some(3), second_file_sequence);
+}
+
+#[tokio::test]
+async fn test_bulk_insert_during_region_editing_is_queued() {
+    let mut env = TestEnv::new().await;
+    let (engine, mut request_rx) = create_engine_with_request_listener(&mut env).await;
+
+    let region_id = RegionId::new(1, 1);
+    engine
+        .handle_request(
+            region_id,
+            RegionRequest::Create(CreateRequestBuilder::new().build()),
+        )
+        .await
+        .unwrap();
+    let region = engine.get_region(region_id).unwrap();
+
+    let manifest_guard = region.manifest_ctx.manifest_manager.write().await;
+    let edit = test_region_edit(region.region_id, FileId::random());
+
+    let edit_engine = engine.clone();
+    let edit_task = tokio::spawn(async move { edit_engine.edit_region(region_id, edit).await });
+
+    tokio::time::timeout(Duration::from_secs(3), async {
+        while region.state() != RegionRoleState::Leader(RegionLeaderState::Editing) {
+            tokio::time::sleep(Duration::from_millis(10)).await;
+        }
+    })
+    .await
+    .unwrap();
+    drain_worker_recv_events(&mut request_rx);
+
+    let bulk_engine = engine.clone();
+    let bulk_task = tokio::spawn(async move {
+        bulk_engine
+            .handle_request(
+                region_id,
+                RegionRequest::BulkInserts(build_bulk_insert_request(region_id, 0, 1)),
+            )
+            .await
+    });
+
+    wait_worker_recv_event(&mut request_rx).await;
+
+    let second_file_id = FileId::random();
+    let second_edit = test_region_edit(region.region_id, second_file_id);
+    let second_edit_engine = engine.clone();
+    let second_edit_task =
+        tokio::spawn(async move { second_edit_engine.edit_region(region_id, second_edit).await });
+    wait_worker_recv_event(&mut request_rx).await;
+
+    drop(manifest_guard);
+    edit_task.await.unwrap().unwrap();
+    let output = bulk_task.await.unwrap().unwrap();
+    assert_eq!(1, output.affected_rows);
+    second_edit_task.await.unwrap().unwrap();
+
+    let second_file_sequence = region.version().ssts.levels()[0]
+        .files
+        .iter()
+        .find(|(file_id, _)| **file_id == second_file_id)
+        .and_then(|(_, file)| file.meta_ref().sequence)
+        .map(|sequence| sequence.get());
+    assert_eq!(Some(3), second_file_sequence);
+}
+
+#[tokio::test]
+async fn test_stalled_write_fails_fast_if_region_closed_during_editing() {
+    let mut env = TestEnv::new().await;
+    let (engine, mut request_rx) = create_engine_with_request_listener(&mut env).await;
+
+    let region_id = RegionId::new(1, 1);
+    let create_request = CreateRequestBuilder::new().build();
+    let column_schemas = rows_schema(&create_request);
+    engine
+        .handle_request(region_id, RegionRequest::Create(create_request))
+        .await
+        .unwrap();
+    let region = engine.get_region(region_id).unwrap();
+
+    let manifest_guard = region.manifest_ctx.manifest_manager.write().await;
+    let edit = test_region_edit(region.region_id, FileId::random());
+
+    let edit_engine = engine.clone();
+    let edit_task = tokio::spawn(async move { edit_engine.edit_region(region_id, edit).await });
+
+    tokio::time::timeout(Duration::from_secs(3), async {
+        while region.state() != RegionRoleState::Leader(RegionLeaderState::Editing) {
+            tokio::time::sleep(Duration::from_millis(10)).await;
+        }
+    })
+    .await
+    .unwrap();
+
+    drain_worker_recv_events(&mut request_rx);
+
+    let write_engine = engine.clone();
+    let rows = Rows {
+        schema: column_schemas,
+        rows: build_rows(0, 1),
+    };
+    let write_task = tokio::spawn(async move {
+        write_engine
+            .handle_request(
+                region_id,
+                RegionRequest::Put(RegionPutRequest {
+                    rows,
+                    hint: None,
+                    partition_expr_version: None,
+                }),
+            )
+            .await
+    });
+
+    wait_worker_recv_event(&mut request_rx).await;
+
+    let second_edit_engine = engine.clone();
+    let second_edit = test_region_edit(region.region_id, FileId::random());
+    let second_edit_task =
+        tokio::spawn(async move { second_edit_engine.edit_region(region_id, second_edit).await });
+    wait_worker_recv_event(&mut request_rx).await;
+
+    let close_engine = engine.clone();
+    let close_task = tokio::spawn(async move {
+        close_engine
+            .handle_request(region_id, RegionRequest::Close(RegionCloseRequest {}))
+            .await
+    });
+
+    wait_worker_recv_event(&mut request_rx).await;
+
+    drop(manifest_guard);
+    close_task.await.unwrap().unwrap();
+
+    let write_result = tokio::time::timeout(Duration::from_secs(3), write_task)
+        .await
+        .expect("stalled write should fail after region is closed")
+        .unwrap();
+    assert_eq!(
+        StatusCode::RegionNotFound,
+        write_result.unwrap_err().status_code()
+    );
+    assert_eq!(
+        StatusCode::RegionNotFound,
+        second_edit_task.await.unwrap().unwrap_err().status_code()
+    );
+    assert!(edit_task.await.unwrap().is_err());
+}
+
+struct RecvRequestListener {
+    tx: mpsc::UnboundedSender<usize>,
+}
+
+impl EventListener for RecvRequestListener {
+    fn on_recv_requests(&self, request_num: usize) {
+        let _ = self.tx.send(request_num);
+    }
+}
+
+async fn create_engine_with_request_listener(
+    env: &mut TestEnv,
+) -> (MitoEngine, mpsc::UnboundedReceiver<usize>) {
+    let (tx, rx) = mpsc::unbounded_channel();
+    let engine = env
+        .create_engine_with(
+            MitoConfig::default(),
+            None,
+            Some(Arc::new(RecvRequestListener { tx })),
+            None,
+        )
+        .await;
+    (engine, rx)
+}
+
+fn drain_worker_recv_events(rx: &mut mpsc::UnboundedReceiver<usize>) {
+    while rx.try_recv().is_ok() {}
+}
+
+async fn wait_worker_recv_event(rx: &mut mpsc::UnboundedReceiver<usize>) {
+    tokio::time::timeout(Duration::from_secs(3), rx.recv())
+        .await
+        .unwrap()
+        .unwrap();
+}
+
+fn test_region_edit(region_id: RegionId, file_id: FileId) -> RegionEdit {
+    RegionEdit {
+        files_to_add: vec![FileMeta {
+            region_id,
+            file_id,
+            level: 0,
+            ..Default::default()
+        }],
+        files_to_remove: vec![],
+        timestamp_ms: None,
+        compaction_time_window: None,
+        flushed_entry_id: None,
+        flushed_sequence: None,
+        committed_sequence: None,
+    }
+}
+
+fn build_bulk_insert_request(
+    region_id: RegionId,
+    start: usize,
+    end: usize,
+) -> RegionBulkInsertsRequest {
+    let schema = Arc::new(Schema::new(vec![
+        Field::new("tag_0", DataType::Utf8, true),
+        Field::new("field_0", DataType::Float64, true),
+        Field::new(
+            "ts",
+            DataType::Timestamp(TimeUnit::Millisecond, None),
+            false,
+        ),
+    ]));
+    let tag = Arc::new(StringArray::from_iter_values(
+        (start..end).map(|value| value.to_string()),
+    )) as ArrayRef;
+    let field = Arc::new(Float64Array::from_iter_values(
+        (start..end).map(|value| value as f64),
+    )) as ArrayRef;
+    let ts = Arc::new(TimestampMillisecondArray::from_iter_values(
+        (start..end).map(|value| value as i64 * 1000),
+    )) as ArrayRef;
+    let payload = DfRecordBatch::try_new(schema, vec![tag, field, ts]).unwrap();
+    let (schema, record_batch) = encode_to_flight_data(payload.clone());
+
+    RegionBulkInsertsRequest {
+        region_id,
+        payload,
+        raw_data: ArrowIpc {
+            schema: schema.data_header,
+            data_header: record_batch.data_header,
+            payload: record_batch.data_body,
+        },
+        partition_expr_version: None,
+    }
+}
+
 #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
 async fn test_edit_region_concurrently() {
     test_edit_region_concurrently_with_format(false).await;
@@ -268,7 +585,7 @@ async fn test_edit_region_concurrently_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             // Suppress the compaction to not impede the speed of this kinda stress testing.
             min_compaction_interval: Duration::from_secs(60 * 60),
             ..Default::default()
diff --git a/src/mito2/src/engine/filter_deleted_test.rs b/src/mito2/src/engine/filter_deleted_test.rs
index c40fc7ba02..319410c505 100644
--- a/src/mito2/src/engine/filter_deleted_test.rs
+++ b/src/mito2/src/engine/filter_deleted_test.rs
@@ -12,8 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use api::v1::Rows;
+use api::v1::value::ValueData;
+use api::v1::{Row, Rows, Value};
 use common_recordbatch::RecordBatches;
+use datafusion_expr::{col, lit};
 use store_api::region_engine::RegionEngine;
 use store_api::region_request::RegionRequest;
 use store_api::storage::{RegionId, ScanRequest};
@@ -24,6 +26,22 @@ use crate::test_util::{
     CreateRequestBuilder, TestEnv, build_rows, delete_rows, flush_region, put_rows, rows_schema,
 };
 
+fn build_row_with_nullable_field(key: &str, field_0: Option<f64>, ts_millis: i64) -> Row {
+    Row {
+        values: vec![
+            Value {
+                value_data: Some(ValueData::StringValue(key.to_string())),
+            },
+            Value {
+                value_data: field_0.map(ValueData::F64Value),
+            },
+            Value {
+                value_data: Some(ValueData::TimestampMillisecondValue(ts_millis)),
+            },
+        ],
+    }
+}
+
 #[tokio::test]
 async fn test_scan_without_filtering_deleted() {
     test_scan_without_filtering_deleted_with_format(false).await;
@@ -36,7 +54,7 @@ async fn test_scan_without_filtering_deleted_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -121,3 +139,84 @@ async fn test_scan_without_filtering_deleted_with_format(flat_format: bool) {
 +-------+---------+---------------------+";
     assert_eq!(expected, batches.pretty_print().unwrap());
 }
+
+#[tokio::test]
+async fn test_filter_field_value_after_last_row_update() {
+    test_filter_field_value_after_last_row_update_with_format(false).await;
+    test_filter_field_value_after_last_row_update_with_format(true).await;
+}
+
+async fn test_filter_field_value_after_last_row_update_with_format(flat_format: bool) {
+    common_telemetry::init_default_ut_logging();
+
+    let mut env = TestEnv::new().await;
+    let engine = env
+        .create_engine(MitoConfig {
+            default_flat_format: flat_format,
+            ..Default::default()
+        })
+        .await;
+
+    let region_id = RegionId::new(1, 1);
+
+    env.get_schema_metadata_manager()
+        .register_region_table_info(
+            region_id.table_id(),
+            "test_table",
+            "test_catalog",
+            "test_schema",
+            None,
+            env.get_kv_backend(),
+        )
+        .await;
+
+    let request = CreateRequestBuilder::new().build();
+    let column_schemas = rows_schema(&request);
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    put_rows(
+        &engine,
+        region_id,
+        Rows {
+            schema: column_schemas.clone(),
+            rows: vec![build_row_with_nullable_field("a", Some(10.0), 0)],
+        },
+    )
+    .await;
+    flush_region(&engine, region_id, None).await;
+
+    put_rows(
+        &engine,
+        region_id,
+        Rows {
+            schema: column_schemas,
+            rows: vec![build_row_with_nullable_field("a", Some(20.0), 0)],
+        },
+    )
+    .await;
+    flush_region(&engine, region_id, None).await;
+
+    // We skip field filters under merge mode because the flushed field values may be stale before
+    // the last-row update is merged.
+    let stream = engine
+        .scan_to_stream(
+            region_id,
+            ScanRequest {
+                filters: vec![col("field_0").eq(lit(10.0))],
+                ..Default::default()
+            },
+        )
+        .await
+        .unwrap();
+    let batches = RecordBatches::try_collect(stream).await.unwrap();
+    let expected = "\
++-------+---------+---------------------+
+| tag_0 | field_0 | ts                  |
++-------+---------+---------------------+
+| a     | 20.0    | 1970-01-01T00:00:00 |
++-------+---------+---------------------+";
+    assert_eq!(expected, batches.pretty_print().unwrap());
+}
diff --git a/src/mito2/src/engine/flush_test.rs b/src/mito2/src/engine/flush_test.rs
index 78bae2b461..b86e75c72a 100644
--- a/src/mito2/src/engine/flush_test.rs
+++ b/src/mito2/src/engine/flush_test.rs
@@ -49,7 +49,7 @@ async fn test_manual_flush_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -112,7 +112,7 @@ async fn test_flush_engine_with_format(flat_format: bool) {
     let engine = env
         .create_engine_with(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             Some(write_buffer_manager.clone()),
@@ -191,7 +191,7 @@ async fn test_write_stall_with_format(flat_format: bool) {
     let engine = env
         .create_engine_with(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             Some(write_buffer_manager.clone()),
@@ -274,7 +274,7 @@ async fn test_flush_empty_with_format(flat_format: bool) {
     let engine = env
         .create_engine_with(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             Some(write_buffer_manager.clone()),
@@ -447,7 +447,7 @@ async fn test_auto_flush_engine_with_format(flat_format: bool) {
         .create_engine_with_time(
             MitoConfig {
                 auto_flush_interval: Duration::from_secs(60 * 5),
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             Some(write_buffer_manager.clone()),
@@ -523,7 +523,7 @@ async fn test_flush_workers_with_format(flat_format: bool) {
         .create_engine_with(
             MitoConfig {
                 num_workers: 2,
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             Some(write_buffer_manager.clone()),
diff --git a/src/mito2/src/engine/merge_mode_test.rs b/src/mito2/src/engine/merge_mode_test.rs
index 097d5e2b91..40a87642ae 100644
--- a/src/mito2/src/engine/merge_mode_test.rs
+++ b/src/mito2/src/engine/merge_mode_test.rs
@@ -39,7 +39,7 @@ async fn test_merge_mode_write_query_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -107,7 +107,7 @@ async fn test_merge_mode_compaction_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -220,7 +220,7 @@ async fn test_merge_mode_compaction_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
diff --git a/src/mito2/src/engine/open_test.rs b/src/mito2/src/engine/open_test.rs
index 5ee25fb9ff..28ad1de71e 100644
--- a/src/mito2/src/engine/open_test.rs
+++ b/src/mito2/src/engine/open_test.rs
@@ -48,7 +48,7 @@ async fn test_engine_open_empty_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("open-empty").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -87,7 +87,7 @@ async fn test_engine_open_existing_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("open-exiting").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -126,7 +126,7 @@ async fn test_engine_reopen_region_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("reopen-region").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -153,7 +153,7 @@ async fn test_engine_open_readonly_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -207,7 +207,7 @@ async fn test_engine_region_open_with_options_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -260,7 +260,7 @@ async fn test_engine_region_open_with_custom_store_with_format(flat_format: bool
     let engine = env
         .create_engine_with_multiple_object_stores(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             None,
@@ -332,7 +332,7 @@ async fn test_open_region_skip_wal_replay_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -376,7 +376,7 @@ async fn test_open_region_skip_wal_replay_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
@@ -415,7 +415,7 @@ async fn test_open_region_skip_wal_replay_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
@@ -462,7 +462,7 @@ async fn test_open_region_wait_for_opening_region_ok_with_format(flat_format: bo
     let mut env = TestEnv::with_prefix("wait-for-opening-region-ok").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -513,7 +513,7 @@ async fn test_open_region_wait_for_opening_region_err_with_format(flat_format: b
     let mut env = TestEnv::with_prefix("wait-for-opening-region-err").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -569,7 +569,7 @@ async fn test_open_compaction_region() {
 async fn test_open_compaction_region_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let mut mito_config = MitoConfig {
-        default_experimental_flat_format: flat_format,
+        default_flat_format: flat_format,
         ..Default::default()
     };
     mito_config
diff --git a/src/mito2/src/engine/parallel_test.rs b/src/mito2/src/engine/parallel_test.rs
index cf5b6491a7..b88a60739b 100644
--- a/src/mito2/src/engine/parallel_test.rs
+++ b/src/mito2/src/engine/parallel_test.rs
@@ -33,13 +33,11 @@ async fn scan_in_parallel(
     region_id: RegionId,
     table_dir: &str,
     parallelism: usize,
-    channel_size: usize,
     flat_format: bool,
 ) {
     let engine = env
         .open_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
-            parallel_scan_channel_size: channel_size,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -85,7 +83,7 @@ async fn test_parallel_scan_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -146,15 +144,13 @@ async fn test_parallel_scan_with_format(flat_format: bool) {
 
     engine.stop().await.unwrap();
 
-    scan_in_parallel(&mut env, region_id, &table_dir, 0, 1, flat_format).await;
+    scan_in_parallel(&mut env, region_id, &table_dir, 0, flat_format).await;
 
-    scan_in_parallel(&mut env, region_id, &table_dir, 1, 1, flat_format).await;
+    scan_in_parallel(&mut env, region_id, &table_dir, 1, flat_format).await;
 
-    scan_in_parallel(&mut env, region_id, &table_dir, 2, 1, flat_format).await;
+    scan_in_parallel(&mut env, region_id, &table_dir, 2, flat_format).await;
 
-    scan_in_parallel(&mut env, region_id, &table_dir, 2, 8, flat_format).await;
+    scan_in_parallel(&mut env, region_id, &table_dir, 4, flat_format).await;
 
-    scan_in_parallel(&mut env, region_id, &table_dir, 4, 8, flat_format).await;
-
-    scan_in_parallel(&mut env, region_id, &table_dir, 8, 2, flat_format).await;
+    scan_in_parallel(&mut env, region_id, &table_dir, 8, flat_format).await;
 }
diff --git a/src/mito2/src/engine/partition_filter_test.rs b/src/mito2/src/engine/partition_filter_test.rs
index fdea7d547f..33f1d93aee 100644
--- a/src/mito2/src/engine/partition_filter_test.rs
+++ b/src/mito2/src/engine/partition_filter_test.rs
@@ -58,7 +58,7 @@ async fn test_partition_filter_basic_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -89,9 +89,7 @@ async fn test_partition_filter_basic_with_format(flat_format: bool) {
     engine
         .handle_request(
             region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -122,16 +120,15 @@ async fn test_partition_filter_basic_with_format(flat_format: bool) {
     engine
         .handle_request(
             region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
 
     // Scan data in staging mode - should only see initial 5 rows (staging SST not visible)
+    let projection_input = Some(vec![1].into());
     let request = ScanRequest {
-        projection: Some(vec![1]),
+        projection_input,
         ..Default::default()
     };
     let scanner = engine.scanner(region_id, request).await.unwrap();
@@ -153,8 +150,9 @@ async fn test_partition_filter_basic_with_format(flat_format: bool) {
     // Scan after exiting staging - the old SST (tag_0 = "0".."4") should have
     // rows filtered by partition expr (tag_0 >= "5"), which means none of them pass.
     // But the staging SST (tag_0 = "5".."10") satisfies the partition expr.
+    let projection_input = Some(vec![1].into());
     let request = ScanRequest {
-        projection: Some(vec![1]),
+        projection_input,
         ..Default::default()
     };
     let scanner = engine.scanner(region_id, request).await.unwrap();
diff --git a/src/mito2/src/engine/projection_test.rs b/src/mito2/src/engine/projection_test.rs
index 7726005b0b..3bc449d07b 100644
--- a/src/mito2/src/engine/projection_test.rs
+++ b/src/mito2/src/engine/projection_test.rs
@@ -84,7 +84,7 @@ async fn test_scan_projection_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -106,8 +106,9 @@ async fn test_scan_projection_with_format(flat_format: bool) {
     put_rows(&engine, region_id, rows).await;
 
     // Scans tag_1, field_1, ts
+    let projection_input = Some(vec![1, 3, 4].into());
     let request = ScanRequest {
-        projection: Some(vec![1, 3, 4]),
+        projection_input,
         filters: Vec::new(),
         ..Default::default()
     };
@@ -141,7 +142,7 @@ async fn test_scan_projection_without_primary_key_with_format(flat_format: bool)
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -183,8 +184,9 @@ async fn test_scan_projection_without_primary_key_with_format(flat_format: bool)
     put_rows(&engine, region_id, rows).await;
 
     // Scan with projection on field_0 and field_1, filter ts >= 2s
+    let projection_input = Some(vec![0, 1].into());
     let request = ScanRequest {
-        projection: Some(vec![0, 1]), // field_0 and field_1 (not ts)
+        projection_input, // field_0 and field_1 (not ts)
         filters: vec![col("ts").gt_eq(lit(ScalarValue::TimestampMillisecond(Some(2000), None)))],
         ..Default::default()
     };
diff --git a/src/mito2/src/engine/prune_test.rs b/src/mito2/src/engine/prune_test.rs
index beb5e2644a..f217151034 100644
--- a/src/mito2/src/engine/prune_test.rs
+++ b/src/mito2/src/engine/prune_test.rs
@@ -32,7 +32,7 @@ async fn check_prune_row_groups(exprs: Vec<Expr>, expected: &str, flat_format: b
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -138,7 +138,6 @@ async fn test_prune_tag_and_field() {
 
 async fn test_prune_tag_and_field_with_format(flat_format: bool) {
     common_telemetry::init_default_ut_logging();
-    // prune result: only row group 1
     check_prune_row_groups(
         vec![
             col("tag_0").gt(lit(ScalarValue::Utf8(Some("4".to_string())))),
@@ -180,7 +179,7 @@ async fn test_prune_memtable_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -264,7 +263,7 @@ async fn test_prune_memtable_complex_expr_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -327,7 +326,7 @@ async fn test_mem_range_prune_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -392,7 +391,7 @@ async fn test_scan_filter_field_after_delete_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -443,7 +442,10 @@ async fn test_scan_filter_field_after_delete_with_format(flat_format: bool) {
     )
     .await;
 
-    // Scans and filter fields, the field should be deleted.
+    // Scans and filters by a field value. The mito reader skips field filters under
+    // `PreFilterMode::SkipFields` (DataFusion re-applies them above the engine), so
+    // the returned batches still contain all non-deleted rows — the reader's job here
+    // is only to ensure the delete op is honored.
     let request = ScanRequest {
         filters: vec![col("field_0").eq(lit(3.0f64))],
         ..Default::default()
@@ -454,10 +456,12 @@ async fn test_scan_filter_field_after_delete_with_format(flat_format: bool) {
         .unwrap();
     let batches = RecordBatches::try_collect(stream).await.unwrap();
     let expected = "\
-+-------+---------+----+
-| tag_0 | field_0 | ts |
-+-------+---------+----+
-+-------+---------+----+";
++-------+---------+---------------------+
+| tag_0 | field_0 | ts                  |
++-------+---------+---------------------+
+| 1     | 1.0     | 1970-01-01T00:00:01 |
+| 4     | 4.0     | 1970-01-01T00:00:04 |
++-------+---------+---------------------+";
     assert_eq!(
         expected,
         sort_batches_and_print(&batches, &["tag_0", "field_0", "ts"])
diff --git a/src/mito2/src/engine/remap_manifests_test.rs b/src/mito2/src/engine/remap_manifests_test.rs
index e3538401aa..467d14a1ef 100644
--- a/src/mito2/src/engine/remap_manifests_test.rs
+++ b/src/mito2/src/engine/remap_manifests_test.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 
 use api::v1::Rows;
 use datatypes::value::Value;
@@ -37,7 +37,7 @@ async fn test_remap_manifests_invalid_partition_expr_with_format(flat_format: bo
     let mut env = TestEnv::with_prefix("invalid-partition-expr").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -83,7 +83,7 @@ async fn test_remap_manifests_invalid_region_state_with_format(flat_format: bool
     let mut env = TestEnv::with_prefix("invalid-region-state").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -123,7 +123,7 @@ async fn test_remap_manifests_invalid_input_regions_with_format(flat_format: boo
     let mut env = TestEnv::with_prefix("invalid-input-regions").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -166,7 +166,7 @@ async fn test_remap_manifests_success_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("engine-stop").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -194,9 +194,7 @@ async fn test_remap_manifests_success_with_format(flat_format: bool) {
         engine
             .handle_request(
                 region_id,
-                RegionRequest::Flush(RegionFlushRequest {
-                    row_group_size: None,
-                }),
+                RegionRequest::Flush(RegionFlushRequest::default()),
             )
             .await
             .unwrap();
diff --git a/src/mito2/src/engine/row_selector_test.rs b/src/mito2/src/engine/row_selector_test.rs
index 317ede5a97..26d7327c2f 100644
--- a/src/mito2/src/engine/row_selector_test.rs
+++ b/src/mito2/src/engine/row_selector_test.rs
@@ -13,18 +13,21 @@
 // limitations under the License.
 
 use api::v1::Rows;
+use common_base::readable_size::ReadableSize;
 use common_recordbatch::RecordBatches;
+use datafusion_expr::{col, lit};
 use store_api::region_engine::RegionEngine;
 use store_api::region_request::RegionRequest;
 use store_api::storage::{RegionId, ScanRequest, TimeSeriesRowSelector};
 
 use crate::config::MitoConfig;
+use crate::engine::MitoEngine;
 use crate::test_util::batch_util::sort_batches_and_print;
 use crate::test_util::{
     CreateRequestBuilder, TestEnv, build_rows_for_key, flush_region, put_rows, rows_schema,
 };
 
-async fn test_last_row(append_mode: bool) {
+async fn test_last_row(append_mode: bool, flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env.create_engine(MitoConfig::default()).await;
     let region_id = RegionId::new(1, 1);
@@ -39,9 +42,12 @@ async fn test_last_row(append_mode: bool) {
             env.get_kv_backend(),
         )
         .await;
-    let request = CreateRequestBuilder::new()
-        .insert_option("append_mode", &append_mode.to_string())
-        .build();
+    let mut request_builder =
+        CreateRequestBuilder::new().insert_option("append_mode", &append_mode.to_string());
+    if flat_format {
+        request_builder = request_builder.insert_option("sst_format", "flat");
+    }
+    let request = request_builder.build();
     let column_schemas = rows_schema(&request);
     engine
         .handle_request(region_id, RegionRequest::Create(request))
@@ -104,12 +110,109 @@ async fn test_last_row(append_mode: bool) {
     assert_eq!(expected, sort_batches_and_print(&batches, &["tag_0", "ts"]));
 }
 
+async fn scan_last_row(
+    engine: &MitoEngine,
+    region_id: RegionId,
+    filters: Vec<datafusion_expr::Expr>,
+) -> String {
+    let scanner = engine
+        .scanner(
+            region_id,
+            ScanRequest {
+                filters,
+                series_row_selector: Some(TimeSeriesRowSelector::LastRow),
+                ..Default::default()
+            },
+        )
+        .await
+        .unwrap();
+    let stream = scanner.scan().await.unwrap();
+    let batches = RecordBatches::try_collect(stream).await.unwrap();
+    sort_batches_and_print(&batches, &["tag_0", "ts"])
+}
+
 #[tokio::test]
 async fn test_last_row_append_mode_disabled() {
-    test_last_row(false).await;
+    test_last_row(false, false).await;
 }
 
 #[tokio::test]
 async fn test_last_row_append_mode_enabled() {
-    test_last_row(true).await;
+    test_last_row(true, false).await;
+}
+
+#[tokio::test]
+async fn test_last_row_flat_format_append_mode_disabled() {
+    test_last_row(false, true).await;
+}
+
+#[tokio::test]
+async fn test_last_row_flat_format_append_mode_enabled() {
+    test_last_row(true, true).await;
+}
+
+#[tokio::test]
+async fn test_last_row_flat_format_prefilter_does_not_poison_selector_cache() {
+    let mut env = TestEnv::new().await;
+    let engine = env
+        .create_engine(MitoConfig {
+            selector_result_cache_size: ReadableSize::mb(1),
+            ..Default::default()
+        })
+        .await;
+    let region_id = RegionId::new(1, 1);
+
+    env.get_schema_metadata_manager()
+        .register_region_table_info(
+            region_id.table_id(),
+            "test_table",
+            "test_catalog",
+            "test_schema",
+            None,
+            env.get_kv_backend(),
+        )
+        .await;
+
+    let request = CreateRequestBuilder::new()
+        .insert_option("sst_format", "flat")
+        .build();
+    let column_schemas = rows_schema(&request);
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    let rows = Rows {
+        schema: column_schemas,
+        rows: [
+            build_rows_for_key("a", 0, 3, 0),
+            build_rows_for_key("b", 0, 3, 10),
+        ]
+        .concat(),
+    };
+    put_rows(&engine, region_id, rows).await;
+    flush_region(&engine, region_id, Some(16)).await;
+
+    let filtered = scan_last_row(&engine, region_id, vec![col("tag_0").eq(lit("a"))]).await;
+    assert_eq!(
+        "\
++-------+---------+---------------------+
+| tag_0 | field_0 | ts                  |
++-------+---------+---------------------+
+| a     | 2.0     | 1970-01-01T00:00:02 |
++-------+---------+---------------------+",
+        filtered
+    );
+
+    let unfiltered = scan_last_row(&engine, region_id, vec![]).await;
+    assert_eq!(
+        "\
++-------+---------+---------------------+
+| tag_0 | field_0 | ts                  |
++-------+---------+---------------------+
+| a     | 2.0     | 1970-01-01T00:00:02 |
+| b     | 12.0    | 1970-01-01T00:00:02 |
++-------+---------+---------------------+",
+        unfiltered
+    );
 }
diff --git a/src/mito2/src/engine/scan_test.rs b/src/mito2/src/engine/scan_test.rs
index 46f4cc6cf2..d4d48b9fe6 100644
--- a/src/mito2/src/engine/scan_test.rs
+++ b/src/mito2/src/engine/scan_test.rs
@@ -12,32 +12,332 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+use std::collections::BTreeMap;
+
 use api::v1::Rows;
 use common_error::ext::ErrorExt;
 use common_error::status_code::StatusCode;
 use common_recordbatch::RecordBatches;
 use datafusion::physical_plan::metrics::ExecutionPlanMetricsSet;
+use datatypes::arrow::array::AsArray;
+use datatypes::arrow::datatypes::{Float64Type, TimestampMillisecondType};
 use futures::TryStreamExt;
 use store_api::region_engine::{PrepareRequest, RegionEngine, RegionScanner};
 use store_api::region_request::RegionRequest;
 use store_api::storage::{RegionId, ScanRequest, TimeSeriesDistribution};
 
 use crate::config::MitoConfig;
+use crate::error::Error;
 use crate::read::scan_region::Scanner;
 use crate::test_util;
 use crate::test_util::{CreateRequestBuilder, TestEnv};
 
+#[tokio::test]
+async fn test_incremental_query_stale_error() {
+    let mut env = TestEnv::with_prefix("test_incremental_query_stale_error").await;
+    let engine = env.create_engine(MitoConfig::default()).await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+    let column_schemas = test_util::rows_schema(&request);
+
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    let rows = Rows {
+        schema: column_schemas.clone(),
+        rows: test_util::build_rows(0, 3),
+    };
+    test_util::put_rows(&engine, region_id, rows).await;
+    test_util::flush_region(&engine, region_id, None).await;
+
+    let err = engine
+        .scanner(
+            region_id,
+            ScanRequest {
+                memtable_min_sequence: Some(0),
+                ..Default::default()
+            },
+        )
+        .await
+        .err()
+        .expect("expect stale incremental error");
+
+    let min_readable_seq = match &err {
+        Error::IncrementalQueryStale {
+            region_id: err_region_id,
+            given_seq,
+            min_readable_seq,
+            ..
+        } => {
+            assert_eq!(*err_region_id, region_id);
+            assert_eq!(*given_seq, 0);
+            assert!(*min_readable_seq > 0);
+            *min_readable_seq
+        }
+        _ => panic!("unexpected err: {err}"),
+    };
+    assert_eq!(StatusCode::RequestOutdated, err.status_code());
+    let err_msg = err.to_string();
+    assert!(err_msg.contains("STALE_CURSOR"));
+    assert!(err_msg.contains(&region_id.to_string()));
+    assert!(err_msg.contains("given_seq: 0"));
+    assert!(err_msg.contains(&format!("min_readable_seq: {min_readable_seq}")));
+
+    let incremental_rows = Rows {
+        schema: column_schemas,
+        rows: test_util::build_rows(3, 5),
+    };
+    test_util::put_rows(&engine, region_id, incremental_rows).await;
+
+    let scanner = engine
+        .scanner(
+            region_id,
+            ScanRequest {
+                memtable_min_sequence: Some(min_readable_seq),
+                sst_min_sequence: Some(u64::MAX),
+                ..Default::default()
+            },
+        )
+        .await
+        .unwrap();
+    let stream = scanner.scan().await.unwrap();
+    let batches = RecordBatches::try_collect(stream).await.unwrap();
+    assert_eq!(
+        batches.pretty_print().unwrap(),
+        "\
++-------+---------+---------------------+
+| tag_0 | field_0 | ts                  |
++-------+---------+---------------------+
+| 3     | 3.0     | 1970-01-01T00:00:03 |
+| 4     | 4.0     | 1970-01-01T00:00:04 |
++-------+---------+---------------------+"
+    );
+}
+
 #[tokio::test]
 async fn test_scan_with_min_sst_sequence() {
     test_scan_with_min_sst_sequence_with_format(false).await;
     test_scan_with_min_sst_sequence_with_format(true).await;
 }
 
+#[tokio::test]
+async fn test_full_snapshot_upper_bound_does_not_constrain_sst_rows() {
+    let mut env =
+        TestEnv::with_prefix("test_full_snapshot_upper_bound_does_not_constrain_sst_rows").await;
+    let engine = env.create_engine(MitoConfig::default()).await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+    let column_schemas = test_util::rows_schema(&request);
+
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    let first_rows = Rows {
+        schema: column_schemas.clone(),
+        rows: test_util::build_rows(0, 3),
+    };
+    test_util::put_rows(&engine, region_id, first_rows).await;
+    test_util::flush_region(&engine, region_id, None).await;
+
+    let snapshot_upper_bound = engine.get_committed_sequence(region_id).await.unwrap();
+
+    let second_rows = Rows {
+        schema: column_schemas,
+        rows: test_util::build_rows(3, 5),
+    };
+    test_util::put_rows(&engine, region_id, second_rows).await;
+    test_util::flush_region(&engine, region_id, None).await;
+
+    let scanner = engine
+        .scanner(
+            region_id,
+            ScanRequest {
+                memtable_max_sequence: Some(snapshot_upper_bound),
+                ..Default::default()
+            },
+        )
+        .await
+        .unwrap();
+
+    let stream = scanner.scan().await.unwrap();
+    let batches = RecordBatches::try_collect(stream).await.unwrap();
+    let pretty = batches.pretty_print().unwrap();
+
+    assert!(pretty.contains("1970-01-01T00:00:03"));
+    assert!(pretty.contains("1970-01-01T00:00:04"));
+}
+
+#[tokio::test]
+async fn test_snapshot_bound_query_binds_memtable_upper_bound_at_scan_open() {
+    let mut env =
+        TestEnv::with_prefix("test_snapshot_bound_query_binds_memtable_upper_bound_at_scan_open")
+            .await;
+    let engine = env.create_engine(MitoConfig::default()).await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+    let column_schemas = test_util::rows_schema(&request);
+
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    let first_rows = Rows {
+        schema: column_schemas.clone(),
+        rows: test_util::build_rows(0, 3),
+    };
+    test_util::put_rows(&engine, region_id, first_rows).await;
+
+    let expected_snapshot = engine.get_committed_sequence(region_id).await.unwrap();
+    let scanner = engine
+        .scanner(
+            region_id,
+            ScanRequest {
+                snapshot_on_scan: true,
+                ..Default::default()
+            },
+        )
+        .await
+        .unwrap();
+    assert_eq!(scanner.snapshot_sequence(), Some(expected_snapshot));
+
+    let second_rows = Rows {
+        schema: column_schemas,
+        rows: test_util::build_rows(3, 5),
+    };
+    test_util::put_rows(&engine, region_id, second_rows).await;
+
+    let stream = scanner.scan().await.unwrap();
+    let batches = RecordBatches::try_collect(stream).await.unwrap();
+    assert_eq!(
+        batches.pretty_print().unwrap(),
+        "\
++-------+---------+---------------------+
+| tag_0 | field_0 | ts                  |
++-------+---------+---------------------+
+| 0     | 0.0     | 1970-01-01T00:00:00 |
+| 1     | 1.0     | 1970-01-01T00:00:01 |
+| 2     | 2.0     | 1970-01-01T00:00:02 |
++-------+---------+---------------------+"
+    );
+}
+
+#[tokio::test]
+async fn test_snapshot_bound_query_keeps_open_snapshot_after_late_flush() {
+    let mut env =
+        TestEnv::with_prefix("test_snapshot_bound_query_keeps_open_snapshot_after_late_flush")
+            .await;
+    let engine = env.create_engine(MitoConfig::default()).await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+    let column_schemas = test_util::rows_schema(&request);
+
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    let first_rows = Rows {
+        schema: column_schemas.clone(),
+        rows: test_util::build_rows(0, 3),
+    };
+    test_util::put_rows(&engine, region_id, first_rows).await;
+
+    let expected_snapshot = engine.get_committed_sequence(region_id).await.unwrap();
+    let scanner = engine
+        .scanner(
+            region_id,
+            ScanRequest {
+                snapshot_on_scan: true,
+                ..Default::default()
+            },
+        )
+        .await
+        .unwrap();
+    assert_eq!(scanner.snapshot_sequence(), Some(expected_snapshot));
+
+    let second_rows = Rows {
+        schema: column_schemas,
+        rows: test_util::build_rows(3, 5),
+    };
+    test_util::put_rows(&engine, region_id, second_rows).await;
+    test_util::flush_region(&engine, region_id, None).await;
+
+    assert_eq!(scanner.snapshot_sequence(), Some(expected_snapshot));
+}
+
+#[tokio::test]
+async fn test_snapshot_bound_query_keeps_correct_result_after_late_flush() {
+    let mut env =
+        TestEnv::with_prefix("test_snapshot_bound_query_keeps_correct_result_after_late_flush")
+            .await;
+    let engine = env.create_engine(MitoConfig::default()).await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+    let column_schemas = test_util::rows_schema(&request);
+
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    let first_rows = Rows {
+        schema: column_schemas.clone(),
+        rows: test_util::build_rows(0, 3),
+    };
+    test_util::put_rows(&engine, region_id, first_rows).await;
+
+    let expected_snapshot = engine.get_committed_sequence(region_id).await.unwrap();
+    let scanner = engine
+        .scanner(
+            region_id,
+            ScanRequest {
+                snapshot_on_scan: true,
+                ..Default::default()
+            },
+        )
+        .await
+        .unwrap();
+    assert_eq!(scanner.snapshot_sequence(), Some(expected_snapshot));
+
+    let second_rows = Rows {
+        schema: column_schemas,
+        rows: test_util::build_rows(3, 5),
+    };
+    test_util::put_rows(&engine, region_id, second_rows).await;
+    test_util::flush_region(&engine, region_id, None).await;
+
+    assert_eq!(scanner.snapshot_sequence(), Some(expected_snapshot));
+
+    let stream = scanner.scan().await.unwrap();
+    let batches = RecordBatches::try_collect(stream).await.unwrap();
+    assert_eq!(
+        batches.pretty_print().unwrap(),
+        "\
++-------+---------+---------------------+
+| tag_0 | field_0 | ts                  |
++-------+---------+---------------------+
+| 0     | 0.0     | 1970-01-01T00:00:00 |
+| 1     | 1.0     | 1970-01-01T00:00:01 |
+| 2     | 2.0     | 1970-01-01T00:00:02 |
++-------+---------+---------------------+"
+    );
+}
+
 async fn test_scan_with_min_sst_sequence_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("test_scan_with_min_sst_sequence").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -172,7 +472,7 @@ async fn test_max_concurrent_scan_files() {
 async fn test_max_concurrent_scan_files_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("test_max_concurrent_scan_files").await;
     let config = MitoConfig {
-        default_experimental_flat_format: flat_format,
+        default_flat_format: flat_format,
         max_concurrent_scan_files: 2,
         ..Default::default()
     };
@@ -222,11 +522,16 @@ async fn test_max_concurrent_scan_files_with_format(flat_format: bool) {
 }
 
 #[tokio::test]
-async fn test_series_scan_primarykey() {
+async fn test_series_scan() {
+    test_series_scan_with_format(false).await;
+    test_series_scan_with_format(true).await;
+}
+
+async fn test_series_scan_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("test_series_scan").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: false,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -295,10 +600,27 @@ async fn test_series_scan_primarykey() {
         })
         .unwrap();
 
+    let actual_rows = collect_partition_rows_round_robin(&scanner, 3).await;
+
+    let mut expected_rows = Vec::new();
+    for value in [0_i64, 1, 2, 3, 4, 5, 3600, 3601, 3602, 7200, 7201, 7202] {
+        expected_rows.push((value.to_string(), value as f64, value * 1000));
+    }
+    expected_rows.sort_by(|a, b| a.0.cmp(&b.0).then(a.2.cmp(&b.2)));
+
+    assert_eq!(expected_rows, actual_rows);
+}
+
+/// Scans all partitions in round-robin fashion and returns rows sorted by (tag, ts).
+/// Also asserts that each series appears in only one partition.
+async fn collect_partition_rows_round_robin(
+    scanner: &dyn RegionScanner,
+    num_partitions: usize,
+) -> Vec<(String, f64, i64)> {
     let metrics_set = ExecutionPlanMetricsSet::default();
 
-    let mut partition_batches = vec![vec![]; 3];
-    let mut streams: Vec<_> = (0..3)
+    let mut partition_batches = vec![vec![]; num_partitions];
+    let mut streams: Vec<_> = (0..num_partitions)
         .map(|partition| {
             let stream = scanner
                 .scan_partition(&Default::default(), &metrics_set, partition)
@@ -309,11 +631,11 @@ async fn test_series_scan_primarykey() {
     let mut num_done = 0;
     let mut schema = None;
     // Pull streams in round-robin fashion to get the consistent output from the sender.
-    while num_done < 3 {
+    while num_done < num_partitions {
         if schema.is_none() {
             schema = Some(streams[0].as_ref().unwrap().schema().clone());
         }
-        for i in 0..3 {
+        for i in 0..num_partitions {
             let Some(mut stream) = streams[i].take() else {
                 continue;
             };
@@ -326,53 +648,66 @@ async fn test_series_scan_primarykey() {
         }
     }
 
-    let mut check_result = |expected| {
-        let batches =
-            RecordBatches::try_new(schema.clone().unwrap(), partition_batches.remove(0)).unwrap();
-        assert_eq!(expected, batches.pretty_print().unwrap());
-    };
-
-    // Output series order is 0, 1, 2, 3, 3600, 3601, 3602, 4, 5, 7200, 7201, 7202
-    let expected = "\
-+-------+---------+---------------------+
-| tag_0 | field_0 | ts                  |
-+-------+---------+---------------------+
-| 0     | 0.0     | 1970-01-01T00:00:00 |
-| 3     | 3.0     | 1970-01-01T00:00:03 |
-| 3602  | 3602.0  | 1970-01-01T01:00:02 |
-| 7200  | 7200.0  | 1970-01-01T02:00:00 |
-+-------+---------+---------------------+";
-    check_result(expected);
-
-    let expected = "\
-+-------+---------+---------------------+
-| tag_0 | field_0 | ts                  |
-+-------+---------+---------------------+
-| 1     | 1.0     | 1970-01-01T00:00:01 |
-| 3600  | 3600.0  | 1970-01-01T01:00:00 |
-| 4     | 4.0     | 1970-01-01T00:00:04 |
-| 7201  | 7201.0  | 1970-01-01T02:00:01 |
-+-------+---------+---------------------+";
-    check_result(expected);
-
-    let expected = "\
-+-------+---------+---------------------+
-| tag_0 | field_0 | ts                  |
-+-------+---------+---------------------+
-| 2     | 2.0     | 1970-01-01T00:00:02 |
-| 3601  | 3601.0  | 1970-01-01T01:00:01 |
-| 5     | 5.0     | 1970-01-01T00:00:05 |
-| 7202  | 7202.0  | 1970-01-01T02:00:02 |
-+-------+---------+---------------------+";
-    check_result(expected);
+    let schema = schema.unwrap();
+    collect_and_assert_partition_rows(schema, partition_batches)
 }
 
+/// Collects rows sorted by (tag, ts) from partition batches.
+/// Also asserts that each series appears in only one partition.
+fn collect_and_assert_partition_rows(
+    schema: datatypes::schema::SchemaRef,
+    partition_batches: Vec<Vec<common_recordbatch::RecordBatch>>,
+) -> Vec<(String, f64, i64)> {
+    let mut series_to_partition = BTreeMap::new();
+    let mut actual_rows = Vec::new();
+
+    for (partition, batches) in partition_batches.into_iter().enumerate() {
+        let batches = RecordBatches::try_new(schema.clone(), batches).unwrap();
+        let mut partition_series = Vec::new();
+
+        for batch in batches.iter() {
+            let tags = batch.column_by_name("tag_0").unwrap().as_string::<i32>();
+            let fields = batch
+                .column_by_name("field_0")
+                .unwrap()
+                .as_primitive::<Float64Type>();
+            let ts = batch
+                .column_by_name("ts")
+                .unwrap()
+                .as_primitive::<TimestampMillisecondType>();
+
+            for row in 0..batch.num_rows() {
+                let tag = tags.value(row).to_string();
+                let field = fields.value(row);
+                let ts = ts.value(row);
+                partition_series.push(tag.clone());
+                actual_rows.push((tag, field, ts));
+            }
+        }
+
+        partition_series.sort();
+        partition_series.dedup();
+        for tag in partition_series {
+            let prev = series_to_partition.insert(tag.clone(), partition);
+            assert_eq!(
+                None, prev,
+                "series {tag} appears in multiple partitions: {prev:?} and {partition}"
+            );
+        }
+    }
+
+    actual_rows.sort_by(|a, b| a.0.cmp(&b.0).then(a.2.cmp(&b.2)));
+    actual_rows
+}
+
+/// Tests series scan with multiple partition ranges (each with multiple overlapping sources)
+/// and small semaphore permits (controlled by num_partitions).
 #[tokio::test]
-async fn test_series_scan_flat() {
-    let mut env = TestEnv::with_prefix("test_series_scan").await;
+async fn test_series_scan_flat_small_permits() {
+    let mut env = TestEnv::with_prefix("test_series_scan_small_permits").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: true,
+            default_flat_format: true,
             ..Default::default()
         })
         .await;
@@ -389,6 +724,7 @@ async fn test_series_scan_flat() {
         .await
         .unwrap();
 
+    // Create overlapping SSTs in each time window so partition ranges have multiple sources.
     let put_flush_rows = async |start, end| {
         let rows = Rows {
             schema: column_schemas.clone(),
@@ -397,118 +733,65 @@ async fn test_series_scan_flat() {
         test_util::put_rows(&engine, region_id, rows).await;
         test_util::flush_region(&engine, region_id, None).await;
     };
-    // generates 3 SST files
+    // Window 0 (0s-999s): 3 overlapping SSTs
     put_flush_rows(0, 3).await;
-    put_flush_rows(2, 6).await;
+    put_flush_rows(1, 5).await;
+    put_flush_rows(3, 7).await;
+    // Window 1 (3600s-4599s): 2 overlapping SSTs
     put_flush_rows(3600, 3603).await;
-    // Put to memtable.
-    let rows = Rows {
-        schema: column_schemas.clone(),
-        rows: test_util::build_rows(7200, 7203),
-    };
-    test_util::put_rows(&engine, region_id, rows).await;
+    put_flush_rows(3601, 3605).await;
+    // Window 2 (7200s-8199s): 2 overlapping SSTs
+    put_flush_rows(7200, 7203).await;
+    put_flush_rows(7201, 7204).await;
 
-    let request = ScanRequest {
-        distribution: Some(TimeSeriesDistribution::PerSeries),
-        ..Default::default()
-    };
-    let scanner = engine.scanner(region_id, request).await.unwrap();
-    let Scanner::Series(mut scanner) = scanner else {
-        panic!("Scanner should be series scan");
-    };
-    // 3 partition ranges for 3 time window.
-    assert_eq!(
-        3,
-        scanner.properties().partitions[0].len(),
-        "unexpected ranges: {:?}",
-        scanner.properties().partitions
-    );
-    let raw_ranges: Vec<_> = scanner
-        .properties()
-        .partitions
-        .iter()
-        .flatten()
-        .cloned()
-        .collect();
-    let mut new_ranges = Vec::with_capacity(3);
-    for range in raw_ranges {
-        new_ranges.push(vec![range]);
+    let mut expected_rows = Vec::new();
+    for value in [
+        0_i64, 1, 2, 3, 4, 5, 6, 3600, 3601, 3602, 3603, 3604, 7200, 7201, 7202, 7203,
+    ] {
+        expected_rows.push((value.to_string(), value as f64, value * 1000));
     }
-    scanner
-        .prepare(PrepareRequest {
-            ranges: Some(new_ranges),
+    expected_rows.sort_by(|a, b| a.0.cmp(&b.0).then(a.2.cmp(&b.2)));
+
+    // Test with different semaphore sizes (num_partitions controls Semaphore::new(num_partitions)).
+    for num_partitions in [1, 2] {
+        let request = ScanRequest {
+            distribution: Some(TimeSeriesDistribution::PerSeries),
             ..Default::default()
-        })
-        .unwrap();
+        };
+        let scanner = engine.scanner(region_id, request).await.unwrap();
+        let Scanner::Series(mut scanner) = scanner else {
+            panic!("Scanner should be series scan");
+        };
 
-    let metrics_set = ExecutionPlanMetricsSet::default();
+        // Collect all partition ranges and redistribute into `num_partitions` partitions.
+        let raw_ranges: Vec<_> = scanner
+            .properties()
+            .partitions
+            .iter()
+            .flatten()
+            .cloned()
+            .collect();
+        assert!(
+            raw_ranges.len() >= 3,
+            "expected at least 3 partition ranges, got {}",
+            raw_ranges.len()
+        );
 
-    let mut partition_batches = vec![vec![]; 3];
-    let mut streams: Vec<_> = (0..3)
-        .map(|partition| {
-            let stream = scanner
-                .scan_partition(&Default::default(), &metrics_set, partition)
-                .unwrap();
-            Some(stream)
-        })
-        .collect();
-    let mut num_done = 0;
-    let mut schema = None;
-    // Pull streams in round-robin fashion to get the consistent output from the sender.
-    while num_done < 3 {
-        if schema.is_none() {
-            schema = Some(streams[0].as_ref().unwrap().schema().clone());
-        }
-        for i in 0..3 {
-            let Some(mut stream) = streams[i].take() else {
-                continue;
-            };
-            let Some(rb) = stream.try_next().await.unwrap() else {
-                num_done += 1;
-                continue;
-            };
-            partition_batches[i].push(rb);
-            streams[i] = Some(stream);
+        let mut new_ranges = vec![vec![]; num_partitions];
+        for (i, range) in raw_ranges.into_iter().enumerate() {
+            new_ranges[i % num_partitions].push(range);
         }
+        scanner
+            .prepare(PrepareRequest {
+                ranges: Some(new_ranges),
+                ..Default::default()
+            })
+            .unwrap();
+
+        let actual_rows = collect_partition_rows_round_robin(&scanner, num_partitions).await;
+        assert_eq!(
+            expected_rows, actual_rows,
+            "mismatch with num_partitions={num_partitions}"
+        );
     }
-
-    let mut check_result = |expected| {
-        let batches =
-            RecordBatches::try_new(schema.clone().unwrap(), partition_batches.remove(0)).unwrap();
-        assert_eq!(expected, batches.pretty_print().unwrap());
-    };
-
-    // Output series order is 0, 1, 2, 3, 3600, 3601, 3602, 4, 5, 7200, 7201, 7202
-    let expected = "\
-+-------+---------+---------------------+
-| tag_0 | field_0 | ts                  |
-+-------+---------+---------------------+
-| 0     | 0.0     | 1970-01-01T00:00:00 |
-| 1     | 1.0     | 1970-01-01T00:00:01 |
-| 2     | 2.0     | 1970-01-01T00:00:02 |
-| 3     | 3.0     | 1970-01-01T00:00:03 |
-| 7200  | 7200.0  | 1970-01-01T02:00:00 |
-| 7201  | 7201.0  | 1970-01-01T02:00:01 |
-| 7202  | 7202.0  | 1970-01-01T02:00:02 |
-+-------+---------+---------------------+";
-    check_result(expected);
-
-    let expected = "\
-+-------+---------+---------------------+
-| tag_0 | field_0 | ts                  |
-+-------+---------+---------------------+
-| 3600  | 3600.0  | 1970-01-01T01:00:00 |
-| 3601  | 3601.0  | 1970-01-01T01:00:01 |
-| 3602  | 3602.0  | 1970-01-01T01:00:02 |
-+-------+---------+---------------------+";
-    check_result(expected);
-
-    let expected = "\
-+-------+---------+---------------------+
-| tag_0 | field_0 | ts                  |
-+-------+---------+---------------------+
-| 4     | 4.0     | 1970-01-01T00:00:04 |
-| 5     | 5.0     | 1970-01-01T00:00:05 |
-+-------+---------+---------------------+";
-    check_result(expected);
 }
diff --git a/src/mito2/src/engine/set_role_state_test.rs b/src/mito2/src/engine/set_role_state_test.rs
index fd90cd99f7..40e03b063a 100644
--- a/src/mito2/src/engine/set_role_state_test.rs
+++ b/src/mito2/src/engine/set_role_state_test.rs
@@ -19,7 +19,9 @@ use store_api::region_engine::{
     RegionEngine, RegionRole, SetRegionRoleStateResponse, SetRegionRoleStateSuccess,
     SettableRegionRoleState,
 };
-use store_api::region_request::{RegionPutRequest, RegionRequest};
+use store_api::region_request::{
+    EnterStagingRequest, RegionPutRequest, RegionRequest, StagingPartitionDirective,
+};
 use store_api::storage::RegionId;
 
 use crate::config::MitoConfig;
@@ -70,7 +72,7 @@ async fn test_set_role_state_gracefully_with_format(flat_format: bool) {
         let mut env = TestEnv::new().await;
         let engine = env
             .create_engine(MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             })
             .await;
@@ -141,7 +143,7 @@ async fn test_set_role_state_gracefully_not_exist_with_format(flat_format: bool)
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -166,7 +168,7 @@ async fn test_write_downgrading_region_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("write-to-downgrading-region").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -220,7 +222,7 @@ async fn test_unified_state_transitions_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -241,12 +243,14 @@ async fn test_unified_state_transitions_with_format(flat_format: bool) {
         .await
         .unwrap();
     assert_success_response(&result, 0);
+    assert_eq!(engine.role(region_id), Some(RegionRole::StagingLeader));
 
     let result = engine
         .set_region_role_state_gracefully(region_id, SettableRegionRoleState::Leader)
         .await
         .unwrap();
     assert_success_response(&result, 0);
+    assert_eq!(engine.role(region_id), Some(RegionRole::Leader));
 
     // Leader -> StagingLeader -> Follower (exit staging via demotion)
     engine
@@ -259,6 +263,7 @@ async fn test_unified_state_transitions_with_format(flat_format: bool) {
         .await
         .unwrap();
     assert_success_response(&result, 0);
+    assert_eq!(engine.role(region_id), Some(RegionRole::Follower));
 
     // Note: Direct Follower -> Leader promotion is no longer allowed
     // Use existing set_region_role method for follower -> leader promotion
@@ -277,6 +282,7 @@ async fn test_unified_state_transitions_with_format(flat_format: bool) {
         .await
         .unwrap();
     assert_success_response(&result, 0);
+    assert_eq!(engine.role(region_id), Some(RegionRole::DowngradingLeader));
 
     // Note: Direct DowngradingLeader -> Leader is no longer allowed
     // Use existing set_region_role method for downgrading -> leader promotion
@@ -325,11 +331,269 @@ async fn test_restricted_state_transitions() {
     test_restricted_state_transitions_with_format(true).await;
 }
 
+#[tokio::test]
+async fn test_direct_set_region_role_staging_leader_is_noop() {
+    let mut env = TestEnv::new().await;
+    let engine = env.create_engine(MitoConfig::default()).await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    engine
+        .set_region_role(region_id, RegionRole::StagingLeader)
+        .unwrap();
+
+    assert_eq!(engine.role(region_id), Some(RegionRole::Leader));
+
+    engine
+        .set_region_role(region_id, RegionRole::Follower)
+        .unwrap();
+    engine
+        .set_region_role(region_id, RegionRole::StagingLeader)
+        .unwrap();
+
+    assert_eq!(engine.role(region_id), Some(RegionRole::Follower));
+}
+
+#[tokio::test]
+async fn test_direct_set_region_role_exits_staging_state_only() {
+    let mut env = TestEnv::new().await;
+    let engine = env.create_engine(MitoConfig::default()).await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    engine
+        .handle_request(
+            region_id,
+            RegionRequest::EnterStaging(EnterStagingRequest {
+                partition_directive: StagingPartitionDirective::RejectAllWrites,
+            }),
+        )
+        .await
+        .unwrap();
+    assert_eq!(engine.role(region_id), Some(RegionRole::StagingLeader));
+    assert!(
+        engine
+            .get_region(region_id)
+            .unwrap()
+            .manifest_ctx
+            .staging_partition_info()
+            .is_some()
+    );
+
+    engine
+        .set_region_role(region_id, RegionRole::Leader)
+        .unwrap();
+    assert_eq!(engine.role(region_id), Some(RegionRole::Leader));
+    assert!(
+        engine
+            .get_region(region_id)
+            .unwrap()
+            .manifest_ctx
+            .staging_partition_info()
+            .is_none()
+    );
+}
+
+#[tokio::test]
+async fn test_set_region_role_can_exit_staging_to_leader() {
+    let mut env = TestEnv::new().await;
+    let engine = env.create_engine(MitoConfig::default()).await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    engine
+        .set_region_role_state_gracefully(region_id, SettableRegionRoleState::StagingLeader)
+        .await
+        .unwrap();
+    assert_eq!(engine.role(region_id), Some(RegionRole::StagingLeader));
+
+    engine
+        .set_region_role(region_id, RegionRole::Leader)
+        .unwrap();
+
+    assert_eq!(engine.role(region_id), Some(RegionRole::Leader));
+    assert!(
+        engine
+            .get_region(region_id)
+            .unwrap()
+            .manifest_ctx
+            .staging_partition_info()
+            .is_none()
+    );
+}
+
+#[tokio::test]
+async fn test_set_region_role_leader_clears_staging_partition_info() {
+    let mut env = TestEnv::new().await;
+    let engine = env.create_engine(MitoConfig::default()).await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    engine
+        .handle_request(
+            region_id,
+            RegionRequest::EnterStaging(EnterStagingRequest {
+                partition_directive: StagingPartitionDirective::RejectAllWrites,
+            }),
+        )
+        .await
+        .unwrap();
+
+    let region = engine.get_region(region_id).unwrap();
+    assert!(region.manifest_ctx.staging_partition_info().is_some());
+
+    engine
+        .set_region_role(region_id, RegionRole::Leader)
+        .unwrap();
+
+    let region = engine.get_region(region_id).unwrap();
+    assert_eq!(engine.role(region_id), Some(RegionRole::Leader));
+    assert!(region.manifest_ctx.staging_partition_info().is_none());
+}
+
+#[tokio::test]
+async fn test_set_region_role_follower_clears_staging_partition_info() {
+    let mut env = TestEnv::new().await;
+    let engine = env.create_engine(MitoConfig::default()).await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    engine
+        .handle_request(
+            region_id,
+            RegionRequest::EnterStaging(EnterStagingRequest {
+                partition_directive: StagingPartitionDirective::RejectAllWrites,
+            }),
+        )
+        .await
+        .unwrap();
+
+    let region = engine.get_region(region_id).unwrap();
+    assert!(region.manifest_ctx.staging_partition_info().is_some());
+
+    engine
+        .set_region_role(region_id, RegionRole::Follower)
+        .unwrap();
+
+    let region = engine.get_region(region_id).unwrap();
+    assert_eq!(engine.role(region_id), Some(RegionRole::Follower));
+    assert!(region.manifest_ctx.staging_partition_info().is_none());
+}
+
+#[tokio::test]
+async fn test_set_region_role_downgrading_leader_clears_staging_partition_info() {
+    let mut env = TestEnv::new().await;
+    let engine = env.create_engine(MitoConfig::default()).await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    engine
+        .handle_request(
+            region_id,
+            RegionRequest::EnterStaging(EnterStagingRequest {
+                partition_directive: StagingPartitionDirective::RejectAllWrites,
+            }),
+        )
+        .await
+        .unwrap();
+
+    let region = engine.get_region(region_id).unwrap();
+    assert!(region.manifest_ctx.staging_partition_info().is_some());
+
+    engine
+        .set_region_role(region_id, RegionRole::DowngradingLeader)
+        .unwrap();
+
+    let region = engine.get_region(region_id).unwrap();
+    assert_eq!(engine.role(region_id), Some(RegionRole::DowngradingLeader));
+    assert!(region.manifest_ctx.staging_partition_info().is_none());
+}
+
+#[tokio::test]
+async fn test_can_reenter_staging_after_direct_exit_cleanup() {
+    let mut env = TestEnv::new().await;
+    let engine = env.create_engine(MitoConfig::default()).await;
+
+    let region_id = RegionId::new(1, 1);
+    let request = CreateRequestBuilder::new().build();
+
+    engine
+        .handle_request(region_id, RegionRequest::Create(request))
+        .await
+        .unwrap();
+
+    engine
+        .handle_request(
+            region_id,
+            RegionRequest::EnterStaging(EnterStagingRequest {
+                partition_directive: StagingPartitionDirective::RejectAllWrites,
+            }),
+        )
+        .await
+        .unwrap();
+    engine
+        .set_region_role(region_id, RegionRole::Follower)
+        .unwrap();
+    engine
+        .set_region_role(region_id, RegionRole::Leader)
+        .unwrap();
+
+    engine
+        .handle_request(
+            region_id,
+            RegionRequest::EnterStaging(EnterStagingRequest {
+                partition_directive: StagingPartitionDirective::RejectAllWrites,
+            }),
+        )
+        .await
+        .unwrap();
+
+    let region = engine.get_region(region_id).unwrap();
+    assert_eq!(engine.role(region_id), Some(RegionRole::StagingLeader));
+    assert!(region.manifest_ctx.staging_partition_info().is_some());
+}
+
 async fn test_restricted_state_transitions_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
diff --git a/src/mito2/src/engine/skip_wal_test.rs b/src/mito2/src/engine/skip_wal_test.rs
index d1b38c47fb..c59be6ba2c 100644
--- a/src/mito2/src/engine/skip_wal_test.rs
+++ b/src/mito2/src/engine/skip_wal_test.rs
@@ -15,7 +15,9 @@
 use api::v1::Rows;
 use common_wal::options::{WAL_OPTIONS_KEY, WalOptions};
 use store_api::region_engine::{RegionEngine, RegionRole};
-use store_api::region_request::{RegionCloseRequest, RegionRequest};
+use store_api::region_request::{
+    RegionCloseRequest, RegionOpenRequest, RegionRequest, RegionTruncateRequest,
+};
 use store_api::storage::{RegionId, ScanRequest};
 
 use crate::config::MitoConfig;
@@ -168,3 +170,76 @@ async fn test_close_follower_region_skip_wal() {
     let total_rows: usize = batches.iter().map(|b| b.num_rows()).sum();
     assert_eq!(0, total_rows);
 }
+
+#[tokio::test]
+async fn test_close_region_after_truncate_skip_wal() {
+    common_telemetry::init_default_ut_logging();
+    let mut env = TestEnv::with_prefix("close-truncate-skip-wal").await;
+    let engine = env.create_engine(MitoConfig::default()).await;
+
+    let region_id = RegionId::new(1, 1);
+    let mut request = CreateRequestBuilder::new().build();
+    let wal_options = WalOptions::Noop;
+    request.options.insert(
+        WAL_OPTIONS_KEY.to_string(),
+        serde_json::to_string(&wal_options).unwrap(),
+    );
+
+    engine
+        .handle_request(region_id, RegionRequest::Create(request.clone()))
+        .await
+        .unwrap();
+
+    engine
+        .handle_request(
+            region_id,
+            RegionRequest::Truncate(RegionTruncateRequest::All),
+        )
+        .await
+        .unwrap();
+
+    let region = engine.get_region(region_id).unwrap();
+    let version_data = region.version_control.current();
+    assert_eq!(
+        version_data.version.truncated_entry_id,
+        Some(version_data.last_entry_id)
+    );
+
+    let rows = Rows {
+        schema: rows_schema(&request),
+        rows: build_rows(0, 3),
+    };
+    put_rows(&engine, region_id, rows).await;
+
+    let region = engine.get_region(region_id).unwrap();
+    assert!(!region.version().memtables.is_empty());
+
+    engine
+        .handle_request(region_id, RegionRequest::Close(RegionCloseRequest {}))
+        .await
+        .unwrap();
+
+    engine
+        .handle_request(
+            region_id,
+            RegionRequest::Open(RegionOpenRequest {
+                engine: String::new(),
+                table_dir: request.table_dir,
+                path_type: store_api::region_request::PathType::Bare,
+                options: request.options,
+                skip_wal_replay: false,
+                checkpoint: None,
+            }),
+        )
+        .await
+        .unwrap();
+    let stream = engine
+        .scan_to_stream(region_id, ScanRequest::default())
+        .await
+        .unwrap();
+    let batches = common_recordbatch::RecordBatches::try_collect(stream)
+        .await
+        .unwrap();
+    let total_rows: usize = batches.iter().map(|b| b.num_rows()).sum();
+    assert_eq!(3, total_rows);
+}
diff --git a/src/mito2/src/engine/staging_test.rs b/src/mito2/src/engine/staging_test.rs
index 2e9c5045ff..71abd67eb2 100644
--- a/src/mito2/src/engine/staging_test.rs
+++ b/src/mito2/src/engine/staging_test.rs
@@ -14,10 +14,9 @@
 
 //! Integration tests for staging state functionality.
 
-use std::assert_matches::assert_matches;
-use std::fs;
 use std::sync::Arc;
 use std::time::Duration;
+use std::{assert_matches, fs};
 
 use api::v1::Rows;
 use common_error::ext::ErrorExt;
@@ -27,8 +26,8 @@ use common_recordbatch::RecordBatches;
 use datatypes::value::Value;
 use object_store::Buffer;
 use object_store::layers::mock::{
-    Entry, Error as MockError, ErrorKind, List, Lister, Metadata, MockLayerBuilder,
-    Result as MockResult, Write, Writer,
+    Delete, Deleter, Entry, Error as MockError, ErrorKind, List, Lister, Metadata,
+    MockLayerBuilder, OpDelete, Result as MockResult, Write, Writer,
 };
 use partition::expr::{PartitionExpr, col};
 use store_api::region_engine::{
@@ -73,7 +72,7 @@ async fn test_staging_state_integration_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -131,7 +130,7 @@ async fn test_staging_blocks_alter_operations_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -172,7 +171,7 @@ async fn test_staging_blocks_truncate_operations_with_format(flat_format: bool)
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -309,7 +308,7 @@ async fn test_staging_write_partition_expr_version_with_format(flat_format: bool
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -403,9 +402,7 @@ async fn test_staging_write_partition_expr_version_with_format(flat_format: bool
     engine
         .handle_request(
             region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -506,7 +503,7 @@ async fn test_staging_manifest_directory_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -548,7 +545,7 @@ async fn test_staging_manifest_directory_with_format(flat_format: bool) {
         .await
         .unwrap();
     let region = engine.get_region(region_id).unwrap();
-    let staging_partition_info = region.staging_partition_info.lock().unwrap().clone();
+    let staging_partition_info = region.manifest_ctx.staging_partition_info();
     assert_eq!(
         staging_partition_info
             .unwrap()
@@ -611,9 +608,7 @@ async fn test_staging_manifest_directory_with_format(flat_format: bool) {
     engine
         .handle_request(
             region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -658,7 +653,7 @@ async fn test_staging_exit_success_with_manifests_with_format(flat_format: bool)
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -704,9 +699,7 @@ async fn test_staging_exit_success_with_manifests_with_format(flat_format: bool)
     engine
         .handle_request(
             region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -721,9 +714,7 @@ async fn test_staging_exit_success_with_manifests_with_format(flat_format: bool)
     engine
         .handle_request(
             region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -884,7 +875,7 @@ async fn test_enter_staging_writes_partition_expr_change_action_with_format(flat
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -948,7 +939,7 @@ async fn test_staging_exit_conflict_partition_expr_change_and_change_with_format
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -1033,7 +1024,7 @@ async fn test_write_stall_on_enter_staging_with_format(flat_format: bool) {
     let engine = env
         .create_engine_with(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             None,
@@ -1153,11 +1144,28 @@ impl Write for MockWriter {
     }
 }
 
+struct MockDeleter {
+    inner: Deleter,
+}
+
+impl Delete for MockDeleter {
+    async fn delete(&mut self, path: &str, args: OpDelete) -> MockResult<()> {
+        if path.contains("staging") {
+            return Err(MockError::new(ErrorKind::Unexpected, "mock error"));
+        }
+        self.inner.delete(path, args).await
+    }
+
+    async fn close(&mut self) -> MockResult<()> {
+        self.inner.close().await
+    }
+}
+
 async fn test_enter_staging_error(env: &mut TestEnv, flat_format: bool) {
     let partition_expr = default_partition_expr();
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -1202,6 +1210,7 @@ async fn test_enter_staging_clean_staging_manifest_error_with_format(flat_format
                 inner: lister,
             })
         }))
+        .deleter_factory(Arc::new(|deleter| Box::new(MockDeleter { inner: deleter })))
         .build()
         .unwrap();
     let mut env = TestEnv::new().await.with_mock_layer(mock_layer);
diff --git a/src/mito2/src/engine/sync_test.rs b/src/mito2/src/engine/sync_test.rs
index 65b997e498..17d73b1848 100644
--- a/src/mito2/src/engine/sync_test.rs
+++ b/src/mito2/src/engine/sync_test.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 
 use api::v1::{Rows, SemanticType};
 use common_error::ext::ErrorExt;
@@ -80,7 +80,7 @@ async fn test_sync_after_flush_region_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -112,7 +112,7 @@ async fn test_sync_after_flush_region_with_format(flat_format: bool) {
     // Open the region on the follower engine
     let follower_engine = env
         .create_follower_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -189,7 +189,7 @@ async fn test_sync_after_alter_region_with_format(flat_format: bool) {
     let mut env = TestEnv::new().await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -224,7 +224,7 @@ async fn test_sync_after_alter_region_with_format(flat_format: bool) {
     // Open the region on the follower engine
     let follower_engine = env
         .create_follower_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
diff --git a/src/mito2/src/engine/truncate_test.rs b/src/mito2/src/engine/truncate_test.rs
index 223cc2b488..8c3fdad75d 100644
--- a/src/mito2/src/engine/truncate_test.rs
+++ b/src/mito2/src/engine/truncate_test.rs
@@ -41,7 +41,7 @@ async fn test_engine_truncate_region_basic_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("truncate-basic").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -104,7 +104,7 @@ async fn test_engine_put_data_after_truncate_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("truncate-put").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -180,7 +180,7 @@ async fn test_engine_truncate_after_flush_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("truncate-flush").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -217,9 +217,7 @@ async fn test_engine_truncate_after_flush_with_format(flat_format: bool) {
     engine
         .handle_request(
             region_id,
-            RegionRequest::Flush(RegionFlushRequest {
-                row_group_size: None,
-            }),
+            RegionRequest::Flush(RegionFlushRequest::default()),
         )
         .await
         .unwrap();
@@ -270,7 +268,7 @@ async fn test_engine_truncate_reopen_with_format(flat_format: bool) {
     let mut env = TestEnv::with_prefix("truncate-reopen").await;
     let engine = env
         .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
+            default_flat_format: flat_format,
             ..Default::default()
         })
         .await;
@@ -310,7 +308,7 @@ async fn test_engine_truncate_reopen_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
@@ -355,7 +353,7 @@ async fn test_engine_truncate_during_flush_with_format(flat_format: bool) {
     let engine = env
         .create_engine_with(
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
             Some(write_buffer_manager),
@@ -395,9 +393,7 @@ async fn test_engine_truncate_during_flush_with_format(flat_format: bool) {
         engine_cloned
             .handle_request(
                 region_id,
-                RegionRequest::Flush(RegionFlushRequest {
-                    row_group_size: None,
-                }),
+                RegionRequest::Flush(RegionFlushRequest::default()),
             )
             .await
     });
@@ -436,7 +432,7 @@ async fn test_engine_truncate_during_flush_with_format(flat_format: bool) {
         .reopen_engine(
             engine,
             MitoConfig {
-                default_experimental_flat_format: flat_format,
+                default_flat_format: flat_format,
                 ..Default::default()
             },
         )
diff --git a/src/mito2/src/error.rs b/src/mito2/src/error.rs
index 923d8a2713..b822c17976 100644
--- a/src/mito2/src/error.rs
+++ b/src/mito2/src/error.rs
@@ -228,6 +228,20 @@ pub enum Error {
         location: Location,
     },
 
+    #[snafu(display(
+        "STALE_CURSOR: incremental query stale, region: {}, given_seq: {}, min_readable_seq: {}, retry_hint: FALLBACK_FULL_RECOMPUTE",
+        region_id,
+        given_seq,
+        min_readable_seq
+    ))]
+    IncrementalQueryStale {
+        region_id: RegionId,
+        given_seq: u64,
+        min_readable_seq: u64,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
     #[snafu(display("Old manifest missing for region {}", region_id))]
     MissingOldManifest {
         region_id: RegionId,
@@ -616,15 +630,6 @@ pub enum Error {
         location: Location,
     },
 
-    #[snafu(display("Failed to read arrow record batch from parquet file {}", path))]
-    ArrowReader {
-        path: String,
-        #[snafu(source)]
-        error: ArrowError,
-        #[snafu(implicit)]
-        location: Location,
-    },
-
     #[snafu(display("Column not found, column: {column}"))]
     ColumnNotFound {
         column: String,
@@ -1082,6 +1087,9 @@ pub enum Error {
     #[snafu(display("Manual compaction is override by following operations."))]
     ManualCompactionOverride {},
 
+    #[snafu(display("Compaction is cancelled."))]
+    CompactionCancelled {},
+
     #[snafu(display("Compaction memory exhausted for region {region_id} (policy: {policy})",))]
     CompactionMemoryExhausted {
         region_id: RegionId,
@@ -1232,6 +1240,14 @@ pub enum Error {
         #[snafu(implicit)]
         location: Location,
     },
+
+    #[snafu(display("Failed to cast column"))]
+    CastColumn {
+        #[snafu(source)]
+        error: datafusion::error::DataFusionError,
+        #[snafu(implicit)]
+        location: Location,
+    },
 }
 
 pub type Result<T, E = Error> = std::result::Result<T, E>;
@@ -1304,6 +1320,8 @@ impl ErrorExt for Error {
             | SerializePartitionExpr { .. }
             | InvalidSourceAndTargetRegion { .. } => StatusCode::InvalidArguments,
 
+            IncrementalQueryStale { .. } => StatusCode::RequestOutdated,
+
             RegionMetadataNotFound { .. }
             | Join { .. }
             | WorkerStopped { .. }
@@ -1319,6 +1337,7 @@ impl ErrorExt for Error {
             | ReadDataPart { .. }
             | BuildEntry { .. }
             | Metadata { .. }
+            | CastColumn { .. }
             | MitoManifestInfo { .. } => StatusCode::Internal,
 
             FetchManifests { source, .. } => source.status_code(),
@@ -1349,7 +1368,6 @@ impl ErrorExt for Error {
             RegionState { .. } | UpdateManifest { .. } => StatusCode::RegionNotReady,
             JsonOptions { .. } => StatusCode::InvalidArguments,
             EmptyRegionDir { .. } | EmptyManifestDir { .. } => StatusCode::RegionNotFound,
-            ArrowReader { .. } => StatusCode::StorageUnavailable,
             ConvertValue { source, .. } => source.status_code(),
             ApplyBloomFilterIndex { source, .. } => source.status_code(),
             InvalidPartitionExpr { source, .. } => source.status_code(),
@@ -1399,7 +1417,7 @@ impl ErrorExt for Error {
             #[cfg(feature = "vector_index")]
             VectorIndexBuild { .. } | VectorIndexFinish { .. } => StatusCode::Internal,
 
-            ManualCompactionOverride {} => StatusCode::Cancelled,
+            ManualCompactionOverride {} | CompactionCancelled {} => StatusCode::Cancelled,
 
             CompactionMemoryExhausted { source, .. } => source.status_code(),
 
diff --git a/src/mito2/src/extension.rs b/src/mito2/src/extension.rs
index 944344b1cf..b31d006c27 100644
--- a/src/mito2/src/extension.rs
+++ b/src/mito2/src/extension.rs
@@ -8,14 +8,23 @@ use common_time::range::TimestampRange;
 use store_api::storage::{ScanRequest, SequenceNumber};
 
 use crate::error::Result;
+use crate::read::BoxedRecordBatchStream;
 use crate::read::range::RowGroupIndex;
 use crate::read::scan_region::StreamContext;
 use crate::read::scan_util::PartitionMetrics;
-use crate::read::{BoxedBatchStream, BoxedRecordBatchStream};
 use crate::region::MitoRegionRef;
+use crate::sst::parquet::file_range::PreFilterMode;
 
 pub type InclusiveTimeRange = (Timestamp, Timestamp);
 
+/// Per-range read options passed to [`ExtensionRange::flat_reader`].
+#[derive(Debug, Clone, Copy)]
+pub struct ExtensionRangeReadOptions {
+    /// How aggressively to pre-filter columns before merging with other sources
+    /// in the same partition range.
+    pub pre_filter_mode: PreFilterMode,
+}
+
 /// [`ExtensionRange`] is used to represent a scannable "range" for mito engine, just like the
 /// memtable range and sst file range, but resides on the outside.
 /// It can be scanned side by side as other ranges to produce the final result, so it's very useful
@@ -30,29 +39,16 @@ pub trait ExtensionRange: Debug + Send + Sync {
     /// The row groups number in this range.
     fn num_row_groups(&self) -> u64;
 
-    /// Create the reader for reading this range.
-    fn reader(&self, context: &StreamContext) -> BoxedExtensionRangeReader;
-
     /// Create the flat reader for reading this range in flat format.
-    fn flat_reader(&self, context: &StreamContext) -> BoxedExtensionFlatRangeReader;
+    fn flat_reader(
+        &self,
+        context: &StreamContext,
+        options: ExtensionRangeReadOptions,
+    ) -> BoxedExtensionFlatRangeReader;
 }
 
 pub type BoxedExtensionRange = Box<dyn ExtensionRange>;
 
-/// The reader to read an extension range.
-#[async_trait]
-pub trait ExtensionRangeReader: Send {
-    /// Read the extension range by creating a stream that produces [`Batch`].
-    async fn read(
-        self: Box<Self>,
-        context: Arc<StreamContext>,
-        metrics: PartitionMetrics,
-        index: RowGroupIndex,
-    ) -> Result<BoxedBatchStream, BoxedError>;
-}
-
-pub type BoxedExtensionRangeReader = Box<dyn ExtensionRangeReader>;
-
 /// The reader to read an extension range in flat format (producing [`RecordBatch`]).
 #[async_trait]
 pub trait ExtensionFlatRangeReader: Send {
diff --git a/src/mito2/src/flush.rs b/src/mito2/src/flush.rs
index 0c16544b6e..c863d15d08 100644
--- a/src/mito2/src/flush.rs
+++ b/src/mito2/src/flush.rs
@@ -20,12 +20,13 @@ use std::sync::Arc;
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::time::Instant;
 
+use bytes::Bytes;
 use common_telemetry::{debug, error, info};
 use datatypes::arrow::datatypes::SchemaRef;
-use either::Either;
 use partition::expr::PartitionExpr;
 use smallvec::{SmallVec, smallvec};
 use snafu::ResultExt;
+use store_api::region_request::RegionFlushReason;
 use store_api::storage::{RegionId, SequenceNumber};
 use strum::IntoStaticStr;
 use tokio::sync::{Semaphore, mpsc, watch};
@@ -41,18 +42,14 @@ use crate::error::{
 };
 use crate::manifest::action::{RegionEdit, RegionMetaAction, RegionMetaActionList};
 use crate::memtable::bulk::ENCODE_ROW_THRESHOLD;
-use crate::memtable::{
-    BoxedRecordBatchIterator, EncodedRange, IterBuilder, MemtableRanges, RangesOptions,
-};
+use crate::memtable::{BoxedRecordBatchIterator, EncodedRange, MemtableRanges, RangesOptions};
 use crate::metrics::{
     FLUSH_BYTES_TOTAL, FLUSH_ELAPSED, FLUSH_FAILURE_TOTAL, FLUSH_FILE_TOTAL, FLUSH_REQUESTS_TOTAL,
     INFLIGHT_FLUSH_COUNT,
 };
-use crate::read::dedup::{DedupReader, LastNonNull, LastRow};
+use crate::read::FlatSource;
 use crate::read::flat_dedup::{FlatDedupIterator, FlatLastNonNull, FlatLastRow};
 use crate::read::flat_merge::FlatMergeIterator;
-use crate::read::merge::MergeReaderBuilder;
-use crate::read::{FlatSource, Source};
 use crate::region::options::{IndexOptions, MergeMode, RegionOptions};
 use crate::region::version::{VersionControlData, VersionControlRef, VersionRef};
 use crate::region::{ManifestContextRef, RegionLeaderState, RegionRoleState, parse_partition_expr};
@@ -62,8 +59,11 @@ use crate::request::{
 };
 use crate::schedule::scheduler::{Job, SchedulerRef};
 use crate::sst::file::FileMeta;
-use crate::sst::parquet::{DEFAULT_READ_BATCH_SIZE, DEFAULT_ROW_GROUP_SIZE, SstInfo, WriteOptions};
-use crate::sst::{FlatSchemaOptions, to_flat_sst_arrow_schema};
+use crate::sst::parquet::metadata::extract_primary_key_range;
+use crate::sst::parquet::{
+    DEFAULT_READ_BATCH_SIZE, DEFAULT_ROW_GROUP_SIZE, SstInfo, WriteOptions, flat_format,
+};
+use crate::sst::{FlatSchemaOptions, FormatType, to_flat_sst_arrow_schema};
 use crate::worker::WorkerListener;
 
 /// Global write buffer (memtable) manager.
@@ -222,6 +222,12 @@ pub enum FlushReason {
     EnterStaging,
     /// Flush when region is closing.
     Closing,
+    /// Flush triggered before region migration.
+    RegionMigration,
+    /// Flush triggered by repartition procedure.
+    Repartition,
+    /// Flush triggered by remote WAL pruning.
+    RemoteWalPrune,
 }
 
 impl FlushReason {
@@ -231,6 +237,16 @@ impl FlushReason {
     }
 }
 
+impl From<RegionFlushReason> for FlushReason {
+    fn from(reason: RegionFlushReason) -> Self {
+        match reason {
+            RegionFlushReason::RegionMigration => FlushReason::RegionMigration,
+            RegionFlushReason::Repartition => FlushReason::Repartition,
+            RegionFlushReason::RemoteWalPrune => FlushReason::RemoteWalPrune,
+        }
+    }
+}
+
 /// Task to flush a region.
 pub(crate) struct RegionFlushTask {
     /// Region to flush.
@@ -480,92 +496,60 @@ impl RegionFlushTask {
             // the counter may have more series than the actual series count.
             series_count += memtable_series_count;
 
-            if mem_ranges.is_record_batch() {
-                let flush_start = Instant::now();
-                let FlushFlatMemResult {
-                    num_encoded,
-                    num_sources,
-                    results,
-                } = self
-                    .flush_flat_mem_ranges(version, &write_opts, mem_ranges)
-                    .await?;
-                encoded_part_count += num_encoded;
-                for (source_idx, result) in results.into_iter().enumerate() {
-                    let (max_sequence, ssts_written, metrics) = result?;
-                    if ssts_written.is_empty() {
-                        // No data written.
-                        continue;
-                    }
-
-                    common_telemetry::debug!(
-                        "Region {} flush one memtable {} {}/{}, metrics: {:?}",
-                        self.region_id,
-                        memtable_id,
-                        source_idx,
-                        num_sources,
-                        metrics
-                    );
-
-                    flush_metrics = flush_metrics.merge(metrics);
-
-                    file_metas.extend(ssts_written.into_iter().map(|sst_info| {
-                        flushed_bytes += sst_info.file_size;
-                        Self::new_file_meta(
-                            self.region_id,
-                            max_sequence,
-                            sst_info,
-                            partition_expr.clone(),
-                        )
-                    }));
-                }
-
-                common_telemetry::debug!(
-                    "Region {} flush {} memtables for {}, num_mem_ranges: {}, num_encoded: {}, num_rows: {}, flush_cost: {:?}, compact_cost: {:?}",
-                    self.region_id,
-                    num_sources,
-                    memtable_id,
-                    num_mem_ranges,
-                    num_encoded,
-                    num_mem_rows,
-                    flush_start.elapsed(),
-                    compact_cost,
-                );
-            } else {
-                let max_sequence = mem_ranges.max_sequence();
-                let source = memtable_source(mem_ranges, &version.options).await?;
-
-                // Flush to level 0.
-                let source = Either::Left(source);
-                let write_request = self.new_write_request(version, max_sequence, source);
-
-                let mut metrics = Metrics::new(WriteType::Flush);
-                let ssts_written = self
-                    .access_layer
-                    .write_sst(write_request, &write_opts, &mut metrics)
-                    .await?;
-                FLUSH_FILE_TOTAL.inc_by(ssts_written.len() as u64);
+            let flush_start = Instant::now();
+            let FlushFlatMemResult {
+                num_encoded,
+                num_sources,
+                results,
+            } = self
+                .flush_flat_mem_ranges(version, &write_opts, mem_ranges)
+                .await?;
+            encoded_part_count += num_encoded;
+            for (source_idx, result) in results.into_iter().enumerate() {
+                let (max_sequence, ssts_written, metrics) = result?;
                 if ssts_written.is_empty() {
                     // No data written.
                     continue;
                 }
 
-                debug!(
-                    "Region {} flush one memtable, num_mem_ranges: {}, num_rows: {}, metrics: {:?}",
-                    self.region_id, num_mem_ranges, num_mem_rows, metrics
+                common_telemetry::debug!(
+                    "Region {} flush one memtable {} {}/{}, metrics: {:?}",
+                    self.region_id,
+                    memtable_id,
+                    source_idx,
+                    num_sources,
+                    metrics
                 );
 
                 flush_metrics = flush_metrics.merge(metrics);
 
-                file_metas.extend(ssts_written.into_iter().map(|sst_info| {
+                for sst_info in ssts_written {
                     flushed_bytes += sst_info.file_size;
-                    Self::new_file_meta(
+                    let pk_range = sst_info
+                        .file_metadata
+                        .as_ref()
+                        .and_then(|meta| extract_primary_key_range(meta, &version.metadata));
+                    file_metas.push(Self::new_file_meta(
                         self.region_id,
                         max_sequence,
                         sst_info,
                         partition_expr.clone(),
-                    )
-                }));
-            };
+                        pk_range,
+                    ));
+                }
+            }
+
+            common_telemetry::debug!(
+                "Region {} flush {} memtables for {}, num_mem_ranges: {}, num_encoded: {}, num_rows: {}, flush_cost: {:?}, compact_cost: {:?}",
+                self.region_id,
+                num_sources,
+                memtable_id,
+                num_mem_ranges,
+                num_encoded,
+                num_mem_rows,
+                flush_start.elapsed(),
+                compact_cost,
+            );
         }
 
         Ok(DoFlushMemtablesResult {
@@ -587,16 +571,17 @@ impl RegionFlushTask {
             &version.metadata,
             &FlatSchemaOptions::from_encoding(version.metadata.primary_key_encoding),
         );
+        let field_column_start =
+            flat_format::field_column_start(&version.metadata, batch_schema.fields().len());
         let flat_sources = memtable_flat_sources(
             batch_schema,
             mem_ranges,
             &version.options,
-            version.metadata.primary_key.len(),
+            field_column_start,
         )?;
         let mut tasks = Vec::with_capacity(flat_sources.encoded.len() + flat_sources.sources.len());
         let num_encoded = flat_sources.encoded.len();
         for (source, max_sequence) in flat_sources.sources {
-            let source = Either::Right(source);
             let write_request = self.new_write_request(version, max_sequence, source);
             let access_layer = self.access_layer.clone();
             let write_opts = write_opts.clone();
@@ -643,7 +628,12 @@ impl RegionFlushTask {
         max_sequence: u64,
         sst_info: SstInfo,
         partition_expr: Option<PartitionExpr>,
+        primary_key_range: Option<(Bytes, Bytes)>,
     ) -> FileMeta {
+        let (primary_key_min, primary_key_max) = match primary_key_range {
+            Some((min, max)) => (Some(min), Some(max)),
+            None => (None, None),
+        };
         FileMeta {
             region_id,
             file_id: sst_info.file_id,
@@ -660,6 +650,8 @@ impl RegionFlushTask {
             sequence: NonZeroU64::new(max_sequence),
             partition_expr,
             num_series: sst_info.num_series,
+            primary_key_min,
+            primary_key_max,
         }
     }
 
@@ -667,8 +659,13 @@ impl RegionFlushTask {
         &self,
         version: &VersionRef,
         max_sequence: u64,
-        source: Either<Source, FlatSource>,
+        source: FlatSource,
     ) -> SstWriteRequest {
+        let flat_format = version
+            .options
+            .sst_format
+            .map(|f| f == FormatType::Flat)
+            .unwrap_or(self.engine_config.default_flat_format);
         SstWriteRequest {
             op_type: OperationType::Flush,
             metadata: version.metadata.clone(),
@@ -676,6 +673,11 @@ impl RegionFlushTask {
             cache_manager: self.cache_manager.clone(),
             storage: version.options.storage.clone(),
             max_sequence: Some(max_sequence),
+            sst_write_format: if flat_format {
+                FormatType::Flat
+            } else {
+                FormatType::PrimaryKey
+            },
             index_options: self.index_options.clone(),
             index_config: self.engine_config.index.clone(),
             inverted_index_config: self.engine_config.inverted_index.clone(),
@@ -722,41 +724,6 @@ struct DoFlushMemtablesResult {
     flush_metrics: Metrics,
 }
 
-/// Returns a [Source] for the given memtable.
-async fn memtable_source(mem_ranges: MemtableRanges, options: &RegionOptions) -> Result<Source> {
-    let source = if mem_ranges.ranges.len() == 1 {
-        let only_range = mem_ranges.ranges.into_values().next().unwrap();
-        let iter = only_range.build_iter()?;
-        Source::Iter(iter)
-    } else {
-        // todo(hl): a workaround since sync version of MergeReader is wip.
-        let sources = mem_ranges
-            .ranges
-            .into_values()
-            .map(|r| r.build_iter().map(Source::Iter))
-            .collect::<Result<Vec<_>>>()?;
-        let merge_reader = MergeReaderBuilder::from_sources(sources).build().await?;
-        let maybe_dedup = if options.append_mode {
-            // no dedup in append mode
-            Box::new(merge_reader) as _
-        } else {
-            // dedup according to merge mode
-            match options.merge_mode.unwrap_or(MergeMode::LastRow) {
-                MergeMode::LastRow => {
-                    Box::new(DedupReader::new(merge_reader, LastRow::new(false), None)) as _
-                }
-                MergeMode::LastNonNull => Box::new(DedupReader::new(
-                    merge_reader,
-                    LastNonNull::new(false),
-                    None,
-                )) as _,
-            }
-        };
-        Source::Reader(maybe_dedup)
-    };
-    Ok(source)
-}
-
 struct FlatSources {
     sources: SmallVec<[(FlatSource, SequenceNumber); 4]>,
     encoded: SmallVec<[(EncodedRange, SequenceNumber); 4]>,
@@ -794,7 +761,7 @@ fn memtable_flat_sources(
             );
             flat_sources
                 .sources
-                .push((FlatSource::Iter(iter), max_sequence));
+                .push((FlatSource::new_iter(schema, iter), max_sequence));
         };
     } else {
         let min_flush_rows = *ENCODE_ROW_THRESHOLD;
@@ -857,9 +824,10 @@ fn memtable_flat_sources(
                     std::mem::replace(&mut input_iters, Vec::with_capacity(num_ranges)),
                 )?;
 
-                flat_sources
-                    .sources
-                    .push((FlatSource::Iter(maybe_dedup), max_sequence));
+                flat_sources.sources.push((
+                    FlatSource::new_iter(schema.clone(), maybe_dedup),
+                    max_sequence,
+                ));
                 last_iter_rows = 0;
                 current_ranges.clear();
             }
@@ -890,7 +858,7 @@ fn memtable_flat_sources(
 
             flat_sources
                 .sources
-                .push((FlatSource::Iter(maybe_dedup), max_sequence));
+                .push((FlatSource::new_iter(schema, maybe_dedup), max_sequence));
         }
     }
 
@@ -1563,14 +1531,10 @@ mod tests {
             // Consume the iterator and count rows
             let mut total_rows = 0usize;
             for (source, _sequence) in flat_sources.sources {
-                match source {
-                    crate::read::FlatSource::Iter(iter) => {
-                        for rb in iter {
-                            total_rows += rb.unwrap().num_rows();
-                        }
-                    }
-                    crate::read::FlatSource::Stream(_) => unreachable!(),
-                }
+                total_rows += source
+                    .take_iter()
+                    .map(|x| x.unwrap().num_rows())
+                    .sum::<usize>();
             }
             assert_eq!(1, total_rows, "dedup should keep a single row");
         }
@@ -1593,14 +1557,10 @@ mod tests {
 
             let mut total_rows = 0usize;
             for (source, _sequence) in flat_sources.sources {
-                match source {
-                    crate::read::FlatSource::Iter(iter) => {
-                        for rb in iter {
-                            total_rows += rb.unwrap().num_rows();
-                        }
-                    }
-                    crate::read::FlatSource::Stream(_) => unreachable!(),
-                }
+                total_rows += source
+                    .take_iter()
+                    .map(|x| x.unwrap().num_rows())
+                    .sum::<usize>();
             }
             assert_eq!(2, total_rows, "append_mode should preserve duplicates");
         }
diff --git a/src/mito2/src/lib.rs b/src/mito2/src/lib.rs
index a15711b34a..7d43685ded 100644
--- a/src/mito2/src/lib.rs
+++ b/src/mito2/src/lib.rs
@@ -16,8 +16,6 @@
 //!
 //! Mito is the a region engine to store timeseries data.
 
-#![feature(assert_matches)]
-#![feature(int_roundings)]
 #![feature(debug_closure_helpers)]
 #![feature(duration_constructors)]
 
diff --git a/src/mito2/src/manifest/checkpointer.rs b/src/mito2/src/manifest/checkpointer.rs
index 1da03dda21..9dd35e189a 100644
--- a/src/mito2/src/manifest/checkpointer.rs
+++ b/src/mito2/src/manifest/checkpointer.rs
@@ -16,7 +16,7 @@ use std::fmt::Debug;
 use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
 
-use common_telemetry::{error, info};
+use common_telemetry::{error, info, warn};
 use store_api::storage::RegionId;
 use store_api::{MIN_VERSION, ManifestVersion};
 
@@ -69,14 +69,18 @@ impl Inner {
             return;
         }
 
-        if let Err(e) = self.manifest_store.delete_until(version, true).await {
-            error!(e; "Failed to delete manifest actions until version {} for region {}", version, region_id);
-            return;
-        }
-
+        // Advance the in-memory checkpoint version as soon as the checkpoint file
+        // is durable. If the subsequent delta cleanup fails, the on-disk state is
+        // still consistent (the `_last_checkpoint` metadata points at the new
+        // checkpoint) and `maybe_do_checkpoint` must not re-checkpoint the same
+        // range.
         self.last_checkpoint_version
             .store(version, Ordering::Relaxed);
 
+        if let Err(e) = self.manifest_store.delete_until(version, true).await {
+            warn!(e; "Failed to delete manifest actions until version {} for region {}, leftover files will be ignored on recovery", version, region_id);
+        }
+
         info!(
             "Checkpoint for region {} success, version: {}",
             region_id, version
diff --git a/src/mito2/src/manifest/manager.rs b/src/mito2/src/manifest/manager.rs
index e489c6756d..a69f08693c 100644
--- a/src/mito2/src/manifest/manager.rs
+++ b/src/mito2/src/manifest/manager.rs
@@ -14,6 +14,7 @@
 
 use std::sync::Arc;
 use std::sync::atomic::{AtomicU64, Ordering};
+use std::time::Instant;
 
 use common_datasource::compression::CompressionType;
 use common_telemetry::{debug, info};
@@ -34,8 +35,8 @@ use crate::manifest::action::{
 };
 use crate::manifest::checkpointer::Checkpointer;
 use crate::manifest::storage::{
-    ManifestObjectStore, file_version, is_checkpoint_file, is_delta_file, manifest_compress_type,
-    manifest_dir,
+    ManifestObjectStore, file_version, is_checkpoint_file, is_delta_file, list_start_after,
+    manifest_compress_type, manifest_dir,
 };
 use crate::metrics::MANIFEST_OP_ELAPSED;
 use crate::region::{ManifestStats, RegionLeaderState, RegionRoleState};
@@ -255,6 +256,7 @@ impl RegionManifestManager {
         let _t = MANIFEST_OP_ELAPSED
             .with_label_values(&["open"])
             .start_timer();
+        let open_start = Instant::now();
 
         // construct storage
         let mut store = ManifestObjectStore::new(
@@ -290,8 +292,15 @@ impl RegionManifestManager {
             RegionManifestBuilder::default()
         };
 
+        let replay_start_version = version;
+        info!(
+            "Replaying region manifest {} from version {}, last checkpoint version: {}",
+            options.manifest_dir, replay_start_version, last_checkpoint_version,
+        );
+
         // apply actions from storage
         let manifests = store.fetch_manifests(version, MAX_VERSION).await?;
+        let replayed_deltas = manifests.len();
 
         for (manifest_version, raw_action_list) in manifests {
             let action_list = RegionMetaActionList::decode(&raw_action_list)?;
@@ -334,6 +343,7 @@ impl RegionManifestManager {
         );
         let version = manifest.manifest_version;
 
+        let manifest_dir = options.manifest_dir.clone();
         let checkpointer = Checkpointer::new(
             manifest.metadata.region_id,
             options,
@@ -344,6 +354,16 @@ impl RegionManifestManager {
         manifest
             .removed_files
             .update_file_removed_cnt_to_stats(stats);
+        info!(
+            "Opened region manifest {}, region_id: {}, start_version: {}, last_checkpoint_version: {}, replayed_deltas: {}, final_version: {}, cost: {:?}",
+            manifest_dir,
+            manifest.metadata.region_id,
+            replay_start_version,
+            last_checkpoint_version,
+            replayed_deltas,
+            version,
+            open_start.elapsed(),
+        );
         Ok(Some(Self {
             store,
             last_version: manifest_version,
@@ -632,13 +652,17 @@ impl RegionManifestManager {
     pub async fn has_update(&self) -> Result<bool> {
         let last_version = self.last_version();
 
-        let streamer =
-            self.store
-                .manifest_lister(false)
-                .await?
-                .context(error::EmptyManifestDirSnafu {
-                    manifest_dir: self.store.manifest_dir(),
-                })?;
+        // Skip older files at the object-store layer. Files for `v == last_version`
+        // may still appear (`{path}{v:020}` sorts before `{path}{v:020}.json`) but
+        // they are filtered out below by the `version > last_version` check.
+        let start_after = list_start_after(self.store.manifest_dir(), last_version);
+        let streamer = self
+            .store
+            .manifest_lister(false, Some(&start_after))
+            .await?
+            .context(error::EmptyManifestDirSnafu {
+                manifest_dir: self.store.manifest_dir(),
+            })?;
 
         let need_update = streamer
             .try_any(|entry| async move {
diff --git a/src/mito2/src/manifest/storage.rs b/src/mito2/src/manifest/storage.rs
index 6013c0393f..5f795cc1c9 100644
--- a/src/mito2/src/manifest/storage.rs
+++ b/src/mito2/src/manifest/storage.rs
@@ -24,18 +24,22 @@ use std::sync::Arc;
 use std::sync::atomic::AtomicU64;
 
 use common_datasource::compression::CompressionType;
-use common_telemetry::debug;
+use common_telemetry::{debug, warn};
 use crc32fast::Hasher;
 use lazy_static::lazy_static;
 use object_store::util::join_dir;
 use object_store::{Lister, ObjectStore, util};
 use regex::Regex;
-use snafu::{ResultExt, ensure};
+#[cfg(test)]
+use snafu::ResultExt;
+use snafu::ensure;
 use store_api::ManifestVersion;
 use store_api::storage::RegionId;
 
 use crate::cache::manifest_cache::ManifestCache;
-use crate::error::{ChecksumMismatchSnafu, OpenDalSnafu, Result};
+#[cfg(test)]
+use crate::error::OpenDalSnafu;
+use crate::error::{ChecksumMismatchSnafu, Result};
 use crate::manifest::storage::checkpoint::CheckpointStorage;
 use crate::manifest::storage::delta::DeltaStorage;
 use crate::manifest::storage::size_tracker::{CheckpointTracker, DeltaTracker, SizeTracker};
@@ -76,6 +80,24 @@ pub fn checkpoint_file(version: ManifestVersion) -> String {
     format!("{version:020}.checkpoint")
 }
 
+/// Returns a lexicographic `start_after` key for an object-store `list`
+/// request over the manifest directory at `path`.
+///
+/// `path` must be the same directory prefix passed to `lister_with(path)`
+/// and must end with `/`. OpenDAL resolves `start_after` against the
+/// operator root, not relative to the listed path, so the caller must
+/// supply the full prefix — otherwise the bound is compared against keys
+/// that already share a longer prefix and is silently a no-op.
+pub(crate) fn list_start_after(path: &str, version: ManifestVersion) -> String {
+    debug_assert!(
+        path.ends_with('/'),
+        "list_start_after: path must end with '/', got {path:?}",
+    );
+    // Manifest files are named `{version:020}.{json,checkpoint}[.gz]` and sort lexicographically;
+    // `{path}{version:020}` is a strict prefix of `{path}{version:020}.{json,checkpoint}[.gz]`.
+    format!("{path}{version:020}")
+}
+
 pub fn gen_path(path: &str, file: &str, compress_type: CompressionType) -> String {
     if compress_type == CompressionType::Uncompressed {
         format!("{}{}", path, file)
@@ -194,11 +216,19 @@ impl ManifestObjectStore {
     }
 
     /// Returns an iterator of manifests from normal or staging directory.
-    pub(crate) async fn manifest_lister(&self, is_staging: bool) -> Result<Option<Lister>> {
+    ///
+    /// `start_after` is forwarded to the non-staging lister to skip entries
+    /// whose name is lexicographically less than or equal to it. It is
+    /// ignored for the staging directory.
+    pub(crate) async fn manifest_lister(
+        &self,
+        is_staging: bool,
+        start_after: Option<&str>,
+    ) -> Result<Option<Lister>> {
         if is_staging {
             self.staging_storage.manifest_lister().await
         } else {
-            self.delta_storage.manifest_lister().await
+            self.delta_storage.manifest_lister(start_after).await
         }
     }
 
@@ -239,9 +269,14 @@ impl ManifestObjectStore {
         keep_last_checkpoint: bool,
     ) -> Result<usize> {
         // Stores (entry, is_checkpoint, version) in a Vec.
+        //
+        // `start_after` is intentionally `None` here: a previous deletion
+        // may have been interrupted and left stale files at versions below
+        // the current checkpoint; we need the lister to surface them so
+        // cleanup can finish.
         let entries: Vec<_> = self
             .delta_storage
-            .get_paths(|entry| {
+            .get_paths(None, |entry| {
                 let file_name = entry.name();
                 let is_checkpoint = is_checkpoint_file(file_name);
                 if is_delta_file(file_name) || is_checkpoint_file(file_name) {
@@ -287,11 +322,11 @@ impl ManifestObjectStore {
             .iter()
             .map(|(e, _, _)| e.path().to_string())
             .collect::<Vec<_>>();
-        let ret = paths.len();
+        let total = paths.len();
 
         debug!(
             "Deleting {} logs from manifest storage path {} until {}, checkpoint_version: {:?}, paths: {:?}",
-            ret, self.path, end, checkpoint_version, paths,
+            total, self.path, end, checkpoint_version, paths,
         );
 
         // Remove from cache first
@@ -299,13 +334,37 @@ impl ManifestObjectStore {
             remove_from_cache(self.manifest_cache.as_ref(), entry.path()).await;
         }
 
-        self.object_store
-            .delete_iter(paths)
-            .await
-            .context(OpenDalSnafu)?;
+        // Try batch delete first. On failure, fall back to per-file deletes.
+        // This is a workaround for S3-compatible object stores that do not support batch delete. See issue #7986.
+        let mut succeeded = vec![false; del_entries.len()];
+        match self.object_store.delete_iter(paths.clone()).await {
+            Ok(()) => succeeded.fill(true),
+            Err(batch_err) => {
+                warn!(
+                    batch_err;
+                    "Batch delete failed for manifest path {}, falling back to per-file delete for {} paths",
+                    self.path, total,
+                );
+                for (i, path) in paths.iter().enumerate() {
+                    if let Err(e) = self.object_store.delete(path).await {
+                        warn!(
+                            e;
+                            "Failed to delete manifest file {} under {}, aborting fallback, {} files will be retried on next checkpoint",
+                            path, self.path, total - i,
+                        );
+                        break;
+                    }
+                    succeeded[i] = true;
+                }
+            }
+        }
 
-        // delete manifest sizes
-        for (_, is_checkpoint, version) in &del_entries {
+        let mut deleted = 0usize;
+        for (i, (_, is_checkpoint, version)) in del_entries.iter().enumerate() {
+            if !succeeded[i] {
+                continue;
+            }
+            deleted += 1;
             if *is_checkpoint {
                 self.size_tracker
                     .remove(&size_tracker::FileKey::Checkpoint(*version));
@@ -315,7 +374,7 @@ impl ManifestObjectStore {
             }
         }
 
-        Ok(ret)
+        Ok(deleted)
     }
 
     /// Save the delta manifest file.
@@ -420,12 +479,16 @@ mod tests {
     use crate::manifest::storage::checkpoint::CheckpointMetadata;
 
     fn new_test_manifest_store() -> ManifestObjectStore {
+        new_test_manifest_store_at("/")
+    }
+
+    fn new_test_manifest_store_at(path: &str) -> ManifestObjectStore {
         common_telemetry::init_default_ut_logging();
         let tmp_dir = create_temp_dir("test_manifest_log_store");
         let builder = Fs::default().root(&tmp_dir.path().to_string_lossy());
         let object_store = ObjectStore::new(builder).unwrap().finish();
         ManifestObjectStore::new(
-            "/",
+            path,
             object_store,
             CompressionType::Uncompressed,
             Default::default(),
@@ -690,4 +753,66 @@ mod tests {
 
         assert_eq!(log_store.total_manifest_size(), 0);
     }
+
+    #[tokio::test]
+    async fn test_scan_with_start_after_uncompress() {
+        let mut log_store = new_test_manifest_store();
+        log_store.set_compress_type(CompressionType::Uncompressed);
+        test_scan_with_start_after_case(log_store).await;
+    }
+
+    #[tokio::test]
+    async fn test_scan_with_start_after_compress() {
+        let mut log_store = new_test_manifest_store();
+        log_store.set_compress_type(CompressionType::Gzip);
+        test_scan_with_start_after_case(log_store).await;
+    }
+
+    // OpenDAL resolves `start_after` against the operator
+    // root, so the bound must embed the manifest directory prefix. Running the
+    // same assertions against a non-root path exercises that composition.
+    #[tokio::test]
+    async fn test_scan_with_start_after_nested_path() {
+        let mut log_store = new_test_manifest_store_at("/nested/region-1/");
+        log_store.set_compress_type(CompressionType::Uncompressed);
+        test_scan_with_start_after_case(log_store).await;
+    }
+
+    async fn test_scan_with_start_after_case(mut log_store: ManifestObjectStore) {
+        for v in 0..10 {
+            log_store
+                .save(v, format!("hello, {v}").as_bytes(), false)
+                .await
+                .unwrap();
+        }
+        // A checkpoint at version 5 shares the directory; scan must still
+        // return only delta files in range.
+        log_store
+            .save_checkpoint(5, "checkpoint".as_bytes())
+            .await
+            .unwrap();
+
+        // start > 0: `start_after` must skip pre-start deltas without losing any.
+        let entries = log_store.delta_storage.scan(3, 10).await.unwrap();
+        let versions: Vec<_> = entries.iter().map(|(v, _)| *v).collect();
+        assert_eq!(versions, vec![3, 4, 5, 6, 7, 8, 9]);
+
+        // start == 0: `start_after` is skipped; every delta is returned.
+        let entries = log_store.delta_storage.scan(0, 10).await.unwrap();
+        let versions: Vec<_> = entries.iter().map(|(v, _)| *v).collect();
+        assert_eq!(versions, (0..10).collect::<Vec<_>>());
+
+        // Upper bound exclusive.
+        let entries = log_store.delta_storage.scan(7, 9).await.unwrap();
+        let versions: Vec<_> = entries.iter().map(|(v, _)| *v).collect();
+        assert_eq!(versions, vec![7, 8]);
+
+        // Start beyond any existing file returns empty.
+        let entries = log_store
+            .delta_storage
+            .scan(10, ManifestVersion::MAX)
+            .await
+            .unwrap();
+        assert!(entries.is_empty());
+    }
 }
diff --git a/src/mito2/src/manifest/storage/delta.rs b/src/mito2/src/manifest/storage/delta.rs
index 3f312f1d07..594b56ddae 100644
--- a/src/mito2/src/manifest/storage/delta.rs
+++ b/src/mito2/src/manifest/storage/delta.rs
@@ -34,7 +34,7 @@ use crate::manifest::storage::utils::{
 };
 use crate::manifest::storage::{
     FETCH_MANIFEST_PARALLELISM, delta_file, file_compress_type, file_version, gen_path,
-    is_delta_file,
+    is_delta_file, list_start_after,
 };
 
 #[derive(Debug, Clone)]
@@ -76,8 +76,18 @@ impl<T: Tracker> DeltaStorage<T> {
     }
 
     /// Returns an iterator of manifests from path directory.
-    pub(crate) async fn manifest_lister(&self) -> Result<Option<Lister>> {
-        match self.object_store.lister_with(&self.path).await {
+    ///
+    /// If `start_after` is `Some`, the lister will skip entries whose name is
+    /// lexicographically less than or equal to it (see OpenDAL's `start_after`).
+    pub(crate) async fn manifest_lister(
+        &self,
+        start_after: Option<&str>,
+    ) -> Result<Option<Lister>> {
+        let mut builder = self.object_store.lister_with(&self.path);
+        if let Some(s) = start_after {
+            builder = builder.start_after(s);
+        }
+        match builder.await {
             Ok(streamer) => Ok(Some(streamer)),
             Err(e) if e.kind() == ErrorKind::NotFound => {
                 debug!("Manifest directory does not exist: {}", self.path);
@@ -90,16 +100,22 @@ impl<T: Tracker> DeltaStorage<T> {
     /// Return all `R`s in the directory that meet the `filter` conditions (that is, the `filter` closure returns `Some(R)`),
     /// and discard `R` that does not meet the conditions (that is, the `filter` closure returns `None`)
     /// Return an empty vector when directory is not found.
-    pub async fn get_paths<F, R>(&self, filter: F) -> Result<Vec<R>>
+    ///
+    /// `start_after` is forwarded to the underlying lister to skip entries
+    /// whose name is lexicographically less than or equal to it.
+    pub async fn get_paths<F, R>(&self, start_after: Option<&str>, mut filter: F) -> Result<Vec<R>>
     where
-        F: Fn(Entry) -> Option<R>,
+        F: FnMut(Entry) -> Option<R>,
     {
-        let Some(streamer) = self.manifest_lister().await? else {
+        let Some(streamer) = self.manifest_lister(start_after).await? else {
             return Ok(vec![]);
         };
 
         streamer
-            .try_filter_map(|e| async { Ok(filter(e)) })
+            .try_filter_map(|e| {
+                let result = filter(e);
+                async { Ok(result) }
+            })
             .try_collect::<Vec<_>>()
             .await
             .context(OpenDalSnafu)
@@ -113,8 +129,13 @@ impl<T: Tracker> DeltaStorage<T> {
     ) -> Result<Vec<(ManifestVersion, Entry)>> {
         ensure!(start <= end, InvalidScanIndexSnafu { start, end });
 
+        // Push the version lower bound into the list request via
+        // `list_start_after`; skip the hint when `start == 0` (nothing to skip).
+        let start_after = (start > 0).then(|| list_start_after(&self.path, start));
+        let mut total_paths = 0;
         let mut entries: Vec<(ManifestVersion, Entry)> = self
-            .get_paths(|entry| {
+            .get_paths(start_after.as_deref(), |entry| {
+                total_paths += 1;
                 let file_name = entry.name();
                 if is_delta_file(file_name) {
                     let version = file_version(file_name);
@@ -128,6 +149,16 @@ impl<T: Tracker> DeltaStorage<T> {
 
         sort_manifests(&mut entries);
 
+        common_telemetry::debug!(
+            "DeltaStorage get paths for {}, start: {}, end: {}, start_after: {:?}, total_paths: {}, entries: {}",
+            self.path,
+            start,
+            end,
+            start_after,
+            total_paths,
+            entries.len()
+        );
+
         Ok(entries)
     }
 
diff --git a/src/mito2/src/manifest/storage/staging.rs b/src/mito2/src/manifest/storage/staging.rs
index 17d341605c..dd0105a091 100644
--- a/src/mito2/src/manifest/storage/staging.rs
+++ b/src/mito2/src/manifest/storage/staging.rs
@@ -156,14 +156,14 @@ impl StagingStorage {
 
     /// Returns an iterator of manifests from staging directory.
     pub(crate) async fn manifest_lister(&self) -> Result<Option<Lister>> {
-        self.delta_storage.manifest_lister().await
+        self.delta_storage.manifest_lister(None).await
     }
 
     /// Fetch all staging manifest files and return them as (version, action_list) pairs.
     pub(crate) async fn fetch_manifests(&self) -> Result<Vec<(ManifestVersion, Vec<u8>)>> {
         let manifest_entries = self
             .delta_storage
-            .get_paths(|entry| {
+            .get_paths(None, |entry| {
                 let file_name = entry.name();
                 if is_delta_file(file_name) {
                     let version = file_version(file_name);
@@ -191,13 +191,15 @@ impl StagingStorage {
     pub(crate) async fn clear(&self) -> Result<()> {
         self.delta_storage
             .object_store()
-            .remove_all(self.delta_storage.path())
+            .delete_with(self.delta_storage.path())
+            .recursive(true)
             .await
             .context(OpenDalSnafu)?;
 
         self.blob_storage
             .object_store()
-            .remove_all(self.blob_storage.path())
+            .delete_with(self.blob_storage.path())
+            .recursive(true)
             .await
             .context(OpenDalSnafu)?;
 
diff --git a/src/mito2/src/manifest/tests/checkpoint.rs b/src/mito2/src/manifest/tests/checkpoint.rs
index 718755cd46..07f879701a 100644
--- a/src/mito2/src/manifest/tests/checkpoint.rs
+++ b/src/mito2/src/manifest/tests/checkpoint.rs
@@ -12,11 +12,15 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 use std::sync::Arc;
+use std::sync::atomic::{AtomicUsize, Ordering};
 use std::time::Duration;
 
 use common_datasource::compression::CompressionType;
+use object_store::layers::mock::{
+    Error as MockError, ErrorKind, MockLayerBuilder, OpDelete, Result as MockResult, oio,
+};
 use store_api::storage::{FileId, RegionId};
 use strum::IntoEnumIterator;
 
@@ -26,6 +30,7 @@ use crate::manifest::action::{
 };
 use crate::manifest::manager::RegionManifestManager;
 use crate::manifest::storage::checkpoint::CheckpointMetadata;
+use crate::manifest::storage::is_delta_file;
 use crate::manifest::tests::utils::basic_region_metadata;
 use crate::sst::file::FileMeta;
 use crate::test_util::TestEnv;
@@ -118,7 +123,7 @@ async fn manager_without_checkpoint() {
     let mut paths = manager
         .store()
         .delta_storage()
-        .get_paths(|e| Some(e.name().to_string()))
+        .get_paths(None, |e| Some(e.name().to_string()))
         .await
         .unwrap();
     paths.sort_unstable();
@@ -161,7 +166,7 @@ async fn manager_with_checkpoint_distance_1() {
     let mut paths = manager
         .store()
         .delta_storage()
-        .get_paths(|e| Some(e.name().to_string()))
+        .get_paths(None, |e| Some(e.name().to_string()))
         .await
         .unwrap();
     paths.sort_unstable();
@@ -257,6 +262,7 @@ async fn checkpoint_with_different_compression_types() {
             sequence: None,
             partition_expr: None,
             num_series: 0,
+            ..Default::default()
         };
         let action = RegionMetaActionList::new(vec![RegionMetaAction::Edit(RegionEdit {
             files_to_add: vec![file_meta],
@@ -323,6 +329,7 @@ fn generate_action_lists(num: usize) -> (Vec<FileId>, Vec<RegionMetaActionList>)
             sequence: None,
             partition_expr: None,
             num_series: 0,
+            ..Default::default()
         };
         let action = RegionMetaActionList::new(vec![RegionMetaAction::Edit(RegionEdit {
             files_to_add: vec![file_meta],
@@ -414,7 +421,7 @@ async fn manifest_install_manifest_to_with_checkpoint() {
     let mut paths = manager
         .store()
         .delta_storage()
-        .get_paths(|e| Some(e.name().to_string()))
+        .get_paths(None, |e| Some(e.name().to_string()))
         .await
         .unwrap();
 
@@ -489,3 +496,96 @@ async fn test_checkpoint_bypass_in_staging_mode() {
     // Checkpoint should include all 16 actions (15 from staging + 1 from writable)
     assert_eq!(last_version, 16);
 }
+
+/// A deleter that fails on `close`, simulating the S3 batch-delete failure
+/// described in issue #7986.
+struct FailingDeleter {
+    inner: oio::Deleter,
+    close_calls: Arc<AtomicUsize>,
+}
+
+impl oio::Delete for FailingDeleter {
+    async fn delete(&mut self, path: &str, args: OpDelete) -> MockResult<()> {
+        self.inner.delete(path, args).await
+    }
+
+    async fn close(&mut self) -> MockResult<()> {
+        self.close_calls.fetch_add(1, Ordering::Relaxed);
+        Err(MockError::new(
+            ErrorKind::Unexpected,
+            "mock manifest delete close failure",
+        ))
+    }
+}
+
+#[tokio::test]
+async fn checkpoint_advances_and_recovery_works_when_delete_fails() {
+    common_telemetry::init_default_ut_logging();
+
+    let close_calls = Arc::new(AtomicUsize::new(0));
+    let factory_close_calls = close_calls.clone();
+    let mock_layer = MockLayerBuilder::default()
+        .deleter_factory(Arc::new(move |inner| {
+            Box::new(FailingDeleter {
+                inner,
+                close_calls: factory_close_calls.clone(),
+            })
+        }))
+        .build()
+        .unwrap();
+
+    let env = TestEnv::new().await.with_mock_layer(mock_layer);
+    let metadata = Arc::new(basic_region_metadata());
+    let mut manager = env
+        .create_manifest_manager(CompressionType::Uncompressed, 1, Some(metadata))
+        .await
+        .unwrap()
+        .unwrap();
+
+    for _ in 0..10 {
+        manager.update(nop_action(), false).await.unwrap();
+        while manager.checkpointer().is_doing_checkpoint() {
+            tokio::time::sleep(Duration::from_millis(10)).await;
+        }
+    }
+
+    // The checkpointer must have attempted to delete stale files at least once.
+    assert!(close_calls.load(Ordering::Relaxed) > 0);
+
+    // Despite delete failures, the in-memory checkpoint marker advances so
+    // subsequent `maybe_do_checkpoint` calls compute correct ranges.
+    assert_eq!(manager.checkpointer().last_checkpoint_version(), 10);
+
+    // And the durable `_last_checkpoint` metadata reflects the latest version.
+    let (last_version, _) = manager
+        .store()
+        .load_last_checkpoint()
+        .await
+        .unwrap()
+        .expect("checkpoint should be durable");
+    assert_eq!(last_version, 10);
+
+    // Stale deltas below the checkpoint version must still be present because
+    // the mocked deleter refused them.
+    let file_names: Vec<String> = manager
+        .store()
+        .delta_storage()
+        .get_paths(None, |e| Some(e.name().to_string()))
+        .await
+        .unwrap();
+    let stale_delta_count = file_names.iter().filter(|name| is_delta_file(name)).count();
+    assert!(
+        stale_delta_count > 0,
+        "expected leftover delta files after failed delete, got {:?}",
+        file_names,
+    );
+
+    // Recovery must succeed despite the leftover deltas.
+    manager.stop().await;
+    let reopened = env
+        .create_manifest_manager(CompressionType::Uncompressed, 1, None)
+        .await
+        .unwrap()
+        .expect("manifest should be recoverable");
+    assert_eq!(reopened.manifest().manifest_version, 10);
+}
diff --git a/src/mito2/src/memtable.rs b/src/mito2/src/memtable.rs
index c39bbfa346..e1494aa47b 100644
--- a/src/mito2/src/memtable.rs
+++ b/src/mito2/src/memtable.rs
@@ -28,6 +28,7 @@ use mito_codec::key_values::KeyValue;
 pub use mito_codec::key_values::KeyValues;
 use mito_codec::row_converter::{PrimaryKeyCodec, build_primary_key_codec};
 use serde::{Deserialize, Serialize};
+use snafu::ensure;
 use store_api::metadata::RegionMetadataRef;
 use store_api::storage::{ColumnId, SequenceNumber, SequenceRange};
 
@@ -231,10 +232,17 @@ impl MemtableRanges {
 
 impl IterBuilder for MemtableRanges {
     fn build(&self, _metrics: Option<MemScanMetrics>) -> Result<BoxedBatchIterator> {
-        UnsupportedOperationSnafu {
-            err_msg: "MemtableRanges does not support build iterator",
-        }
-        .fail()
+        ensure!(
+            self.ranges.len() == 1,
+            UnsupportedOperationSnafu {
+                err_msg: format!(
+                    "Building an iterator from MemtableRanges expects 1 range, but got {}",
+                    self.ranges.len()
+                ),
+            }
+        );
+
+        self.ranges.values().next().unwrap().build_iter()
     }
 
     fn is_record_batch(&self) -> bool {
@@ -256,20 +264,6 @@ pub trait Memtable: Send + Sync + fmt::Debug {
     /// Writes an encoded batch of into memtable.
     fn write_bulk(&self, part: crate::memtable::bulk::part::BulkPart) -> Result<()>;
 
-    /// Scans the memtable.
-    /// `projection` selects columns to read, `None` means reading all columns.
-    /// `filters` are the predicates to be pushed down to memtable.
-    ///
-    /// # Note
-    /// This method should only be used for tests.
-    #[cfg(any(test, feature = "test"))]
-    fn iter(
-        &self,
-        projection: Option<&[ColumnId]>,
-        predicate: Option<table::predicate::Predicate>,
-        sequence: Option<SequenceRange>,
-    ) -> Result<BoxedBatchIterator>;
-
     /// Returns the ranges in the memtable.
     ///
     /// The returned map contains the range id and the range after applying the predicate.
@@ -427,7 +421,7 @@ impl MemtableBuilderProvider {
         let flat_format = options
             .sst_format
             .map(|format| format == FormatType::Flat)
-            .unwrap_or(self.config.default_experimental_flat_format);
+            .unwrap_or(self.config.default_flat_format);
         if flat_format {
             if options.memtable.is_some() {
                 common_telemetry::info!(
@@ -503,6 +497,8 @@ impl MemScanMetrics {
         metrics.num_rows += inner.num_rows;
         metrics.num_batches += inner.num_batches;
         metrics.scan_cost += inner.scan_cost;
+        metrics.prefilter_cost += inner.prefilter_cost;
+        metrics.prefilter_rows_filtered += inner.prefilter_rows_filtered;
     }
 
     /// Gets the metrics data.
@@ -521,6 +517,10 @@ pub(crate) struct MemScanMetricsData {
     pub(crate) num_batches: usize,
     /// Duration to scan the memtable.
     pub(crate) scan_cost: Duration,
+    /// Duration of prefilter in memtable scan.
+    pub(crate) prefilter_cost: Duration,
+    /// Number of rows filtered by prefilter in memtable scan.
+    pub(crate) prefilter_rows_filtered: usize,
 }
 
 /// Encoded range in the memtable.
@@ -543,11 +543,15 @@ pub trait IterBuilder: Send + Sync {
     }
 
     /// Returns the record batch iterator to read the range.
+    /// ## Note
+    /// Implementations should ensure the iterator yields data within given time range.
     fn build_record_batch(
         &self,
+        time_range: Option<(Timestamp, Timestamp)>,
         metrics: Option<MemScanMetrics>,
     ) -> Result<BoxedRecordBatchIterator> {
         let _metrics = metrics;
+        let _ = time_range;
         UnsupportedOperationSnafu {
             err_msg: "Record batch iterator is not supported by this memtable",
         }
@@ -706,7 +710,7 @@ impl MemtableRange {
         metrics: Option<MemScanMetrics>,
     ) -> Result<BoxedRecordBatchIterator> {
         if self.context.builder.is_record_batch() {
-            return self.context.builder.build_record_batch(metrics);
+            return self.context.builder.build_record_batch(time_range, metrics);
         }
 
         if let Some(context) = self.context.batch_to_record_batch.as_ref() {
diff --git a/src/mito2/src/memtable/bulk.rs b/src/mito2/src/memtable/bulk.rs
index cf2ced06fe..50330dd3ac 100644
--- a/src/mito2/src/memtable/bulk.rs
+++ b/src/mito2/src/memtable/bulk.rs
@@ -14,9 +14,8 @@
 
 //! Memtable implementation for bulk load
 
-#[allow(unused)]
+pub(crate) mod chunk_reader;
 pub mod context;
-#[allow(unused)]
 pub mod part;
 pub mod part_reader;
 mod row_group_reader;
@@ -34,6 +33,7 @@ fn env_usize(name: &str, default: usize) -> usize {
         .unwrap_or(default)
 }
 
+use common_time::Timestamp;
 use datatypes::arrow::datatypes::SchemaRef;
 use mito_codec::key_values::KeyValue;
 use rayon::prelude::*;
@@ -46,6 +46,7 @@ use crate::flush::WriteBufferManagerRef;
 use crate::memtable::bulk::context::BulkIterContext;
 use crate::memtable::bulk::part::{
     BulkPart, BulkPartEncodeMetrics, BulkPartEncoder, MultiBulkPart, UnorderedPart,
+    should_prune_bulk_part,
 };
 use crate::memtable::bulk::part_reader::BulkPartBatchIter;
 use crate::memtable::stats::WriteMetrics;
@@ -57,7 +58,7 @@ use crate::memtable::{
 use crate::read::flat_dedup::{FlatDedupIterator, FlatLastNonNull, FlatLastRow};
 use crate::read::flat_merge::FlatMergeIterator;
 use crate::region::options::MergeMode;
-use crate::sst::parquet::format::FIXED_POS_COLUMN_NUM;
+use crate::sst::parquet::flat_format::field_column_start;
 use crate::sst::parquet::{DEFAULT_READ_BATCH_SIZE, DEFAULT_ROW_GROUP_SIZE};
 use crate::sst::{FlatSchemaOptions, to_flat_sst_arrow_schema};
 
@@ -137,7 +138,7 @@ struct CollectedParts {
 /// All parts in a bulk memtable.
 #[derive(Default)]
 struct BulkParts {
-    /// Unordered small parts (< 1024 rows).
+    /// Unordered small parts.
     unordered_part: UnorderedPart,
     /// All parts (raw and encoded).
     parts: Vec<BulkPartWrapper>,
@@ -462,16 +463,6 @@ impl Memtable for BulkMemtable {
         Ok(())
     }
 
-    #[cfg(any(test, feature = "test"))]
-    fn iter(
-        &self,
-        _projection: Option<&[ColumnId]>,
-        _predicate: Option<table::predicate::Predicate>,
-        _sequence: Option<SequenceRange>,
-    ) -> Result<crate::memtable::BoxedBatchIterator> {
-        todo!()
-    }
-
     fn ranges(
         &self,
         projection: Option<&[ColumnId]>,
@@ -802,8 +793,14 @@ impl IterBuilder for BulkRangeIterBuilder {
 
     fn build_record_batch(
         &self,
+        _time_range: Option<(Timestamp, Timestamp)>,
         metrics: Option<MemScanMetrics>,
     ) -> Result<BoxedRecordBatchIterator> {
+        let metadata = self.context.read_format().metadata();
+        if should_prune_bulk_part(&self.part.batch, &self.context, metadata) {
+            return Ok(Box::new(std::iter::empty()));
+        }
+
         let series_count = self.part.estimated_series_count();
         let iter = BulkPartBatchIter::from_single(
             self.part.batch.clone(),
@@ -835,16 +832,17 @@ impl IterBuilder for MultiBulkRangeIterBuilder {
 
     fn build_record_batch(
         &self,
+        _time_range: Option<(Timestamp, Timestamp)>,
         metrics: Option<MemScanMetrics>,
     ) -> Result<BoxedRecordBatchIterator> {
-        self.part
+        match self
+            .part
             .read(self.context.clone(), self.sequence, metrics)?
-            .ok_or_else(|| {
-                UnsupportedOperationSnafu {
-                    err_msg: "Failed to create iterator for multi bulk part",
-                }
-                .build()
-            })
+        {
+            Some(iter) => Ok(iter),
+            // All batches were pruned by the predicate. Return an empty iterator.
+            None => Ok(Box::new(std::iter::empty())),
+        }
     }
 
     fn encoded_range(&self) -> Option<EncodedRange> {
@@ -874,6 +872,7 @@ impl IterBuilder for EncodedBulkRangeIterBuilder {
 
     fn build_record_batch(
         &self,
+        _time_range: Option<(Timestamp, Timestamp)>,
         metrics: Option<MemScanMetrics>,
     ) -> Result<BoxedRecordBatchIterator> {
         if let Some(iter) = self
@@ -1186,13 +1185,8 @@ impl MemtableCompactor {
                     Box::new(dedup_iter)
                 }
                 MergeMode::LastNonNull => {
-                    // Calculates field column start: total columns - fixed columns - field columns
-                    // Field column count = total metadata columns - time index column - primary key columns
-                    let field_column_count =
-                        metadata.column_metadatas.len() - 1 - metadata.primary_key.len();
-                    let total_columns = arrow_schema.fields().len();
                     let field_column_start =
-                        total_columns - FIXED_POS_COLUMN_NUM - field_column_count;
+                        field_column_start(metadata, arrow_schema.fields().len());
 
                     let dedup_iter = FlatDedupIterator::new(
                         merged_iter,
@@ -1244,6 +1238,7 @@ impl MemtableCompactor {
                 max_timestamp,
                 max_sequence,
                 estimated_series_count,
+                metadata,
             );
 
             common_telemetry::trace!(
@@ -2220,4 +2215,57 @@ mod tests {
         }
         assert_eq!(expected_rows, total_rows_read);
     }
+
+    #[test]
+    fn test_multi_bulk_range_iter_builder_all_pruned() {
+        let metadata = metadata_for_test();
+        let merge_threshold = 8;
+        let config = BulkMemtableConfig {
+            merge_threshold,
+            ..Default::default()
+        };
+        let memtable = BulkMemtable::new(
+            2006,
+            config,
+            metadata.clone(),
+            None,
+            None,
+            false,
+            MergeMode::LastRow,
+        );
+        memtable.set_unordered_part_threshold(0);
+
+        // Write enough bulk parts to trigger merge into MultiBulkPart.
+        for i in 0..merge_threshold {
+            let part = create_bulk_part_with_converter(
+                &format!("key_{}", i),
+                i as u32,
+                vec![1000 + i as i64 * 100, 2000 + i as i64 * 100],
+                vec![Some(i as f64 * 10.0), Some(i as f64 * 10.0 + 1.0)],
+                100 + i as u64,
+            )
+            .unwrap();
+            memtable.write_bulk(part).unwrap();
+        }
+        memtable.compact(false).unwrap();
+
+        // Use a predicate that matches no rows so all batches are pruned.
+        let filter = datafusion_expr::col("k0").eq(datafusion_expr::lit("nonexistent"));
+        let predicate_group = PredicateGroup::new(&metadata, &[filter]).unwrap();
+        let ranges = memtable
+            .ranges(
+                None,
+                RangesOptions::default().with_predicate(predicate_group),
+            )
+            .unwrap();
+
+        // Should return ranges but each range should produce an empty iterator
+        // instead of an error.
+        for (_range_id, range) in ranges.ranges.iter() {
+            assert!(range.is_record_batch());
+            let record_batch_iter = range.build_record_batch_iter(None, None).unwrap();
+            let total_rows: usize = record_batch_iter.map(|r| r.unwrap().num_rows()).sum();
+            assert_eq!(0, total_rows);
+        }
+    }
 }
diff --git a/src/mito2/src/memtable/bulk/chunk_reader.rs b/src/mito2/src/memtable/bulk/chunk_reader.rs
new file mode 100644
index 0000000000..e632cd1b37
--- /dev/null
+++ b/src/mito2/src/memtable/bulk/chunk_reader.rs
@@ -0,0 +1,65 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! ChunkReader implementation for in-memory parquet bytes.
+
+use std::io::Cursor;
+
+use bytes::Bytes;
+use parquet::errors::{ParquetError, Result};
+use parquet::file::reader::{ChunkReader, Length};
+
+/// A [ChunkReader] implementation for in-memory parquet bytes.
+///
+/// This provides byte access to parquet data stored in memory (Bytes),
+/// used for reading parquet data from bulk memtable.
+#[derive(Clone)]
+pub struct MemtableChunkReader {
+    /// The in-memory parquet data.
+    data: Bytes,
+}
+
+impl MemtableChunkReader {
+    /// Creates a new [MemtableChunkReader] from the given bytes.
+    pub fn new(data: Bytes) -> Self {
+        Self { data }
+    }
+}
+
+impl Length for MemtableChunkReader {
+    fn len(&self) -> u64 {
+        self.data.len() as u64
+    }
+}
+
+impl ChunkReader for MemtableChunkReader {
+    type T = Cursor<Bytes>;
+
+    fn get_read(&self, start: u64) -> Result<Self::T> {
+        let start = start as usize;
+        if start > self.data.len() {
+            return Err(ParquetError::IndexOutOfBound(start, self.data.len()));
+        }
+        Ok(Cursor::new(self.data.slice(start..)))
+    }
+
+    fn get_bytes(&self, start: u64, length: usize) -> Result<Bytes> {
+        let start = start as usize;
+        let end = start + length;
+        if end > self.data.len() {
+            return Err(ParquetError::IndexOutOfBound(end, self.data.len()));
+        }
+        Ok(self.data.slice(start..end))
+    }
+}
diff --git a/src/mito2/src/memtable/bulk/context.rs b/src/mito2/src/memtable/bulk/context.rs
index c3274d30e9..3ac8e009fe 100644
--- a/src/mito2/src/memtable/bulk/context.rs
+++ b/src/mito2/src/memtable/bulk/context.rs
@@ -17,17 +17,18 @@
 use std::collections::VecDeque;
 use std::sync::Arc;
 
-use mito_codec::row_converter::{DensePrimaryKeyCodec, build_primary_key_codec};
+use common_recordbatch::filter::SimpleFilterEvaluator;
+use mito_codec::row_converter::build_primary_key_codec;
 use parquet::file::metadata::ParquetMetaData;
 use store_api::metadata::RegionMetadataRef;
 use store_api::storage::ColumnId;
 use table::predicate::Predicate;
 
 use crate::error::Result;
+use crate::read::read_columns::ReadColumns;
 use crate::sst::parquet::file_range::{PreFilterMode, RangeBase};
 use crate::sst::parquet::flat_format::FlatReadFormat;
-use crate::sst::parquet::format::ReadFormat;
-use crate::sst::parquet::reader::SimpleFilterContext;
+use crate::sst::parquet::prefilter::{CachedPrimaryKeyFilter, build_bulk_filter_plan};
 use crate::sst::parquet::stats::RowGroupPruningStats;
 
 pub(crate) type BulkIterContextRef = Arc<BulkIterContext>;
@@ -35,6 +36,9 @@ pub(crate) type BulkIterContextRef = Arc<BulkIterContext>;
 pub struct BulkIterContext {
     pub(crate) base: RangeBase,
     pub(crate) predicate: Option<Predicate>,
+    /// Pre-extracted primary key filters for PK prefiltering.
+    /// `None` if PK prefiltering is not applicable.
+    pk_filters: Option<Arc<Vec<SimpleFilterEvaluator>>>,
 }
 
 impl BulkIterContext {
@@ -62,21 +66,19 @@ impl BulkIterContext {
     ) -> Result<Self> {
         let codec = build_primary_key_codec(&region_metadata);
 
-        let simple_filters = predicate
-            .as_ref()
-            .iter()
-            .flat_map(|predicate| {
-                predicate
-                    .exprs()
+        let read_cols = if let Some(col_ids) = projection {
+            ReadColumns::from_deduped_column_ids(col_ids.iter().copied())
+        } else {
+            ReadColumns::from_deduped_column_ids(
+                region_metadata
+                    .column_metadatas
                     .iter()
-                    .filter_map(|expr| SimpleFilterContext::new_opt(&region_metadata, None, expr))
-            })
-            .collect();
-
-        let read_format = ReadFormat::new(
+                    .map(|col| col.column_id),
+            )
+        };
+        let read_format = FlatReadFormat::new(
             region_metadata.clone(),
-            projection,
-            true,
+            read_cols,
             None,
             "memtable",
             skip_auto_convert,
@@ -87,9 +89,11 @@ impl BulkIterContext {
             .map(|pred| pred.dyn_filters().as_ref().clone())
             .unwrap_or_default();
 
+        let filter_plan = build_bulk_filter_plan(&read_format, predicate.as_ref());
+
         Ok(Self {
             base: RangeBase {
-                filters: simple_filters,
+                filters: filter_plan.remaining_simple_filters,
                 dyn_filters,
                 read_format,
                 prune_schema: region_metadata.schema.clone(),
@@ -102,6 +106,7 @@ impl BulkIterContext {
                 partition_filter: None,
             },
             predicate,
+            pk_filters: filter_plan.pk_filters,
         })
     }
 
@@ -133,7 +138,20 @@ impl BulkIterContext {
         }
     }
 
-    pub(crate) fn read_format(&self) -> &ReadFormat {
+    /// Builds a fresh PK filter for a new iterator. Returns `None` if PK
+    /// prefiltering is not applicable.
+    pub(crate) fn build_pk_filter(&self) -> Option<CachedPrimaryKeyFilter> {
+        let pk_filters = self.pk_filters.as_ref()?;
+        let metadata = self.base.read_format.metadata();
+        // Parquet PK prefilter always supports the partition column.
+        let inner = self
+            .base
+            .codec
+            .primary_key_filter(metadata, Arc::clone(pk_filters), false);
+        Some(CachedPrimaryKeyFilter::new(inner))
+    }
+
+    pub(crate) fn read_format(&self) -> &FlatReadFormat {
         &self.base.read_format
     }
 
diff --git a/src/mito2/src/memtable/bulk/part.rs b/src/mito2/src/memtable/bulk/part.rs
index 71e49776c0..2b817dcb3a 100644
--- a/src/mito2/src/memtable/bulk/part.rs
+++ b/src/mito2/src/memtable/bulk/part.rs
@@ -14,66 +14,63 @@
 
 //! Bulk part encoder/decoder.
 
-use std::collections::{HashMap, HashSet, VecDeque};
+use std::collections::{HashMap, HashSet};
 use std::sync::Arc;
 use std::time::{Duration, Instant};
 
 use api::helper::{ColumnDataTypeWrapper, to_grpc_value};
 use api::v1::bulk_wal_entry::Body;
-use api::v1::{ArrowIpc, BulkWalEntry, Mutation, OpType, bulk_wal_entry};
+use api::v1::{ArrowIpc, BulkWalEntry, Mutation, OpType};
 use bytes::Bytes;
 use common_grpc::flight::{FlightDecoder, FlightEncoder, FlightMessage};
 use common_recordbatch::DfRecordBatch as RecordBatch;
 use common_time::Timestamp;
-use common_time::timestamp::TimeUnit;
+use datafusion_common::Column;
+use datafusion_common::pruning::PruningStatistics;
+use datafusion_expr::utils::expr_to_columns;
 use datatypes::arrow;
 use datatypes::arrow::array::{
-    Array, ArrayRef, BinaryBuilder, BinaryDictionaryBuilder, DictionaryArray, StringBuilder,
-    StringDictionaryBuilder, TimestampMicrosecondArray, TimestampMillisecondArray,
-    TimestampNanosecondArray, TimestampSecondArray, UInt8Array, UInt8Builder, UInt32Array,
-    UInt64Array, UInt64Builder,
+    Array, ArrayRef, BinaryArray, BooleanArray, StringDictionaryBuilder, UInt8Array, UInt64Array,
 };
-use datatypes::arrow::compute::{SortColumn, SortOptions, TakeOptions};
+use datatypes::arrow::compute::{SortColumn, SortOptions};
 use datatypes::arrow::datatypes::{
     DataType as ArrowDataType, Field, Schema, SchemaRef, UInt32Type,
 };
-use datatypes::arrow_array::BinaryArray;
 use datatypes::data_type::DataType;
-use datatypes::prelude::{MutableVector, ScalarVectorBuilder, Vector};
-use datatypes::value::{Value, ValueRef};
+use datatypes::extension::json::is_json_extension_type;
+use datatypes::prelude::{MutableVector, Vector};
+use datatypes::schema::ext::ArrowSchemaExt;
+use datatypes::types::JsonType;
+use datatypes::value::ValueRef;
 use datatypes::vectors::Helper;
-use mito_codec::key_values::{KeyValue, KeyValues, KeyValuesRef};
-use mito_codec::row_converter::{
-    DensePrimaryKeyCodec, PrimaryKeyCodec, PrimaryKeyCodecExt, build_primary_key_codec,
-};
+use datatypes::vectors::json::array::JsonArray;
+use mito_codec::key_values::{KeyValue, KeyValues};
+use mito_codec::row_converter::{PrimaryKeyCodec, SortField, build_primary_key_codec_with_fields};
 use parquet::arrow::ArrowWriter;
 use parquet::basic::{Compression, ZstdLevel};
-use parquet::data_type::AsBytes;
 use parquet::file::metadata::ParquetMetaData;
 use parquet::file::properties::WriterProperties;
 use smallvec::SmallVec;
-use snafu::{OptionExt, ResultExt, Snafu};
+use snafu::{OptionExt, ResultExt};
 use store_api::codec::PrimaryKeyEncoding;
-use store_api::metadata::{ColumnMetadata, RegionMetadata, RegionMetadataRef};
+use store_api::metadata::{RegionMetadata, RegionMetadataRef};
 use store_api::storage::consts::PRIMARY_KEY_COLUMN_NAME;
-use store_api::storage::{FileId, RegionId, SequenceNumber, SequenceRange};
-use table::predicate::Predicate;
+use store_api::storage::{ColumnId, FileId, SequenceNumber, SequenceRange};
 
 use crate::error::{
-    self, ColumnNotFoundSnafu, ComputeArrowSnafu, ConvertColumnDataTypeSnafu, CreateDefaultSnafu,
+    self, ColumnNotFoundSnafu, ComputeArrowSnafu, ConvertValueSnafu, CreateDefaultSnafu,
     DataTypeMismatchSnafu, EncodeMemtableSnafu, EncodeSnafu, InvalidMetadataSnafu,
-    InvalidRequestSnafu, NewRecordBatchSnafu, Result, UnexpectedSnafu,
+    InvalidRequestSnafu, NewRecordBatchSnafu, Result,
 };
-use crate::memtable::bulk::context::BulkIterContextRef;
+use crate::memtable::bulk::context::{BulkIterContext, BulkIterContextRef};
 use crate::memtable::bulk::part_reader::EncodedBulkPartIter;
 use crate::memtable::time_series::{ValueBuilder, Values};
 use crate::memtable::{BoxedRecordBatchIterator, MemScanMetrics, MemtableStats};
+use crate::sst::SeriesEstimator;
 use crate::sst::index::IndexOutput;
-use crate::sst::parquet::file_range::{PreFilterMode, row_group_contains_delete};
 use crate::sst::parquet::flat_format::primary_key_column_index;
-use crate::sst::parquet::format::{PrimaryKeyArray, PrimaryKeyArrayBuilder, ReadFormat};
+use crate::sst::parquet::format::{PrimaryKeyArray, PrimaryKeyArrayBuilder};
 use crate::sst::parquet::{PARQUET_METADATA_KEY, SstInfo};
-use crate::sst::{SeriesEstimator, to_sst_arrow_schema};
 
 const INIT_DICT_VALUE_CAPACITY: usize = 8;
 
@@ -438,11 +435,13 @@ impl UnorderedPart {
 
         // Get the schema from the first part
         let schema = self.parts[0].batch.schema();
-
-        // Concatenate all record batches
-        let batches: Vec<RecordBatch> = self.parts.iter().map(|p| p.batch.clone()).collect();
-        let concatenated =
-            arrow::compute::concat_batches(&schema, &batches).context(ComputeArrowSnafu)?;
+        let concatenated = if schema.has_json_extension_field() {
+            let (schema, batches) = align_parts(&self.parts)?;
+            arrow::compute::concat_batches(&schema, &batches).context(ComputeArrowSnafu)?
+        } else {
+            arrow::compute::concat_batches(&schema, self.parts.iter().map(|x| &x.batch))
+                .context(ComputeArrowSnafu)?
+        };
 
         // Sort the concatenated batch
         let sorted_batch = sort_primary_key_record_batch(&concatenated)?;
@@ -479,6 +478,73 @@ impl UnorderedPart {
     }
 }
 
+/// Align the JSON columns in [BulkPart]s, to unified Arrow arrays. So that we can compute (concat,
+/// sort, etc.) on them.
+fn align_parts(parts: &[BulkPart]) -> Result<(SchemaRef, Vec<RecordBatch>)> {
+    debug_assert!(
+        !parts.is_empty()
+            && parts
+                .windows(2)
+                .all(|w| w[0].batch.schema_ref().fields().len()
+                    == w[1].batch.schema_ref().fields().len())
+    );
+
+    let first = &parts[0];
+    let base_schema = first.batch.schema_ref();
+    let rest = &parts[1..];
+
+    let mut merged_types = HashMap::new();
+    let mut aligned_fields = Vec::with_capacity(base_schema.fields().len());
+    for (i, field) in base_schema.fields().iter().enumerate() {
+        if is_json_extension_type(field) {
+            let mut merged = JsonType::from(field.data_type());
+            rest.iter()
+                .try_fold(&mut merged, |acc, x| {
+                    acc.merge(&JsonType::from(x.batch.schema_ref().field(i).data_type()))?;
+                    Ok(acc)
+                })
+                .context(DataTypeMismatchSnafu)?;
+            merged_types.insert(i, merged.as_arrow_type());
+
+            aligned_fields.push(Arc::new(
+                Field::new(
+                    field.name().clone(),
+                    merged.as_arrow_type(),
+                    field.is_nullable(),
+                )
+                .with_metadata(field.metadata().clone()),
+            ));
+        } else {
+            aligned_fields.push(field.clone())
+        };
+    }
+    let aligned_schema = Arc::new(Schema::new_with_metadata(
+        aligned_fields,
+        base_schema.metadata().clone(),
+    ));
+
+    let mut aligned_batches = Vec::with_capacity(parts.len());
+    for part in parts {
+        let mut columns = Vec::with_capacity(part.batch.num_columns());
+        for (i, column) in part.batch.columns().iter().enumerate() {
+            if let Some(expect) = merged_types.get(&i) {
+                columns.push(
+                    JsonArray::from(column)
+                        .try_align(expect)
+                        .context(ConvertValueSnafu)?,
+                );
+            } else {
+                columns.push(column.clone());
+            }
+        }
+        aligned_batches.push(
+            RecordBatch::try_new(aligned_schema.clone(), columns).context(NewRecordBatchSnafu)?,
+        );
+    }
+
+    Ok((aligned_schema, aligned_batches))
+}
+
 /// More accurate estimation of the size of a record batch.
 pub fn record_batch_estimated_size(batch: &RecordBatch) -> usize {
     batch
@@ -527,8 +593,6 @@ impl PrimaryKeyColumnBuilder {
 
 /// Converter that converts structs into [BulkPart].
 pub struct BulkPartConverter {
-    /// Region metadata.
-    region_metadata: RegionMetadataRef,
     /// Schema of the converted batch.
     schema: SchemaRef,
     /// Primary key codec for encoding keys
@@ -577,7 +641,6 @@ impl BulkPartConverter {
         };
 
         Self {
-            region_metadata: region_metadata.clone(),
             schema,
             primary_key_codec,
             key_buf: Vec::new(),
@@ -693,7 +756,10 @@ impl BulkPartConverter {
         columns.push(values.sequence.to_arrow_array());
         columns.push(values.op_type.to_arrow_array());
 
-        let batch = RecordBatch::try_new(self.schema, columns).context(NewRecordBatchSnafu)?;
+        // The actual datatype of JSON array is data oriented, not to be derived from the Region
+        // metadata, which is static. So here we have to align the schema.
+        let schema = align_schema_with_json_array(self.schema, &columns);
+        let batch = RecordBatch::try_new(schema, columns).context(NewRecordBatchSnafu)?;
         // Sorts the record batch.
         let batch = sort_primary_key_record_batch(&batch)?;
 
@@ -708,6 +774,26 @@ impl BulkPartConverter {
     }
 }
 
+fn align_schema_with_json_array(schema: SchemaRef, columns: &[ArrayRef]) -> SchemaRef {
+    if schema.fields().iter().all(|f| !is_json_extension_type(f)) {
+        return schema;
+    }
+
+    let mut fields = Vec::with_capacity(schema.fields().len());
+    for (field, array) in schema.fields().iter().zip(columns) {
+        if !is_json_extension_type(field) {
+            fields.push(field.clone());
+            continue;
+        }
+
+        let mut field = field.as_ref().clone();
+        field.set_data_type(array.data_type().clone());
+        fields.push(Arc::new(field));
+    }
+
+    Arc::new(Schema::new_with_metadata(fields, schema.metadata().clone()))
+}
+
 fn new_primary_key_column_builders(
     metadata: &RegionMetadata,
     capacity: usize,
@@ -967,7 +1053,7 @@ impl EncodedBulkPart {
         Self { data, metadata }
     }
 
-    pub(crate) fn metadata(&self) -> &BulkPartMeta {
+    pub fn metadata(&self) -> &BulkPartMeta {
         &self.metadata
     }
 
@@ -977,7 +1063,7 @@ impl EncodedBulkPart {
     }
 
     /// Returns the encoded data.
-    pub(crate) fn data(&self) -> &Bytes {
+    pub fn data(&self) -> &Bytes {
         &self.data
     }
 
@@ -1042,9 +1128,8 @@ impl EncodedBulkPart {
         sequence: Option<SequenceRange>,
         mem_scan_metrics: Option<MemScanMetrics>,
     ) -> Result<Option<BoxedRecordBatchIterator>> {
-        // Compute skip_fields for row group pruning using the same approach as compute_skip_fields in reader.rs.
-        let skip_fields_for_pruning =
-            Self::compute_skip_fields(context.pre_filter_mode(), &self.metadata.parquet_metadata);
+        // Compute skip_fields for row group pruning from the configured pre-filter mode.
+        let skip_fields_for_pruning = context.pre_filter_mode().skip_fields();
 
         // use predicate to find row groups to read.
         let row_groups_to_read =
@@ -1064,20 +1149,6 @@ impl EncodedBulkPart {
         )?;
         Ok(Some(Box::new(iter) as BoxedRecordBatchIterator))
     }
-
-    /// Computes whether to skip field columns based on PreFilterMode.
-    fn compute_skip_fields(pre_filter_mode: PreFilterMode, parquet_meta: &ParquetMetaData) -> bool {
-        match pre_filter_mode {
-            PreFilterMode::All => false,
-            PreFilterMode::SkipFields => true,
-            PreFilterMode::SkipFieldsOnDelete => {
-                // Check if any row group contains delete op
-                (0..parquet_meta.num_row_groups()).any(|rg_idx| {
-                    row_group_contains_delete(parquet_meta, rg_idx, "memtable").unwrap_or(true)
-                })
-            }
-        }
-    }
 }
 
 // TODO(yingwen): max_sequence
@@ -1116,15 +1187,11 @@ pub struct BulkPartEncodeMetrics {
 
 pub struct BulkPartEncoder {
     metadata: RegionMetadataRef,
-    row_group_size: usize,
     writer_props: Option<WriterProperties>,
 }
 
 impl BulkPartEncoder {
-    pub(crate) fn new(
-        metadata: RegionMetadataRef,
-        row_group_size: usize,
-    ) -> Result<BulkPartEncoder> {
+    pub fn new(metadata: RegionMetadataRef, row_group_size: usize) -> Result<BulkPartEncoder> {
         // TODO(yingwen): Skip arrow schema if needed.
         let json = metadata.to_json().context(InvalidMetadataSnafu)?;
         let key_value_meta =
@@ -1144,7 +1211,6 @@ impl BulkPartEncoder {
 
         Ok(Self {
             metadata,
-            row_group_size,
             writer_props,
         })
     }
@@ -1185,7 +1251,6 @@ impl BulkPartEncoder {
             iter_start = Instant::now();
         }
         metrics.iter_cost += iter_start.elapsed();
-        iter_start = Instant::now();
 
         if total_rows == 0 {
             return Ok(None);
@@ -1216,7 +1281,7 @@ impl BulkPartEncoder {
     }
 
     /// Encodes bulk part to a [EncodedBulkPart], returns the encoded data.
-    fn encode_part(&self, part: &BulkPart) -> Result<Option<EncodedBulkPart>> {
+    pub fn encode_part(&self, part: &BulkPart) -> Result<Option<EncodedBulkPart>> {
         if part.batch.num_rows() == 0 {
             return Ok(None);
         }
@@ -1250,6 +1315,194 @@ impl BulkPartEncoder {
     }
 }
 
+/// Per-batch min/max statistics for the first tag column in a `MultiBulkPart`.
+///
+/// Since batches are sorted by primary key, we can extract the min/max of the first tag
+/// from the first/last row's encoded primary key in each batch. These statistics enable
+/// batch-level pruning using predicates, analogous to row-group pruning in parquet.
+#[derive(Debug, Clone)]
+struct BatchStats {
+    /// Number of batches.
+    num_batches: usize,
+    /// Column id of the first tag.
+    first_tag_id: ColumnId,
+    /// Min values of the first tag, one element per batch.
+    min_values: ArrayRef,
+    /// Max values of the first tag, one element per batch.
+    max_values: ArrayRef,
+}
+
+impl BatchStats {
+    /// Computes batch statistics from a slice of record batches.
+    ///
+    /// Returns `None` if there is no primary key (no first tag to collect stats for)
+    /// or if extracting statistics fails.
+    fn compute(batches: &[RecordBatch], metadata: &RegionMetadata) -> Option<Self> {
+        // `primary_key.first()` is correct for both dense and sparse encodings.
+        // For dense, values follow the order of `metadata.primary_key`.
+        // For sparse, `decode_leftmost` decodes the first value which also
+        // corresponds to `primary_key.first()`. See `SparsePrimaryKeyCodec` for format details.
+        let first_tag_id = *metadata.primary_key.first()?;
+        let first_tag_column = metadata.column_by_id(first_tag_id)?;
+        let data_type = &first_tag_column.column_schema.data_type;
+
+        let converter = build_primary_key_codec_with_fields(
+            metadata.primary_key_encoding,
+            [(first_tag_id, SortField::new(data_type.clone()))].into_iter(),
+        );
+        let pk_index = primary_key_column_index(batches.first()?.num_columns());
+
+        let mut min_builder = data_type.create_mutable_vector(batches.len());
+        let mut max_builder = data_type.create_mutable_vector(batches.len());
+
+        for batch in batches {
+            match Self::extract_first_tag_bounds(batch, pk_index, &*converter) {
+                Some((min_val, max_val)) => {
+                    min_builder.push_value_ref(&min_val.as_value_ref());
+                    max_builder.push_value_ref(&max_val.as_value_ref());
+                }
+                None => {
+                    min_builder.push_null();
+                    max_builder.push_null();
+                }
+            }
+        }
+
+        Some(Self {
+            num_batches: batches.len(),
+            first_tag_id,
+            min_values: min_builder.to_vector().to_arrow_array(),
+            max_values: max_builder.to_vector().to_arrow_array(),
+        })
+    }
+
+    /// Extracts the first tag value from the first and last rows of a batch.
+    fn extract_first_tag_bounds(
+        batch: &RecordBatch,
+        pk_index: usize,
+        converter: &dyn PrimaryKeyCodec,
+    ) -> Option<(datatypes::value::Value, datatypes::value::Value)> {
+        if batch.num_rows() == 0 {
+            return None;
+        }
+
+        let pk_dict = batch
+            .column(pk_index)
+            .as_any()
+            .downcast_ref::<PrimaryKeyArray>()?;
+        let pk_values = pk_dict.values().as_any().downcast_ref::<BinaryArray>()?;
+
+        let keys = pk_dict.keys();
+        let min_key = keys.value(0);
+        let max_key = keys.value(batch.num_rows() - 1);
+        let min_bytes = pk_values.value(min_key as usize);
+        let max_bytes = pk_values.value(max_key as usize);
+
+        Some((
+            converter.decode_leftmost(min_bytes).ok()??,
+            converter.decode_leftmost(max_bytes).ok()??,
+        ))
+    }
+}
+
+/// Adapter implementing `PruningStatistics` for `BatchStats`.
+///
+/// Used with `Predicate::prune_with_stats()` to skip batches whose first-tag
+/// min/max range does not match the query predicate.
+struct BatchPruningStats<'a> {
+    stats: &'a BatchStats,
+    metadata: &'a RegionMetadataRef,
+}
+
+impl PruningStatistics for BatchPruningStats<'_> {
+    fn min_values(&self, column: &Column) -> Option<ArrayRef> {
+        let col = self.metadata.column_by_name(&column.name)?;
+        if col.column_id == self.stats.first_tag_id {
+            Some(self.stats.min_values.clone())
+        } else {
+            None
+        }
+    }
+
+    fn max_values(&self, column: &Column) -> Option<ArrayRef> {
+        let col = self.metadata.column_by_name(&column.name)?;
+        if col.column_id == self.stats.first_tag_id {
+            Some(self.stats.max_values.clone())
+        } else {
+            None
+        }
+    }
+
+    fn num_containers(&self) -> usize {
+        self.stats.num_batches
+    }
+
+    fn null_counts(&self, _column: &Column) -> Option<ArrayRef> {
+        None
+    }
+
+    fn row_counts(&self, _column: &Column) -> Option<ArrayRef> {
+        None
+    }
+
+    fn contained(
+        &self,
+        _column: &Column,
+        _values: &std::collections::HashSet<datafusion_common::ScalarValue>,
+    ) -> Option<BooleanArray> {
+        None
+    }
+}
+
+/// Returns true if the predicate references the given column name.
+fn predicate_references_column(predicate: &table::predicate::Predicate, column_name: &str) -> bool {
+    let mut columns = HashSet::new();
+    for expr in predicate.exprs() {
+        let _ = expr_to_columns(expr, &mut columns);
+    }
+    columns.iter().any(|col| col.name == column_name)
+}
+
+/// Returns true if the batch should be pruned (skipped) based on the first-tag min/max
+/// statistics and the predicate in the context. Returns false if no pruning is possible
+/// (no primary key, no predicate, or the batch matches the predicate).
+pub(crate) fn should_prune_bulk_part(
+    batch: &RecordBatch,
+    context: &BulkIterContext,
+    metadata: &RegionMetadata,
+) -> bool {
+    let predicate = match &context.predicate {
+        Some(p) => p,
+        None => return false,
+    };
+    // Check if the predicate references the first tag column to avoid computing
+    // expensive batch statistics when they won't help with pruning.
+    let first_tag_id = match metadata.primary_key.first() {
+        Some(id) => *id,
+        None => return false,
+    };
+    // Safety: `first_tag_id` comes from `metadata.primary_key` so the column always exists.
+    let first_tag_name = &metadata
+        .column_by_id(first_tag_id)
+        .unwrap()
+        .column_schema
+        .name;
+    if !predicate_references_column(predicate, first_tag_name) {
+        return false;
+    }
+    let stats = match BatchStats::compute(std::slice::from_ref(batch), metadata) {
+        Some(s) => s,
+        None => return false,
+    };
+    let region_meta = context.read_format().metadata();
+    let pruning_stats = BatchPruningStats {
+        stats: &stats,
+        metadata: region_meta,
+    };
+    let mask = predicate.prune_with_stats(&pruning_stats, region_meta.schema.arrow_schema());
+    !mask.first().copied().unwrap_or(true)
+}
+
 /// A collection of ordered RecordBatches representing a bulk part without parquet encoding.
 ///
 /// Similar to `EncodedBulkPart` but stores raw RecordBatches instead of encoded parquet data.
@@ -1269,13 +1522,17 @@ pub struct MultiBulkPart {
     max_sequence: SequenceNumber,
     /// Number of series.
     series_count: usize,
+    /// Pre-computed per-batch statistics for the first tag column.
+    /// `None` if there is no primary key.
+    batch_stats: Option<BatchStats>,
 }
 
 impl MultiBulkPart {
     /// Creates a new MultiBulkPart from a single BulkPart.
-    pub fn from_bulk_part(part: BulkPart) -> Self {
+    pub fn from_bulk_part(part: BulkPart, metadata: &RegionMetadata) -> Self {
         let num_rows = part.num_rows();
         let series_count = part.estimated_series_count();
+        let batch_stats = BatchStats::compute(std::slice::from_ref(&part.batch), metadata);
         let mut batches = SmallVec::new();
         batches.push(part.batch);
 
@@ -1286,6 +1543,7 @@ impl MultiBulkPart {
             min_timestamp: part.min_timestamp,
             max_sequence: part.sequence,
             series_count,
+            batch_stats,
         }
     }
 
@@ -1297,6 +1555,7 @@ impl MultiBulkPart {
     /// * `max_timestamp` - Maximum timestamp across all batches
     /// * `max_sequence` - Maximum sequence number across all batches
     /// * `series_count` - Number of series in the batches
+    /// * `metadata` - Region metadata for computing batch statistics
     ///
     /// # Panics
     /// Panics if batches is empty.
@@ -1306,10 +1565,12 @@ impl MultiBulkPart {
         max_timestamp: i64,
         max_sequence: SequenceNumber,
         series_count: usize,
+        metadata: &RegionMetadata,
     ) -> Self {
         assert!(!batches.is_empty(), "batches must not be empty");
 
         let total_rows = batches.iter().map(|b| b.num_rows()).sum();
+        let batch_stats = BatchStats::compute(&batches, metadata);
 
         Self {
             batches: SmallVec::from_vec(batches),
@@ -1318,6 +1579,7 @@ impl MultiBulkPart {
             min_timestamp,
             max_sequence,
             series_count,
+            batch_stats,
         }
     }
 
@@ -1351,17 +1613,16 @@ impl MultiBulkPart {
         self.batches.len()
     }
 
-    /// Returns an iterator over the record batches.
-    pub(crate) fn batches(&self) -> impl Iterator<Item = &RecordBatch> {
-        self.batches.iter()
-    }
-
     /// Returns the estimated memory size of all batches.
     pub(crate) fn estimated_size(&self) -> usize {
         self.batches.iter().map(record_batch_estimated_size).sum()
     }
 
     /// Reads data from this part with the given context and filters.
+    ///
+    /// If batch-level statistics are available and a predicate is set, prunes
+    /// batches whose first-tag min/max range doesn't match the predicate before
+    /// creating the iterator.
     pub(crate) fn read(
         &self,
         context: BulkIterContextRef,
@@ -1372,8 +1633,14 @@ impl MultiBulkPart {
             return Ok(None);
         }
 
+        let batches_to_read = self.prune_batches(&context);
+
+        if batches_to_read.is_empty() {
+            return Ok(None);
+        }
+
         let iter = crate::memtable::bulk::part_reader::BulkPartBatchIter::new(
-            self.batches.iter().cloned().collect(),
+            batches_to_read,
             context,
             sequence,
             self.series_count,
@@ -1382,6 +1649,33 @@ impl MultiBulkPart {
         Ok(Some(Box::new(iter) as BoxedRecordBatchIterator))
     }
 
+    /// Prunes batches using the first-tag min/max statistics and the predicate.
+    /// Returns all batches if no stats or no predicate is available.
+    fn prune_batches(&self, context: &BulkIterContextRef) -> Vec<RecordBatch> {
+        if let Some(stats) = &self.batch_stats
+            && let Some(predicate) = &context.predicate
+        {
+            let region_meta = context.read_format().metadata();
+            let pruning_stats = BatchPruningStats {
+                stats,
+                metadata: region_meta,
+            };
+            let mask =
+                predicate.prune_with_stats(&pruning_stats, region_meta.schema.arrow_schema());
+            self.batches
+                .iter()
+                .zip(mask.iter())
+                .filter_map(
+                    |(batch, &selected)| {
+                        if selected { Some(batch.clone()) } else { None }
+                    },
+                )
+                .collect()
+        } else {
+            self.batches.iter().cloned().collect()
+        }
+    }
+
     /// Converts this `MultiBulkPart` to `MemtableStats`.
     pub fn to_memtable_stats(&self, region_metadata: &RegionMetadataRef) -> MemtableStats {
         let ts_type = region_metadata.time_index_type();
@@ -1403,19 +1697,22 @@ impl MultiBulkPart {
 mod tests {
     use api::v1::{Row, SemanticType, WriteHint};
     use datafusion_common::ScalarValue;
-    use datatypes::arrow::array::Float64Array;
+    use datatypes::arrow::array::{
+        BinaryArray, DictionaryArray, Float64Array, TimestampMillisecondArray,
+    };
+    use datatypes::arrow::datatypes::UInt32Type;
     use datatypes::prelude::{ConcreteDataType, Value};
     use datatypes::schema::ColumnSchema;
+    use mito_codec::row_converter::build_primary_key_codec;
     use store_api::metadata::{ColumnMetadata, RegionMetadataBuilder};
     use store_api::storage::RegionId;
     use store_api::storage::consts::ReservedColumnId;
+    use table::predicate::Predicate;
 
     use super::*;
     use crate::memtable::bulk::context::BulkIterContext;
     use crate::sst::{FlatSchemaOptions, to_flat_sst_arrow_schema};
-    use crate::test_util::memtable_util::{
-        build_key_values_with_ts_seq_values, metadata_for_test, region_metadata_to_row_schema,
-    };
+    use crate::test_util::memtable_util::{build_key_values_with_ts_seq_values, metadata_for_test};
 
     struct MutationInput<'a> {
         k0: &'a str,
@@ -1425,13 +1722,6 @@ mod tests {
         sequence: u64,
     }
 
-    #[derive(Debug, PartialOrd, PartialEq)]
-    struct BatchOutput<'a> {
-        pk_values: &'a [Value],
-        timestamps: &'a [i64],
-        v1: &'a [Option<f64>],
-    }
-
     fn encode(input: &[MutationInput]) -> EncodedBulkPart {
         let metadata = metadata_for_test();
         let kvs = input
@@ -1485,7 +1775,7 @@ mod tests {
         ]);
 
         let projection = &[4u32];
-        let mut reader = part
+        let reader = part
             .read(
                 Arc::new(
                     BulkIterContext::new(
@@ -1526,7 +1816,7 @@ mod tests {
         let kvs = key_values
             .into_iter()
             .map(|(k0, k1, (start, end), sequence)| {
-                let ts = (start..end);
+                let ts = start..end;
                 let v1 = (start..end).map(|_| None);
                 build_key_values_with_ts_seq_values(&metadata, k0.to_string(), k1, ts, v1, sequence)
             })
@@ -1556,7 +1846,7 @@ mod tests {
             )
             .unwrap(),
         );
-        let mut reader = part
+        let reader = part
             .read(context, None, None)
             .unwrap()
             .expect("expect at least one row group");
@@ -1629,7 +1919,7 @@ mod tests {
             100,
         );
 
-        /// Predicates over field column can do precise filtering.
+        // Predicates over field column can do precise filtering.
         check_prune_row_group(
             &part,
             Some(Predicate::new(vec![
@@ -2474,4 +2764,128 @@ mod tests {
         let timestamps: Vec<i64> = ts_array.values().to_vec();
         assert_eq!(timestamps, vec![1000, 3000, 2000, 4000]);
     }
+
+    /// Helper to create a converted BulkPart (with __primary_key column) from MutationInputs.
+    fn build_converted_bulk_part(inputs: &[MutationInput]) -> BulkPart {
+        let metadata = metadata_for_test();
+        let kvs = inputs
+            .iter()
+            .map(|m| {
+                build_key_values_with_ts_seq_values(
+                    &metadata,
+                    m.k0.to_string(),
+                    m.k1,
+                    m.timestamps.iter().copied(),
+                    m.v1.iter().copied(),
+                    m.sequence,
+                )
+            })
+            .collect::<Vec<_>>();
+        let schema = to_flat_sst_arrow_schema(&metadata, &FlatSchemaOptions::default());
+        let primary_key_codec = build_primary_key_codec(&metadata);
+        let mut converter = BulkPartConverter::new(&metadata, schema, 64, primary_key_codec, true);
+        for kv in kvs {
+            converter.append_key_values(&kv).unwrap();
+        }
+        converter.convert().unwrap()
+    }
+
+    /// Helper to create a MultiBulkPart where each group becomes a separate batch.
+    fn build_multi_bulk_part(groups: &[&[MutationInput]]) -> (MultiBulkPart, RegionMetadataRef) {
+        let metadata = metadata_for_test();
+        let mut all_batches = Vec::new();
+        let mut min_ts = i64::MAX;
+        let mut max_ts = i64::MIN;
+        let mut max_seq = 0u64;
+
+        for inputs in groups {
+            let part = build_converted_bulk_part(inputs);
+            min_ts = min_ts.min(part.min_timestamp);
+            max_ts = max_ts.max(part.max_timestamp);
+            max_seq = max_seq.max(part.sequence);
+            all_batches.push(part.batch);
+        }
+
+        let multi = MultiBulkPart::new(
+            all_batches,
+            min_ts,
+            max_ts,
+            max_seq,
+            groups.len(),
+            &metadata,
+        );
+        (multi, metadata)
+    }
+
+    #[test]
+    fn test_multi_bulk_part_prune_batches() {
+        // Three batches with distinct k0 ranges: ["a"], ["m"], ["z"].
+        let (multi, metadata) = build_multi_bulk_part(&[
+            &[MutationInput {
+                k0: "a",
+                k1: 0,
+                timestamps: &[1, 2],
+                v1: &[Some(1.0), Some(2.0)],
+                sequence: 0,
+            }],
+            &[MutationInput {
+                k0: "m",
+                k1: 0,
+                timestamps: &[3, 4],
+                v1: &[Some(3.0), Some(4.0)],
+                sequence: 1,
+            }],
+            &[MutationInput {
+                k0: "z",
+                k1: 0,
+                timestamps: &[5, 6],
+                v1: &[Some(5.0), Some(6.0)],
+                sequence: 2,
+            }],
+        ]);
+        assert_eq!(multi.num_rows(), 6);
+        assert_eq!(multi.num_batches(), 3);
+
+        // k0 = "m" => only middle batch (2 rows).
+        let context = Arc::new(
+            BulkIterContext::new(
+                metadata.clone(),
+                None,
+                Some(Predicate::new(vec![
+                    datafusion_expr::col("k0").eq(datafusion_expr::lit("m")),
+                ])),
+                false,
+            )
+            .unwrap(),
+        );
+        let reader = multi
+            .read(context, None, None)
+            .unwrap()
+            .expect("should have results");
+        let total_rows: usize = reader.map(|r| r.unwrap().num_rows()).sum();
+        assert_eq!(total_rows, 2);
+
+        // k0 = "nonexistent" => all pruned, returns None.
+        let context = Arc::new(
+            BulkIterContext::new(
+                metadata.clone(),
+                None,
+                Some(Predicate::new(vec![
+                    datafusion_expr::col("k0").eq(datafusion_expr::lit("nonexistent")),
+                ])),
+                false,
+            )
+            .unwrap(),
+        );
+        assert!(multi.read(context, None, None).unwrap().is_none());
+
+        // No predicate => all 6 rows.
+        let context = Arc::new(BulkIterContext::new(metadata.clone(), None, None, false).unwrap());
+        let reader = multi
+            .read(context, None, None)
+            .unwrap()
+            .expect("should have results");
+        let total_rows: usize = reader.map(|r| r.unwrap().num_rows()).sum();
+        assert_eq!(total_rows, 6);
+    }
 }
diff --git a/src/mito2/src/memtable/bulk/part_reader.rs b/src/mito2/src/memtable/bulk/part_reader.rs
index 1e9d955321..7853bd9d9c 100644
--- a/src/mito2/src/memtable/bulk/part_reader.rs
+++ b/src/mito2/src/memtable/bulk/part_reader.rs
@@ -17,6 +17,7 @@ use std::time::Instant;
 
 use datatypes::arrow::array::BooleanArray;
 use datatypes::arrow::record_batch::RecordBatch;
+use mito_codec::row_converter::PrimaryKeyFilter;
 use parquet::arrow::ProjectionMask;
 use parquet::arrow::arrow_reader::ParquetRecordBatchReader;
 use snafu::ResultExt;
@@ -28,9 +29,9 @@ use crate::memtable::bulk::part::EncodedBulkPart;
 use crate::memtable::bulk::row_group_reader::MemtableRowGroupReaderBuilder;
 use crate::memtable::{MemScanMetrics, MemScanMetricsData};
 use crate::metrics::{READ_ROWS_TOTAL, READ_STAGE_ELAPSED};
-use crate::sst::parquet::file_range::{PreFilterMode, TagDecodeState};
-use crate::sst::parquet::flat_format::sequence_column_index;
-use crate::sst::parquet::reader::RowGroupReaderContext;
+use crate::sst::parquet::file_range::TagDecodeState;
+use crate::sst::parquet::flat_format::{primary_key_column_index, sequence_column_index};
+use crate::sst::parquet::prefilter::{CachedPrimaryKeyFilter, prefilter_flat_batch_by_primary_key};
 
 /// Iterator for reading data inside a bulk part.
 pub struct EncodedBulkPartIter {
@@ -42,6 +43,8 @@ pub struct EncodedBulkPartIter {
     sequence: Option<SequenceRange>,
     /// Cached skip_fields for current row group.
     current_skip_fields: bool,
+    /// Primary key filter for prefiltering before convert_batch.
+    pk_filter: Option<CachedPrimaryKeyFilter>,
     /// Metrics for this iterator.
     metrics: MemScanMetricsData,
     /// Optional memory scan metrics to report to.
@@ -50,29 +53,33 @@ pub struct EncodedBulkPartIter {
 
 impl EncodedBulkPartIter {
     /// Creates a new [BulkPartIter].
-    pub(crate) fn try_new(
+    pub fn try_new(
         encoded_part: &EncodedBulkPart,
         context: BulkIterContextRef,
         mut row_groups_to_read: VecDeque<usize>,
         sequence: Option<SequenceRange>,
         mem_scan_metrics: Option<MemScanMetrics>,
     ) -> error::Result<Self> {
-        assert!(context.read_format().as_flat().is_some());
-
         let parquet_meta = encoded_part.metadata().parquet_metadata.clone();
         let data = encoded_part.data().clone();
         let series_count = encoded_part.metadata().num_series as usize;
 
-        let projection_mask = ProjectionMask::roots(
-            parquet_meta.file_metadata().schema_descr(),
-            context.read_format().projection_indices().iter().copied(),
-        );
+        // TODO(fys): Nested projection pushdown to the memtable layer is not supported yet.
+        let root_indices = context
+            .read_format()
+            .parquet_read_columns()
+            .root_indices_iter();
+        let projection_mask =
+            ProjectionMask::roots(parquet_meta.file_metadata().schema_descr(), root_indices);
         let builder =
             MemtableRowGroupReaderBuilder::try_new(&context, projection_mask, parquet_meta, data)?;
 
+        // Build PK filter if applicable (flat format with dictionary-encoded PKs).
+        let pk_filter = context.build_pk_filter();
+
         let (init_reader, current_skip_fields) = match row_groups_to_read.pop_front() {
             Some(first_row_group) => {
-                let skip_fields = builder.compute_skip_fields(&context, first_row_group);
+                let skip_fields = context.pre_filter_mode().skip_fields();
                 let reader = builder.build_row_group_reader(first_row_group, None)?;
                 (Some(reader), skip_fields)
             }
@@ -86,6 +93,7 @@ impl EncodedBulkPartIter {
             builder,
             sequence,
             current_skip_fields,
+            pk_filter,
             metrics: MemScanMetricsData {
                 total_series: series_count,
                 ..Default::default()
@@ -117,6 +125,10 @@ impl EncodedBulkPartIter {
                 &self.sequence,
                 batch,
                 self.current_skip_fields,
+                self.pk_filter
+                    .as_mut()
+                    .map(|f| f as &mut dyn PrimaryKeyFilter),
+                &mut self.metrics,
             )? {
                 // Update metrics
                 self.metrics.num_batches += 1;
@@ -129,9 +141,7 @@ impl EncodedBulkPartIter {
         // Previous row group exhausted, read next row group
         while let Some(next_row_group) = self.row_groups_to_read.pop_front() {
             // Compute skip_fields for this row group
-            self.current_skip_fields = self
-                .builder
-                .compute_skip_fields(&self.context, next_row_group);
+            self.current_skip_fields = self.context.pre_filter_mode().skip_fields();
 
             let next_reader = self.builder.build_row_group_reader(next_row_group, None)?;
             let current = self.current_reader.insert(next_reader);
@@ -143,6 +153,10 @@ impl EncodedBulkPartIter {
                     &self.sequence,
                     batch,
                     self.current_skip_fields,
+                    self.pk_filter
+                        .as_mut()
+                        .map(|f| f as &mut dyn PrimaryKeyFilter),
+                    &mut self.metrics,
                 )? {
                     // Update metrics
                     self.metrics.num_batches += 1;
@@ -176,12 +190,14 @@ impl Iterator for EncodedBulkPartIter {
 impl Drop for EncodedBulkPartIter {
     fn drop(&mut self) {
         common_telemetry::debug!(
-            "EncodedBulkPartIter region: {}, metrics: total_series={}, num_rows={}, num_batches={}, scan_cost={:?}",
+            "EncodedBulkPartIter region: {}, metrics: total_series={}, num_rows={}, num_batches={}, scan_cost={:?}, prefilter_cost={:?}, prefilter_rows_filtered={}",
             self.context.region_id(),
             self.metrics.total_series,
             self.metrics.num_rows,
             self.metrics.num_batches,
-            self.metrics.scan_cost
+            self.metrics.scan_cost,
+            self.metrics.prefilter_cost,
+            self.metrics.prefilter_rows_filtered
         );
 
         // Report MemScanMetrics if not already reported
@@ -206,6 +222,8 @@ pub struct BulkPartBatchIter {
     context: BulkIterContextRef,
     /// Sequence number filter.
     sequence: Option<SequenceRange>,
+    /// Primary key filter for prefiltering before convert_batch.
+    pk_filter: Option<CachedPrimaryKeyFilter>,
     /// Metrics for this iterator.
     metrics: MemScanMetricsData,
     /// Optional memory scan metrics to report to.
@@ -221,12 +239,13 @@ impl BulkPartBatchIter {
         series_count: usize,
         mem_scan_metrics: Option<MemScanMetrics>,
     ) -> Self {
-        assert!(context.read_format().as_flat().is_some());
+        let pk_filter = context.build_pk_filter();
 
         Self {
             batches: VecDeque::from(batches),
             context,
             sequence,
+            pk_filter,
             metrics: MemScanMetricsData {
                 total_series: series_count,
                 ..Default::default()
@@ -260,7 +279,11 @@ impl BulkPartBatchIter {
 
     /// Applies projection to the RecordBatch if needed.
     fn apply_projection(&self, record_batch: RecordBatch) -> error::Result<RecordBatch> {
-        let projection_indices = self.context.read_format().projection_indices();
+        let projection_indices = self
+            .context
+            .read_format()
+            .parquet_read_columns()
+            .root_indices();
         if projection_indices.len() == record_batch.num_columns() {
             return Ok(record_batch);
         }
@@ -277,14 +300,18 @@ impl BulkPartBatchIter {
         let projected_batch = self.apply_projection(record_batch)?;
 
         // Apply combined filtering (both predicate and sequence filters)
-        let skip_fields = match self.context.pre_filter_mode() {
-            PreFilterMode::All => false,
-            PreFilterMode::SkipFields => true,
-            PreFilterMode::SkipFieldsOnDelete => true,
-        };
+        let skip_fields = self.context.pre_filter_mode().skip_fields();
 
-        let Some(filtered_batch) =
-            apply_combined_filters(&self.context, &self.sequence, projected_batch, skip_fields)?
+        let Some(filtered_batch) = apply_combined_filters(
+            &self.context,
+            &self.sequence,
+            projected_batch,
+            skip_fields,
+            self.pk_filter
+                .as_mut()
+                .map(|f| f as &mut dyn PrimaryKeyFilter),
+            &mut self.metrics,
+        )?
         else {
             self.metrics.scan_cost += start.elapsed();
             return Ok(None);
@@ -324,12 +351,14 @@ impl Iterator for BulkPartBatchIter {
 impl Drop for BulkPartBatchIter {
     fn drop(&mut self) {
         common_telemetry::debug!(
-            "BulkPartBatchIter region: {}, metrics: total_series={}, num_rows={}, num_batches={}, scan_cost={:?}",
+            "BulkPartBatchIter region: {}, metrics: total_series={}, num_rows={}, num_batches={}, scan_cost={:?}, prefilter_cost={:?}, prefilter_rows_filtered={}",
             self.context.region_id(),
             self.metrics.total_series,
             self.metrics.num_rows,
             self.metrics.num_batches,
-            self.metrics.scan_cost
+            self.metrics.scan_cost,
+            self.metrics.prefilter_cost,
+            self.metrics.prefilter_rows_filtered
         );
 
         // Report MemScanMetrics if not already reported
@@ -354,16 +383,38 @@ fn apply_combined_filters(
     sequence: &Option<SequenceRange>,
     record_batch: RecordBatch,
     skip_fields: bool,
+    pk_filter: Option<&mut dyn PrimaryKeyFilter>,
+    metrics: &mut MemScanMetricsData,
 ) -> error::Result<Option<RecordBatch>> {
-    // Converts the format to the flat format first.
-    let format = context.read_format().as_flat().unwrap();
-    let record_batch = format.convert_batch(record_batch, None)?;
+    // Apply PK prefilter on raw batch before convert_batch to reduce conversion overhead.
+    let record_batch = if let Some(pk_filter) = pk_filter {
+        let rows_before = record_batch.num_rows();
+        let prefilter_start = Instant::now();
+        let pk_col_idx = primary_key_column_index(record_batch.num_columns());
+        match prefilter_flat_batch_by_primary_key(record_batch, pk_col_idx, pk_filter)? {
+            Some(batch) => {
+                metrics.prefilter_cost += prefilter_start.elapsed();
+                metrics.prefilter_rows_filtered += rows_before - batch.num_rows();
+                batch
+            }
+            None => {
+                metrics.prefilter_cost += prefilter_start.elapsed();
+                metrics.prefilter_rows_filtered += rows_before;
+                return Ok(None);
+            }
+        }
+    } else {
+        record_batch
+    };
+
+    // Converts the format to the flat format.
+    let record_batch = context.read_format().convert_batch(record_batch, None)?;
 
     let num_rows = record_batch.num_rows();
     let mut combined_filter = None;
     let mut tag_decode_state = TagDecodeState::new();
 
-    // First, apply predicate filters using the shared method.
+    // Apply predicate filters using the shared method.
     if !context.base.filters.is_empty() {
         let predicate_mask = context.base.compute_filter_mask_flat(
             &record_batch,
@@ -433,6 +484,7 @@ mod tests {
 
     use super::*;
     use crate::memtable::bulk::context::BulkIterContext;
+    use crate::test_util::sst_util::new_primary_key;
 
     #[test]
     fn test_bulk_part_batch_iter() {
@@ -461,9 +513,16 @@ mod tests {
             vec![1000, 2000, 3000],
         ));
 
-        // Create primary key dictionary array
+        // Create primary key dictionary array with properly encoded PKs
         use datatypes::arrow::array::{BinaryArray, DictionaryArray, UInt32Array};
-        let values = Arc::new(BinaryArray::from_iter_values([b"key1", b"key2", b"key3"]));
+        let pk1 = new_primary_key(&["key1"]);
+        let pk2 = new_primary_key(&["key2"]);
+        let pk3 = new_primary_key(&["key3"]);
+        let values = Arc::new(BinaryArray::from_iter_values([
+            pk1.as_slice(),
+            pk2.as_slice(),
+            pk3.as_slice(),
+        ]));
         let keys = UInt32Array::from(vec![0, 1, 2]);
         let primary_key = Arc::new(DictionaryArray::new(keys, values));
 
@@ -596,12 +655,17 @@ mod tests {
         ]));
 
         // Create first batch with 2 rows
+        let pk1 = new_primary_key(&["key1"]);
+        let pk2 = new_primary_key(&["key2"]);
         let key1_1 = Arc::new(StringArray::from_iter_values(["key1", "key2"]));
         let field1_1 = Arc::new(Int64Array::from(vec![11, 12]));
         let timestamp_1 = Arc::new(datatypes::arrow::array::TimestampMillisecondArray::from(
             vec![1000, 2000],
         ));
-        let values_1 = Arc::new(BinaryArray::from_iter_values([b"key1", b"key2"]));
+        let values_1 = Arc::new(BinaryArray::from_iter_values([
+            pk1.as_slice(),
+            pk2.as_slice(),
+        ]));
         let keys_1 = UInt32Array::from(vec![0, 1]);
         let primary_key_1 = Arc::new(DictionaryArray::new(keys_1, values_1));
         let sequence_1 = Arc::new(UInt64Array::from(vec![1, 2]));
@@ -621,12 +685,19 @@ mod tests {
         .unwrap();
 
         // Create second batch with 3 rows
+        let pk3 = new_primary_key(&["key3"]);
+        let pk4 = new_primary_key(&["key4"]);
+        let pk5 = new_primary_key(&["key5"]);
         let key1_2 = Arc::new(StringArray::from_iter_values(["key3", "key4", "key5"]));
         let field1_2 = Arc::new(Int64Array::from(vec![13, 14, 15]));
         let timestamp_2 = Arc::new(datatypes::arrow::array::TimestampMillisecondArray::from(
             vec![3000, 4000, 5000],
         ));
-        let values_2 = Arc::new(BinaryArray::from_iter_values([b"key3", b"key4", b"key5"]));
+        let values_2 = Arc::new(BinaryArray::from_iter_values([
+            pk3.as_slice(),
+            pk4.as_slice(),
+            pk5.as_slice(),
+        ]));
         let keys_2 = UInt32Array::from(vec![0, 1, 2]);
         let primary_key_2 = Arc::new(DictionaryArray::new(keys_2, values_2));
         let sequence_2 = Arc::new(UInt64Array::from(vec![3, 4, 5]));
diff --git a/src/mito2/src/memtable/bulk/row_group_reader.rs b/src/mito2/src/memtable/bulk/row_group_reader.rs
index fccd22db10..36e8e8aac1 100644
--- a/src/mito2/src/memtable/bulk/row_group_reader.rs
+++ b/src/mito2/src/memtable/bulk/row_group_reader.rs
@@ -12,124 +12,26 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::ops::Range;
 use std::sync::Arc;
 
 use bytes::Bytes;
-use datatypes::arrow::array::RecordBatch;
-use datatypes::arrow::error::ArrowError;
-use parquet::arrow::arrow_reader::{ParquetRecordBatchReader, RowGroups, RowSelection};
-use parquet::arrow::{FieldLevels, ProjectionMask, parquet_to_arrow_field_levels};
-use parquet::column::page::{PageIterator, PageReader};
-use parquet::file::metadata::{ParquetMetaData, RowGroupMetaData};
+use parquet::arrow::ProjectionMask;
+use parquet::arrow::arrow_reader::{
+    ArrowReaderMetadata, ArrowReaderOptions, ParquetRecordBatchReader,
+    ParquetRecordBatchReaderBuilder, RowSelection,
+};
+use parquet::file::metadata::ParquetMetaData;
 use snafu::ResultExt;
 
 use crate::error;
 use crate::error::ReadDataPartSnafu;
+use crate::memtable::bulk::chunk_reader::MemtableChunkReader;
 use crate::memtable::bulk::context::BulkIterContextRef;
 use crate::sst::parquet::DEFAULT_READ_BATCH_SIZE;
-use crate::sst::parquet::format::ReadFormat;
-use crate::sst::parquet::reader::RowGroupReaderContext;
-use crate::sst::parquet::row_group::{ColumnChunkIterator, RowGroupBase};
-
-/// Helper for reading specific row group inside Memtable Parquet parts.
-// This is similar to [mito2::sst::parquet::row_group::InMemoryRowGroup] since
-// it's a workaround for lacking of keyword generics.
-pub struct MemtableRowGroupPageFetcher<'a> {
-    /// Shared structs for reading row group.
-    base: RowGroupBase<'a>,
-    bytes: Bytes,
-}
-
-impl<'a> MemtableRowGroupPageFetcher<'a> {
-    pub(crate) fn create(
-        row_group_idx: usize,
-        parquet_meta: &'a ParquetMetaData,
-        bytes: Bytes,
-    ) -> Self {
-        Self {
-            // the cached `column_uncompressed_pages` would never be used in Memtable readers.
-            base: RowGroupBase::new(parquet_meta, row_group_idx),
-            bytes,
-        }
-    }
-
-    /// Fetches column pages from memory file.
-    pub(crate) fn fetch(&mut self, projection: &ProjectionMask, selection: Option<&RowSelection>) {
-        if let Some((selection, offset_index)) = selection.zip(self.base.offset_index) {
-            // Selection provided.
-            let (fetch_ranges, page_start_offsets) =
-                self.base
-                    .calc_sparse_read_ranges(projection, offset_index, selection);
-            if fetch_ranges.is_empty() {
-                return;
-            }
-            let chunk_data = self.fetch_bytes(&fetch_ranges);
-
-            self.base
-                .assign_sparse_chunk(projection, chunk_data, page_start_offsets);
-        } else {
-            let fetch_ranges = self.base.calc_dense_read_ranges(projection);
-            if fetch_ranges.is_empty() {
-                // Nothing to fetch.
-                return;
-            }
-            let chunk_data = self.fetch_bytes(&fetch_ranges);
-            self.base.assign_dense_chunk(projection, chunk_data);
-        }
-    }
-
-    fn fetch_bytes(&self, ranges: &[Range<u64>]) -> Vec<Bytes> {
-        ranges
-            .iter()
-            .map(|range| self.bytes.slice(range.start as usize..range.end as usize))
-            .collect()
-    }
-
-    /// Creates a page reader to read column at `i`.
-    fn column_page_reader(&self, i: usize) -> parquet::errors::Result<Box<dyn PageReader>> {
-        let reader = self.base.column_reader(i)?;
-        Ok(Box::new(reader))
-    }
-}
-
-impl RowGroups for MemtableRowGroupPageFetcher<'_> {
-    fn num_rows(&self) -> usize {
-        self.base.row_count
-    }
-
-    fn column_chunks(&self, i: usize) -> parquet::errors::Result<Box<dyn PageIterator>> {
-        Ok(Box::new(ColumnChunkIterator {
-            reader: Some(self.column_page_reader(i)),
-        }))
-    }
-
-    fn row_groups(&self) -> Box<dyn Iterator<Item = &RowGroupMetaData> + '_> {
-        Box::new(std::iter::once(self.base.row_group_metadata()))
-    }
-
-    fn metadata(&self) -> &ParquetMetaData {
-        self.base.parquet_metadata()
-    }
-}
-
-impl RowGroupReaderContext for BulkIterContextRef {
-    fn map_result(
-        &self,
-        result: Result<Option<RecordBatch>, ArrowError>,
-    ) -> error::Result<Option<RecordBatch>> {
-        result.context(error::DecodeArrowRowGroupSnafu)
-    }
-
-    fn read_format(&self) -> &ReadFormat {
-        self.as_ref().read_format()
-    }
-}
 
 pub(crate) struct MemtableRowGroupReaderBuilder {
     projection: ProjectionMask,
-    parquet_metadata: Arc<ParquetMetaData>,
-    field_levels: FieldLevels,
+    arrow_metadata: ArrowReaderMetadata,
     data: Bytes,
 }
 
@@ -140,15 +42,15 @@ impl MemtableRowGroupReaderBuilder {
         parquet_metadata: Arc<ParquetMetaData>,
         data: Bytes,
     ) -> error::Result<Self> {
-        let parquet_schema_desc = parquet_metadata.file_metadata().schema_descr();
-        let hint = Some(context.read_format().arrow_schema().fields());
-        let field_levels =
-            parquet_to_arrow_field_levels(parquet_schema_desc, projection.clone(), hint)
+        // Create ArrowReaderMetadata for building the reader.
+        let arrow_reader_options =
+            ArrowReaderOptions::new().with_schema(context.read_format().arrow_schema().clone());
+        let arrow_metadata =
+            ArrowReaderMetadata::try_new(parquet_metadata.clone(), arrow_reader_options)
                 .context(ReadDataPartSnafu)?;
         Ok(Self {
             projection,
-            parquet_metadata,
-            field_levels,
+            arrow_metadata,
             data,
         })
     }
@@ -159,41 +61,20 @@ impl MemtableRowGroupReaderBuilder {
         row_group_idx: usize,
         row_selection: Option<RowSelection>,
     ) -> error::Result<ParquetRecordBatchReader> {
-        let mut row_group = MemtableRowGroupPageFetcher::create(
-            row_group_idx,
-            &self.parquet_metadata,
-            self.data.clone(),
-        );
-        // Fetches data from memory part. Currently, row selection is not supported.
-        row_group.fetch(&self.projection, row_selection.as_ref());
+        let chunk_reader = MemtableChunkReader::new(self.data.clone());
 
-        // Builds the parquet reader.
-        // Now the row selection is None.
-        ParquetRecordBatchReader::try_new_with_row_groups(
-            &self.field_levels,
-            &row_group,
-            DEFAULT_READ_BATCH_SIZE,
-            row_selection,
+        let mut builder = ParquetRecordBatchReaderBuilder::new_with_metadata(
+            chunk_reader,
+            self.arrow_metadata.clone(),
         )
-        .context(ReadDataPartSnafu)
-    }
+        .with_row_groups(vec![row_group_idx])
+        .with_projection(self.projection.clone())
+        .with_batch_size(DEFAULT_READ_BATCH_SIZE);
 
-    /// Computes whether to skip field filters for a specific row group based on PreFilterMode.
-    pub(crate) fn compute_skip_fields(
-        &self,
-        context: &BulkIterContextRef,
-        row_group_idx: usize,
-    ) -> bool {
-        use crate::sst::parquet::file_range::{PreFilterMode, row_group_contains_delete};
-
-        match context.pre_filter_mode() {
-            PreFilterMode::All => false,
-            PreFilterMode::SkipFields => true,
-            PreFilterMode::SkipFieldsOnDelete => {
-                // Check if this specific row group contains delete op
-                row_group_contains_delete(&self.parquet_metadata, row_group_idx, "memtable")
-                    .unwrap_or(true)
-            }
+        if let Some(selection) = row_selection {
+            builder = builder.with_row_selection(selection);
         }
+
+        builder.build().context(ReadDataPartSnafu)
     }
 }
diff --git a/src/mito2/src/memtable/partition_tree.rs b/src/mito2/src/memtable/partition_tree.rs
index febae46784..662bfd99f6 100644
--- a/src/mito2/src/memtable/partition_tree.rs
+++ b/src/mito2/src/memtable/partition_tree.rs
@@ -177,16 +177,6 @@ impl Memtable for PartitionTreeMemtable {
         .fail()
     }
 
-    #[cfg(any(test, feature = "test"))]
-    fn iter(
-        &self,
-        projection: Option<&[ColumnId]>,
-        predicate: Option<Predicate>,
-        sequence: Option<SequenceRange>,
-    ) -> Result<BoxedBatchIterator> {
-        self.tree.read(projection, predicate, sequence, None)
-    }
-
     fn ranges(
         &self,
         projection: Option<&[ColumnId]>,
@@ -396,8 +386,6 @@ mod tests {
     use api::v1::{Mutation, OpType, Rows, SemanticType};
     use common_query::prelude::{greptime_timestamp, greptime_value};
     use common_time::Timestamp;
-    use datafusion_common::Column;
-    use datafusion_expr::{BinaryExpr, Expr, Literal, Operator};
     use datatypes::data_type::ConcreteDataType;
     use datatypes::prelude::Vector;
     use datatypes::scalars::ScalarVector;
@@ -548,7 +536,10 @@ mod tests {
         let expect = (0..100).collect::<Vec<_>>();
         let kvs = memtable_util::build_key_values(&metadata, "hello".to_string(), 10, &expect, 1);
         memtable.write(&kvs).unwrap();
-        let iter = memtable.iter(Some(&[3]), None, None).unwrap();
+        let ranges = memtable
+            .ranges(Some(&[3]), RangesOptions::default())
+            .unwrap();
+        let iter = ranges.build(None).unwrap();
 
         let mut v0_all = vec![];
         for res in iter {
@@ -625,41 +616,6 @@ mod tests {
         assert_eq!(expect, read);
     }
 
-    #[test]
-    fn test_memtable_filter() {
-        let metadata = Arc::new(memtable_util::metadata_with_primary_key(vec![0, 1], false));
-        // Try to build a memtable via the builder.
-        let memtable = PartitionTreeMemtableBuilder::new(
-            PartitionTreeConfig {
-                index_max_keys_per_shard: 40,
-                ..Default::default()
-            },
-            None,
-        )
-        .build(1, &metadata);
-
-        for i in 0..100 {
-            let timestamps: Vec<_> = (0..10).map(|v| i as i64 * 1000 + v).collect();
-            let kvs =
-                memtable_util::build_key_values(&metadata, "hello".to_string(), i, &timestamps, 1);
-            memtable.write(&kvs).unwrap();
-        }
-
-        for i in 0..100 {
-            let timestamps: Vec<_> = (0..10).map(|v| i as i64 * 1000 + v).collect();
-            let expr = Expr::BinaryExpr(BinaryExpr {
-                left: Box::new(Expr::Column(Column::from_name("k1"))),
-                op: Operator::Eq,
-                right: Box::new((i as u32).lit()),
-            });
-            let iter = memtable
-                .iter(None, Some(Predicate::new(vec![expr])), None)
-                .unwrap();
-            let read = collect_iter_timestamps(iter);
-            assert_eq!(timestamps, read);
-        }
-    }
-
     #[test]
     fn test_deserialize_config() {
         let config = PartitionTreeConfig {
@@ -811,7 +767,11 @@ mod tests {
             ))
             .unwrap();
 
-        let mut reader = new_memtable.iter(None, None, None).unwrap();
+        let mut reader = new_memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let batch = reader.next().unwrap().unwrap();
         let pk = codec.decode(batch.primary_key()).unwrap().into_dense();
         if let Value::String(s) = &pk[2] {
@@ -916,7 +876,14 @@ mod tests {
             .unwrap();
         memtable.freeze().unwrap();
         assert_eq!(
-            collect_kvs(memtable.iter(None, None, None).unwrap(), &metadata),
+            collect_kvs(
+                memtable
+                    .ranges(None, RangesOptions::default())
+                    .unwrap()
+                    .build(None)
+                    .unwrap(),
+                &metadata
+            ),
             ('a'..'h').map(|c| (c.to_string(), c.to_string())).collect()
         );
         let forked = memtable.fork(2, &metadata);
@@ -925,7 +892,14 @@ mod tests {
         forked.write(&key_values(&metadata, keys.iter())).unwrap();
         forked.freeze().unwrap();
         assert_eq!(
-            collect_kvs(forked.iter(None, None, None).unwrap(), &metadata),
+            collect_kvs(
+                forked
+                    .ranges(None, RangesOptions::default())
+                    .unwrap()
+                    .build(None)
+                    .unwrap(),
+                &metadata
+            ),
             keys.iter()
                 .map(|c| (c.to_string(), c.to_string()))
                 .collect()
@@ -936,7 +910,14 @@ mod tests {
         let keys = ["g", "e", "a", "f", "b", "c", "h"];
         forked2.write(&key_values(&metadata, keys.iter())).unwrap();
 
-        let kvs = collect_kvs(forked2.iter(None, None, None).unwrap(), &metadata);
+        let kvs = collect_kvs(
+            forked2
+                .ranges(None, RangesOptions::default())
+                .unwrap()
+                .build(None)
+                .unwrap(),
+            &metadata,
+        );
         let expected = keys
             .iter()
             .map(|c| (c.to_string(), c.to_string()))
diff --git a/src/mito2/src/memtable/partition_tree/data.rs b/src/mito2/src/memtable/partition_tree/data.rs
index a6d40bdcbf..f6e2a59bec 100644
--- a/src/mito2/src/memtable/partition_tree/data.rs
+++ b/src/mito2/src/memtable/partition_tree/data.rs
@@ -50,6 +50,7 @@ use crate::memtable::partition_tree::merger::{DataBatchKey, DataNode, DataSource
 use crate::metrics::{
     PARTITION_TREE_DATA_BUFFER_FREEZE_STAGE_ELAPSED, PARTITION_TREE_READ_STAGE_ELAPSED,
 };
+use crate::sst::parquet::DEFAULT_READ_BATCH_SIZE;
 
 const PK_INDEX_COLUMN_NAME: &str = "__pk_index";
 
@@ -821,7 +822,11 @@ impl DataPart {
     /// Reads frozen data part and yields [DataBatch]es.
     pub fn read(&self) -> Result<DataPartReader> {
         match self {
-            DataPart::Parquet(data_bytes) => DataPartReader::new(data_bytes.data.clone(), None),
+            // Keep encoded memtable scans aligned with mito/DataFusion batch sizing instead of
+            // parquet-rs's implicit 1024-row default.
+            DataPart::Parquet(data_bytes) => {
+                DataPartReader::new(data_bytes.data.clone(), Some(DEFAULT_READ_BATCH_SIZE))
+            }
         }
     }
 
diff --git a/src/mito2/src/memtable/partition_tree/partition.rs b/src/mito2/src/memtable/partition_tree/partition.rs
index e6e3b8bf81..0ffbce4867 100644
--- a/src/mito2/src/memtable/partition_tree/partition.rs
+++ b/src/mito2/src/memtable/partition_tree/partition.rs
@@ -152,7 +152,8 @@ impl Partition {
         filters: &Arc<Vec<SimpleFilterEvaluator>>,
     ) -> Option<Box<dyn PrimaryKeyFilter>> {
         if need_prune_key {
-            let filter = row_codec.primary_key_filter(metadata, filters.clone());
+            // TODO(yingwen): Remove `skip_partition_column` after dropping PartitionTreeMemtable.
+            let filter = row_codec.primary_key_filter(metadata, filters.clone(), true);
             Some(filter)
         } else {
             None
diff --git a/src/mito2/src/memtable/partition_tree/shard.rs b/src/mito2/src/memtable/partition_tree/shard.rs
index 162f937a7c..c5dc25f573 100644
--- a/src/mito2/src/memtable/partition_tree/shard.rs
+++ b/src/mito2/src/memtable/partition_tree/shard.rs
@@ -19,9 +19,10 @@ use std::time::{Duration, Instant};
 
 use mito_codec::key_values::KeyValue;
 use mito_codec::row_converter::PrimaryKeyFilter;
+use snafu::ResultExt;
 use store_api::metadata::RegionMetadataRef;
 
-use crate::error::Result;
+use crate::error::{DecodeSnafu, Result};
 use crate::memtable::partition_tree::data::{
     DATA_INIT_CAP, DataBatch, DataParts, DataPartsReader, DataPartsReaderBuilder,
 };
@@ -243,7 +244,7 @@ impl ShardReader {
             // Safety: `key_filter` is some so the shard has primary keys.
             let key = self.key_dict.as_ref().unwrap().key_by_pk_index(pk_index);
             let now = Instant::now();
-            if key_filter.matches(key) {
+            if key_filter.matches(key).context(DecodeSnafu)? {
                 self.prune_pk_cost += now.elapsed();
                 self.last_yield_pk_index = Some(pk_index);
                 self.keys_after_pruning += 1;
diff --git a/src/mito2/src/memtable/partition_tree/shard_builder.rs b/src/mito2/src/memtable/partition_tree/shard_builder.rs
index 26de85767d..78eeb463c6 100644
--- a/src/mito2/src/memtable/partition_tree/shard_builder.rs
+++ b/src/mito2/src/memtable/partition_tree/shard_builder.rs
@@ -20,9 +20,10 @@ use std::time::{Duration, Instant};
 
 use mito_codec::key_values::KeyValue;
 use mito_codec::row_converter::PrimaryKeyFilter;
+use snafu::ResultExt;
 use store_api::metadata::RegionMetadataRef;
 
-use crate::error::Result;
+use crate::error::{DecodeSnafu, Result};
 use crate::memtable::partition_tree::data::{
     DATA_INIT_CAP, DataBatch, DataBuffer, DataBufferReader, DataBufferReaderBuilder, DataParts,
 };
@@ -281,7 +282,7 @@ impl ShardBuilderReader {
             self.keys_before_pruning += 1;
             let key = self.dict_reader.key_by_pk_index(pk_index);
             let now = Instant::now();
-            if key_filter.matches(key) {
+            if key_filter.matches(key).context(DecodeSnafu)? {
                 self.prune_pk_cost += now.elapsed();
                 self.last_yield_pk_index = Some(pk_index);
                 self.keys_after_pruning += 1;
diff --git a/src/mito2/src/memtable/partition_tree/tree.rs b/src/mito2/src/memtable/partition_tree/tree.rs
index 17977db56a..f5863ae0c8 100644
--- a/src/mito2/src/memtable/partition_tree/tree.rs
+++ b/src/mito2/src/memtable/partition_tree/tree.rs
@@ -490,6 +490,7 @@ impl TreeIter {
                 num_rows: self.metrics.rows_fetched,
                 num_batches: self.metrics.batches_fetched,
                 scan_cost: self.metrics.iter_elapsed,
+                ..Default::default()
             };
             mem_scan_metrics.merge_inner(&inner);
         }
diff --git a/src/mito2/src/memtable/simple_bulk_memtable.rs b/src/mito2/src/memtable/simple_bulk_memtable.rs
index 4dcaa2bac0..6ff799ebf5 100644
--- a/src/mito2/src/memtable/simple_bulk_memtable.rs
+++ b/src/mito2/src/memtable/simple_bulk_memtable.rs
@@ -213,22 +213,6 @@ impl Memtable for SimpleBulkMemtable {
         Ok(())
     }
 
-    #[cfg(any(test, feature = "test"))]
-    fn iter(
-        &self,
-        projection: Option<&[ColumnId]>,
-        _predicate: Option<table::predicate::Predicate>,
-        sequence: Option<store_api::storage::SequenceRange>,
-    ) -> error::Result<BoxedBatchIterator> {
-        let iter = self.create_iter(projection, sequence)?.build(None)?;
-        if self.merge_mode == MergeMode::LastNonNull {
-            let iter = LastNonNullIter::new(iter);
-            Ok(Box::new(iter))
-        } else {
-            Ok(Box::new(iter))
-        }
-    }
-
     fn ranges(
         &self,
         projection: Option<&[ColumnId]>,
@@ -388,6 +372,7 @@ impl IterBuilder for BatchRangeBuilder {
                 num_rows: batch.num_rows(),
                 num_batches: 1,
                 scan_cost: self.scan_cost,
+                ..Default::default()
             };
             metrics.merge_inner(&inner);
         }
@@ -436,10 +421,6 @@ mod tests {
     use store_api::storage::{RegionId, SequenceNumber, SequenceRange};
 
     use super::*;
-    use crate::read;
-    use crate::read::dedup::DedupReader;
-    use crate::read::merge::MergeReaderBuilder;
-    use crate::read::{BatchReader, Source};
     use crate::region::options::MergeMode;
     use crate::test_util::column_metadata_to_column_schema;
 
@@ -526,7 +507,11 @@ mod tests {
             ))
             .unwrap();
 
-        let mut iter = memtable.iter(None, None, None).unwrap();
+        let mut iter = memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let batch = iter.next().unwrap().unwrap();
         assert_eq!(2, batch.num_rows());
         assert_eq!(2, batch.fields().len());
@@ -551,7 +536,11 @@ mod tests {
             ))
             .unwrap();
 
-        let mut iter = memtable.iter(None, None, None).unwrap();
+        let mut iter = memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let batch = iter.next().unwrap().unwrap();
         assert_eq!(1, batch.num_rows());
         assert_eq!(2, batch.fields().len());
@@ -565,7 +554,11 @@ mod tests {
 
         // Only project column 2 (f1)
         let projection = vec![2];
-        let mut iter = memtable.iter(Some(&projection), None, None).unwrap();
+        let mut iter = memtable
+            .ranges(Some(&projection), RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let batch = iter.next().unwrap().unwrap();
 
         assert_eq!(1, batch.num_rows());
@@ -592,7 +585,11 @@ mod tests {
                 OpType::Put,
             ))
             .unwrap();
-        let mut iter = memtable.iter(None, None, None).unwrap();
+        let mut iter = memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let batch = iter.next().unwrap().unwrap();
 
         assert_eq!(1, batch.num_rows()); // deduped to 1 row
@@ -611,86 +608,15 @@ mod tests {
         let kv = kvs.iter().next().unwrap();
         memtable.write_one(kv).unwrap();
 
-        let mut iter = memtable.iter(None, None, None).unwrap();
+        let mut iter = memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let batch = iter.next().unwrap().unwrap();
         assert_eq!(1, batch.num_rows());
     }
 
-    #[tokio::test]
-    async fn test_write_dedup() {
-        let memtable = new_test_memtable(true, MergeMode::LastRow);
-        let kvs = build_key_values(
-            &memtable.region_metadata,
-            0,
-            &[(1, 1.0, "a".to_string())],
-            OpType::Put,
-        );
-        let kv = kvs.iter().next().unwrap();
-        memtable.write_one(kv).unwrap();
-        memtable.freeze().unwrap();
-
-        let kvs = build_key_values(
-            &memtable.region_metadata,
-            1,
-            &[(1, 1.0, "a".to_string())],
-            OpType::Delete,
-        );
-        let kv = kvs.iter().next().unwrap();
-        memtable.write_one(kv).unwrap();
-
-        let ranges = memtable.ranges(None, RangesOptions::default()).unwrap();
-        let mut source = vec![];
-        for r in ranges.ranges.values() {
-            source.push(Source::Iter(r.build_iter().unwrap()));
-        }
-
-        let reader = MergeReaderBuilder::from_sources(source)
-            .build()
-            .await
-            .unwrap();
-
-        let mut reader = DedupReader::new(reader, read::dedup::LastRow::new(false), None);
-        let mut num_rows = 0;
-        while let Some(b) = reader.next_batch().await.unwrap() {
-            num_rows += b.num_rows();
-        }
-        assert_eq!(num_rows, 1);
-    }
-
-    #[tokio::test]
-    async fn test_delete_only() {
-        let memtable = new_test_memtable(true, MergeMode::LastRow);
-        let kvs = build_key_values(
-            &memtable.region_metadata,
-            0,
-            &[(1, 1.0, "a".to_string())],
-            OpType::Delete,
-        );
-        let kv = kvs.iter().next().unwrap();
-        memtable.write_one(kv).unwrap();
-        memtable.freeze().unwrap();
-
-        let ranges = memtable.ranges(None, RangesOptions::default()).unwrap();
-        let mut source = vec![];
-        for r in ranges.ranges.values() {
-            source.push(Source::Iter(r.build_iter().unwrap()));
-        }
-
-        let reader = MergeReaderBuilder::from_sources(source)
-            .build()
-            .await
-            .unwrap();
-
-        let mut reader = DedupReader::new(reader, read::dedup::LastRow::new(false), None);
-        let mut num_rows = 0;
-        while let Some(b) = reader.next_batch().await.unwrap() {
-            num_rows += b.num_rows();
-            assert_eq!(b.num_rows(), 1);
-            assert_eq!(b.op_types().get_data(0).unwrap(), OpType::Delete as u8);
-        }
-        assert_eq!(num_rows, 1);
-    }
-
     #[tokio::test]
     async fn test_single_range() {
         let memtable = new_test_memtable(true, MergeMode::LastRow);
@@ -745,7 +671,11 @@ mod tests {
         };
         memtable.write_bulk(part).unwrap();
 
-        let mut iter = memtable.iter(None, None, None).unwrap();
+        let mut iter = memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let batch = iter.next().unwrap().unwrap();
         assert_eq!(2, batch.num_rows());
 
@@ -764,7 +694,11 @@ mod tests {
             OpType::Put,
         );
         memtable.write(&kvs).unwrap();
-        let mut iter = memtable.iter(None, None, None).unwrap();
+        let mut iter = memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let batch = iter.next().unwrap().unwrap();
         assert_eq!(3, batch.num_rows());
         assert_eq!(
@@ -854,7 +788,15 @@ mod tests {
 
         // Filter with sequence 0 should only return first write
         let mut iter = memtable
-            .iter(None, None, Some(SequenceRange::LtEq { max: 0 }))
+            .ranges(
+                None,
+                RangesOptions {
+                    sequence: Some(SequenceRange::LtEq { max: 0 }),
+                    ..Default::default()
+                },
+            )
+            .unwrap()
+            .build(None)
             .unwrap();
         let batch = iter.next().unwrap().unwrap();
         assert_eq!(1, batch.num_rows());
@@ -881,8 +823,8 @@ mod tests {
         .unwrap()
     }
 
-    #[tokio::test]
-    async fn test_write_read_large_string() {
+    #[test]
+    fn test_write_read_large_string() {
         let mut builder = RegionMetadataBuilder::new(RegionId::new(123, 456));
         builder
             .push_column_metadata(ColumnMetadata {
@@ -927,25 +869,12 @@ mod tests {
             .unwrap();
         let MemtableRanges { ranges, .. } =
             memtable.ranges(None, RangesOptions::default()).unwrap();
-        let mut source = if ranges.len() == 1 {
-            let only_range = ranges.into_values().next().unwrap();
-            Source::Iter(only_range.build_iter().unwrap())
-        } else {
-            let sources = ranges
-                .into_values()
-                .map(|r| r.build_iter().map(Source::Iter))
-                .collect::<error::Result<Vec<_>>>()
-                .unwrap();
-            let merge_reader = MergeReaderBuilder::from_sources(sources)
-                .build()
-                .await
-                .unwrap();
-            Source::Reader(Box::new(merge_reader))
-        };
-
         let mut rows = 0;
-        while let Some(b) = source.next_batch().await.unwrap() {
-            rows += b.num_rows();
+        for range in ranges.into_values() {
+            let iter = range.build_iter().unwrap();
+            for batch in iter {
+                rows += batch.unwrap().num_rows();
+            }
         }
         assert_eq!(rows, 2);
     }
diff --git a/src/mito2/src/memtable/simple_bulk_memtable/test_only.rs b/src/mito2/src/memtable/simple_bulk_memtable/test_only.rs
index b71a86c554..08edebdbb2 100644
--- a/src/mito2/src/memtable/simple_bulk_memtable/test_only.rs
+++ b/src/mito2/src/memtable/simple_bulk_memtable/test_only.rs
@@ -12,98 +12,12 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::collections::HashSet;
-use std::time::Instant;
-
 use store_api::metadata::RegionMetadataRef;
-use store_api::storage::{ColumnId, SequenceRange};
 
-use crate::error;
-use crate::memtable::simple_bulk_memtable::{Iter, SimpleBulkMemtable};
-use crate::memtable::time_series::Values;
-use crate::memtable::{BoxedBatchIterator, IterBuilder, MemScanMetrics};
-use crate::read::dedup::LastNonNullIter;
-use crate::region::options::MergeMode;
+use crate::memtable::simple_bulk_memtable::SimpleBulkMemtable;
 
 impl SimpleBulkMemtable {
     pub fn region_metadata(&self) -> RegionMetadataRef {
         self.region_metadata.clone()
     }
-
-    pub(crate) fn create_iter(
-        &self,
-        projection: Option<&[ColumnId]>,
-        sequence: Option<SequenceRange>,
-    ) -> error::Result<BatchIterBuilderDeprecated> {
-        let mut series = self.series.write().unwrap();
-
-        let values = if series.is_empty() {
-            None
-        } else {
-            Some(series.compact(&self.region_metadata)?.clone())
-        };
-        let projection = self.build_projection(projection);
-        Ok(BatchIterBuilderDeprecated {
-            region_metadata: self.region_metadata.clone(),
-            values,
-            projection,
-            dedup: self.dedup,
-            sequence,
-            merge_mode: self.merge_mode,
-        })
-    }
-}
-
-#[derive(Clone)]
-pub(crate) struct BatchIterBuilderDeprecated {
-    region_metadata: RegionMetadataRef,
-    values: Option<Values>,
-    projection: HashSet<ColumnId>,
-    sequence: Option<SequenceRange>,
-    dedup: bool,
-    merge_mode: MergeMode,
-}
-
-impl IterBuilder for BatchIterBuilderDeprecated {
-    fn build(&self, metrics: Option<MemScanMetrics>) -> error::Result<BoxedBatchIterator> {
-        let start_time = Instant::now();
-        let Some(values) = self.values.clone() else {
-            return Ok(Box::new(Iter { batch: None }));
-        };
-
-        let maybe_batch = values
-            .to_batch(
-                &[],
-                &self.region_metadata,
-                &self.projection,
-                self.sequence,
-                self.dedup,
-                self.merge_mode,
-            )
-            .map(Some)
-            .transpose();
-
-        // Collect metrics from the batch
-        if let Some(metrics) = metrics {
-            let (num_rows, num_batches) = match &maybe_batch {
-                Some(Ok(batch)) => (batch.num_rows(), 1),
-                _ => (0, 0),
-            };
-            let inner = crate::memtable::MemScanMetricsData {
-                total_series: 1,
-                num_rows,
-                num_batches,
-                scan_cost: start_time.elapsed(),
-            };
-            metrics.merge_inner(&inner);
-        }
-
-        let iter = Iter { batch: maybe_batch };
-
-        if self.merge_mode == MergeMode::LastNonNull {
-            Ok(Box::new(LastNonNullIter::new(iter)))
-        } else {
-            Ok(Box::new(iter))
-        }
-    }
 }
diff --git a/src/mito2/src/memtable/time_partition.rs b/src/mito2/src/memtable/time_partition.rs
index 6f11c813cb..ee695aceb8 100644
--- a/src/mito2/src/memtable/time_partition.rs
+++ b/src/mito2/src/memtable/time_partition.rs
@@ -827,6 +827,7 @@ mod tests {
     use super::*;
     use crate::memtable::partition_tree::PartitionTreeMemtableBuilder;
     use crate::memtable::time_series::TimeSeriesMemtableBuilder;
+    use crate::memtable::{IterBuilder, RangesOptions};
     use crate::test_util::memtable_util::{self, collect_iter_timestamps};
 
     #[test]
@@ -852,7 +853,11 @@ mod tests {
         partitions.list_memtables(&mut memtables);
         assert_eq!(0, memtables[0].id());
 
-        let iter = memtables[0].iter(None, None, None).unwrap();
+        let iter = memtables[0]
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let timestamps = collect_iter_timestamps(iter);
         assert_eq!(&[1000, 3000, 5000, 6000, 7000], &timestamps[..]);
     }
@@ -890,7 +895,11 @@ mod tests {
 
         let mut memtables = Vec::new();
         partitions.list_memtables(&mut memtables);
-        let iter = memtables[0].iter(None, None, None).unwrap();
+        let iter = memtables[0]
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let timestamps = collect_iter_timestamps(iter);
         assert_eq!(&[0, 2000, 3000, 4000, 5000, 7000], &timestamps[..]);
         let parts = partitions.list_partitions();
@@ -943,7 +952,12 @@ mod tests {
         let partitions = new_multi_partitions(&metadata);
 
         let parts = partitions.list_partitions();
-        let iter = parts[0].memtable.iter(None, None, None).unwrap();
+        let iter = parts[0]
+            .memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let timestamps = collect_iter_timestamps(iter);
         assert_eq!(0, parts[0].memtable.id());
         assert_eq!(
@@ -955,7 +969,12 @@ mod tests {
             parts[0].time_range.max_timestamp
         );
         assert_eq!(&[0, 2000, 3000, 4000], &timestamps[..]);
-        let iter = parts[1].memtable.iter(None, None, None).unwrap();
+        let iter = parts[1]
+            .memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         assert_eq!(1, parts[1].memtable.id());
         let timestamps = collect_iter_timestamps(iter);
         assert_eq!(&[5000, 7000], &timestamps[..]);
@@ -1273,7 +1292,12 @@ mod tests {
 
         let parts = partitions.list_partitions();
         assert_eq!(1, parts.len());
-        let iter = parts[0].memtable.iter(None, None, None).unwrap();
+        let iter = parts[0]
+            .memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let timestamps = collect_iter_timestamps(iter);
         assert_eq!(&[1000, 2000, 3000], &timestamps[..]);
 
@@ -1284,11 +1308,21 @@ mod tests {
         let parts = partitions.list_partitions();
         assert_eq!(2, parts.len());
         // Check first partition [0, 5000)
-        let iter = parts[0].memtable.iter(None, None, None).unwrap();
+        let iter = parts[0]
+            .memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let timestamps = collect_iter_timestamps(iter);
         assert_eq!(&[1000, 2000, 3000, 4000], &timestamps[..]);
         // Check second partition [5000, 10000)
-        let iter = parts[1].memtable.iter(None, None, None).unwrap();
+        let iter = parts[1]
+            .memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let timestamps = collect_iter_timestamps(iter);
         assert_eq!(&[5000, 6000], &timestamps[..]);
 
@@ -1301,7 +1335,12 @@ mod tests {
         assert_eq!(3, parts.len());
 
         // Check new partition [10000, 15000)
-        let iter = parts[2].memtable.iter(None, None, None).unwrap();
+        let iter = parts[2]
+            .memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let timestamps = collect_iter_timestamps(iter);
         assert_eq!(&[11000, 12000], &timestamps[..]);
 
@@ -1314,7 +1353,12 @@ mod tests {
 
         let parts = partitions.list_partitions();
         assert_eq!(1, parts.len());
-        let iter = parts[0].memtable.iter(None, None, None).unwrap();
+        let iter = parts[0]
+            .memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let timestamps = collect_iter_timestamps(iter);
         assert_eq!(&[1000, 5000, 9000], &timestamps[..]);
     }
diff --git a/src/mito2/src/memtable/time_series.rs b/src/mito2/src/memtable/time_series.rs
index 271a9343eb..f5b772b076 100644
--- a/src/mito2/src/memtable/time_series.rs
+++ b/src/mito2/src/memtable/time_series.rs
@@ -51,15 +51,18 @@ use crate::memtable::bulk::part::BulkPart;
 use crate::memtable::simple_bulk_memtable::SimpleBulkMemtable;
 use crate::memtable::stats::WriteMetrics;
 use crate::memtable::{
-    AllocTracker, BatchToRecordBatchContext, BoxedBatchIterator, IterBuilder, KeyValues,
-    MemScanMetrics, Memtable, MemtableBuilder, MemtableId, MemtableRange, MemtableRangeContext,
-    MemtableRanges, MemtableRef, MemtableStats, RangesOptions, read_column_ids_from_projection,
+    AllocTracker, BatchToRecordBatchContext, BoxedBatchIterator, BoxedRecordBatchIterator,
+    IterBuilder, KeyValues, MemScanMetrics, Memtable, MemtableBuilder, MemtableId, MemtableRange,
+    MemtableRangeContext, MemtableRanges, MemtableRef, MemtableStats, RangesOptions,
+    read_column_ids_from_projection,
 };
 use crate::metrics::{
     MEMTABLE_ACTIVE_FIELD_BUILDER_COUNT, MEMTABLE_ACTIVE_SERIES_COUNT, READ_ROWS_TOTAL,
     READ_STAGE_ELAPSED,
 };
 use crate::read::dedup::LastNonNullIter;
+use crate::read::prune::PruneTimeIterator;
+use crate::read::scan_region::PredicateGroup;
 use crate::read::{Batch, BatchBuilder, BatchColumn};
 use crate::region::options::MergeMode;
 
@@ -267,39 +270,6 @@ impl Memtable for TimeSeriesMemtable {
         Ok(())
     }
 
-    #[cfg(any(test, feature = "test"))]
-    fn iter(
-        &self,
-        projection: Option<&[ColumnId]>,
-        filters: Option<Predicate>,
-        sequence: Option<SequenceRange>,
-    ) -> Result<BoxedBatchIterator> {
-        let projection = if let Some(projection) = projection {
-            projection.iter().copied().collect()
-        } else {
-            self.region_metadata
-                .field_columns()
-                .map(|c| c.column_id)
-                .collect()
-        };
-
-        let iter = self.series_set.iter_series(
-            projection,
-            filters,
-            self.dedup,
-            self.merge_mode,
-            sequence,
-            None,
-        )?;
-
-        if self.merge_mode == MergeMode::LastNonNull {
-            let iter = LastNonNullIter::new(iter);
-            Ok(Box::new(iter))
-        } else {
-            Ok(Box::new(iter))
-        }
-    }
-
     fn ranges(
         &self,
         projection: Option<&[ColumnId]>,
@@ -316,25 +286,20 @@ impl Memtable for TimeSeriesMemtable {
                 .map(|c| c.column_id)
                 .collect()
         };
-        let builder = Box::new(TimeSeriesIterBuilder {
-            series_set: self.series_set.clone(),
-            projection,
-            predicate: predicate.predicate().cloned(),
-            dedup: self.dedup,
-            merge_mode: self.merge_mode,
-            sequence,
-        });
-        let adapter_context = Arc::new(BatchToRecordBatchContext::new(
+        let batch_to_record_batch = Arc::new(BatchToRecordBatchContext::new(
             self.region_metadata.clone(),
             read_column_ids,
         ));
-        let context = Arc::new(MemtableRangeContext::new_with_batch_to_record_batch(
-            self.id,
-            builder,
-            predicate,
-            Some(adapter_context),
-        ));
-
+        let builder = Box::new(TimeSeriesIterBuilder {
+            series_set: self.series_set.clone(),
+            projection,
+            predicate: predicate.clone(),
+            dedup: self.dedup,
+            merge_mode: self.merge_mode,
+            sequence,
+            batch_to_record_batch,
+        });
+        let context = Arc::new(MemtableRangeContext::new(self.id, builder, predicate));
         let range_stats = self.stats();
         let range = MemtableRange::new(context, range_stats);
         Ok(MemtableRanges {
@@ -476,7 +441,7 @@ impl SeriesSet {
     fn iter_series(
         &self,
         projection: HashSet<ColumnId>,
-        predicate: Option<Predicate>,
+        predicate: PredicateGroup,
         dedup: bool,
         merge_mode: MergeMode,
         sequence: Option<SequenceRange>,
@@ -493,7 +458,7 @@ impl SeriesSet {
             self.region_metadata.clone(),
             self.series.clone(),
             projection,
-            predicate,
+            predicate.predicate().cloned(),
             primary_key_schema,
             primary_key_datatypes,
             self.codec.clone(),
@@ -602,6 +567,7 @@ impl Iter {
                 num_rows: self.metrics.num_rows,
                 num_batches: self.metrics.num_batches,
                 scan_cost: self.metrics.scan_cost,
+                ..Default::default()
             };
             mem_scan_metrics.merge_inner(&inner);
         }
@@ -927,7 +893,9 @@ impl ValueBuilder {
             size += field_value.data_size();
             if !field_value.is_null() || self.fields[idx].is_some() {
                 if let Some(field) = self.fields[idx].as_mut() {
-                    let _ = field.push(field_value);
+                    field
+                        .push(field_value)
+                        .unwrap_or_else(|e| panic!("Failed to push field value: {e:?}"));
                 } else {
                     let mut mutable_vector =
                         if let ConcreteDataType::String(_) = &self.field_types[idx] {
@@ -1278,10 +1246,11 @@ impl From<ValueBuilder> for Values {
 struct TimeSeriesIterBuilder {
     series_set: SeriesSet,
     projection: HashSet<ColumnId>,
-    predicate: Option<Predicate>,
+    predicate: PredicateGroup,
     dedup: bool,
     sequence: Option<SequenceRange>,
     merge_mode: MergeMode,
+    batch_to_record_batch: Arc<BatchToRecordBatchContext>,
 }
 
 impl IterBuilder for TimeSeriesIterBuilder {
@@ -1301,6 +1270,25 @@ impl IterBuilder for TimeSeriesIterBuilder {
             Ok(Box::new(iter))
         }
     }
+
+    fn is_record_batch(&self) -> bool {
+        true
+    }
+
+    fn build_record_batch(
+        &self,
+        time_range: Option<(Timestamp, Timestamp)>,
+        metrics: Option<MemScanMetrics>,
+    ) -> Result<BoxedRecordBatchIterator> {
+        let iter = self.build(metrics)?;
+        let iter: BoxedBatchIterator = if let Some(time_range) = time_range {
+            let time_filters = self.predicate.time_filters();
+            Box::new(PruneTimeIterator::new(iter, time_range, time_filters))
+        } else {
+            iter
+        };
+        Ok(self.batch_to_record_batch.adapt_iter(iter))
+    }
 }
 
 #[cfg(test)]
@@ -1798,7 +1786,9 @@ mod tests {
             *expected_ts.entry(ts).or_default() += if dedup { 1 } else { 2 };
         }
 
-        let iter = memtable.iter(None, None, None).unwrap();
+        let ranges = memtable.ranges(None, RangesOptions::default()).unwrap();
+        let range = ranges.ranges.into_values().next().unwrap();
+        let iter = range.build_iter().unwrap();
         let mut read = HashMap::new();
 
         for ts in iter
@@ -1838,7 +1828,11 @@ mod tests {
         let memtable = TimeSeriesMemtable::new(schema, 42, None, true, MergeMode::LastRow);
         memtable.write(&kvs).unwrap();
 
-        let iter = memtable.iter(Some(&[3]), None, None).unwrap();
+        let iter = memtable
+            .ranges(Some(&[3]), RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
 
         let mut v0_all = vec![];
 
@@ -1917,7 +1911,11 @@ mod tests {
                 barrier.wait();
 
                 for _ in 0..10 {
-                    let iter = memtable.iter(None, None, None).unwrap();
+                    let iter = memtable
+                        .ranges(None, RangesOptions::default())
+                        .unwrap()
+                        .build(None)
+                        .unwrap();
                     for batch_result in iter {
                         let _ = batch_result.unwrap();
                     }
@@ -1936,7 +1934,11 @@ mod tests {
             handle.join().unwrap();
         }
 
-        let iter = memtable.iter(None, None, None).unwrap();
+        let iter = memtable
+            .ranges(None, RangesOptions::default())
+            .unwrap()
+            .build(None)
+            .unwrap();
         let mut series_count = 0;
         let mut row_count = 0;
 
@@ -2033,4 +2035,265 @@ mod tests {
         all_timestamps.sort();
         assert_eq!(vec![3, 4, 5, 6, 7], all_timestamps);
     }
+
+    /// Helper to create a TimeSeriesIterBuilder from a memtable and schema.
+    fn build_iter_builder(
+        schema: &RegionMetadataRef,
+        memtable: &TimeSeriesMemtable,
+        projection: Option<&[ColumnId]>,
+        dedup: bool,
+        merge_mode: MergeMode,
+        sequence: Option<SequenceRange>,
+    ) -> TimeSeriesIterBuilder {
+        let read_column_ids = read_column_ids_from_projection(schema, projection);
+        let field_projection = if let Some(projection) = projection {
+            projection.iter().copied().collect()
+        } else {
+            schema.field_columns().map(|c| c.column_id).collect()
+        };
+        let adapter_context = Arc::new(BatchToRecordBatchContext::new(
+            schema.clone(),
+            read_column_ids,
+        ));
+        TimeSeriesIterBuilder {
+            series_set: memtable.series_set.clone(),
+            projection: field_projection,
+            predicate: PredicateGroup::default(),
+            dedup,
+            merge_mode,
+            sequence,
+            batch_to_record_batch: adapter_context,
+        }
+    }
+
+    #[test]
+    fn test_iter_builder_build_record_batch_basic() {
+        let schema = schema_for_test();
+        let memtable = TimeSeriesMemtable::new(schema.clone(), 1, None, true, MergeMode::LastRow);
+
+        let kvs = build_key_values(&schema, "hello".to_string(), 42, 10);
+        memtable.write(&kvs).unwrap();
+
+        let builder = build_iter_builder(&schema, &memtable, None, true, MergeMode::LastRow, None);
+
+        let mut iter = builder.build_record_batch(None, None).unwrap();
+        let rb = iter.next().transpose().unwrap().unwrap();
+        assert_eq!(10, rb.num_rows());
+
+        let rb_schema = rb.schema();
+        let col_names: Vec<_> = rb_schema
+            .fields()
+            .iter()
+            .map(|f| f.name().as_str())
+            .collect();
+        assert_eq!(
+            col_names,
+            vec![
+                "k0",
+                "k1",
+                "v0",
+                "v1",
+                "ts",
+                "__primary_key",
+                "__sequence",
+                "__op_type",
+            ]
+        );
+
+        assert!(iter.next().is_none());
+    }
+
+    #[test]
+    fn test_iter_builder_build_record_batch_with_projection() {
+        let schema = schema_for_test();
+        let memtable = TimeSeriesMemtable::new(schema.clone(), 1, None, true, MergeMode::LastRow);
+
+        let kvs = build_key_values(&schema, "test".to_string(), 1, 5);
+        memtable.write(&kvs).unwrap();
+
+        // Project only field v0 (column_id=3) and ts (column_id=2).
+        let projection = vec![2, 3];
+        let builder = build_iter_builder(
+            &schema,
+            &memtable,
+            Some(&projection),
+            true,
+            MergeMode::LastRow,
+            None,
+        );
+
+        let mut iter = builder.build_record_batch(None, None).unwrap();
+        let rb = iter.next().transpose().unwrap().unwrap();
+        assert_eq!(5, rb.num_rows());
+
+        let rb_schema = rb.schema();
+        let col_names: Vec<_> = rb_schema
+            .fields()
+            .iter()
+            .map(|f| f.name().as_str())
+            .collect();
+        // Only projected columns + internal columns.
+        assert_eq!(
+            col_names,
+            vec!["v0", "ts", "__primary_key", "__sequence", "__op_type",]
+        );
+
+        assert!(iter.next().is_none());
+    }
+
+    #[test]
+    fn test_iter_builder_build_record_batch_multiple_series() {
+        let schema = schema_for_test();
+        let memtable = TimeSeriesMemtable::new(schema.clone(), 1, None, true, MergeMode::LastRow);
+
+        let kvs_a = build_key_values(&schema, "aaa".to_string(), 1, 3);
+        let kvs_b = build_key_values(&schema, "bbb".to_string(), 2, 4);
+        memtable.write(&kvs_a).unwrap();
+        memtable.write(&kvs_b).unwrap();
+
+        let builder = build_iter_builder(&schema, &memtable, None, true, MergeMode::LastRow, None);
+
+        let iter = builder.build_record_batch(None, None).unwrap();
+        let mut total_rows = 0;
+        for rb in iter {
+            let rb = rb.unwrap();
+            total_rows += rb.num_rows();
+            assert_eq!(8, rb.num_columns());
+        }
+        assert_eq!(7, total_rows);
+    }
+
+    #[test]
+    fn test_iter_builder_build_record_batch_dedup() {
+        let schema = schema_for_test();
+        let memtable = TimeSeriesMemtable::new(schema.clone(), 1, None, true, MergeMode::LastRow);
+
+        // Write same data twice — dedup should keep only one copy per timestamp.
+        let kvs = build_key_values(&schema, "dup".to_string(), 10, 5);
+        memtable.write(&kvs).unwrap();
+        memtable.write(&kvs).unwrap();
+
+        let builder = build_iter_builder(&schema, &memtable, None, true, MergeMode::LastRow, None);
+
+        let iter = builder.build_record_batch(None, None).unwrap();
+        let total_rows: usize = iter.map(|rb| rb.unwrap().num_rows()).sum();
+        assert_eq!(5, total_rows);
+    }
+
+    #[test]
+    fn test_iter_builder_build_record_batch_no_dedup() {
+        let schema = schema_for_test();
+        let memtable = TimeSeriesMemtable::new(schema.clone(), 1, None, false, MergeMode::LastRow);
+
+        let kvs = build_key_values(&schema, "dup".to_string(), 10, 5);
+        memtable.write(&kvs).unwrap();
+        memtable.write(&kvs).unwrap();
+
+        let builder = build_iter_builder(&schema, &memtable, None, false, MergeMode::LastRow, None);
+
+        let iter = builder.build_record_batch(None, None).unwrap();
+        let total_rows: usize = iter.map(|rb| rb.unwrap().num_rows()).sum();
+        assert_eq!(10, total_rows);
+    }
+
+    #[test]
+    fn test_iter_builder_build_record_batch_with_sequence_filter() {
+        let schema = schema_for_test();
+        let memtable = TimeSeriesMemtable::new(schema.clone(), 1, None, true, MergeMode::LastRow);
+
+        // build_key_values creates a mutation with base sequence=0.
+        // Each row gets sequence = base + row_index, so 5 rows get sequences 0,1,2,3,4.
+        let kvs = build_key_values(&schema, "seq".to_string(), 1, 5);
+        memtable.write(&kvs).unwrap();
+
+        // Filter to sequence > 4 — should yield no rows.
+        let builder = build_iter_builder(
+            &schema,
+            &memtable,
+            None,
+            true,
+            MergeMode::LastRow,
+            Some(SequenceRange::Gt { min: 4 }),
+        );
+
+        let iter = builder.build_record_batch(None, None).unwrap();
+        let total_rows: usize = iter.map(|rb| rb.unwrap().num_rows()).sum();
+        assert_eq!(0, total_rows);
+
+        // Filter to sequence <= 2 — should yield 3 rows (sequences 0, 1, 2).
+        let builder = build_iter_builder(
+            &schema,
+            &memtable,
+            None,
+            true,
+            MergeMode::LastRow,
+            Some(SequenceRange::LtEq { max: 2 }),
+        );
+
+        let iter = builder.build_record_batch(None, None).unwrap();
+        let total_rows: usize = iter.map(|rb| rb.unwrap().num_rows()).sum();
+        assert_eq!(3, total_rows);
+    }
+
+    #[test]
+    fn test_iter_builder_build_record_batch_data_correctness() {
+        use datatypes::arrow::array::{
+            Float64Array, Int64Array, TimestampMillisecondArray, UInt8Array,
+        };
+
+        let schema = schema_for_test();
+        let memtable = TimeSeriesMemtable::new(schema.clone(), 1, None, true, MergeMode::LastRow);
+
+        let kvs = build_key_values(&schema, "check".to_string(), 7, 3);
+        memtable.write(&kvs).unwrap();
+
+        let builder = build_iter_builder(&schema, &memtable, None, true, MergeMode::LastRow, None);
+
+        let mut iter = builder.build_record_batch(None, None).unwrap();
+        let rb = iter.next().transpose().unwrap().unwrap();
+        assert_eq!(3, rb.num_rows());
+
+        // Verify timestamp values.
+        let ts_col = rb
+            .column_by_name("ts")
+            .unwrap()
+            .as_any()
+            .downcast_ref::<TimestampMillisecondArray>()
+            .unwrap();
+        let timestamps: Vec<_> = (0..ts_col.len()).map(|i| ts_col.value(i)).collect();
+        assert_eq!(vec![0, 1, 2], timestamps);
+
+        // Verify field v0 values.
+        let v0_col = rb
+            .column_by_name("v0")
+            .unwrap()
+            .as_any()
+            .downcast_ref::<Int64Array>()
+            .unwrap();
+        let v0_values: Vec<_> = (0..v0_col.len()).map(|i| v0_col.value(i)).collect();
+        assert_eq!(vec![0, 1, 2], v0_values);
+
+        // Verify field v1 values.
+        let v1_col = rb
+            .column_by_name("v1")
+            .unwrap()
+            .as_any()
+            .downcast_ref::<Float64Array>()
+            .unwrap();
+        let v1_values: Vec<_> = (0..v1_col.len()).map(|i| v1_col.value(i)).collect();
+        assert_eq!(vec![0.0, 1.0, 2.0], v1_values);
+
+        // Verify op_type is all Put (1).
+        let op_col = rb
+            .column_by_name("__op_type")
+            .unwrap()
+            .as_any()
+            .downcast_ref::<UInt8Array>()
+            .unwrap();
+        for i in 0..op_col.len() {
+            assert_eq!(OpType::Put as u8, op_col.value(i));
+        }
+
+        assert!(iter.next().is_none());
+    }
 }
diff --git a/src/mito2/src/metrics.rs b/src/mito2/src/metrics.rs
index c0537567f9..30a7ac765c 100644
--- a/src/mito2/src/metrics.rs
+++ b/src/mito2/src/metrics.rs
@@ -244,10 +244,15 @@ lazy_static! {
         "greptime_mito_scan_memory_usage_bytes",
         "current scan memory usage in bytes"
     ).unwrap();
-    /// Counter of rejected scan requests due to memory limit.
+    /// Counter of scan allocation attempts that could not acquire memory immediately.
+    pub static ref SCAN_MEMORY_EXHAUSTED_TOTAL: IntCounter = register_int_counter!(
+        "greptime_mito_scan_memory_exhausted_total",
+        "total number of times scan memory was unavailable for immediate acquisition"
+    ).unwrap();
+    /// Counter of scan requests that ultimately failed due to memory pressure.
     pub static ref SCAN_REQUESTS_REJECTED_TOTAL: IntCounter = register_int_counter!(
         "greptime_mito_scan_requests_rejected_total",
-        "total number of scan requests rejected due to memory limit"
+        "total number of scan requests that ultimately failed due to memory limit"
     ).unwrap();
     /// Gauge for active file range builders in the pruner.
     pub static ref PRUNER_ACTIVE_BUILDERS: IntGauge = register_int_gauge!(
diff --git a/src/mito2/src/read.rs b/src/mito2/src/read.rs
index 5fbd63ce8b..5a71bd1a4e 100644
--- a/src/mito2/src/read.rs
+++ b/src/mito2/src/read.rs
@@ -21,12 +21,15 @@ pub mod flat_dedup;
 pub mod flat_merge;
 pub mod flat_projection;
 pub mod last_row;
-pub mod merge;
-pub mod plain_batch;
 pub mod projection;
 pub(crate) mod prune;
 pub(crate) mod pruner;
 pub mod range;
+#[cfg(feature = "test")]
+pub mod range_cache;
+#[cfg(not(feature = "test"))]
+pub(crate) mod range_cache;
+pub(crate) mod read_columns;
 pub mod scan_region;
 pub mod scan_util;
 pub(crate) mod seq_scan;
@@ -34,11 +37,12 @@ pub mod series_scan;
 pub mod stream;
 pub(crate) mod unordered_scan;
 
-use std::collections::{HashMap, HashSet};
+use std::collections::HashMap;
 use std::sync::Arc;
 use std::time::Duration;
 
 use api::v1::OpType;
+use arrow_schema::SchemaRef;
 use async_trait::async_trait;
 use common_time::Timestamp;
 use datafusion_common::arrow::array::UInt8Array;
@@ -61,7 +65,6 @@ use futures::TryStreamExt;
 use futures::stream::BoxStream;
 use mito_codec::row_converter::{CompositeValues, PrimaryKeyCodec};
 use snafu::{OptionExt, ResultExt, ensure};
-use store_api::metadata::RegionMetadata;
 use store_api::storage::{ColumnId, SequenceNumber, SequenceRange};
 
 use crate::error::{
@@ -69,8 +72,6 @@ use crate::error::{
     Result,
 };
 use crate::memtable::{BoxedBatchIterator, BoxedRecordBatchIterator};
-use crate::read::prune::PruneReader;
-
 /// Storage internal representation of a batch of rows for a primary key (time series).
 ///
 /// Rows are sorted by primary key, timestamp, sequence desc, op_type desc. Fields
@@ -171,6 +172,7 @@ impl Batch {
     }
 
     /// Create an empty [`Batch`].
+    #[allow(dead_code)]
     pub(crate) fn empty() -> Self {
         Self {
             primary_key: vec![],
@@ -570,24 +572,6 @@ impl Batch {
         size
     }
 
-    /// Returns ids and datatypes of fields in the [Batch] after applying the `projection`.
-    pub(crate) fn projected_fields(
-        metadata: &RegionMetadata,
-        projection: &[ColumnId],
-    ) -> Vec<(ColumnId, ConcreteDataType)> {
-        let projected_ids: HashSet<_> = projection.iter().copied().collect();
-        metadata
-            .field_columns()
-            .filter_map(|column| {
-                if projected_ids.contains(&column.column_id) {
-                    Some((column.column_id, column.column_schema.data_type.clone()))
-                } else {
-                    None
-                }
-            })
-            .collect()
-    }
-
     /// Returns timestamps in a native slice or `None` if the batch is empty.
     pub(crate) fn timestamps_native(&self) -> Option<&[i64]> {
         if self.timestamps.is_empty() {
@@ -673,6 +657,7 @@ impl Batch {
 
     /// Checks the batch is monotonic by timestamps.
     #[cfg(debug_assertions)]
+    #[allow(dead_code)]
     pub(crate) fn check_monotonic(&self) -> Result<(), String> {
         use std::cmp::Ordering;
         if self.timestamps_native().is_none() {
@@ -715,6 +700,7 @@ impl Batch {
 
     /// Returns Ok if the given batch is behind the current batch.
     #[cfg(debug_assertions)]
+    #[allow(dead_code)]
     pub(crate) fn check_next_batch(&self, other: &Batch) -> Result<(), String> {
         // Checks the primary key
         if self.primary_key() < other.primary_key() {
@@ -794,6 +780,7 @@ impl Batch {
 /// A struct to check the batch is monotonic.
 #[cfg(debug_assertions)]
 #[derive(Default)]
+#[allow(dead_code)]
 pub(crate) struct BatchChecker {
     last_batch: Option<Batch>,
     start: Option<Timestamp>,
@@ -801,6 +788,7 @@ pub(crate) struct BatchChecker {
 }
 
 #[cfg(debug_assertions)]
+#[allow(dead_code)]
 impl BatchChecker {
     /// Attaches the given start timestamp to the checker.
     pub(crate) fn with_start(mut self, start: Option<Timestamp>) -> Self {
@@ -1104,8 +1092,6 @@ pub enum Source {
     Iter(BoxedBatchIterator),
     /// Source from a [BoxedBatchStream].
     Stream(BoxedBatchStream),
-    /// Source from a [PruneReader].
-    PruneReader(PruneReader),
 }
 
 impl Source {
@@ -1115,25 +1101,64 @@ impl Source {
             Source::Reader(reader) => reader.next_batch().await,
             Source::Iter(iter) => iter.next().transpose(),
             Source::Stream(stream) => stream.try_next().await,
-            Source::PruneReader(reader) => reader.next_batch().await,
         }
     }
 }
 
 /// Async [RecordBatch] reader and iterator wrapper for flat format.
-pub enum FlatSource {
+pub struct FlatSource {
+    schema: SchemaRef,
+    inner: FlatSourceInner,
+}
+
+impl FlatSource {
+    /// Create a [FlatSource] from a [BoxedRecordBatchIterator] and its schema.
+    pub fn new_iter(schema: SchemaRef, iter: BoxedRecordBatchIterator) -> Self {
+        Self {
+            schema,
+            inner: FlatSourceInner::Iter(iter),
+        }
+    }
+
+    /// Create a [FlatSource] from a [BoxedRecordBatchStream] and its schema.
+    pub fn new_stream(schema: SchemaRef, stream: BoxedRecordBatchStream) -> Self {
+        Self {
+            schema,
+            inner: FlatSourceInner::Stream(stream),
+        }
+    }
+
+    #[expect(unused)]
+    fn schema(&self) -> &SchemaRef {
+        &self.schema
+    }
+
+    pub async fn next_batch(&mut self) -> Result<Option<RecordBatch>> {
+        self.inner.next_batch().await
+    }
+
+    #[cfg(test)]
+    pub(crate) fn take_iter(self) -> BoxedRecordBatchIterator {
+        match self.inner {
+            FlatSourceInner::Iter(iter) => iter,
+            FlatSourceInner::Stream(_) => unreachable!(),
+        }
+    }
+}
+
+enum FlatSourceInner {
     /// Source from a [BoxedRecordBatchIterator].
     Iter(BoxedRecordBatchIterator),
     /// Source from a [BoxedRecordBatchStream].
     Stream(BoxedRecordBatchStream),
 }
 
-impl FlatSource {
+impl FlatSourceInner {
     /// Returns next [RecordBatch] from this data source.
     pub async fn next_batch(&mut self) -> Result<Option<RecordBatch>> {
         match self {
-            FlatSource::Iter(iter) => iter.next().transpose(),
-            FlatSource::Stream(stream) => stream.try_next().await,
+            Self::Iter(iter) => iter.next().transpose(),
+            Self::Stream(stream) => stream.try_next().await,
         }
     }
 }
diff --git a/src/mito2/src/read/batch_adapter.rs b/src/mito2/src/read/batch_adapter.rs
index 461dbeba69..1b9821e11b 100644
--- a/src/mito2/src/read/batch_adapter.rs
+++ b/src/mito2/src/read/batch_adapter.rs
@@ -35,7 +35,7 @@ use crate::error::{
 };
 use crate::memtable::BoxedBatchIterator;
 use crate::read::Batch;
-use crate::sst::{internal_fields, tag_maybe_to_dictionary_field};
+use crate::sst::{internal_fields, tag_maybe_to_dictionary_field, with_field_id};
 
 /// Adapts a [`BoxedBatchIterator`] into an `Iterator<Item = Result<RecordBatch>>`
 /// producing flat-format record batches.
@@ -59,7 +59,7 @@ impl BatchToRecordBatchAdapter {
     /// - `metadata`: region metadata describing the schema.
     /// - `codec`: codec for decoding the encoded primary key bytes.
     /// - `read_column_ids`: projected column ids to read.
-    pub(crate) fn new(
+    pub fn new(
         iter: BoxedBatchIterator,
         metadata: RegionMetadataRef,
         codec: Arc<dyn PrimaryKeyCodec>,
@@ -212,38 +212,45 @@ fn compute_output_arrow_schema(
         if !read_column_id_set.contains(&column_metadata.column_id) {
             continue;
         }
-        let field = Arc::new(Field::new(
+        let field = Field::new(
             &column_metadata.column_schema.name,
             column_metadata.column_schema.data_type.as_arrow_type(),
             column_metadata.column_schema.is_nullable(),
-        ));
-        let field = if column_metadata.semantic_type == SemanticType::Tag {
-            tag_maybe_to_dictionary_field(&column_metadata.column_schema.data_type, &field)
+        );
+        let field = with_field_id(field, column_metadata.column_id);
+
+        if column_metadata.semantic_type == SemanticType::Tag {
+            fields.push(tag_maybe_to_dictionary_field(
+                &column_metadata.column_schema.data_type,
+                &Arc::new(field),
+            ));
         } else {
-            field
-        };
-        fields.push(field);
+            fields.push(Arc::new(field));
+        }
     }
 
     for column_metadata in metadata.field_columns() {
         if !read_column_id_set.contains(&column_metadata.column_id) {
             continue;
         }
-        let field = Arc::new(Field::new(
+        let field = Field::new(
             &column_metadata.column_schema.name,
             column_metadata.column_schema.data_type.as_arrow_type(),
             column_metadata.column_schema.is_nullable(),
-        ));
-        fields.push(field);
+        );
+        fields.push(Arc::new(with_field_id(field, column_metadata.column_id)));
     }
 
     let time_index = metadata.time_index_column();
-    let time_index_field = Arc::new(Field::new(
+    let time_index_field = Field::new(
         &time_index.column_schema.name,
         time_index.column_schema.data_type.as_arrow_type(),
         time_index.column_schema.is_nullable(),
-    ));
-    fields.push(time_index_field);
+    );
+    fields.push(Arc::new(with_field_id(
+        time_index_field,
+        time_index.column_id,
+    )));
     fields.extend(internal_fields().iter().cloned());
 
     Arc::new(datatypes::arrow::datatypes::Schema::new(fields))
@@ -681,7 +688,7 @@ mod tests {
             BatchToRecordBatchAdapter::new(iter, metadata.clone(), codec, &read_column_ids);
         let rb = adapter.into_iter().next().unwrap().unwrap();
 
-        let mapper = FlatProjectionMapper::new(&metadata, [0, 3].into_iter()).unwrap();
+        let mapper = FlatProjectionMapper::new(&metadata, [0, 3]).unwrap();
         assert_eq!(rb.schema(), mapper.input_arrow_schema(false));
         // tag_0 + field_1 + ts + 3 internal columns.
         assert_eq!(6, rb.num_columns());
diff --git a/src/mito2/src/read/compat.rs b/src/mito2/src/read/compat.rs
index fd88749827..6d6f535068 100644
--- a/src/mito2/src/read/compat.rs
+++ b/src/mito2/src/read/compat.rs
@@ -18,7 +18,6 @@ use std::collections::HashMap;
 use std::sync::Arc;
 
 use api::v1::SemanticType;
-use common_recordbatch::recordbatch::align_json_array;
 use datatypes::arrow::array::{
     Array, ArrayRef, BinaryArray, BinaryBuilder, DictionaryArray, UInt32Array,
 };
@@ -28,7 +27,8 @@ use datatypes::arrow::record_batch::RecordBatch;
 use datatypes::data_type::ConcreteDataType;
 use datatypes::prelude::DataType;
 use datatypes::value::Value;
-use datatypes::vectors::{Helper, VectorRef};
+use datatypes::vectors::VectorRef;
+use datatypes::vectors::json::array::JsonArray;
 use mito_codec::row_converter::{
     CompositeValues, PrimaryKeyCodec, SortField, build_primary_key_codec,
     build_primary_key_codec_with_fields,
@@ -39,125 +39,13 @@ use store_api::metadata::{RegionMetadata, RegionMetadataRef};
 use store_api::storage::ColumnId;
 
 use crate::error::{
-    CastVectorSnafu, CompatReaderSnafu, ComputeArrowSnafu, ConvertVectorSnafu, CreateDefaultSnafu,
-    DecodeSnafu, EncodeSnafu, NewRecordBatchSnafu, RecordBatchSnafu, Result, UnexpectedSnafu,
-    UnsupportedOperationSnafu,
+    CompatReaderSnafu, ComputeArrowSnafu, ConvertValueSnafu, CreateDefaultSnafu, DecodeSnafu,
+    EncodeSnafu, NewRecordBatchSnafu, Result, UnsupportedOperationSnafu,
 };
 use crate::read::flat_projection::{FlatProjectionMapper, flat_projected_columns};
-use crate::read::projection::{PrimaryKeyProjectionMapper, ProjectionMapper};
-use crate::read::{Batch, BatchColumn, BatchReader};
 use crate::sst::parquet::flat_format::primary_key_column_index;
 use crate::sst::parquet::format::{FormatProjection, INTERNAL_COLUMN_NUM, PrimaryKeyArray};
-use crate::sst::{internal_fields, tag_maybe_to_dictionary_field};
-
-/// Reader to adapt schema of underlying reader to expected schema.
-pub struct CompatReader<R> {
-    /// Underlying reader.
-    reader: R,
-    /// Helper to compat batches.
-    compat: PrimaryKeyCompatBatch,
-}
-
-impl<R> CompatReader<R> {
-    /// Creates a new compat reader.
-    /// - `mapper` is built from the metadata users expect to see.
-    /// - `reader_meta` is the metadata of the input reader.
-    /// - `reader` is the input reader.
-    pub fn new(
-        mapper: &ProjectionMapper,
-        reader_meta: RegionMetadataRef,
-        reader: R,
-    ) -> Result<CompatReader<R>> {
-        Ok(CompatReader {
-            reader,
-            compat: PrimaryKeyCompatBatch::new(mapper, reader_meta)?,
-        })
-    }
-}
-
-#[async_trait::async_trait]
-impl<R: BatchReader> BatchReader for CompatReader<R> {
-    async fn next_batch(&mut self) -> Result<Option<Batch>> {
-        let Some(mut batch) = self.reader.next_batch().await? else {
-            return Ok(None);
-        };
-
-        batch = self.compat.compat_batch(batch)?;
-
-        Ok(Some(batch))
-    }
-}
-
-/// Helper to adapt schema of the batch to an expected schema.
-pub(crate) enum CompatBatch {
-    /// Adapter for primary key format.
-    PrimaryKey(PrimaryKeyCompatBatch),
-    /// Adapter for flat format.
-    Flat(FlatCompatBatch),
-}
-
-impl CompatBatch {
-    /// Returns the inner primary key batch adapter if this is a PrimaryKey format.
-    pub(crate) fn as_primary_key(&self) -> Option<&PrimaryKeyCompatBatch> {
-        match self {
-            CompatBatch::PrimaryKey(batch) => Some(batch),
-            _ => None,
-        }
-    }
-
-    /// Returns the inner flat batch adapter if this is a Flat format.
-    pub(crate) fn as_flat(&self) -> Option<&FlatCompatBatch> {
-        match self {
-            CompatBatch::Flat(batch) => Some(batch),
-            _ => None,
-        }
-    }
-}
-
-/// A helper struct to adapt schema of the batch to an expected schema.
-pub(crate) struct PrimaryKeyCompatBatch {
-    /// Optional primary key adapter.
-    rewrite_pk: Option<RewritePrimaryKey>,
-    /// Optional primary key adapter.
-    compat_pk: Option<CompatPrimaryKey>,
-    /// Optional fields adapter.
-    compat_fields: Option<CompatFields>,
-}
-
-impl PrimaryKeyCompatBatch {
-    /// Creates a new [CompatBatch].
-    /// - `mapper` is built from the metadata users expect to see.
-    /// - `reader_meta` is the metadata of the input reader.
-    pub(crate) fn new(mapper: &ProjectionMapper, reader_meta: RegionMetadataRef) -> Result<Self> {
-        let rewrite_pk = may_rewrite_primary_key(mapper.metadata(), &reader_meta);
-        let compat_pk = may_compat_primary_key(mapper.metadata(), &reader_meta)?;
-        let mapper = mapper.as_primary_key().context(UnexpectedSnafu {
-            reason: "Unexpected format",
-        })?;
-        let compat_fields = may_compat_fields(mapper, &reader_meta)?;
-
-        Ok(Self {
-            rewrite_pk,
-            compat_pk,
-            compat_fields,
-        })
-    }
-
-    /// Adapts the `batch` to the expected schema.
-    pub(crate) fn compat_batch(&self, mut batch: Batch) -> Result<Batch> {
-        if let Some(rewrite_pk) = &self.rewrite_pk {
-            batch = rewrite_pk.compat(batch)?;
-        }
-        if let Some(compat_pk) = &self.compat_pk {
-            batch = compat_pk.compat(batch)?;
-        }
-        if let Some(compat_fields) = &self.compat_fields {
-            batch = compat_fields.compat(batch)?;
-        }
-
-        Ok(batch)
-    }
-}
+use crate::sst::{internal_fields, tag_maybe_to_dictionary_field, with_field_id};
 
 /// Returns true if `left` and `right` have same columns and primary key encoding.
 pub(crate) fn has_same_columns_and_pk_encoding(
@@ -255,12 +143,19 @@ impl FlatCompatBatch {
             let column_field = &expect_metadata.schema.arrow_schema().fields()[column_index];
             // For tag columns, we need to create a dictionary field.
             if expect_column.semantic_type == SemanticType::Tag {
-                fields.push(tag_maybe_to_dictionary_field(
+                let field = tag_maybe_to_dictionary_field(
                     &expect_column.column_schema.data_type,
                     column_field,
-                ));
+                );
+                fields.push(Arc::new(with_field_id(
+                    (*field).clone(),
+                    expect_column.column_id,
+                )));
             } else {
-                fields.push(column_field.clone());
+                fields.push(Arc::new(with_field_id(
+                    (**column_field).clone(),
+                    expect_column.column_id,
+                )));
             };
 
             if let Some((index, actual_data_type)) = actual_schema_index.get(column_id) {
@@ -292,7 +187,6 @@ impl FlatCompatBatch {
                         ),
                     })?;
                 index_or_defaults.push(IndexOrDefault::DefaultValue {
-                    column_id: expect_column.column_id,
                     default_vector,
                     semantic_type: expect_column.semantic_type,
                 });
@@ -353,9 +247,12 @@ impl FlatCompatBatch {
                     let old_column = batch.column(*pos);
 
                     if let Some(ty) = cast_type {
-                        let casted = if let Some(json_type) = ty.as_json() {
-                            align_json_array(old_column, &json_type.as_arrow_type())
-                                .context(RecordBatchSnafu)?
+                        let casted = if let Some(json_type) = ty.as_json()
+                            && json_type.is_json2()
+                        {
+                            JsonArray::from(old_column)
+                                .try_align(&json_type.as_arrow_type())
+                                .context(ConvertValueSnafu)?
                         } else {
                             datatypes::arrow::compute::cast(old_column, &ty.as_arrow_type())
                                 .context(ComputeArrowSnafu)?
@@ -366,7 +263,6 @@ impl FlatCompatBatch {
                     }
                 }
                 IndexOrDefault::DefaultValue {
-                    column_id: _,
                     default_vector,
                     semantic_type,
                 } => repeat_vector(default_vector, len, *semantic_type == SemanticType::Tag),
@@ -414,121 +310,6 @@ fn repeat_vector(vector: &VectorRef, to_len: usize, is_tag: bool) -> Result<Arra
     }
 }
 
-/// Helper to make primary key compatible.
-#[derive(Debug)]
-struct CompatPrimaryKey {
-    /// Row converter to append values to primary keys.
-    converter: Arc<dyn PrimaryKeyCodec>,
-    /// Default values to append.
-    values: Vec<(ColumnId, Value)>,
-}
-
-impl CompatPrimaryKey {
-    /// Make primary key of the `batch` compatible.
-    fn compat(&self, mut batch: Batch) -> Result<Batch> {
-        let mut buffer = Vec::with_capacity(
-            batch.primary_key().len() + self.converter.estimated_size().unwrap_or_default(),
-        );
-        buffer.extend_from_slice(batch.primary_key());
-        self.converter
-            .encode_values(&self.values, &mut buffer)
-            .context(EncodeSnafu)?;
-
-        batch.set_primary_key(buffer);
-
-        // update cache
-        if let Some(pk_values) = &mut batch.pk_values {
-            pk_values.extend(&self.values);
-        }
-
-        Ok(batch)
-    }
-}
-
-/// Helper to make fields compatible.
-#[derive(Debug)]
-struct CompatFields {
-    /// Column Ids and DataTypes the reader actually returns.
-    actual_fields: Vec<(ColumnId, ConcreteDataType)>,
-    /// Indices to convert actual fields to expect fields.
-    index_or_defaults: Vec<IndexOrDefault>,
-}
-
-impl CompatFields {
-    /// Make fields of the `batch` compatible.
-    fn compat(&self, batch: Batch) -> Result<Batch> {
-        debug_assert_eq!(self.actual_fields.len(), batch.fields().len());
-        debug_assert!(
-            self.actual_fields
-                .iter()
-                .zip(batch.fields())
-                .all(|((id, _), batch_column)| *id == batch_column.column_id)
-        );
-
-        let len = batch.num_rows();
-        self.index_or_defaults
-            .iter()
-            .map(|index_or_default| match index_or_default {
-                IndexOrDefault::Index { pos, cast_type } => {
-                    let old_column = &batch.fields()[*pos];
-
-                    let data = if let Some(ty) = cast_type {
-                        if let Some(json_type) = ty.as_json() {
-                            let json_array = old_column.data.to_arrow_array();
-                            let json_array =
-                                align_json_array(&json_array, &json_type.as_arrow_type())
-                                    .context(RecordBatchSnafu)?;
-                            Helper::try_into_vector(&json_array).context(ConvertVectorSnafu)?
-                        } else {
-                            old_column.data.cast(ty).with_context(|_| CastVectorSnafu {
-                                from: old_column.data.data_type(),
-                                to: ty.clone(),
-                            })?
-                        }
-                    } else {
-                        old_column.data.clone()
-                    };
-                    Ok(BatchColumn {
-                        column_id: old_column.column_id,
-                        data,
-                    })
-                }
-                IndexOrDefault::DefaultValue {
-                    column_id,
-                    default_vector,
-                    semantic_type: _,
-                } => {
-                    let data = default_vector.replicate(&[len]);
-                    Ok(BatchColumn {
-                        column_id: *column_id,
-                        data,
-                    })
-                }
-            })
-            .collect::<Result<Vec<_>>>()
-            .and_then(|fields| batch.with_fields(fields))
-    }
-}
-
-fn may_rewrite_primary_key(
-    expect: &RegionMetadata,
-    actual: &RegionMetadata,
-) -> Option<RewritePrimaryKey> {
-    if expect.primary_key_encoding == actual.primary_key_encoding {
-        return None;
-    }
-
-    let fields = expect.primary_key.clone();
-    let original = build_primary_key_codec(actual);
-    let new = build_primary_key_codec(expect);
-
-    Some(RewritePrimaryKey {
-        original,
-        new,
-        fields,
-    })
-}
-
 /// Returns true if the actual primary keys is the same as expected.
 fn is_primary_key_same(expect: &RegionMetadata, actual: &RegionMetadata) -> Result<bool> {
     ensure!(
@@ -556,113 +337,6 @@ fn is_primary_key_same(expect: &RegionMetadata, actual: &RegionMetadata) -> Resu
     Ok(actual.primary_key.len() == expect.primary_key.len())
 }
 
-/// Creates a [CompatPrimaryKey] if needed.
-fn may_compat_primary_key(
-    expect: &RegionMetadata,
-    actual: &RegionMetadata,
-) -> Result<Option<CompatPrimaryKey>> {
-    if is_primary_key_same(expect, actual)? {
-        return Ok(None);
-    }
-
-    // We need to append default values to the primary key.
-    let to_add = &expect.primary_key[actual.primary_key.len()..];
-    let mut fields = Vec::with_capacity(to_add.len());
-    let mut values = Vec::with_capacity(to_add.len());
-    for column_id in to_add {
-        // Safety: The id comes from expect region metadata.
-        let column = expect.column_by_id(*column_id).unwrap();
-        fields.push((
-            *column_id,
-            SortField::new(column.column_schema.data_type.clone()),
-        ));
-        let default_value = column
-            .column_schema
-            .create_default()
-            .context(CreateDefaultSnafu {
-                region_id: expect.region_id,
-                column: &column.column_schema.name,
-            })?
-            .with_context(|| CompatReaderSnafu {
-                region_id: expect.region_id,
-                reason: format!(
-                    "key column {} does not have a default value to read",
-                    column.column_schema.name
-                ),
-            })?;
-        values.push((*column_id, default_value));
-    }
-    // Using expect primary key encoding to build the converter
-    let converter =
-        build_primary_key_codec_with_fields(expect.primary_key_encoding, fields.into_iter());
-
-    Ok(Some(CompatPrimaryKey { converter, values }))
-}
-
-/// Creates a [CompatFields] if needed.
-fn may_compat_fields(
-    mapper: &PrimaryKeyProjectionMapper,
-    actual: &RegionMetadata,
-) -> Result<Option<CompatFields>> {
-    let expect_fields = mapper.batch_fields();
-    let actual_fields = Batch::projected_fields(actual, mapper.column_ids());
-    if expect_fields == actual_fields {
-        return Ok(None);
-    }
-
-    let source_field_index: HashMap<_, _> = actual_fields
-        .iter()
-        .enumerate()
-        .map(|(idx, (column_id, data_type))| (*column_id, (idx, data_type)))
-        .collect();
-
-    let index_or_defaults = expect_fields
-        .iter()
-        .map(|(column_id, expect_data_type)| {
-            if let Some((index, actual_data_type)) = source_field_index.get(column_id) {
-                let mut cast_type = None;
-
-                if expect_data_type != *actual_data_type {
-                    cast_type = Some(expect_data_type.clone())
-                }
-                // Source has this field.
-                Ok(IndexOrDefault::Index {
-                    pos: *index,
-                    cast_type,
-                })
-            } else {
-                // Safety: mapper must have this column.
-                let column = mapper.metadata().column_by_id(*column_id).unwrap();
-                // Create a default vector with 1 element for that column.
-                let default_vector = column
-                    .column_schema
-                    .create_default_vector(1)
-                    .context(CreateDefaultSnafu {
-                        region_id: mapper.metadata().region_id,
-                        column: &column.column_schema.name,
-                    })?
-                    .with_context(|| CompatReaderSnafu {
-                        region_id: mapper.metadata().region_id,
-                        reason: format!(
-                            "column {} does not have a default value to read",
-                            column.column_schema.name
-                        ),
-                    })?;
-                Ok(IndexOrDefault::DefaultValue {
-                    column_id: column.column_id,
-                    default_vector,
-                    semantic_type: SemanticType::Field,
-                })
-            }
-        })
-        .collect::<Result<Vec<_>>>()?;
-
-    Ok(Some(CompatFields {
-        actual_fields,
-        index_or_defaults,
-    }))
-}
-
 /// Index in source batch or a default value to fill a column.
 #[derive(Debug)]
 enum IndexOrDefault {
@@ -673,8 +347,6 @@ enum IndexOrDefault {
     },
     /// Default value for the column.
     DefaultValue {
-        /// Id of the column.
-        column_id: ColumnId,
         /// Default value. The vector has only 1 element.
         default_vector: VectorRef,
         /// Semantic type of the column.
@@ -682,58 +354,6 @@ enum IndexOrDefault {
     },
 }
 
-/// Adapter to rewrite primary key.
-struct RewritePrimaryKey {
-    /// Original primary key codec.
-    original: Arc<dyn PrimaryKeyCodec>,
-    /// New primary key codec.
-    new: Arc<dyn PrimaryKeyCodec>,
-    /// Order of the fields in the new primary key.
-    fields: Vec<ColumnId>,
-}
-
-impl RewritePrimaryKey {
-    /// Make primary key of the `batch` compatible.
-    fn compat(&self, mut batch: Batch) -> Result<Batch> {
-        if batch.pk_values().is_none() {
-            let new_pk_values = self
-                .original
-                .decode(batch.primary_key())
-                .context(DecodeSnafu)?;
-            batch.set_pk_values(new_pk_values);
-        }
-        // Safety: We ensure pk_values is not None.
-        let values = batch.pk_values().unwrap();
-
-        let mut buffer = Vec::with_capacity(
-            batch.primary_key().len() + self.new.estimated_size().unwrap_or_default(),
-        );
-        match values {
-            CompositeValues::Dense(values) => {
-                self.new
-                    .encode_values(values.as_slice(), &mut buffer)
-                    .context(EncodeSnafu)?;
-            }
-            CompositeValues::Sparse(values) => {
-                let values = self
-                    .fields
-                    .iter()
-                    .map(|id| {
-                        let value = values.get_or_null(*id);
-                        (*id, value.as_value_ref())
-                    })
-                    .collect::<Vec<_>>();
-                self.new
-                    .encode_value_refs(&values, &mut buffer)
-                    .context(EncodeSnafu)?;
-            }
-        }
-        batch.set_primary_key(buffer);
-
-        Ok(batch)
-    }
-}
-
 /// Helper to rewrite primary key to another encoding for flat format.
 struct FlatRewritePrimaryKey {
     /// New primary key encoder.
@@ -980,7 +600,6 @@ mod tests {
     use datatypes::prelude::ConcreteDataType;
     use datatypes::schema::ColumnSchema;
     use datatypes::value::ValueRef;
-    use datatypes::vectors::{Int64Vector, TimestampMillisecondVector, UInt8Vector, UInt64Vector};
     use mito_codec::row_converter::{
         DensePrimaryKeyCodec, PrimaryKeyCodecExt, SparsePrimaryKeyCodec,
     };
@@ -990,9 +609,9 @@ mod tests {
 
     use super::*;
     use crate::read::flat_projection::FlatProjectionMapper;
+    use crate::read::read_columns::ReadColumns;
     use crate::sst::parquet::flat_format::FlatReadFormat;
     use crate::sst::{FlatSchemaOptions, to_flat_sst_arrow_schema};
-    use crate::test_util::{VecBatchReader, check_reader_result};
 
     /// Creates a new [RegionMetadata].
     fn new_metadata(
@@ -1053,471 +672,6 @@ mod tests {
         buffer
     }
 
-    /// Creates a batch for specific primary `key`.
-    ///
-    /// `fields`: [(column_id of the field, is null)]
-    fn new_batch(
-        primary_key: &[u8],
-        fields: &[(ColumnId, bool)],
-        start_ts: i64,
-        num_rows: usize,
-    ) -> Batch {
-        let timestamps = Arc::new(TimestampMillisecondVector::from_values(
-            start_ts..start_ts + num_rows as i64,
-        ));
-        let sequences = Arc::new(UInt64Vector::from_values(0..num_rows as u64));
-        let op_types = Arc::new(UInt8Vector::from_vec(vec![OpType::Put as u8; num_rows]));
-        let field_columns = fields
-            .iter()
-            .map(|(id, is_null)| {
-                let data = if *is_null {
-                    Arc::new(Int64Vector::from(vec![None; num_rows]))
-                } else {
-                    Arc::new(Int64Vector::from_vec(vec![*id as i64; num_rows]))
-                };
-                BatchColumn {
-                    column_id: *id,
-                    data,
-                }
-            })
-            .collect();
-        Batch::new(
-            primary_key.to_vec(),
-            timestamps,
-            sequences,
-            op_types,
-            field_columns,
-        )
-        .unwrap()
-    }
-
-    #[test]
-    fn test_invalid_pk_len() {
-        let reader_meta = new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (3, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1, 2],
-        );
-        let expect_meta = new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1],
-        );
-        may_compat_primary_key(&expect_meta, &reader_meta).unwrap_err();
-    }
-
-    #[test]
-    fn test_different_pk() {
-        let reader_meta = new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (3, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[2, 1],
-        );
-        let expect_meta = new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (3, SemanticType::Field, ConcreteDataType::int64_datatype()),
-                (4, SemanticType::Tag, ConcreteDataType::string_datatype()),
-            ],
-            &[1, 2, 4],
-        );
-        may_compat_primary_key(&expect_meta, &reader_meta).unwrap_err();
-    }
-
-    #[test]
-    fn test_same_pk() {
-        let reader_meta = new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1],
-        );
-        assert!(
-            may_compat_primary_key(&reader_meta, &reader_meta)
-                .unwrap()
-                .is_none()
-        );
-    }
-
-    #[test]
-    fn test_same_pk_encoding() {
-        let reader_meta = Arc::new(new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-            ],
-            &[1],
-        ));
-
-        assert!(
-            may_compat_primary_key(&reader_meta, &reader_meta)
-                .unwrap()
-                .is_none()
-        );
-    }
-
-    #[test]
-    fn test_same_fields() {
-        let reader_meta = Arc::new(new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1],
-        ));
-        let mapper = PrimaryKeyProjectionMapper::all(&reader_meta).unwrap();
-        assert!(may_compat_fields(&mapper, &reader_meta).unwrap().is_none())
-    }
-
-    #[tokio::test]
-    async fn test_compat_reader() {
-        let reader_meta = Arc::new(new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1],
-        ));
-        let expect_meta = Arc::new(new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::int64_datatype()),
-                (3, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (4, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1, 3],
-        ));
-        let mapper = ProjectionMapper::all(&expect_meta, false).unwrap();
-        let k1 = encode_key(&[Some("a")]);
-        let k2 = encode_key(&[Some("b")]);
-        let source_reader = VecBatchReader::new(&[
-            new_batch(&k1, &[(2, false)], 1000, 3),
-            new_batch(&k2, &[(2, false)], 1000, 3),
-        ]);
-
-        let mut compat_reader = CompatReader::new(&mapper, reader_meta, source_reader).unwrap();
-        let k1 = encode_key(&[Some("a"), None]);
-        let k2 = encode_key(&[Some("b"), None]);
-        check_reader_result(
-            &mut compat_reader,
-            &[
-                new_batch(&k1, &[(2, false), (4, true)], 1000, 3),
-                new_batch(&k2, &[(2, false), (4, true)], 1000, 3),
-            ],
-        )
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_compat_reader_different_order() {
-        let reader_meta = Arc::new(new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1],
-        ));
-        let expect_meta = Arc::new(new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (3, SemanticType::Field, ConcreteDataType::int64_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::int64_datatype()),
-                (4, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1],
-        ));
-        let mapper = ProjectionMapper::all(&expect_meta, false).unwrap();
-        let k1 = encode_key(&[Some("a")]);
-        let k2 = encode_key(&[Some("b")]);
-        let source_reader = VecBatchReader::new(&[
-            new_batch(&k1, &[(2, false)], 1000, 3),
-            new_batch(&k2, &[(2, false)], 1000, 3),
-        ]);
-
-        let mut compat_reader = CompatReader::new(&mapper, reader_meta, source_reader).unwrap();
-        check_reader_result(
-            &mut compat_reader,
-            &[
-                new_batch(&k1, &[(3, true), (2, false), (4, true)], 1000, 3),
-                new_batch(&k2, &[(3, true), (2, false), (4, true)], 1000, 3),
-            ],
-        )
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_compat_reader_different_types() {
-        let actual_meta = Arc::new(new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1],
-        ));
-        let expect_meta = Arc::new(new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::string_datatype()),
-            ],
-            &[1],
-        ));
-        let mapper = ProjectionMapper::all(&expect_meta, false).unwrap();
-        let k1 = encode_key(&[Some("a")]);
-        let k2 = encode_key(&[Some("b")]);
-        let source_reader = VecBatchReader::new(&[
-            new_batch(&k1, &[(2, false)], 1000, 3),
-            new_batch(&k2, &[(2, false)], 1000, 3),
-        ]);
-
-        let fn_batch_cast = |batch: Batch| {
-            let mut new_fields = batch.fields.clone();
-            new_fields[0].data = new_fields[0]
-                .data
-                .cast(&ConcreteDataType::string_datatype())
-                .unwrap();
-
-            batch.with_fields(new_fields).unwrap()
-        };
-        let mut compat_reader = CompatReader::new(&mapper, actual_meta, source_reader).unwrap();
-        check_reader_result(
-            &mut compat_reader,
-            &[
-                fn_batch_cast(new_batch(&k1, &[(2, false)], 1000, 3)),
-                fn_batch_cast(new_batch(&k2, &[(2, false)], 1000, 3)),
-            ],
-        )
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_compat_reader_projection() {
-        let reader_meta = Arc::new(new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1],
-        ));
-        let expect_meta = Arc::new(new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (3, SemanticType::Field, ConcreteDataType::int64_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::int64_datatype()),
-                (4, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1],
-        ));
-        // tag_1, field_2, field_3
-        let mapper = ProjectionMapper::new(&expect_meta, [1, 3, 2].into_iter(), false).unwrap();
-        let k1 = encode_key(&[Some("a")]);
-        let source_reader = VecBatchReader::new(&[new_batch(&k1, &[(2, false)], 1000, 3)]);
-
-        let mut compat_reader =
-            CompatReader::new(&mapper, reader_meta.clone(), source_reader).unwrap();
-        check_reader_result(
-            &mut compat_reader,
-            &[new_batch(&k1, &[(3, true), (2, false)], 1000, 3)],
-        )
-        .await;
-
-        // tag_1, field_4, field_3
-        let mapper = ProjectionMapper::new(&expect_meta, [1, 4, 2].into_iter(), false).unwrap();
-        let k1 = encode_key(&[Some("a")]);
-        let source_reader = VecBatchReader::new(&[new_batch(&k1, &[], 1000, 3)]);
-
-        let mut compat_reader = CompatReader::new(&mapper, reader_meta, source_reader).unwrap();
-        check_reader_result(
-            &mut compat_reader,
-            &[new_batch(&k1, &[(3, true), (4, true)], 1000, 3)],
-        )
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_compat_reader_projection_read_superset() {
-        let reader_meta = Arc::new(new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1],
-        ));
-        let expect_meta = Arc::new(new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (3, SemanticType::Field, ConcreteDataType::int64_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::int64_datatype()),
-                (4, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1],
-        ));
-        // Output: tag_1, field_3, field_2. Read also includes field_4.
-        let mapper = ProjectionMapper::new_with_read_columns(
-            &expect_meta,
-            [1, 3, 2].into_iter(),
-            false,
-            vec![1, 3, 2, 4],
-        )
-        .unwrap();
-        let k1 = encode_key(&[Some("a")]);
-        let source_reader = VecBatchReader::new(&[new_batch(&k1, &[(2, false)], 1000, 3)]);
-
-        let mut compat_reader = CompatReader::new(&mapper, reader_meta, source_reader).unwrap();
-        check_reader_result(
-            &mut compat_reader,
-            &[new_batch(&k1, &[(3, true), (2, false), (4, true)], 1000, 3)],
-        )
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_compat_reader_different_pk_encoding() {
-        let mut reader_meta = new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1],
-        );
-        reader_meta.primary_key_encoding = PrimaryKeyEncoding::Dense;
-        let reader_meta = Arc::new(reader_meta);
-        let mut expect_meta = new_metadata(
-            &[
-                (
-                    0,
-                    SemanticType::Timestamp,
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                ),
-                (1, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (2, SemanticType::Field, ConcreteDataType::int64_datatype()),
-                (3, SemanticType::Tag, ConcreteDataType::string_datatype()),
-                (4, SemanticType::Field, ConcreteDataType::int64_datatype()),
-            ],
-            &[1, 3],
-        );
-        expect_meta.primary_key_encoding = PrimaryKeyEncoding::Sparse;
-        let expect_meta = Arc::new(expect_meta);
-
-        let mapper = ProjectionMapper::all(&expect_meta, false).unwrap();
-        let k1 = encode_key(&[Some("a")]);
-        let k2 = encode_key(&[Some("b")]);
-        let source_reader = VecBatchReader::new(&[
-            new_batch(&k1, &[(2, false)], 1000, 3),
-            new_batch(&k2, &[(2, false)], 1000, 3),
-        ]);
-
-        let mut compat_reader = CompatReader::new(&mapper, reader_meta, source_reader).unwrap();
-        let k1 = encode_sparse_key(&[(1, Some("a")), (3, None)]);
-        let k2 = encode_sparse_key(&[(1, Some("b")), (3, None)]);
-        check_reader_result(
-            &mut compat_reader,
-            &[
-                new_batch(&k1, &[(2, false), (4, true)], 1000, 3),
-                new_batch(&k2, &[(2, false), (4, true)], 1000, 3),
-            ],
-        )
-        .await;
-    }
-
     /// Creates a primary key array for flat format testing.
     fn build_flat_test_pk_array(primary_keys: &[&[u8]]) -> ArrayRef {
         let mut builder = BinaryDictionaryBuilder::<UInt32Type>::new();
@@ -1560,7 +714,7 @@ mod tests {
         let mapper = FlatProjectionMapper::all(&expected_metadata).unwrap();
         let read_format = FlatReadFormat::new(
             actual_metadata.clone(),
-            [0, 1, 2, 3].into_iter(),
+            ReadColumns::from_deduped_column_ids([0, 1, 2, 3]),
             None,
             "test",
             false,
@@ -1646,16 +800,15 @@ mod tests {
             &[1],
         ));
 
-        // Output projection: tag_1, field_2. Read also includes field_3.
         let mapper = FlatProjectionMapper::new_with_read_columns(
             &expected_metadata,
             vec![1, 2],
-            vec![1, 2, 3],
+            ReadColumns::from_deduped_column_ids([1, 2, 3]),
         )
         .unwrap();
         let read_format = FlatReadFormat::new(
             actual_metadata.clone(),
-            [1, 2, 3].into_iter(),
+            ReadColumns::from_deduped_column_ids([1, 2, 3]),
             None,
             "test",
             false,
@@ -1746,7 +899,7 @@ mod tests {
         let mapper = FlatProjectionMapper::all(&expected_metadata).unwrap();
         let read_format = FlatReadFormat::new(
             actual_metadata.clone(),
-            [0, 1, 2, 3].into_iter(),
+            ReadColumns::from_deduped_column_ids([0, 1, 2, 3]),
             None,
             "test",
             false,
@@ -1840,7 +993,7 @@ mod tests {
         let mapper = FlatProjectionMapper::all(&expected_metadata).unwrap();
         let read_format = FlatReadFormat::new(
             actual_metadata.clone(),
-            [0, 2, 3].into_iter(),
+            ReadColumns::from_deduped_column_ids([0, 2, 3]),
             None,
             "test",
             true,
diff --git a/src/mito2/src/read/dedup.rs b/src/mito2/src/read/dedup.rs
index 5c881459b2..86f6b07ffc 100644
--- a/src/mito2/src/read/dedup.rs
+++ b/src/mito2/src/read/dedup.rs
@@ -19,17 +19,13 @@ use std::sync::Arc;
 use std::time::{Duration, Instant};
 
 use api::v1::OpType;
-use async_trait::async_trait;
-use common_telemetry::debug;
-use common_time::Timestamp;
 use datatypes::data_type::DataType;
 use datatypes::prelude::ScalarVector;
 use datatypes::value::Value;
 use datatypes::vectors::MutableVector;
 
 use crate::error::Result;
-use crate::metrics::MERGE_FILTER_ROWS_TOTAL;
-use crate::read::{Batch, BatchColumn, BatchReader};
+use crate::read::{Batch, BatchColumn};
 
 /// Trait for reporting dedup metrics.
 pub trait DedupMetricsReport: Send + Sync {
@@ -37,80 +33,6 @@ pub trait DedupMetricsReport: Send + Sync {
     fn report(&self, metrics: &mut DedupMetrics);
 }
 
-/// A reader that dedup sorted batches from a source based on the
-/// dedup strategy.
-pub struct DedupReader<R, S> {
-    source: R,
-    strategy: S,
-    metrics: DedupMetrics,
-    /// Optional metrics reporter.
-    metrics_reporter: Option<Arc<dyn DedupMetricsReport>>,
-}
-
-impl<R, S> DedupReader<R, S> {
-    /// Creates a new dedup reader.
-    pub fn new(
-        source: R,
-        strategy: S,
-        metrics_reporter: Option<Arc<dyn DedupMetricsReport>>,
-    ) -> Self {
-        Self {
-            source,
-            strategy,
-            metrics: DedupMetrics::default(),
-            metrics_reporter,
-        }
-    }
-}
-
-impl<R: BatchReader, S: DedupStrategy> DedupReader<R, S> {
-    /// Returns the next deduplicated batch.
-    async fn fetch_next_batch(&mut self) -> Result<Option<Batch>> {
-        while let Some(batch) = self.source.next_batch().await? {
-            if let Some(batch) = self.strategy.push_batch(batch, &mut self.metrics)? {
-                self.metrics.maybe_report(&self.metrics_reporter);
-                return Ok(Some(batch));
-            }
-        }
-
-        let result = self.strategy.finish(&mut self.metrics)?;
-        self.metrics.maybe_report(&self.metrics_reporter);
-        Ok(result)
-    }
-}
-
-#[async_trait]
-impl<R: BatchReader, S: DedupStrategy> BatchReader for DedupReader<R, S> {
-    async fn next_batch(&mut self) -> Result<Option<Batch>> {
-        self.fetch_next_batch().await
-    }
-}
-
-impl<R, S> Drop for DedupReader<R, S> {
-    fn drop(&mut self) {
-        debug!("Dedup reader finished, metrics: {:?}", self.metrics);
-
-        MERGE_FILTER_ROWS_TOTAL
-            .with_label_values(&["dedup"])
-            .inc_by(self.metrics.num_unselected_rows as u64);
-        MERGE_FILTER_ROWS_TOTAL
-            .with_label_values(&["delete"])
-            .inc_by(self.metrics.num_unselected_rows as u64);
-
-        // Report any remaining metrics.
-        if let Some(reporter) = &self.metrics_reporter {
-            reporter.report(&mut self.metrics);
-        }
-    }
-}
-
-#[cfg(test)]
-impl<R, S> DedupReader<R, S> {
-    fn metrics(&self) -> &DedupMetrics {
-        &self.metrics
-    }
-}
-
 /// Strategy to remove duplicate rows from sorted batches.
 pub trait DedupStrategy: Send {
     /// Pushes a batch to the dedup strategy.
@@ -124,114 +46,6 @@ pub trait DedupStrategy: Send {
     fn finish(&mut self, metrics: &mut DedupMetrics) -> Result<Option<Batch>>;
 }
 
-/// State of the last row in a batch for dedup.
-struct BatchLastRow {
-    primary_key: Vec<u8>,
-    /// The last timestamp of the batch.
-    timestamp: Timestamp,
-}
-
-/// Dedup strategy that keeps the row with latest sequence of each key.
-///
-/// This strategy is optimized specially based on the properties of the SST files,
-/// memtables and the merge reader. It assumes that batches from files and memtables
-/// don't contain duplicate rows and the merge reader never concatenates batches from
-/// different source.
-///
-/// We might implement a new strategy if we need to process files with duplicate rows.
-pub struct LastRow {
-    /// Meta of the last row in the previous batch that has the same key
-    /// as the batch to push.
-    prev_batch: Option<BatchLastRow>,
-    /// Filter deleted rows.
-    filter_deleted: bool,
-}
-
-impl LastRow {
-    /// Creates a new strategy with the given `filter_deleted` flag.
-    pub fn new(filter_deleted: bool) -> Self {
-        Self {
-            prev_batch: None,
-            filter_deleted,
-        }
-    }
-}
-
-impl DedupStrategy for LastRow {
-    fn push_batch(
-        &mut self,
-        mut batch: Batch,
-        metrics: &mut DedupMetrics,
-    ) -> Result<Option<Batch>> {
-        let start = Instant::now();
-
-        if batch.is_empty() {
-            return Ok(None);
-        }
-        debug_assert!(batch.first_timestamp().is_some());
-        let prev_timestamp = match &self.prev_batch {
-            Some(prev_batch) => {
-                if prev_batch.primary_key != batch.primary_key() {
-                    // The key has changed. This is the first batch of the
-                    // new key.
-                    None
-                } else {
-                    Some(prev_batch.timestamp)
-                }
-            }
-            None => None,
-        };
-        if batch.first_timestamp() == prev_timestamp {
-            metrics.num_unselected_rows += 1;
-            // This batch contains a duplicate row, skip it.
-            if batch.num_rows() == 1 {
-                // We don't need to update `prev_batch` because they have the same
-                // key and timestamp.
-                metrics.dedup_cost += start.elapsed();
-                return Ok(None);
-            }
-            // Skips the first row.
-            batch = batch.slice(1, batch.num_rows() - 1);
-        }
-
-        // Store current batch to `prev_batch` so we could compare the next batch
-        // with this batch. We store batch before filtering it as rows with `OpType::Delete`
-        // would be removed from the batch after filter, then we may store an incorrect `last row`
-        // of previous batch.
-        match &mut self.prev_batch {
-            Some(prev) => {
-                // Reuse the primary key buffer.
-                prev.primary_key.clone_from(&batch.primary_key);
-                prev.timestamp = batch.last_timestamp().unwrap();
-            }
-            None => {
-                self.prev_batch = Some(BatchLastRow {
-                    primary_key: batch.primary_key().to_vec(),
-                    timestamp: batch.last_timestamp().unwrap(),
-                })
-            }
-        }
-
-        // Filters deleted rows.
-        if self.filter_deleted {
-            filter_deleted_from_batch(&mut batch, metrics)?;
-        }
-
-        metrics.dedup_cost += start.elapsed();
-
-        // The batch can become empty if all rows are deleted.
-        if batch.is_empty() {
-            Ok(None)
-        } else {
-            Ok(Some(batch))
-        }
-    }
-
-    fn finish(&mut self, _metrics: &mut DedupMetrics) -> Result<Option<Batch>> {
-        Ok(None)
-    }
-}
-
 /// Removes deleted rows from the batch and updates metrics.
 fn filter_deleted_from_batch(batch: &mut Batch, metrics: &mut DedupMetrics) -> Result<()> {
     let num_rows = batch.num_rows();
@@ -672,137 +486,10 @@ impl<I: Iterator<Item = Result<Batch>>> Iterator for LastNonNullIter<I> {
 mod tests {
     use std::sync::Arc;
 
-    use api::v1::OpType;
     use datatypes::arrow::array::{TimestampMillisecondArray, UInt8Array, UInt64Array};
 
     use super::*;
     use crate::read::BatchBuilder;
-    use crate::test_util::{VecBatchReader, check_reader_result, new_batch};
-
-    #[tokio::test]
-    async fn test_dedup_reader_no_duplications() {
-        let input = [
-            new_batch(
-                b"k1",
-                &[1, 2],
-                &[11, 12],
-                &[OpType::Put, OpType::Put],
-                &[21, 22],
-            ),
-            new_batch(b"k1", &[3], &[13], &[OpType::Put], &[23]),
-            new_batch(
-                b"k2",
-                &[1, 2],
-                &[111, 112],
-                &[OpType::Put, OpType::Put],
-                &[31, 32],
-            ),
-        ];
-
-        // Test last row.
-        let reader = VecBatchReader::new(&input);
-        let mut reader = DedupReader::new(reader, LastRow::new(true), None);
-        check_reader_result(&mut reader, &input).await;
-        assert_eq!(0, reader.metrics().num_unselected_rows);
-        assert_eq!(0, reader.metrics().num_deleted_rows);
-
-        // Test last non-null.
-        let reader = VecBatchReader::new(&input);
-        let mut reader = DedupReader::new(reader, LastNonNull::new(true), None);
-        check_reader_result(&mut reader, &input).await;
-        assert_eq!(0, reader.metrics().num_unselected_rows);
-        assert_eq!(0, reader.metrics().num_deleted_rows);
-    }
-
-    #[tokio::test]
-    async fn test_dedup_reader_duplications() {
-        let input = [
-            new_batch(
-                b"k1",
-                &[1, 2],
-                &[13, 11],
-                &[OpType::Put, OpType::Put],
-                &[11, 12],
-            ),
-            // empty batch.
-            new_batch(b"k1", &[], &[], &[], &[]),
-            // Duplicate with the previous batch.
-            new_batch(
-                b"k1",
-                &[2, 3, 4],
-                &[10, 13, 13],
-                &[OpType::Put, OpType::Put, OpType::Delete],
-                &[2, 13, 14],
-            ),
-            new_batch(
-                b"k2",
-                &[1, 2],
-                &[20, 20],
-                &[OpType::Put, OpType::Delete],
-                &[101, 0],
-            ),
-            new_batch(b"k2", &[2], &[19], &[OpType::Put], &[102]),
-            new_batch(b"k3", &[2], &[20], &[OpType::Put], &[202]),
-            // This batch won't increase the deleted rows count as it
-            // is filtered out by the previous batch.
-            new_batch(b"k3", &[2], &[19], &[OpType::Delete], &[0]),
-        ];
-        // Filter deleted.
-        let reader = VecBatchReader::new(&input);
-        let mut reader = DedupReader::new(reader, LastRow::new(true), None);
-        check_reader_result(
-            &mut reader,
-            &[
-                new_batch(
-                    b"k1",
-                    &[1, 2],
-                    &[13, 11],
-                    &[OpType::Put, OpType::Put],
-                    &[11, 12],
-                ),
-                new_batch(b"k1", &[3], &[13], &[OpType::Put], &[13]),
-                new_batch(b"k2", &[1], &[20], &[OpType::Put], &[101]),
-                new_batch(b"k3", &[2], &[20], &[OpType::Put], &[202]),
-            ],
-        )
-        .await;
-        assert_eq!(5, reader.metrics().num_unselected_rows);
-        assert_eq!(2, reader.metrics().num_deleted_rows);
-
-        // Does not filter deleted.
-        let reader = VecBatchReader::new(&input);
-        let mut reader = DedupReader::new(reader, LastRow::new(false), None);
-        check_reader_result(
-            &mut reader,
-            &[
-                new_batch(
-                    b"k1",
-                    &[1, 2],
-                    &[13, 11],
-                    &[OpType::Put, OpType::Put],
-                    &[11, 12],
-                ),
-                new_batch(
-                    b"k1",
-                    &[3, 4],
-                    &[13, 13],
-                    &[OpType::Put, OpType::Delete],
-                    &[13, 14],
-                ),
-                new_batch(
-                    b"k2",
-                    &[1, 2],
-                    &[20, 20],
-                    &[OpType::Put, OpType::Delete],
-                    &[101, 0],
-                ),
-                new_batch(b"k3", &[2], &[20], &[OpType::Put], &[202]),
-            ],
-        )
-        .await;
-        assert_eq!(3, reader.metrics().num_unselected_rows);
-        assert_eq!(0, reader.metrics().num_deleted_rows);
-    }
 
     /// Returns a new [Batch] whose field has column id 1, 2.
     fn new_batch_multi_fields(
@@ -839,235 +526,6 @@ mod tests {
         builder.build().unwrap()
     }
 
-    #[tokio::test]
-    async fn test_last_non_null_merge() {
-        let input = [
-            new_batch_multi_fields(
-                b"k1",
-                &[1, 2],
-                &[13, 11],
-                &[OpType::Put, OpType::Put],
-                &[(Some(11), Some(11)), (None, None)],
-            ),
-            // empty batch.
-            new_batch_multi_fields(b"k1", &[], &[], &[], &[]),
-            // Duplicate with the previous batch.
-            new_batch_multi_fields(b"k1", &[2], &[10], &[OpType::Put], &[(Some(12), None)]),
-            new_batch_multi_fields(
-                b"k1",
-                &[2, 3, 4],
-                &[10, 13, 13],
-                &[OpType::Put, OpType::Put, OpType::Delete],
-                &[(Some(2), Some(22)), (Some(13), None), (None, Some(14))],
-            ),
-            new_batch_multi_fields(
-                b"k2",
-                &[1, 2],
-                &[20, 20],
-                &[OpType::Put, OpType::Delete],
-                &[(Some(101), Some(101)), (None, None)],
-            ),
-            new_batch_multi_fields(
-                b"k2",
-                &[2],
-                &[19],
-                &[OpType::Put],
-                &[(Some(102), Some(102))],
-            ),
-            new_batch_multi_fields(
-                b"k3",
-                &[2],
-                &[20],
-                &[OpType::Put],
-                &[(Some(202), Some(202))],
-            ),
-            // This batch won't increase the deleted rows count as it
-            // is filtered out by the previous batch. (All fields are null).
-            new_batch_multi_fields(b"k3", &[2], &[19], &[OpType::Delete], &[(None, None)]),
-        ];
-
-        // Filter deleted.
-        let reader = VecBatchReader::new(&input);
-        let mut reader = DedupReader::new(reader, LastNonNull::new(true), None);
-        check_reader_result(
-            &mut reader,
-            &[
-                new_batch_multi_fields(
-                    b"k1",
-                    &[1, 2],
-                    &[13, 11],
-                    &[OpType::Put, OpType::Put],
-                    &[(Some(11), Some(11)), (Some(12), Some(22))],
-                ),
-                new_batch_multi_fields(b"k1", &[3], &[13], &[OpType::Put], &[(Some(13), None)]),
-                new_batch_multi_fields(
-                    b"k2",
-                    &[1],
-                    &[20],
-                    &[OpType::Put],
-                    &[(Some(101), Some(101))],
-                ),
-                new_batch_multi_fields(
-                    b"k3",
-                    &[2],
-                    &[20],
-                    &[OpType::Put],
-                    &[(Some(202), Some(202))],
-                ),
-            ],
-        )
-        .await;
-        assert_eq!(6, reader.metrics().num_unselected_rows);
-        assert_eq!(2, reader.metrics().num_deleted_rows);
-
-        // Does not filter deleted.
-        let reader = VecBatchReader::new(&input);
-        let mut reader = DedupReader::new(reader, LastNonNull::new(false), None);
-        check_reader_result(
-            &mut reader,
-            &[
-                new_batch_multi_fields(
-                    b"k1",
-                    &[1, 2],
-                    &[13, 11],
-                    &[OpType::Put, OpType::Put],
-                    &[(Some(11), Some(11)), (Some(12), Some(22))],
-                ),
-                new_batch_multi_fields(
-                    b"k1",
-                    &[3, 4],
-                    &[13, 13],
-                    &[OpType::Put, OpType::Delete],
-                    &[(Some(13), None), (None, Some(14))],
-                ),
-                new_batch_multi_fields(
-                    b"k2",
-                    &[1, 2],
-                    &[20, 20],
-                    &[OpType::Put, OpType::Delete],
-                    &[(Some(101), Some(101)), (None, None)],
-                ),
-                new_batch_multi_fields(
-                    b"k3",
-                    &[2],
-                    &[20],
-                    &[OpType::Put],
-                    &[(Some(202), Some(202))],
-                ),
-            ],
-        )
-        .await;
-        assert_eq!(4, reader.metrics().num_unselected_rows);
-        assert_eq!(0, reader.metrics().num_deleted_rows);
-    }
-
-    #[tokio::test]
-    async fn test_last_non_null_skip_merge_single() {
-        let input = [new_batch_multi_fields(
-            b"k1",
-            &[1, 2, 3],
-            &[13, 11, 13],
-            &[OpType::Put, OpType::Delete, OpType::Put],
-            &[(Some(11), Some(11)), (None, None), (Some(13), Some(13))],
-        )];
-
-        let reader = VecBatchReader::new(&input);
-        let mut reader = DedupReader::new(reader, LastNonNull::new(true), None);
-        check_reader_result(
-            &mut reader,
-            &[new_batch_multi_fields(
-                b"k1",
-                &[1, 3],
-                &[13, 13],
-                &[OpType::Put, OpType::Put],
-                &[(Some(11), Some(11)), (Some(13), Some(13))],
-            )],
-        )
-        .await;
-        assert_eq!(1, reader.metrics().num_unselected_rows);
-        assert_eq!(1, reader.metrics().num_deleted_rows);
-
-        let reader = VecBatchReader::new(&input);
-        let mut reader = DedupReader::new(reader, LastNonNull::new(false), None);
-        check_reader_result(&mut reader, &input).await;
-        assert_eq!(0, reader.metrics().num_unselected_rows);
-        assert_eq!(0, reader.metrics().num_deleted_rows);
-    }
-
-    #[tokio::test]
-    async fn test_last_non_null_skip_merge_no_null() {
-        let input = [
-            new_batch_multi_fields(
-                b"k1",
-                &[1, 2],
-                &[13, 11],
-                &[OpType::Put, OpType::Put],
-                &[(Some(11), Some(11)), (Some(12), Some(12))],
-            ),
-            new_batch_multi_fields(b"k1", &[2], &[10], &[OpType::Put], &[(None, Some(22))]),
-            new_batch_multi_fields(
-                b"k1",
-                &[2, 3],
-                &[9, 13],
-                &[OpType::Put, OpType::Put],
-                &[(Some(32), None), (Some(13), Some(13))],
-            ),
-        ];
-
-        let reader = VecBatchReader::new(&input);
-        let mut reader = DedupReader::new(reader, LastNonNull::new(true), None);
-        check_reader_result(
-            &mut reader,
-            &[
-                new_batch_multi_fields(
-                    b"k1",
-                    &[1, 2],
-                    &[13, 11],
-                    &[OpType::Put, OpType::Put],
-                    &[(Some(11), Some(11)), (Some(12), Some(12))],
-                ),
-                new_batch_multi_fields(b"k1", &[3], &[13], &[OpType::Put], &[(Some(13), Some(13))]),
-            ],
-        )
-        .await;
-        assert_eq!(2, reader.metrics().num_unselected_rows);
-        assert_eq!(0, reader.metrics().num_deleted_rows);
-    }
-
-    #[tokio::test]
-    async fn test_last_non_null_merge_null() {
-        let input = [
-            new_batch_multi_fields(
-                b"k1",
-                &[1, 2],
-                &[13, 11],
-                &[OpType::Put, OpType::Put],
-                &[(Some(11), Some(11)), (None, None)],
-            ),
-            new_batch_multi_fields(b"k1", &[2], &[10], &[OpType::Put], &[(None, Some(22))]),
-            new_batch_multi_fields(b"k1", &[3], &[13], &[OpType::Put], &[(Some(33), None)]),
-        ];
-
-        let reader = VecBatchReader::new(&input);
-        let mut reader = DedupReader::new(reader, LastNonNull::new(true), None);
-        check_reader_result(
-            &mut reader,
-            &[
-                new_batch_multi_fields(
-                    b"k1",
-                    &[1, 2],
-                    &[13, 11],
-                    &[OpType::Put, OpType::Put],
-                    &[(Some(11), Some(11)), (None, Some(22))],
-                ),
-                new_batch_multi_fields(b"k1", &[3], &[13], &[OpType::Put], &[(Some(33), None)]),
-            ],
-        )
-        .await;
-        assert_eq!(1, reader.metrics().num_unselected_rows);
-        assert_eq!(0, reader.metrics().num_deleted_rows);
-    }
-
     fn check_dedup_strategy(input: &[Batch], strategy: &mut dyn DedupStrategy, expect: &[Batch]) {
         let mut actual = Vec::new();
         let mut metrics = DedupMetrics::default();
diff --git a/src/mito2/src/read/flat_merge.rs b/src/mito2/src/read/flat_merge.rs
index 90df227ae9..b1c304f244 100644
--- a/src/mito2/src/read/flat_merge.rs
+++ b/src/mito2/src/read/flat_merge.rs
@@ -14,14 +14,16 @@
 
 use std::cmp::Ordering;
 use std::collections::BinaryHeap;
+use std::fmt;
 use std::sync::Arc;
-use std::time::Instant;
+use std::time::{Duration, Instant};
 
 use async_stream::try_stream;
 use common_telemetry::debug;
-use datatypes::arrow::array::{Int64Array, UInt64Array};
+use datatypes::arrow::array::{Array, AsArray, Int64Array, UInt64Array};
 use datatypes::arrow::compute::interleave;
-use datatypes::arrow::datatypes::SchemaRef;
+use datatypes::arrow::datatypes::{ArrowNativeType, BinaryType, DataType, SchemaRef, Utf8Type};
+use datatypes::arrow::error::ArrowError;
 use datatypes::arrow::record_batch::RecordBatch;
 use datatypes::arrow_array::BinaryArray;
 use datatypes::timestamp::timestamp_array_to_primitive;
@@ -33,12 +35,67 @@ use crate::error::{ComputeArrowSnafu, Result};
 use crate::memtable::BoxedRecordBatchIterator;
 use crate::metrics::READ_STAGE_ELAPSED;
 use crate::read::BoxedRecordBatchStream;
-use crate::read::merge::{MergeMetrics, MergeMetricsReport};
 use crate::sst::parquet::flat_format::{
     primary_key_column_index, sequence_column_index, time_index_column_index,
 };
 use crate::sst::parquet::format::PrimaryKeyArray;
 
+/// Checks whether interleaving the selected rows from byte columns would overflow
+/// i32 offsets. Similar to arrow-rs `interleave_bytes()`, accumulates offsets and
+/// returns an error if the capacity exceeds `i32::MAX`.
+///
+/// TODO(yingwen): Remove this after upgrading to arrow >= 58.1.0, which handles
+/// offset overflow in `interleave_bytes()` natively.
+///
+/// See: <https://github.com/apache/arrow-rs/blob/65ad652f2410fc51ad77da1805e85c0a76d9a7ea/arrow-select/src/interleave.rs#L208-L225>
+fn check_interleave_bytes_overflow<T: datatypes::arrow::datatypes::ByteArrayType>(
+    batches: &[(usize, RecordBatch)],
+    col_idx: usize,
+    indices: &[(usize, usize)],
+) -> std::result::Result<(), ArrowError> {
+    // Quick check: if concatenating all value data won't overflow, interleaving
+    // a subset of rows definitely won't either.
+    let total: usize = batches
+        .iter()
+        .map(|(_, batch)| batch.column(col_idx).as_bytes::<T>().value_data().len())
+        .sum();
+    if T::Offset::from_usize(total).is_some() {
+        return Ok(());
+    }
+    // Total exceeds the offset limit, do the precise per-row check.
+    let mut capacity: usize = 0;
+    for &(a, b) in indices {
+        let array = batches[a].1.column(col_idx).as_bytes::<T>();
+        let o = array.value_offsets();
+        let element_len = o[b + 1].as_usize() - o[b].as_usize();
+        capacity += element_len;
+        T::Offset::from_usize(capacity).ok_or(ArrowError::OffsetOverflowError(capacity))?;
+    }
+    Ok(())
+}
+
+/// Checks whether `interleave()` would overflow i32 offsets for `Utf8` or `Binary` columns.
+fn check_interleave_overflow(
+    batches: &[(usize, RecordBatch)],
+    schema: &SchemaRef,
+    indices: &[(usize, usize)],
+) -> Result<()> {
+    for (col_idx, field) in schema.fields.iter().enumerate() {
+        match field.data_type() {
+            DataType::Utf8 => {
+                check_interleave_bytes_overflow::<Utf8Type>(batches, col_idx, indices)
+                    .context(ComputeArrowSnafu)?;
+            }
+            DataType::Binary => {
+                check_interleave_bytes_overflow::<BinaryType>(batches, col_idx, indices)
+                    .context(ComputeArrowSnafu)?;
+            }
+            _ => continue,
+        }
+    }
+    Ok(())
+}
+
 /// Keeps track of the current position in a batch
 #[derive(Debug, Copy, Clone, Default)]
 struct BatchCursor {
@@ -48,6 +105,84 @@ struct BatchCursor {
     row_idx: usize,
 }
 
+/// Trait for reporting merge metrics.
+pub trait MergeMetricsReport: Send + Sync {
+    /// Reports and resets the metrics.
+    fn report(&self, metrics: &mut MergeMetrics);
+}
+
+/// Metrics for the merge reader.
+#[derive(Default)]
+pub struct MergeMetrics {
+    /// Cost to initialize the reader.
+    pub(crate) init_cost: Duration,
+    /// Total scan cost of the reader.
+    pub(crate) scan_cost: Duration,
+    /// Number of times to fetch batches.
+    pub(crate) num_fetch_by_batches: usize,
+    /// Number of times to fetch rows.
+    pub(crate) num_fetch_by_rows: usize,
+    /// Cost to fetch batches from sources.
+    pub(crate) fetch_cost: Duration,
+}
+
+impl fmt::Debug for MergeMetrics {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        if self.scan_cost.is_zero() {
+            return write!(f, "{{}}");
+        }
+
+        write!(f, r#"{{"scan_cost":"{:?}""#, self.scan_cost)?;
+
+        if !self.init_cost.is_zero() {
+            write!(f, r#", "init_cost":"{:?}""#, self.init_cost)?;
+        }
+        if self.num_fetch_by_batches > 0 {
+            write!(
+                f,
+                r#", "num_fetch_by_batches":{}"#,
+                self.num_fetch_by_batches
+            )?;
+        }
+        if self.num_fetch_by_rows > 0 {
+            write!(f, r#", "num_fetch_by_rows":{}"#, self.num_fetch_by_rows)?;
+        }
+        if !self.fetch_cost.is_zero() {
+            write!(f, r#", "fetch_cost":"{:?}""#, self.fetch_cost)?;
+        }
+
+        write!(f, "}}")
+    }
+}
+
+impl MergeMetrics {
+    /// Merges metrics from another MergeMetrics instance.
+    pub(crate) fn merge(&mut self, other: &MergeMetrics) {
+        let MergeMetrics {
+            init_cost,
+            scan_cost,
+            num_fetch_by_batches,
+            num_fetch_by_rows,
+            fetch_cost,
+        } = other;
+
+        self.init_cost += *init_cost;
+        self.scan_cost += *scan_cost;
+        self.num_fetch_by_batches += *num_fetch_by_batches;
+        self.num_fetch_by_rows += *num_fetch_by_rows;
+        self.fetch_cost += *fetch_cost;
+    }
+
+    /// Reports the metrics if scan_cost exceeds 10ms and resets them.
+    pub(crate) fn maybe_report(&mut self, reporter: &Option<Arc<dyn MergeMetricsReport>>) {
+        if self.scan_cost.as_millis() > 10
+            && let Some(r) = reporter
+        {
+            r.report(self);
+        }
+    }
+}
+
 /// Provides an API to incrementally build a [`RecordBatch`] from partitioned [`RecordBatch`]
 // Ports from https://github.com/apache/datafusion/blob/49.0.0/datafusion/physical-plan/src/sorts/builder.rs
 // Adds the `take_remaining_rows()` method.
@@ -121,6 +256,8 @@ impl BatchBuilder {
             return Ok(None);
         }
 
+        check_interleave_overflow(&self.batches, &self.schema, &self.indices)?;
+
         let columns = (0..self.schema.fields.len())
             .map(|column_idx| {
                 let arrays: Vec<_> = self
diff --git a/src/mito2/src/read/flat_projection.rs b/src/mito2/src/read/flat_projection.rs
index 3e0f1169df..523ee6e684 100644
--- a/src/mito2/src/read/flat_projection.rs
+++ b/src/mito2/src/read/flat_projection.rs
@@ -18,22 +18,31 @@ use std::sync::Arc;
 
 use api::v1::SemanticType;
 use common_error::ext::BoxedError;
-use common_recordbatch::error::{ArrowComputeSnafu, ExternalSnafu};
+use common_recordbatch::error::{
+    ArrowComputeSnafu, DataTypesSnafu, ExternalSnafu, NewDfRecordBatchSnafu,
+};
 use common_recordbatch::{DfRecordBatch, RecordBatch};
-use datatypes::arrow::datatypes::Field;
+use datatypes::arrow::array::Array;
+use datatypes::arrow::datatypes::{DataType as ArrowDataType, Field};
+use datatypes::extension::json::is_json_extension_type;
 use datatypes::prelude::{ConcreteDataType, DataType};
 use datatypes::schema::{Schema, SchemaRef};
+use datatypes::value::Value;
 use datatypes::vectors::Helper;
+use datatypes::vectors::json::array::JsonArray;
 use snafu::{OptionExt, ResultExt};
 use store_api::metadata::{RegionMetadata, RegionMetadataRef};
 use store_api::storage::ColumnId;
 
+use crate::cache::CacheStrategy;
 use crate::error::{InvalidRequestSnafu, RecordBatchSnafu, Result};
-use crate::read::projection::read_column_ids_from_projection;
+use crate::read::projection::{read_column_ids_from_projection, repeated_vector_with_cache};
+use crate::read::read_columns::ReadColumns;
 use crate::sst::parquet::flat_format::sst_column_id_indices;
 use crate::sst::parquet::format::FormatProjection;
 use crate::sst::{
     FlatSchemaOptions, internal_fields, tag_maybe_to_dictionary_field, to_flat_sst_arrow_schema,
+    with_field_id,
 };
 
 /// Handles projection and converts batches in flat format with correct schema.
@@ -45,11 +54,11 @@ pub struct FlatProjectionMapper {
     metadata: RegionMetadataRef,
     /// Schema for converted [RecordBatch] to return.
     output_schema: SchemaRef,
-    /// Ids of columns to read from memtables and SSTs.
+    /// The columns to read from memtables and SSTs.
     /// The mapper won't deduplicate the column ids.
     ///
     /// Note that this doesn't contain the `__table_id` and `__tsid`.
-    read_column_ids: Vec<ColumnId>,
+    read_cols: ReadColumns,
     /// Ids and DataTypes of columns of the expected batch.
     /// We can use this to check if the batch is compatible with the expected schema.
     ///
@@ -70,50 +79,48 @@ impl FlatProjectionMapper {
     /// empty `RecordBatch` and only use its row count in this query.
     pub fn new(
         metadata: &RegionMetadataRef,
-        projection: impl Iterator<Item = usize>,
+        projection: impl IntoIterator<Item = usize>,
     ) -> Result<Self> {
-        let projection: Vec<_> = projection.collect();
+        let projection: Vec<_> = projection.into_iter().collect();
         let read_column_ids = read_column_ids_from_projection(metadata, &projection)?;
-        Self::new_with_read_columns(metadata, projection, read_column_ids)
+        let read_cols = ReadColumns::from_deduped_column_ids(read_column_ids);
+        Self::new_with_read_columns(metadata, projection, read_cols)
     }
 
     /// Returns a new mapper with output projection and explicit read columns.
     pub fn new_with_read_columns(
         metadata: &RegionMetadataRef,
         projection: Vec<usize>,
-        read_column_ids: Vec<ColumnId>,
+        read_cols: ReadColumns,
     ) -> Result<Self> {
         // If the original projection is empty.
         let is_empty_projection = projection.is_empty();
 
         // Output column schemas for the projection.
-        let mut column_schemas = Vec::with_capacity(projection.len());
+        let mut col_schemas = Vec::with_capacity(projection.len());
         // Column ids of the output projection without deduplication.
-        let mut output_column_ids = Vec::with_capacity(projection.len());
+        let mut output_col_ids = Vec::with_capacity(projection.len());
         for idx in &projection {
-            // For each projection index, we get the column id for projection.
-            let column =
-                metadata
-                    .column_metadatas
-                    .get(*idx)
-                    .with_context(|| InvalidRequestSnafu {
-                        region_id: metadata.region_id,
-                        reason: format!("projection index {} is out of bound", idx),
-                    })?;
-
-            output_column_ids.push(column.column_id);
-            // Safety: idx is valid.
-            column_schemas.push(metadata.schema.column_schemas()[*idx].clone());
+            let col = metadata
+                .column_metadatas
+                .get(*idx)
+                .with_context(|| InvalidRequestSnafu {
+                    region_id: metadata.region_id,
+                    reason: format!("projection index {} is out of bound", idx),
+                })?;
+            output_col_ids.push(col.column_id);
+            col_schemas.push(col.column_schema.clone());
         }
 
         // Creates a map to lookup index.
         let id_to_index = sst_column_id_indices(metadata);
+
         // TODO(yingwen): Support different flat schema options.
         let format_projection = FormatProjection::compute_format_projection(
             &id_to_index,
             // All columns with internal columns.
             metadata.column_metadatas.len() + 3,
-            read_column_ids.iter().copied(),
+            read_cols.clone(),
         );
 
         let batch_schema = flat_projected_columns(metadata, &format_projection);
@@ -126,13 +133,13 @@ impl FlatProjectionMapper {
             Arc::new(Schema::new(vec![]))
         } else {
             // Safety: Columns come from existing schema.
-            Arc::new(Schema::new(column_schemas))
+            Arc::new(Schema::new(col_schemas))
         };
 
         let batch_indices = if is_empty_projection {
             vec![]
         } else {
-            output_column_ids
+            output_col_ids
                 .iter()
                 .map(|id| {
                     // Safety: The map is computed from the read projection.
@@ -160,7 +167,7 @@ impl FlatProjectionMapper {
         Ok(FlatProjectionMapper {
             metadata: metadata.clone(),
             output_schema,
-            read_column_ids,
+            read_cols,
             batch_schema,
             is_empty_projection,
             batch_indices,
@@ -177,11 +184,9 @@ impl FlatProjectionMapper {
     pub(crate) fn metadata(&self) -> &RegionMetadataRef {
         &self.metadata
     }
-
-    /// Returns ids of projected columns that we need to read
-    /// from memtables and SSTs.
-    pub(crate) fn column_ids(&self) -> &[ColumnId] {
-        &self.read_column_ids
+    /// Returns projected columns that we need to read from memtables and SSTs.
+    pub(crate) fn read_columns(&self) -> &ReadColumns {
+        &self.read_cols
     }
 
     /// Returns the field column start index in output batch.
@@ -237,9 +242,8 @@ impl FlatProjectionMapper {
         self.output_schema.clone()
     }
 
-    /// Returns an empty [RecordBatch].
-    pub(crate) fn empty_record_batch(&self) -> RecordBatch {
-        RecordBatch::new_empty(self.output_schema.clone())
+    pub(crate) fn with_output_schema(&mut self, schema: SchemaRef) {
+        self.output_schema = schema;
     }
 
     /// Converts a flat format [RecordBatch] to a normal [RecordBatch].
@@ -248,12 +252,63 @@ impl FlatProjectionMapper {
     pub(crate) fn convert(
         &self,
         batch: &datatypes::arrow::record_batch::RecordBatch,
+        cache_strategy: &CacheStrategy,
     ) -> common_recordbatch::error::Result<RecordBatch> {
         if self.is_empty_projection {
             return RecordBatch::new_with_count(self.output_schema.clone(), batch.num_rows());
         }
-        let columns = self.project_vectors(batch)?;
-        RecordBatch::new(self.output_schema.clone(), columns)
+        // Construct output record batch directly from Arrow arrays to avoid
+        // Arrow -> Vector -> Arrow roundtrips in the hot path.
+        let mut arrays = Vec::with_capacity(self.output_schema.num_columns());
+        for (output_idx, index) in self.batch_indices.iter().enumerate() {
+            let mut array = batch.column(*index).clone();
+            // Cast dictionary values to the target type.
+            if let ArrowDataType::Dictionary(_key_type, value_type) = array.data_type() {
+                // When a string dictionary column contains only a single value, reuse a cached
+                // repeated vector to avoid repeatedly expanding the dictionary.
+                if let Some(dict_array) = single_value_string_dictionary(
+                    &array,
+                    &self.output_schema.column_schemas()[output_idx].data_type,
+                    value_type.as_ref(),
+                ) {
+                    let dict_values = dict_array.values();
+                    let value = if dict_values.is_null(0) {
+                        Value::Null
+                    } else {
+                        Value::from(datatypes::arrow_array::string_array_value(dict_values, 0))
+                    };
+
+                    let repeated = repeated_vector_with_cache(
+                        &self.output_schema.column_schemas()[output_idx].data_type,
+                        &value,
+                        batch.num_rows(),
+                        cache_strategy,
+                    )?;
+                    array = repeated.to_arrow_array();
+                } else {
+                    let casted = datatypes::arrow::compute::cast(&array, value_type)
+                        .context(ArrowComputeSnafu)?;
+                    array = casted;
+                }
+            }
+
+            let field = &self.output_schema.arrow_schema().fields()[output_idx];
+            if is_json_extension_type(field) {
+                array = JsonArray::from(&array)
+                    .try_align(field.data_type())
+                    .context(DataTypesSnafu)?;
+            }
+
+            arrays.push(array);
+        }
+
+        let df_record_batch =
+            DfRecordBatch::try_new(self.output_schema.arrow_schema().clone(), arrays)
+                .context(NewDfRecordBatchSnafu)?;
+        Ok(RecordBatch::from_df_record_batch(
+            self.output_schema.clone(),
+            df_record_batch,
+        ))
     }
 
     /// Projects columns from the input batch and converts them into vectors.
@@ -281,6 +336,28 @@ impl FlatProjectionMapper {
     }
 }
 
+fn single_value_string_dictionary<'a>(
+    array: &'a Arc<dyn Array>,
+    output_type: &ConcreteDataType,
+    value_type: &ArrowDataType,
+) -> Option<&'a datatypes::arrow::array::DictionaryArray<datatypes::arrow::datatypes::UInt32Type>> {
+    if !matches!(
+        value_type,
+        ArrowDataType::Utf8 | ArrowDataType::LargeUtf8 | ArrowDataType::Utf8View
+    ) || !output_type.is_string()
+    {
+        return None;
+    }
+
+    let dict_array = array
+        .as_any()
+        .downcast_ref::<datatypes::arrow::array::DictionaryArray<
+            datatypes::arrow::datatypes::UInt32Type,
+        >>()?;
+
+    (dict_array.values().len() == 1 && dict_array.null_count() == 0).then_some(dict_array)
+}
+
 /// Returns ids and datatypes of columns of the output batch after applying the `projection`.
 ///
 /// It adds the time index column if it doesn't present in the projection.
@@ -332,17 +409,20 @@ pub(crate) fn compute_input_arrow_schema(
     let mut new_fields = Vec::with_capacity(batch_schema.len() + 3);
     for (column_id, _) in batch_schema {
         let column_metadata = metadata.column_by_id(*column_id).unwrap();
-        let field = Arc::new(Field::new(
+        let field = Field::new(
             &column_metadata.column_schema.name,
             column_metadata.column_schema.data_type.as_arrow_type(),
             column_metadata.column_schema.is_nullable(),
-        ));
-        let field = if column_metadata.semantic_type == SemanticType::Tag {
-            tag_maybe_to_dictionary_field(&column_metadata.column_schema.data_type, &field)
+        );
+        let field = with_field_id(field, *column_id);
+        if column_metadata.semantic_type == SemanticType::Tag {
+            new_fields.push(tag_maybe_to_dictionary_field(
+                &column_metadata.column_schema.data_type,
+                &Arc::new(field),
+            ));
         } else {
-            field
-        };
-        new_fields.push(field);
+            new_fields.push(Arc::new(field));
+        }
     }
     new_fields.extend_from_slice(&internal_fields());
 
@@ -372,15 +452,9 @@ impl CompactionProjectionMapper {
             .chain([metadata.time_index_column_pos()])
             .collect::<Vec<_>>();
 
-        let mapper = FlatProjectionMapper::new_with_read_columns(
-            metadata,
-            projection,
-            metadata
-                .column_metadatas
-                .iter()
-                .map(|col| col.column_id)
-                .collect(),
-        )?;
+        let read_col_ids = metadata.column_metadatas.iter().map(|col| col.column_id);
+        let read_cols = ReadColumns::from_deduped_column_ids(read_col_ids);
+        let mapper = FlatProjectionMapper::new_with_read_columns(metadata, projection, read_cols)?;
         let assembler = DfBatchAssembler::new(mapper.output_schema());
 
         Ok(Self { mapper, assembler })
diff --git a/src/mito2/src/read/last_row.rs b/src/mito2/src/read/last_row.rs
index c2336f218d..301a5cc0b7 100644
--- a/src/mito2/src/read/last_row.rs
+++ b/src/mito2/src/read/last_row.rs
@@ -20,7 +20,7 @@ use async_trait::async_trait;
 use datatypes::arrow::array::{Array, BinaryArray};
 use datatypes::arrow::compute::concat_batches;
 use datatypes::arrow::record_batch::RecordBatch;
-use datatypes::vectors::UInt32Vector;
+use futures::{Stream, TryStreamExt};
 use snafu::ResultExt;
 use store_api::storage::{FileId, TimeSeriesRowSelector};
 
@@ -30,11 +30,12 @@ use crate::cache::{
 };
 use crate::error::{ComputeArrowSnafu, Result};
 use crate::memtable::partition_tree::data::timestamp_array_to_i64_slice;
-use crate::read::{Batch, BatchReader, BoxedBatchReader};
+use crate::read::{Batch, BatchReader, BoxedBatchReader, BoxedRecordBatchStream};
 use crate::sst::parquet::DEFAULT_READ_BATCH_SIZE;
 use crate::sst::parquet::flat_format::{primary_key_column_index, time_index_column_index};
 use crate::sst::parquet::format::{PrimaryKeyArray, primary_key_offsets};
-use crate::sst::parquet::reader::{FlatRowGroupReader, ReaderMetrics, RowGroupReader};
+use crate::sst::parquet::read_columns::ParquetReadColumns;
+use crate::sst::parquet::reader::FlatRowGroupReader;
 
 /// Reader to keep the last row for each time series.
 /// It assumes that batches from the input reader are
@@ -44,6 +45,7 @@ use crate::sst::parquet::reader::{FlatRowGroupReader, ReaderMetrics, RowGroupRea
 ///
 /// This reader is different from the [MergeMode](crate::region::options::MergeMode) as
 /// it focus on time series (the same key).
+#[allow(dead_code)]
 pub(crate) struct LastRowReader {
     /// Inner reader.
     reader: BoxedBatchReader,
@@ -51,6 +53,7 @@ pub(crate) struct LastRowReader {
     selector: LastRowSelector,
 }
 
+#[allow(dead_code)]
 impl LastRowReader {
     /// Creates a new `LastRowReader`.
     pub(crate) fn new(reader: BoxedBatchReader) -> Self {
@@ -78,183 +81,6 @@ impl BatchReader for LastRowReader {
     }
 }
 
-/// Cached last row reader for specific row group.
-/// If the last rows for current row group are already cached, this reader returns the cached value.
-/// If cache misses, [RowGroupLastRowReader] reads last rows from row group and updates the cache
-/// upon finish.
-pub(crate) enum RowGroupLastRowCachedReader {
-    /// Cache hit, reads last rows from cached value.
-    Hit(LastRowCacheReader),
-    /// Cache miss, reads from row group reader and update cache.
-    Miss(RowGroupLastRowReader),
-}
-
-impl RowGroupLastRowCachedReader {
-    pub(crate) fn new(
-        file_id: FileId,
-        row_group_idx: usize,
-        cache_strategy: CacheStrategy,
-        row_group_reader: RowGroupReader,
-    ) -> Self {
-        let key = SelectorResultKey {
-            file_id,
-            row_group_idx,
-            selector: TimeSeriesRowSelector::LastRow,
-        };
-
-        if let Some(value) = cache_strategy.get_selector_result(&key) {
-            let is_primary_key = matches!(&value.result, SelectorResult::PrimaryKey(_));
-            let schema_matches =
-                value.projection == row_group_reader.read_format().projection_indices();
-            if is_primary_key && schema_matches {
-                // Format and schema match, use cache batches.
-                Self::new_hit(value)
-            } else {
-                Self::new_miss(key, row_group_reader, cache_strategy)
-            }
-        } else {
-            Self::new_miss(key, row_group_reader, cache_strategy)
-        }
-    }
-
-    /// Gets the underlying reader metrics if uncached.
-    pub(crate) fn metrics(&self) -> Option<&ReaderMetrics> {
-        match self {
-            RowGroupLastRowCachedReader::Hit(_) => None,
-            RowGroupLastRowCachedReader::Miss(reader) => Some(reader.metrics()),
-        }
-    }
-
-    /// Creates new Hit variant and updates metrics.
-    fn new_hit(value: Arc<SelectorResultValue>) -> Self {
-        selector_result_cache_hit();
-        Self::Hit(LastRowCacheReader { value, idx: 0 })
-    }
-
-    /// Creates new Miss variant and updates metrics.
-    fn new_miss(
-        key: SelectorResultKey,
-        row_group_reader: RowGroupReader,
-        cache_strategy: CacheStrategy,
-    ) -> Self {
-        selector_result_cache_miss();
-        Self::Miss(RowGroupLastRowReader::new(
-            key,
-            row_group_reader,
-            cache_strategy,
-        ))
-    }
-}
-
-#[async_trait]
-impl BatchReader for RowGroupLastRowCachedReader {
-    async fn next_batch(&mut self) -> Result<Option<Batch>> {
-        match self {
-            RowGroupLastRowCachedReader::Hit(r) => r.next_batch().await,
-            RowGroupLastRowCachedReader::Miss(r) => r.next_batch().await,
-        }
-    }
-}
-
-/// Last row reader that returns the cached last rows for row group.
-pub(crate) struct LastRowCacheReader {
-    value: Arc<SelectorResultValue>,
-    idx: usize,
-}
-
-impl LastRowCacheReader {
-    /// Iterates cached last rows.
-    async fn next_batch(&mut self) -> Result<Option<Batch>> {
-        let batches = match &self.value.result {
-            SelectorResult::PrimaryKey(batches) => batches,
-            SelectorResult::Flat(_) => unreachable!(),
-        };
-        if self.idx < batches.len() {
-            let res = Ok(Some(batches[self.idx].clone()));
-            self.idx += 1;
-            res
-        } else {
-            Ok(None)
-        }
-    }
-}
-
-pub(crate) struct RowGroupLastRowReader {
-    key: SelectorResultKey,
-    reader: RowGroupReader,
-    selector: LastRowSelector,
-    yielded_batches: Vec<Batch>,
-    cache_strategy: CacheStrategy,
-    /// Index buffer to take a new batch from the last row.
-    take_index: UInt32Vector,
-}
-
-impl RowGroupLastRowReader {
-    fn new(key: SelectorResultKey, reader: RowGroupReader, cache_strategy: CacheStrategy) -> Self {
-        Self {
-            key,
-            reader,
-            selector: LastRowSelector::default(),
-            yielded_batches: vec![],
-            cache_strategy,
-            take_index: UInt32Vector::from_vec(vec![0]),
-        }
-    }
-
-    async fn next_batch(&mut self) -> Result<Option<Batch>> {
-        while let Some(batch) = self.reader.next_batch().await? {
-            if let Some(yielded) = self.selector.on_next(batch) {
-                push_yielded_batches(yielded.clone(), &self.take_index, &mut self.yielded_batches)?;
-                return Ok(Some(yielded));
-            }
-        }
-        let last_batch = if let Some(last_batch) = self.selector.finish() {
-            push_yielded_batches(
-                last_batch.clone(),
-                &self.take_index,
-                &mut self.yielded_batches,
-            )?;
-            Some(last_batch)
-        } else {
-            None
-        };
-
-        // All last rows in row group are yielded, update cache.
-        self.maybe_update_cache();
-        Ok(last_batch)
-    }
-
-    /// Updates row group's last row cache if cache manager is present.
-    fn maybe_update_cache(&mut self) {
-        if self.yielded_batches.is_empty() {
-            // we always expect that row groups yields batches.
-            return;
-        }
-        let value = Arc::new(SelectorResultValue::new(
-            std::mem::take(&mut self.yielded_batches),
-            self.reader.read_format().projection_indices().to_vec(),
-        ));
-        self.cache_strategy.put_selector_result(self.key, value);
-    }
-
-    fn metrics(&self) -> &ReaderMetrics {
-        self.reader.metrics()
-    }
-}
-
-/// Push last row into `yielded_batches`.
-fn push_yielded_batches(
-    mut batch: Batch,
-    take_index: &UInt32Vector,
-    yielded_batches: &mut Vec<Batch>,
-) -> Result<()> {
-    assert_eq!(1, batch.num_rows());
-    batch.take_in_place(take_index)?;
-    yielded_batches.push(batch);
-
-    Ok(())
-}
-
 /// Common struct that selects only the last row of each time series.
 #[derive(Default)]
 pub struct LastRowSelector {
@@ -309,7 +135,7 @@ impl FlatRowGroupLastRowCachedReader {
         file_id: FileId,
         row_group_idx: usize,
         cache_strategy: CacheStrategy,
-        projection: &[usize],
+        read_cols: &ParquetReadColumns,
         reader: FlatRowGroupReader,
     ) -> Self {
         let key = SelectorResultKey {
@@ -320,22 +146,22 @@ impl FlatRowGroupLastRowCachedReader {
 
         if let Some(value) = cache_strategy.get_selector_result(&key) {
             let is_flat = matches!(&value.result, SelectorResult::Flat(_));
-            let schema_matches = value.projection == projection;
+            let schema_matches = value.read_cols == *read_cols;
             if is_flat && schema_matches {
                 Self::new_hit(value)
             } else {
-                Self::new_miss(key, projection, reader, cache_strategy)
+                Self::new_miss(key, read_cols, reader, cache_strategy)
             }
         } else {
-            Self::new_miss(key, projection, reader, cache_strategy)
+            Self::new_miss(key, read_cols, reader, cache_strategy)
         }
     }
 
     /// Returns the next RecordBatch.
-    pub(crate) fn next_batch(&mut self) -> Result<Option<RecordBatch>> {
+    pub(crate) async fn next_batch(&mut self) -> Result<Option<RecordBatch>> {
         match self {
             FlatRowGroupLastRowCachedReader::Hit(r) => r.next_batch(),
-            FlatRowGroupLastRowCachedReader::Miss(r) => r.next_batch(),
+            FlatRowGroupLastRowCachedReader::Miss(r) => r.next_batch().await,
         }
     }
 
@@ -346,14 +172,14 @@ impl FlatRowGroupLastRowCachedReader {
 
     fn new_miss(
         key: SelectorResultKey,
-        projection: &[usize],
+        read_cols: &ParquetReadColumns,
         reader: FlatRowGroupReader,
         cache_strategy: CacheStrategy,
     ) -> Self {
         selector_result_cache_miss();
         Self::Miss(FlatRowGroupLastRowReader::new(
             key,
-            projection.to_vec(),
+            read_cols.clone(),
             reader,
             cache_strategy,
         ))
@@ -432,7 +258,7 @@ pub(crate) struct FlatRowGroupLastRowReader {
     selector: FlatLastTimestampSelector,
     yielded_batches: Vec<RecordBatch>,
     cache_strategy: CacheStrategy,
-    projection: Vec<usize>,
+    read_cols: ParquetReadColumns,
     /// Accumulates small selector-output batches before concatenating.
     pending: BatchBuffer,
 }
@@ -440,7 +266,7 @@ pub(crate) struct FlatRowGroupLastRowReader {
 impl FlatRowGroupLastRowReader {
     fn new(
         key: SelectorResultKey,
-        projection: Vec<usize>,
+        read_cols: ParquetReadColumns,
         reader: FlatRowGroupReader,
         cache_strategy: CacheStrategy,
     ) -> Self {
@@ -450,7 +276,7 @@ impl FlatRowGroupLastRowReader {
             selector: FlatLastTimestampSelector::default(),
             yielded_batches: vec![],
             cache_strategy,
-            projection,
+            read_cols,
             pending: BatchBuffer::new(),
         }
     }
@@ -465,12 +291,12 @@ impl FlatRowGroupLastRowReader {
         Ok(Some(merged))
     }
 
-    fn next_batch(&mut self) -> Result<Option<RecordBatch>> {
+    async fn next_batch(&mut self) -> Result<Option<RecordBatch>> {
         if self.pending.is_full() {
             return self.flush_pending();
         }
 
-        while let Some(batch) = self.reader.next_batch()? {
+        while let Some(batch) = self.reader.next_batch().await? {
             self.selector.on_next(batch, &mut self.pending)?;
             if self.pending.is_full() {
                 return self.flush_pending();
@@ -498,7 +324,7 @@ impl FlatRowGroupLastRowReader {
         let batches = std::mem::take(&mut self.yielded_batches);
         let value = Arc::new(SelectorResultValue::new_flat(
             batches,
-            self.projection.clone(),
+            self.read_cols.clone(),
         ));
         self.cache_strategy.put_selector_result(self.key, value);
     }
@@ -610,6 +436,41 @@ impl FlatLastTimestampSelector {
     }
 }
 
+/// Reader that keeps only the last row of each time series from a flat RecordBatch stream.
+/// Assumes input is sorted, deduped, and contains no delete operations.
+pub(crate) struct FlatLastRowReader {
+    stream: BoxedRecordBatchStream,
+    selector: FlatLastTimestampSelector,
+    pending: BatchBuffer,
+}
+
+impl FlatLastRowReader {
+    /// Creates a new `FlatLastRowReader`.
+    pub(crate) fn new(stream: BoxedRecordBatchStream) -> Self {
+        Self {
+            stream,
+            selector: FlatLastTimestampSelector::default(),
+            pending: BatchBuffer::new(),
+        }
+    }
+
+    /// Converts the reader into a stream of RecordBatches.
+    pub(crate) fn into_stream(mut self) -> impl Stream<Item = Result<RecordBatch>> {
+        async_stream::try_stream! {
+            while let Some(batch) = self.stream.try_next().await? {
+                self.selector.on_next(batch, &mut self.pending)?;
+                if self.pending.is_full() {
+                    yield self.pending.concat()?;
+                }
+            }
+            self.selector.finish(&mut self.pending)?;
+            if !self.pending.is_empty() {
+                yield self.pending.concat()?;
+            }
+        }
+    }
+}
+
 /// Gets the primary key bytes at `index` from the primary key dictionary column.
 fn primary_key_bytes_at(batch: &RecordBatch, pk_col_idx: usize, index: usize) -> &[u8] {
     let pk_dict = batch
diff --git a/src/mito2/src/read/merge.rs b/src/mito2/src/read/merge.rs
deleted file mode 100644
index 0470e4b01a..0000000000
--- a/src/mito2/src/read/merge.rs
+++ /dev/null
@@ -1,982 +0,0 @@
-// Copyright 2023 Greptime Team
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-//! Merge reader implementation.
-
-use std::cmp::Ordering;
-use std::collections::BinaryHeap;
-use std::sync::Arc;
-use std::time::{Duration, Instant};
-use std::{fmt, mem};
-
-use async_trait::async_trait;
-use common_telemetry::debug;
-
-use crate::error::Result;
-use crate::memtable::BoxedBatchIterator;
-use crate::metrics::READ_STAGE_ELAPSED;
-use crate::read::{Batch, BatchReader, BoxedBatchReader, Source};
-
-/// Trait for reporting merge metrics.
-pub trait MergeMetricsReport: Send + Sync {
-    /// Reports and resets the metrics.
-    fn report(&self, metrics: &mut MergeMetrics);
-}
-
-/// Reader to merge sorted batches.
-///
-/// The merge reader merges [Batch]es from multiple sources that yield sorted batches.
-/// 1. Batch is ordered by primary key, time index, sequence desc, op type desc (we can
-///    ignore op type as sequence is already unique).
-/// 2. Batches from sources **must** not be empty.
-///
-/// The reader won't concatenate batches. Each batch returned by the reader also doesn't
-/// contain duplicate rows. But the last (primary key, timestamp) of a batch may be the same
-/// as the first one in the next batch.
-pub struct MergeReader {
-    /// Holds [Node]s whose key range of current batch **is** overlapped with the merge window.
-    /// Each node yields batches from a `source`.
-    ///
-    /// [Node] in this heap **must** not be empty. A `merge window` is the (primary key, timestamp)
-    /// range of the **root node** in the `hot` heap.
-    hot: BinaryHeap<Node>,
-    /// Holds `Node` whose key range of current batch **isn't** overlapped with the merge window.
-    ///
-    /// `Node` in this heap **must** not be empty.
-    cold: BinaryHeap<Node>,
-    /// Batch to output.
-    output_batch: Option<Batch>,
-    /// Local metrics.
-    metrics: MergeMetrics,
-    /// Optional metrics reporter.
-    metrics_reporter: Option<Arc<dyn MergeMetricsReport>>,
-}
-
-#[async_trait]
-impl BatchReader for MergeReader {
-    async fn next_batch(&mut self) -> Result<Option<Batch>> {
-        let start = Instant::now();
-        while !self.hot.is_empty() && self.output_batch.is_none() {
-            if self.hot.len() == 1 {
-                // No need to do merge sort if only one batch in the hot heap.
-                self.fetch_batch_from_hottest().await?;
-                self.metrics.num_fetch_by_batches += 1;
-            } else {
-                // We could only fetch rows that less than the next node from the hottest node.
-                self.fetch_rows_from_hottest().await?;
-                self.metrics.num_fetch_by_rows += 1;
-            }
-        }
-
-        if let Some(batch) = self.output_batch.take() {
-            self.metrics.scan_cost += start.elapsed();
-            self.metrics.maybe_report(&self.metrics_reporter);
-            Ok(Some(batch))
-        } else {
-            // Nothing fetched.
-            self.metrics.scan_cost += start.elapsed();
-            self.metrics.maybe_report(&self.metrics_reporter);
-            Ok(None)
-        }
-    }
-}
-
-impl Drop for MergeReader {
-    fn drop(&mut self) {
-        debug!("Merge reader finished, metrics: {:?}", self.metrics);
-
-        READ_STAGE_ELAPSED
-            .with_label_values(&["merge"])
-            .observe(self.metrics.scan_cost.as_secs_f64());
-        READ_STAGE_ELAPSED
-            .with_label_values(&["merge_fetch"])
-            .observe(self.metrics.fetch_cost.as_secs_f64());
-
-        // Report any remaining metrics.
-        if let Some(reporter) = &self.metrics_reporter {
-            reporter.report(&mut self.metrics);
-        }
-    }
-}
-
-impl MergeReader {
-    /// Creates and initializes a new [MergeReader].
-    pub async fn new(
-        sources: Vec<Source>,
-        metrics_reporter: Option<Arc<dyn MergeMetricsReport>>,
-    ) -> Result<MergeReader> {
-        let start = Instant::now();
-        let mut metrics = MergeMetrics::default();
-
-        let mut cold = BinaryHeap::with_capacity(sources.len());
-        let hot = BinaryHeap::with_capacity(sources.len());
-        for source in sources {
-            let node = Node::new(source, &mut metrics).await?;
-            if !node.is_eof() {
-                // Ensure `cold` don't have eof nodes.
-                cold.push(node);
-            }
-        }
-
-        let mut reader = MergeReader {
-            hot,
-            cold,
-            output_batch: None,
-            metrics,
-            metrics_reporter,
-        };
-        // Initializes the reader.
-        reader.refill_hot();
-
-        let elapsed = start.elapsed();
-        reader.metrics.init_cost += elapsed;
-        reader.metrics.scan_cost += elapsed;
-        Ok(reader)
-    }
-
-    /// Moves nodes in `cold` heap, whose key range is overlapped with current merge
-    /// window to `hot` heap.
-    fn refill_hot(&mut self) {
-        while !self.cold.is_empty() {
-            if let Some(merge_window) = self.hot.peek() {
-                let warmest = self.cold.peek().unwrap();
-                if warmest.is_behind(merge_window) {
-                    // if the warmest node in the `cold` heap is totally after the
-                    // `merge_window`, then no need to add more nodes into the `hot`
-                    // heap for merge sorting.
-                    break;
-                }
-            }
-
-            let warmest = self.cold.pop().unwrap();
-            self.hot.push(warmest);
-        }
-    }
-
-    /// Fetches one batch from the hottest node.
-    async fn fetch_batch_from_hottest(&mut self) -> Result<()> {
-        assert_eq!(1, self.hot.len());
-
-        let mut hottest = self.hot.pop().unwrap();
-        let batch = hottest.fetch_batch(&mut self.metrics).await?;
-        Self::maybe_output_batch(batch, &mut self.output_batch)?;
-        self.reheap(hottest)
-    }
-
-    /// Fetches non-duplicated rows from the hottest node.
-    async fn fetch_rows_from_hottest(&mut self) -> Result<()> {
-        // Safety: `fetch_batches_to_output()` ensures the hot heap has more than 1 element.
-        // Pop hottest node.
-        let mut top_node = self.hot.pop().unwrap();
-        let top = top_node.current_batch();
-        // Min timestamp and its sequence in the next batch.
-        let next_min_ts = {
-            let next_node = self.hot.peek().unwrap();
-            let next = next_node.current_batch();
-            // top and next have overlapping rows so they must have same primary keys.
-            debug_assert_eq!(top.primary_key(), next.primary_key());
-            // Safety: Batches in the heap is not empty, so we can use unwrap here.
-            next.first_timestamp().unwrap()
-        };
-
-        // Safety: Batches in the heap is not empty, so we can use unwrap here.
-        let timestamps = top.timestamps_native().unwrap();
-        // Binary searches the timestamp in the top batch.
-        // Safety: Batches should have the same timestamp resolution so we can compare the native
-        // value directly.
-        let duplicate_pos = match timestamps.binary_search(&next_min_ts.value()) {
-            Ok(pos) => pos,
-            Err(pos) => {
-                // No duplicate timestamp. Outputs timestamp before `pos`.
-                Self::maybe_output_batch(top.slice(0, pos), &mut self.output_batch)?;
-                top_node.skip_rows(pos, &mut self.metrics).await?;
-                return self.reheap(top_node);
-            }
-        };
-
-        // No need to remove duplicate timestamps.
-        let output_end = if duplicate_pos == 0 {
-            // If the first timestamp of the top node is duplicate. We can simply return the first row
-            // as the heap ensure it is the one with largest sequence.
-            1
-        } else {
-            // We don't know which one has the larger sequence so we use the range before
-            // the duplicate pos.
-            duplicate_pos
-        };
-        Self::maybe_output_batch(top.slice(0, output_end), &mut self.output_batch)?;
-        top_node.skip_rows(output_end, &mut self.metrics).await?;
-        self.reheap(top_node)
-    }
-
-    /// Push the node popped from `hot` back to a proper heap.
-    fn reheap(&mut self, node: Node) -> Result<()> {
-        if node.is_eof() {
-            // If the node is EOF, don't put it into the heap again.
-            // The merge window would be updated, need to refill the hot heap.
-            self.refill_hot();
-        } else {
-            // Find a proper heap for this node.
-            let node_is_cold = if let Some(hottest) = self.hot.peek() {
-                // If key range of this node is behind the hottest node's then we can
-                // push it to the cold heap. Otherwise we should push it to the hot heap.
-                node.is_behind(hottest)
-            } else {
-                // The hot heap is empty, but we don't known whether the current
-                // batch of this node is still the hottest.
-                true
-            };
-
-            if node_is_cold {
-                self.cold.push(node);
-            } else {
-                self.hot.push(node);
-            }
-            // Anyway, the merge window has been changed, we need to refill the hot heap.
-            self.refill_hot();
-        }
-
-        Ok(())
-    }
-
-    /// If `filter_deleted` is set to true, removes deleted entries and sets the `batch` to the `output_batch`.
-    ///
-    /// Ignores the `batch` if it is empty.
-    fn maybe_output_batch(batch: Batch, output_batch: &mut Option<Batch>) -> Result<()> {
-        debug_assert!(output_batch.is_none());
-        if batch.is_empty() {
-            return Ok(());
-        }
-        *output_batch = Some(batch);
-
-        Ok(())
-    }
-}
-
-/// Builder to build and initialize a [MergeReader].
-#[derive(Default)]
-pub struct MergeReaderBuilder {
-    /// Input sources.
-    ///
-    /// All source must yield batches with the same schema.
-    sources: Vec<Source>,
-    /// Optional metrics reporter.
-    metrics_reporter: Option<Arc<dyn MergeMetricsReport>>,
-}
-
-impl MergeReaderBuilder {
-    /// Returns an empty builder.
-    pub fn new() -> MergeReaderBuilder {
-        MergeReaderBuilder::default()
-    }
-
-    /// Creates a builder from sources.
-    pub fn from_sources(sources: Vec<Source>) -> MergeReaderBuilder {
-        MergeReaderBuilder {
-            sources,
-            metrics_reporter: None,
-        }
-    }
-
-    /// Pushes a batch reader to sources.
-    pub fn push_batch_reader(&mut self, reader: BoxedBatchReader) -> &mut Self {
-        self.sources.push(Source::Reader(reader));
-        self
-    }
-
-    /// Pushes a batch iterator to sources.
-    pub fn push_batch_iter(&mut self, iter: BoxedBatchIterator) -> &mut Self {
-        self.sources.push(Source::Iter(iter));
-        self
-    }
-
-    /// Sets the metrics reporter.
-    pub fn with_metrics_reporter(
-        &mut self,
-        reporter: Option<Arc<dyn MergeMetricsReport>>,
-    ) -> &mut Self {
-        self.metrics_reporter = reporter;
-        self
-    }
-
-    /// Builds and initializes the reader, then resets the builder.
-    pub async fn build(&mut self) -> Result<MergeReader> {
-        let sources = mem::take(&mut self.sources);
-        let metrics_reporter = self.metrics_reporter.take();
-        MergeReader::new(sources, metrics_reporter).await
-    }
-}
-
-/// Metrics for the merge reader.
-#[derive(Default)]
-pub struct MergeMetrics {
-    /// Cost to initialize the reader.
-    pub(crate) init_cost: Duration,
-    /// Total scan cost of the reader.
-    pub(crate) scan_cost: Duration,
-    /// Number of times to fetch batches.
-    pub(crate) num_fetch_by_batches: usize,
-    /// Number of times to fetch rows.
-    pub(crate) num_fetch_by_rows: usize,
-    /// Cost to fetch batches from sources.
-    pub(crate) fetch_cost: Duration,
-}
-
-impl fmt::Debug for MergeMetrics {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        // Skip output if scan_cost is zero
-        if self.scan_cost.is_zero() {
-            return write!(f, "{{}}");
-        }
-
-        write!(f, r#"{{"scan_cost":"{:?}""#, self.scan_cost)?;
-
-        if !self.init_cost.is_zero() {
-            write!(f, r#", "init_cost":"{:?}""#, self.init_cost)?;
-        }
-        if self.num_fetch_by_batches > 0 {
-            write!(
-                f,
-                r#", "num_fetch_by_batches":{}"#,
-                self.num_fetch_by_batches
-            )?;
-        }
-        if self.num_fetch_by_rows > 0 {
-            write!(f, r#", "num_fetch_by_rows":{}"#, self.num_fetch_by_rows)?;
-        }
-        if !self.fetch_cost.is_zero() {
-            write!(f, r#", "fetch_cost":"{:?}""#, self.fetch_cost)?;
-        }
-
-        write!(f, "}}")
-    }
-}
-
-impl MergeMetrics {
-    /// Merges metrics from another MergeMetrics instance.
-    pub(crate) fn merge(&mut self, other: &MergeMetrics) {
-        let MergeMetrics {
-            init_cost,
-            scan_cost,
-            num_fetch_by_batches,
-            num_fetch_by_rows,
-            fetch_cost,
-        } = other;
-
-        self.init_cost += *init_cost;
-        self.scan_cost += *scan_cost;
-        self.num_fetch_by_batches += *num_fetch_by_batches;
-        self.num_fetch_by_rows += *num_fetch_by_rows;
-        self.fetch_cost += *fetch_cost;
-    }
-
-    /// Reports the metrics if scan_cost exceeds 10ms and resets them.
-    pub(crate) fn maybe_report(&mut self, reporter: &Option<Arc<dyn MergeMetricsReport>>) {
-        if self.scan_cost.as_millis() > 10
-            && let Some(r) = reporter
-        {
-            r.report(self);
-        }
-    }
-}
-
-/// A `Node` represent an individual input data source to be merged.
-struct Node {
-    /// Data source of this `Node`.
-    source: Source,
-    /// Current batch to be read. The node ensures the batch is not empty.
-    ///
-    /// `None` means the `source` has reached EOF.
-    current_batch: Option<CompareFirst>,
-}
-
-impl Node {
-    /// Initialize a node.
-    ///
-    /// It tries to fetch one batch from the `source`.
-    async fn new(mut source: Source, metrics: &mut MergeMetrics) -> Result<Node> {
-        // Ensures batch is not empty.
-        let start = Instant::now();
-        let current_batch = source.next_batch().await?.map(CompareFirst);
-        metrics.fetch_cost += start.elapsed();
-
-        Ok(Node {
-            source,
-            current_batch,
-        })
-    }
-
-    /// Returns whether the node still has batch to read.
-    fn is_eof(&self) -> bool {
-        self.current_batch.is_none()
-    }
-
-    /// Returns the primary key of current batch.
-    ///
-    /// # Panics
-    /// Panics if the node has reached EOF.
-    fn primary_key(&self) -> &[u8] {
-        self.current_batch().primary_key()
-    }
-
-    /// Returns current batch.
-    ///
-    /// # Panics
-    /// Panics if the node has reached EOF.
-    fn current_batch(&self) -> &Batch {
-        &self.current_batch.as_ref().unwrap().0
-    }
-
-    /// Returns current batch and fetches next batch
-    /// from the source.
-    ///
-    /// # Panics
-    /// Panics if the node has reached EOF.
-    async fn fetch_batch(&mut self, metrics: &mut MergeMetrics) -> Result<Batch> {
-        let current = self.current_batch.take().unwrap();
-        let start = Instant::now();
-        // Ensures batch is not empty.
-        self.current_batch = self.source.next_batch().await?.map(CompareFirst);
-        metrics.fetch_cost += start.elapsed();
-        Ok(current.0)
-    }
-
-    /// Returns true if the key range of current batch in `self` is behind (exclusive) current
-    /// batch in `other`.
-    ///
-    /// # Panics
-    /// Panics if either `self` or `other` is EOF.
-    fn is_behind(&self, other: &Node) -> bool {
-        debug_assert!(!self.current_batch().is_empty());
-        debug_assert!(!other.current_batch().is_empty());
-
-        // We only compare pk and timestamp so nodes in the cold
-        // heap don't have overlapping timestamps with the hottest node
-        // in the hot heap.
-        self.primary_key().cmp(other.primary_key()).then_with(|| {
-            self.current_batch()
-                .first_timestamp()
-                .cmp(&other.current_batch().last_timestamp())
-        }) == Ordering::Greater
-    }
-
-    /// Skips first `num_to_skip` rows from node's current batch. If current batch is empty it fetches
-    /// next batch from the node.
-    ///
-    /// # Panics
-    /// Panics if the node is EOF.
-    async fn skip_rows(&mut self, num_to_skip: usize, metrics: &mut MergeMetrics) -> Result<()> {
-        let batch = self.current_batch();
-        debug_assert!(batch.num_rows() >= num_to_skip);
-
-        let remaining = batch.num_rows() - num_to_skip;
-        if remaining == 0 {
-            // Nothing remains, we need to fetch next batch to ensure the batch is not empty.
-            self.fetch_batch(metrics).await?;
-        } else {
-            debug_assert!(!batch.is_empty());
-            self.current_batch = Some(CompareFirst(batch.slice(num_to_skip, remaining)));
-        }
-
-        Ok(())
-    }
-}
-
-impl PartialEq for Node {
-    fn eq(&self, other: &Node) -> bool {
-        self.current_batch == other.current_batch
-    }
-}
-
-impl Eq for Node {}
-
-impl PartialOrd for Node {
-    fn partial_cmp(&self, other: &Node) -> Option<Ordering> {
-        Some(self.cmp(other))
-    }
-}
-
-impl Ord for Node {
-    fn cmp(&self, other: &Node) -> Ordering {
-        // The std binary heap is a max heap, but we want the nodes are ordered in
-        // ascend order, so we compare the nodes in reverse order.
-        other.current_batch.cmp(&self.current_batch)
-    }
-}
-
-/// Type to compare [Batch] by first row.
-///
-/// It ignores op type as sequence is enough to distinguish different rows.
-struct CompareFirst(Batch);
-
-impl PartialEq for CompareFirst {
-    fn eq(&self, other: &Self) -> bool {
-        self.0.primary_key() == other.0.primary_key()
-            && self.0.first_timestamp() == other.0.first_timestamp()
-            && self.0.first_sequence() == other.0.first_sequence()
-    }
-}
-
-impl Eq for CompareFirst {}
-
-impl PartialOrd for CompareFirst {
-    fn partial_cmp(&self, other: &CompareFirst) -> Option<Ordering> {
-        Some(self.cmp(other))
-    }
-}
-
-impl Ord for CompareFirst {
-    /// Compares by primary key, time index, sequence desc.
-    fn cmp(&self, other: &CompareFirst) -> Ordering {
-        self.0
-            .primary_key()
-            .cmp(other.0.primary_key())
-            .then_with(|| self.0.first_timestamp().cmp(&other.0.first_timestamp()))
-            .then_with(|| other.0.first_sequence().cmp(&self.0.first_sequence()))
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use api::v1::OpType;
-
-    use super::*;
-    use crate::test_util::{VecBatchReader, check_reader_result, new_batch};
-
-    #[tokio::test]
-    async fn test_merge_reader_empty() {
-        let mut reader = MergeReaderBuilder::new().build().await.unwrap();
-        assert!(reader.next_batch().await.unwrap().is_none());
-        assert!(reader.next_batch().await.unwrap().is_none());
-    }
-
-    #[tokio::test]
-    async fn test_merge_non_overlapping() {
-        let reader1 = VecBatchReader::new(&[
-            new_batch(
-                b"k1",
-                &[1, 2],
-                &[11, 12],
-                &[OpType::Put, OpType::Put],
-                &[21, 22],
-            ),
-            new_batch(
-                b"k1",
-                &[7, 8],
-                &[17, 18],
-                &[OpType::Put, OpType::Delete],
-                &[27, 28],
-            ),
-            new_batch(
-                b"k2",
-                &[2, 3],
-                &[12, 13],
-                &[OpType::Delete, OpType::Put],
-                &[22, 23],
-            ),
-        ]);
-        let reader2 = VecBatchReader::new(&[new_batch(
-            b"k1",
-            &[4, 5],
-            &[14, 15],
-            &[OpType::Put, OpType::Put],
-            &[24, 25],
-        )]);
-        let mut reader = MergeReaderBuilder::new()
-            .push_batch_reader(Box::new(reader1))
-            .push_batch_iter(Box::new(reader2))
-            .build()
-            .await
-            .unwrap();
-        check_reader_result(
-            &mut reader,
-            &[
-                new_batch(
-                    b"k1",
-                    &[1, 2],
-                    &[11, 12],
-                    &[OpType::Put, OpType::Put],
-                    &[21, 22],
-                ),
-                new_batch(
-                    b"k1",
-                    &[4, 5],
-                    &[14, 15],
-                    &[OpType::Put, OpType::Put],
-                    &[24, 25],
-                ),
-                new_batch(
-                    b"k1",
-                    &[7, 8],
-                    &[17, 18],
-                    &[OpType::Put, OpType::Delete],
-                    &[27, 28],
-                ),
-                new_batch(
-                    b"k2",
-                    &[2, 3],
-                    &[12, 13],
-                    &[OpType::Delete, OpType::Put],
-                    &[22, 23],
-                ),
-            ],
-        )
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_merge_reheap_hot() {
-        let reader1 = VecBatchReader::new(&[
-            new_batch(
-                b"k1",
-                &[1, 3],
-                &[10, 10],
-                &[OpType::Put, OpType::Put],
-                &[21, 23],
-            ),
-            new_batch(b"k2", &[3], &[10], &[OpType::Put], &[23]),
-        ]);
-        let reader2 = VecBatchReader::new(&[new_batch(
-            b"k1",
-            &[2, 4],
-            &[11, 11],
-            &[OpType::Put, OpType::Put],
-            &[32, 34],
-        )]);
-        let mut reader = MergeReaderBuilder::new()
-            .push_batch_reader(Box::new(reader1))
-            .push_batch_iter(Box::new(reader2))
-            .build()
-            .await
-            .unwrap();
-        check_reader_result(
-            &mut reader,
-            &[
-                new_batch(b"k1", &[1], &[10], &[OpType::Put], &[21]),
-                new_batch(b"k1", &[2], &[11], &[OpType::Put], &[32]),
-                new_batch(b"k1", &[3], &[10], &[OpType::Put], &[23]),
-                new_batch(b"k1", &[4], &[11], &[OpType::Put], &[34]),
-                new_batch(b"k2", &[3], &[10], &[OpType::Put], &[23]),
-            ],
-        )
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_merge_overlapping() {
-        let reader1 = VecBatchReader::new(&[
-            new_batch(
-                b"k1",
-                &[1, 2],
-                &[11, 12],
-                &[OpType::Put, OpType::Put],
-                &[21, 22],
-            ),
-            new_batch(
-                b"k1",
-                &[4, 5],
-                &[14, 15],
-                // This override 4 and deletes 5.
-                &[OpType::Put, OpType::Delete],
-                &[24, 25],
-            ),
-            new_batch(
-                b"k2",
-                &[2, 3],
-                &[12, 13],
-                // This delete 2.
-                &[OpType::Delete, OpType::Put],
-                &[22, 23],
-            ),
-        ]);
-        let reader2 = VecBatchReader::new(&[
-            new_batch(
-                b"k1",
-                &[3, 4, 5],
-                &[10, 10, 10],
-                &[OpType::Put, OpType::Put, OpType::Put],
-                &[33, 34, 35],
-            ),
-            new_batch(
-                b"k2",
-                &[1, 10],
-                &[11, 20],
-                &[OpType::Put, OpType::Put],
-                &[21, 30],
-            ),
-        ]);
-        let mut reader = MergeReaderBuilder::new()
-            .push_batch_reader(Box::new(reader1))
-            .push_batch_iter(Box::new(reader2))
-            .build()
-            .await
-            .unwrap();
-        check_reader_result(
-            &mut reader,
-            &[
-                new_batch(
-                    b"k1",
-                    &[1, 2],
-                    &[11, 12],
-                    &[OpType::Put, OpType::Put],
-                    &[21, 22],
-                ),
-                new_batch(b"k1", &[3], &[10], &[OpType::Put], &[33]),
-                new_batch(b"k1", &[4], &[14], &[OpType::Put], &[24]),
-                new_batch(b"k1", &[4], &[10], &[OpType::Put], &[34]),
-                new_batch(b"k1", &[5], &[15], &[OpType::Delete], &[25]),
-                new_batch(b"k1", &[5], &[10], &[OpType::Put], &[35]),
-                new_batch(b"k2", &[1], &[11], &[OpType::Put], &[21]),
-                new_batch(
-                    b"k2",
-                    &[2, 3],
-                    &[12, 13],
-                    &[OpType::Delete, OpType::Put],
-                    &[22, 23],
-                ),
-                new_batch(b"k2", &[10], &[20], &[OpType::Put], &[30]),
-            ],
-        )
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_merge_deleted() {
-        let reader1 = VecBatchReader::new(&[
-            new_batch(
-                b"k1",
-                &[1, 2],
-                &[11, 12],
-                &[OpType::Delete, OpType::Delete],
-                &[21, 22],
-            ),
-            new_batch(
-                b"k2",
-                &[2, 3],
-                &[12, 13],
-                &[OpType::Delete, OpType::Put],
-                &[22, 23],
-            ),
-        ]);
-        let reader2 = VecBatchReader::new(&[new_batch(
-            b"k1",
-            &[4, 5],
-            &[14, 15],
-            &[OpType::Delete, OpType::Delete],
-            &[24, 25],
-        )]);
-        let mut reader = MergeReaderBuilder::new()
-            .push_batch_reader(Box::new(reader1))
-            .push_batch_iter(Box::new(reader2))
-            .build()
-            .await
-            .unwrap();
-        check_reader_result(
-            &mut reader,
-            &[
-                new_batch(
-                    b"k1",
-                    &[1, 2],
-                    &[11, 12],
-                    &[OpType::Delete, OpType::Delete],
-                    &[21, 22],
-                ),
-                new_batch(
-                    b"k1",
-                    &[4, 5],
-                    &[14, 15],
-                    &[OpType::Delete, OpType::Delete],
-                    &[24, 25],
-                ),
-                new_batch(
-                    b"k2",
-                    &[2, 3],
-                    &[12, 13],
-                    &[OpType::Delete, OpType::Put],
-                    &[22, 23],
-                ),
-            ],
-        )
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_merge_next_node_empty() {
-        let reader1 = VecBatchReader::new(&[new_batch(
-            b"k1",
-            &[1, 2],
-            &[11, 12],
-            &[OpType::Put, OpType::Put],
-            &[21, 22],
-        )]);
-        let reader2 = VecBatchReader::new(&[new_batch(b"k1", &[1], &[10], &[OpType::Put], &[33])]);
-        let mut reader = MergeReaderBuilder::new()
-            .push_batch_reader(Box::new(reader1))
-            .push_batch_iter(Box::new(reader2))
-            .build()
-            .await
-            .unwrap();
-        check_reader_result(
-            &mut reader,
-            &[
-                new_batch(b"k1", &[1], &[11], &[OpType::Put], &[21]),
-                new_batch(b"k1", &[1], &[10], &[OpType::Put], &[33]),
-                new_batch(b"k1", &[2], &[12], &[OpType::Put], &[22]),
-            ],
-        )
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_merge_top_node_empty() {
-        let reader1 = VecBatchReader::new(&[new_batch(
-            b"k1",
-            &[1, 2],
-            &[10, 10],
-            &[OpType::Put, OpType::Put],
-            &[21, 22],
-        )]);
-        let reader2 = VecBatchReader::new(&[new_batch(
-            b"k1",
-            &[2, 3],
-            &[11, 11],
-            &[OpType::Put, OpType::Put],
-            &[32, 33],
-        )]);
-        let mut reader = MergeReaderBuilder::new()
-            .push_batch_reader(Box::new(reader1))
-            .push_batch_iter(Box::new(reader2))
-            .build()
-            .await
-            .unwrap();
-        check_reader_result(
-            &mut reader,
-            &[
-                new_batch(b"k1", &[1], &[10], &[OpType::Put], &[21]),
-                new_batch(b"k1", &[2], &[11], &[OpType::Put], &[32]),
-                new_batch(b"k1", &[2], &[10], &[OpType::Put], &[22]),
-                new_batch(b"k1", &[3], &[11], &[OpType::Put], &[33]),
-            ],
-        )
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_merge_large_range() {
-        let reader1 = VecBatchReader::new(&[new_batch(
-            b"k1",
-            &[1, 10],
-            &[10, 10],
-            &[OpType::Put, OpType::Put],
-            &[21, 30],
-        )]);
-        let reader2 = VecBatchReader::new(&[new_batch(
-            b"k1",
-            &[1, 20],
-            &[11, 11],
-            &[OpType::Put, OpType::Put],
-            &[31, 40],
-        )]);
-        // The hot heap have a node that doesn't have duplicate
-        // timestamps.
-        let reader3 = VecBatchReader::new(&[new_batch(
-            b"k1",
-            &[6, 8],
-            &[11, 11],
-            &[OpType::Put, OpType::Put],
-            &[36, 38],
-        )]);
-        let mut reader = MergeReaderBuilder::new()
-            .push_batch_reader(Box::new(reader1))
-            .push_batch_iter(Box::new(reader2))
-            .push_batch_reader(Box::new(reader3))
-            .build()
-            .await
-            .unwrap();
-        check_reader_result(
-            &mut reader,
-            &[
-                new_batch(b"k1", &[1], &[11], &[OpType::Put], &[31]),
-                new_batch(b"k1", &[1], &[10], &[OpType::Put], &[21]),
-                new_batch(
-                    b"k1",
-                    &[6, 8],
-                    &[11, 11],
-                    &[OpType::Put, OpType::Put],
-                    &[36, 38],
-                ),
-                new_batch(b"k1", &[10], &[10], &[OpType::Put], &[30]),
-                new_batch(b"k1", &[20], &[11], &[OpType::Put], &[40]),
-            ],
-        )
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_merge_many_duplicates() {
-        let mut builder = MergeReaderBuilder::new();
-        for i in 0..10 {
-            let batches: Vec<_> = (0..8)
-                .map(|ts| new_batch(b"k1", &[ts], &[i], &[OpType::Put], &[100]))
-                .collect();
-            let reader = VecBatchReader::new(&batches);
-            builder.push_batch_reader(Box::new(reader));
-        }
-        let mut reader = builder.build().await.unwrap();
-        let mut expect = Vec::with_capacity(80);
-        for ts in 0..8 {
-            for i in 0..10 {
-                let batch = new_batch(b"k1", &[ts], &[9 - i], &[OpType::Put], &[100]);
-                expect.push(batch);
-            }
-        }
-        check_reader_result(&mut reader, &expect).await;
-    }
-
-    #[tokio::test]
-    async fn test_merge_keep_duplicate() {
-        let reader1 = VecBatchReader::new(&[new_batch(
-            b"k1",
-            &[1, 2],
-            &[10, 10],
-            &[OpType::Put, OpType::Put],
-            &[21, 22],
-        )]);
-        let reader2 = VecBatchReader::new(&[new_batch(
-            b"k1",
-            &[2, 3],
-            &[11, 11],
-            &[OpType::Put, OpType::Put],
-            &[32, 33],
-        )]);
-        let sources = vec![
-            Source::Reader(Box::new(reader1)),
-            Source::Iter(Box::new(reader2)),
-        ];
-        let mut reader = MergeReaderBuilder::from_sources(sources)
-            .build()
-            .await
-            .unwrap();
-        check_reader_result(
-            &mut reader,
-            &[
-                new_batch(b"k1", &[1], &[10], &[OpType::Put], &[21]),
-                new_batch(b"k1", &[2], &[11], &[OpType::Put], &[32]),
-                new_batch(b"k1", &[2], &[10], &[OpType::Put], &[22]),
-                new_batch(b"k1", &[3], &[11], &[OpType::Put], &[33]),
-            ],
-        )
-        .await;
-    }
-}
diff --git a/src/mito2/src/read/plain_batch.rs b/src/mito2/src/read/plain_batch.rs
deleted file mode 100644
index f22b6688d6..0000000000
--- a/src/mito2/src/read/plain_batch.rs
+++ /dev/null
@@ -1,505 +0,0 @@
-// Copyright 2023 Greptime Team
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-//! Batch without an encoded primary key.
-
-use std::collections::HashMap;
-use std::sync::Arc;
-
-use api::v1::OpType;
-use datatypes::arrow::array::{ArrayRef, BooleanArray, UInt8Array, UInt64Array};
-use datatypes::arrow::compute::filter_record_batch;
-use datatypes::arrow::datatypes::SchemaRef;
-use datatypes::arrow::record_batch::RecordBatch;
-use snafu::{OptionExt, ResultExt};
-use store_api::metadata::{ColumnMetadata, RegionMetadata};
-use store_api::storage::{RegionId, SequenceNumber};
-
-use crate::error::{
-    ComputeArrowSnafu, CreateDefaultSnafu, InvalidRequestSnafu, NewRecordBatchSnafu, Result,
-    UnexpectedSnafu,
-};
-
-/// Number of columns that have fixed positions.
-///
-/// Contains all internal columns.
-pub(crate) const PLAIN_FIXED_POS_COLUMN_NUM: usize = 2;
-
-/// [PlainBatch] represents a batch of rows.
-/// It is a wrapper around [RecordBatch].
-///
-/// The columns order is the same as the order of the columns read from the SST.
-/// It always contains two internal columns now. We may change modify this behavior
-/// in the future.
-#[derive(Debug)]
-pub struct PlainBatch {
-    /// The original record batch.
-    record_batch: RecordBatch,
-}
-
-impl PlainBatch {
-    /// Creates a new [PlainBatch] from a [RecordBatch].
-    pub fn new(record_batch: RecordBatch) -> Self {
-        assert!(
-            record_batch.num_columns() >= 2,
-            "record batch missing internal columns, num_columns: {}",
-            record_batch.num_columns()
-        );
-
-        Self { record_batch }
-    }
-
-    /// Returns a new [PlainBatch] with the given columns.
-    pub fn with_new_columns(&self, columns: Vec<ArrayRef>) -> Result<Self> {
-        let record_batch = RecordBatch::try_new(self.record_batch.schema(), columns)
-            .context(NewRecordBatchSnafu)?;
-        Ok(Self::new(record_batch))
-    }
-
-    /// Returns the number of columns in the batch.
-    pub fn num_columns(&self) -> usize {
-        self.record_batch.num_columns()
-    }
-
-    /// Returns the number of rows in the batch.
-    pub fn num_rows(&self) -> usize {
-        self.record_batch.num_rows()
-    }
-
-    /// Returns true if the batch is empty.
-    pub fn is_empty(&self) -> bool {
-        self.num_rows() == 0
-    }
-
-    /// Returns all columns.
-    pub fn columns(&self) -> &[ArrayRef] {
-        self.record_batch.columns()
-    }
-
-    /// Returns the array of column at index `idx`.
-    pub fn column(&self, idx: usize) -> &ArrayRef {
-        self.record_batch.column(idx)
-    }
-
-    /// Returns the slice of internal columns.
-    pub fn internal_columns(&self) -> &[ArrayRef] {
-        &self.record_batch.columns()[self.record_batch.num_columns() - PLAIN_FIXED_POS_COLUMN_NUM..]
-    }
-
-    /// Returns the inner record batch.
-    pub fn as_record_batch(&self) -> &RecordBatch {
-        &self.record_batch
-    }
-
-    /// Converts this batch into a record batch.
-    pub fn into_record_batch(self) -> RecordBatch {
-        self.record_batch
-    }
-
-    /// Filters this batch by the boolean array.
-    pub fn filter(&self, predicate: &BooleanArray) -> Result<Self> {
-        let record_batch =
-            filter_record_batch(&self.record_batch, predicate).context(ComputeArrowSnafu)?;
-        Ok(Self::new(record_batch))
-    }
-
-    /// Returns the column index of the sequence column.
-    #[allow(dead_code)]
-    pub(crate) fn sequence_column_index(&self) -> usize {
-        self.record_batch.num_columns() - PLAIN_FIXED_POS_COLUMN_NUM
-    }
-}
-
-/// Helper struct to fill default values and internal columns.
-pub struct ColumnFiller<'a> {
-    /// Region metadata information
-    metadata: &'a RegionMetadata,
-    /// Schema for the output record batch
-    schema: SchemaRef,
-    /// Map of column names to indices in the input record batch
-    name_to_index: HashMap<String, usize>,
-}
-
-impl<'a> ColumnFiller<'a> {
-    /// Creates a new ColumnFiller
-    /// The `schema` is the sst schema of the `metadata`.
-    pub fn new(
-        metadata: &'a RegionMetadata,
-        schema: SchemaRef,
-        record_batch: &RecordBatch,
-    ) -> Self {
-        debug_assert_eq!(metadata.column_metadatas.len() + 2, schema.fields().len());
-
-        // Pre-construct the name to index map
-        let name_to_index: HashMap<_, _> = record_batch
-            .schema()
-            .fields()
-            .iter()
-            .enumerate()
-            .map(|(i, field)| (field.name().clone(), i))
-            .collect();
-
-        Self {
-            metadata,
-            schema,
-            name_to_index,
-        }
-    }
-
-    /// Fills default values and internal columns for a [RecordBatch].
-    pub fn fill_missing_columns(
-        &self,
-        record_batch: &RecordBatch,
-        sequence: SequenceNumber,
-        op_type: OpType,
-    ) -> Result<RecordBatch> {
-        let num_rows = record_batch.num_rows();
-        let mut new_columns =
-            Vec::with_capacity(record_batch.num_columns() + PLAIN_FIXED_POS_COLUMN_NUM);
-
-        // Fills default values.
-        // Implementation based on `WriteRequest::fill_missing_columns()`.
-        for column in &self.metadata.column_metadatas {
-            let array = match self.name_to_index.get(&column.column_schema.name) {
-                Some(index) => record_batch.column(*index).clone(),
-                None => match op_type {
-                    OpType::Put => {
-                        // For put requests, we use the default value from column schema.
-                        fill_column_put_default(self.metadata.region_id, column, num_rows)?
-                    }
-                    OpType::Delete => {
-                        // For delete requests, we need default value for padding.
-                        fill_column_delete_default(column, num_rows)?
-                    }
-                },
-            };
-
-            new_columns.push(array);
-        }
-
-        // Adds internal columns.
-        // Adds the sequence number.
-        let sequence_array = Arc::new(UInt64Array::from(vec![sequence; num_rows]));
-        // Adds the op type.
-        let op_type_array = Arc::new(UInt8Array::from(vec![op_type as u8; num_rows]));
-        new_columns.push(sequence_array);
-        new_columns.push(op_type_array);
-
-        RecordBatch::try_new(self.schema.clone(), new_columns).context(NewRecordBatchSnafu)
-    }
-}
-
-fn fill_column_put_default(
-    region_id: RegionId,
-    column: &ColumnMetadata,
-    num_rows: usize,
-) -> Result<ArrayRef> {
-    if column.column_schema.is_default_impure() {
-        return UnexpectedSnafu {
-            reason: format!(
-                "unexpected impure default value with region_id: {}, column: {}, default_value: {:?}",
-                region_id,
-                column.column_schema.name,
-                column.column_schema.default_constraint(),
-            ),
-        }
-        .fail();
-    }
-    let vector = column
-        .column_schema
-        .create_default_vector(num_rows)
-        .context(CreateDefaultSnafu {
-            region_id,
-            column: &column.column_schema.name,
-        })?
-        // This column doesn't have default value.
-        .with_context(|| InvalidRequestSnafu {
-            region_id,
-            reason: format!(
-                "column {} does not have default value",
-                column.column_schema.name
-            ),
-        })?;
-    Ok(vector.to_arrow_array())
-}
-
-fn fill_column_delete_default(column: &ColumnMetadata, num_rows: usize) -> Result<ArrayRef> {
-    // For delete requests, we need a default value for padding
-    let vector = column
-        .column_schema
-        .create_default_vector_for_padding(num_rows);
-    Ok(vector.to_arrow_array())
-}
-
-#[cfg(test)]
-mod tests {
-    use api::v1::SemanticType;
-    use datatypes::arrow::array::{
-        Float64Array, Int32Array, StringArray, TimestampMillisecondArray,
-    };
-    use datatypes::arrow::datatypes::{DataType, Field, Schema, TimeUnit};
-    use datatypes::schema::ColumnSchema;
-    use datatypes::schema::constraint::ColumnDefaultConstraint;
-    use datatypes::value::Value;
-    use store_api::metadata::{ColumnMetadata, RegionMetadataBuilder};
-    use store_api::storage::consts::{OP_TYPE_COLUMN_NAME, SEQUENCE_COLUMN_NAME};
-    use store_api::storage::{ConcreteDataType, RegionId};
-
-    use super::*;
-    use crate::sst::to_plain_sst_arrow_schema;
-
-    /// Creates a test region metadata with schema: k0(string), ts(timestamp), v1(float64)
-    fn create_test_region_metadata() -> RegionMetadata {
-        let mut builder = RegionMetadataBuilder::new(RegionId::new(100, 200));
-        builder
-            // Add string key column
-            .push_column_metadata(ColumnMetadata {
-                column_schema: ColumnSchema::new("k0", ConcreteDataType::string_datatype(), false)
-                    .with_default_constraint(None)
-                    .unwrap(),
-                semantic_type: SemanticType::Tag,
-                column_id: 0,
-            })
-            // Add timestamp column
-            .push_column_metadata(ColumnMetadata {
-                column_schema: ColumnSchema::new(
-                    "ts",
-                    ConcreteDataType::timestamp_millisecond_datatype(),
-                    false,
-                )
-                .with_time_index(true)
-                .with_default_constraint(None)
-                .unwrap(),
-                semantic_type: SemanticType::Timestamp,
-                column_id: 1,
-            })
-            // Add float value column with default
-            .push_column_metadata(ColumnMetadata {
-                column_schema: ColumnSchema::new("v1", ConcreteDataType::float64_datatype(), true)
-                    .with_default_constraint(Some(ColumnDefaultConstraint::Value(Value::Float64(
-                        datatypes::value::OrderedFloat::from(42.0),
-                    ))))
-                    .unwrap(),
-                semantic_type: SemanticType::Field,
-                column_id: 2,
-            })
-            .primary_key(vec![0]);
-
-        builder.build().unwrap()
-    }
-
-    #[test]
-    fn test_column_filler_put() {
-        let region_metadata = create_test_region_metadata();
-        let output_schema = to_plain_sst_arrow_schema(&region_metadata);
-
-        // Create input record batch with only k0 and ts columns (v1 is missing)
-        let input_schema = Arc::new(Schema::new(vec![
-            Field::new("k0", DataType::Utf8, false),
-            Field::new(
-                "ts",
-                DataType::Timestamp(TimeUnit::Millisecond, None),
-                false,
-            ),
-        ]));
-
-        let k0_values: ArrayRef = Arc::new(StringArray::from(vec!["key1", "key2"]));
-        let ts_values: ArrayRef = Arc::new(TimestampMillisecondArray::from(vec![1000, 2000]));
-
-        let input_batch =
-            RecordBatch::try_new(input_schema, vec![k0_values.clone(), ts_values.clone()]).unwrap();
-
-        // Create column filler
-        let filler = ColumnFiller::new(&region_metadata, output_schema.clone(), &input_batch);
-
-        // Fill missing columns with OpType::Put
-        let result = filler
-            .fill_missing_columns(&input_batch, 100, OpType::Put)
-            .unwrap();
-
-        // Verify the result
-        // Create an expected record batch to compare against
-        let expected_columns = vec![
-            k0_values.clone(),
-            ts_values.clone(),
-            Arc::new(Float64Array::from(vec![42.0, 42.0])),
-            Arc::new(UInt64Array::from(vec![100, 100])),
-            Arc::new(UInt8Array::from(vec![OpType::Put as u8, OpType::Put as u8])),
-        ];
-        let expected_batch = RecordBatch::try_new(output_schema.clone(), expected_columns).unwrap();
-        assert_eq!(expected_batch, result);
-    }
-
-    #[test]
-    fn test_column_filler_delete() {
-        let region_metadata = create_test_region_metadata();
-        let output_schema = to_plain_sst_arrow_schema(&region_metadata);
-
-        // Create input record batch with only k0 and ts columns (v1 is missing)
-        let input_schema = Arc::new(Schema::new(vec![
-            Field::new("k0", DataType::Utf8, false),
-            Field::new(
-                "ts",
-                DataType::Timestamp(TimeUnit::Millisecond, None),
-                false,
-            ),
-        ]));
-
-        let k0_values: ArrayRef = Arc::new(StringArray::from(vec!["key1", "key2"]));
-        let ts_values: ArrayRef = Arc::new(TimestampMillisecondArray::from(vec![1000, 2000]));
-
-        let input_batch =
-            RecordBatch::try_new(input_schema, vec![k0_values.clone(), ts_values.clone()]).unwrap();
-
-        // Create column filler
-        let filler = ColumnFiller::new(&region_metadata, output_schema.clone(), &input_batch);
-
-        // Fill missing columns with OpType::Delete
-        let result = filler
-            .fill_missing_columns(&input_batch, 200, OpType::Delete)
-            .unwrap();
-
-        // Verify the result by creating an expected record batch to compare against
-        let v1_default = Arc::new(Float64Array::from(vec![None, None]));
-        let expected_columns = vec![
-            k0_values.clone(),
-            ts_values.clone(),
-            v1_default,
-            Arc::new(UInt64Array::from(vec![200, 200])),
-            Arc::new(UInt8Array::from(vec![
-                OpType::Delete as u8,
-                OpType::Delete as u8,
-            ])),
-        ];
-        let expected_batch = RecordBatch::try_new(output_schema.clone(), expected_columns).unwrap();
-        assert_eq!(expected_batch, result);
-    }
-
-    fn create_test_record_batch() -> RecordBatch {
-        let schema = Arc::new(Schema::new(vec![
-            Field::new("col1", DataType::Int32, false),
-            Field::new("col2", DataType::Utf8, false),
-            Field::new(SEQUENCE_COLUMN_NAME, DataType::UInt64, false),
-            Field::new(OP_TYPE_COLUMN_NAME, DataType::UInt8, false),
-        ]));
-
-        let col1 = Arc::new(Int32Array::from(vec![1, 2, 3]));
-        let col2 = Arc::new(StringArray::from(vec!["a", "b", "c"]));
-        let sequence = Arc::new(UInt64Array::from(vec![100, 101, 102]));
-        let op_type = Arc::new(UInt8Array::from(vec![1, 1, 1]));
-
-        RecordBatch::try_new(schema, vec![col1, col2, sequence, op_type]).unwrap()
-    }
-
-    #[test]
-    fn test_plain_batch_basic_methods() {
-        let record_batch = create_test_record_batch();
-        let plain_batch = PlainBatch::new(record_batch.clone());
-
-        // Test basic properties
-        assert_eq!(plain_batch.num_columns(), 4);
-        assert_eq!(plain_batch.num_rows(), 3);
-        assert!(!plain_batch.is_empty());
-        assert_eq!(plain_batch.columns().len(), 4);
-
-        // Test internal columns access
-        let internal_columns = plain_batch.internal_columns();
-        assert_eq!(internal_columns.len(), PLAIN_FIXED_POS_COLUMN_NUM);
-        assert_eq!(internal_columns[0].len(), 3);
-        assert_eq!(internal_columns[1].len(), 3);
-
-        // Test column access
-        let col1 = plain_batch.column(0);
-        assert_eq!(col1.len(), 3);
-        assert_eq!(
-            col1.as_any().downcast_ref::<Int32Array>().unwrap().value(0),
-            1
-        );
-
-        // Test sequence column index
-        assert_eq!(plain_batch.sequence_column_index(), 2);
-
-        // Test to record batch.
-        assert_eq!(record_batch, *plain_batch.as_record_batch());
-        assert_eq!(record_batch, plain_batch.into_record_batch());
-    }
-
-    #[test]
-    fn test_with_new_columns() {
-        let record_batch = create_test_record_batch();
-        let plain_batch = PlainBatch::new(record_batch);
-
-        // Create new columns
-        let col1 = Arc::new(Int32Array::from(vec![10, 20, 30]));
-        let col2 = Arc::new(StringArray::from(vec!["x", "y", "z"]));
-        let sequence = Arc::new(UInt64Array::from(vec![200, 201, 202]));
-        let op_type = Arc::new(UInt8Array::from(vec![0, 0, 0]));
-
-        let new_batch = plain_batch
-            .with_new_columns(vec![col1, col2, sequence, op_type])
-            .unwrap();
-
-        assert_eq!(new_batch.num_columns(), 4);
-        assert_eq!(new_batch.num_rows(), 3);
-        assert_eq!(
-            new_batch
-                .column(0)
-                .as_any()
-                .downcast_ref::<Int32Array>()
-                .unwrap()
-                .value(0),
-            10
-        );
-        assert_eq!(
-            new_batch
-                .column(1)
-                .as_any()
-                .downcast_ref::<StringArray>()
-                .unwrap()
-                .value(0),
-            "x"
-        );
-    }
-
-    #[test]
-    fn test_filter() {
-        let record_batch = create_test_record_batch();
-        let plain_batch = PlainBatch::new(record_batch);
-
-        // Create a predicate that selects the first and third rows
-        let predicate = BooleanArray::from(vec![true, false, true]);
-
-        let filtered_batch = plain_batch.filter(&predicate).unwrap();
-
-        assert_eq!(filtered_batch.num_rows(), 2);
-        assert_eq!(
-            filtered_batch
-                .column(0)
-                .as_any()
-                .downcast_ref::<Int32Array>()
-                .unwrap()
-                .value(0),
-            1
-        );
-        assert_eq!(
-            filtered_batch
-                .column(0)
-                .as_any()
-                .downcast_ref::<Int32Array>()
-                .unwrap()
-                .value(1),
-            3
-        );
-    }
-}
diff --git a/src/mito2/src/read/projection.rs b/src/mito2/src/read/projection.rs
index 2c000e7bdc..c3d8e4323d 100644
--- a/src/mito2/src/read/projection.rs
+++ b/src/mito2/src/read/projection.rs
@@ -12,381 +12,23 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-//! Utilities for projection operations.
+//! Projection helpers shared by flat projection code.
 
 use std::cmp::Ordering;
-use std::collections::HashMap;
-use std::sync::Arc;
 
-use api::v1::SemanticType;
-use common_error::ext::BoxedError;
-use common_recordbatch::RecordBatch;
-use common_recordbatch::error::ExternalSnafu;
+use common_recordbatch::error::DataTypesSnafu;
 use datatypes::prelude::{ConcreteDataType, DataType};
-use datatypes::schema::{Schema, SchemaRef};
 use datatypes::value::Value;
 use datatypes::vectors::VectorRef;
-use mito_codec::row_converter::{CompositeValues, PrimaryKeyCodec, build_primary_key_codec};
 use snafu::{OptionExt, ResultExt};
 use store_api::metadata::RegionMetadataRef;
 use store_api::storage::ColumnId;
 
 use crate::cache::CacheStrategy;
 use crate::error::{InvalidRequestSnafu, Result};
-use crate::read::Batch;
-use crate::read::flat_projection::FlatProjectionMapper;
 
 /// Only cache vector when its length `<=` this value.
-const MAX_VECTOR_LENGTH_TO_CACHE: usize = 16384;
-
-/// Wrapper enum for different projection mapper implementations.
-pub enum ProjectionMapper {
-    /// Projection mapper for primary key format.
-    PrimaryKey(PrimaryKeyProjectionMapper),
-    /// Projection mapper for flat format.
-    Flat(FlatProjectionMapper),
-}
-
-impl ProjectionMapper {
-    /// Returns a new mapper with projection.
-    pub fn new(
-        metadata: &RegionMetadataRef,
-        projection: impl Iterator<Item = usize> + Clone,
-        flat_format: bool,
-    ) -> Result<Self> {
-        if flat_format {
-            Ok(ProjectionMapper::Flat(FlatProjectionMapper::new(
-                metadata, projection,
-            )?))
-        } else {
-            Ok(ProjectionMapper::PrimaryKey(
-                PrimaryKeyProjectionMapper::new(metadata, projection)?,
-            ))
-        }
-    }
-
-    /// Returns a new mapper with output projection and explicit read columns.
-    pub fn new_with_read_columns(
-        metadata: &RegionMetadataRef,
-        projection: impl Iterator<Item = usize>,
-        flat_format: bool,
-        read_column_ids: Vec<ColumnId>,
-    ) -> Result<Self> {
-        let projection: Vec<_> = projection.collect();
-        if flat_format {
-            Ok(ProjectionMapper::Flat(
-                FlatProjectionMapper::new_with_read_columns(metadata, projection, read_column_ids)?,
-            ))
-        } else {
-            Ok(ProjectionMapper::PrimaryKey(
-                PrimaryKeyProjectionMapper::new_with_read_columns(
-                    metadata,
-                    projection,
-                    read_column_ids,
-                )?,
-            ))
-        }
-    }
-
-    /// Returns a new mapper without projection.
-    pub fn all(metadata: &RegionMetadataRef, flat_format: bool) -> Result<Self> {
-        if flat_format {
-            Ok(ProjectionMapper::Flat(FlatProjectionMapper::all(metadata)?))
-        } else {
-            Ok(ProjectionMapper::PrimaryKey(
-                PrimaryKeyProjectionMapper::all(metadata)?,
-            ))
-        }
-    }
-
-    /// Returns the metadata that created the mapper.
-    pub(crate) fn metadata(&self) -> &RegionMetadataRef {
-        match self {
-            ProjectionMapper::PrimaryKey(m) => m.metadata(),
-            ProjectionMapper::Flat(m) => m.metadata(),
-        }
-    }
-
-    /// Returns true if the projection includes any tag columns.
-    pub(crate) fn has_tags(&self) -> bool {
-        match self {
-            ProjectionMapper::PrimaryKey(m) => m.has_tags(),
-            ProjectionMapper::Flat(_) => false,
-        }
-    }
-
-    /// Returns ids of projected columns that we need to read
-    /// from memtables and SSTs.
-    pub(crate) fn column_ids(&self) -> &[ColumnId] {
-        match self {
-            ProjectionMapper::PrimaryKey(m) => m.column_ids(),
-            ProjectionMapper::Flat(m) => m.column_ids(),
-        }
-    }
-
-    /// Returns the schema of converted [RecordBatch].
-    pub(crate) fn output_schema(&self) -> SchemaRef {
-        match self {
-            ProjectionMapper::PrimaryKey(m) => m.output_schema(),
-            ProjectionMapper::Flat(m) => m.output_schema(),
-        }
-    }
-
-    /// Returns the primary key projection mapper or None if this is not a primary key mapper.
-    pub fn as_primary_key(&self) -> Option<&PrimaryKeyProjectionMapper> {
-        match self {
-            ProjectionMapper::PrimaryKey(m) => Some(m),
-            ProjectionMapper::Flat(_) => None,
-        }
-    }
-
-    /// Returns the flat projection mapper or None if this is not a flat mapper.
-    pub fn as_flat(&self) -> Option<&FlatProjectionMapper> {
-        match self {
-            ProjectionMapper::PrimaryKey(_) => None,
-            ProjectionMapper::Flat(m) => Some(m),
-        }
-    }
-
-    /// Returns an empty [RecordBatch].
-    // TODO(yingwen): This is unused now. Use it after we finishing the flat format.
-    pub fn empty_record_batch(&self) -> RecordBatch {
-        match self {
-            ProjectionMapper::PrimaryKey(m) => m.empty_record_batch(),
-            ProjectionMapper::Flat(m) => m.empty_record_batch(),
-        }
-    }
-}
-
-/// Handles projection and converts a projected [Batch] to a projected [RecordBatch].
-pub struct PrimaryKeyProjectionMapper {
-    /// Metadata of the region.
-    metadata: RegionMetadataRef,
-    /// Maps column in [RecordBatch] to index in [Batch].
-    batch_indices: Vec<BatchIndex>,
-    /// Output record batch contains tags.
-    has_tags: bool,
-    /// Decoder for primary key.
-    codec: Arc<dyn PrimaryKeyCodec>,
-    /// Schema for converted [RecordBatch].
-    output_schema: SchemaRef,
-    /// Ids of columns to read from memtables and SSTs.
-    read_column_ids: Vec<ColumnId>,
-    /// Ids and DataTypes of field columns in the read [Batch].
-    batch_fields: Vec<(ColumnId, ConcreteDataType)>,
-    /// `true` If the original projection is empty.
-    is_empty_projection: bool,
-}
-
-impl PrimaryKeyProjectionMapper {
-    /// Returns a new mapper with projection.
-    /// If `projection` is empty, it outputs [RecordBatch] without any column but only a row count.
-    /// `SELECT COUNT(*) FROM table` is an example that uses an empty projection. DataFusion accepts
-    /// empty `RecordBatch` and only use its row count in this query.
-    pub fn new(
-        metadata: &RegionMetadataRef,
-        projection: impl Iterator<Item = usize>,
-    ) -> Result<PrimaryKeyProjectionMapper> {
-        let projection: Vec<_> = projection.collect();
-        let read_column_ids = read_column_ids_from_projection(metadata, &projection)?;
-        Self::new_with_read_columns(metadata, projection, read_column_ids)
-    }
-
-    /// Returns a new mapper with output projection and explicit read columns.
-    pub fn new_with_read_columns(
-        metadata: &RegionMetadataRef,
-        projection: Vec<usize>,
-        read_column_ids: Vec<ColumnId>,
-    ) -> Result<PrimaryKeyProjectionMapper> {
-        // If the original projection is empty.
-        let is_empty_projection = projection.is_empty();
-
-        let mut column_schemas = Vec::with_capacity(projection.len());
-        for idx in &projection {
-            // For each projection index, we get the column schema for projection
-            column_schemas.push(
-                metadata
-                    .schema
-                    .column_schemas()
-                    .get(*idx)
-                    .with_context(|| InvalidRequestSnafu {
-                        region_id: metadata.region_id,
-                        reason: format!("projection index {} is out of bound", idx),
-                    })?
-                    .clone(),
-            );
-        }
-
-        let codec = build_primary_key_codec(metadata);
-        // If projection is empty, we don't output any column.
-        let output_schema = if is_empty_projection {
-            Arc::new(Schema::new(vec![]))
-        } else {
-            // Safety: Columns come from existing schema.
-            Arc::new(Schema::new(column_schemas))
-        };
-        // Get fields in each read batch.
-        let batch_fields = Batch::projected_fields(metadata, &read_column_ids);
-
-        // Field column id to its index in batch.
-        let field_id_to_index: HashMap<_, _> = batch_fields
-            .iter()
-            .enumerate()
-            .map(|(index, (column_id, _))| (*column_id, index))
-            .collect();
-        // For each projected column, compute its index in batches.
-        let mut batch_indices = Vec::with_capacity(projection.len());
-        let mut has_tags = false;
-        if !is_empty_projection {
-            for idx in &projection {
-                // Safety: idx is valid.
-                let column = &metadata.column_metadatas[*idx];
-                // Get column index in a batch by its semantic type and column id.
-                let batch_index = match column.semantic_type {
-                    SemanticType::Tag => {
-                        // Safety: It is a primary key column.
-                        let index = metadata.primary_key_index(column.column_id).unwrap();
-                        // We need to output a tag.
-                        has_tags = true;
-                        // We always read all primary key so the column always exists and the tag
-                        // index is always valid.
-                        BatchIndex::Tag((index, column.column_id))
-                    }
-                    SemanticType::Timestamp => BatchIndex::Timestamp,
-                    SemanticType::Field => {
-                        let index = *field_id_to_index.get(&column.column_id).context(
-                            InvalidRequestSnafu {
-                                region_id: metadata.region_id,
-                                reason: format!(
-                                    "field column {} is missing in read projection",
-                                    column.column_schema.name
-                                ),
-                            },
-                        )?;
-                        BatchIndex::Field(index)
-                    }
-                };
-                batch_indices.push(batch_index);
-            }
-        }
-
-        Ok(PrimaryKeyProjectionMapper {
-            metadata: metadata.clone(),
-            batch_indices,
-            has_tags,
-            codec,
-            output_schema,
-            read_column_ids,
-            batch_fields,
-            is_empty_projection,
-        })
-    }
-
-    /// Returns a new mapper without projection.
-    pub fn all(metadata: &RegionMetadataRef) -> Result<PrimaryKeyProjectionMapper> {
-        PrimaryKeyProjectionMapper::new(metadata, 0..metadata.column_metadatas.len())
-    }
-
-    /// Returns the metadata that created the mapper.
-    pub(crate) fn metadata(&self) -> &RegionMetadataRef {
-        &self.metadata
-    }
-
-    /// Returns true if the projection includes any tag columns.
-    pub(crate) fn has_tags(&self) -> bool {
-        self.has_tags
-    }
-
-    /// Returns ids of projected columns that we need to read
-    /// from memtables and SSTs.
-    pub(crate) fn column_ids(&self) -> &[ColumnId] {
-        &self.read_column_ids
-    }
-
-    /// Returns ids of fields in [Batch]es the mapper expects to convert.
-    pub(crate) fn batch_fields(&self) -> &[(ColumnId, ConcreteDataType)] {
-        &self.batch_fields
-    }
-
-    /// Returns the schema of converted [RecordBatch].
-    /// This is the schema that the stream will output. This schema may contain
-    /// less columns than [PrimaryKeyProjectionMapper::column_ids()].
-    pub(crate) fn output_schema(&self) -> SchemaRef {
-        self.output_schema.clone()
-    }
-
-    /// Returns an empty [RecordBatch].
-    pub(crate) fn empty_record_batch(&self) -> RecordBatch {
-        RecordBatch::new_empty(self.output_schema.clone())
-    }
-
-    /// Converts a [Batch] to a [RecordBatch].
-    ///
-    /// The batch must match the `projection` using to build the mapper.
-    pub(crate) fn convert(
-        &self,
-        batch: &Batch,
-        cache_strategy: &CacheStrategy,
-    ) -> common_recordbatch::error::Result<RecordBatch> {
-        if self.is_empty_projection {
-            return RecordBatch::new_with_count(self.output_schema.clone(), batch.num_rows());
-        }
-
-        debug_assert_eq!(self.batch_fields.len(), batch.fields().len());
-        debug_assert!(
-            self.batch_fields
-                .iter()
-                .zip(batch.fields())
-                .all(|((id, _), batch_col)| *id == batch_col.column_id)
-        );
-
-        // Skips decoding pk if we don't need to output it.
-        let pk_values = if self.has_tags {
-            match batch.pk_values() {
-                Some(v) => v.clone(),
-                None => self
-                    .codec
-                    .decode(batch.primary_key())
-                    .map_err(BoxedError::new)
-                    .context(ExternalSnafu)?,
-            }
-        } else {
-            CompositeValues::Dense(vec![])
-        };
-
-        let mut columns = Vec::with_capacity(self.output_schema.num_columns());
-        let num_rows = batch.num_rows();
-        for (index, column_schema) in self
-            .batch_indices
-            .iter()
-            .zip(self.output_schema.column_schemas())
-        {
-            match index {
-                BatchIndex::Tag((idx, column_id)) => {
-                    let value = match &pk_values {
-                        CompositeValues::Dense(v) => &v[*idx].1,
-                        CompositeValues::Sparse(v) => v.get_or_null(*column_id),
-                    };
-                    let vector = repeated_vector_with_cache(
-                        &column_schema.data_type,
-                        value,
-                        num_rows,
-                        cache_strategy,
-                    )?;
-                    columns.push(vector);
-                }
-                BatchIndex::Timestamp => {
-                    columns.push(batch.timestamps().clone());
-                }
-                BatchIndex::Field(idx) => {
-                    columns.push(batch.fields()[*idx].data.clone());
-                }
-            }
-        }
-
-        RecordBatch::new(self.output_schema.clone(), columns)
-    }
-}
+pub(crate) const MAX_VECTOR_LENGTH_TO_CACHE: usize = 16384;
 
 pub(crate) fn read_column_ids_from_projection(
     metadata: &RegionMetadataRef,
@@ -411,27 +53,14 @@ pub(crate) fn read_column_ids_from_projection(
     Ok(column_ids)
 }
 
-/// Index of a vector in a [Batch].
-#[derive(Debug, Clone, Copy)]
-enum BatchIndex {
-    /// Index in primary keys.
-    Tag((usize, ColumnId)),
-    /// The time index column.
-    Timestamp,
-    /// Index in fields.
-    Field(usize),
-}
-
 /// Gets a vector with repeated values from specific cache or creates a new one.
-fn repeated_vector_with_cache(
+pub(crate) fn repeated_vector_with_cache(
     data_type: &ConcreteDataType,
     value: &Value,
     num_rows: usize,
     cache_strategy: &CacheStrategy,
 ) -> common_recordbatch::error::Result<VectorRef> {
     if let Some(vector) = cache_strategy.get_repeated_vector(data_type, value) {
-        // Tries to get the vector from cache manager. If the vector doesn't
-        // have enough length, creates a new one.
         match vector.len().cmp(&num_rows) {
             Ordering::Less => (),
             Ordering::Equal => return Ok(vector),
@@ -439,9 +68,7 @@ fn repeated_vector_with_cache(
         }
     }
 
-    // Creates a new one.
     let vector = new_repeated_vector(data_type, value, num_rows)?;
-    // Updates cache.
     if vector.len() <= MAX_VECTOR_LENGTH_TO_CACHE {
         cache_strategy.put_repeated_vector(value.clone(), vector.clone());
     }
@@ -450,7 +77,7 @@ fn repeated_vector_with_cache(
 }
 
 /// Returns a vector with repeated values.
-fn new_repeated_vector(
+pub(crate) fn new_repeated_vector(
     data_type: &ConcreteDataType,
     value: &Value,
     num_rows: usize,
@@ -458,9 +85,7 @@ fn new_repeated_vector(
     let mut mutable_vector = data_type.create_mutable_vector(1);
     mutable_vector
         .try_push_value_ref(&value.as_value_ref())
-        .map_err(BoxedError::new)
-        .context(ExternalSnafu)?;
-    // This requires an additional allocation.
+        .context(DataTypesSnafu)?;
     let base_vector = mutable_vector.to_vector();
     Ok(base_vector.replicate(&[num_rows]))
 }
@@ -470,6 +95,7 @@ mod tests {
     use std::sync::Arc;
 
     use api::v1::OpType;
+    use common_recordbatch::RecordBatch;
     use datatypes::arrow::array::{Int64Array, TimestampMillisecondArray, UInt8Array, UInt64Array};
     use datatypes::arrow::datatypes::Field;
     use datatypes::arrow::util::pretty;
@@ -481,53 +107,8 @@ mod tests {
     };
 
     use super::*;
-    use crate::cache::CacheManager;
-    use crate::read::BatchBuilder;
-
-    fn new_batch(
-        ts_start: i64,
-        tags: &[i64],
-        fields: &[(ColumnId, i64)],
-        num_rows: usize,
-    ) -> Batch {
-        let converter = DensePrimaryKeyCodec::with_fields(
-            (0..tags.len())
-                .map(|idx| {
-                    (
-                        idx as u32,
-                        SortField::new(ConcreteDataType::int64_datatype()),
-                    )
-                })
-                .collect(),
-        );
-        let primary_key = converter
-            .encode(tags.iter().map(|v| ValueRef::Int64(*v)))
-            .unwrap();
-
-        let mut builder = BatchBuilder::new(primary_key);
-        builder
-            .timestamps_array(Arc::new(TimestampMillisecondArray::from_iter_values(
-                (0..num_rows).map(|i| ts_start + i as i64 * 1000),
-            )))
-            .unwrap()
-            .sequences_array(Arc::new(UInt64Array::from_iter_values(0..num_rows as u64)))
-            .unwrap()
-            .op_types_array(Arc::new(UInt8Array::from_iter_values(
-                (0..num_rows).map(|_| OpType::Put as u8),
-            )))
-            .unwrap();
-        for (column_id, field) in fields {
-            builder
-                .push_field_array(
-                    *column_id,
-                    Arc::new(Int64Array::from_iter_values(std::iter::repeat_n(
-                        *field, num_rows,
-                    ))),
-                )
-                .unwrap();
-        }
-        builder.build().unwrap()
-    }
+    use crate::read::flat_projection::FlatProjectionMapper;
+    use crate::read::read_columns::ReadColumns;
 
     fn print_record_batch(record_batch: RecordBatch) -> String {
         pretty::pretty_format_batches(&[record_batch.into_df_record_batch()])
@@ -535,166 +116,6 @@ mod tests {
             .to_string()
     }
 
-    #[test]
-    fn test_projection_mapper_all() {
-        let metadata = Arc::new(
-            TestRegionMetadataBuilder::default()
-                .num_tags(2)
-                .num_fields(2)
-                .build(),
-        );
-        // Create the enum wrapper with default format (primary key)
-        let mapper = ProjectionMapper::all(&metadata, false).unwrap();
-        assert_eq!([0, 1, 2, 3, 4], mapper.column_ids());
-        assert_eq!(
-            [
-                (3, ConcreteDataType::int64_datatype()),
-                (4, ConcreteDataType::int64_datatype())
-            ],
-            mapper.as_primary_key().unwrap().batch_fields()
-        );
-
-        // With vector cache.
-        let cache = CacheManager::builder().vector_cache_size(1024).build();
-        let cache = CacheStrategy::EnableAll(Arc::new(cache));
-        let batch = new_batch(0, &[1, 2], &[(3, 3), (4, 4)], 3);
-        let record_batch = mapper
-            .as_primary_key()
-            .unwrap()
-            .convert(&batch, &cache)
-            .unwrap();
-        let expect = "\
-+---------------------+----+----+----+----+
-| ts                  | k0 | k1 | v0 | v1 |
-+---------------------+----+----+----+----+
-| 1970-01-01T00:00:00 | 1  | 2  | 3  | 4  |
-| 1970-01-01T00:00:01 | 1  | 2  | 3  | 4  |
-| 1970-01-01T00:00:02 | 1  | 2  | 3  | 4  |
-+---------------------+----+----+----+----+";
-        assert_eq!(expect, print_record_batch(record_batch));
-
-        assert!(
-            cache
-                .get_repeated_vector(&ConcreteDataType::int64_datatype(), &Value::Int64(1))
-                .is_some()
-        );
-        assert!(
-            cache
-                .get_repeated_vector(&ConcreteDataType::int64_datatype(), &Value::Int64(2))
-                .is_some()
-        );
-        assert!(
-            cache
-                .get_repeated_vector(&ConcreteDataType::int64_datatype(), &Value::Int64(3))
-                .is_none()
-        );
-        let record_batch = mapper
-            .as_primary_key()
-            .unwrap()
-            .convert(&batch, &cache)
-            .unwrap();
-        assert_eq!(expect, print_record_batch(record_batch));
-    }
-
-    #[test]
-    fn test_projection_mapper_with_projection() {
-        let metadata = Arc::new(
-            TestRegionMetadataBuilder::default()
-                .num_tags(2)
-                .num_fields(2)
-                .build(),
-        );
-        // Columns v1, k0
-        let mapper = ProjectionMapper::new(&metadata, [4, 1].into_iter(), false).unwrap();
-        assert_eq!([4, 1], mapper.column_ids());
-        assert_eq!(
-            [(4, ConcreteDataType::int64_datatype())],
-            mapper.as_primary_key().unwrap().batch_fields()
-        );
-
-        let batch = new_batch(0, &[1, 2], &[(4, 4)], 3);
-        let cache = CacheManager::builder().vector_cache_size(1024).build();
-        let cache = CacheStrategy::EnableAll(Arc::new(cache));
-        let record_batch = mapper
-            .as_primary_key()
-            .unwrap()
-            .convert(&batch, &cache)
-            .unwrap();
-        let expect = "\
-+----+----+
-| v1 | k0 |
-+----+----+
-| 4  | 1  |
-| 4  | 1  |
-| 4  | 1  |
-+----+----+";
-        assert_eq!(expect, print_record_batch(record_batch));
-    }
-
-    #[test]
-    fn test_projection_mapper_read_superset() {
-        let metadata = Arc::new(
-            TestRegionMetadataBuilder::default()
-                .num_tags(2)
-                .num_fields(2)
-                .build(),
-        );
-        // Output columns v1, k0. Read also includes v0.
-        let mapper = ProjectionMapper::new_with_read_columns(
-            &metadata,
-            [4, 1].into_iter(),
-            false,
-            vec![4, 1, 3],
-        )
-        .unwrap();
-        assert_eq!([4, 1, 3], mapper.column_ids());
-
-        let batch = new_batch(0, &[1, 2], &[(3, 3), (4, 4)], 3);
-        let cache = CacheManager::builder().vector_cache_size(1024).build();
-        let cache = CacheStrategy::EnableAll(Arc::new(cache));
-        let record_batch = mapper
-            .as_primary_key()
-            .unwrap()
-            .convert(&batch, &cache)
-            .unwrap();
-        let expect = "\
-+----+----+
-| v1 | k0 |
-+----+----+
-| 4  | 1  |
-| 4  | 1  |
-| 4  | 1  |
-+----+----+";
-        assert_eq!(expect, print_record_batch(record_batch));
-    }
-
-    #[test]
-    fn test_projection_mapper_empty_projection() {
-        let metadata = Arc::new(
-            TestRegionMetadataBuilder::default()
-                .num_tags(2)
-                .num_fields(2)
-                .build(),
-        );
-        // Empty projection
-        let mapper = ProjectionMapper::new(&metadata, [].into_iter(), false).unwrap();
-        assert_eq!([0], mapper.column_ids()); // Should still read the time index column
-        assert!(mapper.output_schema().is_empty());
-        let pk_mapper = mapper.as_primary_key().unwrap();
-        assert!(pk_mapper.batch_fields().is_empty());
-        assert!(!pk_mapper.has_tags);
-        assert!(pk_mapper.batch_indices.is_empty());
-        assert!(pk_mapper.is_empty_projection);
-
-        let batch = new_batch(0, &[1, 2], &[], 3);
-        let cache = CacheManager::builder().vector_cache_size(1024).build();
-        let cache = CacheStrategy::EnableAll(Arc::new(cache));
-        let record_batch = pk_mapper.convert(&batch, &cache).unwrap();
-        assert_eq!(3, record_batch.num_rows());
-        assert_eq!(0, record_batch.num_columns());
-        assert!(record_batch.schema.is_empty());
-    }
-
     fn new_flat_batch(
         ts_start: Option<i64>,
         idx_tags: &[(usize, i64)],
@@ -704,9 +125,6 @@ mod tests {
         let mut columns = Vec::with_capacity(1 + idx_tags.len() + idx_fields.len() + 3);
         let mut fields = Vec::with_capacity(1 + idx_tags.len() + idx_fields.len() + 3);
 
-        // Flat format: primary key columns, field columns, time index, __primary_key, __sequence, __op_type
-
-        // Primary key columns first
         for (i, tag) in idx_tags {
             let array = Arc::new(Int64Array::from_iter_values(std::iter::repeat_n(
                 *tag, num_rows,
@@ -719,7 +137,6 @@ mod tests {
             ));
         }
 
-        // Field columns
         for (i, field) in idx_fields {
             let array = Arc::new(Int64Array::from_iter_values(std::iter::repeat_n(
                 *field, num_rows,
@@ -732,7 +149,6 @@ mod tests {
             ));
         }
 
-        // Time index
         if let Some(ts_start) = ts_start {
             let timestamps = Arc::new(TimestampMillisecondArray::from_iter_values(
                 (0..num_rows).map(|i| ts_start + i as i64 * 1000),
@@ -748,8 +164,6 @@ mod tests {
             ));
         }
 
-        // __primary_key column (encoded primary key as dictionary)
-        // Create encoded primary key
         let converter = DensePrimaryKeyCodec::with_fields(
             (0..idx_tags.len())
                 .map(|idx| {
@@ -764,7 +178,6 @@ mod tests {
             .encode(idx_tags.iter().map(|(_, v)| ValueRef::Int64(*v)))
             .unwrap();
 
-        // Create dictionary array for the encoded primary key
         let pk_values: Vec<&[u8]> = std::iter::repeat_n(encoded_pk.as_slice(), num_rows).collect();
         let keys = datatypes::arrow::array::UInt32Array::from_iter(0..num_rows as u32);
         let values = Arc::new(datatypes::arrow::array::BinaryArray::from_vec(pk_values));
@@ -778,7 +191,6 @@ mod tests {
             false,
         ));
 
-        // __sequence column
         columns.push(Arc::new(UInt64Array::from_iter_values(0..num_rows as u64)) as _);
         fields.push(Field::new(
             SEQUENCE_COLUMN_NAME,
@@ -786,7 +198,6 @@ mod tests {
             false,
         ));
 
-        // __op_type column
         columns.push(Arc::new(UInt8Array::from_iter_values(
             (0..num_rows).map(|_| OpType::Put as u8),
         )) as _);
@@ -809,8 +220,12 @@ mod tests {
                 .num_fields(2)
                 .build(),
         );
-        let mapper = ProjectionMapper::all(&metadata, true).unwrap();
-        assert_eq!([0, 1, 2, 3, 4], mapper.column_ids());
+        let cache = CacheStrategy::Disabled;
+        let mapper = FlatProjectionMapper::all(&metadata).unwrap();
+        assert_eq!(
+            &[0, 1, 2, 3, 4],
+            mapper.read_columns().column_ids().as_slice()
+        );
         assert_eq!(
             [
                 (1, ConcreteDataType::int64_datatype()),
@@ -819,11 +234,11 @@ mod tests {
                 (4, ConcreteDataType::int64_datatype()),
                 (0, ConcreteDataType::timestamp_millisecond_datatype())
             ],
-            mapper.as_flat().unwrap().batch_schema()
+            mapper.batch_schema()
         );
 
         let batch = new_flat_batch(Some(0), &[(1, 1), (2, 2)], &[(3, 3), (4, 4)], 3);
-        let record_batch = mapper.as_flat().unwrap().convert(&batch).unwrap();
+        let record_batch = mapper.convert(&batch, &cache).unwrap();
         let expect = "\
 +---------------------+----+----+----+----+
 | ts                  | k0 | k1 | v0 | v1 |
@@ -843,20 +258,20 @@ mod tests {
                 .num_fields(2)
                 .build(),
         );
-        // Columns v1, k0
-        let mapper = ProjectionMapper::new(&metadata, [4, 1].into_iter(), true).unwrap();
-        assert_eq!([4, 1], mapper.column_ids());
+        let cache = CacheStrategy::Disabled;
+        let mapper = FlatProjectionMapper::new(&metadata, [4, 1]).unwrap();
+        assert_eq!(&[4, 1], mapper.read_columns().column_ids().as_slice());
         assert_eq!(
             [
                 (1, ConcreteDataType::int64_datatype()),
                 (4, ConcreteDataType::int64_datatype()),
                 (0, ConcreteDataType::timestamp_millisecond_datatype())
             ],
-            mapper.as_flat().unwrap().batch_schema()
+            mapper.batch_schema()
         );
 
         let batch = new_flat_batch(None, &[(1, 1)], &[(4, 4)], 3);
-        let record_batch = mapper.as_flat().unwrap().convert(&batch).unwrap();
+        let record_batch = mapper.convert(&batch, &cache).unwrap();
         let expect = "\
 +----+----+
 | v1 | k0 |
@@ -876,18 +291,17 @@ mod tests {
                 .num_fields(2)
                 .build(),
         );
-        // Output columns v1, k0. Read also includes v0.
-        let mapper = ProjectionMapper::new_with_read_columns(
+        let cache = CacheStrategy::Disabled;
+        let mapper = FlatProjectionMapper::new_with_read_columns(
             &metadata,
-            [4, 1].into_iter(),
-            true,
-            vec![4, 1, 3],
+            vec![4, 1],
+            ReadColumns::from_deduped_column_ids([4, 1, 3]),
         )
         .unwrap();
-        assert_eq!([4, 1, 3], mapper.column_ids());
+        assert_eq!(&[4, 1, 3], mapper.read_columns().column_ids().as_slice());
 
         let batch = new_flat_batch(None, &[(1, 1)], &[(3, 3), (4, 4)], 3);
-        let record_batch = mapper.as_flat().unwrap().convert(&batch).unwrap();
+        let record_batch = mapper.convert(&batch, &cache).unwrap();
         let expect = "\
 +----+----+
 | v1 | k0 |
@@ -907,18 +321,17 @@ mod tests {
                 .num_fields(2)
                 .build(),
         );
-        // Empty projection
-        let mapper = ProjectionMapper::new(&metadata, [].into_iter(), true).unwrap();
-        assert_eq!([0], mapper.column_ids()); // Should still read the time index column
+        let cache = CacheStrategy::Disabled;
+        let mapper = FlatProjectionMapper::new(&metadata, []).unwrap();
+        assert_eq!(&[0], mapper.read_columns().column_ids().as_slice());
         assert!(mapper.output_schema().is_empty());
-        let flat_mapper = mapper.as_flat().unwrap();
         assert_eq!(
             [(0, ConcreteDataType::timestamp_millisecond_datatype())],
-            flat_mapper.batch_schema()
+            mapper.batch_schema()
         );
 
         let batch = new_flat_batch(Some(0), &[], &[], 3);
-        let record_batch = flat_mapper.convert(&batch).unwrap();
+        let record_batch = mapper.convert(&batch, &cache).unwrap();
         assert_eq!(3, record_batch.num_rows());
         assert_eq!(0, record_batch.num_columns());
         assert!(record_batch.schema.is_empty());
diff --git a/src/mito2/src/read/prune.rs b/src/mito2/src/read/prune.rs
index 29ded3d49a..623f7bbc03 100644
--- a/src/mito2/src/read/prune.rs
+++ b/src/mito2/src/read/prune.rs
@@ -24,123 +24,11 @@ use snafu::ResultExt;
 
 use crate::error::{RecordBatchSnafu, Result};
 use crate::memtable::BoxedBatchIterator;
-use crate::read::last_row::{FlatRowGroupLastRowCachedReader, RowGroupLastRowCachedReader};
-use crate::read::{Batch, BatchReader};
+use crate::read::Batch;
+use crate::read::last_row::FlatRowGroupLastRowCachedReader;
 use crate::sst::file::FileTimeRange;
 use crate::sst::parquet::file_range::FileRangeContextRef;
-use crate::sst::parquet::reader::{FlatRowGroupReader, ReaderMetrics, RowGroupReader};
-
-pub enum Source {
-    RowGroup(RowGroupReader),
-    LastRow(RowGroupLastRowCachedReader),
-}
-
-impl Source {
-    async fn next_batch(&mut self) -> Result<Option<Batch>> {
-        match self {
-            Source::RowGroup(r) => r.next_batch().await,
-            Source::LastRow(r) => r.next_batch().await,
-        }
-    }
-}
-
-pub struct PruneReader {
-    /// Context for file ranges.
-    context: FileRangeContextRef,
-    source: Source,
-    metrics: ReaderMetrics,
-    /// Whether to skip field filters for this row group.
-    skip_fields: bool,
-}
-
-impl PruneReader {
-    pub(crate) fn new_with_row_group_reader(
-        ctx: FileRangeContextRef,
-        reader: RowGroupReader,
-        skip_fields: bool,
-    ) -> Self {
-        Self {
-            context: ctx,
-            source: Source::RowGroup(reader),
-            metrics: Default::default(),
-            skip_fields,
-        }
-    }
-
-    pub(crate) fn new_with_last_row_reader(
-        ctx: FileRangeContextRef,
-        reader: RowGroupLastRowCachedReader,
-        skip_fields: bool,
-    ) -> Self {
-        Self {
-            context: ctx,
-            source: Source::LastRow(reader),
-            metrics: Default::default(),
-            skip_fields,
-        }
-    }
-
-    pub(crate) fn reset_source(&mut self, source: Source, skip_fields: bool) {
-        self.source = source;
-        self.skip_fields = skip_fields;
-    }
-
-    /// Merge metrics with the inner reader and return the merged metrics.
-    pub(crate) fn metrics(&self) -> ReaderMetrics {
-        let mut metrics = self.metrics.clone();
-        match &self.source {
-            Source::RowGroup(r) => {
-                metrics.merge_from(r.metrics());
-            }
-            Source::LastRow(r) => {
-                if let Some(inner_metrics) = r.metrics() {
-                    metrics.merge_from(inner_metrics);
-                }
-            }
-        }
-
-        metrics
-    }
-
-    pub(crate) async fn next_batch(&mut self) -> Result<Option<Batch>> {
-        while let Some(b) = self.source.next_batch().await? {
-            match self.prune(b)? {
-                Some(b) => {
-                    return Ok(Some(b));
-                }
-                None => {
-                    continue;
-                }
-            }
-        }
-        Ok(None)
-    }
-
-    /// Prunes batches by the pushed down predicate.
-    fn prune(&mut self, batch: Batch) -> Result<Option<Batch>> {
-        // fast path
-        if self.context.filters().is_empty() && !self.context.has_partition_filter() {
-            return Ok(Some(batch));
-        }
-
-        let num_rows_before_filter = batch.num_rows();
-        let Some(batch_filtered) = self.context.precise_filter(batch, self.skip_fields)? else {
-            // the entire batch is filtered out
-            self.metrics.filter_metrics.rows_precise_filtered += num_rows_before_filter;
-            return Ok(None);
-        };
-
-        // update metric
-        let filtered_rows = num_rows_before_filter - batch_filtered.num_rows();
-        self.metrics.filter_metrics.rows_precise_filtered += filtered_rows;
-
-        if !batch_filtered.is_empty() {
-            Ok(Some(batch_filtered))
-        } else {
-            Ok(None)
-        }
-    }
-}
+use crate::sst::parquet::reader::{FlatRowGroupReader, ReaderMetrics};
 
 /// An iterator that prunes batches by time range.
 pub(crate) struct PruneTimeIterator {
@@ -252,10 +140,10 @@ pub enum FlatSource {
 }
 
 impl FlatSource {
-    fn next_batch(&mut self) -> Result<Option<RecordBatch>> {
+    async fn next_batch(&mut self) -> Result<Option<RecordBatch>> {
         match self {
-            FlatSource::RowGroup(r) => r.next_batch(),
-            FlatSource::LastRow(r) => r.next_batch(),
+            FlatSource::RowGroup(r) => r.next_batch().await,
+            FlatSource::LastRow(r) => r.next_batch().await,
         }
     }
 }
@@ -302,13 +190,16 @@ impl FlatPruneReader {
         self.metrics.clone()
     }
 
-    pub(crate) fn next_batch(&mut self) -> Result<Option<RecordBatch>> {
-        while let Some(record_batch) = {
+    pub(crate) async fn next_batch(&mut self) -> Result<Option<RecordBatch>> {
+        loop {
             let start = std::time::Instant::now();
-            let batch = self.source.next_batch()?;
+            let batch = self.source.next_batch().await?;
             self.metrics.scan_cost += start.elapsed();
-            batch
-        } {
+
+            let Some(record_batch) = batch else {
+                return Ok(None);
+            };
+
             // Update metrics for the received batch
             self.metrics.num_rows += record_batch.num_rows();
             self.metrics.num_batches += 1;
@@ -322,8 +213,6 @@ impl FlatPruneReader {
                 }
             }
         }
-
-        Ok(None)
     }
 
     /// Prunes batches by the pushed down predicate and returns RecordBatch.
diff --git a/src/mito2/src/read/pruner.rs b/src/mito2/src/read/pruner.rs
index b7fbc36cc0..320144484d 100644
--- a/src/mito2/src/read/pruner.rs
+++ b/src/mito2/src/read/pruner.rs
@@ -32,7 +32,7 @@ use crate::metrics::PRUNER_ACTIVE_BUILDERS;
 use crate::read::range::{FileRangeBuilder, RowGroupIndex};
 use crate::read::scan_region::StreamContext;
 use crate::read::scan_util::{FileScanMetrics, PartitionMetrics};
-use crate::sst::parquet::file_range::FileRange;
+use crate::sst::parquet::file_range::{FileRange, PreFilterMode};
 use crate::sst::parquet::reader::ReaderMetrics;
 
 /// Number of files to pre-fetch ahead of the current position.
@@ -43,6 +43,8 @@ pub struct PartitionPruner {
     pruner: Arc<Pruner>,
     /// Files to prune, in the order to scan.
     file_indices: Vec<usize>,
+    /// Per-file pre-filter mode lookup indexed by file_index.
+    pre_filter_modes: Vec<PreFilterMode>,
     /// Current position for tracking pre-fetch progress.
     current_position: AtomicUsize,
 }
@@ -50,12 +52,15 @@ pub struct PartitionPruner {
 impl PartitionPruner {
     /// Creates a new `PartitionPruner` for the given partition ranges.
     pub fn new(pruner: Arc<Pruner>, partition_ranges: &[PartitionRange]) -> Self {
-        let mut file_indices = Vec::with_capacity(pruner.inner.stream_ctx.input.num_files());
+        let num_files = pruner.inner.stream_ctx.input.num_files();
+        let mut file_indices = Vec::with_capacity(num_files);
+        let mut pre_filter_modes = vec![PreFilterMode::SkipFields; num_files];
         let mut dedup_set = HashSet::with_capacity(pruner.inner.stream_ctx.input.num_files());
 
         let num_memtables = pruner.inner.stream_ctx.input.num_memtables();
         for part_range in partition_ranges {
             let range_meta = &pruner.inner.stream_ctx.ranges[part_range.identifier];
+            let pre_filter_mode = pruner.inner.stream_ctx.range_pre_filter_mode(part_range);
             for row_group_index in &range_meta.row_group_indices {
                 if pruner
                     .inner
@@ -67,6 +72,7 @@ impl PartitionPruner {
                         continue;
                     } else {
                         file_indices.push(file_index);
+                        pre_filter_modes[file_index] = pre_filter_mode;
                         dedup_set.insert(file_index);
                     }
                 }
@@ -76,6 +82,7 @@ impl PartitionPruner {
         Self {
             pruner,
             file_indices,
+            pre_filter_modes,
             current_position: AtomicUsize::new(0),
         }
     }
@@ -91,11 +98,12 @@ impl PartitionPruner {
         reader_metrics: &mut ReaderMetrics,
     ) -> Result<SmallVec<[FileRange; 2]>> {
         let file_index = index.index - self.pruner.inner.stream_ctx.input.num_memtables();
+        let pre_filter_mode = self.pre_filter_mode(file_index);
 
         // Delegate to underlying Pruner
         let ranges = self
             .pruner
-            .build_file_ranges(index, partition_metrics, reader_metrics)
+            .build_file_ranges(index, pre_filter_mode, partition_metrics, reader_metrics)
             .await?;
 
         // Find position and trigger pre-fetch for upcoming files
@@ -115,10 +123,22 @@ impl PartitionPruner {
         let end = (start + PREFETCH_COUNT).min(self.file_indices.len());
 
         for i in start..end {
-            self.pruner
-                .get_file_builder_background(self.file_indices[i], Some(partition_metrics.clone()));
+            let file_index = self.file_indices[i];
+            let pre_filter_mode = self.pre_filter_mode(file_index);
+            self.pruner.get_file_builder_background(
+                file_index,
+                pre_filter_mode,
+                Some(partition_metrics.clone()),
+            );
         }
     }
+
+    fn pre_filter_mode(&self, file_index: usize) -> PreFilterMode {
+        self.pre_filter_modes
+            .get(file_index)
+            .copied()
+            .unwrap_or(PreFilterMode::SkipFields)
+    }
 }
 
 /// A pruner that prunes files for all partitions of a scanner.
@@ -152,6 +172,8 @@ struct FileBuilderEntry {
 struct PruneRequest {
     /// Index of the file in ScanInput.files.
     file_index: usize,
+    /// Pre-filter mode to use for the file.
+    pre_filter_mode: PreFilterMode,
     /// Oneshot channel to send back the result.
     response_tx: Option<oneshot::Sender<Result<Arc<FileRangeBuilder>>>>,
     /// Partition metrics for merging reader metrics.
@@ -174,7 +196,6 @@ impl Pruner {
                 })
             })
             .collect();
-
         // Create channels and collect senders
         let mut worker_senders = Vec::with_capacity(num_workers);
         let mut receivers = Vec::with_capacity(num_workers);
@@ -230,6 +251,7 @@ impl Pruner {
     pub async fn build_file_ranges(
         &self,
         index: RowGroupIndex,
+        pre_filter_mode: PreFilterMode,
         partition_metrics: &PartitionMetrics,
         reader_metrics: &mut ReaderMetrics,
     ) -> Result<SmallVec<[FileRange; 2]>> {
@@ -237,7 +259,12 @@ impl Pruner {
 
         // Get builder (from cache or by pruning)
         let builder = self
-            .get_file_builder(file_index, partition_metrics, reader_metrics)
+            .get_file_builder(
+                file_index,
+                pre_filter_mode,
+                partition_metrics,
+                reader_metrics,
+            )
             .await?;
 
         // Build ranges
@@ -254,6 +281,7 @@ impl Pruner {
     async fn get_file_builder(
         &self,
         file_index: usize,
+        pre_filter_mode: PreFilterMode,
         partition_metrics: &PartitionMetrics,
         reader_metrics: &mut ReaderMetrics,
     ) -> Result<Arc<FileRangeBuilder>> {
@@ -275,6 +303,7 @@ impl Pruner {
         let (response_tx, response_rx) = oneshot::channel();
         let request = PruneRequest {
             file_index,
+            pre_filter_mode,
             response_tx: Some(response_tx),
             partition_metrics: Some(partition_metrics.clone()),
         };
@@ -282,7 +311,8 @@ impl Pruner {
         let result = if self.worker_senders[worker_idx].send(request).await.is_err() {
             common_telemetry::warn!("Worker channel closed, falling back to direct pruning");
             // Worker channel closed, falls back to direct pruning
-            self.prune_file_directly(file_index, reader_metrics).await
+            self.prune_file_directly(file_index, pre_filter_mode, reader_metrics)
+                .await
         } else {
             // Waits for response
             match response_rx.await {
@@ -292,7 +322,8 @@ impl Pruner {
                         "Response channel closed, falling back to direct pruning"
                     );
                     // Channel closed, falls back to direct pruning
-                    self.prune_file_directly(file_index, reader_metrics).await
+                    self.prune_file_directly(file_index, pre_filter_mode, reader_metrics)
+                        .await
                 }
             }
         };
@@ -304,6 +335,7 @@ impl Pruner {
     pub fn get_file_builder_background(
         &self,
         file_index: usize,
+        pre_filter_mode: PreFilterMode,
         partition_metrics: Option<PartitionMetrics>,
     ) {
         // Fast path: checks cache
@@ -320,6 +352,7 @@ impl Pruner {
 
         let request = PruneRequest {
             file_index,
+            pre_filter_mode,
             response_tx: None,
             partition_metrics,
         };
@@ -338,6 +371,7 @@ impl Pruner {
     async fn prune_file_directly(
         &self,
         file_index: usize,
+        pre_filter_mode: PreFilterMode,
         reader_metrics: &mut ReaderMetrics,
     ) -> Result<Arc<FileRangeBuilder>> {
         let file = &self.inner.stream_ctx.input.files[file_index];
@@ -345,7 +379,7 @@ impl Pruner {
             .inner
             .stream_ctx
             .input
-            .prune_file(file, reader_metrics)
+            .prune_file(file, pre_filter_mode, reader_metrics)
             .await?;
 
         let arc_builder = Arc::new(builder);
@@ -391,6 +425,7 @@ impl Pruner {
         while let Some(request) = rx.recv().await {
             let PruneRequest {
                 file_index,
+                pre_filter_mode,
                 response_tx,
                 partition_metrics,
             } = request;
@@ -398,7 +433,6 @@ impl Pruner {
             // Check if already cached or in-progress
             {
                 let entry = inner.file_entries[file_index].lock().unwrap();
-
                 if let Some(builder) = &entry.builder {
                     // Cache hit - send immediately
                     if let Some(response_tx) = response_tx {
@@ -414,7 +448,11 @@ impl Pruner {
             let file = &inner.stream_ctx.input.files[file_index];
             pruned_files.push(file.file_id().file_id());
             let mut metrics = ReaderMetrics::default();
-            let result = inner.stream_ctx.input.prune_file(file, &mut metrics).await;
+            let result = inner
+                .stream_ctx
+                .input
+                .prune_file(file, pre_filter_mode, &mut metrics)
+                .await;
 
             // Update state and notify waiters
             let mut entry = inner.file_entries[file_index].lock().unwrap();
diff --git a/src/mito2/src/read/range_cache.rs b/src/mito2/src/read/range_cache.rs
new file mode 100644
index 0000000000..208715f57d
--- /dev/null
+++ b/src/mito2/src/read/range_cache.rs
@@ -0,0 +1,1034 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Utilities for the partition range scan result cache.
+
+use std::mem;
+use std::sync::Arc;
+
+use async_stream::try_stream;
+use common_telemetry::warn;
+use common_time::range::TimestampRange;
+use datatypes::arrow::compute::concat_batches;
+use datatypes::arrow::record_batch::RecordBatch;
+use datatypes::prelude::ConcreteDataType;
+use futures::TryStreamExt;
+use snafu::ResultExt;
+use store_api::region_engine::PartitionRange;
+use store_api::storage::{FileId, RegionId, TimeSeriesRowSelector};
+use tokio::sync::{mpsc, oneshot};
+
+use crate::cache::CacheStrategy;
+use crate::error::{ComputeArrowSnafu, Result};
+use crate::read::BoxedRecordBatchStream;
+use crate::read::read_columns::ReadColumns;
+use crate::read::scan_region::StreamContext;
+use crate::read::scan_util::PartitionMetrics;
+use crate::region::options::MergeMode;
+use crate::sst::file::FileTimeRange;
+use crate::sst::parquet::DEFAULT_READ_BATCH_SIZE;
+
+const RANGE_CACHE_COMPACT_THRESHOLD_BYTES: usize = 8 * 1024 * 1024;
+
+/// Fingerprint of the scan request fields that affect partition range cache reuse.
+///
+/// It records a normalized view of the projected columns and filters, plus
+/// scan options that can change the returned rows. Schema-dependent metadata
+/// and the partition expression version are included so cached results are not
+/// reused across incompatible schema or partitioning changes.
+#[derive(Debug, Clone, PartialEq, Eq, Hash)]
+pub(crate) struct ScanRequestFingerprint {
+    /// Projection and filters without the time index and partition exprs.
+    inner: Arc<SharedScanRequestFingerprint>,
+    /// Filters with the time index column.
+    time_filters: Option<Arc<Vec<String>>>,
+    series_row_selector: Option<TimeSeriesRowSelector>,
+    append_mode: bool,
+    filter_deleted: bool,
+    merge_mode: MergeMode,
+    /// We keep the partition expr version to ensure we won't reuse the fingerprint after we change the partition expr.
+    /// We store the version instead of the whole partition expr or partition expr filters.
+    partition_expr_version: u64,
+}
+
+#[derive(Debug)]
+pub(crate) struct ScanRequestFingerprintBuilder {
+    pub(crate) read_columns: ReadColumns,
+    pub(crate) read_column_types: Vec<Option<ConcreteDataType>>,
+    pub(crate) filters: Vec<String>,
+    pub(crate) time_filters: Vec<String>,
+    pub(crate) series_row_selector: Option<TimeSeriesRowSelector>,
+    pub(crate) append_mode: bool,
+    pub(crate) filter_deleted: bool,
+    pub(crate) merge_mode: MergeMode,
+    pub(crate) partition_expr_version: u64,
+}
+
+impl ScanRequestFingerprintBuilder {
+    pub(crate) fn build(self) -> ScanRequestFingerprint {
+        let Self {
+            read_columns,
+            read_column_types,
+            filters,
+            time_filters,
+            series_row_selector,
+            append_mode,
+            filter_deleted,
+            merge_mode,
+            partition_expr_version,
+        } = self;
+
+        ScanRequestFingerprint {
+            inner: Arc::new(SharedScanRequestFingerprint {
+                read_columns,
+                read_column_types,
+                filters,
+            }),
+            time_filters: (!time_filters.is_empty()).then(|| Arc::new(time_filters)),
+            series_row_selector,
+            append_mode,
+            filter_deleted,
+            merge_mode,
+            partition_expr_version,
+        }
+    }
+}
+
+/// Non-copiable struct of the fingerprint.
+#[derive(Debug, PartialEq, Eq, Hash)]
+struct SharedScanRequestFingerprint {
+    /// Logical columns of the projection.
+    read_columns: ReadColumns,
+    /// Column types of the projection.
+    /// We keep this to ensure we won't reuse the fingerprint after a schema change.
+    read_column_types: Vec<Option<ConcreteDataType>>,
+    /// Filters without the time index column and region partition exprs.
+    filters: Vec<String>,
+}
+
+impl ScanRequestFingerprint {
+    #[cfg(test)]
+    pub(crate) fn read_columns(&self) -> &ReadColumns {
+        &self.inner.read_columns
+    }
+
+    #[cfg(test)]
+    pub(crate) fn read_column_types(&self) -> &[Option<ConcreteDataType>] {
+        &self.inner.read_column_types
+    }
+
+    #[cfg(test)]
+    pub(crate) fn filters(&self) -> &[String] {
+        &self.inner.filters
+    }
+
+    #[cfg(test)]
+    pub(crate) fn time_filters(&self) -> &[String] {
+        self.time_filters
+            .as_deref()
+            .map(Vec::as_slice)
+            .unwrap_or(&[])
+    }
+
+    pub(crate) fn without_time_filters(&self) -> Self {
+        Self {
+            inner: Arc::clone(&self.inner),
+            time_filters: None,
+            series_row_selector: self.series_row_selector,
+            append_mode: self.append_mode,
+            filter_deleted: self.filter_deleted,
+            merge_mode: self.merge_mode,
+            partition_expr_version: self.partition_expr_version,
+        }
+    }
+
+    pub(crate) fn estimated_size(&self) -> usize {
+        mem::size_of::<SharedScanRequestFingerprint>()
+            + self.inner.read_columns.estimated_size()
+            + self.inner.read_column_types.capacity() * mem::size_of::<Option<ConcreteDataType>>()
+            + self.inner.filters.capacity() * mem::size_of::<String>()
+            + self
+                .inner
+                .filters
+                .iter()
+                .map(|filter| filter.capacity())
+                .sum::<usize>()
+            + self.time_filters.as_ref().map_or(0, |filters| {
+                mem::size_of::<Vec<String>>()
+                    + filters.capacity() * mem::size_of::<String>()
+                    + filters
+                        .iter()
+                        .map(|filter| filter.capacity())
+                        .sum::<usize>()
+            })
+    }
+}
+
+/// Cache key for range scan outputs.
+#[derive(Debug, Clone, PartialEq, Eq, Hash)]
+pub(crate) struct RangeScanCacheKey {
+    pub(crate) region_id: RegionId,
+    /// Sorted (file_id, row_group_index) pairs that uniquely identify the data this range covers.
+    pub(crate) row_groups: Vec<(FileId, i64)>,
+    pub(crate) scan: ScanRequestFingerprint,
+}
+
+impl RangeScanCacheKey {
+    pub(crate) fn estimated_size(&self) -> usize {
+        mem::size_of::<Self>()
+            + self.row_groups.capacity() * mem::size_of::<(FileId, i64)>()
+            + self.scan.estimated_size()
+    }
+}
+
+/// Cached result for one range scan.
+#[derive(Debug)]
+pub(crate) struct CachedBatchSlice {
+    batch: RecordBatch,
+    slice_lengths: Vec<usize>,
+}
+
+impl CachedBatchSlice {
+    fn metadata_size(&self) -> usize {
+        self.slice_lengths.capacity() * mem::size_of::<usize>()
+    }
+}
+
+pub(crate) struct RangeScanCacheValue {
+    cached_batches: Vec<CachedBatchSlice>,
+    /// Precomputed size of all compacted batches.
+    estimated_batches_size: usize,
+}
+
+impl RangeScanCacheValue {
+    pub(crate) fn new(
+        cached_batches: Vec<CachedBatchSlice>,
+        estimated_batches_size: usize,
+    ) -> Self {
+        Self {
+            cached_batches,
+            estimated_batches_size,
+        }
+    }
+
+    pub(crate) fn estimated_size(&self) -> usize {
+        mem::size_of::<Self>()
+            + self.cached_batches.capacity() * mem::size_of::<CachedBatchSlice>()
+            + self
+                .cached_batches
+                .iter()
+                .map(CachedBatchSlice::metadata_size)
+                .sum::<usize>()
+            + self.estimated_batches_size
+    }
+}
+
+/// Row groups and whether all sources are file-only for a partition range.
+pub(crate) struct PartitionRangeRowGroups {
+    /// Sorted (file_id, row_group_index) pairs.
+    pub(crate) row_groups: Vec<(FileId, i64)>,
+    pub(crate) only_file_sources: bool,
+}
+
+/// Collects (file_id, row_group_index) pairs from a partition range's row group indices.
+pub(crate) fn collect_partition_range_row_groups(
+    stream_ctx: &StreamContext,
+    part_range: &PartitionRange,
+) -> PartitionRangeRowGroups {
+    let range_meta = &stream_ctx.ranges[part_range.identifier];
+    let mut row_groups = Vec::new();
+    let mut only_file_sources = true;
+
+    for index in &range_meta.row_group_indices {
+        if stream_ctx.is_file_range_index(*index) {
+            let file_id = stream_ctx.input.file_from_index(*index).file_id().file_id();
+            row_groups.push((file_id, index.row_group_index));
+        } else {
+            only_file_sources = false;
+        }
+    }
+
+    row_groups.sort_unstable_by(|a, b| a.0.as_bytes().cmp(b.0.as_bytes()).then(a.1.cmp(&b.1)));
+
+    PartitionRangeRowGroups {
+        row_groups,
+        only_file_sources,
+    }
+}
+
+/// Builds a cache key for the given partition range if it is eligible for caching.
+pub(crate) fn build_range_cache_key(
+    stream_ctx: &StreamContext,
+    part_range: &PartitionRange,
+) -> Option<RangeScanCacheKey> {
+    if !stream_ctx.input.cache_strategy.has_range_result_cache() {
+        return None;
+    }
+
+    let fingerprint = stream_ctx.scan_fingerprint.as_ref()?;
+
+    // Dyn filters can change at runtime, so we can't cache when they're present.
+    let has_dyn_filters = stream_ctx
+        .input
+        .predicate_group()
+        .predicate_without_region()
+        .is_some_and(|p| !p.dyn_filters().is_empty());
+    if has_dyn_filters {
+        return None;
+    }
+
+    let rg = collect_partition_range_row_groups(stream_ctx, part_range);
+    if !rg.only_file_sources || rg.row_groups.is_empty() {
+        return None;
+    }
+
+    let range_meta = &stream_ctx.ranges[part_range.identifier];
+    let scan = if query_time_range_covers_partition_range(
+        stream_ctx.input.time_range.as_ref(),
+        range_meta.time_range,
+    ) {
+        fingerprint.without_time_filters()
+    } else {
+        fingerprint.clone()
+    };
+
+    Some(RangeScanCacheKey {
+        region_id: stream_ctx.input.region_metadata().region_id,
+        row_groups: rg.row_groups,
+        scan,
+    })
+}
+
+fn query_time_range_covers_partition_range(
+    query_time_range: Option<&TimestampRange>,
+    partition_time_range: FileTimeRange,
+) -> bool {
+    let Some(query_time_range) = query_time_range else {
+        return true;
+    };
+
+    let (part_start, part_end) = partition_time_range;
+    query_time_range.contains(&part_start) && query_time_range.contains(&part_end)
+}
+
+/// Returns a stream that replays cached record batches.
+pub(crate) fn cached_flat_range_stream(value: Arc<RangeScanCacheValue>) -> BoxedRecordBatchStream {
+    Box::pin(try_stream! {
+        for cached_batch in &value.cached_batches {
+            let mut offset = 0;
+            for &len in &cached_batch.slice_lengths {
+                yield cached_batch.batch.slice(offset, len);
+                offset += len;
+            }
+        }
+    })
+}
+
+enum CacheConcatCommand {
+    Compact(Vec<RecordBatch>),
+    Finish {
+        pending: Vec<RecordBatch>,
+        key: RangeScanCacheKey,
+        cache_strategy: CacheStrategy,
+        part_metrics: PartitionMetrics,
+        result_tx: Option<oneshot::Sender<Result<Arc<RangeScanCacheValue>>>>,
+    },
+}
+
+#[derive(Default)]
+struct CacheConcatState {
+    cached_batches: Vec<CachedBatchSlice>,
+    estimated_size: usize,
+}
+
+impl CacheConcatState {
+    async fn compact(
+        &mut self,
+        batches: Vec<RecordBatch>,
+        limiter: &crate::cache::RangeResultMemoryLimiter,
+    ) -> Result<()> {
+        if batches.is_empty() {
+            return Ok(());
+        }
+
+        let input_size = batches
+            .iter()
+            .map(RecordBatch::get_array_memory_size)
+            .sum::<usize>();
+        let _permit = limiter.acquire(input_size).await?;
+
+        let compacted = compact_record_batches(batches)?;
+        self.estimated_size += compacted.batch.get_array_memory_size();
+        self.cached_batches.push(compacted);
+        Ok(())
+    }
+
+    fn finish(self) -> RangeScanCacheValue {
+        RangeScanCacheValue::new(self.cached_batches, self.estimated_size)
+    }
+}
+
+fn compact_record_batches(batches: Vec<RecordBatch>) -> Result<CachedBatchSlice> {
+    debug_assert!(!batches.is_empty());
+
+    let slice_lengths = batches.iter().map(RecordBatch::num_rows).collect();
+    build_cached_batch_slice(batches, slice_lengths)
+}
+
+fn build_cached_batch_slice(
+    batches: Vec<RecordBatch>,
+    slice_lengths: Vec<usize>,
+) -> Result<CachedBatchSlice> {
+    let batch = if batches.len() == 1 {
+        batches.into_iter().next().unwrap()
+    } else {
+        let schema = batches[0].schema();
+        concat_batches(&schema, &batches).context(ComputeArrowSnafu)?
+    };
+
+    Ok(CachedBatchSlice {
+        batch,
+        slice_lengths,
+    })
+}
+
+async fn run_cache_concat_task(
+    mut rx: mpsc::UnboundedReceiver<CacheConcatCommand>,
+    limiter: Arc<crate::cache::RangeResultMemoryLimiter>,
+    skip_threshold_bytes: usize,
+) {
+    let mut state = CacheConcatState::default();
+
+    while let Some(cmd) = rx.recv().await {
+        match cmd {
+            CacheConcatCommand::Compact(batches) => {
+                if let Err(err) = state.compact(batches, &limiter).await {
+                    warn!(err; "Failed to compact range cache batches");
+                    return;
+                }
+                // Close the channel to stop further work as soon as the cached
+                // size exceeds the configured cache budget.
+                if state.estimated_size > skip_threshold_bytes {
+                    return;
+                }
+            }
+            CacheConcatCommand::Finish {
+                pending,
+                key,
+                cache_strategy,
+                part_metrics,
+                result_tx,
+            } => {
+                let compact_result = state
+                    .compact(pending, &limiter)
+                    .await
+                    .map(|()| state.finish());
+                let result = match compact_result {
+                    Ok(v) => {
+                        let value = Arc::new(v);
+                        part_metrics
+                            .inc_range_cache_size(key.estimated_size() + value.estimated_size());
+                        cache_strategy.put_range_result(key, value.clone());
+
+                        Ok(value)
+                    }
+                    Err(e) => {
+                        warn!(e; "Failed to finalize range cache batches");
+
+                        Err(e)
+                    }
+                };
+
+                if let Some(tx) = result_tx {
+                    let _ = tx.send(result);
+                }
+
+                break;
+            }
+        }
+    }
+}
+
+struct CacheBatchBuffer {
+    buffered_batches: Vec<RecordBatch>,
+    buffered_rows: usize,
+    buffered_size: usize,
+    sender: Option<mpsc::UnboundedSender<CacheConcatCommand>>,
+}
+
+impl CacheBatchBuffer {
+    fn new(cache_strategy: &CacheStrategy) -> Self {
+        let sender = cache_strategy.range_result_memory_limiter().map(|limiter| {
+            let skip_threshold_bytes = cache_strategy.range_result_cache_size().unwrap_or(0);
+            let (tx, rx) = mpsc::unbounded_channel();
+            common_runtime::spawn_global(run_cache_concat_task(
+                rx,
+                limiter.clone(),
+                skip_threshold_bytes,
+            ));
+            tx
+        });
+
+        Self {
+            buffered_batches: Vec::new(),
+            buffered_rows: 0,
+            buffered_size: 0,
+            sender,
+        }
+    }
+
+    fn push(&mut self, batch: RecordBatch) -> Result<()> {
+        if self.sender.is_none() {
+            return Ok(());
+        }
+
+        self.buffered_rows += batch.num_rows();
+        self.buffered_size += batch.get_array_memory_size();
+        self.buffered_batches.push(batch);
+
+        if self.buffered_batches.len() > 1
+            && (self.buffered_rows > DEFAULT_READ_BATCH_SIZE
+                || self.buffered_size > RANGE_CACHE_COMPACT_THRESHOLD_BYTES)
+        {
+            self.notify_compact();
+        }
+
+        Ok(())
+    }
+
+    fn notify_compact(&mut self) {
+        if self.buffered_batches.is_empty() || self.sender.is_none() {
+            return;
+        }
+
+        let batches = mem::take(&mut self.buffered_batches);
+        self.buffered_rows = 0;
+        self.buffered_size = 0;
+
+        let Some(sender) = &self.sender else {
+            return;
+        };
+        if sender.send(CacheConcatCommand::Compact(batches)).is_err() {
+            self.sender = None;
+        }
+    }
+
+    fn finish(
+        mut self,
+        key: RangeScanCacheKey,
+        cache_strategy: CacheStrategy,
+        part_metrics: PartitionMetrics,
+        result_tx: Option<oneshot::Sender<Result<Arc<RangeScanCacheValue>>>>,
+    ) {
+        let Some(sender) = self.sender.take() else {
+            return;
+        };
+
+        if sender
+            .send(CacheConcatCommand::Finish {
+                pending: mem::take(&mut self.buffered_batches),
+                key,
+                cache_strategy,
+                part_metrics,
+                result_tx,
+            })
+            .is_err()
+        {
+            self.sender = None;
+        }
+    }
+}
+
+/// Wraps a stream to cache its output for future range cache hits.
+pub(crate) fn cache_flat_range_stream(
+    mut stream: BoxedRecordBatchStream,
+    cache_strategy: CacheStrategy,
+    key: RangeScanCacheKey,
+    part_metrics: PartitionMetrics,
+) -> BoxedRecordBatchStream {
+    Box::pin(try_stream! {
+        let mut buffer = CacheBatchBuffer::new(&cache_strategy);
+        while let Some(batch) = stream.try_next().await? {
+            buffer.push(batch.clone())?;
+            yield batch;
+        }
+
+        buffer.finish(key, cache_strategy, part_metrics, None);
+    })
+}
+
+/// Creates a `cache_flat_range_stream` with dummy internals for benchmarking.
+///
+/// This avoids exposing `RangeScanCacheKey`, `ScanRequestFingerprint`, and
+/// `PartitionMetrics` publicly.
+#[cfg(feature = "test")]
+pub fn bench_cache_flat_range_stream(
+    stream: BoxedRecordBatchStream,
+    cache_size_bytes: u64,
+    region_id: RegionId,
+) -> BoxedRecordBatchStream {
+    use std::time::Instant;
+
+    use datafusion::physical_plan::metrics::ExecutionPlanMetricsSet;
+
+    use crate::region::options::MergeMode;
+
+    let cache_manager = Arc::new(
+        crate::cache::CacheManager::builder()
+            .range_result_cache_size(cache_size_bytes)
+            .build(),
+    );
+    let cache_strategy = CacheStrategy::EnableAll(cache_manager);
+
+    let fingerprint = ScanRequestFingerprintBuilder {
+        read_columns: ReadColumns::from_deduped_column_ids(std::iter::empty()),
+        read_column_types: vec![],
+        filters: vec![],
+        time_filters: vec![],
+        series_row_selector: None,
+        append_mode: false,
+        filter_deleted: false,
+        merge_mode: MergeMode::LastRow,
+        partition_expr_version: 0,
+    }
+    .build();
+
+    let key = RangeScanCacheKey {
+        region_id,
+        row_groups: vec![],
+        scan: fingerprint,
+    };
+
+    let metrics_set = ExecutionPlanMetricsSet::new();
+    let part_metrics =
+        PartitionMetrics::new(region_id, 0, "bench", Instant::now(), false, &metrics_set);
+
+    cache_flat_range_stream(stream, cache_strategy, key, part_metrics)
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+    use std::time::Instant;
+
+    use common_time::Timestamp;
+    use common_time::range::TimestampRange;
+    use common_time::timestamp::TimeUnit;
+    use datafusion::physical_plan::metrics::ExecutionPlanMetricsSet;
+    use datafusion_common::ScalarValue;
+    use datafusion_expr::{Expr, col, lit};
+    use smallvec::smallvec;
+    use store_api::storage::{FileId, RegionId};
+
+    use super::*;
+    use crate::cache::CacheManager;
+    use crate::read::flat_projection::FlatProjectionMapper;
+    use crate::read::range::{RangeMeta, RowGroupIndex, SourceIndex};
+    use crate::read::scan_region::{PredicateGroup, ScanInput};
+    use crate::test_util::memtable_util::metadata_with_primary_key;
+    use crate::test_util::scheduler_util::SchedulerEnv;
+    use crate::test_util::sst_util::sst_file_handle_with_file_id;
+
+    fn test_cache_strategy() -> CacheStrategy {
+        CacheStrategy::EnableAll(Arc::new(
+            CacheManager::builder()
+                .range_result_cache_size(1024 * 1024)
+                .build(),
+        ))
+    }
+
+    fn test_scan_fingerprint(
+        filters: Vec<String>,
+        time_filters: Vec<String>,
+        series_row_selector: Option<TimeSeriesRowSelector>,
+        filter_deleted: bool,
+        partition_expr_version: u64,
+    ) -> ScanRequestFingerprint {
+        let read_columns = ReadColumns::from_deduped_column_ids([1, 2]);
+        ScanRequestFingerprintBuilder {
+            read_columns,
+            read_column_types: vec![None, None],
+            filters,
+            time_filters,
+            series_row_selector,
+            append_mode: false,
+            filter_deleted,
+            merge_mode: MergeMode::LastRow,
+            partition_expr_version,
+        }
+        .build()
+    }
+
+    fn test_cache_context(strategy: &CacheStrategy) -> (RangeScanCacheKey, PartitionMetrics) {
+        let region_id = RegionId::new(1, 1);
+        let key = RangeScanCacheKey {
+            region_id,
+            row_groups: vec![],
+            scan: test_scan_fingerprint(vec![], vec![], None, false, 0),
+        };
+
+        let metrics_set = ExecutionPlanMetricsSet::new();
+        let part_metrics =
+            PartitionMetrics::new(region_id, 0, "test", Instant::now(), false, &metrics_set);
+
+        assert!(strategy.get_range_result(&key).is_none());
+        (key, part_metrics)
+    }
+
+    async fn finish_cache_batch_buffer(
+        buffer: CacheBatchBuffer,
+        key: RangeScanCacheKey,
+        cache_strategy: CacheStrategy,
+        part_metrics: PartitionMetrics,
+    ) -> Result<Arc<RangeScanCacheValue>> {
+        let (tx, rx) = oneshot::channel();
+        common_telemetry::info!("finish start");
+        buffer.finish(key, cache_strategy, part_metrics, Some(tx));
+        common_telemetry::info!("finish end");
+        rx.await.context(crate::error::RecvSnafu)?
+    }
+
+    async fn new_stream_context(
+        filters: Vec<Expr>,
+        query_time_range: Option<TimestampRange>,
+        partition_time_range: FileTimeRange,
+    ) -> (StreamContext, PartitionRange) {
+        let env = SchedulerEnv::new().await;
+        let metadata = Arc::new(metadata_with_primary_key(vec![0, 1], false));
+        let mapper = FlatProjectionMapper::new(&metadata, [0, 2, 3]).unwrap();
+        let predicate = PredicateGroup::new(metadata.as_ref(), &filters).unwrap();
+        let file_id = FileId::random();
+        let file = sst_file_handle_with_file_id(
+            file_id,
+            partition_time_range.0.value(),
+            partition_time_range.1.value(),
+        );
+        let input = ScanInput::new(env.access_layer.clone(), mapper)
+            .with_predicate(predicate)
+            .with_time_range(query_time_range)
+            .with_files(vec![file])
+            .with_cache(test_cache_strategy());
+        let range_meta = RangeMeta {
+            time_range: partition_time_range,
+            indices: smallvec![SourceIndex {
+                index: 0,
+                num_row_groups: 1,
+            }],
+            row_group_indices: smallvec![RowGroupIndex {
+                index: 0,
+                row_group_index: 0,
+            }],
+            num_rows: 10,
+        };
+        let partition_range = range_meta.new_partition_range(0);
+        let scan_fingerprint = crate::read::scan_region::build_scan_fingerprint(&input);
+        let stream_ctx = StreamContext {
+            input,
+            ranges: vec![range_meta],
+            scan_fingerprint,
+            query_start: Instant::now(),
+        };
+
+        (stream_ctx, partition_range)
+    }
+
+    /// Helper to create a timestamp millisecond literal.
+    fn ts_lit(val: i64) -> Expr {
+        lit(ScalarValue::TimestampMillisecond(Some(val), None))
+    }
+
+    fn normalized_exprs(exprs: impl IntoIterator<Item = Expr>) -> Vec<String> {
+        let mut exprs = exprs
+            .into_iter()
+            .map(|expr| expr.to_string())
+            .collect::<Vec<_>>();
+        exprs.sort_unstable();
+        exprs
+    }
+
+    async fn assert_range_cache_filters(
+        filters: Vec<Expr>,
+        query_time_range: Option<TimestampRange>,
+        partition_time_range: FileTimeRange,
+        expected_filters: Vec<Expr>,
+        expected_time_filters: Vec<Expr>,
+    ) {
+        let (stream_ctx, part_range) =
+            new_stream_context(filters, query_time_range, partition_time_range).await;
+
+        let key = build_range_cache_key(&stream_ctx, &part_range).unwrap();
+
+        assert_eq!(
+            key.scan.filters(),
+            normalized_exprs(expected_filters).as_slice()
+        );
+        assert_eq!(
+            key.scan.time_filters(),
+            normalized_exprs(expected_time_filters).as_slice()
+        );
+    }
+
+    #[tokio::test]
+    async fn strips_time_only_filters_when_query_covers_partition_range() {
+        assert_range_cache_filters(
+            vec![
+                col("ts").gt_eq(ts_lit(1000)),
+                col("ts").lt(ts_lit(2001)),
+                col("ts").is_not_null(),
+                col("k0").eq(lit("foo")),
+            ],
+            TimestampRange::with_unit(1000, 2002, TimeUnit::Millisecond),
+            (
+                Timestamp::new_millisecond(1000),
+                Timestamp::new_millisecond(2000),
+            ),
+            vec![col("k0").eq(lit("foo")), col("ts").is_not_null()],
+            vec![],
+        )
+        .await;
+    }
+
+    #[tokio::test]
+    async fn preserves_time_filters_when_query_does_not_cover_partition_range() {
+        assert_range_cache_filters(
+            vec![col("ts").gt_eq(ts_lit(1000)), col("k0").eq(lit("foo"))],
+            TimestampRange::with_unit(1000, 1500, TimeUnit::Millisecond),
+            (
+                Timestamp::new_millisecond(1000),
+                Timestamp::new_millisecond(2000),
+            ),
+            vec![col("k0").eq(lit("foo"))],
+            vec![col("ts").gt_eq(ts_lit(1000))],
+        )
+        .await;
+    }
+
+    #[tokio::test]
+    async fn strips_time_only_filters_when_query_has_no_time_range_limit() {
+        assert_range_cache_filters(
+            vec![
+                col("ts").gt_eq(ts_lit(1000)),
+                col("ts").is_not_null(),
+                col("k0").eq(lit("foo")),
+            ],
+            None,
+            (
+                Timestamp::new_millisecond(1000),
+                Timestamp::new_millisecond(2000),
+            ),
+            vec![col("k0").eq(lit("foo")), col("ts").is_not_null()],
+            vec![],
+        )
+        .await;
+    }
+
+    #[test]
+    fn normalizes_and_clears_time_filters() {
+        let normalized =
+            test_scan_fingerprint(vec!["k0 = 'foo'".to_string()], vec![], None, true, 0);
+
+        assert!(normalized.time_filters().is_empty());
+
+        let fingerprint = test_scan_fingerprint(
+            vec!["k0 = 'foo'".to_string()],
+            vec!["ts >= 1000".to_string()],
+            Some(TimeSeriesRowSelector::LastRow),
+            true,
+            7,
+        );
+
+        let reset = fingerprint.without_time_filters();
+
+        assert_eq!(reset.read_columns(), fingerprint.read_columns());
+        assert_eq!(reset.read_column_types(), fingerprint.read_column_types());
+        assert_eq!(reset.filters(), fingerprint.filters());
+        assert!(reset.time_filters().is_empty());
+        assert_eq!(reset.series_row_selector, fingerprint.series_row_selector);
+        assert_eq!(reset.append_mode, fingerprint.append_mode);
+        assert_eq!(reset.filter_deleted, fingerprint.filter_deleted);
+        assert_eq!(reset.merge_mode, fingerprint.merge_mode);
+        assert_eq!(
+            reset.partition_expr_version,
+            fingerprint.partition_expr_version
+        );
+    }
+
+    fn test_schema() -> Arc<datatypes::arrow::datatypes::Schema> {
+        use datatypes::arrow::datatypes::{DataType as ArrowDataType, Field, Schema};
+
+        Arc::new(Schema::new(vec![Field::new(
+            "value",
+            ArrowDataType::Int64,
+            false,
+        )]))
+    }
+
+    fn make_batch(values: &[i64]) -> RecordBatch {
+        use datatypes::arrow::array::Int64Array;
+
+        RecordBatch::try_new(
+            test_schema(),
+            vec![Arc::new(Int64Array::from(values.to_vec()))],
+        )
+        .unwrap()
+    }
+
+    fn make_large_binary_batch(rows: usize, bytes_per_row: usize) -> RecordBatch {
+        use datatypes::arrow::array::BinaryArray;
+        use datatypes::arrow::datatypes::{DataType as ArrowDataType, Field, Schema};
+
+        let schema = Arc::new(Schema::new(vec![Field::new(
+            "value",
+            ArrowDataType::Binary,
+            false,
+        )]));
+        let payload = vec![b'x'; bytes_per_row];
+        let values = (0..rows).map(|_| payload.as_slice()).collect::<Vec<_>>();
+
+        RecordBatch::try_new(schema, vec![Arc::new(BinaryArray::from_vec(values))]).unwrap()
+    }
+
+    #[test]
+    fn compact_record_batches_keeps_original_boundaries() {
+        let batches = vec![make_batch(&[1, 2]), make_batch(&[3]), make_batch(&[4, 5])];
+
+        let compacted = compact_record_batches(batches).unwrap();
+
+        assert_eq!(compacted.batch.num_rows(), 5);
+        assert_eq!(compacted.slice_lengths, vec![2, 1, 2]);
+    }
+
+    #[tokio::test]
+    async fn cached_flat_range_stream_replays_original_batches() {
+        let value = Arc::new(RangeScanCacheValue::new(
+            vec![CachedBatchSlice {
+                batch: make_batch(&[1, 2, 3]),
+                slice_lengths: vec![2, 1],
+            }],
+            make_batch(&[1, 2, 3]).get_array_memory_size(),
+        ));
+
+        let replayed = cached_flat_range_stream(value)
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+
+        assert_eq!(replayed.len(), 2);
+        assert_eq!(replayed[0].num_rows(), 2);
+        assert_eq!(replayed[1].num_rows(), 1);
+    }
+
+    #[tokio::test]
+    async fn cache_batch_buffer_finishes_pending_batches() {
+        let strategy = test_cache_strategy();
+        let batch = make_batch(&[1, 2, 3]);
+        let expected_size = batch.get_array_memory_size();
+        let (key, part_metrics) = test_cache_context(&strategy);
+
+        let mut buffer = CacheBatchBuffer::new(&strategy);
+        buffer.push(batch).unwrap();
+
+        let value = finish_cache_batch_buffer(buffer, key.clone(), strategy.clone(), part_metrics)
+            .await
+            .unwrap();
+        assert_eq!(value.cached_batches.len(), 1);
+        assert_eq!(value.cached_batches[0].slice_lengths, vec![3]);
+        assert_eq!(value.estimated_batches_size, expected_size);
+        assert!(Arc::ptr_eq(
+            &value,
+            &strategy.get_range_result(&key).unwrap()
+        ));
+    }
+
+    #[tokio::test]
+    async fn cache_batch_buffer_compacts_when_rows_exceed_default_batch_size() {
+        let strategy = test_cache_strategy();
+        let batch = make_batch(&vec![1; DEFAULT_READ_BATCH_SIZE / 2 + 1]);
+        let (key, part_metrics) = test_cache_context(&strategy);
+
+        let mut buffer = CacheBatchBuffer::new(&strategy);
+        buffer.push(batch.clone()).unwrap();
+        buffer.push(batch).unwrap();
+
+        assert_eq!(buffer.buffered_rows, 0);
+        assert!(buffer.buffered_batches.is_empty());
+
+        let value = finish_cache_batch_buffer(buffer, key, strategy, part_metrics)
+            .await
+            .unwrap();
+        assert_eq!(value.cached_batches.len(), 1);
+        assert_eq!(
+            value.cached_batches[0].slice_lengths,
+            vec![
+                DEFAULT_READ_BATCH_SIZE / 2 + 1,
+                DEFAULT_READ_BATCH_SIZE / 2 + 1
+            ]
+        );
+    }
+
+    #[tokio::test]
+    async fn cache_batch_buffer_compacts_when_buffered_size_exceeds_threshold() {
+        let large_batch = make_large_binary_batch(DEFAULT_READ_BATCH_SIZE, 4096);
+        let strategy = CacheStrategy::EnableAll(Arc::new(
+            CacheManager::builder()
+                .range_result_cache_size((large_batch.get_array_memory_size() * 3) as u64)
+                .build(),
+        ));
+        let (key, part_metrics) = test_cache_context(&strategy);
+
+        let mut buffer = CacheBatchBuffer::new(&strategy);
+        buffer.push(large_batch.clone()).unwrap();
+
+        assert_eq!(buffer.buffered_rows, large_batch.num_rows());
+        assert_eq!(buffer.buffered_batches.len(), 1);
+
+        buffer.push(large_batch.clone()).unwrap();
+
+        assert_eq!(buffer.buffered_rows, 0);
+        assert!(buffer.buffered_batches.is_empty());
+
+        let value = finish_cache_batch_buffer(buffer, key, strategy, part_metrics)
+            .await
+            .unwrap();
+        assert_eq!(value.cached_batches.len(), 1);
+        assert_eq!(
+            value.cached_batches[0].slice_lengths,
+            vec![large_batch.num_rows(), large_batch.num_rows()]
+        );
+    }
+
+    #[tokio::test]
+    async fn cache_batch_buffer_skips_cache_when_compacted_size_exceeds_limit() {
+        let large_batch = make_large_binary_batch(DEFAULT_READ_BATCH_SIZE / 2 + 1, 4096);
+        // Budget only fits two large batches.
+        let budget = (large_batch.get_array_memory_size() as u64) * 2 + 1;
+        let strategy = CacheStrategy::EnableAll(Arc::new(
+            CacheManager::builder()
+                .range_result_cache_size(budget)
+                .build(),
+        ));
+        let (key, part_metrics) = test_cache_context(&strategy);
+
+        let mut buffer = CacheBatchBuffer::new(&strategy);
+        for _ in 0..4 {
+            buffer.push(large_batch.clone()).unwrap();
+        }
+        assert!(
+            finish_cache_batch_buffer(buffer, key.clone(), strategy.clone(), part_metrics)
+                .await
+                .is_err()
+        );
+        assert!(strategy.get_range_result(&key).is_none());
+    }
+}
diff --git a/src/mito2/src/read/read_columns.rs b/src/mito2/src/read/read_columns.rs
new file mode 100644
index 0000000000..ae5876237b
--- /dev/null
+++ b/src/mito2/src/read/read_columns.rs
@@ -0,0 +1,557 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::{BTreeMap, HashSet};
+use std::mem;
+
+use datafusion_common::HashMap;
+use datafusion_expr::utils::expr_to_columns;
+use snafu::OptionExt;
+use store_api::metadata::RegionMetadataRef;
+use store_api::storage::{ColumnId, NestedPath, ProjectionInput};
+
+use crate::error::{InvalidRequestSnafu, Result};
+use crate::read::scan_region::PredicateGroup;
+
+/// Logical columns to read from a region.
+///
+/// Read columns describe which logical columns and nested fields should be read
+/// from storage. Each read column is identified by its [`ColumnId`],
+/// which represents the root column in the storage schema.
+///
+/// Nested fields under the column are specified by [`NestedPath`] entries.
+/// Each path includes the root column name as its first element.
+///
+/// For example, assume column id `9` corresponds to a root column named `j`
+/// with nested fields:
+///
+/// ```text
+/// j
+/// ├── a
+/// └── b
+///     └── c
+/// ```
+///
+/// The following SQL:
+///
+/// SELECT j.a, j.b.c FROM t
+///
+/// may produce read columns like:
+///
+/// ```text
+/// ReadColumn {
+///     column_id: 9,
+///     nested_paths: [
+///         ["j", "a"],
+///         ["j", "b", "c"],
+///     ]
+/// }
+/// ```
+///
+/// If `nested_paths` is empty, the whole column will be read.
+#[derive(Debug, Default, Clone, PartialEq, Eq, Hash)]
+pub struct ReadColumns {
+    pub cols: Vec<ReadColumn>,
+}
+
+impl ReadColumns {
+    pub fn from_deduped_column_ids<I>(column_ids: I) -> Self
+    where
+        I: IntoIterator<Item = ColumnId>,
+    {
+        let cols = column_ids
+            .into_iter()
+            .map(|col_id| ReadColumn::new(col_id, vec![]))
+            .collect();
+        ReadColumns { cols }
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.cols.is_empty()
+    }
+
+    pub fn column_ids_iter(&self) -> impl Iterator<Item = ColumnId> + '_ {
+        self.cols.iter().map(|column| column.column_id)
+    }
+
+    pub fn column_ids(&self) -> Vec<ColumnId> {
+        self.column_ids_iter().collect()
+    }
+
+    pub fn columns(&self) -> &[ReadColumn] {
+        &self.cols
+    }
+
+    pub fn estimated_size(&self) -> usize {
+        self.cols.capacity() * mem::size_of::<ReadColumn>()
+            + self
+                .cols
+                .iter()
+                .map(ReadColumn::estimated_size)
+                .sum::<usize>()
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Hash)]
+pub struct ReadColumn {
+    pub column_id: ColumnId,
+    /// Nested field paths under this column.
+    /// Empty means reading the whole column.
+    pub nested_paths: Vec<NestedPath>,
+}
+
+impl ReadColumn {
+    pub fn new(column_id: ColumnId, nested_paths: Vec<NestedPath>) -> Self {
+        Self {
+            column_id,
+            nested_paths,
+        }
+    }
+
+    pub fn nested_paths(&self) -> &[NestedPath] {
+        &self.nested_paths
+    }
+
+    pub fn estimated_size(&self) -> usize {
+        mem::size_of::<ColumnId>()
+            + self.nested_paths.capacity() * mem::size_of::<NestedPath>()
+            + self
+                .nested_paths
+                .iter()
+                .map(|path| {
+                    path.capacity() * mem::size_of::<String>()
+                        + path.iter().map(|node| node.capacity()).sum::<usize>()
+                })
+                .sum::<usize>()
+    }
+}
+
+pub fn merge(a: ReadColumns, b: ReadColumns) -> ReadColumns {
+    let mut merged = BTreeMap::<ColumnId, Vec<NestedPath>>::new();
+
+    for col in a.cols.into_iter().chain(b.cols) {
+        if let Some(nested_paths) = merged.get_mut(&col.column_id) {
+            if nested_paths.is_empty() || col.nested_paths.is_empty() {
+                *nested_paths = vec![];
+            } else {
+                merge_nested_paths(nested_paths, col.nested_paths);
+            }
+            continue;
+        }
+
+        merged.insert(col.column_id, normalize_nested_paths(col.nested_paths));
+    }
+
+    ReadColumns {
+        cols: merged
+            .into_iter()
+            .map(|(column_id, nested_paths)| ReadColumn {
+                column_id,
+                nested_paths,
+            })
+            .collect(),
+    }
+}
+
+fn normalize_nested_paths(nested_paths: Vec<NestedPath>) -> Vec<NestedPath> {
+    let mut normalized = Vec::with_capacity(nested_paths.len());
+    merge_nested_paths(&mut normalized, nested_paths);
+    normalized
+}
+
+fn merge_nested_paths(merged: &mut Vec<NestedPath>, incoming: Vec<NestedPath>) {
+    for path in incoming {
+        if merged
+            .iter()
+            .any(|existing| path.starts_with(existing.as_slice()))
+        {
+            continue;
+        }
+
+        merged.retain(|existing| !existing.starts_with(path.as_slice()));
+        merged.push(path);
+    }
+}
+
+/// Build [`ReadColumns`] from [`ProjectionInput`].
+///
+/// Note: If `projection.projection` is empty, this function still reads the
+/// time index column so the scan can preserve row counts for empty-output
+/// queries such as `SELECT COUNT(*)`.
+///
+/// Order:
+/// - This function keeps the first-seen order from `projection.projection`
+///   (duplicate indices are skipped).
+/// - Keeping a stable order makes [`ReadColumns`] comparisons deterministic
+///   (`Eq`/`Hash`) and avoids cache-key instability in upper layers.
+pub fn read_columns_from_projection(
+    projection: ProjectionInput,
+    metadata: &RegionMetadataRef,
+) -> Result<ReadColumns> {
+    let root_indices = if projection.projection.is_empty() {
+        vec![metadata.time_index_column_pos()]
+    } else {
+        projection.projection
+    };
+
+    let mut paths_by_col: HashMap<String, Vec<NestedPath>> =
+        HashMap::with_capacity(projection.nested_paths.len());
+    for path in projection.nested_paths {
+        let Some((root_name, _)) = path.split_first() else {
+            continue;
+        };
+        paths_by_col
+            .entry(root_name.clone())
+            .or_default()
+            .push(path);
+    }
+
+    let mut read_cols = Vec::with_capacity(root_indices.len());
+    let mut seen = HashSet::with_capacity(root_indices.len());
+    for root_idx in root_indices {
+        if !seen.insert(root_idx) {
+            continue;
+        }
+
+        let col = metadata
+            .column_metadatas
+            .get(root_idx)
+            .with_context(|| InvalidRequestSnafu {
+                region_id: metadata.region_id,
+                reason: format!("projection index {} is out of bounds", root_idx),
+            })?;
+        let col_id = col.column_id;
+
+        let nested_paths = paths_by_col
+            .remove(&col.column_schema.name)
+            .unwrap_or_default();
+
+        read_cols.push(ReadColumn {
+            column_id: col_id,
+            nested_paths,
+        });
+    }
+
+    Ok(ReadColumns { cols: read_cols })
+}
+
+/// Build [`ReadColumns`] from [`PredicateGroup`].
+///
+/// Order:
+/// - This function follows `metadata.column_metadatas` order when materializing
+///   columns from predicate-referenced names.
+/// - Using metadata order keeps the output deterministic for [`ReadColumns`]
+///   equality/hash checks and for cache keys derived from read columns.
+pub fn read_columns_from_predicate(
+    predicate: &PredicateGroup,
+    metadata: &RegionMetadataRef,
+) -> ReadColumns {
+    let mut root_names = HashSet::new();
+    let mut columns = HashSet::new();
+
+    if let Some(p) = predicate.predicate_without_region() {
+        for expr in p.exprs() {
+            columns.clear();
+            if expr_to_columns(expr, &mut columns).is_err() {
+                continue;
+            }
+            root_names.extend(columns.drain().map(|column| column.name));
+        }
+    }
+
+    if let Some(expr) = predicate.region_partition_expr() {
+        expr.collect_column_names(&mut root_names);
+    }
+
+    // TODO(fys): Parse nested paths from predicate expressions and attach them
+    // to read columns instead of always reading the whole root column.
+    let mut cols = Vec::with_capacity(root_names.len());
+    for column in &metadata.column_metadatas {
+        if root_names.contains(&column.column_schema.name) {
+            cols.push(ReadColumn::new(column.column_id, vec![]));
+        }
+    }
+
+    ReadColumns { cols }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use api::v1::SemanticType;
+    use datafusion_expr::{col, lit};
+    use datatypes::prelude::ConcreteDataType;
+    use datatypes::schema::ColumnSchema;
+    use store_api::metadata::{ColumnMetadata, RegionMetadataBuilder};
+    use store_api::storage::RegionId;
+
+    use super::*;
+
+    #[test]
+    fn test_read_columns_from_empty_projection() {
+        let metadata = new_test_metadata();
+
+        let read_columns =
+            read_columns_from_projection(ProjectionInput::default(), &metadata).unwrap();
+
+        let expected = ReadColumns {
+            cols: vec![ReadColumn::new(2, vec![])],
+        };
+        assert_eq!(expected, read_columns);
+
+        let projection_input =
+            ProjectionInput::new(vec![]).with_nested_paths(vec![vec!["1".to_string()]]);
+        let read_columns = read_columns_from_projection(projection_input, &metadata).unwrap();
+
+        let expected = ReadColumns {
+            cols: vec![ReadColumn::new(2, vec![])],
+        };
+        assert_eq!(expected, read_columns);
+    }
+
+    #[test]
+    fn test_read_columns_from_projection_with_nested_paths() {
+        let metadata = new_test_metadata();
+        let projection = ProjectionInput::new(vec![1, 0]).with_nested_paths(vec![
+            nested_path(&["field_0", "a"]),
+            nested_path(&["field_0", "b", "c"]),
+        ]);
+
+        let read_columns = read_columns_from_projection(projection, &metadata).unwrap();
+
+        let expected = ReadColumns {
+            cols: vec![
+                ReadColumn::new(
+                    3,
+                    vec![
+                        nested_path(&["field_0", "a"]),
+                        nested_path(&["field_0", "b", "c"]),
+                    ],
+                ),
+                ReadColumn::new(0, vec![]),
+            ],
+        };
+        assert_eq!(expected, read_columns,);
+    }
+
+    #[test]
+    fn test_read_columns_from_projection_dedups_duplicate_indices() {
+        let metadata = new_test_metadata();
+        let projection = ProjectionInput::new(vec![1, 1, 0]).with_nested_paths(vec![
+            nested_path(&["field_0", "a"]),
+            nested_path(&["field_0", "b", "c"]),
+        ]);
+
+        let read_columns = read_columns_from_projection(projection, &metadata).unwrap();
+
+        let expected = ReadColumns {
+            cols: vec![
+                ReadColumn::new(
+                    3,
+                    vec![
+                        nested_path(&["field_0", "a"]),
+                        nested_path(&["field_0", "b", "c"]),
+                    ],
+                ),
+                ReadColumn::new(0, vec![]),
+            ],
+        };
+        assert_eq!(expected, read_columns);
+    }
+
+    #[test]
+    fn test_read_columns_from_projection_out_of_bound() {
+        let metadata = new_test_metadata();
+        let projection = ProjectionInput::new(vec![3]);
+
+        let err = read_columns_from_projection(projection, &metadata).unwrap_err();
+
+        assert!(
+            err.to_string()
+                .contains("projection index 3 is out of bound")
+        );
+    }
+
+    #[test]
+    fn test_read_columns_from_predicate_reads_root_columns_only() {
+        let metadata = new_test_metadata();
+        let predicate = PredicateGroup::new(
+            metadata.as_ref(),
+            &[col("field_0").gt(lit(1)), col("tag_0").eq(lit("a"))],
+        )
+        .unwrap();
+
+        let read_columns = read_columns_from_predicate(&predicate, &metadata);
+
+        let expected = ReadColumns {
+            cols: vec![ReadColumn::new(0, vec![]), ReadColumn::new(3, vec![])],
+        };
+        assert_eq!(expected, read_columns);
+    }
+
+    #[test]
+    fn test_read_columns_from_predicate_empty() {
+        let metadata = new_test_metadata();
+        let predicate = PredicateGroup::new(metadata.as_ref(), &[]).unwrap();
+
+        let read_columns = read_columns_from_predicate(&predicate, &metadata);
+
+        assert!(read_columns.is_empty());
+    }
+
+    #[test]
+    fn test_merge_read_cols_with_only_root() {
+        let a = ReadColumns {
+            cols: vec![ReadColumn::new(3, vec![]), ReadColumn::new(1, vec![])],
+        };
+        let b = ReadColumns {
+            cols: vec![ReadColumn::new(2, vec![])],
+        };
+
+        let merged = merge(a, b);
+
+        assert_eq!(
+            merged,
+            ReadColumns {
+                cols: vec![
+                    ReadColumn::new(1, vec![]),
+                    ReadColumn::new(2, vec![]),
+                    ReadColumn::new(3, vec![]),
+                ],
+            }
+        );
+    }
+
+    #[test]
+    fn test_merge_read_cols_with_nested_paths() {
+        let a = ReadColumns {
+            cols: vec![ReadColumn::new(1, vec![nested_path(&["j", "a"])])],
+        };
+        let b = ReadColumns {
+            cols: vec![ReadColumn::new(
+                1,
+                vec![nested_path(&["j", "b"]), nested_path(&["j", "c"])],
+            )],
+        };
+
+        let merged = merge(a, b);
+
+        assert_eq!(
+            merged,
+            ReadColumns {
+                cols: vec![ReadColumn::new(
+                    1,
+                    vec![
+                        nested_path(&["j", "a"]),
+                        nested_path(&["j", "b"]),
+                        nested_path(&["j", "c"]),
+                    ],
+                )],
+            }
+        );
+    }
+
+    #[test]
+    fn test_merge_read_cols_with_column_override() {
+        let a = ReadColumns {
+            cols: vec![
+                ReadColumn::new(1, vec![nested_path(&["j", "a"])]),
+                ReadColumn::new(2, vec![nested_path(&["k", "b"])]),
+            ],
+        };
+        let b = ReadColumns {
+            cols: vec![
+                ReadColumn::new(1, vec![]),
+                ReadColumn::new(2, vec![nested_path(&["k", "b", "c"])]),
+            ],
+        };
+
+        let merged = merge(a, b);
+
+        assert_eq!(
+            merged,
+            ReadColumns {
+                cols: vec![
+                    ReadColumn::new(1, vec![]),
+                    ReadColumn::new(2, vec![nested_path(&["k", "b"])])
+                ],
+            }
+        );
+    }
+
+    #[test]
+    fn test_merge_read_cols_dedups_redundant_nested_paths() {
+        let a = ReadColumns {
+            cols: vec![ReadColumn::new(
+                1,
+                vec![
+                    nested_path(&["j", "a", "b"]),
+                    nested_path(&["j", "a"]),
+                    nested_path(&["j", "a", "b", "c"]),
+                ],
+            )],
+        };
+        let b = ReadColumns {
+            cols: vec![ReadColumn::new(1, vec![nested_path(&["j", "a"])])],
+        };
+
+        let merged = merge(a, b);
+
+        assert_eq!(
+            merged,
+            ReadColumns {
+                cols: vec![ReadColumn::new(1, vec![nested_path(&["j", "a"])])],
+            }
+        );
+    }
+
+    fn new_test_metadata() -> RegionMetadataRef {
+        let mut builder = RegionMetadataBuilder::new(RegionId::new(1, 1));
+        builder
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "tag_0".to_string(),
+                    ConcreteDataType::string_datatype(),
+                    true,
+                ),
+                semantic_type: SemanticType::Tag,
+                column_id: 0,
+            })
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "field_0".to_string(),
+                    ConcreteDataType::string_datatype(),
+                    true,
+                ),
+                semantic_type: SemanticType::Field,
+                column_id: 3,
+            })
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "ts".to_string(),
+                    ConcreteDataType::timestamp_millisecond_datatype(),
+                    false,
+                ),
+                semantic_type: SemanticType::Timestamp,
+                column_id: 2,
+            });
+        builder.primary_key(vec![0]);
+        Arc::new(builder.build().unwrap())
+    }
+
+    fn nested_path(parts: &[&str]) -> NestedPath {
+        parts.iter().map(|part| (*part).to_string()).collect()
+    }
+}
diff --git a/src/mito2/src/read/scan_region.rs b/src/mito2/src/read/scan_region.rs
index 5d934afd2d..19360f7208 100644
--- a/src/mito2/src/read/scan_region.rs
+++ b/src/mito2/src/read/scan_region.rs
@@ -14,7 +14,7 @@
 
 //! Scans a region according to the scan request.
 
-use std::collections::HashSet;
+use std::collections::{HashMap, HashSet};
 use std::fmt;
 use std::num::NonZeroU64;
 use std::sync::Arc;
@@ -31,38 +31,47 @@ use datafusion::physical_plan::expressions::DynamicFilterPhysicalExpr;
 use datafusion_common::Column;
 use datafusion_expr::Expr;
 use datafusion_expr::utils::expr_to_columns;
+use datatypes::data_type::ConcreteDataType;
+use datatypes::extension::json::is_json_extension_type;
+use datatypes::schema::Schema;
+use datatypes::schema::ext::ArrowSchemaExt;
+use datatypes::types::json_type::JsonNativeType;
 use futures::StreamExt;
 use partition::expr::PartitionExpr;
 use smallvec::SmallVec;
-use snafu::{OptionExt as _, ResultExt};
+use snafu::ResultExt;
 use store_api::metadata::{RegionMetadata, RegionMetadataRef};
 use store_api::region_engine::{PartitionRange, RegionScannerRef};
 use store_api::storage::{
-    ColumnId, RegionId, ScanRequest, SequenceRange, TimeSeriesDistribution, TimeSeriesRowSelector,
+    RegionId, ScanRequest, SequenceNumber, SequenceRange, TimeSeriesDistribution,
+    TimeSeriesRowSelector,
 };
-use table::predicate::{Predicate, build_time_range_predicate};
+use table::predicate::{Predicate, build_time_range_predicate, extract_time_range_from_expr};
 use tokio::sync::{Semaphore, mpsc};
 use tokio_stream::wrappers::ReceiverStream;
 
 use crate::access_layer::AccessLayerRef;
 use crate::cache::CacheStrategy;
-use crate::config::{DEFAULT_MAX_CONCURRENT_SCAN_FILES, DEFAULT_SCAN_CHANNEL_SIZE};
-use crate::error::{InvalidPartitionExprSnafu, InvalidRequestSnafu, Result};
+use crate::config::DEFAULT_MAX_CONCURRENT_SCAN_FILES;
+use crate::error::{InvalidPartitionExprSnafu, Result};
 #[cfg(feature = "enterprise")]
 use crate::extension::{BoxedExtensionRange, BoxedExtensionRangeProvider};
 use crate::memtable::{MemtableRange, RangesOptions};
 use crate::metrics::READ_SST_COUNT;
-use crate::read::compat::{self, CompatBatch, FlatCompatBatch, PrimaryKeyCompatBatch};
-use crate::read::projection::ProjectionMapper;
+use crate::read::compat::{self, FlatCompatBatch};
+use crate::read::flat_projection::FlatProjectionMapper;
 use crate::read::range::{FileRangeBuilder, MemRangeBuilder, RangeMeta, RowGroupIndex};
+use crate::read::range_cache::ScanRequestFingerprint;
+use crate::read::read_columns::{
+    ReadColumns, merge, read_columns_from_predicate, read_columns_from_projection,
+};
 use crate::read::seq_scan::SeqScan;
 use crate::read::series_scan::SeriesScan;
 use crate::read::stream::ScanBatchStream;
 use crate::read::unordered_scan::UnorderedScan;
-use crate::read::{Batch, BoxedRecordBatchStream, RecordBatch, Source};
+use crate::read::{BoxedRecordBatchStream, RecordBatch};
 use crate::region::options::MergeMode;
 use crate::region::version::VersionRef;
-use crate::sst::FormatType;
 use crate::sst::file::FileHandle;
 use crate::sst::index::bloom_filter::applier::{
     BloomFilterIndexApplierBuilder, BloomFilterIndexApplierRef,
@@ -76,8 +85,6 @@ use crate::sst::index::vector_index::applier::{VectorIndexApplier, VectorIndexAp
 use crate::sst::parquet::file_range::PreFilterMode;
 use crate::sst::parquet::reader::ReaderMetrics;
 
-/// Parallel scan channel size for flat format.
-const FLAT_SCAN_CHANNEL_SIZE: usize = 2;
 #[cfg(feature = "vector_index")]
 const VECTOR_INDEX_OVERFETCH_MULTIPLIER: usize = 2;
 
@@ -149,6 +156,14 @@ impl Scanner {
         }
     }
 
+    pub(crate) fn snapshot_sequence(&self) -> Option<SequenceNumber> {
+        match self {
+            Scanner::Seq(seq_scan) => seq_scan.input().snapshot_sequence,
+            Scanner::Unordered(unordered_scan) => unordered_scan.input().snapshot_sequence,
+            Scanner::Series(series_scan) => series_scan.input().snapshot_sequence,
+        }
+    }
+
     /// Sets the target partitions for the scanner. It can controls the parallelism of the scanner.
     pub(crate) fn set_target_partitions(&mut self, target_partitions: usize) {
         use store_api::region_engine::{PrepareRequest, RegionScanner};
@@ -221,8 +236,6 @@ pub(crate) struct ScanRegion {
     request: ScanRequest,
     /// Cache.
     cache_strategy: CacheStrategy,
-    /// Capacity of the channel to send data from parallel scan tasks to the main task.
-    parallel_scan_channel_size: usize,
     /// Maximum number of SST files to scan concurrently.
     max_concurrent_scan_files: usize,
     /// Whether to ignore inverted index.
@@ -253,7 +266,6 @@ impl ScanRegion {
             access_layer,
             request,
             cache_strategy,
-            parallel_scan_channel_size: DEFAULT_SCAN_CHANNEL_SIZE,
             max_concurrent_scan_files: DEFAULT_MAX_CONCURRENT_SCAN_FILES,
             ignore_inverted_index: false,
             ignore_fulltext_index: false,
@@ -265,16 +277,6 @@ impl ScanRegion {
         }
     }
 
-    /// Sets parallel scan task channel size.
-    #[must_use]
-    pub(crate) fn with_parallel_scan_channel_size(
-        mut self,
-        parallel_scan_channel_size: usize,
-    ) -> Self {
-        self.parallel_scan_channel_size = parallel_scan_channel_size;
-        self
-    }
-
     /// Sets maximum number of SST files to scan concurrently.
     #[must_use]
     pub(crate) fn with_max_concurrent_scan_files(
@@ -398,41 +400,44 @@ impl ScanRegion {
         self.request.distribution == Some(TimeSeriesDistribution::PerSeries)
     }
 
-    /// Returns true if the region use flat format.
-    fn use_flat_format(&self) -> bool {
-        self.request.force_flat_format
-            || self.version.options.sst_format.unwrap_or_default() == FormatType::Flat
-    }
-
     /// Creates a scan input.
     #[tracing::instrument(skip_all, fields(region_id = %self.region_id()))]
-    async fn scan_input(mut self) -> Result<ScanInput> {
+    async fn scan_input(self) -> Result<ScanInput> {
         let sst_min_sequence = self.request.sst_min_sequence.and_then(NonZeroU64::new);
         let time_range = self.build_time_range_predicate();
         let predicate = PredicateGroup::new(&self.version.metadata, &self.request.filters)?;
-        let flat_format = self.use_flat_format();
 
-        let read_column_ids = match &self.request.projection {
-            Some(p) => self.build_read_column_ids(p, &predicate)?,
-            None => self
-                .version
-                .metadata
-                .column_metadatas
-                .iter()
-                .map(|col| col.column_id)
-                .collect(),
+        let read_cols = match &self.request.projection_input {
+            Some(p) => {
+                // Read columns include the pushed-down projection and columns
+                // resolved from the predicate.
+                let metadata = &self.version.metadata;
+                let from_projection = read_columns_from_projection(p.clone(), metadata)?;
+                let from_predicate = read_columns_from_predicate(&predicate, metadata);
+                merge(from_projection, from_predicate)
+            }
+            None => {
+                let read_col_ids = self
+                    .version
+                    .metadata
+                    .column_metadatas
+                    .iter()
+                    .map(|col| col.column_id);
+                ReadColumns::from_deduped_column_ids(read_col_ids)
+            }
         };
+        let read_col_ids = read_cols.column_ids();
 
         // The mapper always computes projected column ids as the schema of SSTs may change.
-        let mapper = match &self.request.projection {
-            Some(p) => ProjectionMapper::new_with_read_columns(
+        let mut mapper = match self.request.projection_indices() {
+            Some(p) => FlatProjectionMapper::new_with_read_columns(
                 &self.version.metadata,
-                p.iter().copied(),
-                flat_format,
-                read_column_ids.clone(),
+                p.to_vec(),
+                read_cols,
             )?,
-            None => ProjectionMapper::all(&self.version.metadata, flat_format)?,
+            None => FlatProjectionMapper::all(&self.version.metadata)?,
         };
+        concretize_json_types(&mut mapper, &self.request.json_type_hint);
 
         let ssts = &self.version.ssts;
         let mut files = Vec::new();
@@ -478,7 +483,7 @@ impl ScanRegion {
                 continue;
             }
             let ranges_in_memtable = m.ranges(
-                Some(read_column_ids.as_slice()),
+                Some(&read_col_ids),
                 RangesOptions::default()
                     .with_predicate(predicate.clone())
                     .with_sequence(SequenceRange::new(
@@ -495,14 +500,13 @@ impl ScanRegion {
 
         let region_id = self.region_id();
         debug!(
-            "Scan region {}, request: {:?}, time range: {:?}, memtables: {}, ssts_to_read: {}, append_mode: {}, flat_format: {}",
+            "Scan region {}, request: {:?}, time range: {:?}, memtables: {}, ssts_to_read: {}, append_mode: {}",
             region_id,
             self.request,
             time_range,
             mem_range_builders.len(),
             files.len(),
             self.version.options.append_mode,
-            flat_format,
         );
 
         let (non_field_filters, field_filters) = self.partition_by_field_filters();
@@ -529,11 +533,6 @@ impl ScanRegion {
             }
         });
 
-        if flat_format {
-            // The batch is already large enough so we use a small channel size here.
-            self.parallel_scan_channel_size = FLAT_SCAN_CHANNEL_SIZE;
-        }
-
         let input = ScanInput::new(self.access_layer, mapper)
             .with_time_range(Some(time_range))
             .with_predicate(predicate)
@@ -543,7 +542,6 @@ impl ScanRegion {
             .with_inverted_index_appliers(inverted_index_appliers)
             .with_bloom_filter_index_appliers(bloom_filter_appliers)
             .with_fulltext_index_appliers(fulltext_index_appliers)
-            .with_parallel_scan_channel_size(self.parallel_scan_channel_size)
             .with_max_concurrent_scan_files(self.max_concurrent_scan_files)
             .with_start_time(self.start_time)
             .with_append_mode(self.version.options.append_mode)
@@ -551,7 +549,15 @@ impl ScanRegion {
             .with_merge_mode(self.version.options.merge_mode())
             .with_series_row_selector(self.request.series_row_selector)
             .with_distribution(self.request.distribution)
-            .with_flat_format(flat_format);
+            .with_explain_flat_format(
+                self.version.options.sst_format == Some(crate::sst::FormatType::Flat),
+            )
+            .with_snapshot_sequence(
+                self.request
+                    .snapshot_on_scan
+                    .then_some(self.request.memtable_max_sequence)
+                    .flatten(),
+            );
         #[cfg(feature = "vector_index")]
         let input = input
             .with_vector_index_applier(vector_index_applier)
@@ -586,72 +592,6 @@ impl ScanRegion {
         build_time_range_predicate(&time_index.column_schema.name, unit, &self.request.filters)
     }
 
-    /// Return all columns id to read according to the projection and filters.
-    fn build_read_column_ids(
-        &self,
-        projection: &[usize],
-        predicate: &PredicateGroup,
-    ) -> Result<Vec<ColumnId>> {
-        let metadata = &self.version.metadata;
-        // use Vec for read_column_ids to keep the order of columns.
-        let mut read_column_ids = Vec::new();
-        let mut seen = HashSet::new();
-
-        for idx in projection {
-            let column =
-                metadata
-                    .column_metadatas
-                    .get(*idx)
-                    .with_context(|| InvalidRequestSnafu {
-                        region_id: metadata.region_id,
-                        reason: format!("projection index {} is out of bound", idx),
-                    })?;
-            seen.insert(column.column_id);
-            // keep the projection order
-            read_column_ids.push(column.column_id);
-        }
-
-        if projection.is_empty() {
-            let time_index = metadata.time_index_column().column_id;
-            if seen.insert(time_index) {
-                read_column_ids.push(time_index);
-            }
-        }
-
-        let mut extra_names = HashSet::new();
-        let mut columns = HashSet::new();
-
-        for expr in &self.request.filters {
-            columns.clear();
-            if expr_to_columns(expr, &mut columns).is_err() {
-                continue;
-            }
-            extra_names.extend(columns.iter().map(|column| column.name.clone()));
-        }
-
-        if let Some(expr) = predicate.region_partition_expr() {
-            expr.collect_column_names(&mut extra_names);
-        }
-
-        if !extra_names.is_empty() {
-            for column in &metadata.column_metadatas {
-                if extra_names.contains(column.column_schema.name.as_str())
-                    && !seen.contains(&column.column_id)
-                {
-                    read_column_ids.push(column.column_id);
-                }
-                extra_names.remove(column.column_schema.name.as_str());
-            }
-            if !extra_names.is_empty() {
-                warn!(
-                    "Some columns in filters are not found in region {}: {:?}",
-                    metadata.region_id, extra_names
-                );
-            }
-        }
-        Ok(read_column_ids)
-    }
-
     /// Partitions filters into two groups: non-field filters and field filters.
     /// Returns `(non_field_filters, field_filters)`.
     fn partition_by_field_filters(&self) -> (Vec<Expr>, Vec<Expr>) {
@@ -793,6 +733,34 @@ impl ScanRegion {
     }
 }
 
+fn concretize_json_types(
+    mapper: &mut FlatProjectionMapper,
+    json_type_hint: &HashMap<String, JsonNativeType>,
+) {
+    let output_schema = mapper.output_schema();
+    let output_arrow_schema = output_schema.arrow_schema();
+    if !output_arrow_schema.has_json_extension_field() {
+        return;
+    }
+
+    let mut column_schemas = output_schema.column_schemas().to_vec();
+    for (idx, column_schema) in column_schemas.iter_mut().enumerate() {
+        if !is_json_extension_type(&output_arrow_schema.fields()[idx]) {
+            continue;
+        }
+        let Some(json_type) = json_type_hint.get(&column_schema.name) else {
+            continue;
+        };
+        column_schema.data_type = ConcreteDataType::from(json_type);
+    }
+
+    let output_schema = Arc::new(Schema::new_with_version(
+        column_schemas,
+        output_schema.version(),
+    ));
+    mapper.with_output_schema(output_schema);
+}
+
 /// Returns true if the time range of a SST `file` matches the `predicate`.
 fn file_in_range(file: &FileHandle, predicate: &TimestampRange) -> bool {
     if predicate == &TimestampRange::min_to_max() {
@@ -809,13 +777,13 @@ pub struct ScanInput {
     /// Region SST access layer.
     access_layer: AccessLayerRef,
     /// Maps projected Batches to RecordBatches.
-    pub(crate) mapper: Arc<ProjectionMapper>,
-    /// Column ids to read from memtables and SSTs.
+    pub(crate) mapper: Arc<FlatProjectionMapper>,
+    /// The columns to read from memtables and SSTs.
     /// Notice this is different from the columns in `mapper` which are projected columns.
     /// But this read columns might also include non-projected columns needed for filtering.
-    pub(crate) read_column_ids: Vec<ColumnId>,
+    pub(crate) read_cols: ReadColumns,
     /// Time range filter for time index.
-    time_range: Option<TimestampRange>,
+    pub(crate) time_range: Option<TimestampRange>,
     /// Predicate to push down.
     pub(crate) predicate: PredicateGroup,
     /// Region partition expr applied at read time.
@@ -828,8 +796,6 @@ pub struct ScanInput {
     pub(crate) cache_strategy: CacheStrategy,
     /// Ignores file not found error.
     ignore_file_not_found: bool,
-    /// Capacity of the channel to send data from parallel scan tasks to the main task.
-    pub(crate) parallel_scan_channel_size: usize,
     /// Maximum number of SST files to scan concurrently.
     pub(crate) max_concurrent_scan_files: usize,
     /// Index appliers.
@@ -854,8 +820,10 @@ pub struct ScanInput {
     pub(crate) series_row_selector: Option<TimeSeriesRowSelector>,
     /// Hint for the required distribution of the scanner.
     pub(crate) distribution: Option<TimeSeriesDistribution>,
-    /// Whether to use flat format.
-    pub(crate) flat_format: bool,
+    /// Whether the region's configured SST format is flat.
+    explain_flat_format: bool,
+    /// Snapshot upper bound bound at scan open and propagated back to the caller.
+    pub(crate) snapshot_sequence: Option<SequenceNumber>,
     /// Whether this scan is for compaction.
     pub(crate) compaction: bool,
     #[cfg(feature = "enterprise")]
@@ -865,10 +833,10 @@ pub struct ScanInput {
 impl ScanInput {
     /// Creates a new [ScanInput].
     #[must_use]
-    pub(crate) fn new(access_layer: AccessLayerRef, mapper: ProjectionMapper) -> ScanInput {
+    pub(crate) fn new(access_layer: AccessLayerRef, mapper: FlatProjectionMapper) -> ScanInput {
         ScanInput {
             access_layer,
-            read_column_ids: mapper.column_ids().to_vec(),
+            read_cols: mapper.read_columns().clone(),
             mapper: Arc::new(mapper),
             time_range: None,
             predicate: PredicateGroup::default(),
@@ -877,7 +845,6 @@ impl ScanInput {
             files: Vec::new(),
             cache_strategy: CacheStrategy::Disabled,
             ignore_file_not_found: false,
-            parallel_scan_channel_size: DEFAULT_SCAN_CHANNEL_SIZE,
             max_concurrent_scan_files: DEFAULT_MAX_CONCURRENT_SCAN_FILES,
             inverted_index_appliers: [None, None],
             bloom_filter_index_appliers: [None, None],
@@ -892,7 +859,8 @@ impl ScanInput {
             merge_mode: MergeMode::default(),
             series_row_selector: None,
             distribution: None,
-            flat_format: false,
+            explain_flat_format: false,
+            snapshot_sequence: None,
             compaction: false,
             #[cfg(feature = "enterprise")]
             extension_ranges: Vec::new(),
@@ -942,16 +910,6 @@ impl ScanInput {
         self
     }
 
-    /// Sets scan task channel size.
-    #[must_use]
-    pub(crate) fn with_parallel_scan_channel_size(
-        mut self,
-        parallel_scan_channel_size: usize,
-    ) -> Self {
-        self.parallel_scan_channel_size = parallel_scan_channel_size;
-        self
-    }
-
     /// Sets maximum number of SST files to scan concurrently.
     #[must_use]
     pub(crate) fn with_max_concurrent_scan_files(
@@ -1048,6 +1006,13 @@ impl ScanInput {
         self
     }
 
+    /// Sets whether the region's configured SST format is flat for explain output.
+    #[must_use]
+    pub(crate) fn with_explain_flat_format(mut self, explain_flat_format: bool) -> Self {
+        self.explain_flat_format = explain_flat_format;
+        self
+    }
+
     /// Sets the time series row selector.
     #[must_use]
     pub(crate) fn with_series_row_selector(
@@ -1058,10 +1023,12 @@ impl ScanInput {
         self
     }
 
-    /// Sets whether to use flat format.
     #[must_use]
-    pub(crate) fn with_flat_format(mut self, flat_format: bool) -> Self {
-        self.flat_format = flat_format;
+    pub(crate) fn with_snapshot_sequence(
+        mut self,
+        snapshot_sequence: Option<SequenceNumber>,
+    ) -> Self {
+        self.snapshot_sequence = snapshot_sequence;
         self
     }
 
@@ -1072,38 +1039,6 @@ impl ScanInput {
         self
     }
 
-    /// Scans sources in parallel.
-    ///
-    /// # Panics if the input doesn't allow parallel scan.
-    #[tracing::instrument(
-        skip(self, sources, semaphore),
-        fields(
-            region_id = %self.region_metadata().region_id,
-            source_count = sources.len()
-        )
-    )]
-    pub(crate) fn create_parallel_sources(
-        &self,
-        sources: Vec<Source>,
-        semaphore: Arc<Semaphore>,
-    ) -> Result<Vec<Source>> {
-        if sources.len() <= 1 {
-            return Ok(sources);
-        }
-
-        // Spawn a task for each source.
-        let sources = sources
-            .into_iter()
-            .map(|source| {
-                let (sender, receiver) = mpsc::channel(self.parallel_scan_channel_size);
-                self.spawn_scan_task(source, semaphore.clone(), sender);
-                let stream = Box::pin(ReceiverStream::new(receiver));
-                Source::Stream(stream)
-            })
-            .collect();
-        Ok(sources)
-    }
-
     /// Builds memtable ranges to scan by `index`.
     pub(crate) fn build_mem_ranges(&self, index: RowGroupIndex) -> SmallVec<[MemtableRange; 2]> {
         let memtable = &self.memtables[index.index];
@@ -1141,16 +1076,21 @@ impl ScanInput {
     pub async fn prune_file(
         &self,
         file: &FileHandle,
+        pre_filter_mode: PreFilterMode,
         reader_metrics: &mut ReaderMetrics,
     ) -> Result<FileRangeBuilder> {
         let predicate = self.predicate_for_file(file);
-        let filter_mode = pre_filter_mode(self.append_mode, self.merge_mode);
-        let decode_pk_values = !self.compaction && self.mapper.has_tags();
+        let decode_pk_values = !self.compaction
+            && self
+                .mapper
+                .read_columns()
+                .column_ids_iter()
+                .any(|column_id| self.mapper.metadata().primary_key.contains(&column_id));
         let reader = self
             .access_layer
             .read_sst(file.clone())
             .predicate(predicate)
-            .projection(Some(self.read_column_ids.clone()))
+            .projection(Some(self.read_cols.clone()))
             .cache(self.cache_strategy.clone())
             .inverted_index_appliers(self.inverted_index_appliers.clone())
             .bloom_filter_index_appliers(self.bloom_filter_index_appliers.clone())
@@ -1164,9 +1104,8 @@ impl ScanInput {
         };
         let res = reader
             .expected_metadata(Some(self.mapper.metadata().clone()))
-            .flat_format(self.flat_format)
             .compaction(self.compaction)
-            .pre_filter_mode(filter_mode)
+            .pre_filter_mode(pre_filter_mode)
             .decode_primary_key_values(decode_pk_values)
             .build_reader_input(reader_metrics)
             .await;
@@ -1193,70 +1132,17 @@ impl ScanInput {
         if need_compat {
             // They have different schema. We need to adapt the batch first so the
             // mapper can convert it.
-            let compat = if let Some(flat_format) = file_range_ctx.read_format().as_flat() {
-                let mapper = self.mapper.as_flat().unwrap();
-                FlatCompatBatch::try_new(
-                    mapper,
-                    flat_format.metadata(),
-                    flat_format.format_projection(),
-                    self.compaction,
-                )?
-                .map(CompatBatch::Flat)
-            } else {
-                let compact_batch = PrimaryKeyCompatBatch::new(
-                    &self.mapper,
-                    file_range_ctx.read_format().metadata().clone(),
-                )?;
-                Some(CompatBatch::PrimaryKey(compact_batch))
-            };
+            let compat = FlatCompatBatch::try_new(
+                &self.mapper,
+                file_range_ctx.read_format().metadata(),
+                file_range_ctx.read_format().format_projection(),
+                self.compaction,
+            )?;
             file_range_ctx.set_compat_batch(compat);
         }
         Ok(FileRangeBuilder::new(Arc::new(file_range_ctx), selection))
     }
 
-    /// Scans the input source in another task and sends batches to the sender.
-    #[tracing::instrument(
-        skip(self, input, semaphore, sender),
-        fields(region_id = %self.region_metadata().region_id)
-    )]
-    pub(crate) fn spawn_scan_task(
-        &self,
-        mut input: Source,
-        semaphore: Arc<Semaphore>,
-        sender: mpsc::Sender<Result<Batch>>,
-    ) {
-        let region_id = self.region_metadata().region_id;
-        let span = tracing::info_span!(
-            "ScanInput::parallel_scan_task",
-            region_id = %region_id,
-            stream_kind = "batch"
-        );
-        common_runtime::spawn_global(
-            async move {
-                loop {
-                    // We release the permit before sending result to avoid the task waiting on
-                    // the channel with the permit held.
-                    let maybe_batch = {
-                        // Safety: We never close the semaphore.
-                        let _permit = semaphore.acquire().await.unwrap();
-                        input.next_batch().await
-                    };
-                    match maybe_batch {
-                        Ok(Some(batch)) => {
-                            let _ = sender.send(Ok(batch)).await;
-                        }
-                        Ok(None) => break,
-                        Err(e) => {
-                            let _ = sender.send(Err(e)).await;
-                            break;
-                        }
-                    }
-                }
-            }
-            .instrument(span),
-        );
-    }
-
     /// Scans flat sources (RecordBatch streams) in parallel.
     ///
     /// # Panics if the input doesn't allow parallel scan.
@@ -1271,6 +1157,7 @@ impl ScanInput {
         &self,
         sources: Vec<BoxedRecordBatchStream>,
         semaphore: Arc<Semaphore>,
+        channel_size: usize,
     ) -> Result<Vec<BoxedRecordBatchStream>> {
         if sources.len() <= 1 {
             return Ok(sources);
@@ -1280,7 +1167,7 @@ impl ScanInput {
         let sources = sources
             .into_iter()
             .map(|source| {
-                let (sender, receiver) = mpsc::channel(self.parallel_scan_channel_size);
+                let (sender, receiver) = mpsc::channel(channel_size);
                 self.spawn_flat_scan_task(source, semaphore.clone(), sender);
                 let stream = Box::pin(ReceiverStream::new(receiver));
                 Box::pin(stream) as _
@@ -1370,6 +1257,18 @@ impl ScanInput {
     pub fn region_metadata(&self) -> &RegionMetadataRef {
         self.mapper.metadata()
     }
+
+    fn range_pre_filter_mode(&self, source_count: usize) -> PreFilterMode {
+        if source_count <= 1 {
+            // Duplicated rows in the same source is not a normal case and we don't provide
+            // strict dedup semantic (last_row/last_non_null) for it. We expect the duplicated rows
+            // are exactly identical in the same source so we use PreFilterMode::All for
+            // performance reason.
+            return PreFilterMode::All;
+        }
+
+        pre_filter_mode(self.append_mode, self.merge_mode)
+    }
 }
 
 #[cfg(feature = "enterprise")]
@@ -1412,11 +1311,108 @@ fn pre_filter_mode(append_mode: bool, merge_mode: MergeMode) -> PreFilterMode {
     }
 
     match merge_mode {
-        MergeMode::LastRow => PreFilterMode::SkipFieldsOnDelete,
+        MergeMode::LastRow => PreFilterMode::SkipFields,
         MergeMode::LastNonNull => PreFilterMode::SkipFields,
     }
 }
 
+/// Builds a [ScanRequestFingerprint] from a [ScanInput] if the scan is eligible
+/// for partition range caching.
+pub(crate) fn build_scan_fingerprint(input: &ScanInput) -> Option<ScanRequestFingerprint> {
+    let eligible = !input.compaction
+        && !input.files.is_empty()
+        && matches!(input.cache_strategy, CacheStrategy::EnableAll(_));
+
+    if !eligible {
+        return None;
+    }
+
+    let metadata = input.region_metadata();
+    let tag_names: HashSet<&str> = metadata
+        .column_metadatas
+        .iter()
+        .filter(|col| col.semantic_type == SemanticType::Tag)
+        .map(|col| col.column_schema.name.as_str())
+        .collect();
+
+    let time_index = metadata.time_index_column();
+    let time_index_name = time_index.column_schema.name.clone();
+    let ts_col_unit = time_index
+        .column_schema
+        .data_type
+        .as_timestamp()
+        .expect("Time index must have timestamp-compatible type")
+        .unit();
+
+    let exprs = input
+        .predicate_group()
+        .predicate_without_region()
+        .map(|predicate| predicate.exprs())
+        .unwrap_or_default();
+
+    let mut filters = Vec::new();
+    let mut time_filters = Vec::new();
+    let mut has_tag_filter = false;
+    let mut columns = HashSet::new();
+
+    for expr in exprs {
+        columns.clear();
+        let is_time_only = match expr_to_columns(expr, &mut columns) {
+            Ok(()) if !columns.is_empty() => {
+                has_tag_filter |= columns
+                    .iter()
+                    .any(|col| tag_names.contains(col.name.as_str()));
+                columns.iter().all(|col| col.name == time_index_name)
+            }
+            _ => false,
+        };
+
+        if is_time_only
+            && extract_time_range_from_expr(&time_index_name, ts_col_unit, expr).is_some()
+        {
+            // Range-reducible time predicates can be safely dropped from the
+            // cache key when the query time range covers the partition range.
+            time_filters.push(expr.to_string());
+        } else {
+            // Non-time filters and non-range time predicates (those that
+            // extract_time_range_from_expr cannot convert to a TimestampRange)
+            // always stay in the cache key.
+            filters.push(expr.to_string());
+        }
+    }
+
+    if !has_tag_filter {
+        // We only cache requests that have tag filters to avoid caching all series.
+        return None;
+    }
+
+    // Ensure the filters are sorted for consistent fingerprinting.
+    filters.sort_unstable();
+    time_filters.sort_unstable();
+    let read_columns = input.read_cols.clone();
+    Some(
+        crate::read::range_cache::ScanRequestFingerprintBuilder {
+            read_column_types: read_columns
+                .column_ids_iter()
+                .map(|id| {
+                    metadata
+                        .column_by_id(id)
+                        .map(|col| col.column_schema.data_type.clone())
+                })
+                .collect(),
+            read_columns,
+            filters,
+            time_filters,
+            series_row_selector: input.series_row_selector,
+            append_mode: input.append_mode,
+            filter_deleted: input.filter_deleted,
+            merge_mode: input.merge_mode,
+            partition_expr_version: metadata.partition_expr_version,
+        }
+        .build(),
+    )
+}
+
 /// Context shared by different streams from a scanner.
 /// It contains the input and ranges to scan.
 pub struct StreamContext {
@@ -1424,6 +1420,10 @@ pub struct StreamContext {
     pub input: ScanInput,
     /// Metadata for partition ranges.
     pub(crate) ranges: Vec<RangeMeta>,
+    /// Precomputed scan fingerprint for partition range caching.
+    /// `None` when the scan is not eligible for caching.
+    #[allow(dead_code)]
+    pub(crate) scan_fingerprint: Option<ScanRequestFingerprint>,
 
     // Metrics:
     /// The start time of the query.
@@ -1436,10 +1436,12 @@ impl StreamContext {
         let query_start = input.query_start.unwrap_or_else(Instant::now);
         let ranges = RangeMeta::seq_scan_ranges(&input);
         READ_SST_COUNT.observe(input.num_files() as f64);
+        let scan_fingerprint = build_scan_fingerprint(&input);
 
         Self {
             input,
             ranges,
+            scan_fingerprint,
             query_start,
         }
     }
@@ -1449,10 +1451,12 @@ impl StreamContext {
         let query_start = input.query_start.unwrap_or_else(Instant::now);
         let ranges = RangeMeta::unordered_scan_ranges(&input);
         READ_SST_COUNT.observe(input.num_files() as f64);
+        let scan_fingerprint = build_scan_fingerprint(&input);
 
         Self {
             input,
             ranges,
+            scan_fingerprint,
             query_start,
         }
     }
@@ -1467,6 +1471,13 @@ impl StreamContext {
             && index.index < self.input.num_files() + self.input.num_memtables()
     }
 
+    pub(crate) fn range_pre_filter_mode(&self, part_range: &PartitionRange) -> PreFilterMode {
+        let range_meta = &self.ranges[part_range.identifier];
+        let source_count = range_meta.indices.len();
+
+        self.input.range_pre_filter_mode(source_count)
+    }
+
     /// Retrieves the partition ranges.
     pub(crate) fn partition_ranges(&self) -> Vec<PartitionRange> {
         self.ranges
@@ -1601,8 +1612,7 @@ impl StreamContext {
                         .entries(self.input.files.iter().map(|file| FileWrapper { file }))
                         .finish()?;
                 }
-                write!(f, ", \"flat_format\": {}", self.input.flat_format)?;
-
+                write!(f, ", \"flat_format\": {}", self.input.explain_flat_format)?;
                 #[cfg(feature = "enterprise")]
                 self.format_extension_ranges(f)?;
 
@@ -1759,168 +1769,276 @@ impl PredicateGroup {
 
 #[cfg(test)]
 mod tests {
+    use std::collections::HashMap;
     use std::sync::Arc;
 
     use datafusion::physical_plan::expressions::lit as physical_lit;
+    use datafusion_common::ScalarValue;
     use datafusion_expr::{col, lit};
-    use store_api::storage::ScanRequest;
+    use datatypes::extension::json::{JsonExtensionType, JsonMetadata};
+    use datatypes::schema::ColumnSchema;
+    use datatypes::types::json_type::JsonObjectType;
+    use datatypes::value::Value;
+    use partition::expr::col as partition_col;
+    use store_api::metadata::{ColumnMetadata, RegionMetadataBuilder};
+    use store_api::storage::{RegionId, TimeSeriesDistribution, TimeSeriesRowSelector};
 
     use super::*;
-    use crate::memtable::time_partition::TimePartitions;
-    use crate::region::options::RegionOptions;
-    use crate::region::version::VersionBuilder;
-    use crate::sst::FormatType;
-    use crate::test_util::memtable_util::{EmptyMemtableBuilder, metadata_with_primary_key};
+    use crate::cache::CacheManager;
+    use crate::read::range_cache::ScanRequestFingerprintBuilder;
+    use crate::test_util::memtable_util::metadata_with_primary_key;
     use crate::test_util::scheduler_util::SchedulerEnv;
 
-    fn new_version(metadata: RegionMetadataRef) -> VersionRef {
-        let mutable = Arc::new(TimePartitions::new(
-            metadata.clone(),
-            Arc::new(EmptyMemtableBuilder::default()),
-            0,
-            None,
-        ));
-        Arc::new(VersionBuilder::new(metadata, mutable).build())
+    async fn new_scan_input(metadata: RegionMetadataRef, filters: Vec<Expr>) -> ScanInput {
+        let env = SchedulerEnv::new().await;
+        let mapper = FlatProjectionMapper::new(&metadata, [0, 2, 3]).unwrap();
+        let predicate = PredicateGroup::new(metadata.as_ref(), &filters).unwrap();
+        let file = FileHandle::new(
+            crate::sst::file::FileMeta::default(),
+            Arc::new(crate::sst::file_purger::NoopFilePurger),
+        );
+
+        ScanInput::new(env.access_layer.clone(), mapper)
+            .with_predicate(predicate)
+            .with_cache(CacheStrategy::EnableAll(Arc::new(
+                CacheManager::builder()
+                    .range_result_cache_size(1024)
+                    .build(),
+            )))
+            .with_files(vec![file])
     }
 
-    fn new_version_with_sst_format(
-        metadata: RegionMetadataRef,
-        sst_format: Option<FormatType>,
-    ) -> VersionRef {
-        let mutable = Arc::new(TimePartitions::new(
-            metadata.clone(),
-            Arc::new(EmptyMemtableBuilder::default()),
-            0,
-            None,
-        ));
-        let options = RegionOptions {
-            sst_format,
-            ..Default::default()
-        };
-        Arc::new(
-            VersionBuilder::new(metadata, mutable)
-                .options(options)
-                .build(),
-        )
+    /// Helper to create a timestamp millisecond literal.
+    fn ts_lit(val: i64) -> datafusion_expr::Expr {
+        lit(ScalarValue::TimestampMillisecond(Some(val), None))
+    }
+
+    fn metadata_with_json_field() -> RegionMetadataRef {
+        let mut json_column = ColumnSchema::new(
+            "payload",
+            ConcreteDataType::from(&JsonNativeType::Object(JsonObjectType::new())),
+            true,
+        );
+        json_column
+            .with_extension_type(&JsonExtensionType::new(Arc::new(JsonMetadata::default())))
+            .unwrap();
+
+        let mut builder = RegionMetadataBuilder::new(RegionId::new(1024, 1));
+        builder
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "host",
+                    ConcreteDataType::string_datatype(),
+                    false,
+                ),
+                semantic_type: SemanticType::Tag,
+                column_id: 1,
+            })
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "ts",
+                    ConcreteDataType::timestamp_millisecond_datatype(),
+                    false,
+                ),
+                semantic_type: SemanticType::Timestamp,
+                column_id: 2,
+            })
+            .push_column_metadata(ColumnMetadata {
+                column_schema: json_column,
+                semantic_type: SemanticType::Field,
+                column_id: 3,
+            })
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new("value", ConcreteDataType::int64_datatype(), true),
+                semantic_type: SemanticType::Field,
+                column_id: 4,
+            })
+            .primary_key(vec![1]);
+
+        Arc::new(builder.build().unwrap())
+    }
+
+    fn concrete_json_type_hint() -> JsonNativeType {
+        JsonNativeType::Object(JsonObjectType::from([
+            ("active".to_string(), JsonNativeType::Bool),
+            ("name".to_string(), JsonNativeType::String),
+        ]))
+    }
+
+    #[test]
+    fn test_concretize_json_types_rewrites_json_output_schema() -> Result<()> {
+        let metadata = metadata_with_json_field();
+        let mut mapper = FlatProjectionMapper::new(&metadata, [0, 2, 3])?;
+        let output_schema = mapper.output_schema();
+        let original_arrow_schema = output_schema.arrow_schema();
+        assert!(original_arrow_schema.has_json_extension_field());
+        assert!(is_json_extension_type(&original_arrow_schema.fields()[1]));
+
+        let expected_type = concrete_json_type_hint();
+        concretize_json_types(
+            &mut mapper,
+            &HashMap::from([("payload".to_string(), expected_type.clone())]),
+        );
+
+        let output_schema = mapper.output_schema();
+        let output_arrow_schema = output_schema.arrow_schema();
+        assert!(output_arrow_schema.has_json_extension_field());
+        assert_eq!(
+            ConcreteDataType::string_datatype(),
+            output_schema.column_schemas()[0].data_type
+        );
+        assert_eq!(
+            ConcreteDataType::from(&expected_type),
+            output_schema.column_schemas()[1].data_type
+        );
+        assert_eq!(
+            ConcreteDataType::int64_datatype(),
+            output_schema.column_schemas()[2].data_type
+        );
+        Ok(())
+    }
+
+    #[test]
+    fn test_concretize_json_types_keeps_json_without_hint() {
+        let metadata = metadata_with_json_field();
+        let mut mapper = FlatProjectionMapper::new(&metadata, [0, 2, 3]).unwrap();
+
+        concretize_json_types(
+            &mut mapper,
+            &HashMap::from([("missing".to_string(), JsonNativeType::String)]),
+        );
+
+        let output_schema = mapper.output_schema();
+        let output_arrow_schema = output_schema.arrow_schema();
+        assert!(output_arrow_schema.has_json_extension_field());
+        assert!(is_json_extension_type(&output_arrow_schema.fields()[1]));
+
+        // Assert that the expected JSON type stays un-concretized: empty object.
+        assert_eq!(
+            ConcreteDataType::from(&JsonNativeType::Object(JsonObjectType::new())),
+            output_schema.column_schemas()[1].data_type
+        );
     }
 
     #[tokio::test]
-    async fn test_build_read_column_ids_includes_filters() {
+    async fn test_build_scan_fingerprint_for_eligible_scan() {
         let metadata = Arc::new(metadata_with_primary_key(vec![0, 1], false));
-        let version = new_version(metadata.clone());
-        let env = SchedulerEnv::new().await;
-        let request = ScanRequest {
-            projection: Some(vec![4]),
-            filters: vec![
-                col("v0").gt(lit(1)),
-                col("ts").gt(lit(0)),
+        let input = new_scan_input(
+            metadata.clone(),
+            vec![
+                col("ts").gt_eq(ts_lit(1000)),
                 col("k0").eq(lit("foo")),
+                col("v0").gt(lit(1)),
             ],
-            ..Default::default()
-        };
-        let scan_region = ScanRegion::new(
-            version,
-            env.access_layer.clone(),
-            request,
-            CacheStrategy::Disabled,
-        );
-        let predicate =
-            PredicateGroup::new(metadata.as_ref(), &scan_region.request.filters).unwrap();
-        let projection = scan_region.request.projection.as_ref().unwrap();
-        let read_ids = scan_region
-            .build_read_column_ids(projection, &predicate)
-            .unwrap();
-        assert_eq!(vec![4, 0, 2, 3], read_ids);
+        )
+        .await
+        .with_distribution(Some(TimeSeriesDistribution::PerSeries))
+        .with_series_row_selector(Some(TimeSeriesRowSelector::LastRow))
+        .with_merge_mode(MergeMode::LastNonNull)
+        .with_filter_deleted(false);
+
+        let fingerprint = build_scan_fingerprint(&input).unwrap();
+
+        let expected = ScanRequestFingerprintBuilder {
+            read_columns: input.read_cols,
+            read_column_types: vec![
+                metadata
+                    .column_by_id(0)
+                    .map(|col| col.column_schema.data_type.clone()),
+                metadata
+                    .column_by_id(2)
+                    .map(|col| col.column_schema.data_type.clone()),
+                metadata
+                    .column_by_id(3)
+                    .map(|col| col.column_schema.data_type.clone()),
+            ],
+            filters: vec![
+                col("k0").eq(lit("foo")).to_string(),
+                col("v0").gt(lit(1)).to_string(),
+            ],
+            time_filters: vec![col("ts").gt_eq(ts_lit(1000)).to_string()],
+            series_row_selector: Some(TimeSeriesRowSelector::LastRow),
+            append_mode: false,
+            filter_deleted: false,
+            merge_mode: MergeMode::LastNonNull,
+            partition_expr_version: 0,
+        }
+        .build();
+        assert_eq!(expected, fingerprint);
     }
 
     #[tokio::test]
-    async fn test_build_read_column_ids_empty_projection() {
+    async fn test_build_scan_fingerprint_requires_tag_filter() {
         let metadata = Arc::new(metadata_with_primary_key(vec![0, 1], false));
-        let version = new_version(metadata.clone());
-        let env = SchedulerEnv::new().await;
-        let request = ScanRequest {
-            projection: Some(vec![]),
-            ..Default::default()
-        };
-        let scan_region = ScanRegion::new(
-            version,
-            env.access_layer.clone(),
-            request,
-            CacheStrategy::Disabled,
-        );
-        let predicate =
-            PredicateGroup::new(metadata.as_ref(), &scan_region.request.filters).unwrap();
-        let projection = scan_region.request.projection.as_ref().unwrap();
-        let read_ids = scan_region
-            .build_read_column_ids(projection, &predicate)
-            .unwrap();
-        // Empty projection should still read the time index column (id 2 in this test schema).
-        assert_eq!(vec![2], read_ids);
+        let input = new_scan_input(
+            metadata,
+            vec![col("ts").gt_eq(lit(1000)), col("v0").gt(lit(1))],
+        )
+        .await;
+
+        assert!(build_scan_fingerprint(&input).is_none());
     }
 
     #[tokio::test]
-    async fn test_build_read_column_ids_keeps_projection_order() {
+    async fn test_build_scan_fingerprint_respects_scan_eligibility() {
         let metadata = Arc::new(metadata_with_primary_key(vec![0, 1], false));
-        let version = new_version(metadata.clone());
-        let env = SchedulerEnv::new().await;
-        let request = ScanRequest {
-            projection: Some(vec![4, 1]),
-            filters: vec![col("v0").gt(lit(1))],
-            ..Default::default()
-        };
-        let scan_region = ScanRegion::new(
-            version,
-            env.access_layer.clone(),
-            request,
-            CacheStrategy::Disabled,
-        );
-        let predicate =
-            PredicateGroup::new(metadata.as_ref(), &scan_region.request.filters).unwrap();
-        let projection = scan_region.request.projection.as_ref().unwrap();
-        let read_ids = scan_region
-            .build_read_column_ids(projection, &predicate)
-            .unwrap();
-        // Projection order preserved, extra columns appended in schema order.
-        assert_eq!(vec![4, 1, 3], read_ids);
+        let filters = vec![col("k0").eq(lit("foo"))];
+
+        let disabled = ScanInput::new(
+            SchedulerEnv::new().await.access_layer.clone(),
+            FlatProjectionMapper::new(&metadata, [0, 2, 3].into_iter()).unwrap(),
+        )
+        .with_predicate(PredicateGroup::new(metadata.as_ref(), &filters).unwrap());
+        assert!(build_scan_fingerprint(&disabled).is_none());
+
+        let compaction = new_scan_input(metadata.clone(), filters.clone())
+            .await
+            .with_compaction(true);
+        assert!(build_scan_fingerprint(&compaction).is_none());
+
+        // No files to read.
+        let no_files = new_scan_input(metadata, filters).await.with_files(vec![]);
+        assert!(build_scan_fingerprint(&no_files).is_none());
     }
 
     #[tokio::test]
-    async fn test_use_flat_format_honors_request_override() {
-        let metadata = Arc::new(metadata_with_primary_key(vec![0, 1], false));
-        let env = SchedulerEnv::new().await;
+    async fn test_build_scan_fingerprint_tracks_schema_and_partition_expr_changes() {
+        let base = metadata_with_primary_key(vec![0, 1], false);
+        let mut builder = RegionMetadataBuilder::from_existing(base);
+        let partition_expr = partition_col("k0")
+            .gt_eq(Value::String("foo".into()))
+            .as_json_str()
+            .unwrap();
+        builder.partition_expr_json(Some(partition_expr));
+        let metadata = Arc::new(builder.build_without_validation().unwrap());
 
-        let primary_key_version =
-            new_version_with_sst_format(metadata.clone(), Some(FormatType::PrimaryKey));
-        let request = ScanRequest::default();
-        let scan_region = ScanRegion::new(
-            primary_key_version.clone(),
-            env.access_layer.clone(),
-            request,
-            CacheStrategy::Disabled,
-        );
-        assert!(!scan_region.use_flat_format());
+        let input = new_scan_input(metadata.clone(), vec![col("k0").eq(lit("foo"))]).await;
+        let fingerprint = build_scan_fingerprint(&input).unwrap();
 
-        let request = ScanRequest {
-            force_flat_format: true,
-            ..Default::default()
-        };
-        let scan_region = ScanRegion::new(
-            primary_key_version,
-            env.access_layer.clone(),
-            request,
-            CacheStrategy::Disabled,
-        );
-        assert!(scan_region.use_flat_format());
-
-        let flat_version = new_version_with_sst_format(metadata, Some(FormatType::Flat));
-        let request = ScanRequest::default();
-        let scan_region = ScanRegion::new(
-            flat_version,
-            env.access_layer.clone(),
-            request,
-            CacheStrategy::Disabled,
-        );
-        assert!(scan_region.use_flat_format());
+        let expected = ScanRequestFingerprintBuilder {
+            read_columns: input.read_cols,
+            read_column_types: vec![
+                metadata
+                    .column_by_id(0)
+                    .map(|col| col.column_schema.data_type.clone()),
+                metadata
+                    .column_by_id(2)
+                    .map(|col| col.column_schema.data_type.clone()),
+                metadata
+                    .column_by_id(3)
+                    .map(|col| col.column_schema.data_type.clone()),
+            ],
+            filters: vec![col("k0").eq(lit("foo")).to_string()],
+            time_filters: vec![],
+            series_row_selector: None,
+            append_mode: false,
+            filter_deleted: true,
+            merge_mode: MergeMode::LastRow,
+            partition_expr_version: metadata.partition_expr_version,
+        }
+        .build();
+        assert_eq!(expected, fingerprint);
+        assert_ne!(0, metadata.partition_expr_version);
     }
 
     #[test]
@@ -1941,4 +2059,24 @@ mod tests {
         assert!(predicate_without_region.exprs().is_empty());
         assert_eq!(1, predicate_without_region.dyn_filters().len());
     }
+
+    #[tokio::test]
+    async fn test_range_pre_filter_mode() {
+        let metadata = Arc::new(metadata_with_primary_key(vec![0, 1], false));
+        let cases = [
+            (true, MergeMode::LastRow, 1, PreFilterMode::All),
+            (false, MergeMode::LastNonNull, 1, PreFilterMode::All),
+            (false, MergeMode::LastRow, 2, PreFilterMode::SkipFields),
+            (true, MergeMode::LastRow, 2, PreFilterMode::All),
+        ];
+
+        for (append_mode, merge_mode, source_count, expected_mode) in cases {
+            let input = new_scan_input(metadata.clone(), vec![])
+                .await
+                .with_append_mode(append_mode)
+                .with_merge_mode(merge_mode);
+
+            assert_eq!(expected_mode, input.range_pre_filter_mode(source_count));
+        }
+    }
 }
diff --git a/src/mito2/src/read/scan_util.rs b/src/mito2/src/read/scan_util.rs
index 0ee6a4437d..5b7e46b0c1 100644
--- a/src/mito2/src/read/scan_util.rs
+++ b/src/mito2/src/read/scan_util.rs
@@ -29,30 +29,29 @@ use datatypes::timestamp::timestamp_array_to_primitive;
 use futures::Stream;
 use prometheus::IntGauge;
 use smallvec::SmallVec;
-use snafu::OptionExt;
 use store_api::storage::RegionId;
 
-use crate::error::{Result, UnexpectedSnafu};
+use crate::error::Result;
 use crate::memtable::MemScanMetrics;
 use crate::metrics::{
     IN_PROGRESS_SCAN, PRECISE_FILTER_ROWS_TOTAL, READ_BATCHES_RETURN, READ_ROW_GROUPS_TOTAL,
     READ_ROWS_IN_ROW_GROUP_TOTAL, READ_ROWS_RETURN, READ_STAGE_ELAPSED,
 };
 use crate::read::dedup::{DedupMetrics, DedupMetricsReport};
-use crate::read::merge::{MergeMetrics, MergeMetricsReport};
+use crate::read::flat_merge::{MergeMetrics, MergeMetricsReport};
 use crate::read::pruner::PartitionPruner;
 use crate::read::range::{RangeMeta, RowGroupIndex};
 use crate::read::scan_region::StreamContext;
-use crate::read::{Batch, BoxedBatchStream, BoxedRecordBatchStream, ScannerMetrics, Source};
+use crate::read::{BoxedRecordBatchStream, ScannerMetrics};
 use crate::sst::file::{FileTimeRange, RegionFileId};
 use crate::sst::index::bloom_filter::applier::BloomFilterIndexApplyMetrics;
 use crate::sst::index::fulltext_index::applier::FulltextIndexApplyMetrics;
 use crate::sst::index::inverted_index::applier::InvertedIndexApplyMetrics;
-use crate::sst::parquet::DEFAULT_ROW_GROUP_SIZE;
-use crate::sst::parquet::file_range::FileRange;
+use crate::sst::parquet::file_range::{FileRange, PreFilterMode};
 use crate::sst::parquet::flat_format::time_index_column_index;
 use crate::sst::parquet::reader::{MetadataCacheMetrics, ReaderFilterMetrics, ReaderMetrics};
 use crate::sst::parquet::row_group::ParquetFetchMetrics;
+use crate::sst::parquet::{DEFAULT_READ_BATCH_SIZE, DEFAULT_ROW_GROUP_SIZE};
 
 /// Per-file scan metrics.
 #[derive(Default, Clone)]
@@ -138,6 +137,10 @@ pub(crate) struct ScanMetricsSet {
     mem_batches: usize,
     /// Number of series read from memtables.
     mem_series: usize,
+    /// Duration of prefilter in memtable scan.
+    mem_prefilter_cost: Duration,
+    /// Number of rows filtered by prefilter in memtable scan.
+    mem_prefilter_rows_filtered: usize,
 
     // SST related metrics:
     /// Duration to build file ranges.
@@ -247,6 +250,12 @@ pub(crate) struct ScanMetricsSet {
     num_range_builders: isize,
     /// Peak number of file range builders.
     num_peak_range_builders: isize,
+    /// Total bytes added to the range cache during this scan.
+    range_cache_size: usize,
+    /// Number of range cache hits during this scan.
+    range_cache_hit: usize,
+    /// Number of range cache misses during this scan.
+    range_cache_miss: usize,
 }
 
 /// Wrapper for file metrics that compares by total cost in reverse order.
@@ -335,6 +344,8 @@ impl fmt::Debug for ScanMetricsSet {
             mem_rows,
             mem_batches,
             mem_series,
+            mem_prefilter_cost,
+            mem_prefilter_rows_filtered,
             inverted_index_apply_metrics,
             bloom_filter_apply_metrics,
             fulltext_index_apply_metrics,
@@ -345,6 +356,9 @@ impl fmt::Debug for ScanMetricsSet {
             build_ranges_peak_mem_size,
             num_range_builders: _,
             num_peak_range_builders,
+            range_cache_size,
+            range_cache_hit,
+            range_cache_miss,
         } = self;
 
         // Write core metrics
@@ -500,6 +514,15 @@ impl fmt::Debug for ScanMetricsSet {
         if !mem_scan_cost.is_zero() {
             write!(f, ", \"mem_scan_cost\":\"{mem_scan_cost:?}\"")?;
         }
+        if !mem_prefilter_cost.is_zero() {
+            write!(f, ", \"mem_prefilter_cost\":\"{mem_prefilter_cost:?}\"")?;
+        }
+        if *mem_prefilter_rows_filtered > 0 {
+            write!(
+                f,
+                ", \"mem_prefilter_rows_filtered\":{mem_prefilter_rows_filtered}"
+            )?;
+        }
 
         // Write optional verbose metrics if they are not empty
         if let Some(metrics) = inverted_index_apply_metrics
@@ -590,6 +613,16 @@ impl fmt::Debug for ScanMetricsSet {
             write!(f, "}}")?;
         }
 
+        if *range_cache_size > 0 {
+            write!(f, ", \"range_cache_size\":{range_cache_size}")?;
+        }
+        if *range_cache_hit > 0 {
+            write!(f, ", \"range_cache_hit\":{range_cache_hit}")?;
+        }
+        if *range_cache_miss > 0 {
+            write!(f, ", \"range_cache_miss\":{range_cache_miss}")?;
+        }
+
         write!(
             f,
             ", \"build_ranges_peak_mem_size\":{build_ranges_peak_mem_size}, \
@@ -1042,6 +1075,8 @@ impl PartitionMetrics {
         metrics.mem_rows += data.num_rows;
         metrics.mem_batches += data.num_batches;
         metrics.mem_series += data.total_series;
+        metrics.mem_prefilter_cost += data.prefilter_cost;
+        metrics.mem_prefilter_rows_filtered += data.prefilter_rows_filtered;
     }
 
     /// Merges [ScannerMetrics], `build_reader_cost`, `scan_cost` and `yield_cost`.
@@ -1097,6 +1132,27 @@ impl PartitionMetrics {
     pub(crate) fn dedup_metrics_reporter(&self) -> Arc<dyn DedupMetricsReport> {
         self.0.clone()
     }
+
+    /// Increments the total bytes added to the range cache.
+    #[allow(dead_code)]
+    pub(crate) fn inc_range_cache_size(&self, size: usize) {
+        let mut metrics = self.0.metrics.lock().unwrap();
+        metrics.range_cache_size += size;
+    }
+
+    /// Increments the range cache hit counter.
+    #[allow(dead_code)]
+    pub(crate) fn inc_range_cache_hit(&self) {
+        let mut metrics = self.0.metrics.lock().unwrap();
+        metrics.range_cache_hit += 1;
+    }
+
+    /// Increments the range cache miss counter.
+    #[allow(dead_code)]
+    pub(crate) fn inc_range_cache_miss(&self) {
+        let mut metrics = self.0.metrics.lock().unwrap();
+        metrics.range_cache_miss += 1;
+    }
 }
 
 impl fmt::Debug for PartitionMetrics {
@@ -1129,45 +1185,6 @@ pub(crate) struct SeriesDistributorMetrics {
     pub(crate) divider_cost: Duration,
 }
 
-/// Scans memtable ranges at `index`.
-#[tracing::instrument(
-    skip_all,
-    fields(
-        region_id = %stream_ctx.input.region_metadata().region_id,
-        file_or_mem_index = %index.index,
-        row_group_index = %index.row_group_index,
-        source = "mem"
-    )
-)]
-pub(crate) fn scan_mem_ranges(
-    stream_ctx: Arc<StreamContext>,
-    part_metrics: PartitionMetrics,
-    index: RowGroupIndex,
-    time_range: FileTimeRange,
-) -> impl Stream<Item = Result<Batch>> {
-    try_stream! {
-        let ranges = stream_ctx.input.build_mem_ranges(index);
-        part_metrics.inc_num_mem_ranges(ranges.len());
-        for range in ranges {
-            let build_reader_start = Instant::now();
-            let mem_scan_metrics = Some(MemScanMetrics::default());
-            let iter = range.build_prune_iter(time_range, mem_scan_metrics.clone())?;
-            part_metrics.inc_build_reader_cost(build_reader_start.elapsed());
-
-            let mut source = Source::Iter(iter);
-            while let Some(batch) = source.next_batch().await? {
-                yield batch;
-            }
-
-            // Report the memtable scan metrics to partition metrics
-            if let Some(ref metrics) = mem_scan_metrics {
-                let data = metrics.data();
-                part_metrics.report_mem_scan_metrics(&data);
-            }
-        }
-    }
-}
-
 /// Scans memtable ranges at `index` using flat format that returns RecordBatch.
 #[tracing::instrument(
     skip_all,
@@ -1213,15 +1230,19 @@ const NUM_SERIES_THRESHOLD: u64 = 10240;
 /// 60 samples per hour.
 const BATCH_SIZE_THRESHOLD: u64 = 50;
 
-/// Returns true if splitting flat record batches may improve merge performance.
+/// Returns the estimated rows per batch after splitting if splitting flat record batches
+/// may improve merge performance. Returns `None` if splitting is not beneficial.
 pub(crate) fn should_split_flat_batches_for_merge(
     stream_ctx: &Arc<StreamContext>,
     range_meta: &RangeMeta,
-) -> bool {
+) -> Option<usize> {
     // Number of files to split and scan.
     let mut num_files_to_split = 0;
     let mut num_mem_rows = 0;
     let mut num_mem_series = 0;
+    // Total rows and series for estimating batch size after splitting.
+    let mut total_rows: u64 = 0;
+    let mut total_series: u64 = 0;
     // Checks each file range, returns early if any range is not splittable.
     // For mem ranges, we collect the total number of rows and series because the number of rows in a
     // mem range may be too small.
@@ -1236,40 +1257,172 @@ pub(crate) fn should_split_flat_batches_for_merge(
             // This is a file range.
             let file_index = index.index - stream_ctx.input.num_memtables();
             let file = &stream_ctx.input.files[file_index];
-            if file.meta_ref().num_rows < SPLIT_ROW_THRESHOLD || file.meta_ref().num_series == 0 {
+            let file_meta = file.meta_ref();
+            if file_meta.level == 0 {
+                // Always split level 0 files.
+                num_files_to_split += 1;
+                continue;
+            } else if file_meta.num_rows < SPLIT_ROW_THRESHOLD || file_meta.num_series == 0 {
                 // If the file doesn't have enough rows, or the number of series is unavailable, skips it.
                 continue;
             }
-            debug_assert!(file.meta_ref().num_rows > 0);
-            if !can_split_series(file.meta_ref().num_rows, file.meta_ref().num_series) {
+            debug_assert!(file_meta.num_rows > 0);
+            if !can_split_series(file_meta.num_rows, file_meta.num_series) {
                 // We can't split batches in a file.
-                return false;
+                common_telemetry::trace!(
+                    "Can't split series for file {}, level: {}, num_rows: {}, num_series: {}",
+                    file_meta.file_id,
+                    file_meta.level,
+                    file_meta.num_rows,
+                    file_meta.num_series,
+                );
+                return None;
             } else {
                 num_files_to_split += 1;
+                total_rows += file.meta_ref().num_rows;
+                total_series += file.meta_ref().num_series;
             }
         }
         // Skips non-file and non-mem ranges.
     }
 
-    if num_files_to_split > 0 {
+    let should_split = if num_files_to_split > 0 {
         // We mainly consider file ranges because they have enough data for sampling.
         true
-    } else if num_mem_series > 0 && num_mem_rows > 0 {
-        // If we don't have files to scan, we check whether to split by the memtable.
-        can_split_series(num_mem_rows as u64, num_mem_series as u64)
+    } else if num_mem_series > 0
+        && num_mem_rows > 0
+        && can_split_series(num_mem_rows as u64, num_mem_series as u64)
+    {
+        total_rows += num_mem_rows as u64;
+        total_series += num_mem_series as u64;
+        true
     } else {
         false
+    };
+
+    if !should_split {
+        return None;
     }
+
+    // Estimate rows per batch after splitting.
+    let estimated_batch_size = if total_series > 0 && total_rows > 0 {
+        ((total_rows / total_series) as usize).clamp(1, DEFAULT_READ_BATCH_SIZE)
+    } else {
+        // No valid estimate available, use a conservative fallback.
+        DEFAULT_READ_BATCH_SIZE / 4
+    };
+    Some(estimated_batch_size)
+}
+
+/// Computes the channel size for parallel scan based on the estimated rows per batch.
+/// The channel should buffer approximately `2 * DEFAULT_READ_BATCH_SIZE` rows.
+pub(crate) fn compute_parallel_channel_size(estimated_rows_per_batch: usize) -> usize {
+    let size = 2 * DEFAULT_READ_BATCH_SIZE / estimated_rows_per_batch.max(1);
+    size.clamp(2, 64)
+}
+
+/// Computes the average estimated rows per batch across multiple range readers.
+pub(crate) fn compute_average_batch_size(
+    estimated_rows_per_batch: impl IntoIterator<Item = usize>,
+) -> usize {
+    let mut total = 0usize;
+    let mut count = 0usize;
+    for size in estimated_rows_per_batch {
+        total += size;
+        count += 1;
+    }
+
+    if count == 0 {
+        return DEFAULT_READ_BATCH_SIZE;
+    }
+
+    (total / count).clamp(1, DEFAULT_READ_BATCH_SIZE)
 }
 
 fn can_split_series(num_rows: u64, num_series: u64) -> bool {
-    assert!(num_series > 0);
-    assert!(num_rows > 0);
+    if num_rows == 0 || num_series == 0 {
+        return false;
+    }
 
     // It doesn't have too many series or it will have enough rows for each batch.
     num_series < NUM_SERIES_THRESHOLD || num_rows / num_series >= BATCH_SIZE_THRESHOLD
 }
 
+#[cfg(test)]
+mod split_tests {
+    use std::sync::Arc;
+
+    use common_time::Timestamp;
+    use smallvec::smallvec;
+    use store_api::storage::FileId;
+
+    use super::*;
+    use crate::read::flat_projection::FlatProjectionMapper;
+    use crate::read::range::{RangeMeta, RowGroupIndex, SourceIndex};
+    use crate::read::scan_region::{ScanInput, StreamContext};
+    use crate::sst::file::FileHandle;
+    use crate::test_util::memtable_util::metadata_with_primary_key;
+    use crate::test_util::scheduler_util::SchedulerEnv;
+    use crate::test_util::sst_util::sst_file_handle_with_file_id;
+
+    async fn new_stream_context_with_files(files: Vec<FileHandle>) -> StreamContext {
+        let env = SchedulerEnv::new().await;
+        let metadata = Arc::new(metadata_with_primary_key(vec![0, 1], false));
+        let mapper = FlatProjectionMapper::new(&metadata, [0, 2, 3]).unwrap();
+        let input = ScanInput::new(env.access_layer.clone(), mapper).with_files(files);
+
+        StreamContext {
+            input,
+            ranges: vec![],
+            scan_fingerprint: None,
+            query_start: std::time::Instant::now(),
+        }
+    }
+
+    fn single_file_range_meta() -> RangeMeta {
+        RangeMeta {
+            time_range: (
+                Timestamp::new_millisecond(0),
+                Timestamp::new_millisecond(1000),
+            ),
+            indices: smallvec![SourceIndex {
+                index: 0,
+                num_row_groups: 1,
+            }],
+            row_group_indices: smallvec![RowGroupIndex {
+                index: 0,
+                row_group_index: 0,
+            }],
+            num_rows: 1024,
+        }
+    }
+
+    #[tokio::test]
+    async fn should_split_level_zero_file_even_when_series_stats_are_missing() {
+        let mut file = sst_file_handle_with_file_id(FileId::random(), 0, 1000)
+            .meta_ref()
+            .clone();
+        file.level = 0;
+        file.num_rows = DEFAULT_ROW_GROUP_SIZE as u64;
+        file.num_row_groups = 1;
+        file.num_series = 0;
+
+        let file = FileHandle::new(file, crate::test_util::new_noop_file_purger());
+        let stream_ctx = Arc::new(new_stream_context_with_files(vec![file]).await);
+
+        assert!(
+            should_split_flat_batches_for_merge(&stream_ctx, &single_file_range_meta()).is_some()
+        );
+    }
+
+    #[test]
+    fn can_split_series_returns_false_for_zero_inputs() {
+        assert!(!can_split_series(0, 1));
+        assert!(!can_split_series(1, 0));
+        assert!(!can_split_series(0, 0));
+    }
+}
+
 /// Creates a new [ReaderFilterMetrics] with optional apply metrics initialized
 /// based on the `explain_verbose` flag.
 fn new_filter_metrics(explain_verbose: bool) -> ReaderFilterMetrics {
@@ -1285,59 +1438,6 @@ fn new_filter_metrics(explain_verbose: bool) -> ReaderFilterMetrics {
     }
 }
 
-/// Scans file ranges at `index`.
-#[tracing::instrument(
-    skip_all,
-    fields(
-        region_id = %stream_ctx.input.region_metadata().region_id,
-        row_group_index = %index.index,
-        source = read_type
-    )
-)]
-pub(crate) async fn scan_file_ranges(
-    stream_ctx: Arc<StreamContext>,
-    part_metrics: PartitionMetrics,
-    index: RowGroupIndex,
-    read_type: &'static str,
-    partition_pruner: Arc<PartitionPruner>,
-) -> Result<impl Stream<Item = Result<Batch>>> {
-    let mut reader_metrics = ReaderMetrics {
-        filter_metrics: new_filter_metrics(part_metrics.explain_verbose()),
-        ..Default::default()
-    };
-    let ranges = partition_pruner
-        .build_file_ranges(index, &part_metrics, &mut reader_metrics)
-        .await?;
-    part_metrics.inc_num_file_ranges(ranges.len());
-    part_metrics.merge_reader_metrics(&reader_metrics, None);
-
-    // Creates initial per-file metrics with build_part_cost.
-    let init_per_file_metrics = if part_metrics.explain_verbose() {
-        let file = stream_ctx.input.file_from_index(index);
-        let file_id = file.file_id();
-
-        let mut map = HashMap::new();
-        map.insert(
-            file_id,
-            FileScanMetrics {
-                build_part_cost: reader_metrics.build_cost,
-                ..Default::default()
-            },
-        );
-        Some(map)
-    } else {
-        None
-    };
-
-    Ok(build_file_range_scan_stream(
-        stream_ctx,
-        part_metrics,
-        read_type,
-        ranges,
-        init_per_file_metrics,
-    ))
-}
-
 /// Scans file ranges at `index` using flat reader that returns RecordBatch.
 #[tracing::instrument(
     skip_all,
@@ -1391,77 +1491,13 @@ pub(crate) async fn scan_flat_file_ranges(
     ))
 }
 
-/// Build the stream of scanning the input [`FileRange`]s.
-#[tracing::instrument(
-    skip_all,
-    fields(read_type = read_type, range_count = ranges.len())
-)]
-pub fn build_file_range_scan_stream(
-    stream_ctx: Arc<StreamContext>,
-    part_metrics: PartitionMetrics,
-    read_type: &'static str,
-    ranges: SmallVec<[FileRange; 2]>,
-    mut per_file_metrics: Option<HashMap<RegionFileId, FileScanMetrics>>,
-) -> impl Stream<Item = Result<Batch>> {
-    try_stream! {
-        let fetch_metrics = if part_metrics.explain_verbose() {
-            Some(Arc::new(ParquetFetchMetrics::default()))
-        } else {
-            None
-        };
-        let reader_metrics = &mut ReaderMetrics {
-            fetch_metrics: fetch_metrics.clone(),
-            ..Default::default()
-        };
-        for range in ranges {
-            let build_reader_start = Instant::now();
-            let Some(reader) = range.reader(stream_ctx.input.series_row_selector, fetch_metrics.as_deref()).await? else {
-                continue;
-            };
-            let build_cost = build_reader_start.elapsed();
-            part_metrics.inc_build_reader_cost(build_cost);
-            let compat_batch = range.compat_batch();
-            let mut source = Source::PruneReader(reader);
-            while let Some(mut batch) = source.next_batch().await? {
-                if let Some(compact_batch) = compat_batch {
-                    batch = compact_batch.as_primary_key().unwrap().compat_batch(batch)?;
-                }
-                yield batch;
-            }
-            if let Source::PruneReader(reader) = source {
-                let prune_metrics = reader.metrics();
-
-                // Update per-file metrics if tracking is enabled
-                if let Some(file_metrics_map) = per_file_metrics.as_mut() {
-                    let file_id = range.file_handle().file_id();
-                    let file_metrics = file_metrics_map
-                        .entry(file_id)
-                        .or_insert_with(FileScanMetrics::default);
-
-                    file_metrics.num_ranges += 1;
-                    file_metrics.num_rows += prune_metrics.num_rows;
-                    file_metrics.build_reader_cost += build_cost;
-                    file_metrics.scan_cost += prune_metrics.scan_cost;
-                }
-
-                reader_metrics.merge_from(&prune_metrics);
-            }
-        }
-
-        // Reports metrics.
-        reader_metrics.observe_rows(read_type);
-        reader_metrics.filter_metrics.observe();
-        part_metrics.merge_reader_metrics(reader_metrics, per_file_metrics.as_ref());
-    }
-}
-
 /// Build the stream of scanning the input [`FileRange`]s using flat reader that returns RecordBatch.
 #[tracing::instrument(
     skip_all,
     fields(read_type = read_type, range_count = ranges.len())
 )]
 pub fn build_flat_file_range_scan_stream(
-    _stream_ctx: Arc<StreamContext>,
+    stream_ctx: Arc<StreamContext>,
     part_metrics: PartitionMetrics,
     read_type: &'static str,
     ranges: SmallVec<[FileRange; 2]>,
@@ -1479,21 +1515,22 @@ pub fn build_flat_file_range_scan_stream(
         };
         for range in ranges {
             let build_reader_start = Instant::now();
-            let Some(mut reader) = range.flat_reader(_stream_ctx.input.series_row_selector, fetch_metrics.as_deref()).await? else{continue};
+            let Some(mut reader) = range
+                .flat_reader(
+                    stream_ctx.input.series_row_selector,
+                    fetch_metrics.as_deref(),
+                )
+                .await?
+            else {
+                continue;
+            };
             let build_cost = build_reader_start.elapsed();
             part_metrics.inc_build_reader_cost(build_cost);
 
-            let may_compat = range
-                .compat_batch()
-                .map(|compat| {
-                    compat.as_flat().context(UnexpectedSnafu {
-                        reason: "Invalid compat for flat format",
-                    })
-                })
-                .transpose()?;
+            let may_compat = range.compat_batch();
 
             let mapper = range.compaction_projection_mapper();
-            while let Some(record_batch) = reader.next_batch()? {
+            while let Some(record_batch) = reader.next_batch().await? {
                 let record_batch = if let Some(mapper) = mapper {
                     let batch = mapper.project(record_batch)?;
                     batch
@@ -1534,58 +1571,18 @@ pub fn build_flat_file_range_scan_stream(
     }
 }
 
-/// Build the stream of scanning the extension range denoted by the [`RowGroupIndex`].
-#[cfg(feature = "enterprise")]
-pub(crate) async fn scan_extension_range(
-    context: Arc<StreamContext>,
-    index: RowGroupIndex,
-    partition_metrics: PartitionMetrics,
-) -> Result<BoxedBatchStream> {
-    use snafu::ResultExt;
-
-    let range = context.input.extension_range(index.index);
-    let reader = range.reader(context.as_ref());
-    let stream = reader
-        .read(context, partition_metrics, index)
-        .await
-        .context(crate::error::ScanExternalRangeSnafu)?;
-    Ok(stream)
-}
-
-pub(crate) async fn maybe_scan_other_ranges(
-    context: &Arc<StreamContext>,
-    index: RowGroupIndex,
-    metrics: &PartitionMetrics,
-) -> Result<BoxedBatchStream> {
-    #[cfg(feature = "enterprise")]
-    {
-        scan_extension_range(context.clone(), index, metrics.clone()).await
-    }
-
-    #[cfg(not(feature = "enterprise"))]
-    {
-        let _ = context;
-        let _ = index;
-        let _ = metrics;
-
-        crate::error::UnexpectedSnafu {
-            reason: "no other ranges scannable",
-        }
-        .fail()
-    }
-}
-
 /// Build the stream of scanning the extension range in flat format denoted by the [`RowGroupIndex`].
 #[cfg(feature = "enterprise")]
 pub(crate) async fn scan_flat_extension_range(
     context: Arc<StreamContext>,
     index: RowGroupIndex,
     partition_metrics: PartitionMetrics,
+    options: crate::extension::ExtensionRangeReadOptions,
 ) -> Result<BoxedRecordBatchStream> {
     use snafu::ResultExt;
 
     let range = context.input.extension_range(index.index);
-    let reader = range.flat_reader(context.as_ref());
+    let reader = range.flat_reader(context.as_ref(), options);
     let stream = reader
         .read(context, partition_metrics, index)
         .await
@@ -1597,10 +1594,12 @@ pub(crate) async fn maybe_scan_flat_other_ranges(
     context: &Arc<StreamContext>,
     index: RowGroupIndex,
     metrics: &PartitionMetrics,
+    pre_filter_mode: PreFilterMode,
 ) -> Result<BoxedRecordBatchStream> {
     #[cfg(feature = "enterprise")]
     {
-        scan_flat_extension_range(context.clone(), index, metrics.clone()).await
+        let options = crate::extension::ExtensionRangeReadOptions { pre_filter_mode };
+        scan_flat_extension_range(context.clone(), index, metrics.clone(), options).await
     }
 
     #[cfg(not(feature = "enterprise"))]
@@ -1608,6 +1607,7 @@ pub(crate) async fn maybe_scan_flat_other_ranges(
         let _ = context;
         let _ = index;
         let _ = metrics;
+        let _ = pre_filter_mode;
 
         crate::error::UnexpectedSnafu {
             reason: "no other ranges scannable in flat format",
@@ -1695,3 +1695,258 @@ pub(crate) fn split_record_batch(record_batch: RecordBatch, batches: &mut VecDeq
         batches.push_back(record_batch.slice(start, rows_in_batch));
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+    use std::time::Instant;
+
+    use common_time::Timestamp;
+    use smallvec::{SmallVec, smallvec};
+    use store_api::storage::RegionId;
+
+    use super::*;
+    use crate::cache::CacheStrategy;
+    use crate::memtable::{
+        BoxedBatchIterator, BoxedRecordBatchIterator, IterBuilder, MemtableRange,
+        MemtableRangeContext, MemtableStats,
+    };
+    use crate::read::flat_projection::FlatProjectionMapper;
+    use crate::read::range::{MemRangeBuilder, SourceIndex};
+    use crate::read::scan_region::ScanInput;
+    use crate::sst::file::{FileHandle, FileMeta};
+    use crate::sst::file_purger::NoopFilePurger;
+    use crate::test_util::memtable_util::metadata_for_test;
+    use crate::test_util::scheduler_util::SchedulerEnv;
+
+    struct EmptyIterBuilder;
+
+    impl IterBuilder for EmptyIterBuilder {
+        fn build(&self, _metrics: Option<MemScanMetrics>) -> Result<BoxedBatchIterator> {
+            Ok(Box::new(std::iter::empty()))
+        }
+
+        fn is_record_batch(&self) -> bool {
+            true
+        }
+
+        fn build_record_batch(
+            &self,
+            _time_range: Option<(Timestamp, Timestamp)>,
+            _metrics: Option<MemScanMetrics>,
+        ) -> Result<BoxedRecordBatchIterator> {
+            Ok(Box::new(std::iter::empty()))
+        }
+    }
+
+    async fn new_test_stream_ctx(
+        files: Vec<FileHandle>,
+        memtables: Vec<MemRangeBuilder>,
+    ) -> Arc<StreamContext> {
+        let env = SchedulerEnv::new().await;
+        let metadata = metadata_for_test();
+        let mapper = FlatProjectionMapper::new(&metadata, [0, 2, 3]).unwrap();
+        let input = ScanInput::new(env.access_layer.clone(), mapper)
+            .with_cache(CacheStrategy::Disabled)
+            .with_memtables(memtables)
+            .with_files(files);
+
+        Arc::new(StreamContext {
+            input,
+            ranges: Vec::new(),
+            scan_fingerprint: None,
+            query_start: Instant::now(),
+        })
+    }
+
+    fn new_test_file(num_rows: u64, num_series: u64) -> FileHandle {
+        let meta = FileMeta {
+            region_id: RegionId::new(123, 456),
+            file_id: Default::default(),
+            level: 1,
+            time_range: (
+                Timestamp::new_millisecond(0),
+                Timestamp::new_millisecond(1000),
+            ),
+            num_rows,
+            num_series,
+            ..Default::default()
+        };
+        FileHandle::new(meta, Arc::new(NoopFilePurger))
+    }
+
+    fn new_test_memtable(num_rows: usize, series_count: usize) -> MemRangeBuilder {
+        let context = Arc::new(MemtableRangeContext::new(
+            0,
+            Box::new(EmptyIterBuilder),
+            Default::default(),
+        ));
+        let stats = MemtableStats {
+            time_range: Some((
+                Timestamp::new_millisecond(0),
+                Timestamp::new_millisecond(1000),
+            )),
+            num_rows,
+            num_ranges: 1,
+            series_count,
+            ..Default::default()
+        };
+        let range = MemtableRange::new(context, stats.clone());
+        MemRangeBuilder::new(range, stats)
+    }
+
+    fn new_test_range_meta(row_group_indices: SmallVec<[RowGroupIndex; 2]>) -> RangeMeta {
+        let indices = row_group_indices
+            .iter()
+            .map(|row_group_index| SourceIndex {
+                index: row_group_index.index,
+                num_row_groups: 1,
+            })
+            .collect();
+
+        RangeMeta {
+            time_range: (
+                Timestamp::new_millisecond(0),
+                Timestamp::new_millisecond(1000),
+            ),
+            indices,
+            row_group_indices,
+            num_rows: 0,
+        }
+    }
+
+    #[tokio::test]
+    async fn test_should_split_flat_batches_for_merge_uses_splittable_file_rows_per_series() {
+        let num_rows = SPLIT_ROW_THRESHOLD * 2;
+        let num_series = (num_rows / 100).max(1);
+        let stream_ctx =
+            new_test_stream_ctx(vec![new_test_file(num_rows, num_series)], vec![]).await;
+        let range_meta = new_test_range_meta(smallvec![RowGroupIndex {
+            index: 0,
+            row_group_index: 0,
+        }]);
+
+        assert_eq!(
+            Some((num_rows / num_series) as usize),
+            should_split_flat_batches_for_merge(&stream_ctx, &range_meta)
+        );
+    }
+
+    #[tokio::test]
+    async fn test_should_split_flat_batches_for_merge_skips_small_or_unknown_series_files() {
+        let stream_ctx = new_test_stream_ctx(
+            vec![
+                new_test_file(SPLIT_ROW_THRESHOLD.saturating_sub(1), 1),
+                new_test_file(SPLIT_ROW_THRESHOLD * 2, 0),
+            ],
+            vec![],
+        )
+        .await;
+        let range_meta = new_test_range_meta(smallvec![
+            RowGroupIndex {
+                index: 0,
+                row_group_index: 0,
+            },
+            RowGroupIndex {
+                index: 1,
+                row_group_index: 0,
+            }
+        ]);
+
+        assert_eq!(
+            None,
+            should_split_flat_batches_for_merge(&stream_ctx, &range_meta)
+        );
+    }
+
+    #[tokio::test]
+    async fn test_should_split_flat_batches_for_merge_returns_none_for_unsplittable_file() {
+        let num_series =
+            (SPLIT_ROW_THRESHOLD / (BATCH_SIZE_THRESHOLD - 1)).max(NUM_SERIES_THRESHOLD) + 1;
+        let stream_ctx =
+            new_test_stream_ctx(vec![new_test_file(SPLIT_ROW_THRESHOLD, num_series)], vec![]).await;
+        let range_meta = new_test_range_meta(smallvec![RowGroupIndex {
+            index: 0,
+            row_group_index: 0,
+        }]);
+
+        assert_eq!(
+            None,
+            should_split_flat_batches_for_merge(&stream_ctx, &range_meta)
+        );
+    }
+
+    #[tokio::test]
+    async fn test_should_split_flat_batches_for_merge_falls_back_to_memtables() {
+        let stream_ctx = new_test_stream_ctx(vec![], vec![new_test_memtable(5_000, 100)]).await;
+        let range_meta = new_test_range_meta(smallvec![RowGroupIndex {
+            index: 0,
+            row_group_index: 0,
+        }]);
+
+        assert_eq!(
+            Some(50),
+            should_split_flat_batches_for_merge(&stream_ctx, &range_meta)
+        );
+    }
+
+    #[tokio::test]
+    async fn test_should_split_flat_batches_for_merge_clamps_estimate() {
+        let stream_ctx =
+            new_test_stream_ctx(vec![new_test_file(SPLIT_ROW_THRESHOLD * 2, 1)], vec![]).await;
+        let range_meta = new_test_range_meta(smallvec![RowGroupIndex {
+            index: 0,
+            row_group_index: 0,
+        }]);
+
+        assert_eq!(
+            Some(DEFAULT_READ_BATCH_SIZE),
+            should_split_flat_batches_for_merge(&stream_ctx, &range_meta)
+        );
+    }
+
+    #[test]
+    fn test_compute_parallel_channel_size_clamps_to_max_for_small_batches() {
+        assert_eq!(64, compute_parallel_channel_size(0));
+        assert_eq!(64, compute_parallel_channel_size(1));
+    }
+
+    #[test]
+    fn test_compute_parallel_channel_size_returns_expected_mid_range_size() {
+        assert_eq!(
+            4,
+            compute_parallel_channel_size(DEFAULT_READ_BATCH_SIZE / 2)
+        );
+    }
+
+    #[test]
+    fn test_compute_parallel_channel_size_clamps_to_min_for_large_batches() {
+        assert_eq!(2, compute_parallel_channel_size(DEFAULT_READ_BATCH_SIZE));
+        assert_eq!(
+            2,
+            compute_parallel_channel_size(DEFAULT_READ_BATCH_SIZE * 2)
+        );
+    }
+
+    #[test]
+    fn test_compute_average_batch_size_uses_arithmetic_mean() {
+        assert_eq!(24, compute_average_batch_size([16, 24, 32]));
+    }
+
+    #[test]
+    fn test_compute_average_batch_size_clamps_values() {
+        assert_eq!(
+            DEFAULT_READ_BATCH_SIZE,
+            compute_average_batch_size([DEFAULT_READ_BATCH_SIZE, DEFAULT_READ_BATCH_SIZE * 2])
+        );
+        assert_eq!(1, compute_average_batch_size([0, 1]));
+    }
+
+    #[test]
+    fn test_compute_average_batch_size_falls_back_when_empty() {
+        assert_eq!(
+            DEFAULT_READ_BATCH_SIZE,
+            compute_average_batch_size(std::iter::empty())
+        );
+    }
+}
diff --git a/src/mito2/src/read/seq_scan.rs b/src/mito2/src/read/seq_scan.rs
index c13b40d111..432099dbcf 100644
--- a/src/mito2/src/read/seq_scan.rs
+++ b/src/mito2/src/read/seq_scan.rs
@@ -27,7 +27,7 @@ use datafusion::physical_plan::metrics::ExecutionPlanMetricsSet;
 use datafusion::physical_plan::{DisplayAs, DisplayFormatType};
 use datatypes::schema::SchemaRef;
 use futures::{StreamExt, TryStreamExt};
-use snafu::{OptionExt, ensure};
+use snafu::ensure;
 use store_api::metadata::RegionMetadataRef;
 use store_api::region_engine::{
     PartitionRange, PrepareRequest, QueryScanContext, RegionScanner, ScannerProperties,
@@ -35,24 +35,22 @@ use store_api::region_engine::{
 use store_api::storage::TimeSeriesRowSelector;
 use tokio::sync::Semaphore;
 
-use crate::error::{PartitionOutOfRangeSnafu, Result, TooManyFilesToReadSnafu, UnexpectedSnafu};
-use crate::read::dedup::{DedupReader, LastNonNull, LastRow};
+use crate::error::{PartitionOutOfRangeSnafu, Result, TooManyFilesToReadSnafu};
 use crate::read::flat_dedup::{FlatDedupReader, FlatLastNonNull, FlatLastRow};
 use crate::read::flat_merge::FlatMergeReader;
-use crate::read::last_row::LastRowReader;
-use crate::read::merge::MergeReaderBuilder;
+use crate::read::last_row::FlatLastRowReader;
 use crate::read::pruner::{PartitionPruner, Pruner};
 use crate::read::range::RangeMeta;
+use crate::read::range_cache::{
+    build_range_cache_key, cache_flat_range_stream, cached_flat_range_stream,
+};
 use crate::read::scan_region::{ScanInput, StreamContext};
 use crate::read::scan_util::{
-    PartitionMetrics, PartitionMetricsList, SplitRecordBatchStream, scan_file_ranges,
-    scan_flat_file_ranges, scan_flat_mem_ranges, scan_mem_ranges,
-    should_split_flat_batches_for_merge,
+    PartitionMetrics, PartitionMetricsList, SplitRecordBatchStream, compute_parallel_channel_size,
+    scan_flat_file_ranges, scan_flat_mem_ranges, should_split_flat_batches_for_merge,
 };
 use crate::read::stream::{ConvertBatchStream, ScanBatch, ScanBatchStream};
-use crate::read::{
-    Batch, BatchReader, BoxedBatchReader, BoxedRecordBatchStream, ScannerMetrics, Source, scan_util,
-};
+use crate::read::{BoxedRecordBatchStream, ScannerMetrics, scan_util};
 use crate::region::options::MergeMode;
 use crate::sst::parquet::DEFAULT_READ_BATCH_SIZE;
 
@@ -121,35 +119,13 @@ impl SeqScan {
         let streams = (0..self.properties.partitions.len())
             .map(|partition| {
                 let metrics = self.new_partition_metrics(false, &metrics_set, partition);
-                self.scan_batch_in_partition(partition, metrics)
+                self.scan_flat_batch_in_partition(partition, metrics)
             })
             .collect::<Result<Vec<_>>>()?;
 
         Ok(Box::pin(futures::stream::iter(streams).flatten()))
     }
 
-    /// Builds a [BoxedBatchReader] from sequential scan for compaction.
-    ///
-    /// # Panics
-    /// Panics if the compaction flag is not set.
-    pub async fn build_reader_for_compaction(&self) -> Result<BoxedBatchReader> {
-        assert!(self.stream_ctx.input.compaction);
-
-        let metrics_set = ExecutionPlanMetricsSet::new();
-        let part_metrics = self.new_partition_metrics(false, &metrics_set, 0);
-        debug_assert_eq!(1, self.properties.partitions.len());
-        let partition_ranges = &self.properties.partitions[0];
-
-        let reader = Self::merge_all_ranges_for_compaction(
-            &self.stream_ctx,
-            partition_ranges,
-            &part_metrics,
-            self.pruner.clone(),
-        )
-        .await?;
-        Ok(Box::new(reader))
-    }
-
     /// Builds a [BoxedRecordBatchStream] from sequential scan for flat format compaction.
     ///
     /// # Panics
@@ -172,40 +148,6 @@ impl SeqScan {
         Ok(reader)
     }
 
-    /// Builds a merge reader that reads all ranges.
-    /// Callers MUST not split ranges before calling this method.
-    async fn merge_all_ranges_for_compaction(
-        stream_ctx: &Arc<StreamContext>,
-        partition_ranges: &[PartitionRange],
-        part_metrics: &PartitionMetrics,
-        pruner: Arc<Pruner>,
-    ) -> Result<BoxedBatchReader> {
-        pruner.add_partition_ranges(partition_ranges);
-        let partition_pruner = Arc::new(PartitionPruner::new(pruner, partition_ranges));
-
-        let mut sources = Vec::new();
-        for part_range in partition_ranges {
-            build_sources(
-                stream_ctx,
-                part_range,
-                true,
-                part_metrics,
-                partition_pruner.clone(),
-                &mut sources,
-                None,
-            )
-            .await?;
-        }
-
-        common_telemetry::debug!(
-            "Build reader to read all parts, region_id: {}, num_part_ranges: {}, num_sources: {}",
-            stream_ctx.input.mapper.metadata().region_id,
-            partition_ranges.len(),
-            sources.len()
-        );
-        Self::build_reader_from_sources(stream_ctx, sources, None, None).await
-    }
-
     /// Builds a merge reader that reads all flat ranges.
     /// Callers MUST not split ranges before calling this method.
     async fn merge_all_flat_ranges_for_compaction(
@@ -237,93 +179,61 @@ impl SeqScan {
             partition_ranges.len(),
             sources.len()
         );
-        Self::build_flat_reader_from_sources(stream_ctx, sources, None, None).await
+        Self::build_flat_reader_from_sources(
+            stream_ctx,
+            sources,
+            None,
+            None,
+            false,
+            compute_parallel_channel_size(DEFAULT_READ_BATCH_SIZE),
+        )
+        .await
     }
 
-    /// Builds a reader to read sources. If `semaphore` is provided, reads sources in parallel
-    /// if possible.
-    #[tracing::instrument(level = tracing::Level::DEBUG, skip_all)]
-    pub(crate) async fn build_reader_from_sources(
-        stream_ctx: &StreamContext,
-        mut sources: Vec<Source>,
-        semaphore: Option<Arc<Semaphore>>,
-        part_metrics: Option<&PartitionMetrics>,
-    ) -> Result<BoxedBatchReader> {
-        if let Some(semaphore) = semaphore.as_ref() {
-            // Read sources in parallel.
-            if sources.len() > 1 {
-                sources = stream_ctx
-                    .input
-                    .create_parallel_sources(sources, semaphore.clone())?;
-            }
-        }
-
-        let mut builder = MergeReaderBuilder::from_sources(sources);
-        if let Some(metrics) = part_metrics {
-            builder.with_metrics_reporter(Some(metrics.merge_metrics_reporter()));
-        }
-        let reader = builder.build().await?;
-
-        let dedup = !stream_ctx.input.append_mode;
-        let dedup_metrics_reporter = part_metrics.map(|m| m.dedup_metrics_reporter());
-        let reader = if dedup {
-            match stream_ctx.input.merge_mode {
-                MergeMode::LastRow => Box::new(DedupReader::new(
-                    reader,
-                    LastRow::new(stream_ctx.input.filter_deleted),
-                    dedup_metrics_reporter,
-                )) as _,
-                MergeMode::LastNonNull => Box::new(DedupReader::new(
-                    reader,
-                    LastNonNull::new(stream_ctx.input.filter_deleted),
-                    dedup_metrics_reporter,
-                )) as _,
-            }
-        } else {
-            Box::new(reader) as _
-        };
-
-        let reader = match &stream_ctx.input.series_row_selector {
-            Some(TimeSeriesRowSelector::LastRow) => Box::new(LastRowReader::new(reader)) as _,
-            None => reader,
-        };
-
-        Ok(reader)
-    }
-
-    /// Builds a flat reader to read sources that returns RecordBatch. If `semaphore` is provided, reads sources in parallel
-    /// if possible.
+    /// Builds a flat reader to read sources that returns RecordBatch.
+    /// If `semaphore` is provided, reads sources in parallel if possible.
+    /// If `skip_dedup` is true, the merged stream is returned without applying flat dedup.
     #[tracing::instrument(level = tracing::Level::DEBUG, skip_all)]
     pub(crate) async fn build_flat_reader_from_sources(
         stream_ctx: &StreamContext,
         mut sources: Vec<BoxedRecordBatchStream>,
         semaphore: Option<Arc<Semaphore>>,
         part_metrics: Option<&PartitionMetrics>,
+        skip_dedup: bool,
+        channel_size: usize,
     ) -> Result<BoxedRecordBatchStream> {
         if let Some(semaphore) = semaphore.as_ref() {
             // Read sources in parallel.
             if sources.len() > 1 {
-                sources = stream_ctx
-                    .input
-                    .create_parallel_flat_sources(sources, semaphore.clone())?;
+                sources = stream_ctx.input.create_parallel_flat_sources(
+                    sources,
+                    semaphore.clone(),
+                    channel_size,
+                )?;
             }
         }
 
-        let mapper = stream_ctx.input.mapper.as_flat().unwrap();
-        let schema = mapper.input_arrow_schema(stream_ctx.input.compaction);
+        let mapper = &stream_ctx.input.mapper;
+        let reader: BoxedRecordBatchStream = if sources.len() == 1 {
+            // Currently, we can't skip dedup when there is only one source because
+            // that source may have duplicate rows.
+            sources.pop().unwrap()
+        } else {
+            let schema = mapper.input_arrow_schema(stream_ctx.input.compaction);
+            let metrics_reporter = part_metrics.map(|m| m.merge_metrics_reporter());
+            let reader =
+                FlatMergeReader::new(schema, sources, DEFAULT_READ_BATCH_SIZE, metrics_reporter)
+                    .await?;
+            Box::pin(reader.into_stream())
+        };
 
-        let metrics_reporter = part_metrics.map(|m| m.merge_metrics_reporter());
-        let reader =
-            FlatMergeReader::new(schema, sources, DEFAULT_READ_BATCH_SIZE, metrics_reporter)
-                .await?;
-
-        let dedup = !stream_ctx.input.append_mode;
+        let dedup = !skip_dedup && !stream_ctx.input.append_mode;
         let dedup_metrics_reporter = part_metrics.map(|m| m.dedup_metrics_reporter());
         let reader = if dedup {
             match stream_ctx.input.merge_mode {
                 MergeMode::LastRow => Box::pin(
                     FlatDedupReader::new(
-                        reader.into_stream().boxed(),
+                        reader,
                         FlatLastRow::new(stream_ctx.input.filter_deleted),
                         dedup_metrics_reporter,
                     )
@@ -331,7 +241,7 @@ impl SeqScan {
                 ) as _,
                 MergeMode::LastNonNull => Box::pin(
                     FlatDedupReader::new(
-                        reader.into_stream().boxed(),
+                        reader,
                         FlatLastNonNull::new(
                             mapper.field_column_start(),
                             stream_ctx.input.filter_deleted,
@@ -342,12 +252,75 @@ impl SeqScan {
                 ) as _,
             }
         } else {
-            Box::pin(reader.into_stream()) as _
+            reader
+        };
+
+        let reader = match &stream_ctx.input.series_row_selector {
+            Some(TimeSeriesRowSelector::LastRow) => {
+                Box::pin(FlatLastRowReader::new(reader).into_stream()) as _
+            }
+            None => reader,
         };
 
         Ok(reader)
     }
 
+    /// Builds a flat read stream for one partition range.
+    pub(crate) async fn build_flat_partition_range_read(
+        stream_ctx: &Arc<StreamContext>,
+        part_range: &PartitionRange,
+        compaction: bool,
+        part_metrics: &PartitionMetrics,
+        partition_pruner: Arc<PartitionPruner>,
+        file_scan_semaphore: Option<Arc<Semaphore>>,
+        merge_semaphore: Option<Arc<Semaphore>>,
+    ) -> Result<(BoxedRecordBatchStream, usize)> {
+        let cache_key = build_range_cache_key(stream_ctx, part_range);
+
+        if let Some(key) = cache_key.as_ref() {
+            if let Some(value) = stream_ctx.input.cache_strategy.get_range_result(key) {
+                part_metrics.inc_range_cache_hit();
+                return Ok((cached_flat_range_stream(value), DEFAULT_READ_BATCH_SIZE));
+            }
+            part_metrics.inc_range_cache_miss();
+        }
+
+        let mut sources = Vec::new();
+        let split_batch_size = build_flat_sources(
+            stream_ctx,
+            part_range,
+            compaction,
+            part_metrics,
+            partition_pruner,
+            &mut sources,
+            file_scan_semaphore,
+        )
+        .await?;
+        let estimated_rows_per_batch = split_batch_size.unwrap_or(DEFAULT_READ_BATCH_SIZE);
+        let channel_size = compute_parallel_channel_size(estimated_rows_per_batch);
+        let stream = Self::build_flat_reader_from_sources(
+            stream_ctx,
+            sources,
+            merge_semaphore,
+            Some(part_metrics),
+            false,
+            channel_size,
+        )
+        .await?;
+
+        let stream = match cache_key {
+            Some(key) => cache_flat_range_stream(
+                stream,
+                stream_ctx.input.cache_strategy.clone(),
+                key,
+                part_metrics.clone(),
+            ),
+            None => stream,
+        };
+
+        Ok((stream, estimated_rows_per_batch))
+    }
+
     /// Scans the given partition when the part list is set properly.
     /// Otherwise the returned stream might not contains any data.
     fn scan_partition_impl(
@@ -367,13 +340,7 @@ impl SeqScan {
         let metrics = self.new_partition_metrics(ctx.explain_verbose, metrics_set, partition);
         let input = &self.stream_ctx.input;
 
-        let batch_stream = if input.flat_format {
-            // Use flat scan for bulk memtables
-            self.scan_flat_batch_in_partition(partition, metrics.clone())?
-        } else {
-            // Use regular batch scan for normal memtables
-            self.scan_batch_in_partition(partition, metrics.clone())?
-        };
+        let batch_stream = self.scan_flat_batch_in_partition(partition, metrics.clone())?;
         let record_batch_stream = ConvertBatchStream::new(
             batch_stream,
             input.mapper.clone(),
@@ -387,125 +354,6 @@ impl SeqScan {
         )))
     }
 
-    #[tracing::instrument(
-        skip_all,
-        fields(
-            region_id = %self.stream_ctx.input.mapper.metadata().region_id,
-            partition = partition
-        )
-    )]
-    fn scan_batch_in_partition(
-        &self,
-        partition: usize,
-        part_metrics: PartitionMetrics,
-    ) -> Result<ScanBatchStream> {
-        ensure!(
-            partition < self.properties.partitions.len(),
-            PartitionOutOfRangeSnafu {
-                given: partition,
-                all: self.properties.partitions.len(),
-            }
-        );
-
-        if self.properties.partitions[partition].is_empty() {
-            return Ok(Box::pin(futures::stream::empty()));
-        }
-
-        let stream_ctx = self.stream_ctx.clone();
-        let semaphore = self.new_semaphore();
-        let partition_ranges = self.properties.partitions[partition].clone();
-        let compaction = self.stream_ctx.input.compaction;
-        let distinguish_range = self.properties.distinguish_partition_range;
-        let file_scan_semaphore = if compaction { None } else { semaphore.clone() };
-        let pruner = self.pruner.clone();
-        // Initializes ref counts for the pruner.
-        // If we call scan_batch_in_partition() multiple times but don't read all batches from the stream,
-        // then the ref count won't be decremented.
-        // This is a rare case and keeping all remaining entries still uses less memory than a per partition cache.
-        pruner.add_partition_ranges(&partition_ranges);
-        let partition_pruner = Arc::new(PartitionPruner::new(pruner, &partition_ranges));
-
-        let stream = try_stream! {
-            part_metrics.on_first_poll();
-            // Start fetch time before building sources so scan cost contains
-            // build part cost.
-            let mut fetch_start = Instant::now();
-
-            let _mapper = stream_ctx.input.mapper.as_primary_key().context(UnexpectedSnafu {
-                reason: "Unexpected format",
-            })?;
-            // Scans each part.
-            for part_range in partition_ranges {
-                let mut sources = Vec::new();
-                build_sources(
-                    &stream_ctx,
-                    &part_range,
-                    compaction,
-                    &part_metrics,
-                    partition_pruner.clone(),
-                    &mut sources,
-                    file_scan_semaphore.clone(),
-                ).await?;
-
-                let mut reader =
-                    Self::build_reader_from_sources(&stream_ctx, sources, semaphore.clone(), Some(&part_metrics))
-                        .await?;
-                #[cfg(debug_assertions)]
-                let mut checker = crate::read::BatchChecker::default()
-                    .with_start(Some(part_range.start))
-                    .with_end(Some(part_range.end));
-
-                let mut metrics = ScannerMetrics {
-                    scan_cost: fetch_start.elapsed(),
-                    ..Default::default()
-                };
-                fetch_start = Instant::now();
-
-                while let Some(batch) = reader.next_batch().await? {
-                    metrics.scan_cost += fetch_start.elapsed();
-                    metrics.num_batches += 1;
-                    metrics.num_rows += batch.num_rows();
-
-                    debug_assert!(!batch.is_empty());
-                    if batch.is_empty() {
-                        fetch_start = Instant::now();
-                        continue;
-                    }
-
-                    #[cfg(debug_assertions)]
-                    checker.ensure_part_range_batch(
-                        "SeqScan",
-                        _mapper.metadata().region_id,
-                        partition,
-                        part_range,
-                        &batch,
-                    );
-
-                    let yield_start = Instant::now();
-                    yield ScanBatch::Normal(batch);
-                    metrics.yield_cost += yield_start.elapsed();
-
-                    fetch_start = Instant::now();
-                }
-
-                // Yields an empty part to indicate this range is terminated.
-                // The query engine can use this to optimize some queries.
-                if distinguish_range {
-                    let yield_start = Instant::now();
-                    yield ScanBatch::Normal(Batch::empty());
-                    metrics.yield_cost += yield_start.elapsed();
-                }
-
-                metrics.scan_cost += fetch_start.elapsed();
-                fetch_start = Instant::now();
-                part_metrics.merge_metrics(&metrics);
-            }
-
-            part_metrics.on_finish();
-        };
-        Ok(Box::pin(stream))
-    }
-
     #[tracing::instrument(
         skip_all,
         fields(
@@ -551,20 +399,16 @@ impl SeqScan {
 
             // Scans each part.
             for part_range in partition_ranges {
-                let mut sources = Vec::new();
-                build_flat_sources(
+                let (mut reader, _) = Self::build_flat_partition_range_read(
                     &stream_ctx,
                     &part_range,
                     compaction,
                     &part_metrics,
                     partition_pruner.clone(),
-                    &mut sources,
                     file_scan_semaphore.clone(),
-                ).await?;
-
-                let mut reader =
-                    Self::build_flat_reader_from_sources(&stream_ctx, sources, semaphore.clone(), Some(&part_metrics))
-                        .await?;
+                    semaphore.clone(),
+                )
+                .await?;
 
                 let mut metrics = ScannerMetrics {
                     scan_cost: fetch_start.elapsed(),
@@ -728,6 +572,10 @@ impl RegionScanner for SeqScan {
     fn set_logical_region(&mut self, logical_region: bool) {
         self.properties.set_logical_region(logical_region);
     }
+
+    fn snapshot_sequence(&self) -> Option<u64> {
+        self.stream_ctx.input.snapshot_sequence
+    }
 }
 
 impl DisplayAs for SeqScan {
@@ -758,109 +606,8 @@ impl fmt::Debug for SeqScan {
     }
 }
 
-/// Builds sources for the partition range and push them to the `sources` vector.
-pub(crate) async fn build_sources(
-    stream_ctx: &Arc<StreamContext>,
-    part_range: &PartitionRange,
-    compaction: bool,
-    part_metrics: &PartitionMetrics,
-    partition_pruner: Arc<PartitionPruner>,
-    sources: &mut Vec<Source>,
-    semaphore: Option<Arc<Semaphore>>,
-) -> Result<()> {
-    // Gets range meta.
-    let range_meta = &stream_ctx.ranges[part_range.identifier];
-    #[cfg(debug_assertions)]
-    if compaction {
-        // Compaction expects input sources are not been split.
-        debug_assert_eq!(range_meta.indices.len(), range_meta.row_group_indices.len());
-        for (i, row_group_idx) in range_meta.row_group_indices.iter().enumerate() {
-            // It should scan all row groups.
-            debug_assert_eq!(
-                -1, row_group_idx.row_group_index,
-                "Expect {} range scan all row groups, given: {}",
-                i, row_group_idx.row_group_index,
-            );
-        }
-    }
-
-    let read_type = if compaction {
-        "compaction"
-    } else {
-        "seq_scan_files"
-    };
-    let num_indices = range_meta.row_group_indices.len();
-    if num_indices == 0 {
-        return Ok(());
-    }
-
-    sources.reserve(num_indices);
-    let mut ordered_sources = Vec::with_capacity(num_indices);
-    ordered_sources.resize_with(num_indices, || None);
-    let mut file_scan_tasks = Vec::new();
-
-    for (position, index) in range_meta.row_group_indices.iter().enumerate() {
-        if stream_ctx.is_mem_range_index(*index) {
-            let stream = scan_mem_ranges(
-                stream_ctx.clone(),
-                part_metrics.clone(),
-                *index,
-                range_meta.time_range,
-            );
-            ordered_sources[position] = Some(Source::Stream(Box::pin(stream) as _));
-        } else if stream_ctx.is_file_range_index(*index) {
-            if let Some(semaphore_ref) = semaphore.as_ref() {
-                // run in parallel, controlled by semaphore
-                let stream_ctx = stream_ctx.clone();
-                let part_metrics = part_metrics.clone();
-                let partition_pruner = partition_pruner.clone();
-                let semaphore = Arc::clone(semaphore_ref);
-                let row_group_index = *index;
-                file_scan_tasks.push(async move {
-                    let _permit = semaphore.acquire().await.unwrap();
-                    let stream = scan_file_ranges(
-                        stream_ctx,
-                        part_metrics,
-                        row_group_index,
-                        read_type,
-                        partition_pruner,
-                    )
-                    .await?;
-                    Ok((position, Source::Stream(Box::pin(stream) as _)))
-                });
-            } else {
-                // no semaphore, run sequentially
-                let stream = scan_file_ranges(
-                    stream_ctx.clone(),
-                    part_metrics.clone(),
-                    *index,
-                    read_type,
-                    partition_pruner.clone(),
-                )
-                .await?;
-                ordered_sources[position] = Some(Source::Stream(Box::pin(stream) as _));
-            }
-        } else {
-            let stream =
-                scan_util::maybe_scan_other_ranges(stream_ctx, *index, part_metrics).await?;
-            ordered_sources[position] = Some(Source::Stream(stream));
-        }
-    }
-
-    if !file_scan_tasks.is_empty() {
-        let results = futures::future::try_join_all(file_scan_tasks).await?;
-        for (position, source) in results {
-            ordered_sources[position] = Some(source);
-        }
-    }
-
-    for source in ordered_sources.into_iter().flatten() {
-        sources.push(source);
-    }
-    Ok(())
-}
-
 /// Builds flat sources for the partition range and push them to the `sources` vector.
+/// Returns the estimated rows per batch after splitting if splitting is applied, or `None`.
 pub(crate) async fn build_flat_sources(
     stream_ctx: &Arc<StreamContext>,
     part_range: &PartitionRange,
@@ -869,7 +616,7 @@ pub(crate) async fn build_flat_sources(
     partition_pruner: Arc<PartitionPruner>,
     sources: &mut Vec<BoxedRecordBatchStream>,
     semaphore: Option<Arc<Semaphore>>,
-) -> Result<()> {
+) -> Result<Option<usize>> {
     // Gets range meta.
     let range_meta = &stream_ctx.ranges[part_range.identifier];
     #[cfg(debug_assertions)]
@@ -893,14 +640,16 @@ pub(crate) async fn build_flat_sources(
     };
     let num_indices = range_meta.row_group_indices.len();
     if num_indices == 0 {
-        return Ok(());
+        return Ok(None);
     }
 
-    let should_split = should_split_flat_batches_for_merge(stream_ctx, range_meta);
+    let split_batch_size = should_split_flat_batches_for_merge(stream_ctx, range_meta);
+    let should_split = split_batch_size.is_some();
     sources.reserve(num_indices);
     let mut ordered_sources = Vec::with_capacity(num_indices);
     ordered_sources.resize_with(num_indices, || None);
     let mut file_scan_tasks = Vec::new();
+    let pre_filter_mode = stream_ctx.range_pre_filter_mode(part_range);
 
     for (position, index) in range_meta.row_group_indices.iter().enumerate() {
         if stream_ctx.is_mem_range_index(*index) {
@@ -944,8 +693,13 @@ pub(crate) async fn build_flat_sources(
                 ordered_sources[position] = Some(Box::pin(stream) as _);
             }
         } else {
-            let stream =
-                scan_util::maybe_scan_flat_other_ranges(stream_ctx, *index, part_metrics).await?;
+            let stream = scan_util::maybe_scan_flat_other_ranges(
+                stream_ctx,
+                *index,
+                part_metrics,
+                pre_filter_mode,
+            )
+            .await?;
             ordered_sources[position] = Some(stream);
         }
     }
@@ -974,7 +728,7 @@ pub(crate) async fn build_flat_sources(
         );
     }
 
-    Ok(())
+    Ok(split_batch_size)
 }
 
 #[cfg(test)]
diff --git a/src/mito2/src/read/series_scan.rs b/src/mito2/src/read/series_scan.rs
index 2d6994d0af..003a754363 100644
--- a/src/mito2/src/read/series_scan.rs
+++ b/src/mito2/src/read/series_scan.rs
@@ -30,7 +30,7 @@ use datatypes::arrow::array::BinaryArray;
 use datatypes::arrow::record_batch::RecordBatch;
 use datatypes::schema::SchemaRef;
 use futures::{StreamExt, TryStreamExt};
-use smallvec::{SmallVec, smallvec};
+use smallvec::SmallVec;
 use snafu::{OptionExt, ResultExt, ensure};
 use store_api::metadata::RegionMetadataRef;
 use store_api::region_engine::{
@@ -41,15 +41,18 @@ use tokio::sync::mpsc::error::{SendTimeoutError, TrySendError};
 use tokio::sync::mpsc::{self, Receiver, Sender};
 
 use crate::error::{
-    Error, InvalidSenderSnafu, PartitionOutOfRangeSnafu, Result, ScanMultiTimesSnafu,
+    Error, InvalidSenderSnafu, JoinSnafu, PartitionOutOfRangeSnafu, Result, ScanMultiTimesSnafu,
     ScanSeriesSnafu, TooManyFilesToReadSnafu,
 };
+use crate::read::ScannerMetrics;
 use crate::read::pruner::{PartitionPruner, Pruner};
 use crate::read::scan_region::{ScanInput, StreamContext};
-use crate::read::scan_util::{PartitionMetrics, PartitionMetricsList, SeriesDistributorMetrics};
-use crate::read::seq_scan::{SeqScan, build_flat_sources, build_sources};
+use crate::read::scan_util::{
+    PartitionMetrics, PartitionMetricsList, SeriesDistributorMetrics, compute_average_batch_size,
+    compute_parallel_channel_size,
+};
+use crate::read::seq_scan::SeqScan;
 use crate::read::stream::{ConvertBatchStream, ScanBatch, ScanBatchStream};
-use crate::read::{Batch, ScannerMetrics};
 use crate::sst::parquet::flat_format::primary_key_column_index;
 use crate::sst::parquet::format::PrimaryKeyArray;
 
@@ -168,7 +171,7 @@ impl SeriesScan {
             }
         );
 
-        self.maybe_start_distributor(metrics_set, &self.metrics_list);
+        self.maybe_start_distributor(metrics_set, &self.metrics_list, ctx.explain_verbose);
 
         let mut receiver = self.take_receiver(partition)?;
         let stream = try_stream! {
@@ -214,6 +217,7 @@ impl SeriesScan {
         &self,
         metrics_set: &ExecutionPlanMetricsSet,
         metrics_list: &Arc<PartitionMetricsList>,
+        explain_verbose: bool,
     ) {
         let mut rx_list = self.receivers.lock().unwrap();
         if !rx_list.is_empty() {
@@ -223,12 +227,14 @@ impl SeriesScan {
         let (senders, receivers) = new_channel_list(self.properties.num_partitions());
         let mut distributor = SeriesDistributor {
             stream_ctx: self.stream_ctx.clone(),
-            semaphore: Some(Arc::new(Semaphore::new(self.properties.num_partitions()))),
+            range_semaphore: Some(Arc::new(Semaphore::new(self.properties.num_partitions()))),
+            final_merge_semaphore: Some(Arc::new(Semaphore::new(self.properties.num_partitions()))),
             partitions: self.properties.partitions.clone(),
             pruner: self.pruner.clone(),
             senders,
             metrics_set: metrics_set.clone(),
             metrics_list: metrics_list.clone(),
+            explain_verbose,
         };
         let region_id = distributor.stream_ctx.input.mapper.metadata().region_id;
         let span = tracing::info_span!("SeriesScan::distributor", region_id = %region_id);
@@ -374,6 +380,10 @@ impl RegionScanner for SeriesScan {
     fn set_logical_region(&mut self, logical_region: bool) {
         self.properties.set_logical_region(logical_region);
     }
+
+    fn snapshot_sequence(&self) -> Option<u64> {
+        self.stream_ctx.input.snapshot_sequence
+    }
 }
 
 impl DisplayAs for SeriesScan {
@@ -415,8 +425,13 @@ impl SeriesScan {
 struct SeriesDistributor {
     /// Context for the scan stream.
     stream_ctx: Arc<StreamContext>,
-    /// Optional semaphore for limiting the number of concurrent scans.
-    semaphore: Option<Arc<Semaphore>>,
+    /// Semaphore for file scanning and range-level merging.
+    range_semaphore: Option<Arc<Semaphore>>,
+    /// Semaphore for the final merge across all range streams.
+    /// Must be separate from `range_semaphore` to avoid deadlock: final merge tasks
+    /// hold a permit while waiting for data from range-level merge tasks, which also
+    /// need permits to produce data.
+    final_merge_semaphore: Option<Arc<Semaphore>>,
     /// Partition ranges to scan.
     partitions: Vec<Vec<PartitionRange>>,
     /// Shared pruner for file range building.
@@ -430,6 +445,8 @@ struct SeriesDistributor {
     /// distributor.
     metrics_set: ExecutionPlanMetricsSet,
     metrics_list: Arc<PartitionMetricsList>,
+    /// Whether to use verbose logging and collect detailed metrics.
+    explain_verbose: bool,
 }
 
 impl SeriesDistributor {
@@ -439,11 +456,7 @@ impl SeriesDistributor {
         fields(region_id = %self.stream_ctx.input.mapper.metadata().region_id)
     )]
     async fn execute(&mut self) {
-        let result = if self.stream_ctx.input.flat_format {
-            self.scan_partitions_flat().await
-        } else {
-            self.scan_partitions().await
-        };
+        let result = self.scan_partitions_flat().await;
 
         if let Err(e) = result {
             self.senders.send_error(e).await;
@@ -470,7 +483,7 @@ impl SeriesDistributor {
 
         let part_metrics = new_partition_metrics(
             &self.stream_ctx,
-            false,
+            self.explain_verbose,
             &self.metrics_set,
             self.partitions.len(),
             &self.metrics_list,
@@ -480,30 +493,58 @@ impl SeriesDistributor {
         // build part cost.
         let mut fetch_start = Instant::now();
 
-        // Scans all parts.
-        let mut sources = Vec::with_capacity(self.partitions.len());
+        // Builds one deduped stream per partition range, then merges across ranges.
+        let build_start = Instant::now();
+        let mut tasks = Vec::new();
         for partition in &self.partitions {
-            sources.reserve(partition.len());
             for part_range in partition {
-                build_flat_sources(
-                    &self.stream_ctx,
-                    part_range,
-                    false,
-                    &part_metrics,
-                    partition_pruner.clone(),
-                    &mut sources,
-                    self.semaphore.clone(),
-                )
-                .await?;
+                let stream_ctx = self.stream_ctx.clone();
+                let part_range = *part_range;
+                let part_metrics = part_metrics.clone();
+                let partition_pruner = partition_pruner.clone();
+                let file_scan_semaphore = self.range_semaphore.clone();
+                let merge_semaphore = self.range_semaphore.clone();
+                tasks.push(common_runtime::spawn_global(async move {
+                    SeqScan::build_flat_partition_range_read(
+                        &stream_ctx,
+                        &part_range,
+                        false,
+                        &part_metrics,
+                        partition_pruner,
+                        file_scan_semaphore,
+                        merge_semaphore,
+                    )
+                    .await
+                }));
             }
         }
+        let mut range_streams = Vec::with_capacity(tasks.len());
+        let mut estimated_batch_sizes = Vec::with_capacity(tasks.len());
+        for task in tasks {
+            let (stream, estimated_batch_size) = task.await.context(JoinSnafu)??;
+            range_streams.push(stream);
+            estimated_batch_sizes.push(estimated_batch_size);
+        }
+        let channel_size =
+            compute_parallel_channel_size(compute_average_batch_size(estimated_batch_sizes));
+        common_telemetry::debug!(
+            "SeriesDistributor built {} range_streams, region: {}, build cost: {:?}, channel_size: {}",
+            range_streams.len(),
+            self.stream_ctx.input.region_metadata().region_id,
+            build_start.elapsed(),
+            channel_size,
+        );
 
-        // Builds a flat reader that merge sources from all parts.
+        // Each partition range stream is already deduped, so skip dedup here.
+        // Use a separate semaphore for the final merge to avoid deadlock with
+        // range-level merge tasks that share the range_semaphore.
         let mut reader = SeqScan::build_flat_reader_from_sources(
             &self.stream_ctx,
-            sources,
-            self.semaphore.clone(),
+            range_streams,
+            self.final_merge_semaphore.clone(),
             Some(&part_metrics),
+            true,
+            channel_size,
         )
         .await?;
         let mut metrics = SeriesDistributorMetrics::default();
@@ -555,151 +596,11 @@ impl SeriesDistributor {
 
         Ok(())
     }
-
-    /// Scans all parts.
-    #[tracing::instrument(
-        skip_all,
-        fields(region_id = %self.stream_ctx.input.mapper.metadata().region_id)
-    )]
-    async fn scan_partitions(&mut self) -> Result<()> {
-        // Initialize reference counts for all partition ranges.
-        for partition_ranges in &self.partitions {
-            self.pruner.add_partition_ranges(partition_ranges);
-        }
-
-        // Create PartitionPruner covering all partitions
-        let all_partition_ranges: Vec<_> = self.partitions.iter().flatten().cloned().collect();
-        let partition_pruner = Arc::new(PartitionPruner::new(
-            self.pruner.clone(),
-            &all_partition_ranges,
-        ));
-
-        let part_metrics = new_partition_metrics(
-            &self.stream_ctx,
-            false,
-            &self.metrics_set,
-            self.partitions.len(),
-            &self.metrics_list,
-        );
-        part_metrics.on_first_poll();
-        // Start fetch time before building sources so scan cost contains
-        // build part cost.
-        let mut fetch_start = Instant::now();
-
-        // Scans all parts.
-        let mut sources = Vec::with_capacity(self.partitions.len());
-        for partition in &self.partitions {
-            sources.reserve(partition.len());
-            for part_range in partition {
-                build_sources(
-                    &self.stream_ctx,
-                    part_range,
-                    false,
-                    &part_metrics,
-                    partition_pruner.clone(),
-                    &mut sources,
-                    self.semaphore.clone(),
-                )
-                .await?;
-            }
-        }
-
-        // Builds a reader that merge sources from all parts.
-        let mut reader = SeqScan::build_reader_from_sources(
-            &self.stream_ctx,
-            sources,
-            self.semaphore.clone(),
-            Some(&part_metrics),
-        )
-        .await?;
-        let mut metrics = SeriesDistributorMetrics::default();
-
-        let mut current_series = PrimaryKeySeriesBatch::default();
-        while let Some(batch) = reader.next_batch().await? {
-            metrics.scan_cost += fetch_start.elapsed();
-            metrics.num_batches += 1;
-            metrics.num_rows += batch.num_rows();
-
-            debug_assert!(!batch.is_empty());
-            if batch.is_empty() {
-                fetch_start = Instant::now();
-                continue;
-            }
-
-            let Some(last_key) = current_series.current_key() else {
-                current_series.push(batch);
-                fetch_start = Instant::now();
-                continue;
-            };
-
-            if last_key == batch.primary_key() {
-                current_series.push(batch);
-                fetch_start = Instant::now();
-                continue;
-            }
-
-            // We find a new series, send the current one.
-            let to_send =
-                std::mem::replace(&mut current_series, PrimaryKeySeriesBatch::single(batch));
-            let yield_start = Instant::now();
-            self.senders
-                .send_batch(SeriesBatch::PrimaryKey(to_send))
-                .await?;
-            metrics.yield_cost += yield_start.elapsed();
-            fetch_start = Instant::now();
-        }
-
-        if !current_series.is_empty() {
-            let yield_start = Instant::now();
-            self.senders
-                .send_batch(SeriesBatch::PrimaryKey(current_series))
-                .await?;
-            metrics.yield_cost += yield_start.elapsed();
-        }
-
-        metrics.scan_cost += fetch_start.elapsed();
-        metrics.num_series_send_timeout = self.senders.num_timeout;
-        metrics.num_series_send_full = self.senders.num_full;
-        part_metrics.set_distributor_metrics(&metrics);
-
-        part_metrics.on_finish();
-
-        Ok(())
-    }
-}
-
-/// Batches of the same series in primary key format.
-#[derive(Default, Debug)]
-pub struct PrimaryKeySeriesBatch {
-    pub batches: SmallVec<[Batch; 4]>,
-}
-
-impl PrimaryKeySeriesBatch {
-    /// Creates a new [PrimaryKeySeriesBatch] from a single [Batch].
-    fn single(batch: Batch) -> Self {
-        Self {
-            batches: smallvec![batch],
-        }
-    }
-
-    fn current_key(&self) -> Option<&[u8]> {
-        self.batches.first().map(|batch| batch.primary_key())
-    }
-
-    fn push(&mut self, batch: Batch) {
-        self.batches.push(batch);
-    }
-
-    /// Returns true if there is no batch.
-    fn is_empty(&self) -> bool {
-        self.batches.is_empty()
-    }
 }
 
 /// Batches of the same series.
 #[derive(Debug)]
 pub enum SeriesBatch {
-    PrimaryKey(PrimaryKeySeriesBatch),
     Flat(FlatSeriesBatch),
 }
 
@@ -707,7 +608,6 @@ impl SeriesBatch {
     /// Returns the number of batches.
     pub fn num_batches(&self) -> usize {
         match self {
-            SeriesBatch::PrimaryKey(primary_key_batch) => primary_key_batch.batches.len(),
             SeriesBatch::Flat(flat_batch) => flat_batch.batches.len(),
         }
     }
@@ -715,9 +615,6 @@ impl SeriesBatch {
     /// Returns the total number of rows across all batches.
     pub fn num_rows(&self) -> usize {
         match self {
-            SeriesBatch::PrimaryKey(primary_key_batch) => {
-                primary_key_batch.batches.iter().map(|x| x.num_rows()).sum()
-            }
             SeriesBatch::Flat(flat_batch) => flat_batch.batches.iter().map(|x| x.num_rows()).sum(),
         }
     }
diff --git a/src/mito2/src/read/stream.rs b/src/mito2/src/read/stream.rs
index dd85616241..27ae0e8c3a 100644
--- a/src/mito2/src/read/stream.rs
+++ b/src/mito2/src/read/stream.rs
@@ -27,14 +27,12 @@ use snafu::ResultExt;
 
 use crate::cache::CacheStrategy;
 use crate::error::Result;
-use crate::read::Batch;
-use crate::read::projection::ProjectionMapper;
+use crate::read::flat_projection::FlatProjectionMapper;
 use crate::read::scan_util::PartitionMetrics;
 use crate::read::series_scan::SeriesBatch;
 
 /// All kinds of [`Batch`]es to produce in scanner.
 pub enum ScanBatch {
-    Normal(Batch),
     Series(SeriesBatch),
     RecordBatch(DfRecordBatch),
 }
@@ -44,7 +42,8 @@ pub type ScanBatchStream = BoxStream<'static, Result<ScanBatch>>;
 /// A stream that takes [`ScanBatch`]es and produces (converts them to) [`RecordBatch`]es.
 pub(crate) struct ConvertBatchStream {
     inner: ScanBatchStream,
-    projection_mapper: Arc<ProjectionMapper>,
+    projection_mapper: Arc<FlatProjectionMapper>,
+    #[allow(dead_code)]
     cache_strategy: CacheStrategy,
     partition_metrics: PartitionMetrics,
     pending: VecDeque<RecordBatch>,
@@ -53,7 +52,7 @@ pub(crate) struct ConvertBatchStream {
 impl ConvertBatchStream {
     pub(crate) fn new(
         inner: ScanBatchStream,
-        projection_mapper: Arc<ProjectionMapper>,
+        projection_mapper: Arc<FlatProjectionMapper>,
         cache_strategy: CacheStrategy,
         partition_metrics: PartitionMetrics,
     ) -> Self {
@@ -68,40 +67,19 @@ impl ConvertBatchStream {
 
     fn convert(&mut self, batch: ScanBatch) -> common_recordbatch::error::Result<RecordBatch> {
         match batch {
-            ScanBatch::Normal(batch) => {
-                // Safety: Only primary key format returns this batch.
-                let mapper = self.projection_mapper.as_primary_key().unwrap();
-
-                if batch.is_empty() {
-                    Ok(mapper.empty_record_batch())
-                } else {
-                    mapper.convert(&batch, &self.cache_strategy)
-                }
-            }
             ScanBatch::Series(series) => {
                 debug_assert!(
                     self.pending.is_empty(),
                     "ConvertBatchStream should not convert a new SeriesBatch when pending batches exist"
                 );
 
-                match series {
-                    SeriesBatch::PrimaryKey(primary_key_batch) => {
-                        // Safety: Only primary key format returns this batch.
-                        let mapper = self.projection_mapper.as_primary_key().unwrap();
-
-                        for batch in primary_key_batch.batches {
-                            self.pending
-                                .push_back(mapper.convert(&batch, &self.cache_strategy)?);
-                        }
-                    }
-                    SeriesBatch::Flat(flat_batch) => {
-                        // Safety: Only flat format returns this batch.
-                        let mapper = self.projection_mapper.as_flat().unwrap();
-
-                        for batch in flat_batch.batches {
-                            self.pending.push_back(mapper.convert(&batch)?);
-                        }
-                    }
+                let SeriesBatch::Flat(flat_batch) = series;
+                // Safety: Only flat format returns this batch.
+                for batch in flat_batch.batches {
+                    self.pending.push_back(
+                        self.projection_mapper
+                            .convert(&batch, &self.cache_strategy)?,
+                    );
                 }
 
                 let output_schema = self.projection_mapper.output_schema();
@@ -112,9 +90,8 @@ impl ConvertBatchStream {
             }
             ScanBatch::RecordBatch(df_record_batch) => {
                 // Safety: Only flat format returns this batch.
-                let mapper = self.projection_mapper.as_flat().unwrap();
-
-                mapper.convert(&df_record_batch)
+                self.projection_mapper
+                    .convert(&df_record_batch, &self.cache_strategy)
             }
         }
     }
diff --git a/src/mito2/src/read/unordered_scan.rs b/src/mito2/src/read/unordered_scan.rs
index 2d557e8871..a48a7048bd 100644
--- a/src/mito2/src/read/unordered_scan.rs
+++ b/src/mito2/src/read/unordered_scan.rs
@@ -37,11 +37,10 @@ use crate::error::{PartitionOutOfRangeSnafu, Result};
 use crate::read::pruner::{PartitionPruner, Pruner};
 use crate::read::scan_region::{ScanInput, StreamContext};
 use crate::read::scan_util::{
-    PartitionMetrics, PartitionMetricsList, scan_file_ranges, scan_flat_file_ranges,
-    scan_flat_mem_ranges, scan_mem_ranges,
+    PartitionMetrics, PartitionMetricsList, scan_flat_file_ranges, scan_flat_mem_ranges,
 };
 use crate::read::stream::{ConvertBatchStream, ScanBatch, ScanBatchStream};
-use crate::read::{Batch, ScannerMetrics, scan_util};
+use crate::read::{ScannerMetrics, scan_util};
 
 /// Scans a region without providing any output ordering guarantee.
 ///
@@ -103,59 +102,6 @@ impl UnorderedScan {
         Ok(stream)
     }
 
-    /// Scans a [PartitionRange] by its `identifier` and returns a stream.
-    #[tracing::instrument(
-        skip_all,
-        fields(
-            region_id = %stream_ctx.input.region_metadata().region_id,
-            part_range_id = part_range_id
-        )
-    )]
-    fn scan_partition_range(
-        stream_ctx: Arc<StreamContext>,
-        part_range_id: usize,
-        part_metrics: PartitionMetrics,
-        partition_pruner: Arc<PartitionPruner>,
-    ) -> impl Stream<Item = Result<Batch>> {
-        try_stream! {
-            // Gets range meta.
-            let range_meta = &stream_ctx.ranges[part_range_id];
-            for index in &range_meta.row_group_indices {
-                if stream_ctx.is_mem_range_index(*index) {
-                    let stream = scan_mem_ranges(
-                        stream_ctx.clone(),
-                        part_metrics.clone(),
-                        *index,
-                        range_meta.time_range,
-                    );
-                    for await batch in stream {
-                        yield batch?;
-                    }
-                } else if stream_ctx.is_file_range_index(*index) {
-                    let stream = scan_file_ranges(
-                        stream_ctx.clone(),
-                        part_metrics.clone(),
-                        *index,
-                        "unordered_scan_files",
-                        partition_pruner.clone(),
-                    ).await?;
-                    for await batch in stream {
-                        yield batch?;
-                    }
-                } else {
-                    let stream = scan_util::maybe_scan_other_ranges(
-                        &stream_ctx,
-                        *index,
-                        &part_metrics,
-                    ).await?;
-                    for await batch in stream {
-                        yield batch?;
-                    }
-                }
-            }
-        }
-    }
-
     /// Scans a [PartitionRange] by its `identifier` and returns a flat stream of RecordBatch.
     #[tracing::instrument(
         skip_all,
@@ -173,6 +119,8 @@ impl UnorderedScan {
         try_stream! {
             // Gets range meta.
             let range_meta = &stream_ctx.ranges[part_range_id];
+            let part_range = range_meta.new_partition_range(part_range_id);
+            let pre_filter_mode = stream_ctx.range_pre_filter_mode(&part_range);
             for index in &range_meta.row_group_indices {
                 if stream_ctx.is_mem_range_index(*index) {
                     let stream = scan_flat_mem_ranges(
@@ -200,6 +148,7 @@ impl UnorderedScan {
                         &stream_ctx,
                         *index,
                         &part_metrics,
+                        pre_filter_mode,
                     ).await?;
                     for await record_batch in stream {
                         yield record_batch?;
@@ -216,7 +165,7 @@ impl UnorderedScan {
         let streams = (0..self.properties.partitions.len())
             .map(|partition| {
                 let metrics = self.partition_metrics(false, partition, &metrics_set);
-                self.scan_batch_in_partition(partition, metrics)
+                self.scan_flat_batch_in_partition(partition, metrics)
             })
             .collect::<Result<Vec<_>>>()?;
 
@@ -265,13 +214,7 @@ impl UnorderedScan {
         let metrics = self.partition_metrics(ctx.explain_verbose, partition, metrics_set);
         let input = &self.stream_ctx.input;
 
-        let batch_stream = if input.flat_format {
-            // Use flat scan for bulk memtables
-            self.scan_flat_batch_in_partition(partition, metrics.clone())?
-        } else {
-            // Use regular batch scan for normal memtables
-            self.scan_batch_in_partition(partition, metrics.clone())?
-        };
+        let batch_stream = self.scan_flat_batch_in_partition(partition, metrics.clone())?;
 
         let record_batch_stream = ConvertBatchStream::new(
             batch_stream,
@@ -286,100 +229,6 @@ impl UnorderedScan {
         )))
     }
 
-    #[tracing::instrument(
-        skip_all,
-        fields(
-            region_id = %self.stream_ctx.input.mapper.metadata().region_id,
-            partition = partition
-        )
-    )]
-    fn scan_batch_in_partition(
-        &self,
-        partition: usize,
-        part_metrics: PartitionMetrics,
-    ) -> Result<ScanBatchStream> {
-        ensure!(
-            partition < self.properties.partitions.len(),
-            PartitionOutOfRangeSnafu {
-                given: partition,
-                all: self.properties.partitions.len(),
-            }
-        );
-
-        let stream_ctx = self.stream_ctx.clone();
-        let part_ranges = self.properties.partitions[partition].clone();
-        let distinguish_range = self.properties.distinguish_partition_range;
-        let pruner = self.pruner.clone();
-        // Initializes ref counts for the pruner.
-        // If we call scan_batch_in_partition() multiple times but don't read all batches from the stream,
-        // then the ref count won't be decremented.
-        // This is a rare case and keeping all remaining entries still uses less memory than a per partition cache.
-        pruner.add_partition_ranges(&part_ranges);
-        let partition_pruner = Arc::new(PartitionPruner::new(pruner, &part_ranges));
-
-        let stream = try_stream! {
-            part_metrics.on_first_poll();
-
-            // Scans each part.
-            for part_range in part_ranges {
-                let mut metrics = ScannerMetrics::default();
-                let mut fetch_start = Instant::now();
-                let _mapper = &stream_ctx.input.mapper;
-                #[cfg(debug_assertions)]
-                let mut checker = crate::read::BatchChecker::default()
-                    .with_start(Some(part_range.start))
-                    .with_end(Some(part_range.end));
-
-                let stream = Self::scan_partition_range(
-                    stream_ctx.clone(),
-                    part_range.identifier,
-                    part_metrics.clone(),
-                    partition_pruner.clone(),
-                );
-                for await batch in stream {
-                    let batch = batch?;
-                    metrics.scan_cost += fetch_start.elapsed();
-                    metrics.num_batches += 1;
-                    metrics.num_rows += batch.num_rows();
-
-                    debug_assert!(!batch.is_empty());
-                    if batch.is_empty() {
-                        continue;
-                    }
-
-                    #[cfg(debug_assertions)]
-                    checker.ensure_part_range_batch(
-                        "UnorderedScan",
-                        _mapper.metadata().region_id,
-                        partition,
-                        part_range,
-                        &batch,
-                    );
-
-                    let yield_start = Instant::now();
-                    yield ScanBatch::Normal(batch);
-                    metrics.yield_cost += yield_start.elapsed();
-
-                    fetch_start = Instant::now();
-                }
-
-                // Yields an empty part to indicate this range is terminated.
-                // The query engine can use this to optimize some queries.
-                if distinguish_range {
-                    let yield_start = Instant::now();
-                    yield ScanBatch::Normal(Batch::empty());
-                    metrics.yield_cost += yield_start.elapsed();
-                }
-
-                metrics.scan_cost += fetch_start.elapsed();
-                part_metrics.merge_metrics(&metrics);
-            }
-
-            part_metrics.on_finish();
-        };
-        Ok(Box::pin(stream))
-    }
-
     #[tracing::instrument(
         skip_all,
         fields(
@@ -505,6 +354,10 @@ impl RegionScanner for UnorderedScan {
     fn set_logical_region(&mut self, logical_region: bool) {
         self.properties.set_logical_region(logical_region);
     }
+
+    fn snapshot_sequence(&self) -> Option<u64> {
+        self.stream_ctx.input.snapshot_sequence
+    }
 }
 
 impl DisplayAs for UnorderedScan {
diff --git a/src/mito2/src/region.rs b/src/mito2/src/region.rs
index de8927c4de..2e7b7a8ae5 100644
--- a/src/mito2/src/region.rs
+++ b/src/mito2/src/region.rs
@@ -156,11 +156,6 @@ pub struct MitoRegion {
     pub(crate) topic_latest_entry_id: AtomicU64,
     /// The total bytes written to the region.
     pub(crate) written_bytes: Arc<AtomicU64>,
-    /// Partition info of the region in staging mode.
-    ///
-    /// During the staging mode, the region metadata in [`VersionControlRef`] is not updated,
-    /// so we need to store the partition info separately.
-    pub(crate) staging_partition_info: Mutex<Option<StagingPartitionInfo>>,
     /// manifest stats
     stats: ManifestStats,
 }
@@ -318,6 +313,15 @@ impl MitoRegion {
         self.version_control.committed_sequence()
     }
 
+    /// Returns the latest sequence that has already been persisted into SSTs.
+    ///
+    /// Incremental memtable-only reads must use a cursor greater than or equal to
+    /// this boundary; older cursors are stale because the corresponding updates may
+    /// already have been flushed out of memtables.
+    pub fn flushed_sequence(&self) -> SequenceNumber {
+        self.version_control.current().version.flushed_sequence
+    }
+
     /// Returns whether the region is readonly.
     pub fn is_follower(&self) -> bool {
         self.manifest_ctx.state.load() == RegionRoleState::Follower
@@ -333,6 +337,17 @@ impl MitoRegion {
         self.manifest_ctx.set_role(next_role, self.region_id);
     }
 
+    pub(crate) fn region_role(&self) -> RegionRole {
+        match self.state() {
+            RegionRoleState::Follower => RegionRole::Follower,
+            RegionRoleState::Leader(RegionLeaderState::Staging) => RegionRole::StagingLeader,
+            RegionRoleState::Leader(RegionLeaderState::Downgrading) => {
+                RegionRole::DowngradingLeader
+            }
+            RegionRoleState::Leader(_) => RegionRole::Leader,
+        }
+    }
+
     /// Sets the altering state.
     /// You should call this method in the worker loop.
     pub(crate) fn set_altering(&self) -> Result<()> {
@@ -393,9 +408,8 @@ impl MitoRegion {
     /// You should call this method in the worker loop.
     /// Transitions from Staging to Writable state.
     pub fn exit_staging(&self) -> Result<()> {
-        *self.staging_partition_info.lock().unwrap() = None;
-        self.compare_exchange_state(
-            RegionLeaderState::Staging,
+        self.manifest_ctx.exit_staging(
+            self.region_id,
             RegionRoleState::Leader(RegionLeaderState::Writable),
         )
     }
@@ -656,7 +670,7 @@ impl MitoRegion {
             .unwrap_or_default();
         let files = manifest_files
             .into_iter()
-            .chain(staging_files.into_iter())
+            .chain(staging_files)
             .collect::<HashMap<_, _>>();
 
         files
@@ -819,7 +833,7 @@ impl MitoRegion {
     pub fn maybe_staging_partition_expr_str(&self) -> Option<String> {
         let is_staging = self.is_staging();
         if is_staging {
-            let staging_partition_info = self.staging_partition_info.lock().unwrap();
+            let staging_partition_info = self.manifest_ctx.staging_partition_info();
             if staging_partition_info.is_none() {
                 warn!(
                     "Staging partition expr is none for region {} in staging state",
@@ -837,8 +851,8 @@ impl MitoRegion {
 
     pub fn expected_partition_expr_version(&self) -> u64 {
         if self.is_staging() {
-            let staging_partition_info = self.staging_partition_info.lock().unwrap();
-            staging_partition_info
+            self.manifest_ctx
+                .staging_partition_info()
                 .as_ref()
                 .map(|info| info.partition_rule_version)
                 .unwrap_or_default()
@@ -852,8 +866,8 @@ impl MitoRegion {
         if !self.is_staging() {
             return false;
         }
-        let staging_partition_info = self.staging_partition_info.lock().unwrap();
-        staging_partition_info
+        self.manifest_ctx
+            .staging_partition_info()
             .as_ref()
             .map(|info| {
                 matches!(
@@ -873,6 +887,11 @@ pub(crate) struct ManifestContext {
     /// The state of the region. The region checks the state before updating
     /// manifest.
     state: AtomicCell<RegionRoleState>,
+    /// Partition info of the region in staging mode.
+    ///
+    /// During the staging mode, the region metadata in [`VersionControlRef`] is not updated,
+    /// so we need to store the partition info separately.
+    staging_partition_info: Mutex<Option<StagingPartitionInfo>>,
 }
 
 impl ManifestContext {
@@ -880,9 +899,46 @@ impl ManifestContext {
         ManifestContext {
             manifest_manager: tokio::sync::RwLock::new(manager),
             state: AtomicCell::new(state),
+            staging_partition_info: Mutex::new(None),
         }
     }
 
+    pub(crate) fn staging_partition_info(&self) -> Option<StagingPartitionInfo> {
+        self.staging_partition_info.lock().unwrap().clone()
+    }
+
+    pub(crate) fn set_staging_partition_info(&self, staging_partition_info: StagingPartitionInfo) {
+        let mut current = self.staging_partition_info.lock().unwrap();
+        debug_assert!(current.is_none());
+        *current = Some(staging_partition_info);
+    }
+
+    fn clear_staging_partition_info(&self) {
+        *self.staging_partition_info.lock().unwrap() = None;
+    }
+
+    pub(crate) fn exit_staging(
+        &self,
+        region_id: RegionId,
+        next_state: RegionRoleState,
+    ) -> Result<()> {
+        self.state
+            .compare_exchange(
+                RegionRoleState::Leader(RegionLeaderState::Staging),
+                next_state,
+            )
+            .map_err(|actual| {
+                RegionStateSnafu {
+                    region_id,
+                    state: actual,
+                    expect: RegionRoleState::Leader(RegionLeaderState::Staging),
+                }
+                .build()
+            })?;
+        self.clear_staging_partition_info();
+        Ok(())
+    }
+
     pub(crate) async fn manifest_version(&self) -> ManifestVersion {
         self.manifest_manager
             .read()
@@ -973,8 +1029,23 @@ impl ManifestContext {
 
             // This is an edit from flush.
             if let Some(flushed_entry_id) = edit.flushed_entry_id {
+                // A flush edit is valid after truncate in two cases:
+                // 1. `flushed_entry_id` moves past `truncated_entry_id`, meaning it definitely
+                //    flushed data newer than the truncate point.
+                // 2. `flushed_entry_id` equals `truncated_entry_id`, but `flushed_sequence`
+                //    increases. This happens in skip-WAL tables where entry id can stay at 0,
+                //    while sequence still advances for post-truncate writes.
+                //
+                // We still reject stale flushes from before truncate:
+                // if entry id is equal and sequence does not advance, the flush is outdated.
+                let is_newer_entry = truncated_entry_id < flushed_entry_id;
+                let is_same_entry_with_newer_sequence = truncated_entry_id == flushed_entry_id
+                    && edit.flushed_sequence.is_some_and(|flushed_sequence| {
+                        manifest.flushed_sequence < flushed_sequence
+                    });
+
                 ensure!(
-                    truncated_entry_id < flushed_entry_id,
+                    is_newer_entry || is_same_entry_with_newer_sequence,
                     RegionTruncatedSnafu {
                         region_id: manifest.metadata.region_id,
                     }
@@ -1013,27 +1084,50 @@ impl ManifestContext {
     /// Sets the [`RegionRole`].
     ///
     /// ```text
-    ///     +------------------------------------------+
-    ///     |                      +-----------------+ |
-    ///     |                      |                 | |
-    /// +---+------+       +-------+-----+        +--v-v---+
-    /// | Follower |       | Downgrading |        | Leader |
-    /// +---^-^----+       +-----+-^-----+        +--+-+---+
-    ///     | |                  | |                 | |
-    ///     | +------------------+ +-----------------+ |
-    ///     +------------------------------------------+
-    ///
-    /// Transition:
-    /// - Follower -> Leader
-    /// - Downgrading Leader -> Leader
-    /// - Leader -> Follower
-    /// - Downgrading Leader -> Follower
-    /// - Leader -> Downgrading Leader
+    ///                  +---------------------+
+    ///                  |   Staging Leader    |
+    ///                  +----------+----------+
+    ///                             |
+    ///                             v
+    ///     +----------+     +------+-------+     +-------------+
+    ///     | Follower | <-> |    Leader    | <-> | Downgrading |
+    ///     +-----+----+     +------+-------+     +------+------+
+    ///           ^                 ^                    |
+    ///           +-----------------+--------------------+
     ///
     /// ```
+    ///
+    /// # State Transitions
+    ///
+    /// From `Follower`:
+    /// - `Follower -> Leader`
+    ///
+    /// From `Leader`:
+    /// - `Leader -> Follower`
+    /// - `Leader -> Downgrading Leader`
+    ///
+    /// From `Staging Leader`:
+    /// - `Staging Leader -> Leader`
+    /// - `Staging Leader -> Follower`
+    /// - `Staging Leader -> Downgrading Leader`
+    ///
+    /// From `Downgrading Leader`:
+    /// - `Downgrading Leader -> Leader`
+    /// - `Downgrading Leader -> Follower`
     pub(crate) fn set_role(&self, next_role: RegionRole, region_id: RegionId) {
         match next_role {
             RegionRole::Follower => {
+                if self
+                    .exit_staging(region_id, RegionRoleState::Follower)
+                    .is_ok()
+                {
+                    info!(
+                        "Convert region {} to follower, previous role state: {:?}",
+                        region_id,
+                        RegionRoleState::Leader(RegionLeaderState::Staging)
+                    );
+                    return;
+                }
                 match self.state.fetch_update(|state| {
                     if !matches!(state, RegionRoleState::Follower) {
                         Some(RegionRoleState::Follower)
@@ -1056,6 +1150,20 @@ impl ManifestContext {
                 }
             }
             RegionRole::Leader => {
+                if self
+                    .exit_staging(
+                        region_id,
+                        RegionRoleState::Leader(RegionLeaderState::Writable),
+                    )
+                    .is_ok()
+                {
+                    info!(
+                        "Convert region {} to leader, previous role state: {:?}",
+                        region_id,
+                        RegionRoleState::Leader(RegionLeaderState::Staging)
+                    );
+                    return;
+                }
                 match self.state.fetch_update(|state| {
                     if matches!(
                         state,
@@ -1081,7 +1189,27 @@ impl ManifestContext {
                     }
                 }
             }
+            RegionRole::StagingLeader => {
+                info!(
+                    "Ignore direct conversion of region {} to staging leader; staging requires the dedicated workflow",
+                    region_id
+                );
+            }
             RegionRole::DowngradingLeader => {
+                if self
+                    .exit_staging(
+                        region_id,
+                        RegionRoleState::Leader(RegionLeaderState::Downgrading),
+                    )
+                    .is_ok()
+                {
+                    info!(
+                        "Convert region {} to downgrading region, previous role state: {:?}",
+                        region_id,
+                        RegionRoleState::Leader(RegionLeaderState::Staging)
+                    );
+                    return;
+                }
                 match self.state.compare_exchange(
                     RegionRoleState::Leader(RegionLeaderState::Writable),
                     RegionRoleState::Leader(RegionLeaderState::Downgrading),
@@ -1423,8 +1551,8 @@ pub fn parse_partition_expr(partition_expr_str: Option<&str>) -> Result<Option<P
 
 #[cfg(test)]
 mod tests {
+    use std::sync::Arc;
     use std::sync::atomic::AtomicU64;
-    use std::sync::{Arc, Mutex};
 
     use common_datasource::compression::CompressionType;
     use common_test_util::temp_dir::create_temp_dir;
@@ -1497,7 +1625,6 @@ mod tests {
             topic_latest_entry_id: Default::default(),
             written_bytes: Arc::new(AtomicU64::new(0)),
             stats: ManifestStats::default(),
-            staging_partition_info: Mutex::new(None),
         }
     }
 
@@ -1669,6 +1796,13 @@ mod tests {
             RegionRoleState::Leader(RegionLeaderState::Writable)
         );
 
+        // Direct Leader -> StagingLeader should be ignored.
+        manifest_ctx.set_role(RegionRole::StagingLeader, region_id);
+        assert_eq!(
+            manifest_ctx.state.load(),
+            RegionRoleState::Leader(RegionLeaderState::Writable)
+        );
+
         // Leader -> Downgrading Leader
         manifest_ctx.set_role(RegionRole::DowngradingLeader, region_id);
         assert_eq!(
@@ -1810,7 +1944,6 @@ mod tests {
             topic_latest_entry_id: Default::default(),
             written_bytes: Arc::new(AtomicU64::new(0)),
             stats: ManifestStats::default(),
-            staging_partition_info: Mutex::new(None),
         };
 
         // Test initial state
diff --git a/src/mito2/src/region/opener.rs b/src/mito2/src/region/opener.rs
index 014c50820f..dca26185ca 100644
--- a/src/mito2/src/region/opener.rs
+++ b/src/mito2/src/region/opener.rs
@@ -17,7 +17,7 @@
 use std::any::TypeId;
 use std::collections::HashMap;
 use std::sync::atomic::{AtomicI64, AtomicU64};
-use std::sync::{Arc, LazyLock, Mutex};
+use std::sync::{Arc, LazyLock};
 use std::time::Instant;
 
 use common_telemetry::{debug, error, info, warn};
@@ -70,7 +70,7 @@ use crate::sst::file_ref::FileReferenceManagerRef;
 use crate::sst::index::intermediate::IntermediateManager;
 use crate::sst::index::puffin_manager::PuffinManagerFactory;
 use crate::sst::location::{self, region_dir_from_table_dir};
-use crate::sst::parquet::metadata::MetadataLoader;
+use crate::sst::parquet::metadata::{MetadataLoader, extract_primary_key_range};
 use crate::sst::parquet::reader::MetadataCacheMetrics;
 use crate::time_provider::TimeProviderRef;
 use crate::wal::entry_reader::WalEntryReader;
@@ -269,7 +269,7 @@ impl RegionOpener {
         // Sets the sst_format based on options or flat_format flag
         let sst_format = if let Some(format) = options.sst_format {
             format
-        } else if config.default_experimental_flat_format {
+        } else if config.default_flat_format {
             options.sst_format = Some(FormatType::Flat);
             FormatType::Flat
         } else {
@@ -309,7 +309,7 @@ impl RegionOpener {
 
         debug!(
             "Create region {} with options: {:?}, default_flat_format: {}",
-            region_id, options, config.default_experimental_flat_format
+            region_id, options, config.default_flat_format
         );
 
         let version = VersionBuilder::new(metadata, mutable)
@@ -349,7 +349,6 @@ impl RegionOpener {
             topic_latest_entry_id: AtomicU64::new(0),
             written_bytes: Arc::new(AtomicU64::new(0)),
             stats: self.stats,
-            staging_partition_info: Mutex::new(None),
         }))
     }
 
@@ -586,8 +585,6 @@ impl RegionOpener {
             topic_latest_entry_id: AtomicU64::new(topic_latest_entry_id),
             written_bytes: Arc::new(AtomicU64::new(0)),
             stats: self.stats.clone(),
-            // TODO(weny): reload the staging partition info from the manifest.
-            staging_partition_info: Mutex::new(None),
         };
 
         let region = Arc::new(region);
@@ -626,8 +623,10 @@ pub(crate) fn sanitize_region_options(manifest: &RegionManifest, options: &mut R
             manifest.sst_format,
             manifest.metadata.region_id,
         );
-        options.sst_format = Some(manifest.sst_format);
     }
+    // Always set sst_format from manifest to ensure it's explicitly stored,
+    // even when the default matches the manifest value.
+    options.sst_format = Some(manifest.sst_format);
     if let Some(manifest_append_mode) = manifest.append_mode
         && options.append_mode != manifest_append_mode
     {
@@ -881,7 +880,7 @@ impl RegionLoadCacheTask {
         }
 
         // Sorts files by max timestamp in descending order to loads latest files first
-        files_to_download.sort_by(|a, b| b.2.cmp(&a.2));
+        files_to_download.sort_by_key(|b| std::cmp::Reverse(b.2));
 
         let total_files = files_to_download.len() as i64;
 
@@ -992,6 +991,7 @@ fn maybe_load_cache(
 /// - If the region storage backend is local filesystem (`Scheme::Fs`), it may also load metadata
 ///   directly from the local store.
 /// - It will not fetch metadata from remote object stores (S3/GCS/OSS/...).
+#[allow(clippy::too_many_arguments)]
 async fn preload_parquet_meta_cache_for_files(
     region_id: RegionId,
     cache_manager: CacheManagerRef,
@@ -999,6 +999,7 @@ async fn preload_parquet_meta_cache_for_files(
     table_dir: String,
     path_type: PathType,
     object_store: object_store::ObjectStore,
+    region_metadata: RegionMetadataRef,
     mut files: Vec<FileHandle>,
 ) -> usize {
     if !cache_manager.sst_meta_cache_enabled()
@@ -1008,10 +1009,10 @@ async fn preload_parquet_meta_cache_for_files(
         return 0;
     }
 
-    let allow_direct_load = matches!(object_store.info().scheme(), object_store::Scheme::Fs);
+    let allow_direct_load = object_store.info().scheme() == object_store::services::FS_SCHEME;
 
     // Sort by time range so we can prefer preloading newer files first.
-    files.sort_by(|a, b| b.meta_ref().time_range.1.cmp(&a.meta_ref().time_range.1));
+    files.sort_by_key(|b| std::cmp::Reverse(b.meta_ref().time_range.1));
 
     let mut loaded = 0usize;
     for file_handle in files {
@@ -1022,11 +1023,16 @@ async fn preload_parquet_meta_cache_for_files(
 
         let file_id = file_handle.file_id();
         let mut cache_metrics = MetadataCacheMetrics::default();
-        if cache_manager
+        if let Some(metadata) = cache_manager
             .get_parquet_meta_data(file_id, &mut cache_metrics, Default::default())
             .await
-            .is_some()
         {
+            if file_handle.primary_key_range().is_none()
+                && let Some(primary_key_range) =
+                    extract_primary_key_range(&metadata, &region_metadata)
+            {
+                file_handle.set_primary_key_range(primary_key_range);
+            }
             // Metadata is either already in memory or loaded from file cache.
             if cache_metrics.mem_cache_hit == 0 {
                 loaded += 1;
@@ -1043,7 +1049,12 @@ async fn preload_parquet_meta_cache_for_files(
         let loader = MetadataLoader::new(object_store.clone(), &file_path, file_size);
         match loader.load(&mut cache_metrics).await {
             Ok(metadata) => {
-                cache_manager.put_parquet_meta_data(file_id, Arc::new(metadata));
+                if let Some(primary_key_range) =
+                    extract_primary_key_range(&metadata, &region_metadata)
+                {
+                    file_handle.set_primary_key_range(primary_key_range);
+                }
+                cache_manager.put_parquet_meta_data(file_id, Arc::new(metadata), None);
                 loaded += 1;
             }
             Err(err) => {
@@ -1091,6 +1102,7 @@ fn maybe_preload_parquet_meta_cache(
         let table_dir = region.access_layer.table_dir().to_string();
         let path_type = region.access_layer.path_type();
         let object_store = region.access_layer.object_store().clone();
+        let region_metadata = region.version_control.current().version.metadata.clone();
 
         // Collect SST files. Do not hold the version longer than needed.
         let mut files = Vec::new();
@@ -1110,6 +1122,7 @@ fn maybe_preload_parquet_meta_cache(
             table_dir,
             path_type,
             object_store,
+            region_metadata,
             files,
         )
         .await;
@@ -1148,11 +1161,13 @@ mod tests {
     use common_base::readable_size::ReadableSize;
     use common_test_util::temp_dir::create_temp_dir;
     use common_time::Timestamp;
-    use datatypes::arrow::array::{ArrayRef, Int64Array};
+    use datatypes::arrow::array::{ArrayRef, BinaryArray, Int64Array};
     use datatypes::arrow::record_batch::RecordBatch;
     use object_store::ObjectStore;
     use object_store::services::{Fs, Memory};
     use parquet::arrow::ArrowWriter;
+    use parquet::file::metadata::KeyValue;
+    use parquet::file::properties::WriterProperties;
     use store_api::region_request::PathType;
     use store_api::storage::{FileId, RegionId};
 
@@ -1161,7 +1176,27 @@ mod tests {
     use crate::cache::file_cache::{FileType, IndexKey};
     use crate::sst::file::{FileHandle, FileMeta};
     use crate::sst::file_purger::NoopFilePurger;
+    use crate::sst::parquet::PARQUET_METADATA_KEY;
     use crate::test_util::TestEnv;
+    use crate::test_util::sst_util::sst_region_metadata;
+
+    fn sst_parquet_bytes(batch: &RecordBatch) -> Vec<u8> {
+        let key_value_meta = KeyValue::new(
+            PARQUET_METADATA_KEY.to_string(),
+            sst_region_metadata().to_json().unwrap(),
+        );
+        let props = WriterProperties::builder()
+            .set_key_value_metadata(Some(vec![key_value_meta]))
+            .build();
+
+        let mut parquet_bytes = Vec::new();
+        let mut writer =
+            ArrowWriter::try_new(&mut parquet_bytes, batch.schema(), Some(props)).unwrap();
+        writer.write(batch).unwrap();
+        writer.close().unwrap();
+
+        parquet_bytes
+    }
 
     #[tokio::test]
     async fn test_preload_parquet_meta_cache_uses_file_cache() {
@@ -1182,11 +1217,16 @@ mod tests {
         let file_id = FileId::random();
 
         let col = Arc::new(Int64Array::from_iter_values([1, 2, 3])) as ArrayRef;
-        let batch = RecordBatch::try_from_iter([("col", col)]).unwrap();
-        let mut parquet_bytes = Vec::new();
-        let mut writer = ArrowWriter::try_new(&mut parquet_bytes, batch.schema(), None).unwrap();
-        writer.write(&batch).unwrap();
-        writer.close().unwrap();
+        let primary_key = Arc::new(BinaryArray::from_iter_values([b"a", b"b", b"c"])) as ArrayRef;
+        let batch = RecordBatch::try_from_iter([
+            ("col", col),
+            (
+                store_api::storage::consts::PRIMARY_KEY_COLUMN_NAME,
+                primary_key,
+            ),
+        ])
+        .unwrap();
+        let parquet_bytes = sst_parquet_bytes(&batch);
         let file_size = parquet_bytes.len() as u64;
 
         let file_meta = FileMeta {
@@ -1205,6 +1245,7 @@ mod tests {
             sequence: None,
             partition_expr: None,
             num_series: 0,
+            ..Default::default()
         };
         let file_handle = FileHandle::new(file_meta, Arc::new(NoopFilePurger));
 
@@ -1240,7 +1281,8 @@ mod tests {
             table_dir.to_string(),
             path_type,
             source_store.clone(),
-            vec![file_handle],
+            Arc::new(sst_region_metadata()),
+            vec![file_handle.clone()],
         )
         .await;
 
@@ -1251,6 +1293,7 @@ mod tests {
                 .get_parquet_meta_data_from_mem_cache(region_file_id)
                 .is_some()
         );
+        assert!(file_handle.primary_key_range().is_some());
     }
 
     #[tokio::test]
@@ -1281,6 +1324,7 @@ mod tests {
             sequence: None,
             partition_expr: None,
             num_series: 0,
+            ..Default::default()
         };
         let file_handle = FileHandle::new(file_meta, Arc::new(NoopFilePurger));
 
@@ -1309,6 +1353,7 @@ mod tests {
             table_dir.to_string(),
             path_type,
             object_store,
+            Arc::new(sst_region_metadata()),
             vec![file_handle],
         )
         .await;
@@ -1334,10 +1379,7 @@ mod tests {
 
         let col = Arc::new(Int64Array::from_iter_values([1, 2, 3])) as ArrayRef;
         let batch = RecordBatch::try_from_iter([("col", col)]).unwrap();
-        let mut parquet_bytes = Vec::new();
-        let mut writer = ArrowWriter::try_new(&mut parquet_bytes, batch.schema(), None).unwrap();
-        writer.write(&batch).unwrap();
-        writer.close().unwrap();
+        let parquet_bytes = sst_parquet_bytes(&batch);
 
         // file_size is 0 when it's missing/defaulted in manifests; MetadataLoader::load will stat
         // the local filesystem to retrieve it.
@@ -1357,6 +1399,7 @@ mod tests {
             sequence: None,
             partition_expr: None,
             num_series: 0,
+            ..Default::default()
         };
         let file_handle = FileHandle::new(file_meta, Arc::new(NoopFilePurger));
 
@@ -1384,6 +1427,7 @@ mod tests {
             table_dir.to_string(),
             path_type,
             object_store,
+            Arc::new(sst_region_metadata()),
             vec![file_handle],
         )
         .await;
diff --git a/src/mito2/src/region/options.rs b/src/mito2/src/region/options.rs
index 0fe0a8f12a..f9241bd287 100644
--- a/src/mito2/src/region/options.rs
+++ b/src/mito2/src/region/options.rs
@@ -38,6 +38,8 @@ use crate::memtable::partition_tree::{DEFAULT_FREEZE_THRESHOLD, DEFAULT_MAX_KEYS
 use crate::sst::FormatType;
 
 const DEFAULT_INDEX_SEGMENT_ROW_COUNT: usize = 1024;
+const COMPACTION_TWCS_PREFIX: &str = "compaction.twcs.";
+const MEMTABLE_PARTITION_TREE_PREFIX: &str = "memtable.partition_tree.";
 
 pub(crate) fn parse_wal_options(
     options_map: &HashMap<String, String>,
@@ -50,7 +52,7 @@ pub(crate) fn parse_wal_options(
 }
 
 /// Mode to handle duplicate rows while merging.
-#[derive(Debug, Default, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, EnumString)]
+#[derive(Debug, Default, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize, EnumString)]
 #[serde(rename_all = "snake_case")]
 #[strum(serialize_all = "snake_case")]
 pub enum MergeMode {
@@ -96,9 +98,10 @@ impl RegionOptions {
     pub fn validate(&self) -> Result<()> {
         if self.append_mode {
             ensure!(
-                self.merge_mode.is_none(),
+                self.merge_mode
+                    .is_none_or(|mode| mode == MergeMode::LastRow),
                 InvalidRegionOptionsSnafu {
-                    reason: "merge_mode is not allowed when append_mode is enabled",
+                    reason: "only last_row merge_mode is allowed when append_mode is enabled",
                 }
             );
         }
@@ -137,7 +140,8 @@ impl TryFrom<&HashMap<String, String>> for RegionOptions {
         // See https://github.com/serde-rs/serde/issues/1626
         let options: RegionOptionsWithoutEnum =
             serde_json::from_str(&json).context(JsonOptionsSnafu)?;
-        let has_compaction_type = validate_enum_options(options_map, "compaction.type")?;
+        let has_compaction_type =
+            validate_enum_options(options_map, "compaction.type", &[COMPACTION_TWCS_PREFIX])?;
         let compaction = if has_compaction_type {
             serde_json::from_str(&json).context(JsonOptionsSnafu)?
         } else {
@@ -147,7 +151,11 @@ impl TryFrom<&HashMap<String, String>> for RegionOptions {
         let wal_options = parse_wal_options(options_map).context(JsonOptionsSnafu)?;
 
         let index_options: IndexOptions = serde_json::from_str(&json).context(JsonOptionsSnafu)?;
-        let memtable = if validate_enum_options(options_map, "memtable.type")? {
+        let memtable = if validate_enum_options(
+            options_map,
+            "memtable.type",
+            &[MEMTABLE_PARTITION_TREE_PREFIX],
+        )? {
             Some(serde_json::from_str(&json).context(JsonOptionsSnafu)?)
         } else {
             None
@@ -436,25 +444,36 @@ fn options_map_to_value(options: &HashMap<String, String>) -> Value {
 // `#[serde(default)]` doesn't support enum (https://github.com/serde-rs/serde/issues/1799) so we
 // check the type key first.
 /// Validates whether the `options_map` has valid options for specific `enum_tag_key`
-/// and returns `true` if the map contains enum options.
+/// and returns `true` if the map contains the enum tag.
+///
+/// Variant options must start with one of `enum_option_prefixes`. If variant options
+/// are provided, the tagged enum type key must also be provided.
 fn validate_enum_options(
     options_map: &HashMap<String, String>,
     enum_tag_key: &str,
+    enum_option_prefixes: &[&str],
 ) -> Result<bool> {
-    let enum_type = enum_tag_key.split('.').next().unwrap();
-    let mut has_other_options = false;
+    let mut has_enum_options = false;
     let mut has_tag = false;
     for key in options_map.keys() {
         if key == enum_tag_key {
             has_tag = true;
-        } else if key.starts_with(enum_type) {
-            has_other_options = true;
+        } else if !has_enum_options
+            && enum_option_prefixes
+                .iter()
+                .any(|prefix| key.starts_with(prefix))
+        {
+            has_enum_options = true;
+        }
+
+        if has_tag && has_enum_options {
+            break;
         }
     }
 
     // If tag is not provided, then other options for the enum should not exist.
     ensure!(
-        has_tag || !has_other_options,
+        has_tag || !has_enum_options,
         InvalidRegionOptionsSnafu {
             reason: format!("missing key {} in options", enum_tag_key),
         }
@@ -537,6 +556,34 @@ mod tests {
         assert_eq!(expect, options);
     }
 
+    #[test]
+    fn test_with_compaction_override_true_without_compaction_type() {
+        let map = make_map(&[(COMPACTION_OVERRIDE, "true")]);
+        let options = RegionOptions::try_from(&map).unwrap();
+        let expect = RegionOptions {
+            compaction_override: true,
+            ..Default::default()
+        };
+        assert_eq!(expect, options);
+    }
+
+    #[test]
+    fn test_with_compaction_override_false_without_compaction_type() {
+        let map = make_map(&[(COMPACTION_OVERRIDE, "false")]);
+        let options = RegionOptions::try_from(&map).unwrap();
+        assert_eq!(RegionOptions::default(), options);
+    }
+
+    #[test]
+    fn test_compaction_twcs_options_still_require_compaction_type_with_override() {
+        let map = make_map(&[
+            (COMPACTION_OVERRIDE, "true"),
+            ("compaction.twcs.time_window", "2h"),
+        ]);
+        let err = RegionOptions::try_from(&map).unwrap_err();
+        assert_eq!(StatusCode::InvalidArguments, err.status_code());
+    }
+
     fn test_with_wal_options(wal_options: &WalOptions) -> bool {
         let encoded_wal_options = serde_json::to_string(&wal_options).unwrap();
         let map = make_map(&[(WAL_OPTIONS_KEY, &encoded_wal_options)]);
@@ -607,6 +654,13 @@ mod tests {
         assert_eq!(StatusCode::InvalidArguments, err.status_code());
     }
 
+    #[test]
+    fn test_without_memtable_type() {
+        let map = make_map(&[("memtable.partition_tree.index_max_keys_per_shard", "2048")]);
+        let err = RegionOptions::try_from(&map).unwrap_err();
+        assert_eq!(StatusCode::InvalidArguments, err.status_code());
+    }
+
     #[test]
     fn test_with_merge_mode() {
         let map = make_map(&[("merge_mode", "last_row")]);
@@ -622,6 +676,18 @@ mod tests {
         assert_eq!(StatusCode::InvalidArguments, err.status_code());
     }
 
+    #[test]
+    fn test_append_mode_allows_last_row_merge_mode() {
+        let map = make_map(&[("append_mode", "true"), ("merge_mode", "last_row")]);
+        let options = RegionOptions::try_from(&map).unwrap();
+        assert!(options.append_mode);
+        assert_eq!(MergeMode::LastRow, options.merge_mode());
+
+        let map = make_map(&[("append_mode", "true"), ("merge_mode", "last_non_null")]);
+        let err = RegionOptions::try_from(&map).unwrap_err();
+        assert_eq!(StatusCode::InvalidArguments, err.status_code());
+    }
+
     #[test]
     fn test_with_all() {
         let wal_options = WalOptions::Kafka(KafkaWalOptions {
diff --git a/src/mito2/src/remap_manifest.rs b/src/mito2/src/remap_manifest.rs
index 49676be8be..596e60c6dd 100644
--- a/src/mito2/src/remap_manifest.rs
+++ b/src/mito2/src/remap_manifest.rs
@@ -436,6 +436,7 @@ mod tests {
             sequence: NonZeroU64::new(1),
             partition_expr,
             num_series: 1,
+            ..Default::default()
         }
     }
 
diff --git a/src/mito2/src/request.rs b/src/mito2/src/request.rs
index 4f8f1b0d51..fefac39e6a 100644
--- a/src/mito2/src/request.rs
+++ b/src/mito2/src/request.rs
@@ -46,9 +46,9 @@ use store_api::storage::{FileId, RegionId};
 use tokio::sync::oneshot::{self, Receiver, Sender};
 
 use crate::error::{
-    CompactRegionSnafu, ConvertColumnDataTypeSnafu, CreateDefaultSnafu, Error, FillDefaultSnafu,
-    FlushRegionSnafu, InvalidPartitionExprSnafu, InvalidRequestSnafu, MissingPartitionExprSnafu,
-    Result, UnexpectedSnafu,
+    CompactRegionSnafu, CompactionCancelledSnafu, ConvertColumnDataTypeSnafu, CreateDefaultSnafu,
+    Error, FillDefaultSnafu, FlushRegionSnafu, InvalidPartitionExprSnafu, InvalidRequestSnafu,
+    MissingPartitionExprSnafu, Result, UnexpectedSnafu,
 };
 use crate::flush::FlushReason;
 use crate::manifest::action::{RegionEdit, TruncateKind};
@@ -556,6 +556,13 @@ pub(crate) struct SenderBulkRequest {
     pub(crate) partition_expr_version: Option<u64>,
 }
 
+#[derive(Debug)]
+pub(crate) struct BulkInsertRequest {
+    pub(crate) metadata: Option<RegionMetadataRef>,
+    pub(crate) request: RegionBulkInsertsRequest,
+    pub(crate) sender: OptionOutputTx,
+}
+
 /// Request sent to a worker with timestamp
 #[derive(Debug)]
 pub(crate) struct WorkerRequestWithTime {
@@ -609,11 +616,7 @@ pub(crate) enum WorkerRequest {
     SyncRegion(RegionSyncRequest),
 
     /// Bulk inserts request and region metadata.
-    BulkInserts {
-        metadata: Option<RegionMetadataRef>,
-        request: RegionBulkInsertsRequest,
-        sender: OptionOutputTx,
-    },
+    BulkInserts(BulkInsertRequest),
 
     /// Remap manifests request.
     RemapManifests(RemapManifestsRequest),
@@ -748,11 +751,13 @@ impl WorkerRequest {
                 sender: sender.into(),
                 request: DdlRequest::EnterStaging(v),
             }),
-            RegionRequest::BulkInserts(region_bulk_inserts_request) => WorkerRequest::BulkInserts {
-                metadata: region_metadata,
-                sender: sender.into(),
-                request: region_bulk_inserts_request,
-            },
+            RegionRequest::BulkInserts(region_bulk_inserts_request) => {
+                WorkerRequest::BulkInserts(BulkInsertRequest {
+                    metadata: region_metadata,
+                    sender: sender.into(),
+                    request: region_bulk_inserts_request,
+                })
+            }
             RegionRequest::ApplyStagingManifest(v) => WorkerRequest::Ddl(SenderDdlRequest {
                 region_id,
                 sender: sender.into(),
@@ -895,6 +900,8 @@ pub(crate) enum BackgroundNotify {
     IndexBuildFailed(IndexBuildFailed),
     /// Compaction has finished.
     CompactionFinished(CompactionFinished),
+    /// Compaction has been cancelled cooperatively.
+    CompactionCancelled(CompactionCancelled),
     /// Compaction has failed.
     CompactionFailed(CompactionFailed),
     /// Truncate result.
@@ -991,6 +998,24 @@ pub(crate) struct CompactionFinished {
     pub(crate) edit: RegionEdit,
 }
 
+/// Notifies a compaction job has been cancelled cooperatively.
+#[derive(Debug)]
+pub(crate) struct CompactionCancelled {
+    /// Region id.
+    pub(crate) region_id: RegionId,
+    /// Waiters to wake once the cancellation has been observed by the worker.
+    pub(crate) senders: Vec<OutputTx>,
+}
+
+impl CompactionCancelled {
+    pub(crate) fn on_success(self) {
+        for sender in self.senders {
+            sender.send(CompactionCancelledSnafu {}.fail());
+        }
+        info!("Compaction cancelled for region: {}", self.region_id);
+    }
+}
+
 impl CompactionFinished {
     pub fn on_success(self) {
         // only update compaction time on success
@@ -1149,10 +1174,13 @@ pub(crate) struct CopyRegionFromRequest {
 mod tests {
     use api::v1::value::ValueData;
     use api::v1::{Row, SemanticType};
+    use common_error::ext::ErrorExt;
+    use common_error::status_code::StatusCode;
     use datatypes::prelude::ConcreteDataType;
     use datatypes::schema::ColumnDefaultConstraint;
     use mito_codec::test_util::i64_value;
     use store_api::metadata::RegionMetadataBuilder;
+    use tokio::sync::oneshot;
 
     use super::*;
     use crate::error::Error;
@@ -1216,6 +1244,21 @@ mod tests {
         assert_eq!(None, request.column_index_by_name("c2"));
     }
 
+    #[test]
+    fn test_compaction_cancelled_sends_cancelled_error() {
+        let (tx, rx) = oneshot::channel();
+        let request = CompactionCancelled {
+            region_id: RegionId::new(1, 1),
+            senders: vec![OutputTx::new(tx)],
+        };
+
+        request.on_success();
+
+        let err = rx.blocking_recv().unwrap().unwrap_err();
+        assert!(matches!(err, Error::CompactionCancelled { .. }));
+        assert_eq!(err.status_code(), StatusCode::Cancelled);
+    }
+
     #[test]
     fn test_write_request_column_num() {
         let rows = Rows {
diff --git a/src/mito2/src/sst.rs b/src/mito2/src/sst.rs
index 78e4c563b1..1007b57668 100644
--- a/src/mito2/src/sst.rs
+++ b/src/mito2/src/sst.rs
@@ -31,7 +31,6 @@ use store_api::storage::consts::{
     OP_TYPE_COLUMN_NAME, PRIMARY_KEY_COLUMN_NAME, SEQUENCE_COLUMN_NAME,
 };
 
-use crate::read::Batch;
 use crate::sst::parquet::flat_format::time_index_column_index;
 
 pub mod file;
@@ -60,6 +59,17 @@ pub enum FormatType {
     Flat,
 }
 
+/// Iceberg-compatible column field ID key stored in Parquet column metadata.
+pub const PARQUET_FIELD_ID_KEY: &str = "PARQUET:field_id";
+
+/// Adds `PARQUET:field_id` metadata to an Arrow field.
+pub fn with_field_id(mut field: Field, column_id: u32) -> Field {
+    field
+        .metadata_mut()
+        .insert(PARQUET_FIELD_ID_KEY.to_string(), column_id.to_string());
+    field
+}
+
 /// Gets the arrow schema to store in parquet.
 pub fn to_sst_arrow_schema(metadata: &RegionMetadata) -> SchemaRef {
     let fields = Fields::from_iter(
@@ -71,13 +81,19 @@ pub fn to_sst_arrow_schema(metadata: &RegionMetadata) -> SchemaRef {
             .zip(&metadata.column_metadatas)
             .filter_map(|(field, column_meta)| {
                 if column_meta.semantic_type == SemanticType::Field {
-                    Some(field.clone())
+                    Some(Arc::new(with_field_id(
+                        (**field).clone(),
+                        column_meta.column_id,
+                    )))
                 } else {
                     // We have fixed positions for tags (primary key) and time index.
                     None
                 }
             })
-            .chain([metadata.time_index_field()])
+            .chain([Arc::new(with_field_id(
+                (*metadata.time_index_field()).clone(),
+                metadata.time_index_column().column_id,
+            ))])
             .chain(internal_fields()),
     );
 
@@ -136,13 +152,14 @@ pub fn to_flat_sst_arrow_schema(
     if options.raw_pk_columns {
         for pk_id in &metadata.primary_key {
             let pk_index = metadata.column_index_by_id(*pk_id).unwrap();
+            let column_id = metadata.column_metadatas[pk_index].column_id;
             if options.string_pk_use_dict {
                 let old_field = &schema.fields[pk_index];
                 let new_field = tag_maybe_to_dictionary_field(
                     &metadata.column_metadatas[pk_index].column_schema.data_type,
                     old_field,
                 );
-                fields.push(new_field);
+                fields.push(Arc::new(with_field_id((*new_field).clone(), column_id)));
             }
         }
     }
@@ -152,12 +169,18 @@ pub fn to_flat_sst_arrow_schema(
         .zip(&metadata.column_metadatas)
         .filter_map(|(field, column_meta)| {
             if column_meta.semantic_type == SemanticType::Field {
-                Some(field.clone())
+                Some(Arc::new(with_field_id(
+                    (**field).clone(),
+                    column_meta.column_id,
+                )))
             } else {
                 None
             }
         })
-        .chain([metadata.time_index_field()])
+        .chain([Arc::new(with_field_id(
+            (*metadata.time_index_field()).clone(),
+            metadata.time_index_column().column_id,
+        ))])
         .chain(internal_fields());
     for field in remaining_fields {
         fields.push(field);
@@ -180,12 +203,21 @@ pub fn flat_sst_arrow_schema_column_num(
 
 /// Helper function to create a dictionary field from a field.
 fn to_dictionary_field(field: &Field) -> Field {
-    Field::new_dictionary(
+    let mut new_field = Field::new_dictionary(
         field.name(),
         datatypes::arrow::datatypes::DataType::UInt32,
         field.data_type().clone(),
         field.is_nullable(),
-    )
+    );
+
+    // retain field_id metadata
+    if let Some(field_id) = field.metadata().get(PARQUET_FIELD_ID_KEY) {
+        new_field
+            .metadata_mut()
+            .insert(PARQUET_FIELD_ID_KEY.to_string(), field_id.clone());
+    }
+
+    new_field
 }
 
 /// Helper function to create a dictionary field from a field if it is a string column.
@@ -219,34 +251,6 @@ pub(crate) fn internal_fields() -> [FieldRef; 3] {
     ]
 }
 
-/// Gets the arrow schema to store in parquet.
-pub fn to_plain_sst_arrow_schema(metadata: &RegionMetadata) -> SchemaRef {
-    let fields = Fields::from_iter(
-        metadata
-            .schema
-            .arrow_schema()
-            .fields()
-            .iter()
-            .cloned()
-            .chain(plain_internal_fields()),
-    );
-
-    Arc::new(Schema::new(fields))
-}
-
-/// Fields for internal columns.
-fn plain_internal_fields() -> [FieldRef; 2] {
-    // Internal columns are always not null.
-    [
-        Arc::new(Field::new(
-            SEQUENCE_COLUMN_NAME,
-            ArrowDataType::UInt64,
-            false,
-        )),
-        Arc::new(Field::new(OP_TYPE_COLUMN_NAME, ArrowDataType::UInt8, false)),
-    ]
-}
-
 /// Gets the estimated number of series from record batches.
 ///
 /// This struct tracks the last timestamp value to detect series boundaries
@@ -260,33 +264,6 @@ pub(crate) struct SeriesEstimator {
 }
 
 impl SeriesEstimator {
-    /// Updates the estimator with a new Batch.
-    ///
-    /// Since each Batch contains only one series, this increments the series count
-    /// and updates the last timestamp.
-    pub(crate) fn update(&mut self, batch: &Batch) {
-        let Some(last_ts) = batch.last_timestamp() else {
-            return;
-        };
-
-        // Checks if there's a boundary between the last batch and this batch
-        if let Some(prev_last_ts) = self.last_timestamp {
-            // If the first timestamp of this batch is less than the last timestamp
-            // we've seen, it indicates a new series
-            if let Some(first_ts) = batch.first_timestamp()
-                && first_ts.value() <= prev_last_ts
-            {
-                self.series_count += 1;
-            }
-        } else {
-            // First batch, counts as first series
-            self.series_count = 1;
-        }
-
-        // Updates the last timestamp
-        self.last_timestamp = Some(last_ts.value());
-    }
-
     /// Updates the estimator with a new record batch in flat format.
     ///
     /// This method examines the time index column to detect series boundaries.
@@ -340,43 +317,14 @@ impl SeriesEstimator {
 mod tests {
     use std::sync::Arc;
 
-    use api::v1::OpType;
     use datatypes::arrow::array::{
-        BinaryArray, DictionaryArray, TimestampMillisecondArray, UInt8Array, UInt8Builder,
-        UInt32Array, UInt64Array,
+        BinaryArray, DictionaryArray, TimestampMillisecondArray, UInt8Array, UInt32Array,
+        UInt64Array,
     };
     use datatypes::arrow::datatypes::{DataType as ArrowDataType, Field, Schema, TimeUnit};
     use datatypes::arrow::record_batch::RecordBatch;
 
     use super::*;
-    use crate::read::{Batch, BatchBuilder};
-
-    fn new_batch(
-        primary_key: &[u8],
-        timestamps: &[i64],
-        sequences: &[u64],
-        op_types: &[OpType],
-    ) -> Batch {
-        let timestamps = Arc::new(TimestampMillisecondArray::from(timestamps.to_vec()));
-        let sequences = Arc::new(UInt64Array::from(sequences.to_vec()));
-        let mut op_type_builder = UInt8Builder::with_capacity(op_types.len());
-        for op_type in op_types {
-            op_type_builder.append_value(*op_type as u8);
-        }
-        let op_types = Arc::new(UInt8Array::from(
-            op_types.iter().map(|op| *op as u8).collect::<Vec<_>>(),
-        ));
-
-        let mut builder = BatchBuilder::new(primary_key.to_vec());
-        builder
-            .timestamps_array(timestamps)
-            .unwrap()
-            .sequences_array(sequences)
-            .unwrap()
-            .op_types_array(op_types)
-            .unwrap();
-        builder.build().unwrap()
-    }
 
     fn new_flat_record_batch(timestamps: &[i64]) -> RecordBatch {
         // Flat format has: [fields..., time_index, __primary_key, __sequence, __op_type]
@@ -411,128 +359,6 @@ mod tests {
         RecordBatch::try_new(schema, vec![time_array, pk_array, seq_array, op_array]).unwrap()
     }
 
-    #[test]
-    fn test_series_estimator_empty_batch() {
-        let mut estimator = SeriesEstimator::default();
-        let batch = new_batch(b"test", &[], &[], &[]);
-        estimator.update(&batch);
-        assert_eq!(0, estimator.finish());
-    }
-
-    #[test]
-    fn test_series_estimator_single_batch() {
-        let mut estimator = SeriesEstimator::default();
-        let batch = new_batch(
-            b"test",
-            &[1, 2, 3],
-            &[1, 2, 3],
-            &[OpType::Put, OpType::Put, OpType::Put],
-        );
-        estimator.update(&batch);
-        assert_eq!(1, estimator.finish());
-    }
-
-    #[test]
-    fn test_series_estimator_multiple_batches_same_series() {
-        let mut estimator = SeriesEstimator::default();
-
-        // First batch with timestamps 1, 2, 3
-        let batch1 = new_batch(
-            b"test",
-            &[1, 2, 3],
-            &[1, 2, 3],
-            &[OpType::Put, OpType::Put, OpType::Put],
-        );
-        estimator.update(&batch1);
-
-        // Second batch with timestamps 4, 5, 6 (continuation)
-        let batch2 = new_batch(
-            b"test",
-            &[4, 5, 6],
-            &[4, 5, 6],
-            &[OpType::Put, OpType::Put, OpType::Put],
-        );
-        estimator.update(&batch2);
-
-        assert_eq!(1, estimator.finish());
-    }
-
-    #[test]
-    fn test_series_estimator_new_series_detected() {
-        let mut estimator = SeriesEstimator::default();
-
-        // First batch with timestamps 1, 2, 3
-        let batch1 = new_batch(
-            b"pk0",
-            &[1, 2, 3],
-            &[1, 2, 3],
-            &[OpType::Put, OpType::Put, OpType::Put],
-        );
-        estimator.update(&batch1);
-
-        // Second batch with timestamps 2, 3, 4 (timestamp goes back, new series)
-        let batch2 = new_batch(
-            b"pk1",
-            &[2, 3, 4],
-            &[4, 5, 6],
-            &[OpType::Put, OpType::Put, OpType::Put],
-        );
-        estimator.update(&batch2);
-
-        assert_eq!(2, estimator.finish());
-    }
-
-    #[test]
-    fn test_series_estimator_equal_timestamp_boundary() {
-        let mut estimator = SeriesEstimator::default();
-
-        // First batch ending at timestamp 5
-        let batch1 = new_batch(
-            b"test",
-            &[1, 2, 5],
-            &[1, 2, 3],
-            &[OpType::Put, OpType::Put, OpType::Put],
-        );
-        estimator.update(&batch1);
-
-        // Second batch starting at timestamp 5 (equal, indicates new series)
-        let batch2 = new_batch(
-            b"test",
-            &[5, 6, 7],
-            &[4, 5, 6],
-            &[OpType::Put, OpType::Put, OpType::Put],
-        );
-        estimator.update(&batch2);
-
-        assert_eq!(2, estimator.finish());
-    }
-
-    #[test]
-    fn test_series_estimator_finish_resets_state() {
-        let mut estimator = SeriesEstimator::default();
-
-        let batch1 = new_batch(
-            b"test",
-            &[1, 2, 3],
-            &[1, 2, 3],
-            &[OpType::Put, OpType::Put, OpType::Put],
-        );
-        estimator.update(&batch1);
-
-        assert_eq!(1, estimator.finish());
-
-        // After finish, state should be reset
-        let batch2 = new_batch(
-            b"test",
-            &[4, 5, 6],
-            &[4, 5, 6],
-            &[OpType::Put, OpType::Put, OpType::Put],
-        );
-        estimator.update(&batch2);
-
-        assert_eq!(1, estimator.finish());
-    }
-
     #[test]
     fn test_series_estimator_flat_empty_batch() {
         let mut estimator = SeriesEstimator::default();
diff --git a/src/mito2/src/sst/file.rs b/src/mito2/src/sst/file.rs
index 74176831b6..213cd4f714 100644
--- a/src/mito2/src/sst/file.rs
+++ b/src/mito2/src/sst/file.rs
@@ -17,9 +17,11 @@
 use std::fmt;
 use std::fmt::{Debug, Formatter};
 use std::num::NonZeroU64;
-use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::{Arc, RwLock};
 
+use base64::prelude::{BASE64_STANDARD, Engine};
+use bytes::Bytes;
 use common_base::readable_size::ReadableSize;
 use common_telemetry::{debug, error};
 use common_time::Timestamp;
@@ -36,6 +38,33 @@ use crate::cache::file_cache::{FileType, IndexKey};
 use crate::sst::file_purger::FilePurgerRef;
 use crate::sst::location;
 
+/// Custom serde functions for Bytes fields serialized as base64 strings.
+fn serialize_bytes_option<S>(bytes: &Option<Bytes>, serializer: S) -> Result<S::Ok, S::Error>
+where
+    S: serde::Serializer,
+{
+    match bytes {
+        None => serializer.serialize_none(),
+        Some(b) => serializer.serialize_some(&BASE64_STANDARD.encode(b)),
+    }
+}
+
+fn deserialize_bytes_option<'de, D>(deserializer: D) -> Result<Option<Bytes>, D::Error>
+where
+    D: serde::Deserializer<'de>,
+{
+    let opt: Option<String> = Option::deserialize(deserializer)?;
+    match opt {
+        None => Ok(None),
+        Some(s) => {
+            let decoded = BASE64_STANDARD
+                .decode(&s)
+                .map_err(serde::de::Error::custom)?;
+            Ok(Some(Bytes::from(decoded)))
+        }
+    }
+}
+
 /// Custom serde functions for partition_expr field in FileMeta
 fn serialize_partition_expr<S>(
     partition_expr: &Option<PartitionExpr>,
@@ -233,6 +262,24 @@ pub struct FileMeta {
     ///
     /// The number is 0 if the series number is not available.
     pub num_series: u64,
+    /// Minimum primary key value in the file, encoded as bytes.
+    /// `None` if the primary key range is not available (e.g., legacy files).
+    #[serde(
+        default,
+        skip_serializing_if = "Option::is_none",
+        serialize_with = "serialize_bytes_option",
+        deserialize_with = "deserialize_bytes_option"
+    )]
+    pub primary_key_min: Option<Bytes>,
+    /// Maximum primary key value in the file, encoded as bytes.
+    /// `None` if the primary key range is not available (e.g., legacy files).
+    #[serde(
+        default,
+        skip_serializing_if = "Option::is_none",
+        serialize_with = "serialize_bytes_option",
+        deserialize_with = "deserialize_bytes_option"
+    )]
+    pub primary_key_max: Option<Bytes>,
 }
 
 impl Debug for FileMeta {
@@ -273,8 +320,19 @@ impl Debug for FileMeta {
                 }
             })
             .field("partition_expr", &self.partition_expr)
-            .field("num_series", &self.num_series)
-            .finish()
+            .field("num_series", &self.num_series);
+        if self.primary_key_min.is_some() || self.primary_key_max.is_some() {
+            debug_struct
+                .field(
+                    "primary_key_min",
+                    &self.primary_key_min.as_ref().map(|b| b.len()),
+                )
+                .field(
+                    "primary_key_max",
+                    &self.primary_key_max.as_ref().map(|b| b.len()),
+                );
+        }
+        debug_struct.finish()
     }
 }
 
@@ -311,6 +369,14 @@ pub struct ColumnIndexMetadata {
 }
 
 impl FileMeta {
+    /// Returns the primary key range if both min and max are present.
+    pub fn primary_key_range(&self) -> Option<(Bytes, Bytes)> {
+        match (&self.primary_key_min, &self.primary_key_max) {
+            (Some(min), Some(max)) => Some((min.clone(), max.clone())),
+            _ => None,
+        }
+    }
+
     pub fn exists_index(&self) -> bool {
         !self.available_indexes.is_empty()
     }
@@ -323,7 +389,7 @@ impl FileMeta {
         }
     }
 
-    /// Whether the index file is up-to-date comparing to another file meta.    
+    /// Whether the index file is up-to-date comparing to another file meta.
     pub fn is_index_up_to_date(&self, other: &FileMeta) -> bool {
         self.exists_index() && other.exists_index() && self.index_version >= other.index_version
     }
@@ -417,8 +483,20 @@ impl fmt::Debug for FileHandle {
 
 impl FileHandle {
     pub fn new(meta: FileMeta, file_purger: FilePurgerRef) -> FileHandle {
+        let pk_range = meta.primary_key_range();
         FileHandle {
-            inner: Arc::new(FileHandleInner::new(meta, file_purger)),
+            inner: Arc::new(FileHandleInner::new(meta, file_purger, pk_range)),
+        }
+    }
+
+    #[cfg(test)]
+    pub fn new_with_primary_key_range(
+        meta: FileMeta,
+        file_purger: FilePurgerRef,
+        primary_key_range: Option<(Bytes, Bytes)>,
+    ) -> FileHandle {
+        FileHandle {
+            inner: Arc::new(FileHandleInner::new(meta, file_purger, primary_key_range)),
         }
     }
 
@@ -498,6 +576,14 @@ impl FileHandle {
     pub fn is_deleted(&self) -> bool {
         self.inner.deleted.load(Ordering::Relaxed)
     }
+
+    pub fn primary_key_range(&self) -> Option<(Bytes, Bytes)> {
+        self.inner.primary_key_range.read().unwrap().clone()
+    }
+
+    pub(crate) fn set_primary_key_range(&self, primary_key_range: (Bytes, Bytes)) {
+        *self.inner.primary_key_range.write().unwrap() = Some(primary_key_range);
+    }
 }
 
 /// Inner data of [FileHandle].
@@ -508,6 +594,7 @@ struct FileHandleInner {
     compacting: AtomicBool,
     deleted: AtomicBool,
     index_outdated: AtomicBool,
+    primary_key_range: RwLock<Option<(Bytes, Bytes)>>,
     file_purger: FilePurgerRef,
 }
 
@@ -523,13 +610,18 @@ impl Drop for FileHandleInner {
 
 impl FileHandleInner {
     /// There should only be one `FileHandleInner` for each file on a datanode
-    fn new(meta: FileMeta, file_purger: FilePurgerRef) -> FileHandleInner {
+    fn new(
+        meta: FileMeta,
+        file_purger: FilePurgerRef,
+        primary_key_range: Option<(Bytes, Bytes)>,
+    ) -> FileHandleInner {
         file_purger.new_file(&meta);
         FileHandleInner {
             meta,
             compacting: AtomicBool::new(false),
             deleted: AtomicBool::new(false),
             index_outdated: AtomicBool::new(false),
+            primary_key_range: RwLock::new(primary_key_range),
             file_purger,
         }
     }
@@ -734,6 +826,7 @@ mod tests {
             sequence: None,
             partition_expr: None,
             num_series: 0,
+            ..Default::default()
         }
     }
 
@@ -786,6 +879,7 @@ mod tests {
             sequence: None,
             partition_expr: Some(partition_expr.clone()),
             num_series: 0,
+            ..Default::default()
         };
 
         // Test serialization/deserialization
diff --git a/src/mito2/src/sst/file_purger.rs b/src/mito2/src/sst/file_purger.rs
index 7c1598787e..70c14915af 100644
--- a/src/mito2/src/sst/file_purger.rs
+++ b/src/mito2/src/sst/file_purger.rs
@@ -70,20 +70,14 @@ impl fmt::Debug for LocalFilePurger {
 
 #[cfg(not(debug_assertions))]
 /// Whether to enable GC for the file purger.
-pub fn should_enable_gc(
-    global_gc_enabled: bool,
-    object_store_scheme: object_store::Scheme,
-) -> bool {
-    global_gc_enabled && object_store_scheme != object_store::Scheme::Fs
+pub fn should_enable_gc(global_gc_enabled: bool, object_store_scheme: &'static str) -> bool {
+    global_gc_enabled && object_store_scheme != object_store::services::FS_SCHEME
 }
 
 #[cfg(debug_assertions)]
 /// For debug build, we may use Fs as the object store scheme,
 /// so we need to enable GC for local file system.
-pub fn should_enable_gc(
-    global_gc_enabled: bool,
-    _object_store_scheme: object_store::Scheme,
-) -> bool {
+pub fn should_enable_gc(global_gc_enabled: bool, _object_store_scheme: &'static str) -> bool {
     global_gc_enabled
 }
 
@@ -283,6 +277,7 @@ mod tests {
                     sequence: None,
                     partition_expr: None,
                     num_series: 0,
+                    ..Default::default()
                 },
                 file_purger,
             );
@@ -357,6 +352,7 @@ mod tests {
                     sequence: NonZeroU64::new(4096),
                     partition_expr: None,
                     num_series: 0,
+                    ..Default::default()
                 },
                 file_purger,
             );
diff --git a/src/mito2/src/sst/file_ref.rs b/src/mito2/src/sst/file_ref.rs
index 6490d2d1dc..2f067cf3c1 100644
--- a/src/mito2/src/sst/file_ref.rs
+++ b/src/mito2/src/sst/file_ref.rs
@@ -285,6 +285,7 @@ mod tests {
             sequence: NonZeroU64::new(4096),
             partition_expr: None,
             num_series: 0,
+            ..Default::default()
         };
 
         file_ref_mgr.add_file(&file_meta);
diff --git a/src/mito2/src/sst/index.rs b/src/mito2/src/sst/index.rs
index 0df3229e9c..31a96eecea 100644
--- a/src/mito2/src/sst/index.rs
+++ b/src/mito2/src/sst/index.rs
@@ -58,7 +58,7 @@ use crate::error::{
 };
 use crate::manifest::action::{RegionEdit, RegionMetaAction, RegionMetaActionList};
 use crate::metrics::INDEX_CREATE_MEMORY_USAGE;
-use crate::read::{Batch, BatchReader};
+use crate::read::Batch;
 use crate::region::options::IndexOptions;
 use crate::region::version::VersionControlRef;
 use crate::region::{ManifestContextRef, RegionLeaderState};
@@ -802,9 +802,9 @@ impl IndexBuildTask {
         if let Some(mut parquet_reader) = parquet_reader {
             // TODO(SNC123): optimize index batch
             loop {
-                match parquet_reader.next_batch().await {
-                    Ok(Some(mut batch)) => {
-                        indexer.update(&mut batch).await;
+                match parquet_reader.next_record_batch().await {
+                    Ok(Some(batch)) => {
+                        indexer.update_flat(&batch).await;
                     }
                     Ok(None) => break,
                     Err(e) => {
@@ -1083,8 +1083,8 @@ impl IndexBuildScheduler {
     /// Find the next task which has the highest priority to run.
     fn find_next_task(&self) -> Option<IndexBuildTask> {
         self.region_status
-            .iter()
-            .filter_map(|(_, status)| status.pending_tasks.peek())
+            .values()
+            .filter_map(|status| status.pending_tasks.peek())
             .max()
             .cloned()
     }
@@ -1227,7 +1227,9 @@ mod tests {
     use crate::sst::parquet::WriteOptions;
     use crate::test_util::memtable_util::EmptyMemtableBuilder;
     use crate::test_util::scheduler_util::SchedulerEnv;
-    use crate::test_util::sst_util::{new_batch_by_range, new_source, sst_region_metadata};
+    use crate::test_util::sst_util::{
+        new_flat_source_from_record_batches, new_record_batch_by_range, sst_region_metadata,
+    };
 
     struct MetaConfig {
         with_inverted: bool,
@@ -1358,19 +1360,20 @@ mod tests {
         env: &SchedulerEnv,
         build_mode: IndexBuildMode,
     ) -> SstInfo {
-        let source = new_source(&[
-            new_batch_by_range(&["a", "d"], 0, 60),
-            new_batch_by_range(&["b", "f"], 0, 40),
-            new_batch_by_range(&["b", "h"], 100, 200),
+        let source = new_flat_source_from_record_batches(vec![
+            new_record_batch_by_range(&["a", "d"], 0, 60),
+            new_record_batch_by_range(&["b", "f"], 0, 40),
+            new_record_batch_by_range(&["b", "h"], 100, 200),
         ]);
         let mut index_config = MitoConfig::default().index;
         index_config.build_mode = build_mode;
         let write_request = SstWriteRequest {
             op_type: OperationType::Flush,
             metadata: metadata.clone(),
-            source: either::Left(source),
+            source,
             storage: None,
             max_sequence: None,
+            sst_write_format: Default::default(),
             cache_manager: Default::default(),
             index_options: IndexOptions::default(),
             index_config,
diff --git a/src/mito2/src/sst/index/inverted_index/applier.rs b/src/mito2/src/sst/index/inverted_index/applier.rs
index f4c80051dd..a75962ec66 100644
--- a/src/mito2/src/sst/index/inverted_index/applier.rs
+++ b/src/mito2/src/sst/index/inverted_index/applier.rs
@@ -20,15 +20,18 @@ use std::time::Instant;
 
 use common_base::range_read::RangeReader;
 use common_telemetry::warn;
+use datatypes::data_type::ConcreteDataType;
 use index::inverted_index::format::reader::{InvertedIndexBlobReader, InvertedIndexReadMetrics};
 use index::inverted_index::search::index_apply::{
-    ApplyOutput, IndexApplier, IndexNotFoundStrategy, SearchContext,
+    ApplyOutput, IndexApplier, IndexNotFoundStrategy, PredicatesIndexApplier, SearchContext,
 };
 use index::inverted_index::search::predicate::Predicate;
+use index::target::IndexTarget;
 use object_store::ObjectStore;
 use puffin::puffin_manager::cache::PuffinMetadataCacheRef;
 use puffin::puffin_manager::{PuffinManager, PuffinReader};
 use snafu::ResultExt;
+use store_api::metadata::RegionMetadataRef;
 use store_api::region_request::PathType;
 use store_api::storage::ColumnId;
 
@@ -37,7 +40,8 @@ use crate::cache::file_cache::{FileCacheRef, FileType, IndexKey};
 use crate::cache::index::inverted_index::{CachedInvertedIndexBlobReader, InvertedIndexCacheRef};
 use crate::cache::index::result_cache::PredicateKey;
 use crate::error::{
-    ApplyInvertedIndexSnafu, MetadataSnafu, PuffinBuildReaderSnafu, PuffinReadBlobSnafu, Result,
+    ApplyInvertedIndexSnafu, BuildIndexApplierSnafu, MetadataSnafu, PuffinBuildReaderSnafu,
+    PuffinReadBlobSnafu, Result,
 };
 use crate::metrics::{INDEX_APPLY_ELAPSED, INDEX_APPLY_MEMORY_USAGE};
 use crate::sst::file::RegionIndexId;
@@ -121,10 +125,6 @@ pub(crate) struct InvertedIndexApplier {
     /// The cache of index files.
     file_cache: Option<FileCacheRef>,
 
-    /// Predefined index applier used to apply predicates to index files
-    /// and return the relevant row group ids for further scan.
-    index_applier: Box<dyn IndexApplier>,
-
     /// The puffin manager factory.
     puffin_manager_factory: PuffinManagerFactory,
 
@@ -134,35 +134,49 @@ pub(crate) struct InvertedIndexApplier {
     /// Puffin metadata cache.
     puffin_metadata_cache: Option<PuffinMetadataCacheRef>,
 
-    /// Predicate key. Used to identify the predicate and fetch result from cache.
-    predicate_key: PredicateKey,
+    /// All collected predicates.
+    predicates: BTreeMap<ColumnId, Vec<Predicate>>,
+
+    /// Default apply plan built from all collected predicates.
+    default_plan: SstApplyPlan,
+
+    /// Expected predicate column types from the latest region metadata.
+    expected_predicate_col_types: BTreeMap<ColumnId, ConcreteDataType>,
 }
 
 pub(crate) type InvertedIndexApplierRef = Arc<InvertedIndexApplier>;
 
+#[derive(Clone)]
+pub(crate) struct SstApplyPlan {
+    pub predicate_key: PredicateKey,
+    pub index_applier: Arc<PredicatesIndexApplier>,
+}
+
 impl InvertedIndexApplier {
     /// Creates a new `InvertedIndexApplier`.
     pub fn new(
         table_dir: String,
         path_type: PathType,
         store: ObjectStore,
-        index_applier: Box<dyn IndexApplier>,
         puffin_manager_factory: PuffinManagerFactory,
         predicates: BTreeMap<ColumnId, Vec<Predicate>>,
-    ) -> Self {
-        INDEX_APPLY_MEMORY_USAGE.add(index_applier.memory_usage() as i64);
+        expected_predicate_col_types: BTreeMap<ColumnId, ConcreteDataType>,
+    ) -> Result<Self> {
+        let default_plan = Self::build_apply_plan(&predicates)?;
+        INDEX_APPLY_MEMORY_USAGE.add(default_plan.index_applier.memory_usage() as i64);
 
-        Self {
+        Ok(Self {
             table_dir,
             path_type,
             store,
             file_cache: None,
-            index_applier,
             puffin_manager_factory,
             inverted_index_cache: None,
             puffin_metadata_cache: None,
-            predicate_key: PredicateKey::new_inverted(Arc::new(predicates)),
-        }
+            predicates,
+            default_plan,
+            expected_predicate_col_types,
+        })
     }
 
     /// Sets the file cache.
@@ -186,11 +200,12 @@ impl InvertedIndexApplier {
         self
     }
 
-    /// Applies predicates to the provided SST file id and returns the relevant row group ids.
+    /// Applies predicates to one SST file with the provided index applier.
     ///
     /// # Arguments
     /// * `file_id` - The region file ID to apply predicates to
     /// * `file_size_hint` - Optional hint for file size to avoid extra metadata reads
+    /// * `index_applier` - Inverted index applier produced by `plan_for_sst`.
     /// * `metrics` - Optional mutable reference to collect metrics on demand
     #[tracing::instrument(
         skip_all,
@@ -200,6 +215,7 @@ impl InvertedIndexApplier {
         &self,
         file_id: RegionIndexId,
         file_size_hint: Option<u64>,
+        index_applier: &PredicatesIndexApplier,
         mut metrics: Option<&mut InvertedIndexApplyMetrics>,
     ) -> Result<ApplyOutput> {
         let start = Instant::now();
@@ -231,7 +247,7 @@ impl InvertedIndexApplier {
                 InvertedIndexBlobReader::new(blob),
                 index_cache.clone(),
             );
-            self.index_applier
+            index_applier
                 .apply(
                     context,
                     &mut index_reader,
@@ -243,7 +259,7 @@ impl InvertedIndexApplier {
                 .context(ApplyInvertedIndexSnafu)
         } else {
             let mut index_reader = InvertedIndexBlobReader::new(blob);
-            self.index_applier
+            index_applier
                 .apply(
                     context,
                     &mut index_reader,
@@ -344,82 +360,141 @@ impl InvertedIndexApplier {
             .context(PuffinBuildReaderSnafu)
     }
 
-    /// Returns the predicate key.
-    pub fn predicate_key(&self) -> &PredicateKey {
-        &self.predicate_key
+    /// Builds a per-SST apply plan.
+    ///
+    /// Returns `None` when no compatible predicate remains for this SST.
+    pub fn plan_for_sst(&self, sst_metadata: &RegionMetadataRef) -> Result<Option<SstApplyPlan>> {
+        let mut compatible_predicates = BTreeMap::new();
+        let mut has_type_mismatch = false;
+
+        for (col_id, expected) in &self.expected_predicate_col_types {
+            if let Some(sst_col) = sst_metadata.column_by_id(*col_id)
+                && sst_col.column_schema.data_type != *expected
+            {
+                has_type_mismatch = true;
+                continue;
+            }
+
+            if let Some(predicates) = self.predicates.get(col_id) {
+                compatible_predicates.insert(*col_id, predicates.clone());
+            }
+        }
+
+        if compatible_predicates.is_empty() {
+            return Ok(None);
+        }
+
+        if !has_type_mismatch {
+            return Ok(Some(self.default_plan.clone()));
+        }
+
+        let plan = Self::build_apply_plan(&compatible_predicates)?;
+        Ok(Some(plan))
+    }
+
+    fn build_apply_plan(
+        predicates_by_col: &BTreeMap<ColumnId, Vec<Predicate>>,
+    ) -> Result<SstApplyPlan> {
+        let predicates = predicates_by_col
+            .iter()
+            .map(|(col_id, preds)| (format!("{}", IndexTarget::ColumnId(*col_id)), preds.clone()))
+            .collect();
+
+        let index_applier =
+            PredicatesIndexApplier::try_from(predicates).context(BuildIndexApplierSnafu)?;
+
+        let predicate_key = PredicateKey::new_inverted(Arc::new(predicates_by_col.clone()));
+        Ok(SstApplyPlan {
+            predicate_key,
+            index_applier: Arc::new(index_applier),
+        })
     }
 }
 
 impl Drop for InvertedIndexApplier {
     fn drop(&mut self) {
-        INDEX_APPLY_MEMORY_USAGE.sub(self.index_applier.memory_usage() as i64);
+        INDEX_APPLY_MEMORY_USAGE.sub(self.default_plan.index_applier.memory_usage() as i64);
     }
 }
 
 #[cfg(test)]
 mod tests {
+    use api::v1::SemanticType;
+    use datatypes::data_type::ConcreteDataType;
+    use datatypes::schema::ColumnSchema;
     use futures::io::Cursor;
-    use index::bitmap::Bitmap;
-    use index::inverted_index::search::index_apply::MockIndexApplier;
+    use index::inverted_index::search::predicate::RegexMatchPredicate;
     use object_store::services::Memory;
     use puffin::puffin_manager::PuffinWriter;
-    use store_api::storage::FileId;
+    use store_api::metadata::{ColumnMetadata, RegionMetadataBuilder};
+    use store_api::storage::{FileId, RegionId};
 
     use super::*;
     use crate::sst::index::RegionFileId;
 
     #[tokio::test]
-    async fn test_index_applier_apply_basic() {
+    async fn test_plan_for_sst() {
         let (_d, puffin_manager_factory) =
-            PuffinManagerFactory::new_for_test_async("test_index_applier_apply_basic_").await;
+            PuffinManagerFactory::new_for_test_async("test_plan_for_sst_basic_").await;
         let object_store = ObjectStore::new(Memory::default()).unwrap().finish();
-        let file_id = RegionFileId::new(0.into(), FileId::random());
-        let index_id = RegionIndexId::new(file_id, 0);
         let table_dir = "table_dir".to_string();
 
-        let puffin_manager = puffin_manager_factory.build(
-            object_store.clone(),
-            RegionFilePathFactory::new(table_dir.clone(), PathType::Bare),
+        let mut predicates = BTreeMap::new();
+        predicates.insert(
+            1,
+            vec![Predicate::RegexMatch(RegexMatchPredicate {
+                pattern: "foo".to_string(),
+            })],
         );
-        let mut writer = puffin_manager.writer(&index_id).await.unwrap();
-        writer
-            .put_blob(
-                INDEX_BLOB_TYPE,
-                Cursor::new(vec![]),
-                Default::default(),
-                Default::default(),
-            )
-            .await
-            .unwrap();
-        writer.finish().await.unwrap();
-
-        let mut mock_index_applier = MockIndexApplier::new();
-        mock_index_applier.expect_memory_usage().returning(|| 100);
-        mock_index_applier.expect_apply().returning(|_, _, _| {
-            Ok(ApplyOutput {
-                matched_segment_ids: Bitmap::new_bitvec(),
-                total_row_count: 100,
-                segment_row_count: 10,
-            })
-        });
+        let expected_predicate_col_types =
+            BTreeMap::from_iter([(1, ConcreteDataType::string_datatype())]);
 
         let sst_index_applier = InvertedIndexApplier::new(
-            table_dir.clone(),
+            table_dir,
             PathType::Bare,
             object_store,
-            Box::new(mock_index_applier),
             puffin_manager_factory,
-            Default::default(),
-        );
-        let output = sst_index_applier.apply(index_id, None, None).await.unwrap();
-        assert_eq!(
-            output,
-            ApplyOutput {
-                matched_segment_ids: Bitmap::new_bitvec(),
-                total_row_count: 100,
-                segment_row_count: 10,
-            }
+            predicates,
+            expected_predicate_col_types,
+        )
+        .unwrap();
+        let plan = sst_index_applier
+            .plan_for_sst(&mock_region_metadata())
+            .unwrap();
+        assert!(plan.is_some());
+    }
+
+    #[tokio::test]
+    async fn test_plan_for_sst_type_mismatch() {
+        let (_d, puffin_manager_factory) =
+            PuffinManagerFactory::new_for_test_async("test_plan_for_sst_type_mismatch_").await;
+        let object_store = ObjectStore::new(Memory::default()).unwrap().finish();
+        let table_dir = "table_dir".to_string();
+
+        let mut predicates = BTreeMap::new();
+        predicates.insert(
+            1,
+            vec![Predicate::RegexMatch(RegexMatchPredicate {
+                pattern: "foo".to_string(),
+            })],
         );
+        // Column id 1 is String in `mock_region_metadata`, set expected type to Int64.
+        let expected_predicate_col_types =
+            BTreeMap::from_iter([(1, ConcreteDataType::int64_datatype())]);
+
+        let sst_index_applier = InvertedIndexApplier::new(
+            table_dir,
+            PathType::Bare,
+            object_store,
+            puffin_manager_factory,
+            predicates,
+            expected_predicate_col_types,
+        )
+        .unwrap();
+        let plan = sst_index_applier
+            .plan_for_sst(&mock_region_metadata())
+            .unwrap();
+        assert!(plan.is_none());
     }
 
     #[tokio::test]
@@ -448,19 +523,52 @@ mod tests {
             .unwrap();
         writer.finish().await.unwrap();
 
-        let mut mock_index_applier = MockIndexApplier::new();
-        mock_index_applier.expect_memory_usage().returning(|| 100);
-        mock_index_applier.expect_apply().never();
-
+        let mut predicates = BTreeMap::new();
+        predicates.insert(
+            1,
+            vec![Predicate::RegexMatch(RegexMatchPredicate {
+                pattern: "foo".to_string(),
+            })],
+        );
+        let expected_predicate_col_types =
+            BTreeMap::from_iter([(1, ConcreteDataType::string_datatype())]);
         let sst_index_applier = InvertedIndexApplier::new(
             table_dir.clone(),
             PathType::Bare,
             object_store,
-            Box::new(mock_index_applier),
             puffin_manager_factory,
-            Default::default(),
-        );
-        let res = sst_index_applier.apply(index_id, None, None).await;
+            predicates,
+            expected_predicate_col_types,
+        )
+        .unwrap();
+        let plan = sst_index_applier
+            .plan_for_sst(&mock_region_metadata())
+            .unwrap()
+            .unwrap();
+        let res = sst_index_applier
+            .apply(index_id, None, &plan.index_applier, None)
+            .await;
         assert!(format!("{:?}", res.unwrap_err()).contains("Blob not found"));
     }
+
+    fn mock_region_metadata() -> RegionMetadataRef {
+        let mut builder = RegionMetadataBuilder::new(RegionId::new(1024, 1));
+        builder
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new("tag", ConcreteDataType::string_datatype(), false),
+                semantic_type: SemanticType::Tag,
+                column_id: 1,
+            })
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "ts",
+                    ConcreteDataType::timestamp_millisecond_datatype(),
+                    false,
+                ),
+                semantic_type: SemanticType::Timestamp,
+                column_id: 2,
+            })
+            .primary_key(vec![1]);
+        Arc::new(builder.build().unwrap())
+    }
 }
diff --git a/src/mito2/src/sst/index/inverted_index/applier/builder.rs b/src/mito2/src/sst/index/inverted_index/applier/builder.rs
index 078d1837fb..4938352792 100644
--- a/src/mito2/src/sst/index/inverted_index/applier/builder.rs
+++ b/src/mito2/src/sst/index/inverted_index/applier/builder.rs
@@ -25,9 +25,7 @@ use datafusion_common::ScalarValue;
 use datafusion_expr::{BinaryExpr, Expr, Operator};
 use datatypes::data_type::ConcreteDataType;
 use datatypes::value::Value;
-use index::inverted_index::search::index_apply::PredicatesIndexApplier;
 use index::inverted_index::search::predicate::Predicate;
-use index::target::IndexTarget;
 use mito_codec::index::IndexValueCodec;
 use mito_codec::row_converter::SortField;
 use object_store::ObjectStore;
@@ -39,9 +37,7 @@ use store_api::storage::ColumnId;
 
 use crate::cache::file_cache::FileCacheRef;
 use crate::cache::index::inverted_index::InvertedIndexCacheRef;
-use crate::error::{
-    BuildIndexApplierSnafu, ColumnNotFoundSnafu, ConvertValueSnafu, EncodeSnafu, Result,
-};
+use crate::error::{ColumnNotFoundSnafu, ConvertValueSnafu, EncodeSnafu, Result};
 use crate::sst::index::inverted_index::applier::InvertedIndexApplier;
 use crate::sst::index::puffin_manager::PuffinManagerFactory;
 
@@ -137,33 +133,38 @@ impl<'a> InvertedIndexApplierBuilder<'a> {
             return Ok(None);
         }
 
-        let predicates = self
-            .output
-            .iter()
-            .map(|(column_id, predicates)| {
-                (
-                    format!("{}", IndexTarget::ColumnId(*column_id)),
-                    predicates.clone(),
-                )
-            })
-            .collect::<Vec<_>>();
-        let applier = PredicatesIndexApplier::try_from(predicates);
+        let expected_predicate_column_types = self.expected_predicate_column_types();
 
         Ok(Some(
             InvertedIndexApplier::new(
                 self.table_dir,
                 self.path_type,
                 self.object_store,
-                Box::new(applier.context(BuildIndexApplierSnafu)?),
                 self.puffin_manager_factory,
                 self.output,
-            )
+                expected_predicate_column_types,
+            )?
             .with_file_cache(self.file_cache)
             .with_puffin_metadata_cache(self.puffin_metadata_cache)
             .with_index_cache(self.inverted_index_cache),
         ))
     }
 
+    /// Returns `(column_id, data_type)` pairs for predicate columns
+    /// collected in `self.output`.
+    ///
+    /// The data types are resolved from the latest region manifest. Columns
+    /// that no longer exist in the latest metadata are skipped.
+    fn expected_predicate_column_types(&self) -> BTreeMap<ColumnId, ConcreteDataType> {
+        self.output
+            .keys()
+            .filter_map(|col_id| {
+                let col = self.metadata.column_by_id(*col_id)?;
+                Some((*col_id, col.column_schema.data_type.clone()))
+            })
+            .collect()
+    }
+
     /// Recursively traverses expressions to collect predicates.
     /// Results are stored in `self.output`.
     fn traverse_and_collect(&mut self, expr: &Expr) {
diff --git a/src/mito2/src/sst/index/inverted_index/creator.rs b/src/mito2/src/sst/index/inverted_index/creator.rs
index 386ee11b9b..952303f33f 100644
--- a/src/mito2/src/sst/index/inverted_index/creator.rs
+++ b/src/mito2/src/sst/index/inverted_index/creator.rs
@@ -614,9 +614,11 @@ mod tests {
             .build(&[expr])
             .unwrap()
             .unwrap();
+            let sst_metadata = Arc::new(region_metadata.clone());
+            let plan = applier.plan_for_sst(&sst_metadata).unwrap().unwrap();
             Box::pin(async move {
                 applier
-                    .apply(index_id, None, None)
+                    .apply(index_id, None, &plan.index_applier, None)
                     .await
                     .unwrap()
                     .matched_segment_ids
diff --git a/src/mito2/src/sst/index/store.rs b/src/mito2/src/sst/index/store.rs
index 1662c6d876..62a5981298 100644
--- a/src/mito2/src/sst/index/store.rs
+++ b/src/mito2/src/sst/index/store.rs
@@ -142,10 +142,11 @@ impl InstrumentedStore {
         Ok(list)
     }
 
-    /// Proxies to [`ObjectStore::remove_all`].
+    /// Recursively deletes all objects under the given path.
     pub async fn remove_all(&self, path: &str) -> Result<()> {
         self.object_store
-            .remove_all(path)
+            .delete_with(path)
+            .recursive(true)
             .await
             .context(OpenDalSnafu)
     }
diff --git a/src/mito2/src/sst/parquet.rs b/src/mito2/src/sst/parquet.rs
index aa98b69176..cd22d2e551 100644
--- a/src/mito2/src/sst/parquet.rs
+++ b/src/mito2/src/sst/parquet.rs
@@ -24,11 +24,14 @@ use crate::sst::DEFAULT_WRITE_BUFFER_SIZE;
 use crate::sst::file::FileTimeRange;
 use crate::sst::index::IndexOutput;
 
+pub(crate) mod async_reader;
 pub mod file_range;
 pub mod flat_format;
 pub mod format;
 pub(crate) mod helper;
 pub(crate) mod metadata;
+pub mod prefilter;
+pub mod read_columns;
 pub mod reader;
 pub mod row_group;
 pub mod row_selection;
@@ -39,9 +42,17 @@ pub mod writer;
 pub const PARQUET_METADATA_KEY: &str = "greptime:metadata";
 
 /// Default batch size to read parquet files.
-pub(crate) const DEFAULT_READ_BATCH_SIZE: usize = 1024;
+///
+/// This is a runtime-only scan granularity, so we align it with DataFusion's
+/// default execution batch size to reduce rebatching and concatenation in the
+/// query pipeline.
+pub(crate) const DEFAULT_READ_BATCH_SIZE: usize = 8 * 1024;
 /// Default row group size for parquet files.
-pub const DEFAULT_ROW_GROUP_SIZE: usize = 100 * DEFAULT_READ_BATCH_SIZE;
+///
+/// Keep the existing persisted/on-disk default stable. It intentionally stays
+/// decoupled from [`DEFAULT_READ_BATCH_SIZE`] so we can tune runtime scan
+/// batching without changing the row group layout of newly written SSTs.
+pub const DEFAULT_ROW_GROUP_SIZE: usize = 100 * 1024;
 
 /// Parquet write options.
 #[derive(Debug, Clone)]
@@ -110,6 +121,7 @@ mod tests {
         TimestampMillisecondArray, UInt8Array, UInt64Array,
     };
     use datatypes::arrow::datatypes::{DataType, Field, Schema, UInt32Type};
+    use datatypes::arrow::util::pretty::pretty_format_batches;
     use datatypes::prelude::ConcreteDataType;
     use datatypes::schema::{FulltextAnalyzer, FulltextBackend, FulltextOptions};
     use object_store::ObjectStore;
@@ -129,7 +141,7 @@ mod tests {
     use crate::cache::test_util::assert_parquet_metadata_equal;
     use crate::cache::{CacheManager, CacheStrategy, PageKey};
     use crate::config::IndexConfig;
-    use crate::read::{BatchBuilder, BatchReader, FlatSource};
+    use crate::read::FlatSource;
     use crate::region::options::{IndexOptions, InvertedIndexOptions};
     use crate::sst::file::{FileHandle, FileMeta, RegionFileId, RegionIndexId};
     use crate::sst::file_purger::NoopFilePurger;
@@ -137,19 +149,20 @@ mod tests {
     use crate::sst::index::fulltext_index::applier::builder::FulltextIndexApplierBuilder;
     use crate::sst::index::inverted_index::applier::builder::InvertedIndexApplierBuilder;
     use crate::sst::index::{IndexBuildType, Indexer, IndexerBuilder, IndexerBuilderImpl};
-    use crate::sst::parquet::format::PrimaryKeyWriteFormat;
+    use crate::sst::parquet::flat_format::FlatWriteFormat;
     use crate::sst::parquet::reader::{ParquetReader, ParquetReaderBuilder, ReaderMetrics};
+    use crate::sst::parquet::row_selection::RowGroupSelection;
     use crate::sst::parquet::writer::ParquetWriter;
     use crate::sst::{
         DEFAULT_WRITE_CONCURRENCY, FlatSchemaOptions, location, to_flat_sst_arrow_schema,
     };
+    use crate::test_util::TestEnv;
     use crate::test_util::sst_util::{
-        build_test_binary_test_region_metadata, new_batch_by_range, new_batch_with_binary,
-        new_batch_with_custom_sequence, new_primary_key, new_source, new_sparse_primary_key,
-        sst_file_handle, sst_file_handle_with_file_id, sst_region_metadata,
+        build_test_binary_test_region_metadata, new_flat_source_from_record_batches,
+        new_primary_key, new_record_batch_by_range, new_record_batch_with_custom_sequence,
+        new_sparse_primary_key, sst_file_handle, sst_file_handle_with_file_id, sst_region_metadata,
         sst_region_metadata_with_encoding,
     };
-    use crate::test_util::{TestEnv, check_reader_result};
 
     const FILE_DIR: &str = "/";
     const REGION_ID: RegionId = RegionId::new(0, 0);
@@ -191,10 +204,10 @@ mod tests {
             region_file_id: handle.file_id(),
         };
         let metadata = Arc::new(sst_region_metadata());
-        let source = new_source(&[
-            new_batch_by_range(&["a", "d"], 0, 60),
-            new_batch_by_range(&["b", "f"], 0, 40),
-            new_batch_by_range(&["b", "h"], 100, 200),
+        let source = new_flat_source_from_record_batches(vec![
+            new_record_batch_by_range(&["a", "d"], 0, 60),
+            new_record_batch_by_range(&["b", "f"], 0, 40),
+            new_record_batch_by_range(&["b", "h"], 100, 200),
         ]);
         // Use a small row group size for test.
         let write_opts = WriteOptions {
@@ -214,7 +227,7 @@ mod tests {
         .await;
 
         let info = writer
-            .write_all(source, None, &write_opts)
+            .write_all_flat_as_primary_key(source, None, &write_opts)
             .await
             .unwrap()
             .remove(0);
@@ -235,14 +248,14 @@ mod tests {
             object_store,
         );
         let mut reader = builder.build().await.unwrap().unwrap();
-        check_reader_result(
+        check_record_batch_reader_result(
             &mut reader,
             &[
-                new_batch_by_range(&["a", "d"], 0, 50),
-                new_batch_by_range(&["a", "d"], 50, 60),
-                new_batch_by_range(&["b", "f"], 0, 40),
-                new_batch_by_range(&["b", "h"], 100, 150),
-                new_batch_by_range(&["b", "h"], 150, 200),
+                new_record_batch_by_range(&["a", "d"], 0, 50),
+                new_record_batch_by_range(&["a", "d"], 50, 60),
+                new_record_batch_by_range(&["b", "f"], 0, 40),
+                new_record_batch_by_range(&["b", "h"], 100, 150),
+                new_record_batch_by_range(&["b", "h"], 150, 200),
             ],
         )
         .await;
@@ -254,10 +267,10 @@ mod tests {
         let object_store = env.init_object_store_manager();
         let handle = sst_file_handle(0, 1000);
         let metadata = Arc::new(sst_region_metadata());
-        let source = new_source(&[
-            new_batch_by_range(&["a", "d"], 0, 60),
-            new_batch_by_range(&["b", "f"], 0, 40),
-            new_batch_by_range(&["b", "h"], 100, 200),
+        let source = new_flat_source_from_record_batches(vec![
+            new_record_batch_by_range(&["a", "d"], 0, 60),
+            new_record_batch_by_range(&["b", "f"], 0, 40),
+            new_record_batch_by_range(&["b", "h"], 100, 200),
         ]);
         // Use a small row group size for test.
         let write_opts = WriteOptions {
@@ -279,7 +292,7 @@ mod tests {
         .await;
 
         let sst_info = writer
-            .write_all(source, None, &write_opts)
+            .write_all_flat_as_primary_key(source, None, &write_opts)
             .await
             .unwrap()
             .remove(0);
@@ -299,14 +312,14 @@ mod tests {
         .cache(cache.clone());
         for _ in 0..3 {
             let mut reader = builder.build().await.unwrap().unwrap();
-            check_reader_result(
+            check_record_batch_reader_result(
                 &mut reader,
                 &[
-                    new_batch_by_range(&["a", "d"], 0, 50),
-                    new_batch_by_range(&["a", "d"], 50, 60),
-                    new_batch_by_range(&["b", "f"], 0, 40),
-                    new_batch_by_range(&["b", "h"], 100, 150),
-                    new_batch_by_range(&["b", "h"], 150, 200),
+                    new_record_batch_by_range(&["a", "d"], 0, 50),
+                    new_record_batch_by_range(&["a", "d"], 50, 60),
+                    new_record_batch_by_range(&["b", "f"], 0, 40),
+                    new_record_batch_by_range(&["b", "h"], 100, 150),
+                    new_record_batch_by_range(&["b", "h"], 150, 200),
                 ],
             )
             .await;
@@ -340,10 +353,10 @@ mod tests {
         let object_store = env.init_object_store_manager();
         let handle = sst_file_handle(0, 1000);
         let metadata = Arc::new(sst_region_metadata());
-        let source = new_source(&[
-            new_batch_by_range(&["a", "d"], 0, 60),
-            new_batch_by_range(&["b", "f"], 0, 40),
-            new_batch_by_range(&["b", "h"], 100, 200),
+        let source = new_flat_source_from_record_batches(vec![
+            new_record_batch_by_range(&["a", "d"], 0, 60),
+            new_record_batch_by_range(&["b", "f"], 0, 40),
+            new_record_batch_by_range(&["b", "h"], 100, 200),
         ]);
         let write_opts = WriteOptions {
             row_group_size: 50,
@@ -366,7 +379,7 @@ mod tests {
         .await;
 
         let sst_info = writer
-            .write_all(source, None, &write_opts)
+            .write_all_flat_as_primary_key(source, None, &write_opts)
             .await
             .unwrap()
             .remove(0);
@@ -382,8 +395,12 @@ mod tests {
         .page_index_policy(PageIndexPolicy::Optional);
         let reader = builder.build().await.unwrap().unwrap();
         let reader_metadata = reader.parquet_metadata();
+        let cached_writer_metadata =
+            crate::cache::CachedSstMeta::try_new("test.sst", Arc::unwrap_or_clone(writer_metadata))
+                .unwrap()
+                .parquet_metadata();
 
-        assert_parquet_metadata_equal(writer_metadata, reader_metadata);
+        assert_parquet_metadata_equal(cached_writer_metadata, reader_metadata);
     }
 
     #[tokio::test]
@@ -392,10 +409,10 @@ mod tests {
         let object_store = env.init_object_store_manager();
         let handle = sst_file_handle(0, 1000);
         let metadata = Arc::new(sst_region_metadata());
-        let source = new_source(&[
-            new_batch_by_range(&["a", "d"], 0, 60),
-            new_batch_by_range(&["b", "f"], 0, 40),
-            new_batch_by_range(&["b", "h"], 100, 200),
+        let source = new_flat_source_from_record_batches(vec![
+            new_record_batch_by_range(&["a", "d"], 0, 60),
+            new_record_batch_by_range(&["b", "f"], 0, 40),
+            new_record_batch_by_range(&["b", "h"], 100, 200),
         ]);
         // Use a small row group size for test.
         let write_opts = WriteOptions {
@@ -416,7 +433,7 @@ mod tests {
         )
         .await;
         writer
-            .write_all(source, None, &write_opts)
+            .write_all_flat_as_primary_key(source, None, &write_opts)
             .await
             .unwrap()
             .remove(0);
@@ -436,11 +453,11 @@ mod tests {
         )
         .predicate(predicate);
         let mut reader = builder.build().await.unwrap().unwrap();
-        check_reader_result(
+        check_record_batch_reader_result(
             &mut reader,
             &[
-                new_batch_by_range(&["a", "d"], 0, 50),
-                new_batch_by_range(&["a", "d"], 50, 60),
+                new_record_batch_by_range(&["a", "d"], 0, 50),
+                new_record_batch_by_range(&["a", "d"], 50, 60),
             ],
         )
         .await;
@@ -452,10 +469,10 @@ mod tests {
         let object_store = env.init_object_store_manager();
         let handle = sst_file_handle(0, 1000);
         let metadata = Arc::new(sst_region_metadata());
-        let source = new_source(&[
-            new_batch_by_range(&["a", "z"], 0, 0),
-            new_batch_by_range(&["a", "z"], 100, 100),
-            new_batch_by_range(&["a", "z"], 200, 230),
+        let source = new_flat_source_from_record_batches(vec![
+            new_record_batch_by_range(&["a", "z"], 0, 0),
+            new_record_batch_by_range(&["a", "z"], 100, 100),
+            new_record_batch_by_range(&["a", "z"], 200, 230),
         ]);
         // Use a small row group size for test.
         let write_opts = WriteOptions {
@@ -476,7 +493,7 @@ mod tests {
         )
         .await;
         writer
-            .write_all(source, None, &write_opts)
+            .write_all_flat_as_primary_key(source, None, &write_opts)
             .await
             .unwrap()
             .remove(0);
@@ -488,7 +505,11 @@ mod tests {
             object_store,
         );
         let mut reader = builder.build().await.unwrap().unwrap();
-        check_reader_result(&mut reader, &[new_batch_by_range(&["a", "z"], 200, 230)]).await;
+        check_record_batch_reader_result(
+            &mut reader,
+            &[new_record_batch_by_range(&["a", "z"], 200, 230)],
+        )
+        .await;
     }
 
     #[tokio::test]
@@ -497,10 +518,10 @@ mod tests {
         let object_store = env.init_object_store_manager();
         let handle = sst_file_handle(0, 1000);
         let metadata = Arc::new(sst_region_metadata());
-        let source = new_source(&[
-            new_batch_by_range(&["a", "d"], 0, 60),
-            new_batch_by_range(&["b", "f"], 0, 40),
-            new_batch_by_range(&["b", "h"], 100, 200),
+        let source = new_flat_source_from_record_batches(vec![
+            new_record_batch_by_range(&["a", "d"], 0, 60),
+            new_record_batch_by_range(&["b", "f"], 0, 40),
+            new_record_batch_by_range(&["b", "h"], 100, 200),
         ]);
         // Use a small row group size for test.
         let write_opts = WriteOptions {
@@ -522,7 +543,7 @@ mod tests {
         .await;
 
         writer
-            .write_all(source, None, &write_opts)
+            .write_all_flat_as_primary_key(source, None, &write_opts)
             .await
             .unwrap()
             .remove(0);
@@ -542,7 +563,11 @@ mod tests {
         )
         .predicate(predicate);
         let mut reader = builder.build().await.unwrap().unwrap();
-        check_reader_result(&mut reader, &[new_batch_by_range(&["b", "h"], 150, 200)]).await;
+        check_record_batch_reader_result(
+            &mut reader,
+            &[new_record_batch_by_range(&["b", "h"], 150, 200)],
+        )
+        .await;
     }
 
     #[tokio::test]
@@ -569,7 +594,7 @@ mod tests {
 
         let writer_props = props_builder.build();
 
-        let write_format = PrimaryKeyWriteFormat::new(metadata);
+        let write_format = FlatWriteFormat::new(metadata, &FlatSchemaOptions::default());
         let fields: Vec<_> = write_format
             .arrow_schema()
             .fields()
@@ -603,9 +628,8 @@ mod tests {
         )
         .unwrap();
 
-        let batch = new_batch_with_binary(&["a"], 0, 60);
-        let arrow_batch = write_format.convert_batch(&batch).unwrap();
-        let arrays: Vec<_> = arrow_batch
+        let batch = new_record_batch_with_binary(&["a"], 0, 60);
+        let arrays: Vec<_> = batch
             .columns()
             .iter()
             .map(|array| {
@@ -629,11 +653,11 @@ mod tests {
             object_store,
         );
         let mut reader = builder.build().await.unwrap().unwrap();
-        check_reader_result(
+        check_record_batch_reader_result(
             &mut reader,
             &[
-                new_batch_with_binary(&["a"], 0, 50),
-                new_batch_with_binary(&["a"], 50, 60),
+                new_record_batch_with_binary(&["a"], 0, 50),
+                new_record_batch_with_binary(&["a"], 50, 60),
             ],
         )
         .await;
@@ -646,17 +670,17 @@ mod tests {
         let mut env = TestEnv::new().await;
         let object_store = env.init_object_store_manager();
         let metadata = Arc::new(sst_region_metadata());
-        let batches = &[
-            new_batch_by_range(&["a", "d"], 0, 1000),
-            new_batch_by_range(&["b", "f"], 0, 1000),
-            new_batch_by_range(&["c", "g"], 0, 1000),
-            new_batch_by_range(&["b", "h"], 100, 200),
-            new_batch_by_range(&["b", "h"], 200, 300),
-            new_batch_by_range(&["b", "h"], 300, 1000),
+        let batches = vec![
+            new_record_batch_by_range(&["a", "d"], 0, 1000),
+            new_record_batch_by_range(&["b", "f"], 0, 1000),
+            new_record_batch_by_range(&["c", "g"], 0, 1000),
+            new_record_batch_by_range(&["b", "h"], 100, 200),
+            new_record_batch_by_range(&["b", "h"], 200, 300),
+            new_record_batch_by_range(&["b", "h"], 300, 1000),
         ];
         let total_rows: usize = batches.iter().map(|batch| batch.num_rows()).sum();
 
-        let source = new_source(batches);
+        let source = new_flat_source_from_record_batches(batches);
         let write_opts = WriteOptions {
             row_group_size: 50,
             max_file_size: Some(1024 * 16),
@@ -678,7 +702,10 @@ mod tests {
         )
         .await;
 
-        let files = writer.write_all(source, None, &write_opts).await.unwrap();
+        let files = writer
+            .write_all_flat_as_primary_key(source, None, &write_opts)
+            .await
+            .unwrap();
         assert_eq!(2, files.len());
 
         let mut rows_read = 0;
@@ -695,7 +722,7 @@ mod tests {
                 object_store.clone(),
             );
             let mut reader = builder.build().await.unwrap().unwrap();
-            while let Some(batch) = reader.next_batch().await.unwrap() {
+            while let Some(batch) = reader.next_record_batch().await.unwrap() {
                 rows_read += batch.num_rows();
             }
         }
@@ -710,12 +737,12 @@ mod tests {
         let metadata = Arc::new(sst_region_metadata());
         let row_group_size = 50;
 
-        let source = new_source(&[
-            new_batch_by_range(&["a", "d"], 0, 20),
-            new_batch_by_range(&["b", "d"], 0, 20),
-            new_batch_by_range(&["c", "d"], 0, 20),
-            new_batch_by_range(&["c", "f"], 0, 40),
-            new_batch_by_range(&["c", "h"], 100, 200),
+        let source = new_flat_source_from_record_batches(vec![
+            new_record_batch_by_range(&["a", "d"], 0, 20),
+            new_record_batch_by_range(&["b", "d"], 0, 20),
+            new_record_batch_by_range(&["c", "d"], 0, 20),
+            new_record_batch_by_range(&["c", "f"], 0, 40),
+            new_record_batch_by_range(&["c", "h"], 100, 200),
         ]);
         // Use a small row group size for test.
         let write_opts = WriteOptions {
@@ -760,7 +787,7 @@ mod tests {
         .await;
 
         let info = writer
-            .write_all(source, None, &write_opts)
+            .write_all_flat_as_primary_key(source, None, &write_opts)
             .await
             .unwrap()
             .remove(0);
@@ -805,6 +832,7 @@ mod tests {
                     None => None,
                 },
                 num_series: 0,
+                ..Default::default()
             },
             Arc::new(NoopFilePurger),
         );
@@ -891,17 +919,24 @@ mod tests {
         let mut reader = ParquetReader::new(Arc::new(context), selection)
             .await
             .unwrap();
-        check_reader_result(&mut reader, &[new_batch_by_range(&["b", "d"], 0, 20)]).await;
+        check_record_batch_reader_result(
+            &mut reader,
+            &[new_record_batch_by_range(&["b", "d"], 0, 20)],
+        )
+        .await;
 
         assert_eq!(metrics.filter_metrics.rg_total, 4);
         assert_eq!(metrics.filter_metrics.rg_minmax_filtered, 3);
         assert_eq!(metrics.filter_metrics.rg_inverted_filtered, 0);
         assert_eq!(metrics.filter_metrics.rows_inverted_filtered, 30);
+        let plan = inverted_index_applier
+            .as_ref()
+            .unwrap()
+            .plan_for_sst(&metadata)
+            .unwrap()
+            .unwrap();
         let cached = index_result_cache
-            .get(
-                inverted_index_applier.unwrap().predicate_key(),
-                handle.file_id().file_id(),
-            )
+            .get(&plan.predicate_key, handle.file_id().file_id())
             .unwrap();
         // inverted index will search all row groups
         assert!(cached.contains_row_group(0));
@@ -1005,13 +1040,13 @@ mod tests {
         let mut reader = ParquetReader::new(Arc::new(context), selection)
             .await
             .unwrap();
-        check_reader_result(
+        check_record_batch_reader_result(
             &mut reader,
             &[
-                new_batch_by_range(&["a", "d"], 0, 20),
-                new_batch_by_range(&["b", "d"], 0, 20),
-                new_batch_by_range(&["c", "d"], 0, 10),
-                new_batch_by_range(&["c", "d"], 10, 20),
+                new_record_batch_by_range(&["a", "d"], 0, 20),
+                new_record_batch_by_range(&["b", "d"], 0, 20),
+                new_record_batch_by_range(&["c", "d"], 0, 10),
+                new_record_batch_by_range(&["c", "d"], 10, 20),
             ],
         )
         .await;
@@ -1032,37 +1067,32 @@ mod tests {
         assert!(cached.contains_row_group(3));
     }
 
-    /// Creates a flat format RecordBatch for testing.
-    /// Similar to `new_batch_by_range` but returns a RecordBatch in flat format.
-    fn new_record_batch_by_range(tags: &[&str], start: usize, end: usize) -> RecordBatch {
+    fn new_record_batch_with_binary(tags: &[&str], start: usize, end: usize) -> RecordBatch {
         assert!(end >= start);
-        let metadata = Arc::new(sst_region_metadata());
+        let metadata = build_test_binary_test_region_metadata();
         let flat_schema = to_flat_sst_arrow_schema(&metadata, &FlatSchemaOptions::default());
 
         let num_rows = end - start;
         let mut columns = Vec::new();
 
-        // Add primary key columns (tag_0, tag_1) as dictionary arrays
         let mut tag_0_builder = StringDictionaryBuilder::<UInt32Type>::new();
-        let mut tag_1_builder = StringDictionaryBuilder::<UInt32Type>::new();
-
         for _ in 0..num_rows {
             tag_0_builder.append_value(tags[0]);
-            tag_1_builder.append_value(tags[1]);
         }
-
         columns.push(Arc::new(tag_0_builder.finish()) as ArrayRef);
-        columns.push(Arc::new(tag_1_builder.finish()) as ArrayRef);
 
-        // Add field column (field_0)
-        let field_values: Vec<u64> = (start..end).map(|v| v as u64).collect();
-        columns.push(Arc::new(UInt64Array::from(field_values)));
+        let values = (0..num_rows)
+            .map(|_| "some data".as_bytes())
+            .collect::<Vec<_>>();
+        columns.push(
+            Arc::new(datatypes::arrow::array::BinaryArray::from_iter_values(
+                values,
+            )) as ArrayRef,
+        );
 
-        // Add time index column (ts)
         let timestamps: Vec<i64> = (start..end).map(|v| v as i64).collect();
         columns.push(Arc::new(TimestampMillisecondArray::from(timestamps)));
 
-        // Add encoded primary key column
         let pk = new_primary_key(tags);
         let mut pk_builder = BinaryDictionaryBuilder::<UInt32Type>::new();
         for _ in 0..num_rows {
@@ -1070,10 +1100,7 @@ mod tests {
         }
         columns.push(Arc::new(pk_builder.finish()));
 
-        // Add sequence column
         columns.push(Arc::new(UInt64Array::from_value(1000, num_rows)));
-
-        // Add op_type column
         columns.push(Arc::new(UInt8Array::from_value(
             OpType::Put as u8,
             num_rows,
@@ -1082,9 +1109,52 @@ mod tests {
         RecordBatch::try_new(flat_schema, columns).unwrap()
     }
 
-    /// Creates a FlatSource from flat format RecordBatches.
-    fn new_flat_source_from_record_batches(batches: Vec<RecordBatch>) -> FlatSource {
-        FlatSource::Iter(Box::new(batches.into_iter().map(Ok)))
+    async fn check_record_batch_reader_result(
+        reader: &mut ParquetReader,
+        expected: &[RecordBatch],
+    ) {
+        let mut actual = Vec::new();
+        while let Some(batch) = reader.next_record_batch().await.unwrap() {
+            actual.push(batch);
+        }
+        assert_eq!(
+            pretty_format_batches(expected).unwrap().to_string(),
+            pretty_format_batches(&actual).unwrap().to_string()
+        );
+        assert!(reader.next_record_batch().await.unwrap().is_none());
+    }
+
+    fn new_record_batch_from_rows(rows: &[(&str, &str, i64)]) -> RecordBatch {
+        let metadata = Arc::new(sst_region_metadata());
+        let flat_schema = to_flat_sst_arrow_schema(&metadata, &FlatSchemaOptions::default());
+
+        let mut tag_0_builder = StringDictionaryBuilder::<UInt32Type>::new();
+        let mut tag_1_builder = StringDictionaryBuilder::<UInt32Type>::new();
+        let mut pk_builder = BinaryDictionaryBuilder::<UInt32Type>::new();
+        let mut field_values = Vec::with_capacity(rows.len());
+        let mut timestamps = Vec::with_capacity(rows.len());
+
+        for (tag_0, tag_1, ts) in rows {
+            tag_0_builder.append_value(*tag_0);
+            tag_1_builder.append_value(*tag_1);
+            pk_builder.append(new_primary_key(&[tag_0, tag_1])).unwrap();
+            field_values.push(*ts as u64);
+            timestamps.push(*ts);
+        }
+
+        RecordBatch::try_new(
+            flat_schema,
+            vec![
+                Arc::new(tag_0_builder.finish()) as ArrayRef,
+                Arc::new(tag_1_builder.finish()) as ArrayRef,
+                Arc::new(UInt64Array::from(field_values)) as ArrayRef,
+                Arc::new(TimestampMillisecondArray::from(timestamps)) as ArrayRef,
+                Arc::new(pk_builder.finish()) as ArrayRef,
+                Arc::new(UInt64Array::from_value(1000, rows.len())) as ArrayRef,
+                Arc::new(UInt8Array::from_value(OpType::Put as u8, rows.len())) as ArrayRef,
+            ],
+        )
+        .unwrap()
     }
 
     /// Creates a flat format RecordBatch for testing with sparse primary key encoding.
@@ -1219,6 +1289,7 @@ mod tests {
                     None => None,
                 },
                 num_series: 0,
+                ..Default::default()
             },
             Arc::new(NoopFilePurger),
         )
@@ -1333,10 +1404,11 @@ mod tests {
         };
         let metadata = Arc::new(sst_region_metadata());
 
-        // Create batches with sequence 0 to trigger override functionality
-        let batch1 = new_batch_with_custom_sequence(&["a", "d"], 0, 60, 0);
-        let batch2 = new_batch_with_custom_sequence(&["b", "f"], 0, 40, 0);
-        let source = new_source(&[batch1, batch2]);
+        // Create batches with sequence 0 to trigger override functionality.
+        let source = new_flat_source_from_record_batches(vec![
+            new_record_batch_with_custom_sequence(&["a", "d"], 0, 60, 0),
+            new_record_batch_with_custom_sequence(&["b", "f"], 0, 40, 0),
+        ]);
 
         let write_opts = WriteOptions {
             row_group_size: 50,
@@ -1355,7 +1427,7 @@ mod tests {
         .await;
 
         writer
-            .write_all(source, None, &write_opts)
+            .write_all_flat_as_primary_key(source, None, &write_opts)
             .await
             .unwrap()
             .remove(0);
@@ -1369,7 +1441,7 @@ mod tests {
         );
         let mut reader = builder.build().await.unwrap().unwrap();
         let mut normal_batches = Vec::new();
-        while let Some(batch) = reader.next_batch().await.unwrap() {
+        while let Some(batch) = reader.next_record_batch().await.unwrap() {
             normal_batches.push(batch);
         }
 
@@ -1391,22 +1463,19 @@ mod tests {
         );
         let mut reader = builder.build().await.unwrap().unwrap();
         let mut override_batches = Vec::new();
-        while let Some(batch) = reader.next_batch().await.unwrap() {
+        while let Some(batch) = reader.next_record_batch().await.unwrap() {
             override_batches.push(batch);
         }
 
         // Compare the results
         assert_eq!(normal_batches.len(), override_batches.len());
         for (normal, override_batch) in normal_batches.into_iter().zip(override_batches.iter()) {
-            // Create expected batch with override sequence
             let expected_batch = {
-                let num_rows = normal.num_rows();
-                let mut builder = BatchBuilder::from(normal);
-                builder
-                    .sequences_array(Arc::new(UInt64Array::from_value(custom_sequence, num_rows)))
-                    .unwrap();
-
-                builder.build().unwrap()
+                let mut columns = normal.columns().to_vec();
+                let num_cols = columns.len();
+                columns[num_cols - 2] =
+                    Arc::new(UInt64Array::from_value(custom_sequence, normal.num_rows()));
+                RecordBatch::try_new(normal.schema(), columns).unwrap()
             };
 
             // Override batch should match expected batch
@@ -1491,7 +1560,6 @@ mod tests {
             handle.clone(),
             object_store.clone(),
         )
-        .flat_format(true)
         .predicate(Some(Predicate::new(preds)))
         .inverted_index_appliers([inverted_index_applier.clone(), None])
         .cache(CacheStrategy::EnableAll(cache.clone()));
@@ -1594,7 +1662,6 @@ mod tests {
             handle.clone(),
             object_store.clone(),
         )
-        .flat_format(true)
         .predicate(Some(Predicate::new(preds)))
         .bloom_filter_index_appliers([None, bloom_filter_applier.clone()])
         .cache(CacheStrategy::EnableAll(cache.clone()));
@@ -1618,6 +1685,131 @@ mod tests {
         assert_eq!(metrics.filter_metrics.rows_bloom_filtered, 100);
     }
 
+    #[tokio::test]
+    async fn test_reader_prefilter_with_outer_selection_and_trailing_filtered_rows() {
+        let mut env = TestEnv::new().await;
+        let object_store = env.init_object_store_manager();
+        let file_path = RegionFilePathFactory::new(FILE_DIR.to_string(), PathType::Bare);
+        let metadata = Arc::new(sst_region_metadata());
+        let row_group_size = 10;
+
+        let flat_source = new_flat_source_from_record_batches(vec![
+            new_record_batch_by_range(&["a", "d"], 0, 3),
+            new_record_batch_by_range(&["b", "d"], 3, 10),
+        ]);
+        let write_opts = WriteOptions {
+            row_group_size,
+            ..Default::default()
+        };
+        let indexer_builder = create_test_indexer_builder(
+            &env,
+            object_store.clone(),
+            file_path.clone(),
+            metadata.clone(),
+            row_group_size,
+        );
+        let info = write_flat_sst(
+            object_store.clone(),
+            metadata.clone(),
+            indexer_builder,
+            file_path,
+            flat_source,
+            &write_opts,
+        )
+        .await;
+        let handle = create_file_handle_from_sst_info(&info, &metadata);
+
+        let builder =
+            ParquetReaderBuilder::new(FILE_DIR.to_string(), PathType::Bare, handle, object_store)
+                .predicate(Some(Predicate::new(vec![col("tag_0").eq(lit("a"))])));
+
+        let mut metrics = ReaderMetrics::default();
+        let (context, _) = builder
+            .build_reader_input(&mut metrics)
+            .await
+            .unwrap()
+            .unwrap();
+        let selection = RowGroupSelection::from_row_ranges(
+            vec![(0, std::iter::once(0..6).collect())],
+            row_group_size,
+        );
+
+        let mut reader = ParquetReader::new(Arc::new(context), selection)
+            .await
+            .unwrap();
+        check_record_batch_reader_result(
+            &mut reader,
+            &[new_record_batch_by_range(&["a", "d"], 0, 3)],
+        )
+        .await;
+    }
+
+    #[tokio::test]
+    async fn test_reader_prefilter_with_outer_selection_disjoint_matches_and_trailing_gap() {
+        let mut env = TestEnv::new().await;
+        let object_store = env.init_object_store_manager();
+        let file_path = RegionFilePathFactory::new(FILE_DIR.to_string(), PathType::Bare);
+        let metadata = Arc::new(sst_region_metadata());
+        let row_group_size = 8;
+
+        let flat_source = new_flat_source_from_record_batches(vec![
+            new_record_batch_by_range(&["a", "d"], 0, 2),
+            new_record_batch_by_range(&["b", "d"], 2, 4),
+            new_record_batch_by_range(&["a", "d"], 4, 6),
+            new_record_batch_by_range(&["c", "d"], 6, 8),
+        ]);
+        let write_opts = WriteOptions {
+            row_group_size,
+            ..Default::default()
+        };
+        let indexer_builder = create_test_indexer_builder(
+            &env,
+            object_store.clone(),
+            file_path.clone(),
+            metadata.clone(),
+            row_group_size,
+        );
+        let info = write_flat_sst(
+            object_store.clone(),
+            metadata.clone(),
+            indexer_builder,
+            file_path,
+            flat_source,
+            &write_opts,
+        )
+        .await;
+        let handle = create_file_handle_from_sst_info(&info, &metadata);
+
+        let builder =
+            ParquetReaderBuilder::new(FILE_DIR.to_string(), PathType::Bare, handle, object_store)
+                .predicate(Some(Predicate::new(vec![col("tag_0").eq(lit("a"))])));
+
+        let mut metrics = ReaderMetrics::default();
+        let (context, _) = builder
+            .build_reader_input(&mut metrics)
+            .await
+            .unwrap()
+            .unwrap();
+        let selection = RowGroupSelection::from_row_ranges(
+            vec![(0, std::iter::once(0..8).collect())],
+            row_group_size,
+        );
+
+        let mut reader = ParquetReader::new(Arc::new(context), selection)
+            .await
+            .unwrap();
+        check_record_batch_reader_result(
+            &mut reader,
+            &[new_record_batch_from_rows(&[
+                ("a", "d", 0),
+                ("a", "d", 1),
+                ("a", "d", 4),
+                ("a", "d", 5),
+            ])],
+        )
+        .await;
+    }
+
     #[tokio::test]
     async fn test_write_flat_read_with_inverted_index_sparse() {
         common_telemetry::init_default_ut_logging();
@@ -1699,7 +1891,6 @@ mod tests {
             handle.clone(),
             object_store.clone(),
         )
-        .flat_format(true)
         .predicate(Some(Predicate::new(preds)))
         .inverted_index_appliers([inverted_index_applier.clone(), None])
         .cache(CacheStrategy::EnableAll(cache.clone()));
@@ -1806,7 +1997,6 @@ mod tests {
             handle.clone(),
             object_store.clone(),
         )
-        .flat_format(true)
         .predicate(Some(Predicate::new(preds)))
         .bloom_filter_index_appliers([None, bloom_filter_applier.clone()])
         .cache(CacheStrategy::EnableAll(cache.clone()));
@@ -2070,7 +2260,6 @@ mod tests {
             handle.clone(),
             object_store.clone(),
         )
-        .flat_format(true)
         .predicate(Some(Predicate::new(preds)))
         .fulltext_index_appliers([None, fulltext_applier.clone()])
         .cache(CacheStrategy::EnableAll(cache.clone()));
@@ -2119,7 +2308,6 @@ mod tests {
             handle.clone(),
             object_store.clone(),
         )
-        .flat_format(true)
         .predicate(Some(Predicate::new(preds)))
         .fulltext_index_appliers([None, fulltext_applier.clone()])
         .cache(CacheStrategy::EnableAll(cache.clone()));
diff --git a/src/mito2/src/sst/parquet/async_reader.rs b/src/mito2/src/sst/parquet/async_reader.rs
new file mode 100644
index 0000000000..a060fd367d
--- /dev/null
+++ b/src/mito2/src/sst/parquet/async_reader.rs
@@ -0,0 +1,221 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Async file reader implementation for SST parquet files.
+
+use std::ops::Range;
+use std::sync::Arc;
+
+use bytes::Bytes;
+use futures::FutureExt;
+use futures::future::BoxFuture;
+use object_store::ObjectStore;
+use parquet::arrow::async_reader::AsyncFileReader;
+use parquet::errors::{ParquetError, Result as ParquetResult};
+use parquet::file::metadata::ParquetMetaData;
+
+use crate::cache::file_cache::{FileType, IndexKey};
+use crate::cache::{CacheStrategy, PageKey, PageValue};
+use crate::metrics::{READ_STAGE_ELAPSED, READ_STAGE_FETCH_PAGES};
+use crate::sst::file::RegionFileId;
+use crate::sst::parquet::helper::fetch_byte_ranges;
+use crate::sst::parquet::row_group::{ParquetFetchMetrics, compute_total_range_size};
+
+/// An [AsyncFileReader] implementation for SST parquet files.
+///
+/// This reader provides async byte access to parquet data in object storage,
+/// with caching support (page cache and write cache).
+pub struct SstAsyncFileReader {
+    /// Region file ID for cache key.
+    region_file_id: RegionFileId,
+    /// Path to the parquet file in object storage.
+    file_path: String,
+    /// Object store for reading data.
+    object_store: ObjectStore,
+    /// Cache strategy for reading pages.
+    cache_strategy: CacheStrategy,
+    /// Cached parquet metadata.
+    metadata: Arc<ParquetMetaData>,
+    /// Row group index for cache key.
+    row_group_idx: usize,
+    /// Optional metrics for tracking fetch operations.
+    fetch_metrics: Option<ParquetFetchMetrics>,
+}
+
+impl SstAsyncFileReader {
+    /// Creates a new [SstAsyncFileReader].
+    pub fn new(
+        region_file_id: RegionFileId,
+        file_path: String,
+        object_store: ObjectStore,
+        cache_strategy: CacheStrategy,
+        metadata: Arc<ParquetMetaData>,
+        row_group_idx: usize,
+    ) -> Self {
+        Self {
+            region_file_id,
+            file_path,
+            object_store,
+            cache_strategy,
+            metadata,
+            row_group_idx,
+            fetch_metrics: None,
+        }
+    }
+
+    /// Sets the fetch metrics.
+    pub fn with_fetch_metrics(mut self, metrics: Option<ParquetFetchMetrics>) -> Self {
+        self.fetch_metrics = metrics;
+        self
+    }
+
+    /// Fetches byte ranges from page cache, write cache, or object store.
+    async fn fetch_bytes_with_cache(&self, ranges: Vec<Range<u64>>) -> ParquetResult<Vec<Bytes>> {
+        let fetch_start = self
+            .fetch_metrics
+            .as_ref()
+            .map(|_| std::time::Instant::now());
+        let _timer = READ_STAGE_FETCH_PAGES.start_timer();
+
+        let page_key = PageKey::new(
+            self.region_file_id.file_id(),
+            self.row_group_idx,
+            ranges.clone(),
+        );
+
+        // Check page cache first.
+        if let Some(pages) = self.cache_strategy.get_pages(&page_key) {
+            if let Some(metrics) = &self.fetch_metrics {
+                let total_size: u64 = ranges.iter().map(|r| r.end - r.start).sum();
+                let mut metrics_data = metrics.data.lock().unwrap();
+                metrics_data.page_cache_hit += 1;
+                metrics_data.pages_to_fetch_mem += ranges.len();
+                metrics_data.page_size_to_fetch_mem += total_size;
+                metrics_data.page_size_needed += total_size;
+                if let Some(start) = fetch_start {
+                    metrics_data.total_fetch_elapsed += start.elapsed();
+                }
+            }
+            return Ok(pages.compressed.clone());
+        }
+
+        // Calculate total range size for metrics.
+        let (total_range_size, unaligned_size) = compute_total_range_size(&ranges);
+
+        // Check write cache.
+        let key = IndexKey::new(
+            self.region_file_id.region_id(),
+            self.region_file_id.file_id(),
+            FileType::Parquet,
+        );
+        let fetch_write_cache_start = self
+            .fetch_metrics
+            .as_ref()
+            .map(|_| std::time::Instant::now());
+        let write_cache_result = self.fetch_ranges_from_write_cache(key, &ranges).await;
+
+        let pages = match write_cache_result {
+            Some(data) => {
+                if let Some(metrics) = &self.fetch_metrics {
+                    let elapsed = fetch_write_cache_start
+                        .map(|start| start.elapsed())
+                        .unwrap_or_default();
+                    let range_size_needed: u64 = ranges.iter().map(|r| r.end - r.start).sum();
+                    let mut metrics_data = metrics.data.lock().unwrap();
+                    metrics_data.write_cache_fetch_elapsed += elapsed;
+                    metrics_data.write_cache_hit += 1;
+                    metrics_data.pages_to_fetch_write_cache += ranges.len();
+                    metrics_data.page_size_to_fetch_write_cache += unaligned_size;
+                    metrics_data.page_size_needed += range_size_needed;
+                }
+                data
+            }
+            None => {
+                // Fetch data from object store.
+                let _timer = READ_STAGE_ELAPSED
+                    .with_label_values(&["cache_miss_read"])
+                    .start_timer();
+
+                let start = self
+                    .fetch_metrics
+                    .as_ref()
+                    .map(|_| std::time::Instant::now());
+                let data = fetch_byte_ranges(&self.file_path, self.object_store.clone(), &ranges)
+                    .await
+                    .map_err(|e| ParquetError::External(Box::new(e)))?;
+
+                if let Some(metrics) = &self.fetch_metrics {
+                    let elapsed = start.map(|start| start.elapsed()).unwrap_or_default();
+                    let range_size_needed: u64 = ranges.iter().map(|r| r.end - r.start).sum();
+                    let mut metrics_data = metrics.data.lock().unwrap();
+                    metrics_data.store_fetch_elapsed += elapsed;
+                    metrics_data.cache_miss += 1;
+                    metrics_data.pages_to_fetch_store += ranges.len();
+                    metrics_data.page_size_to_fetch_store += unaligned_size;
+                    metrics_data.page_size_needed += range_size_needed;
+                }
+                data
+            }
+        };
+
+        // Put pages back to the cache.
+        let page_value = PageValue::new(pages.clone(), total_range_size);
+        self.cache_strategy
+            .put_pages(page_key, Arc::new(page_value));
+
+        if let (Some(metrics), Some(start)) = (&self.fetch_metrics, fetch_start) {
+            metrics.data.lock().unwrap().total_fetch_elapsed += start.elapsed();
+        }
+
+        Ok(pages)
+    }
+
+    /// Fetches data from write cache.
+    /// Returns `None` if the data is not in the cache.
+    async fn fetch_ranges_from_write_cache(
+        &self,
+        key: IndexKey,
+        ranges: &[Range<u64>],
+    ) -> Option<Vec<Bytes>> {
+        if let Some(cache) = self.cache_strategy.write_cache() {
+            return cache.file_cache().read_ranges(key, ranges).await;
+        }
+        None
+    }
+}
+
+impl AsyncFileReader for SstAsyncFileReader {
+    fn get_bytes(&mut self, range: Range<u64>) -> BoxFuture<'_, ParquetResult<Bytes>> {
+        async move {
+            let mut result = self.fetch_bytes_with_cache(vec![range]).await?;
+            Ok(result.pop().unwrap_or_default())
+        }
+        .boxed()
+    }
+
+    fn get_byte_ranges(
+        &mut self,
+        ranges: Vec<Range<u64>>,
+    ) -> BoxFuture<'_, ParquetResult<Vec<Bytes>>> {
+        async move { self.fetch_bytes_with_cache(ranges).await }.boxed()
+    }
+
+    fn get_metadata(
+        &mut self,
+        _options: Option<&parquet::arrow::arrow_reader::ArrowReaderOptions>,
+    ) -> BoxFuture<'_, ParquetResult<Arc<ParquetMetaData>>> {
+        // Metadata is already cached, return it immediately.
+        std::future::ready(Ok(self.metadata.clone())).boxed()
+    }
+}
diff --git a/src/mito2/src/sst/parquet/file_range.rs b/src/mito2/src/sst/parquet/file_range.rs
index 3a5251cb1a..ec7e038a08 100644
--- a/src/mito2/src/sst/parquet/file_range.rs
+++ b/src/mito2/src/sst/parquet/file_range.rs
@@ -26,9 +26,8 @@ use datafusion::physical_plan::expressions::DynamicFilterPhysicalExpr;
 use datatypes::arrow::array::{Array as _, ArrayRef, BooleanArray};
 use datatypes::arrow::buffer::BooleanBuffer;
 use datatypes::arrow::record_batch::RecordBatch;
-use datatypes::prelude::ConcreteDataType;
 use datatypes::schema::Schema;
-use mito_codec::row_converter::{CompositeValues, PrimaryKeyCodec};
+use mito_codec::row_converter::PrimaryKeyCodec;
 use parquet::arrow::arrow_reader::RowSelection;
 use parquet::file::metadata::ParquetMetaData;
 use snafu::{OptionExt, ResultExt};
@@ -37,23 +36,22 @@ use store_api::metadata::RegionMetadataRef;
 use store_api::storage::{ColumnId, TimeSeriesRowSelector};
 use table::predicate::Predicate;
 
+use crate::cache::CacheStrategy;
 use crate::error::{
-    ComputeArrowSnafu, DataTypeMismatchSnafu, DecodeSnafu, DecodeStatsSnafu,
-    EvalPartitionFilterSnafu, NewRecordBatchSnafu, RecordBatchSnafu, Result, StatsNotPresentSnafu,
-    UnexpectedSnafu,
+    ComputeArrowSnafu, DecodeStatsSnafu, EvalPartitionFilterSnafu, NewRecordBatchSnafu,
+    RecordBatchSnafu, Result, StatsNotPresentSnafu, UnexpectedSnafu,
 };
-use crate::read::Batch;
-use crate::read::compat::CompatBatch;
+use crate::read::compat::FlatCompatBatch;
 use crate::read::flat_projection::CompactionProjectionMapper;
-use crate::read::last_row::{FlatRowGroupLastRowCachedReader, RowGroupLastRowCachedReader};
-use crate::read::prune::{FlatPruneReader, PruneReader};
+use crate::read::last_row::FlatRowGroupLastRowCachedReader;
+use crate::read::prune::FlatPruneReader;
 use crate::sst::file::FileHandle;
 use crate::sst::parquet::flat_format::{
-    DecodedPrimaryKeys, decode_primary_keys, time_index_column_index,
+    DecodedPrimaryKeys, FlatReadFormat, decode_primary_keys, time_index_column_index,
 };
-use crate::sst::parquet::format::ReadFormat;
 use crate::sst::parquet::reader::{
-    FlatRowGroupReader, MaybeFilter, RowGroupReader, RowGroupReaderBuilder, SimpleFilterContext,
+    FlatRowGroupReader, MaybeFilter, RowGroupBuildContext, RowGroupReaderBuilder,
+    SimpleFilterContext,
 };
 use crate::sst::parquet::row_group::ParquetFetchMetrics;
 use crate::sst::parquet::stats::RowGroupPruningStats;
@@ -144,7 +142,7 @@ impl FileRange {
             std::slice::from_ref(curr_row_group),
             read_format,
             self.context.base.expected_metadata.clone(),
-            self.compute_skip_fields(),
+            self.context.base.pre_filter_mode.skip_fields(),
         );
 
         // not costly to create a predicate here since dynamic filters are wrapped in Arc
@@ -156,84 +154,6 @@ impl FileRange {
             .unwrap_or(true) // unexpected, not skip just in case
     }
 
-    fn compute_skip_fields(&self) -> bool {
-        match self.context.base.pre_filter_mode {
-            PreFilterMode::All => false,
-            PreFilterMode::SkipFields => true,
-            PreFilterMode::SkipFieldsOnDelete => {
-                // Check if this specific row group contains delete op
-                row_group_contains_delete(
-                    self.context.reader_builder.parquet_metadata(),
-                    self.row_group_idx,
-                    self.context.reader_builder.file_path(),
-                )
-                .unwrap_or(true)
-            }
-        }
-    }
-
-    /// Returns a reader to read the [FileRange].
-    pub(crate) async fn reader(
-        &self,
-        selector: Option<TimeSeriesRowSelector>,
-        fetch_metrics: Option<&ParquetFetchMetrics>,
-    ) -> Result<Option<PruneReader>> {
-        if !self.in_dynamic_filter_range() {
-            return Ok(None);
-        }
-        let parquet_reader = self
-            .context
-            .reader_builder
-            .build(
-                self.row_group_idx,
-                self.row_selection.clone(),
-                fetch_metrics,
-            )
-            .await?;
-
-        let use_last_row_reader = if selector
-            .map(|s| s == TimeSeriesRowSelector::LastRow)
-            .unwrap_or(false)
-        {
-            // Only use LastRowReader if row group does not contain DELETE
-            // and all rows are selected.
-            let put_only = !self
-                .context
-                .contains_delete(self.row_group_idx)
-                .inspect_err(|e| {
-                    error!(e; "Failed to decode min value of op_type, fallback to RowGroupReader");
-                })
-                .unwrap_or(true);
-            put_only && self.select_all()
-        } else {
-            // No selector provided, use RowGroupReader
-            false
-        };
-
-        // Compute skip_fields once for this row group
-        let skip_fields = self.context.should_skip_fields(self.row_group_idx);
-
-        let prune_reader = if use_last_row_reader {
-            // Row group is PUT only, use LastRowReader to skip unnecessary rows.
-            let reader = RowGroupLastRowCachedReader::new(
-                self.file_handle().file_id().file_id(),
-                self.row_group_idx,
-                self.context.reader_builder.cache_strategy().clone(),
-                RowGroupReader::new(self.context.clone(), parquet_reader),
-            );
-            PruneReader::new_with_last_row_reader(self.context.clone(), reader, skip_fields)
-        } else {
-            // Row group contains DELETE, fallback to default reader.
-            PruneReader::new_with_row_group_reader(
-                self.context.clone(),
-                RowGroupReader::new(self.context.clone(), parquet_reader),
-                skip_fields,
-            )
-        };
-
-        Ok(Some(prune_reader))
-    }
-
     /// Creates a flat reader that returns RecordBatch.
     pub(crate) async fn flat_reader(
         &self,
@@ -243,14 +163,16 @@ impl FileRange {
         if !self.in_dynamic_filter_range() {
             return Ok(None);
         }
+        // Compute skip_fields once for this row group
+        let skip_fields = self.context.base.pre_filter_mode.skip_fields();
         let parquet_reader = self
             .context
             .reader_builder
-            .build(
+            .build(self.context.build_context(
                 self.row_group_idx,
                 self.row_selection.clone(),
                 fetch_metrics,
-            )
+            ))
             .await?;
 
         let use_last_row_reader = if selector
@@ -271,17 +193,21 @@ impl FileRange {
             false
         };
 
-        // Compute skip_fields once for this row group
-        let skip_fields = self.context.should_skip_fields(self.row_group_idx);
-
         let flat_prune_reader = if use_last_row_reader {
             let flat_row_group_reader =
                 FlatRowGroupReader::new(self.context.clone(), parquet_reader);
+            // Flat PK prefilter makes the input stream predicate-dependent, so cached
+            // selector results are not reusable across queries with different filters.
+            let cache_strategy = if self.context.reader_builder.has_predicate_prefilter() {
+                CacheStrategy::Disabled
+            } else {
+                self.context.reader_builder.cache_strategy().clone()
+            };
             let reader = FlatRowGroupLastRowCachedReader::new(
                 self.file_handle().file_id().file_id(),
                 self.row_group_idx,
-                self.context.reader_builder.cache_strategy().clone(),
-                self.context.read_format().projection_indices(),
+                cache_strategy,
+                self.context.read_format().parquet_read_columns(),
                 flat_row_group_reader,
             );
             FlatPruneReader::new_with_last_row_reader(self.context.clone(), reader, skip_fields)
@@ -299,7 +225,7 @@ impl FileRange {
     }
 
     /// Returns the helper to compat batches.
-    pub(crate) fn compat_batch(&self) -> Option<&CompatBatch> {
+    pub(crate) fn compat_batch(&self) -> Option<&FlatCompatBatch> {
         self.context.compat_batch()
     }
 
@@ -333,11 +259,6 @@ impl FileRangeContext {
         }
     }
 
-    /// Returns the path of the file to read.
-    pub(crate) fn file_path(&self) -> &str {
-        self.reader_builder.file_path()
-    }
-
     /// Returns filters pushed down.
     pub(crate) fn filters(&self) -> &[SimpleFilterContext] {
         &self.base.filters
@@ -349,7 +270,7 @@ impl FileRangeContext {
     }
 
     /// Returns the format helper.
-    pub(crate) fn read_format(&self) -> &ReadFormat {
+    pub(crate) fn read_format(&self) -> &FlatReadFormat {
         &self.base.read_format
     }
 
@@ -359,7 +280,7 @@ impl FileRangeContext {
     }
 
     /// Returns the helper to compat batches.
-    pub(crate) fn compat_batch(&self) -> Option<&CompatBatch> {
+    pub(crate) fn compat_batch(&self) -> Option<&FlatCompatBatch> {
         self.base.compat_batch.as_ref()
     }
 
@@ -368,20 +289,14 @@ impl FileRangeContext {
         self.base.compaction_projection_mapper.as_ref()
     }
 
-    /// Sets the `CompatBatch` to the context.
-    pub(crate) fn set_compat_batch(&mut self, compat: Option<CompatBatch>) {
+    /// Sets the compat helper to the context.
+    pub(crate) fn set_compat_batch(&mut self, compat: Option<FlatCompatBatch>) {
         self.base.compat_batch = compat;
     }
 
-    /// TRY THE BEST to perform pushed down predicate precisely on the input batch.
-    /// Return the filtered batch. If the entire batch is filtered out, return None.
-    /// If a partition expr filter is configured, it is also applied.
-    pub(crate) fn precise_filter(&self, input: Batch, skip_fields: bool) -> Result<Option<Batch>> {
-        self.base.precise_filter(input, skip_fields)
-    }
-
     /// Filters the input RecordBatch by the pushed down predicate and returns RecordBatch.
     /// If a partition expr filter is configured, it is also applied.
+    /// Physical filter exprs are not evaluated here; they are only applied during prefiltering.
     pub(crate) fn precise_filter_flat(
         &self,
         input: RecordBatch,
@@ -390,16 +305,8 @@ impl FileRangeContext {
         self.base.precise_filter_flat(input, skip_fields)
     }
 
-    /// Determines whether to skip field filters based on PreFilterMode and row group delete status.
-    pub(crate) fn should_skip_fields(&self, row_group_idx: usize) -> bool {
-        match self.base.pre_filter_mode {
-            PreFilterMode::All => false,
-            PreFilterMode::SkipFields => true,
-            PreFilterMode::SkipFieldsOnDelete => {
-                // Check if this specific row group contains delete op
-                self.contains_delete(row_group_idx).unwrap_or(true)
-            }
-        }
+    pub(crate) fn pre_filter_mode(&self) -> PreFilterMode {
+        self.base.pre_filter_mode
     }
 
     //// Decodes parquet metadata and finds if row group contains delete op.
@@ -408,6 +315,20 @@ impl FileRangeContext {
         row_group_contains_delete(metadata, row_group_index, self.reader_builder.file_path())
     }
 
+    /// Creates a [RowGroupBuildContext] for building row group readers with prefiltering.
+    pub(crate) fn build_context<'a>(
+        &'a self,
+        row_group_idx: usize,
+        row_selection: Option<RowSelection>,
+        fetch_metrics: Option<&'a ParquetFetchMetrics>,
+    ) -> RowGroupBuildContext<'a> {
+        RowGroupBuildContext {
+            row_group_idx,
+            row_selection,
+            fetch_metrics,
+        }
+    }
+
     /// Returns the estimated memory size of this context.
     /// Mainly accounts for the parquet metadata size.
     pub(crate) fn memory_size(&self) -> usize {
@@ -416,17 +337,20 @@ impl FileRangeContext {
 }
 
 /// Mode to pre-filter columns in a range.
-#[derive(Debug, Clone, Copy)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum PreFilterMode {
     /// Filters all columns.
     All,
-    /// If the range doesn't contain delete op or doesn't have statistics, filters all columns.
-    /// Otherwise, skips filtering fields.
-    SkipFieldsOnDelete,
     /// Always skip fields.
     SkipFields,
 }
 
+impl PreFilterMode {
+    pub(crate) fn skip_fields(self) -> bool {
+        matches!(self, Self::SkipFields)
+    }
+}
+
 /// Context for partition expression filtering.
 pub(crate) struct PartitionFilterContext {
     pub(crate) region_partition_physical_expr: Arc<dyn PhysicalExpr>,
@@ -442,14 +366,14 @@ pub(crate) struct RangeBase {
     /// Dynamic filter physical exprs.
     pub(crate) dyn_filters: Vec<Arc<DynamicFilterPhysicalExpr>>,
     /// Helper to read the SST.
-    pub(crate) read_format: ReadFormat,
+    pub(crate) read_format: FlatReadFormat,
     pub(crate) expected_metadata: Option<RegionMetadataRef>,
     /// Schema used for pruning with dynamic filters.
     pub(crate) prune_schema: Arc<Schema>,
     /// Decoder for primary keys
     pub(crate) codec: Arc<dyn PrimaryKeyCodec>,
     /// Optional helper to compat batches.
-    pub(crate) compat_batch: Option<CompatBatch>,
+    pub(crate) compat_batch: Option<FlatCompatBatch>,
     /// Optional helper to project batches.
     pub(crate) compaction_projection_mapper: Option<CompactionProjectionMapper>,
     /// Mode to pre-filter columns.
@@ -473,129 +397,13 @@ impl TagDecodeState {
 }
 
 impl RangeBase {
-    /// TRY THE BEST to perform pushed down predicate precisely on the input batch.
-    /// Return the filtered batch. If the entire batch is filtered out, return None.
-    ///
-    /// Supported filter expr type is defined in [SimpleFilterEvaluator].
-    ///
-    /// When a filter is referencing primary key column, this method will decode
-    /// the primary key and put it into the batch.
-    ///
-    /// # Arguments
-    /// * `input` - The batch to filter
-    /// * `skip_fields` - Whether to skip field filters based on PreFilterMode and row group delete status
-    pub(crate) fn precise_filter(
-        &self,
-        mut input: Batch,
-        skip_fields: bool,
-    ) -> Result<Option<Batch>> {
-        let mut mask = BooleanBuffer::new_set(input.num_rows());
-
-        // Run filter one by one and combine them result
-        // TODO(ruihang): run primary key filter first. It may short circuit other filters
-        for filter_ctx in &self.filters {
-            let filter = match filter_ctx.filter() {
-                MaybeFilter::Filter(f) => f,
-                // Column matches.
-                MaybeFilter::Matched => continue,
-                // Column doesn't match, filter the entire batch.
-                MaybeFilter::Pruned => return Ok(None),
-            };
-            let result = match filter_ctx.semantic_type() {
-                SemanticType::Tag => {
-                    let pk_values = if let Some(pk_values) = input.pk_values() {
-                        pk_values
-                    } else {
-                        input.set_pk_values(
-                            self.codec
-                                .decode(input.primary_key())
-                                .context(DecodeSnafu)?,
-                        );
-                        input.pk_values().unwrap()
-                    };
-                    let pk_value = match pk_values {
-                        CompositeValues::Dense(v) => {
-                            // Safety: this is a primary key
-                            let pk_index = self
-                                .read_format
-                                .metadata()
-                                .primary_key_index(filter_ctx.column_id())
-                                .unwrap();
-                            v[pk_index]
-                                .1
-                                .try_to_scalar_value(filter_ctx.data_type())
-                                .context(DataTypeMismatchSnafu)?
-                        }
-                        CompositeValues::Sparse(v) => {
-                            let v = v.get_or_null(filter_ctx.column_id());
-                            v.try_to_scalar_value(filter_ctx.data_type())
-                                .context(DataTypeMismatchSnafu)?
-                        }
-                    };
-                    if filter
-                        .evaluate_scalar(&pk_value)
-                        .context(RecordBatchSnafu)?
-                    {
-                        continue;
-                    } else {
-                        // PK not match means the entire batch is filtered out.
-                        return Ok(None);
-                    }
-                }
-                SemanticType::Field => {
-                    // Skip field filters if skip_fields is true
-                    if skip_fields {
-                        continue;
-                    }
-                    // Safety: Input is Batch so we are using primary key format.
-                    let Some(field_index) = self
-                        .read_format
-                        .as_primary_key()
-                        .unwrap()
-                        .field_index_by_id(filter_ctx.column_id())
-                    else {
-                        continue;
-                    };
-                    let field_col = &input.fields()[field_index].data;
-                    filter
-                        .evaluate_vector(field_col)
-                        .context(RecordBatchSnafu)?
-                }
-                SemanticType::Timestamp => filter
-                    .evaluate_vector(input.timestamps())
-                    .context(RecordBatchSnafu)?,
-            };
-
-            mask = mask.bitand(&result);
-        }
-
-        if mask.count_set_bits() == 0 {
-            return Ok(None);
-        }
-
-        // Apply partition filter
-        if let Some(partition_filter) = &self.partition_filter {
-            let record_batch = self
-                .build_record_batch_for_pruning(&mut input, &partition_filter.partition_schema)?;
-            let partition_mask = self.evaluate_partition_filter(&record_batch, partition_filter)?;
-            mask = mask.bitand(&partition_mask);
-        }
-
-        if mask.count_set_bits() == 0 {
-            Ok(None)
-        } else {
-            input.filter(&BooleanArray::from(mask).into())?;
-            Ok(Some(input))
-        }
-    }
-
     /// Filters the input RecordBatch by the pushed down predicate and returns RecordBatch.
     ///
     /// It assumes all necessary tags are already decoded from the primary key.
     ///
     /// # Arguments
     /// * `input` - The RecordBatch to filter
-    /// * `skip_fields` - Whether to skip field filters based on PreFilterMode and row group delete status
+    /// * `skip_fields` - Whether to skip field filters based on PreFilterMode
     pub(crate) fn precise_filter_flat(
         &self,
         input: RecordBatch,
@@ -637,12 +445,13 @@ impl RangeBase {
 
     /// Computes the filter mask for the input RecordBatch based on pushed down predicates.
     /// If a partition expr filter is configured, it is applied later in `precise_filter_flat` but **NOT** in this function.
+    /// Physical filter exprs are excluded here and only apply during prefiltering.
     ///
     /// Returns `None` if the entire batch is filtered out, otherwise returns the boolean mask.
     ///
     /// # Arguments
     /// * `input` - The RecordBatch to compute mask for
-    /// * `skip_fields` - Whether to skip field filters based on PreFilterMode and row group delete status
+    /// * `skip_fields` - Whether to skip field filters based on PreFilterMode
     pub(crate) fn compute_filter_mask_flat(
         &self,
         input: &RecordBatch,
@@ -651,13 +460,7 @@ impl RangeBase {
     ) -> Result<Option<BooleanBuffer>> {
         let mut mask = BooleanBuffer::new_set(input.num_rows());
 
-        let flat_format = self
-            .read_format
-            .as_flat()
-            .context(crate::error::UnexpectedSnafu {
-                reason: "Expected flat format for precise_filter_flat",
-            })?;
-        let metadata = flat_format.metadata();
+        let metadata = self.read_format.metadata();
 
         // Run filter one by one and combine them result
         for filter_ctx in &self.filters {
@@ -677,7 +480,9 @@ impl RangeBase {
             // Get the column directly by its projected index.
             // If the column is missing and it's not a tag/time column, this filter is skipped.
             // Assumes the projection indices align with the input batch schema.
-            let column_idx = flat_format.projected_index_by_id(filter_ctx.column_id());
+            let column_idx = self
+                .read_format
+                .projected_index_by_id(filter_ctx.column_id());
             if let Some(idx) = column_idx {
                 let column = &input.columns().get(idx).unwrap();
                 let result = filter.evaluate_array(column).context(RecordBatchSnafu)?;
@@ -781,84 +586,6 @@ impl RangeBase {
         Ok(mask)
     }
 
-    /// Builds a `RecordBatch` from the input `Batch` matching the given schema.
-    ///
-    /// This is used for partition expression evaluation. The schema should only contain
-    /// the columns referenced by the partition expression to minimize overhead.
-    fn build_record_batch_for_pruning(
-        &self,
-        input: &mut Batch,
-        schema: &Arc<Schema>,
-    ) -> Result<RecordBatch> {
-        let arrow_schema = schema.arrow_schema();
-        let mut columns = Vec::with_capacity(arrow_schema.fields().len());
-
-        // Decode primary key if necessary.
-        if input.pk_values().is_none() {
-            input.set_pk_values(
-                self.codec
-                    .decode(input.primary_key())
-                    .context(DecodeSnafu)?,
-            );
-        }
-
-        for field in arrow_schema.fields() {
-            let metadata = self.read_format.metadata();
-            let column_id = metadata.column_by_name(field.name()).map(|c| c.column_id);
-
-            // Partition pruning schema should be a subset of the input batch schema.
-            let Some(column_id) = column_id else {
-                return UnexpectedSnafu {
-                    reason: format!(
-                        "Partition pruning schema expects column '{}' but it is missing in \
-                         region metadata",
-                        field.name()
-                    ),
-                }
-                .fail();
-            };
-
-            // 1. Check if it's a tag.
-            if let Some(pk_index) = metadata.primary_key_index(column_id) {
-                let pk_values = input.pk_values().unwrap();
-                let value = match pk_values {
-                    CompositeValues::Dense(v) => &v[pk_index].1,
-                    CompositeValues::Sparse(v) => v.get_or_null(column_id),
-                };
-                let concrete_type = ConcreteDataType::from_arrow_type(field.data_type());
-                let arrow_scalar = value
-                    .try_to_scalar_value(&concrete_type)
-                    .context(DataTypeMismatchSnafu)?;
-                let array = arrow_scalar
-                    .to_array_of_size(input.num_rows())
-                    .context(EvalPartitionFilterSnafu)?;
-                columns.push(array);
-            } else if metadata.time_index_column().column_id == column_id {
-                // 2. Check if it's the timestamp column.
-                columns.push(input.timestamps().to_arrow_array());
-            } else if let Some(field_index) = self
-                .read_format
-                .as_primary_key()
-                .and_then(|f| f.field_index_by_id(column_id))
-            {
-                // 3. Check if it's a field column.
-                columns.push(input.fields()[field_index].data.to_arrow_array());
-            } else {
-                return UnexpectedSnafu {
-                    reason: format!(
-                        "Partition pruning schema expects column '{}' (id {}) but it is not \
-                         present in input batch",
-                        field.name(),
-                        column_id
-                    ),
-                }
-                .fail();
-            }
-        }
-
-        RecordBatch::try_new(arrow_schema.clone(), columns).context(NewRecordBatchSnafu)
-    }
-
     /// Projects the input `RecordBatch` to match the given schema.
     ///
     /// This is used for partition expression evaluation. The schema should only contain
@@ -872,13 +599,7 @@ impl RangeBase {
         let arrow_schema = schema.arrow_schema();
         let mut columns = Vec::with_capacity(arrow_schema.fields().len());
 
-        let flat_format = self
-            .read_format
-            .as_flat()
-            .context(crate::error::UnexpectedSnafu {
-                reason: "Expected flat format for precise_filter_flat",
-            })?;
-        let metadata = flat_format.metadata();
+        let metadata = self.read_format.metadata();
 
         for field in arrow_schema.fields() {
             let column_id = metadata.column_by_name(field.name()).map(|c| c.column_id);
@@ -894,7 +615,7 @@ impl RangeBase {
                 .fail();
             };
 
-            if let Some(idx) = flat_format.projected_index_by_id(column_id) {
+            if let Some(idx) = self.read_format.projected_index_by_id(column_id) {
                 columns.push(input.column(idx).clone());
                 continue;
             }
@@ -926,3 +647,90 @@ impl RangeBase {
         RecordBatch::try_new(arrow_schema.clone(), columns).context(NewRecordBatchSnafu)
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use datafusion_expr::{col, lit};
+
+    use super::*;
+    use crate::read::read_columns::ReadColumns;
+    use crate::sst::parquet::flat_format::FlatReadFormat;
+    use crate::test_util::sst_util::{new_record_batch_with_custom_sequence, sst_region_metadata};
+
+    fn new_test_range_base(filters: Vec<SimpleFilterContext>) -> RangeBase {
+        let metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
+
+        let read_format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids(
+                metadata.column_metadatas.iter().map(|c| c.column_id),
+            ),
+            None,
+            "test",
+            true,
+        )
+        .unwrap();
+
+        RangeBase {
+            filters,
+            dyn_filters: vec![],
+            read_format,
+            expected_metadata: None,
+            prune_schema: metadata.schema.clone(),
+            codec: mito_codec::row_converter::build_primary_key_codec(metadata.as_ref()),
+            compat_batch: None,
+            compaction_projection_mapper: None,
+            pre_filter_mode: PreFilterMode::All,
+            partition_filter: None,
+        }
+    }
+
+    #[test]
+    fn test_compute_filter_mask_flat_applies_remaining_simple_filters() {
+        let metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
+        let filters = vec![
+            SimpleFilterContext::new_opt(&metadata, None, &col("tag_0").eq(lit("a"))).unwrap(),
+            SimpleFilterContext::new_opt(&metadata, None, &col("field_0").gt(lit(1_u64))).unwrap(),
+        ];
+        let base = new_test_range_base(filters);
+        let batch = new_record_batch_with_custom_sequence(&["b", "x"], 0, 4, 1);
+
+        let mask = base
+            .compute_filter_mask_flat(&batch, false, &mut TagDecodeState::new())
+            .unwrap()
+            .unwrap();
+        assert_eq!(mask.count_set_bits(), 0);
+    }
+
+    #[test]
+    fn test_compute_filter_mask_flat_does_not_postfilter_physical_filters() {
+        let metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
+        let read_format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids(
+                metadata.column_metadatas.iter().map(|c| c.column_id),
+            ),
+            None,
+            "test",
+            true,
+        )
+        .unwrap();
+        let physical_filter = crate::sst::parquet::reader::PhysicalFilterContext::new_opt(
+            &metadata,
+            None,
+            &read_format,
+            &col("field_0").in_list(vec![lit(1_u64), lit(2_u64)], false),
+        );
+        assert!(physical_filter.is_some());
+        let base = new_test_range_base(vec![]);
+        let batch = new_record_batch_with_custom_sequence(&["b", "x"], 0, 4, 1);
+
+        let mask = base
+            .compute_filter_mask_flat(&batch, false, &mut TagDecodeState::new())
+            .unwrap()
+            .unwrap();
+        assert_eq!(mask.count_set_bits(), 4);
+    }
+}
diff --git a/src/mito2/src/sst/parquet/flat_format.rs b/src/mito2/src/sst/parquet/flat_format.rs
index d6b061e468..d8111dbe62 100644
--- a/src/mito2/src/sst/parquet/flat_format.rs
+++ b/src/mito2/src/sst/parquet/flat_format.rs
@@ -51,13 +51,15 @@ use crate::error::{
     ComputeArrowSnafu, DecodeSnafu, InvalidParquetSnafu, InvalidRecordBatchSnafu,
     NewRecordBatchSnafu, Result,
 };
+use crate::read::read_columns::ReadColumns;
 use crate::sst::parquet::format::{
-    FormatProjection, INTERNAL_COLUMN_NUM, PrimaryKeyArray, PrimaryKeyReadFormat, ReadFormat,
-    StatValues,
+    FIXED_POS_COLUMN_NUM, FormatProjection, INTERNAL_COLUMN_NUM, PrimaryKeyArray,
+    PrimaryKeyReadFormat, StatValues, column_null_counts, column_values,
 };
+use crate::sst::parquet::read_columns::ParquetReadColumns;
 use crate::sst::{
     FlatSchemaOptions, flat_sst_arrow_schema_column_num, tag_maybe_to_dictionary_field,
-    to_flat_sst_arrow_schema,
+    to_flat_sst_arrow_schema, with_field_id,
 };
 
 /// Helper for writing the SST format.
@@ -87,6 +89,7 @@ impl FlatWriteFormat {
     }
 
     /// Gets the arrow schema to store in parquet.
+    #[cfg(test)]
     pub(crate) fn arrow_schema(&self) -> &SchemaRef {
         &self.arrow_schema
     }
@@ -103,7 +106,7 @@ impl FlatWriteFormat {
         let sequence_array = Arc::new(UInt64Array::from(vec![override_sequence; batch.num_rows()]));
         columns[sequence_column_index(batch.num_columns())] = sequence_array;
 
-        RecordBatch::try_new(self.arrow_schema.clone(), columns).context(NewRecordBatchSnafu)
+        RecordBatch::try_new(batch.schema(), columns).context(NewRecordBatchSnafu)
     }
 }
 
@@ -127,6 +130,21 @@ pub(crate) fn op_type_column_index(num_columns: usize) -> usize {
     num_columns - 1
 }
 
+/// Returns the start index of field columns in a flat batch.
+///
+/// `num_columns` is the total number of columns in the flat batch schema,
+/// including tag columns (if present), field columns, and fixed position columns
+/// (time index, primary key, sequence, op type).
+///
+/// For Dense encoding (raw PK columns included): field_column_start = primary_key.len()
+/// For Sparse encoding (no raw PK columns): field_column_start = 0
+pub(crate) fn field_column_start(metadata: &RegionMetadata, num_columns: usize) -> usize {
+    // Calculates field column start: total columns - fixed columns - field columns
+    // Field column count = total metadata columns - time index column - primary key columns
+    let field_column_count = metadata.column_metadatas.len() - 1 - metadata.primary_key.len();
+    num_columns - FIXED_POS_COLUMN_NUM - field_column_count
+}
+
 // TODO(yingwen): Add an option to skip reading internal columns if the region is
 // append only and doesn't use sparse encoding (We need to check the table id under
 // sparse encoding).
@@ -146,7 +164,7 @@ impl FlatReadFormat {
     /// If `skip_auto_convert` is true, skips auto conversion of format when the encoding is sparse encoding.
     pub fn new(
         metadata: RegionMetadataRef,
-        column_ids: impl Iterator<Item = ColumnId>,
+        read_cols: ReadColumns,
         num_columns: Option<usize>,
         file_path: &str,
         skip_auto_convert: bool,
@@ -162,16 +180,16 @@ impl FlatReadFormat {
                 // Only skip auto convert when the primary key encoding is sparse.
                 ParquetAdapter::PrimaryKeyToFlat(ParquetPrimaryKeyToFlat::new(
                     metadata,
-                    column_ids,
+                    read_cols,
                     skip_auto_convert,
                 ))
             } else {
                 ParquetAdapter::PrimaryKeyToFlat(ParquetPrimaryKeyToFlat::new(
-                    metadata, column_ids, false,
+                    metadata, read_cols, false,
                 ))
             }
         } else {
-            ParquetAdapter::Flat(ParquetFlat::new(metadata, column_ids))
+            ParquetAdapter::Flat(ParquetFlat::new(metadata, read_cols))
         };
 
         Ok(FlatReadFormat {
@@ -240,6 +258,16 @@ impl FlatReadFormat {
         }
     }
 
+    /// Gets the projected output schema produced by parquet reading.
+    pub(crate) fn output_arrow_schema(&self) -> Result<SchemaRef> {
+        let projection = self.parquet_read_columns().root_indices();
+        let schema = self
+            .arrow_schema()
+            .project(projection)
+            .context(ComputeArrowSnafu)?;
+        Ok(Arc::new(schema))
+    }
+
     /// Gets the metadata of the SST.
     pub(crate) fn metadata(&self) -> &RegionMetadataRef {
         match &self.parquet_adapter {
@@ -248,11 +276,11 @@ impl FlatReadFormat {
         }
     }
 
-    /// Gets sorted projection indices to read from the SST file.
-    pub(crate) fn projection_indices(&self) -> &[usize] {
+    /// Get the sorted read columns to read from the sst file.
+    pub(crate) fn parquet_read_columns(&self) -> &ParquetReadColumns {
         match &self.parquet_adapter {
-            ParquetAdapter::Flat(p) => &p.format_projection.projection_indices,
-            ParquetAdapter::PrimaryKeyToFlat(p) => p.format.projection_indices(),
+            ParquetAdapter::Flat(p) => &p.format_projection.parquet_read_cols,
+            ParquetAdapter::PrimaryKeyToFlat(p) => p.format.parquet_read_columns(),
         }
     }
 
@@ -267,6 +295,13 @@ impl FlatReadFormat {
         }
     }
 
+    /// Returns `true` if raw batches from parquet use the flat layout and
+    /// stores primary key columns as raw columns.
+    /// Returns `false` for the legacy primary-key-to-flat conversion path.
+    pub(crate) fn batch_has_raw_pk_columns(&self) -> bool {
+        matches!(&self.parquet_adapter, ParquetAdapter::Flat(_))
+    }
+
     /// Creates a sequence array to override.
     pub(crate) fn new_override_sequence_array(&self, length: usize) -> Option<ArrayRef> {
         self.override_sequence
@@ -381,7 +416,7 @@ impl ParquetPrimaryKeyToFlat {
     /// Creates a helper with existing `metadata` and `column_ids` to read.
     fn new(
         metadata: RegionMetadataRef,
-        column_ids: impl Iterator<Item = ColumnId>,
+        read_cols: ReadColumns,
         skip_auto_convert: bool,
     ) -> ParquetPrimaryKeyToFlat {
         assert!(if skip_auto_convert {
@@ -390,20 +425,18 @@ impl ParquetPrimaryKeyToFlat {
             true
         });
 
-        let column_ids: Vec<_> = column_ids.collect();
-
         // Creates a map to lookup index based on the new format.
         let id_to_index = sst_column_id_indices(&metadata);
         let sst_column_num =
             flat_sst_arrow_schema_column_num(&metadata, &FlatSchemaOptions::default());
 
         let codec = build_primary_key_codec(&metadata);
-        let format = PrimaryKeyReadFormat::new(metadata.clone(), column_ids.iter().copied());
+        let format = PrimaryKeyReadFormat::new(metadata.clone(), read_cols.clone());
         let (convert_format, format_projection) = if skip_auto_convert {
             (
                 None,
                 FormatProjection {
-                    projection_indices: format.projection_indices().to_vec(),
+                    parquet_read_cols: format.parquet_read_columns().clone(),
                     column_id_to_projected_index: format.field_id_to_projected_index().clone(),
                 },
             )
@@ -412,7 +445,7 @@ impl ParquetPrimaryKeyToFlat {
             let format_projection = FormatProjection::compute_format_projection(
                 &id_to_index,
                 sst_column_num,
-                column_ids.iter().copied(),
+                read_cols.clone(),
             );
             (
                 FlatConvertFormat::new(Arc::clone(&metadata), &format_projection, codec),
@@ -450,14 +483,14 @@ struct ParquetFlat {
 
 impl ParquetFlat {
     /// Creates a helper with existing `metadata` and `column_ids` to read.
-    fn new(metadata: RegionMetadataRef, column_ids: impl Iterator<Item = ColumnId>) -> ParquetFlat {
+    fn new(metadata: RegionMetadataRef, read_cols: ReadColumns) -> ParquetFlat {
         // Creates a map to lookup index.
         let id_to_index = sst_column_id_indices(&metadata);
         let arrow_schema = to_flat_sst_arrow_schema(&metadata, &FlatSchemaOptions::default());
         let sst_column_num =
             flat_sst_arrow_schema_column_num(&metadata, &FlatSchemaOptions::default());
         let format_projection =
-            FormatProjection::compute_format_projection(&id_to_index, sst_column_num, column_ids);
+            FormatProjection::compute_format_projection(&id_to_index, sst_column_num, read_cols);
 
         Self {
             metadata,
@@ -496,7 +529,7 @@ impl ParquetFlat {
             return StatValues::NoColumn;
         };
 
-        let stats = ReadFormat::column_null_counts(row_groups, *index);
+        let stats = column_null_counts(row_groups, *index);
         StatValues::from_stats_opt(stats)
     }
 
@@ -513,7 +546,7 @@ impl ParquetFlat {
         // Safety: `column_id_to_sst_index` is built from `metadata`.
         let index = self.column_id_to_sst_index.get(&column_id).unwrap();
 
-        let stats = ReadFormat::column_values(row_groups, column, *index, is_min);
+        let stats = column_values(row_groups, column, *index, is_min);
         StatValues::from_stats_opt(stats)
     }
 }
@@ -737,12 +770,12 @@ impl FlatConvertFormat {
         // Builds new schema
         let mut new_fields =
             Vec::with_capacity(batch.schema().fields().len() + self.projected_primary_keys.len());
-        for (_, _, column_index) in &self.projected_primary_keys {
+        for (column_id, _, column_index) in &self.projected_primary_keys {
             let column_metadata = &self.metadata.column_metadatas[*column_index];
             let old_field = &self.metadata.schema.arrow_schema().fields()[*column_index];
             let field =
                 tag_maybe_to_dictionary_field(&column_metadata.column_schema.data_type, old_field);
-            new_fields.push(field);
+            new_fields.push(Arc::new(with_field_id((*field).clone(), *column_id)));
         }
         new_fields.extend(batch.schema().fields().iter().cloned());
 
@@ -757,7 +790,9 @@ impl FlatReadFormat {
     pub fn new_with_all_columns(metadata: RegionMetadataRef) -> FlatReadFormat {
         Self::new(
             Arc::clone(&metadata),
-            metadata.column_metadatas.iter().map(|c| c.column_id),
+            ReadColumns::from_deduped_column_ids(
+                metadata.column_metadatas.iter().map(|c| c.column_id),
+            ),
             None,
             "test",
             false,
@@ -765,3 +800,117 @@ impl FlatReadFormat {
         .unwrap()
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use api::v1::SemanticType;
+    use datatypes::prelude::ConcreteDataType;
+    use datatypes::schema::ColumnSchema;
+    use store_api::codec::PrimaryKeyEncoding;
+    use store_api::metadata::{ColumnMetadata, RegionMetadata, RegionMetadataBuilder};
+    use store_api::storage::RegionId;
+
+    use super::{FlatReadFormat, field_column_start};
+    use crate::read::read_columns::ReadColumns;
+    use crate::sst::{
+        FlatSchemaOptions, flat_sst_arrow_schema_column_num, to_flat_sst_arrow_schema,
+    };
+
+    /// Builds a `RegionMetadata` with the given number of tags and fields.
+    fn build_metadata(
+        num_tags: usize,
+        num_fields: usize,
+        encoding: PrimaryKeyEncoding,
+    ) -> RegionMetadata {
+        let mut builder = RegionMetadataBuilder::new(RegionId::new(0, 0));
+        let mut col_id = 0u32;
+
+        for i in 0..num_tags {
+            builder.push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    format!("tag_{i}"),
+                    ConcreteDataType::string_datatype(),
+                    true,
+                ),
+                semantic_type: SemanticType::Tag,
+                column_id: col_id,
+            });
+            col_id += 1;
+        }
+
+        for i in 0..num_fields {
+            builder.push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    format!("field_{i}"),
+                    ConcreteDataType::uint64_datatype(),
+                    true,
+                ),
+                semantic_type: SemanticType::Field,
+                column_id: col_id,
+            });
+            col_id += 1;
+        }
+
+        builder.push_column_metadata(ColumnMetadata {
+            column_schema: ColumnSchema::new(
+                "ts".to_string(),
+                ConcreteDataType::timestamp_millisecond_datatype(),
+                false,
+            ),
+            semantic_type: SemanticType::Timestamp,
+            column_id: col_id,
+        });
+
+        let primary_key: Vec<u32> = (0..num_tags as u32).collect();
+        builder.primary_key(primary_key);
+        builder.primary_key_encoding(encoding);
+        builder.build().unwrap()
+    }
+
+    #[test]
+    fn test_field_column_start() {
+        // (num_tags, num_fields, encoding, expected)
+        let cases = [
+            (1, 1, PrimaryKeyEncoding::Dense, 1),
+            (2, 2, PrimaryKeyEncoding::Dense, 2),
+            (0, 2, PrimaryKeyEncoding::Dense, 0),
+            (2, 2, PrimaryKeyEncoding::Sparse, 0),
+        ];
+
+        for (num_tags, num_fields, encoding, expected) in cases {
+            let metadata = build_metadata(num_tags, num_fields, encoding);
+            let options = FlatSchemaOptions::from_encoding(encoding);
+            let num_columns = flat_sst_arrow_schema_column_num(&metadata, &options);
+            let result = field_column_start(&metadata, num_columns);
+            assert_eq!(
+                result, expected,
+                "num_tags={num_tags}, num_fields={num_fields}, encoding={encoding:?}"
+            );
+        }
+    }
+
+    #[test]
+    fn test_output_arrow_schema_uses_projection() {
+        let metadata = Arc::new(build_metadata(1, 2, PrimaryKeyEncoding::Dense));
+        let read_format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids([0_u32, 2_u32]),
+            None,
+            "test",
+            false,
+        )
+        .unwrap();
+
+        let output_schema = read_format.output_arrow_schema().unwrap();
+        let projection = read_format.parquet_read_columns().root_indices();
+        let expected = Arc::new(
+            to_flat_sst_arrow_schema(&metadata, &FlatSchemaOptions::default())
+                .project(projection)
+                .unwrap(),
+        );
+
+        assert_eq!(expected, output_schema);
+    }
+}
diff --git a/src/mito2/src/sst/parquet/format.rs b/src/mito2/src/sst/parquet/format.rs
index 70d026e6db..b9afdfaac7 100644
--- a/src/mito2/src/sst/parquet/format.rs
+++ b/src/mito2/src/sst/parquet/format.rs
@@ -34,15 +34,14 @@ use api::v1::SemanticType;
 use common_time::Timestamp;
 use datafusion_common::ScalarValue;
 use datatypes::arrow::array::{
-    ArrayRef, BinaryArray, BinaryDictionaryBuilder, DictionaryArray, UInt32Array, UInt64Array,
+    ArrayRef, BinaryArray, BinaryDictionaryBuilder, DictionaryArray, UInt64Array,
 };
 use datatypes::arrow::datatypes::{SchemaRef, UInt32Type};
 use datatypes::arrow::record_batch::RecordBatch;
 use datatypes::prelude::DataType;
-use datatypes::vectors::{Helper, Vector};
+use datatypes::vectors::Helper;
 use mito_codec::row_converter::{
-    CompositeValues, PrimaryKeyCodec, SortField, build_primary_key_codec,
-    build_primary_key_codec_with_fields,
+    CompositeValues, PrimaryKeyCodec, SortField, build_primary_key_codec_with_fields,
 };
 use parquet::file::metadata::{ParquetMetaData, RowGroupMetaData};
 use parquet::file::statistics::Statistics;
@@ -51,12 +50,12 @@ use store_api::metadata::{ColumnMetadata, RegionMetadataRef};
 use store_api::storage::{ColumnId, SequenceNumber};
 
 use crate::error::{
-    ConvertVectorSnafu, DecodeSnafu, InvalidBatchSnafu, InvalidRecordBatchSnafu,
-    NewRecordBatchSnafu, Result,
+    ConvertVectorSnafu, DecodeSnafu, InvalidRecordBatchSnafu, NewRecordBatchSnafu, Result,
 };
+use crate::read::read_columns::ReadColumns;
 use crate::read::{Batch, BatchBuilder, BatchColumn};
 use crate::sst::file::{FileMeta, FileTimeRange};
-use crate::sst::parquet::flat_format::FlatReadFormat;
+use crate::sst::parquet::read_columns::{ParquetReadColumn, ParquetReadColumns};
 use crate::sst::to_sst_arrow_schema;
 
 /// Arrow array type for the primary key dictionary.
@@ -73,7 +72,6 @@ pub(crate) const INTERNAL_COLUMN_NUM: usize = 3;
 
 /// Helper for writing the SST format with primary key.
 pub(crate) struct PrimaryKeyWriteFormat {
-    metadata: RegionMetadataRef,
     /// SST file schema.
     arrow_schema: SchemaRef,
     override_sequence: Option<SequenceNumber>,
@@ -84,7 +82,6 @@ impl PrimaryKeyWriteFormat {
     pub(crate) fn new(metadata: RegionMetadataRef) -> PrimaryKeyWriteFormat {
         let arrow_schema = to_sst_arrow_schema(&metadata);
         PrimaryKeyWriteFormat {
-            metadata,
             arrow_schema,
             override_sequence: None,
         }
@@ -100,310 +97,111 @@ impl PrimaryKeyWriteFormat {
     }
 
     /// Gets the arrow schema to store in parquet.
+    #[cfg(test)]
     pub(crate) fn arrow_schema(&self) -> &SchemaRef {
         &self.arrow_schema
     }
 
-    /// Convert `batch` to a arrow record batch to store in parquet.
-    pub(crate) fn convert_batch(&self, batch: &Batch) -> Result<RecordBatch> {
-        debug_assert_eq!(
-            batch.fields().len() + FIXED_POS_COLUMN_NUM,
-            self.arrow_schema.fields().len()
-        );
-        let mut columns = Vec::with_capacity(batch.fields().len() + FIXED_POS_COLUMN_NUM);
-        // Store all fields first.
-        for (column, column_metadata) in batch.fields().iter().zip(self.metadata.field_columns()) {
-            ensure!(
-                column.column_id == column_metadata.column_id,
-                InvalidBatchSnafu {
-                    reason: format!(
-                        "Batch has column {} but metadata has column {}",
-                        column.column_id, column_metadata.column_id
-                    ),
-                }
-            );
-
-            columns.push(column.data.to_arrow_array());
-        }
-        // Add time index column.
-        columns.push(batch.timestamps().to_arrow_array());
-        // Add internal columns: primary key, sequences, op types.
-        columns.push(new_primary_key_array(batch.primary_key(), batch.num_rows()));
+    /// Convert a flat `RecordBatch` to primary-key format, retaining only
+    /// field columns, time index, and internal columns.
+    ///
+    /// `num_fields` is the number of field columns. The method strips
+    /// leading tag columns: `num_tag_columns = batch.num_columns() - num_fields - FIXED_POS_COLUMN_NUM`.
+    pub(crate) fn convert_flat_batch(
+        &self,
+        batch: &RecordBatch,
+        num_fields: usize,
+    ) -> Result<RecordBatch> {
+        let num_tag_columns = batch.num_columns() - num_fields - FIXED_POS_COLUMN_NUM;
+        let mut columns: Vec<ArrayRef> = batch.columns()[num_tag_columns..].to_vec();
 
         if let Some(override_sequence) = self.override_sequence {
-            let sequence_array =
+            let num_cols = columns.len();
+            // sequence is at num_cols - 2 (before op_type)
+            columns[num_cols - 2] =
                 Arc::new(UInt64Array::from(vec![override_sequence; batch.num_rows()]));
-            columns.push(sequence_array);
-        } else {
-            columns.push(batch.sequences().to_arrow_array());
         }
-        columns.push(batch.op_types().to_arrow_array());
 
         RecordBatch::try_new(self.arrow_schema.clone(), columns).context(NewRecordBatchSnafu)
     }
 }
 
-/// Helper to read parquet formats.
-pub enum ReadFormat {
-    /// The parquet is in the old primary key format.
-    PrimaryKey(PrimaryKeyReadFormat),
-    /// The parquet is in the new flat format.
-    Flat(FlatReadFormat),
+/// Returns min/max values of specific columns.
+/// Returns None if the column does not have statistics.
+/// The column should not be encoded as a part of a primary key.
+pub(crate) fn column_values(
+    row_groups: &[impl Borrow<RowGroupMetaData>],
+    column: &ColumnMetadata,
+    column_index: usize,
+    is_min: bool,
+) -> Option<ArrayRef> {
+    let null_scalar: ScalarValue = column
+        .column_schema
+        .data_type
+        .as_arrow_type()
+        .try_into()
+        .ok()?;
+    let scalar_values = row_groups
+        .iter()
+        .map(|meta| {
+            let stats = meta.borrow().column(column_index).statistics()?;
+            match stats {
+                Statistics::Boolean(s) => Some(ScalarValue::Boolean(Some(if is_min {
+                    *s.min_opt()?
+                } else {
+                    *s.max_opt()?
+                }))),
+                Statistics::Int32(s) => Some(ScalarValue::Int32(Some(if is_min {
+                    *s.min_opt()?
+                } else {
+                    *s.max_opt()?
+                }))),
+                Statistics::Int64(s) => Some(ScalarValue::Int64(Some(if is_min {
+                    *s.min_opt()?
+                } else {
+                    *s.max_opt()?
+                }))),
+                Statistics::Int96(_) => None,
+                Statistics::Float(s) => Some(ScalarValue::Float32(Some(if is_min {
+                    *s.min_opt()?
+                } else {
+                    *s.max_opt()?
+                }))),
+                Statistics::Double(s) => Some(ScalarValue::Float64(Some(if is_min {
+                    *s.min_opt()?
+                } else {
+                    *s.max_opt()?
+                }))),
+                Statistics::ByteArray(s) => {
+                    let bytes = if is_min {
+                        s.min_bytes_opt()?
+                    } else {
+                        s.max_bytes_opt()?
+                    };
+                    let s = String::from_utf8(bytes.to_vec()).ok();
+                    Some(ScalarValue::Utf8(s))
+                }
+                Statistics::FixedLenByteArray(_) => None,
+            }
+        })
+        .map(|maybe_scalar| maybe_scalar.unwrap_or_else(|| null_scalar.clone()))
+        .collect::<Vec<ScalarValue>>();
+    debug_assert_eq!(scalar_values.len(), row_groups.len());
+    ScalarValue::iter_to_array(scalar_values).ok()
 }
 
-impl ReadFormat {
-    /// Creates a helper to read the primary key format.
-    pub fn new_primary_key(
-        metadata: RegionMetadataRef,
-        column_ids: impl Iterator<Item = ColumnId>,
-    ) -> Self {
-        ReadFormat::PrimaryKey(PrimaryKeyReadFormat::new(metadata, column_ids))
-    }
-
-    /// Creates a helper to read the flat format.
-    pub fn new_flat(
-        metadata: RegionMetadataRef,
-        column_ids: impl Iterator<Item = ColumnId>,
-        num_columns: Option<usize>,
-        file_path: &str,
-        skip_auto_convert: bool,
-    ) -> Result<Self> {
-        Ok(ReadFormat::Flat(FlatReadFormat::new(
-            metadata,
-            column_ids,
-            num_columns,
-            file_path,
-            skip_auto_convert,
-        )?))
-    }
-
-    /// Creates a new read format.
-    pub fn new(
-        region_metadata: RegionMetadataRef,
-        projection: Option<&[ColumnId]>,
-        flat_format: bool,
-        num_columns: Option<usize>,
-        file_path: &str,
-        skip_auto_convert: bool,
-    ) -> Result<ReadFormat> {
-        if flat_format {
-            if let Some(column_ids) = projection {
-                ReadFormat::new_flat(
-                    region_metadata,
-                    column_ids.iter().copied(),
-                    num_columns,
-                    file_path,
-                    skip_auto_convert,
-                )
-            } else {
-                // No projection, lists all column ids to read.
-                ReadFormat::new_flat(
-                    region_metadata.clone(),
-                    region_metadata
-                        .column_metadatas
-                        .iter()
-                        .map(|col| col.column_id),
-                    num_columns,
-                    file_path,
-                    skip_auto_convert,
-                )
-            }
-        } else if let Some(column_ids) = projection {
-            Ok(ReadFormat::new_primary_key(
-                region_metadata,
-                column_ids.iter().copied(),
-            ))
-        } else {
-            // No projection, lists all column ids to read.
-            Ok(ReadFormat::new_primary_key(
-                region_metadata.clone(),
-                region_metadata
-                    .column_metadatas
-                    .iter()
-                    .map(|col| col.column_id),
-            ))
-        }
-    }
-
-    pub(crate) fn as_primary_key(&self) -> Option<&PrimaryKeyReadFormat> {
-        match self {
-            ReadFormat::PrimaryKey(format) => Some(format),
-            _ => None,
-        }
-    }
-
-    pub(crate) fn as_flat(&self) -> Option<&FlatReadFormat> {
-        match self {
-            ReadFormat::Flat(format) => Some(format),
-            _ => None,
-        }
-    }
-
-    /// Gets the arrow schema of the SST file.
-    ///
-    /// This schema is computed from the region metadata but should be the same
-    /// as the arrow schema decoded from the file metadata.
-    pub(crate) fn arrow_schema(&self) -> &SchemaRef {
-        match self {
-            ReadFormat::PrimaryKey(format) => format.arrow_schema(),
-            ReadFormat::Flat(format) => format.arrow_schema(),
-        }
-    }
-
-    /// Gets the metadata of the SST.
-    pub(crate) fn metadata(&self) -> &RegionMetadataRef {
-        match self {
-            ReadFormat::PrimaryKey(format) => format.metadata(),
-            ReadFormat::Flat(format) => format.metadata(),
-        }
-    }
-
-    /// Gets sorted projection indices to read.
-    pub(crate) fn projection_indices(&self) -> &[usize] {
-        match self {
-            ReadFormat::PrimaryKey(format) => format.projection_indices(),
-            ReadFormat::Flat(format) => format.projection_indices(),
-        }
-    }
-
-    /// Returns min values of specific column in row groups.
-    pub fn min_values(
-        &self,
-        row_groups: &[impl Borrow<RowGroupMetaData>],
-        column_id: ColumnId,
-    ) -> StatValues {
-        match self {
-            ReadFormat::PrimaryKey(format) => format.min_values(row_groups, column_id),
-            ReadFormat::Flat(format) => format.min_values(row_groups, column_id),
-        }
-    }
-
-    /// Returns max values of specific column in row groups.
-    pub fn max_values(
-        &self,
-        row_groups: &[impl Borrow<RowGroupMetaData>],
-        column_id: ColumnId,
-    ) -> StatValues {
-        match self {
-            ReadFormat::PrimaryKey(format) => format.max_values(row_groups, column_id),
-            ReadFormat::Flat(format) => format.max_values(row_groups, column_id),
-        }
-    }
-
-    /// Returns null counts of specific column in row groups.
-    pub fn null_counts(
-        &self,
-        row_groups: &[impl Borrow<RowGroupMetaData>],
-        column_id: ColumnId,
-    ) -> StatValues {
-        match self {
-            ReadFormat::PrimaryKey(format) => format.null_counts(row_groups, column_id),
-            ReadFormat::Flat(format) => format.null_counts(row_groups, column_id),
-        }
-    }
-
-    /// Returns min/max values of specific columns.
-    /// Returns None if the column does not have statistics.
-    /// The column should not be encoded as a part of a primary key.
-    pub(crate) fn column_values(
-        row_groups: &[impl Borrow<RowGroupMetaData>],
-        column: &ColumnMetadata,
-        column_index: usize,
-        is_min: bool,
-    ) -> Option<ArrayRef> {
-        let null_scalar: ScalarValue = column
-            .column_schema
-            .data_type
-            .as_arrow_type()
-            .try_into()
-            .ok()?;
-        let scalar_values = row_groups
-            .iter()
-            .map(|meta| {
-                let stats = meta.borrow().column(column_index).statistics()?;
-                match stats {
-                    Statistics::Boolean(s) => Some(ScalarValue::Boolean(Some(if is_min {
-                        *s.min_opt()?
-                    } else {
-                        *s.max_opt()?
-                    }))),
-                    Statistics::Int32(s) => Some(ScalarValue::Int32(Some(if is_min {
-                        *s.min_opt()?
-                    } else {
-                        *s.max_opt()?
-                    }))),
-                    Statistics::Int64(s) => Some(ScalarValue::Int64(Some(if is_min {
-                        *s.min_opt()?
-                    } else {
-                        *s.max_opt()?
-                    }))),
-
-                    Statistics::Int96(_) => None,
-                    Statistics::Float(s) => Some(ScalarValue::Float32(Some(if is_min {
-                        *s.min_opt()?
-                    } else {
-                        *s.max_opt()?
-                    }))),
-                    Statistics::Double(s) => Some(ScalarValue::Float64(Some(if is_min {
-                        *s.min_opt()?
-                    } else {
-                        *s.max_opt()?
-                    }))),
-                    Statistics::ByteArray(s) => {
-                        let bytes = if is_min {
-                            s.min_bytes_opt()?
-                        } else {
-                            s.max_bytes_opt()?
-                        };
-                        let s = String::from_utf8(bytes.to_vec()).ok();
-                        Some(ScalarValue::Utf8(s))
-                    }
-
-                    Statistics::FixedLenByteArray(_) => None,
-                }
-            })
-            .map(|maybe_scalar| maybe_scalar.unwrap_or_else(|| null_scalar.clone()))
-            .collect::<Vec<ScalarValue>>();
-        debug_assert_eq!(scalar_values.len(), row_groups.len());
-        ScalarValue::iter_to_array(scalar_values).ok()
-    }
-
-    /// Returns null counts of specific columns.
-    /// The column should not be encoded as a part of a primary key.
-    pub(crate) fn column_null_counts(
-        row_groups: &[impl Borrow<RowGroupMetaData>],
-        column_index: usize,
-    ) -> Option<ArrayRef> {
-        let values = row_groups.iter().map(|meta| {
-            let col = meta.borrow().column(column_index);
-            let stat = col.statistics()?;
-            stat.null_count_opt()
-        });
-        Some(Arc::new(UInt64Array::from_iter(values)))
-    }
-
-    /// Sets the sequence number to override.
-    pub(crate) fn set_override_sequence(&mut self, sequence: Option<SequenceNumber>) {
-        match self {
-            ReadFormat::PrimaryKey(format) => format.set_override_sequence(sequence),
-            ReadFormat::Flat(format) => format.set_override_sequence(sequence),
-        }
-    }
-
-    /// Enables or disables eager decoding of primary key values into batches.
-    pub(crate) fn set_decode_primary_key_values(&mut self, decode: bool) {
-        if let ReadFormat::PrimaryKey(format) = self {
-            format.set_decode_primary_key_values(decode);
-        }
-    }
-
-    /// Creates a sequence array to override.
-    pub(crate) fn new_override_sequence_array(&self, length: usize) -> Option<ArrayRef> {
-        match self {
-            ReadFormat::PrimaryKey(format) => format.new_override_sequence_array(length),
-            ReadFormat::Flat(format) => format.new_override_sequence_array(length),
-        }
-    }
+/// Returns null counts of specific columns.
+/// The column should not be encoded as a part of a primary key.
+pub(crate) fn column_null_counts(
+    row_groups: &[impl Borrow<RowGroupMetaData>],
+    column_index: usize,
+) -> Option<ArrayRef> {
+    let values = row_groups.iter().map(|meta| {
+        let col = meta.borrow().column(column_index);
+        let stat = col.statistics()?;
+        stat.null_count_opt()
+    });
+    Some(Arc::new(UInt64Array::from_iter(values)))
 }
 
 /// Helper for reading the SST format.
@@ -416,22 +214,17 @@ pub struct PrimaryKeyReadFormat {
     /// In SST schema, fields are stored in the front of the schema.
     field_id_to_index: HashMap<ColumnId, usize>,
     /// Indices of columns to read from the SST. It contains all internal columns.
-    projection_indices: Vec<usize>,
+    parquet_read_cols: ParquetReadColumns,
     /// Field column id to their index in the projected schema (
     /// the schema of [Batch]).
     field_id_to_projected_index: HashMap<ColumnId, usize>,
-    /// Sequence number to override the sequence read from the SST.
-    override_sequence: Option<SequenceNumber>,
     /// Codec used to decode primary key values if eager decoding is enabled.
     primary_key_codec: Option<Arc<dyn PrimaryKeyCodec>>,
 }
 
 impl PrimaryKeyReadFormat {
     /// Creates a helper with existing `metadata` and `column_ids` to read.
-    pub fn new(
-        metadata: RegionMetadataRef,
-        column_ids: impl Iterator<Item = ColumnId>,
-    ) -> PrimaryKeyReadFormat {
+    pub fn new(metadata: RegionMetadataRef, read_cols: ReadColumns) -> PrimaryKeyReadFormat {
         let field_id_to_index: HashMap<_, _> = metadata
             .field_columns()
             .enumerate()
@@ -442,34 +235,19 @@ impl PrimaryKeyReadFormat {
         let format_projection = FormatProjection::compute_format_projection(
             &field_id_to_index,
             arrow_schema.fields.len(),
-            column_ids,
+            read_cols,
         );
 
         PrimaryKeyReadFormat {
             metadata,
             arrow_schema,
             field_id_to_index,
-            projection_indices: format_projection.projection_indices,
+            parquet_read_cols: format_projection.parquet_read_cols,
             field_id_to_projected_index: format_projection.column_id_to_projected_index,
-            override_sequence: None,
             primary_key_codec: None,
         }
     }
 
-    /// Sets the sequence number to override.
-    pub(crate) fn set_override_sequence(&mut self, sequence: Option<SequenceNumber>) {
-        self.override_sequence = sequence;
-    }
-
-    /// Enables or disables eager decoding of primary key values into batches.
-    pub(crate) fn set_decode_primary_key_values(&mut self, decode: bool) {
-        self.primary_key_codec = if decode {
-            Some(build_primary_key_codec(&self.metadata))
-        } else {
-            None
-        };
-    }
-
     /// Gets the arrow schema of the SST file.
     ///
     /// This schema is computed from the region metadata but should be the same
@@ -483,9 +261,8 @@ impl PrimaryKeyReadFormat {
         &self.metadata
     }
 
-    /// Gets sorted projection indices to read.
-    pub(crate) fn projection_indices(&self) -> &[usize] {
-        &self.projection_indices
+    pub(crate) fn parquet_read_columns(&self) -> &ParquetReadColumns {
+        &self.parquet_read_cols
     }
 
     /// Gets the field id to projected index.
@@ -493,12 +270,6 @@ impl PrimaryKeyReadFormat {
         &self.field_id_to_projected_index
     }
 
-    /// Creates a sequence array to override.
-    pub(crate) fn new_override_sequence_array(&self, length: usize) -> Option<ArrayRef> {
-        self.override_sequence
-            .map(|seq| Arc::new(UInt64Array::from_value(seq, length)) as ArrayRef)
-    }
-
     /// Convert a arrow record batch into `batches`.
     ///
     /// The length of `override_sequence_array` must be larger than the length of the record batch.
@@ -616,12 +387,12 @@ impl PrimaryKeyReadFormat {
             SemanticType::Field => {
                 // Safety: `field_id_to_index` is initialized by the semantic type.
                 let index = self.field_id_to_index.get(&column_id).unwrap();
-                let stats = ReadFormat::column_values(row_groups, column, *index, true);
+                let stats = column_values(row_groups, column, *index, true);
                 StatValues::from_stats_opt(stats)
             }
             SemanticType::Timestamp => {
                 let index = self.time_index_position();
-                let stats = ReadFormat::column_values(row_groups, column, index, true);
+                let stats = column_values(row_groups, column, index, true);
                 StatValues::from_stats_opt(stats)
             }
         }
@@ -642,12 +413,12 @@ impl PrimaryKeyReadFormat {
             SemanticType::Field => {
                 // Safety: `field_id_to_index` is initialized by the semantic type.
                 let index = self.field_id_to_index.get(&column_id).unwrap();
-                let stats = ReadFormat::column_values(row_groups, column, *index, false);
+                let stats = column_values(row_groups, column, *index, false);
                 StatValues::from_stats_opt(stats)
             }
             SemanticType::Timestamp => {
                 let index = self.time_index_position();
-                let stats = ReadFormat::column_values(row_groups, column, index, false);
+                let stats = column_values(row_groups, column, index, false);
                 StatValues::from_stats_opt(stats)
             }
         }
@@ -668,12 +439,12 @@ impl PrimaryKeyReadFormat {
             SemanticType::Field => {
                 // Safety: `field_id_to_index` is initialized by the semantic type.
                 let index = self.field_id_to_index.get(&column_id).unwrap();
-                let stats = ReadFormat::column_null_counts(row_groups, *index);
+                let stats = column_null_counts(row_groups, *index);
                 StatValues::from_stats_opt(stats)
             }
             SemanticType::Timestamp => {
                 let index = self.time_index_position();
-                let stats = ReadFormat::column_null_counts(row_groups, index);
+                let stats = column_null_counts(row_groups, index);
                 StatValues::from_stats_opt(stats)
             }
         }
@@ -807,8 +578,8 @@ impl PrimaryKeyReadFormat {
 
 /// Helper to compute the projection for the SST.
 pub(crate) struct FormatProjection {
-    /// Indices of columns to read from the SST. It contains all internal columns.
-    pub(crate) projection_indices: Vec<usize>,
+    /// The columns to read from the SST. It contains all internal columns.
+    pub(crate) parquet_read_cols: ParquetReadColumns,
     /// Column id to their index in the projected schema (
     /// the schema after projection).
     ///
@@ -823,50 +594,91 @@ impl FormatProjection {
     pub(crate) fn compute_format_projection(
         id_to_index: &HashMap<ColumnId, usize>,
         sst_column_num: usize,
-        column_ids: impl Iterator<Item = ColumnId>,
+        cols: ReadColumns,
     ) -> Self {
-        // Maps column id of a projected column to its index in SST.
-        // It also ignores columns not in the SST.
-        // [(column id, index in SST)]
-        let mut projected_schema: Vec<_> = column_ids
-            .filter_map(|column_id| {
+        let mut projected_columns: Vec<_> = cols
+            .cols
+            .into_iter()
+            .filter_map(|col| {
                 id_to_index
-                    .get(&column_id)
+                    .get(&col.column_id)
                     .copied()
-                    .map(|index| (column_id, index))
+                    .map(|index_of_sst| (col.column_id, index_of_sst, col.nested_paths))
             })
             .collect();
         // Sorts columns by their indices in the SST. SST uses a bitmap for projection.
-        // This ensures the schema of `projected_schema` is the same as the batch returned from the SST.
-        projected_schema.sort_unstable_by_key(|x| x.1);
-        // Dedups the entries to avoid the case that `column_ids` has duplicated columns.
-        projected_schema.dedup_by_key(|x| x.1);
-
-        // Collects all projected indices.
-        // It contains the positions of all columns we need to read.
-        let mut projection_indices: Vec<_> = projected_schema
-            .iter()
-            .map(|(_column_id, index)| *index)
-            // We need to add all fixed position columns.
-            .chain(sst_column_num - FIXED_POS_COLUMN_NUM..sst_column_num)
-            .collect();
-        projection_indices.sort_unstable();
-        // Removes duplications.
-        projection_indices.dedup();
+        // This ensures the schema of `projected_columns` is the same as the batch returned from the SST.
+        projected_columns.sort_unstable_by_key(|(_, index, _)| *index);
 
+        let mut parquet_read_cols: Vec<ParquetReadColumn> =
+            Vec::with_capacity(projected_columns.len() + FIXED_POS_COLUMN_NUM);
         // Creates a map from column id to the index of that column in the projected record batch.
-        let column_id_to_projected_index = projected_schema
-            .into_iter()
-            .map(|(column_id, _)| column_id)
-            .enumerate()
-            .map(|(index, column_id)| (column_id, index))
-            .collect();
+        let mut column_id_to_projected_index = HashMap::with_capacity(projected_columns.len());
+
+        for (col_id, index_of_sst, nested_paths) in projected_columns {
+            Self::merge_or_push_parquet_column(&mut parquet_read_cols, index_of_sst, nested_paths);
+
+            column_id_to_projected_index
+                .entry(col_id)
+                .or_insert_with(|| parquet_read_cols.len() - 1);
+        }
+
+        // In SST schema, fixed-position columns are always in the tail:
+        // `time index, __primary_key, __sequence, __op_type`.
+        Self::append_time_index_if_needed(&mut parquet_read_cols, sst_column_num);
+        Self::append_fixed_internal_columns(&mut parquet_read_cols, sst_column_num);
 
         Self {
-            projection_indices,
+            parquet_read_cols: ParquetReadColumns::from_deduped(parquet_read_cols),
             column_id_to_projected_index,
         }
     }
+
+    fn merge_or_push_parquet_column(
+        parquet_read_cols: &mut Vec<ParquetReadColumn>,
+        index_of_sst: usize,
+        nested_paths: Vec<Vec<String>>,
+    ) {
+        // `projected_columns` is sorted by parquet root index, so repeated reads
+        // for the same root column are always adjacent.
+        if let Some(last_col) = parquet_read_cols.last_mut()
+            && last_col.root_index() == index_of_sst
+        {
+            last_col.merge_nested_paths(nested_paths);
+            return;
+        }
+
+        let parquet_col = ParquetReadColumn::new(index_of_sst).with_nested_paths(nested_paths);
+        parquet_read_cols.push(parquet_col);
+    }
+
+    fn append_time_index_if_needed(
+        parquet_read_cols: &mut Vec<ParquetReadColumn>,
+        sst_column_num: usize,
+    ) {
+        let time_index = sst_column_num - FIXED_POS_COLUMN_NUM;
+        // Existing projected roots are already sorted by SST root index, and may
+        // already include the time index, so we compare against the last root to
+        // decide whether we still need to append `time index`.
+        let needs_time_index = parquet_read_cols
+            .last()
+            .map(|col| col.root_index() != time_index)
+            .unwrap_or(true);
+        if needs_time_index {
+            parquet_read_cols.push(ParquetReadColumn::new(time_index));
+        }
+    }
+
+    // Append internal columns in fixed order: `__primary_key`, `__sequence`,
+    // `__op_type`.
+    fn append_fixed_internal_columns(
+        parquet_read_cols: &mut Vec<ParquetReadColumn>,
+        sst_column_num: usize,
+    ) {
+        for index in sst_column_num - INTERNAL_COLUMN_NUM..sst_column_num {
+            parquet_read_cols.push(ParquetReadColumn::new(index));
+        }
+    }
 }
 
 /// Values of column statistics of the SST.
@@ -898,7 +710,9 @@ impl PrimaryKeyReadFormat {
     pub fn new_with_all_columns(metadata: RegionMetadataRef) -> PrimaryKeyReadFormat {
         Self::new(
             Arc::clone(&metadata),
-            metadata.column_metadatas.iter().map(|c| c.column_id),
+            ReadColumns::from_deduped_column_ids(
+                metadata.column_metadatas.iter().map(|c| c.column_id),
+            ),
         )
     }
 }
@@ -926,15 +740,6 @@ pub(crate) fn primary_key_offsets(pk_dict_array: &PrimaryKeyArray) -> Result<Vec
     Ok(offsets)
 }
 
-/// Creates a new array for specific `primary_key`.
-fn new_primary_key_array(primary_key: &[u8], num_rows: usize) -> ArrayRef {
-    let values = Arc::new(BinaryArray::from_iter_values([primary_key]));
-    let keys = UInt32Array::from_value(0, num_rows);
-
-    // Safety: The key index is valid.
-    Arc::new(DictionaryArray::new(keys, values))
-}
-
 /// Gets the min/max time index of the row group from the parquet meta.
 /// It assumes the parquet is created by the mito engine.
 pub(crate) fn parquet_row_group_time_range(
@@ -1017,7 +822,7 @@ mod tests {
 
     use api::v1::OpType;
     use datatypes::arrow::array::{
-        Int64Array, StringArray, TimestampMillisecondArray, UInt8Array, UInt64Array,
+        Int64Array, StringArray, TimestampMillisecondArray, UInt8Array, UInt32Array, UInt64Array,
     };
     use datatypes::arrow::datatypes::{DataType as ArrowDataType, Field, Schema, TimeUnit};
     use datatypes::prelude::ConcreteDataType;
@@ -1033,8 +838,10 @@ mod tests {
     use store_api::storage::consts::ReservedColumnId;
 
     use super::*;
-    use crate::sst::parquet::flat_format::{FlatWriteFormat, sequence_column_index};
-    use crate::sst::{FlatSchemaOptions, to_flat_sst_arrow_schema};
+    use crate::sst::parquet::flat_format::{
+        FlatReadFormat, FlatWriteFormat, sequence_column_index,
+    };
+    use crate::sst::{FlatSchemaOptions, to_flat_sst_arrow_schema, with_field_id};
 
     const TEST_SEQUENCE: u64 = 1;
     const TEST_OP_TYPE: u8 = OpType::Put as u8;
@@ -1085,23 +892,25 @@ mod tests {
 
     fn build_test_arrow_schema() -> SchemaRef {
         let fields = vec![
-            Field::new("field1", ArrowDataType::Int64, true),
-            Field::new("field0", ArrowDataType::Int64, true),
-            Field::new(
+            make_field("field1", ArrowDataType::Int64, true, Some(4)),
+            make_field("field0", ArrowDataType::Int64, true, Some(2)),
+            make_field(
                 "ts",
                 ArrowDataType::Timestamp(TimeUnit::Millisecond, None),
                 false,
+                Some(5),
             ),
-            Field::new(
+            make_field(
                 "__primary_key",
                 ArrowDataType::Dictionary(
                     Box::new(ArrowDataType::UInt32),
                     Box::new(ArrowDataType::Binary),
                 ),
                 false,
+                None,
             ),
-            Field::new("__sequence", ArrowDataType::UInt64, false),
-            Field::new("__op_type", ArrowDataType::UInt8, false),
+            make_field("__sequence", ArrowDataType::UInt64, false, None),
+            make_field("__op_type", ArrowDataType::UInt8, false, None),
         ];
         Arc::new(Schema::new(fields))
     }
@@ -1145,13 +954,6 @@ mod tests {
         assert_eq!(&build_test_arrow_schema(), write_format.arrow_schema());
     }
 
-    #[test]
-    fn test_new_primary_key_array() {
-        let array = new_primary_key_array(b"test", 3);
-        let expect = build_test_pk_array(&[(b"test".to_vec(), 3)]) as ArrayRef;
-        assert_eq!(&expect, &array);
-    }
-
     fn build_test_pk_array(pk_row_nums: &[(Vec<u8>, usize)]) -> Arc<PrimaryKeyArray> {
         let values = Arc::new(BinaryArray::from_iter_values(
             pk_row_nums.iter().map(|v| &v.0),
@@ -1164,64 +966,37 @@ mod tests {
         Arc::new(DictionaryArray::new(keys, values))
     }
 
-    #[test]
-    fn test_convert_batch() {
-        let metadata = build_test_region_metadata();
-        let write_format = PrimaryKeyWriteFormat::new(metadata);
-
-        let num_rows = 4;
-        let batch = new_batch(b"test", 1, 2, num_rows);
-        let columns: Vec<ArrayRef> = vec![
-            Arc::new(Int64Array::from(vec![2; num_rows])), // field1
-            Arc::new(Int64Array::from(vec![3; num_rows])), // field0
-            Arc::new(TimestampMillisecondArray::from(vec![1, 2, 3, 4])), // ts
-            build_test_pk_array(&[(b"test".to_vec(), num_rows)]), // primary key
-            Arc::new(UInt64Array::from(vec![TEST_SEQUENCE; num_rows])), // sequence
-            Arc::new(UInt8Array::from(vec![TEST_OP_TYPE; num_rows])), // op type
-        ];
-        let expect_record = RecordBatch::try_new(build_test_arrow_schema(), columns).unwrap();
-
-        let actual = write_format.convert_batch(&batch).unwrap();
-        assert_eq!(expect_record, actual);
-    }
-
-    #[test]
-    fn test_convert_batch_with_override_sequence() {
-        let metadata = build_test_region_metadata();
-        let write_format =
-            PrimaryKeyWriteFormat::new(metadata).with_override_sequence(Some(415411));
-
-        let num_rows = 4;
-        let batch = new_batch(b"test", 1, 2, num_rows);
-        let columns: Vec<ArrayRef> = vec![
-            Arc::new(Int64Array::from(vec![2; num_rows])), // field1
-            Arc::new(Int64Array::from(vec![3; num_rows])), // field0
-            Arc::new(TimestampMillisecondArray::from(vec![1, 2, 3, 4])), // ts
-            build_test_pk_array(&[(b"test".to_vec(), num_rows)]), // primary key
-            Arc::new(UInt64Array::from(vec![415411; num_rows])), // sequence
-            Arc::new(UInt8Array::from(vec![TEST_OP_TYPE; num_rows])), // op type
-        ];
-        let expect_record = RecordBatch::try_new(build_test_arrow_schema(), columns).unwrap();
-
-        let actual = write_format.convert_batch(&batch).unwrap();
-        assert_eq!(expect_record, actual);
-    }
-
     #[test]
     fn test_projection_indices() {
         let metadata = build_test_region_metadata();
         // Only read tag1
-        let read_format = ReadFormat::new_primary_key(metadata.clone(), [3].iter().copied());
-        assert_eq!(&[2, 3, 4, 5], read_format.projection_indices());
+        let read_format =
+            PrimaryKeyReadFormat::new(metadata.clone(), ReadColumns::from_deduped_column_ids([3]));
+        assert_eq!(
+            &[2, 3, 4, 5],
+            read_format.parquet_read_columns().root_indices()
+        );
         // Only read field1
-        let read_format = ReadFormat::new_primary_key(metadata.clone(), [4].iter().copied());
-        assert_eq!(&[0, 2, 3, 4, 5], read_format.projection_indices());
+        let read_format =
+            PrimaryKeyReadFormat::new(metadata.clone(), ReadColumns::from_deduped_column_ids([4]));
+        assert_eq!(
+            &[0, 2, 3, 4, 5],
+            read_format.parquet_read_columns().root_indices()
+        );
         // Only read ts
-        let read_format = ReadFormat::new_primary_key(metadata.clone(), [5].iter().copied());
-        assert_eq!(&[2, 3, 4, 5], read_format.projection_indices());
+        let read_format =
+            PrimaryKeyReadFormat::new(metadata.clone(), ReadColumns::from_deduped_column_ids([5]));
+        assert_eq!(
+            &[2, 3, 4, 5],
+            read_format.parquet_read_columns().root_indices()
+        );
         // Read field0, tag0, ts
-        let read_format = ReadFormat::new_primary_key(metadata, [2, 1, 5].iter().copied());
-        assert_eq!(&[1, 2, 3, 4, 5], read_format.projection_indices());
+        let read_format =
+            PrimaryKeyReadFormat::new(metadata, ReadColumns::from_deduped_column_ids([2, 1, 5]));
+        assert_eq!(
+            &[1, 2, 3, 4, 5],
+            read_format.parquet_read_columns().root_indices()
+        );
     }
 
     #[test]
@@ -1267,7 +1042,8 @@ mod tests {
             .iter()
             .map(|col| col.column_id)
             .collect();
-        let read_format = PrimaryKeyReadFormat::new(metadata, column_ids.iter().copied());
+        let read_format =
+            PrimaryKeyReadFormat::new(metadata, ReadColumns::from_deduped_column_ids(column_ids));
         assert_eq!(arrow_schema, *read_format.arrow_schema());
 
         let record_batch = RecordBatch::new_empty(arrow_schema);
@@ -1286,7 +1062,8 @@ mod tests {
             .iter()
             .map(|col| col.column_id)
             .collect();
-        let read_format = PrimaryKeyReadFormat::new(metadata, column_ids.iter().copied());
+        let read_format =
+            PrimaryKeyReadFormat::new(metadata, ReadColumns::from_deduped_column_ids(column_ids));
 
         let columns: Vec<ArrayRef> = vec![
             Arc::new(Int64Array::from(vec![1, 1, 10, 10])), // field1
@@ -1312,13 +1089,12 @@ mod tests {
     #[test]
     fn test_convert_record_batch_with_override_sequence() {
         let metadata = build_test_region_metadata();
-        let column_ids: Vec<_> = metadata
-            .column_metadatas
-            .iter()
-            .map(|col| col.column_id)
-            .collect();
-        let read_format =
-            ReadFormat::new(metadata, Some(&column_ids), false, None, "test", false).unwrap();
+        let read_format = PrimaryKeyReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids(
+                metadata.column_metadatas.iter().map(|c| c.column_id),
+            ),
+        );
 
         let columns: Vec<ArrayRef> = vec![
             Arc::new(Int64Array::from(vec![1, 1, 10, 10])), // field1
@@ -1338,8 +1114,6 @@ mod tests {
 
         let mut batches = VecDeque::new();
         read_format
-            .as_primary_key()
-            .unwrap()
             .convert_record_batch(&record_batch, Some(&override_sequence_array), &mut batches)
             .unwrap();
 
@@ -1353,6 +1127,14 @@ mod tests {
         );
     }
 
+    fn make_field(name: &str, dt: ArrowDataType, nullable: bool, field_id: Option<u32>) -> Field {
+        let mut field = Field::new(name, dt, nullable);
+        if let Some(id) = field_id {
+            field = with_field_id(field, id);
+        }
+        field
+    }
+
     fn build_test_flat_sst_schema() -> SchemaRef {
         let fields = vec![
             Field::new("tag0", ArrowDataType::Int64, true), // primary key columns first
@@ -1378,11 +1160,37 @@ mod tests {
         Arc::new(Schema::new(fields))
     }
 
+    fn build_test_flat_sst_schema_with_field_ids() -> SchemaRef {
+        let ids = [
+            Some(1u32),
+            Some(3),
+            Some(4),
+            Some(2),
+            Some(5),
+            None,
+            None,
+            None,
+        ];
+        let fields: Vec<_> = build_test_flat_sst_schema()
+            .fields()
+            .iter()
+            .zip(ids)
+            .map(|(f, id)| match id {
+                Some(id) => Arc::new(with_field_id((**f).clone(), id)) as _,
+                None => f.clone(),
+            })
+            .collect();
+        Arc::new(Schema::new(fields))
+    }
+
     #[test]
     fn test_flat_to_sst_arrow_schema() {
         let metadata = build_test_region_metadata();
         let format = FlatWriteFormat::new(metadata, &FlatSchemaOptions::default());
-        assert_eq!(&build_test_flat_sst_schema(), format.arrow_schema());
+        assert_eq!(
+            &build_test_flat_sst_schema_with_field_ids(),
+            format.arrow_schema()
+        );
     }
 
     fn input_columns_for_flat_batch(num_rows: usize) -> Vec<ArrayRef> {
@@ -1405,8 +1213,11 @@ mod tests {
 
         let num_rows = 4;
         let columns: Vec<ArrayRef> = input_columns_for_flat_batch(num_rows);
-        let batch = RecordBatch::try_new(build_test_flat_sst_schema(), columns.clone()).unwrap();
-        let expect_record = RecordBatch::try_new(build_test_flat_sst_schema(), columns).unwrap();
+        let batch =
+            RecordBatch::try_new(build_test_flat_sst_schema_with_field_ids(), columns.clone())
+                .unwrap();
+        let expect_record =
+            RecordBatch::try_new(build_test_flat_sst_schema_with_field_ids(), columns).unwrap();
 
         let actual = format.convert_batch(&batch).unwrap();
         assert_eq!(expect_record, actual);
@@ -1420,7 +1231,8 @@ mod tests {
 
         let num_rows = 4;
         let columns: Vec<ArrayRef> = input_columns_for_flat_batch(num_rows);
-        let batch = RecordBatch::try_new(build_test_flat_sst_schema(), columns).unwrap();
+        let batch =
+            RecordBatch::try_new(build_test_flat_sst_schema_with_field_ids(), columns).unwrap();
 
         let expected_columns: Vec<ArrayRef> = vec![
             Arc::new(Int64Array::from(vec![1; num_rows])), // tag0
@@ -1432,8 +1244,11 @@ mod tests {
             Arc::new(UInt64Array::from(vec![415411; num_rows])), // overridden sequence
             Arc::new(UInt8Array::from(vec![TEST_OP_TYPE; num_rows])), // op type
         ];
-        let expected_record =
-            RecordBatch::try_new(build_test_flat_sst_schema(), expected_columns).unwrap();
+        let expected_record = RecordBatch::try_new(
+            build_test_flat_sst_schema_with_field_ids(),
+            expected_columns,
+        )
+        .unwrap();
 
         let actual = format.convert_batch(&batch).unwrap();
         assert_eq!(expected_record, actual);
@@ -1446,27 +1261,60 @@ mod tests {
         // The projection includes all "fixed position" columns: ts(4), __primary_key(5), __sequence(6), __op_type(7)
 
         // Only read tag1 (column_id=3, index=1) + fixed columns
-        let read_format =
-            ReadFormat::new_flat(metadata.clone(), [3].iter().copied(), None, "test", false)
-                .unwrap();
-        assert_eq!(&[1, 4, 5, 6, 7], read_format.projection_indices());
+        let read_format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids([3]),
+            None,
+            "test",
+            false,
+        )
+        .unwrap();
+        assert_eq!(
+            &[1, 4, 5, 6, 7],
+            read_format.parquet_read_columns().root_indices()
+        );
 
         // Only read field1 (column_id=4, index=2) + fixed columns
-        let read_format =
-            ReadFormat::new_flat(metadata.clone(), [4].iter().copied(), None, "test", false)
-                .unwrap();
-        assert_eq!(&[2, 4, 5, 6, 7], read_format.projection_indices());
+        let read_format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids([4]),
+            None,
+            "test",
+            false,
+        )
+        .unwrap();
+        assert_eq!(
+            &[2, 4, 5, 6, 7],
+            read_format.parquet_read_columns().root_indices()
+        );
 
         // Only read ts (column_id=5, index=4) + fixed columns (ts is already included in fixed)
-        let read_format =
-            ReadFormat::new_flat(metadata.clone(), [5].iter().copied(), None, "test", false)
-                .unwrap();
-        assert_eq!(&[4, 5, 6, 7], read_format.projection_indices());
+        let read_format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids([5]),
+            None,
+            "test",
+            false,
+        )
+        .unwrap();
+        assert_eq!(
+            &[4, 5, 6, 7],
+            read_format.parquet_read_columns().root_indices()
+        );
 
         // Read field0(column_id=2, index=3), tag0(column_id=1, index=0), ts(column_id=5, index=4) + fixed columns
-        let read_format =
-            ReadFormat::new_flat(metadata, [2, 1, 5].iter().copied(), None, "test", false).unwrap();
-        assert_eq!(&[0, 3, 4, 5, 6, 7], read_format.projection_indices());
+        let read_format = FlatReadFormat::new(
+            metadata,
+            ReadColumns::from_deduped_column_ids([2, 1, 5]),
+            None,
+            "test",
+            false,
+        )
+        .unwrap();
+        assert_eq!(
+            &[0, 3, 4, 5, 6, 7],
+            read_format.parquet_read_columns().root_indices()
+        );
     }
 
     #[test]
@@ -1474,7 +1322,7 @@ mod tests {
         let metadata = build_test_region_metadata();
         let mut format = FlatReadFormat::new(
             metadata,
-            std::iter::once(1), // Just read tag0
+            ReadColumns::from_deduped_column_ids(std::iter::once(1)), // Just read tag0
             Some(8),
             "test",
             false,
@@ -1691,7 +1539,7 @@ mod tests {
             .collect();
         let format = FlatReadFormat::new(
             metadata.clone(),
-            column_ids.into_iter(),
+            ReadColumns::from_deduped_column_ids(column_ids),
             Some(6),
             "test",
             false,
@@ -1719,26 +1567,7 @@ mod tests {
         ];
 
         // Create schema for old format (without primary key columns)
-        let old_format_fields = vec![
-            Field::new("field1", ArrowDataType::Int64, true),
-            Field::new("field0", ArrowDataType::Int64, true),
-            Field::new(
-                "ts",
-                ArrowDataType::Timestamp(TimeUnit::Millisecond, None),
-                false,
-            ),
-            Field::new(
-                "__primary_key",
-                ArrowDataType::Dictionary(
-                    Box::new(ArrowDataType::UInt32),
-                    Box::new(ArrowDataType::Binary),
-                ),
-                false,
-            ),
-            Field::new("__sequence", ArrowDataType::UInt64, false),
-            Field::new("__op_type", ArrowDataType::UInt8, false),
-        ];
-        let old_schema = Arc::new(Schema::new(old_format_fields));
+        let old_schema = build_test_arrow_schema();
         let record_batch = RecordBatch::try_new(old_schema, columns).unwrap();
 
         // Test conversion with dense encoding
@@ -1755,8 +1584,11 @@ mod tests {
             Arc::new(UInt64Array::from(vec![original_sequence; num_rows])), // sequence
             Arc::new(UInt8Array::from(vec![TEST_OP_TYPE; num_rows])), // op type
         ];
-        let expected_record_batch =
-            RecordBatch::try_new(build_test_flat_sst_schema(), expected_columns).unwrap();
+        let expected_record_batch = RecordBatch::try_new(
+            build_test_flat_sst_schema_with_field_ids(),
+            expected_columns,
+        )
+        .unwrap();
 
         // Compare the actual result with the expected record batch
         assert_eq!(expected_record_batch, result);
@@ -1773,7 +1605,7 @@ mod tests {
             .collect();
         let format = FlatReadFormat::new(
             metadata.clone(),
-            column_ids.clone().into_iter(),
+            ReadColumns::from_deduped_column_ids(column_ids.clone()),
             None,
             "test",
             false,
@@ -1807,26 +1639,7 @@ mod tests {
         ];
 
         // Create schema for old format (without primary key columns)
-        let old_format_fields = vec![
-            Field::new("field1", ArrowDataType::Int64, true),
-            Field::new("field0", ArrowDataType::Int64, true),
-            Field::new(
-                "ts",
-                ArrowDataType::Timestamp(TimeUnit::Millisecond, None),
-                false,
-            ),
-            Field::new(
-                "__primary_key",
-                ArrowDataType::Dictionary(
-                    Box::new(ArrowDataType::UInt32),
-                    Box::new(ArrowDataType::Binary),
-                ),
-                false,
-            ),
-            Field::new("__sequence", ArrowDataType::UInt64, false),
-            Field::new("__op_type", ArrowDataType::UInt8, false),
-        ];
-        let old_schema = Arc::new(Schema::new(old_format_fields));
+        let old_schema = build_test_arrow_schema();
         let record_batch = RecordBatch::try_new(old_schema, columns).unwrap();
 
         // Test conversion with sparse encoding
@@ -1860,11 +1673,112 @@ mod tests {
         // Compare the actual result with the expected record batch
         assert_eq!(expected_record_batch, result);
 
-        let format =
-            FlatReadFormat::new(metadata.clone(), column_ids.into_iter(), None, "test", true)
-                .unwrap();
+        let format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids(column_ids),
+            None,
+            "test",
+            true,
+        )
+        .unwrap();
         // Test conversion with sparse encoding and skip convert.
         let result = format.convert_batch(record_batch.clone(), None).unwrap();
         assert_eq!(record_batch, result);
     }
+
+    #[test]
+    fn test_convert_flat_batch() {
+        let metadata = build_test_region_metadata();
+        let write_format = PrimaryKeyWriteFormat::new(metadata);
+
+        let num_rows = 4;
+        // Build a flat record batch: tag0, tag1, field1, field0, ts, __primary_key, __sequence, __op_type
+        let flat_columns: Vec<ArrayRef> = input_columns_for_flat_batch(num_rows);
+        let flat_batch = RecordBatch::try_new(build_test_flat_sst_schema(), flat_columns).unwrap();
+
+        // num_fields = 2 (field1, field0)
+        let result = write_format.convert_flat_batch(&flat_batch, 2).unwrap();
+
+        // Expected: tag columns stripped, only field1, field0, ts, __primary_key, __sequence, __op_type
+        let expected_columns: Vec<ArrayRef> = vec![
+            Arc::new(Int64Array::from(vec![2; num_rows])), // field1
+            Arc::new(Int64Array::from(vec![3; num_rows])), // field0
+            Arc::new(TimestampMillisecondArray::from(vec![1, 2, 3, 4])), // ts
+            build_test_pk_array(&[(b"test".to_vec(), num_rows)]), // __primary_key
+            Arc::new(UInt64Array::from(vec![TEST_SEQUENCE; num_rows])), // __sequence
+            Arc::new(UInt8Array::from(vec![TEST_OP_TYPE; num_rows])), // __op_type
+        ];
+        let expected = RecordBatch::try_new(build_test_arrow_schema(), expected_columns).unwrap();
+
+        assert_eq!(expected, result);
+    }
+
+    #[test]
+    fn test_convert_flat_batch_with_override_sequence() {
+        let metadata = build_test_region_metadata();
+        let write_format = PrimaryKeyWriteFormat::new(metadata).with_override_sequence(Some(999));
+
+        let num_rows = 4;
+        let flat_columns: Vec<ArrayRef> = input_columns_for_flat_batch(num_rows);
+        let flat_batch = RecordBatch::try_new(build_test_flat_sst_schema(), flat_columns).unwrap();
+
+        let result = write_format.convert_flat_batch(&flat_batch, 2).unwrap();
+
+        let expected_columns: Vec<ArrayRef> = vec![
+            Arc::new(Int64Array::from(vec![2; num_rows])), // field1
+            Arc::new(Int64Array::from(vec![3; num_rows])), // field0
+            Arc::new(TimestampMillisecondArray::from(vec![1, 2, 3, 4])), // ts
+            build_test_pk_array(&[(b"test".to_vec(), num_rows)]), // __primary_key
+            Arc::new(UInt64Array::from(vec![999; num_rows])), // overridden __sequence
+            Arc::new(UInt8Array::from(vec![TEST_OP_TYPE; num_rows])), // __op_type
+        ];
+        let expected = RecordBatch::try_new(build_test_arrow_schema(), expected_columns).unwrap();
+
+        assert_eq!(expected, result);
+    }
+
+    #[test]
+    fn test_convert_flat_batch_no_tags() {
+        // Test with a region that has no primary key columns (no tags to strip).
+        let mut builder = RegionMetadataBuilder::new(RegionId::new(1, 1));
+        builder
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "field0",
+                    ConcreteDataType::int64_datatype(),
+                    true,
+                ),
+                semantic_type: SemanticType::Field,
+                column_id: 1,
+            })
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "ts",
+                    ConcreteDataType::timestamp_millisecond_datatype(),
+                    false,
+                ),
+                semantic_type: SemanticType::Timestamp,
+                column_id: 2,
+            });
+        let metadata = Arc::new(builder.build().unwrap());
+        let write_format = PrimaryKeyWriteFormat::new(metadata);
+
+        let num_rows = 3;
+        // No tag columns, so flat batch is: field0, ts, __primary_key, __sequence, __op_type
+        let sst_schema = write_format.arrow_schema().clone();
+        let columns: Vec<ArrayRef> = vec![
+            Arc::new(Int64Array::from(vec![10; num_rows])), // field0
+            Arc::new(TimestampMillisecondArray::from(vec![1, 2, 3])), // ts
+            build_test_pk_array(&[(b"".to_vec(), num_rows)]), // __primary_key
+            Arc::new(UInt64Array::from(vec![TEST_SEQUENCE; num_rows])), // __sequence
+            Arc::new(UInt8Array::from(vec![TEST_OP_TYPE; num_rows])), // __op_type
+        ];
+        let flat_batch = RecordBatch::try_new(sst_schema.clone(), columns.clone()).unwrap();
+
+        // num_fields = 1, num_tag_columns = 5 - 1 - 4 = 0, so nothing is stripped
+        let result = write_format.convert_flat_batch(&flat_batch, 1).unwrap();
+        let expected = RecordBatch::try_new(sst_schema, columns).unwrap();
+
+        assert_eq!(expected, result);
+    }
 }
diff --git a/src/mito2/src/sst/parquet/metadata.rs b/src/mito2/src/sst/parquet/metadata.rs
index d13f318fb5..442ecab70c 100644
--- a/src/mito2/src/sst/parquet/metadata.rs
+++ b/src/mito2/src/sst/parquet/metadata.rs
@@ -22,7 +22,10 @@ use object_store::ObjectStore;
 use parquet::arrow::async_reader::MetadataFetch;
 use parquet::errors::{ParquetError, Result as ParquetResult};
 use parquet::file::metadata::{PageIndexPolicy, ParquetMetaData, ParquetMetaDataReader};
+use parquet::file::statistics::Statistics;
 use snafu::{IntoError as _, ResultExt};
+use store_api::metadata::RegionMetadata;
+use store_api::storage::consts::PRIMARY_KEY_COLUMN_NAME;
 
 use crate::error::{self, Result};
 use crate::sst::parquet::reader::MetadataCacheMetrics;
@@ -115,6 +118,44 @@ fn unbox_external_error(e: ParquetError) -> StdResult<object_store::Error, Parqu
     }
 }
 
+pub(crate) fn extract_primary_key_range(
+    parquet_meta: &ParquetMetaData,
+    region_metadata: &RegionMetadata,
+) -> Option<(Bytes, Bytes)> {
+    if region_metadata.primary_key.is_empty() {
+        return None;
+    }
+
+    let pk_column_idx = parquet_meta
+        .file_metadata()
+        .schema_descr()
+        .columns()
+        .iter()
+        .position(|column| column.name() == PRIMARY_KEY_COLUMN_NAME)?;
+
+    let mut min: Option<Bytes> = None;
+    let mut max: Option<Bytes> = None;
+
+    for row_group in parquet_meta.row_groups() {
+        let Statistics::ByteArray(stats) = row_group.column(pk_column_idx).statistics()? else {
+            return None;
+        };
+
+        let row_group_min = Bytes::copy_from_slice(stats.min_bytes_opt()?);
+        let row_group_max = Bytes::copy_from_slice(stats.max_bytes_opt()?);
+        min = Some(match min {
+            Some(current) => current.min(row_group_min),
+            None => row_group_min,
+        });
+        max = Some(match max {
+            Some(current) => current.max(row_group_max),
+            None => row_group_max,
+        });
+    }
+
+    min.zip(max)
+}
+
 struct ObjectStoreFetch<'a> {
     object_store: &'a ObjectStore,
     file_path: &'a str,
@@ -139,3 +180,120 @@ impl MetadataFetch for ObjectStoreFetch<'_> {
         .boxed()
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use datatypes::arrow::array::{
+        ArrayRef, BinaryArray, DictionaryArray, Int64Array, UInt32Array,
+    };
+    use datatypes::arrow::datatypes::{DataType as ArrowDataType, Field, Schema};
+    use datatypes::arrow::record_batch::RecordBatch;
+    use parquet::arrow::ArrowWriter;
+    use parquet::arrow::arrow_reader::ParquetRecordBatchReaderBuilder;
+    use parquet::file::metadata::{KeyValue, ParquetMetaData};
+    use parquet::file::properties::{EnabledStatistics, WriterProperties};
+
+    use super::*;
+    use crate::sst::parquet::PARQUET_METADATA_KEY;
+    use crate::test_util::sst_util::sst_region_metadata;
+
+    fn build_test_metadata(
+        include_primary_key: bool,
+        primary_keys: &[&[u8]],
+        row_group_sizes: &[usize],
+        stats_enabled: EnabledStatistics,
+    ) -> ParquetMetaData {
+        let total_rows = row_group_sizes.iter().sum::<usize>();
+        let mut fields = vec![Field::new("field", ArrowDataType::Int64, true)];
+        let mut columns: Vec<ArrayRef> =
+            vec![Arc::new(Int64Array::from_iter_values(0..total_rows as i64))];
+        if include_primary_key {
+            assert_eq!(total_rows, primary_keys.len());
+            fields.push(Field::new(
+                "__primary_key",
+                ArrowDataType::Dictionary(
+                    Box::new(ArrowDataType::UInt32),
+                    Box::new(ArrowDataType::Binary),
+                ),
+                false,
+            ));
+            let values = Arc::new(BinaryArray::from_iter_values(primary_keys.iter().copied()));
+            let keys = UInt32Array::from_iter_values(0..primary_keys.len() as u32);
+            columns.push(Arc::new(DictionaryArray::new(keys, values)));
+        }
+
+        let schema = Arc::new(Schema::new(fields));
+        let region_metadata = Arc::new(sst_region_metadata());
+        let key_value = KeyValue::new(
+            PARQUET_METADATA_KEY.to_string(),
+            region_metadata.to_json().unwrap(),
+        );
+        let props = WriterProperties::builder()
+            .set_key_value_metadata(Some(vec![key_value]))
+            .set_statistics_enabled(stats_enabled)
+            .build();
+
+        let mut parquet_bytes = Vec::new();
+        let mut writer =
+            ArrowWriter::try_new(&mut parquet_bytes, schema.clone(), Some(props)).unwrap();
+        let mut offset = 0;
+        for row_group_size in row_group_sizes {
+            let batch = RecordBatch::try_new(
+                schema.clone(),
+                columns
+                    .iter()
+                    .map(|column| column.slice(offset, *row_group_size))
+                    .collect(),
+            )
+            .unwrap();
+            writer.write(&batch).unwrap();
+            offset += row_group_size;
+        }
+        writer.close().unwrap();
+
+        ParquetRecordBatchReaderBuilder::try_new(Bytes::from(parquet_bytes))
+            .unwrap()
+            .metadata()
+            .as_ref()
+            .clone()
+    }
+
+    #[test]
+    fn test_extract_primary_key_range_returns_none_when_column_absent() {
+        let metadata = build_test_metadata(false, &[], &[1], EnabledStatistics::Page);
+        let region_metadata = sst_region_metadata();
+
+        assert_eq!(None, extract_primary_key_range(&metadata, &region_metadata));
+    }
+
+    #[test]
+    fn test_extract_primary_key_range_folds_row_group_stats() {
+        let metadata = build_test_metadata(
+            true,
+            &[b"bbb", b"ccc", b"aaa", b"zzz"],
+            &[2, 2],
+            EnabledStatistics::Page,
+        );
+        let region_metadata = sst_region_metadata();
+
+        assert_eq!(
+            Some((Bytes::from_static(b"aaa"), Bytes::from_static(b"zzz"))),
+            extract_primary_key_range(&metadata, &region_metadata)
+        );
+    }
+
+    #[test]
+    fn test_extract_primary_key_range_returns_none_when_any_rg_stats_missing() {
+        let metadata = build_test_metadata(
+            true,
+            &[b"bbb", b"ccc", b"aaa", b"zzz"],
+            &[2, 2],
+            EnabledStatistics::None,
+        );
+        let region_metadata = sst_region_metadata();
+
+        assert_eq!(None, extract_primary_key_range(&metadata, &region_metadata));
+    }
+}
diff --git a/src/mito2/src/sst/parquet/prefilter.rs b/src/mito2/src/sst/parquet/prefilter.rs
new file mode 100644
index 0000000000..241a56e8ed
--- /dev/null
+++ b/src/mito2/src/sst/parquet/prefilter.rs
@@ -0,0 +1,1220 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Prefilter framework for parquet reader.
+//!
+//! Prefilter optimization reduces I/O by reading only a subset of columns first
+//! (the prefilter phase), applying filters to compute a refined row selection,
+//! then reading the remaining columns with the refined selection.
+
+use std::collections::HashSet;
+use std::ops::{BitAnd, Range};
+use std::sync::Arc;
+
+use api::v1::SemanticType;
+use common_recordbatch::filter::SimpleFilterEvaluator;
+use datatypes::arrow::array::{Array, BinaryArray, BooleanArray, BooleanBufferBuilder};
+use datatypes::arrow::buffer::BooleanBuffer;
+use datatypes::arrow::record_batch::RecordBatch;
+use futures::StreamExt;
+use mito_codec::row_converter::{PrimaryKeyCodec, PrimaryKeyFilter};
+use parquet::arrow::ProjectionMask;
+use parquet::arrow::arrow_reader::RowSelection;
+use parquet::schema::types::SchemaDescriptor;
+use snafu::{OptionExt, ResultExt};
+use store_api::metadata::{RegionMetadata, RegionMetadataRef};
+use store_api::storage::consts::PRIMARY_KEY_COLUMN_NAME;
+use table::predicate::Predicate;
+
+use crate::error::{
+    ComputeArrowSnafu, DecodeSnafu, EvalPartitionFilterSnafu, NewRecordBatchSnafu,
+    ReadParquetSnafu, RecordBatchSnafu, Result, UnexpectedSnafu,
+};
+use crate::sst::parquet::file_range::PreFilterMode;
+use crate::sst::parquet::flat_format::FlatReadFormat;
+use crate::sst::parquet::format::PrimaryKeyArray;
+use crate::sst::parquet::reader::{
+    MaybeFilter, PhysicalFilterContext, RowGroupBuildContext, RowGroupReaderBuilder,
+    SimpleFilterContext,
+};
+
+pub(crate) fn matching_row_ranges_by_primary_key(
+    input: &RecordBatch,
+    pk_column_index: usize,
+    pk_filter: &mut dyn PrimaryKeyFilter,
+) -> Result<Vec<Range<usize>>> {
+    let pk_dict_array = input
+        .column(pk_column_index)
+        .as_any()
+        .downcast_ref::<PrimaryKeyArray>()
+        .context(UnexpectedSnafu {
+            reason: "Primary key column is not a dictionary array",
+        })?;
+    let pk_values = pk_dict_array
+        .values()
+        .as_any()
+        .downcast_ref::<BinaryArray>()
+        .context(UnexpectedSnafu {
+            reason: "Primary key values are not binary array",
+        })?;
+    let keys = pk_dict_array.keys();
+    let key_values = keys.values();
+
+    if key_values.is_empty() {
+        return Ok(std::iter::once(0..input.num_rows()).collect());
+    }
+
+    let mut matched_row_ranges: Vec<Range<usize>> = Vec::new();
+    let mut start = 0;
+    while start < key_values.len() {
+        let key = key_values[start];
+        let mut end = start + 1;
+        while end < key_values.len() && key_values[end] == key {
+            end += 1;
+        }
+
+        if pk_filter
+            .matches(pk_values.value(key as usize))
+            .context(DecodeSnafu)?
+        {
+            if let Some(last) = matched_row_ranges.last_mut()
+                && last.end == start
+            {
+                last.end = end;
+            } else {
+                matched_row_ranges.push(start..end);
+            }
+        }
+
+        start = end;
+    }
+
+    Ok(matched_row_ranges)
+}
+
+/// Filters a flat-format record batch by primary key, returning only rows whose
+/// primary key matches the filter. Returns `None` if all rows are filtered out.
+pub(crate) fn prefilter_flat_batch_by_primary_key(
+    input: RecordBatch,
+    pk_column_index: usize,
+    pk_filter: &mut dyn PrimaryKeyFilter,
+) -> Result<Option<RecordBatch>> {
+    if input.num_rows() == 0 {
+        return Ok(Some(input));
+    }
+
+    let matched_row_ranges =
+        matching_row_ranges_by_primary_key(&input, pk_column_index, pk_filter)?;
+    if matched_row_ranges.is_empty() {
+        return Ok(None);
+    }
+
+    if matched_row_ranges.len() == 1
+        && matched_row_ranges[0].start == 0
+        && matched_row_ranges[0].end == input.num_rows()
+    {
+        return Ok(Some(input));
+    }
+
+    if matched_row_ranges.len() == 1 {
+        let span = &matched_row_ranges[0];
+        return Ok(Some(input.slice(span.start, span.end - span.start)));
+    }
+
+    let mut builder = BooleanBufferBuilder::new(input.num_rows());
+    builder.append_n(input.num_rows(), false);
+    for span in matched_row_ranges {
+        for i in span {
+            builder.set_bit(i, true);
+        }
+    }
+
+    let filtered = datatypes::arrow::compute::filter_record_batch(
+        &input,
+        &BooleanArray::new(builder.finish(), None),
+    )
+    .context(ComputeArrowSnafu)?;
+    if filtered.num_rows() == 0 {
+        Ok(None)
+    } else {
+        Ok(Some(filtered))
+    }
+}
+
+pub(crate) struct CachedPrimaryKeyFilter {
+    inner: Box<dyn PrimaryKeyFilter>,
+    last_primary_key: Vec<u8>,
+    last_match: Option<bool>,
+}
+
+impl CachedPrimaryKeyFilter {
+    pub(crate) fn new(inner: Box<dyn PrimaryKeyFilter>) -> Self {
+        Self {
+            inner,
+            last_primary_key: Vec::new(),
+            last_match: None,
+        }
+    }
+}
+
+impl PrimaryKeyFilter for CachedPrimaryKeyFilter {
+    fn matches(&mut self, pk: &[u8]) -> mito_codec::error::Result<bool> {
+        if let Some(last_match) = self.last_match
+            && self.last_primary_key == pk
+        {
+            return Ok(last_match);
+        }
+
+        let matched = self.inner.matches(pk)?;
+        self.last_primary_key.clear();
+        self.last_primary_key.extend_from_slice(pk);
+        self.last_match = Some(matched);
+        Ok(matched)
+    }
+}
+
+/// How the bulk-memtable read should apply each predicate.
+///
+/// Unlike the parquet reader, the bulk path has no prefilter pass; predicates
+/// either run row-wise inside the iterator or are pushed down to encoded-PK
+/// matching when the batch still carries the primary-key column.
+pub(crate) struct BulkFilterPlan {
+    /// Simple filters the iterator still has to evaluate row-wise on each batch.
+    pub(crate) remaining_simple_filters: Vec<SimpleFilterContext>,
+    /// Tag predicates lowered to encoded-PK filters. `None` when the batch
+    /// already exposes raw tag columns or there are no tag predicates.
+    pub(crate) pk_filters: Option<Arc<Vec<SimpleFilterEvaluator>>>,
+}
+
+/// How the parquet reader should apply each predicate.
+///
+/// The reader runs in two phases. Predicates routed into `prefilter_builder`
+/// execute on a reduced column set first to compute a refined row selection;
+/// `remaining_simple_filters` execute alongside the full projection on the
+/// normal read path. The contract for what is precise vs best-effort is
+/// documented on [`build_reader_filter_plan`].
+pub(crate) struct ReaderFilterPlan {
+    /// Simple filters that must run on the normal read path: predicates with
+    /// `Matched` / `Pruned` outcomes (which carry expected-metadata
+    /// compatibility decisions later phases rely on), and predicates whose
+    /// column cannot be read directly during the prefilter pass.
+    pub(crate) remaining_simple_filters: Vec<SimpleFilterContext>,
+    /// Pre-built state for the prefilter pass, or `None` when prefiltering is
+    /// not worthwhile (no prefilter columns selected, or the prefilter
+    /// projection would cover nearly the full read).
+    pub(crate) prefilter_builder: Option<PrefilterContextBuilder>,
+}
+
+pub(crate) fn build_bulk_filter_plan(
+    read_format: &FlatReadFormat,
+    predicate: Option<&Predicate>,
+) -> BulkFilterPlan {
+    let metadata = read_format.metadata();
+    // Bulk memtable only needs simple binary filters here. Any filter that
+    // cannot be reduced to a SimpleFilterContext stays out of this fast path.
+    let simple_filters: Vec<SimpleFilterContext> = predicate
+        .into_iter()
+        .flat_map(|predicate| {
+            predicate
+                .exprs()
+                .iter()
+                .filter_map(|expr| SimpleFilterContext::new_opt(metadata, None, expr))
+        })
+        .collect();
+
+    // PK prefilter only works when flat batches still carry the encoded PK
+    // column. If tags have already been expanded to raw columns, the iterator
+    // can apply those filters directly and there is nothing to extract here.
+    if read_format.batch_has_raw_pk_columns() || metadata.primary_key.is_empty() {
+        return BulkFilterPlan {
+            remaining_simple_filters: simple_filters,
+            pk_filters: None,
+        };
+    }
+
+    let mut remaining_simple_filters = Vec::new();
+    let mut pk_filters = Vec::new();
+
+    for filter_ctx in simple_filters {
+        // Split tag predicates that can be evaluated against the encoded PK
+        // from filters that still need normal row-wise evaluation later.
+        let pk_filter = filter_ctx.filter().as_filter().and_then(|filter| {
+            (filter_ctx.semantic_type() == SemanticType::Tag).then(|| filter.clone())
+        });
+
+        if let Some(pk_filter) = pk_filter {
+            pk_filters.push(pk_filter);
+        } else {
+            remaining_simple_filters.push(filter_ctx);
+        }
+    }
+
+    BulkFilterPlan {
+        remaining_simple_filters,
+        pk_filters: (!pk_filters.is_empty()).then_some(Arc::new(pk_filters)),
+    }
+}
+
+/// Splits a query [`Predicate`] into a [`ReaderFilterPlan`]: predicates that can run
+/// during the prefilter pass (on a reduced projection, to compute a refined row
+/// selection) versus predicates that must run on the normal read path (alongside the
+/// full projection).
+///
+/// The prefilter pass is *best-effort pruning*: a physical-filter predicate is silently
+/// dropped when [`PhysicalFilterContext::new_opt`] returns `None` (column not in the
+/// projected arrow schema). This is safe because the DataFusion `FilterExec` above the
+/// reader always re-applies the original predicate, so the prefilter pass is purely a
+/// pruning hint.
+///
+/// Tag and timestamp predicates that lower to [`SimpleFilterEvaluator`] are an
+/// exception — the engine enforces them precisely, so the prefilter pass is the only
+/// place they execute. They are never silently dropped.
+pub(crate) fn build_reader_filter_plan(
+    predicate: Option<&Predicate>,
+    expected_metadata: Option<&RegionMetadata>,
+    pre_filter_mode: PreFilterMode,
+    read_format: &FlatReadFormat,
+    parquet_schema: &SchemaDescriptor,
+    codec: &Arc<dyn PrimaryKeyCodec>,
+) -> ReaderFilterPlan {
+    let Some(predicate) = predicate else {
+        return ReaderFilterPlan {
+            remaining_simple_filters: Vec::new(),
+            prefilter_builder: None,
+        };
+    };
+
+    let metadata = read_format.metadata();
+    let mut prefilter_simple_filters = Vec::new();
+    let mut remaining_simple_filters = Vec::new();
+    let mut prefilter_physical_filters = Vec::new();
+    let mut primary_key_filters = Vec::new();
+    let mut pk_filter_contexts = Vec::new();
+
+    // `SkipFields` keeps field predicates in the normal read path to avoid a
+    // second read of projected field columns, while tags/timestamp can still
+    // participate in prefiltering.
+    let field_prefilter_enabled = pre_filter_mode == PreFilterMode::All;
+    // When true, tag columns are encoded in the primary key column and are NOT
+    // stored as separate parquet columns. Tag predicates must go through PK
+    // decoding rather than direct column reads.
+    let need_pk_prefilter = !read_format.batch_has_raw_pk_columns();
+
+    // Whether a column can be read directly from parquet for prefiltering,
+    // based on its semantic type and the current mode/format.
+    let can_direct_prefilter = |semantic_type: SemanticType| -> bool {
+        match semantic_type {
+            SemanticType::Tag => !need_pk_prefilter,
+            SemanticType::Field => field_prefilter_enabled,
+            SemanticType::Timestamp => true,
+        }
+    };
+
+    for expr in predicate.exprs() {
+        // Prefer cheap simple filters first. They also preserve `Matched` /
+        // `Pruned` states for columns that only exist in expected metadata.
+        if let Some(filter_ctx) = SimpleFilterContext::new_opt(metadata, expected_metadata, expr) {
+            // `Matched` and `Pruned` come from expected-metadata compatibility
+            // and must stay in the main filter list so later phases keep that
+            // outcome.
+            let Some(filter) = filter_ctx.filter().as_filter() else {
+                remaining_simple_filters.push(filter_ctx);
+                continue;
+            };
+
+            // If the column is stored as a separate parquet column and is already projected in the main read,
+            // we can evaluate the simple filter directly during prefilter.
+            let direct_prefilter = can_direct_prefilter(filter_ctx.semantic_type());
+            if direct_prefilter {
+                assert!(
+                    read_format
+                        .arrow_schema()
+                        .column_with_name(filter.column_name())
+                        .is_some(),
+                    "Column '{}' is not present in the arrow schema {:?}",
+                    filter.column_name(),
+                    read_format.arrow_schema(),
+                );
+                prefilter_simple_filters.push(filter_ctx);
+                continue;
+            }
+
+            // Otherwise try to filter through encoded-PK matching.
+            if need_pk_prefilter && filter_ctx.semantic_type() == SemanticType::Tag {
+                primary_key_filters.push(filter.clone());
+                pk_filter_contexts.push(filter_ctx);
+            } else {
+                remaining_simple_filters.push(filter_ctx);
+            }
+            continue;
+        }
+
+        // Best-effort physical-filter prefilter (see fn-level doc): `new_opt`
+        // returning `None` means the column is not in the projected arrow
+        // schema, and dropping the predicate is safe because the upper
+        // `FilterExec` re-applies it.
+        if let Some(filter) =
+            PhysicalFilterContext::new_opt(metadata, expected_metadata, read_format, expr)
+            && can_direct_prefilter(filter.semantic_type())
+        {
+            prefilter_physical_filters.push(filter);
+        }
+    }
+
+    let pk_filter_exprs =
+        (!primary_key_filters.is_empty()).then_some(Arc::new(primary_key_filters));
+    let prefilter_builder = PrefilterContextBuilder::new(
+        read_format,
+        codec,
+        pk_filter_exprs,
+        prefilter_simple_filters.clone(),
+        prefilter_physical_filters,
+        parquet_schema,
+    );
+
+    if prefilter_builder.is_some() {
+        ReaderFilterPlan {
+            remaining_simple_filters,
+            prefilter_builder,
+        }
+    } else {
+        // If prefilter setup is not worthwhile, keep the original simple
+        // filters on the normal path so behavior is unchanged.
+        remaining_simple_filters.extend(prefilter_simple_filters);
+        remaining_simple_filters.extend(pk_filter_contexts);
+        ReaderFilterPlan {
+            remaining_simple_filters,
+            prefilter_builder: None,
+        }
+    }
+}
+
+/// Context for prefiltering a row group.
+pub(crate) struct PrefilterContext {
+    /// Projection mask for reading prefilter columns.
+    projection: ProjectionMask,
+    /// Optional PK filter for legacy primary-key-format parquet.
+    pk_filter: Option<Box<dyn PrimaryKeyFilter>>,
+    /// Simple filters that can be evaluated directly from the prefilter batch.
+    filters: Vec<SimpleFilterContext>,
+    /// Physical filters that can be evaluated directly from the prefilter batch.
+    /// Physical expressions are only applied in the prefilter phase.
+    physical_filters: Vec<PhysicalFilterContext>,
+}
+
+/// Pre-built state for constructing [PrefilterContext] per row group.
+///
+/// Fields invariant across row groups (projection mask, codec, metadata, filters)
+/// are computed once. A fresh [PrefilterContext] with its own mutable PK filter
+/// is created via [PrefilterContextBuilder::build()] for each row group.
+pub(crate) struct PrefilterContextBuilder {
+    projection: ProjectionMask,
+    pk_filters: Option<Arc<Vec<SimpleFilterEvaluator>>>,
+    filters: Vec<SimpleFilterContext>,
+    physical_filters: Vec<PhysicalFilterContext>,
+    codec: Arc<dyn PrimaryKeyCodec>,
+    metadata: RegionMetadataRef,
+}
+
+impl PrefilterContextBuilder {
+    /// Creates a builder if prefiltering is applicable.
+    ///
+    /// Returns `None` if:
+    /// - The read format doesn't use flat layout
+    /// - No prefilter columns are selected
+    /// - Prefilter would read the full projection without any PK filter
+    pub(crate) fn new(
+        read_format: &FlatReadFormat,
+        codec: &Arc<dyn PrimaryKeyCodec>,
+        primary_key_filters: Option<Arc<Vec<SimpleFilterEvaluator>>>,
+        filters: Vec<SimpleFilterContext>,
+        physical_filters: Vec<PhysicalFilterContext>,
+        parquet_schema: &SchemaDescriptor,
+    ) -> Option<Self> {
+        let metadata = read_format.metadata();
+        let use_raw_tag_columns = read_format.batch_has_raw_pk_columns();
+        let pk_filters = (!use_raw_tag_columns)
+            .then_some(primary_key_filters)
+            .flatten()
+            .filter(|filters| !filters.is_empty());
+
+        let mut prefilter_column_names = HashSet::new();
+        for filter_ctx in &filters {
+            if let MaybeFilter::Filter(filter) = filter_ctx.filter() {
+                prefilter_column_names.insert(filter.column_name().to_string());
+            }
+        }
+
+        if pk_filters.is_some() {
+            prefilter_column_names.insert(PRIMARY_KEY_COLUMN_NAME.to_string());
+        }
+
+        for filter_ctx in &physical_filters {
+            prefilter_column_names.insert(filter_ctx.column_name().to_string());
+        }
+
+        let (projection, prefilter_count) = compute_projection_mask(
+            &prefilter_column_names,
+            read_format.arrow_schema(),
+            parquet_schema,
+        );
+
+        if prefilter_count == 0 {
+            return None;
+        }
+
+        let total_count = read_format.parquet_read_columns().root_indices().len();
+        let remaining_count = total_count.saturating_sub(prefilter_count);
+        if pk_filters.is_none() && prefilter_count >= total_count {
+            return None;
+        }
+
+        if pk_filters.is_none()
+            && !should_use_prefilter(prefilter_count, remaining_count, total_count)
+        {
+            return None;
+        }
+
+        Some(Self {
+            projection,
+            pk_filters,
+            filters,
+            physical_filters,
+            codec: Arc::clone(codec),
+            metadata: metadata.clone(),
+        })
+    }
+
+    /// Builds a [PrefilterContext] for a specific row group.
+    pub(crate) fn build(&self) -> PrefilterContext {
+        let pk_filter = self.pk_filters.as_ref().map(|pk_filters| {
+            let pk_filter =
+                self.codec
+                    .primary_key_filter(&self.metadata, Arc::clone(pk_filters), false);
+            Box::new(CachedPrimaryKeyFilter::new(pk_filter)) as Box<dyn PrimaryKeyFilter>
+        });
+        PrefilterContext {
+            projection: self.projection.clone(),
+            pk_filter,
+            filters: self.filters.clone(),
+            physical_filters: self.physical_filters.clone(),
+        }
+    }
+}
+
+const PREFILTER_COLUMN_RATIO_THRESHOLD: f64 = 0.5;
+const PREFILTER_MIN_REMAINING_COLUMNS: usize = 2;
+
+/// Result of prefiltering a row group.
+pub(crate) struct PrefilterResult {
+    /// Refined row selection after prefiltering.
+    pub(crate) refined_selection: RowSelection,
+    /// Number of rows filtered out by prefiltering.
+    pub(crate) filtered_rows: usize,
+}
+
+/// Executes prefiltering on a row group.
+///
+/// Reads only the prefilter columns (currently the PK dictionary column),
+/// applies filters, and returns a refined [RowSelection].
+fn compute_projection_mask(
+    column_names: &HashSet<String>,
+    arrow_schema: &datatypes::arrow::datatypes::SchemaRef,
+    parquet_schema: &SchemaDescriptor,
+) -> (ProjectionMask, usize) {
+    let mut projection_indices: Vec<usize> = column_names
+        .iter()
+        .filter_map(|name| arrow_schema.column_with_name(name).map(|(index, _)| index))
+        .collect();
+    projection_indices.sort_unstable();
+    projection_indices.dedup();
+    let count = projection_indices.len();
+    (
+        ProjectionMask::roots(parquet_schema, projection_indices.iter().copied()),
+        count,
+    )
+}
+
+fn should_use_prefilter(
+    prefilter_count: usize,
+    remaining_count: usize,
+    total_count: usize,
+) -> bool {
+    if remaining_count == 0 {
+        return false;
+    }
+
+    if remaining_count < PREFILTER_MIN_REMAINING_COLUMNS {
+        return false;
+    }
+
+    let ratio = prefilter_count as f64 / total_count as f64;
+    ratio <= PREFILTER_COLUMN_RATIO_THRESHOLD
+}
+
+pub(crate) async fn execute_prefilter(
+    prefilter_ctx: &mut PrefilterContext,
+    reader_builder: &RowGroupReaderBuilder,
+    build_ctx: &RowGroupBuildContext<'_>,
+) -> Result<PrefilterResult> {
+    let mut stream = reader_builder
+        .build_with_projection(
+            build_ctx.row_group_idx,
+            build_ctx.row_selection.clone(),
+            prefilter_ctx.projection.clone(),
+            build_ctx.fetch_metrics,
+        )
+        .await?;
+
+    let mut filter_arrays = Vec::new();
+    let mut rows_before_filter = 0usize;
+    let mut rows_selected = 0usize;
+
+    while let Some(batch_result) = stream.next().await {
+        let batch = batch_result.context(ReadParquetSnafu {
+            path: reader_builder.file_path(),
+        })?;
+        let num_rows = batch.num_rows();
+        if num_rows == 0 {
+            continue;
+        }
+        rows_before_filter += num_rows;
+
+        let batch_mask = match apply_filters_to_batch(
+            &batch,
+            &mut prefilter_ctx.pk_filter,
+            &prefilter_ctx.filters,
+            &prefilter_ctx.physical_filters,
+            reader_builder.file_path(),
+        )? {
+            Some(mask) => mask,
+            None => BooleanBuffer::new_unset(num_rows),
+        };
+        rows_selected += batch_mask.count_set_bits();
+        filter_arrays.push(BooleanArray::from(batch_mask));
+    }
+
+    let filtered_rows = rows_before_filter.saturating_sub(rows_selected);
+    let refined_selection = if filter_arrays.is_empty() || rows_selected == 0 {
+        RowSelection::from(vec![])
+    } else {
+        let prefilter_selection = RowSelection::from_filters(&filter_arrays);
+        match &build_ctx.row_selection {
+            Some(original) => original.and_then(&prefilter_selection),
+            None => prefilter_selection,
+        }
+    };
+
+    Ok(PrefilterResult {
+        refined_selection,
+        filtered_rows,
+    })
+}
+
+fn apply_filters_to_batch(
+    batch: &RecordBatch,
+    pk_filter: &mut Option<Box<dyn PrimaryKeyFilter>>,
+    filters: &[SimpleFilterContext],
+    physical_filters: &[PhysicalFilterContext],
+    file_path: &str,
+) -> Result<Option<BooleanBuffer>> {
+    let mut mask = BooleanBuffer::new_set(batch.num_rows());
+
+    if let Some(pk_filter) = pk_filter.as_mut() {
+        // Prefilter reads a reduced projection. For PK prefilter, the encoded
+        // primary key column is always appended as the last projected column,
+        // while `__sequence` and `__op_type` are not read.
+        let pk_column_index = batch.num_columns() - 1;
+        let matched_row_ranges =
+            matching_row_ranges_by_primary_key(batch, pk_column_index, pk_filter.as_mut())?;
+        let mut builder = BooleanBufferBuilder::new(batch.num_rows());
+        builder.append_n(batch.num_rows(), false);
+        for range in matched_row_ranges {
+            for row in range {
+                builder.set_bit(row, true);
+            }
+        }
+        mask = mask.bitand(&builder.finish());
+    }
+
+    for filter_ctx in filters {
+        let filter = match filter_ctx.filter() {
+            MaybeFilter::Filter(filter) => filter,
+            MaybeFilter::Matched => continue,
+            MaybeFilter::Pruned => return Ok(None),
+        };
+
+        let (idx, _) = batch
+            .schema()
+            .column_with_name(filter.column_name())
+            .with_context(|| UnexpectedSnafu {
+                reason: format!(
+                    "Prefilter column '{}' (id {}) not found in batch for file {}",
+                    filter.column_name(),
+                    filter_ctx.column_id(),
+                    file_path
+                ),
+            })?;
+        let column = batch.column(idx).clone();
+        let result = filter.evaluate_array(&column).context(RecordBatchSnafu)?;
+        mask = mask.bitand(&result);
+    }
+
+    for filter_ctx in physical_filters {
+        let filter = filter_ctx.filter();
+
+        let (idx, _) = batch
+            .schema()
+            .column_with_name(filter_ctx.column_name())
+            .with_context(|| UnexpectedSnafu {
+                reason: format!(
+                    "Prefilter physical column '{}' (id {}) not found in batch for file {}",
+                    filter_ctx.column_name(),
+                    filter_ctx.column_id(),
+                    file_path
+                ),
+            })?;
+        let column = batch.column(idx).clone();
+
+        let record_batch = RecordBatch::try_new(filter_ctx.schema().clone(), vec![column])
+            .context(NewRecordBatchSnafu)?;
+        let evaluated = filter
+            .evaluate(&record_batch)
+            .context(EvalPartitionFilterSnafu)?;
+        let array = evaluated
+            .into_array(record_batch.num_rows())
+            .context(EvalPartitionFilterSnafu)?;
+        let boolean_array =
+            array
+                .as_any()
+                .downcast_ref::<BooleanArray>()
+                .context(UnexpectedSnafu {
+                    reason: "Failed to downcast physical filter result to BooleanArray",
+                })?;
+        // Treat null results as false (filtered out); value bits are not guaranteed
+        // to be false for invalid entries.
+        let mut result = boolean_array.values().clone();
+        if let Some(nulls) = boolean_array.nulls() {
+            result = result.bitand(nulls.inner());
+        }
+        mask = mask.bitand(&result);
+    }
+
+    if mask.count_set_bits() == 0 {
+        Ok(None)
+    } else {
+        Ok(Some(mask))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+    use std::sync::atomic::{AtomicUsize, Ordering};
+
+    use common_recordbatch::filter::SimpleFilterEvaluator;
+    use datafusion_expr::{col, lit};
+    use datatypes::arrow::array::{
+        ArrayRef, DictionaryArray, TimestampMillisecondArray, UInt8Array, UInt32Array, UInt64Array,
+    };
+    use datatypes::arrow::datatypes::{Schema, UInt32Type};
+    use mito_codec::row_converter::{PrimaryKeyFilter, build_primary_key_codec};
+    use parquet::arrow::ArrowSchemaConverter;
+    use store_api::codec::PrimaryKeyEncoding;
+
+    use super::*;
+    use crate::read::read_columns::ReadColumns;
+    use crate::sst::internal_fields;
+    use crate::sst::parquet::flat_format::{FlatReadFormat, primary_key_column_index};
+    use crate::test_util::sst_util::{
+        new_primary_key, new_record_batch_with_custom_sequence, sst_region_metadata,
+        sst_region_metadata_with_encoding,
+    };
+
+    struct CountingPrimaryKeyFilter {
+        hits: Arc<AtomicUsize>,
+        expected: Vec<u8>,
+    }
+
+    impl PrimaryKeyFilter for CountingPrimaryKeyFilter {
+        fn matches(&mut self, pk: &[u8]) -> mito_codec::error::Result<bool> {
+            self.hits.fetch_add(1, Ordering::Relaxed);
+            Ok(pk == self.expected.as_slice())
+        }
+    }
+
+    #[test]
+    fn test_cached_primary_key_filter_reuses_previous_result() {
+        let expected = new_primary_key(&["a", "x"]);
+        let hits = Arc::new(AtomicUsize::new(0));
+        let mut filter = CachedPrimaryKeyFilter::new(Box::new(CountingPrimaryKeyFilter {
+            hits: Arc::clone(&hits),
+            expected: expected.clone(),
+        }));
+
+        assert!(filter.matches(expected.as_slice()).unwrap());
+        assert!(filter.matches(expected.as_slice()).unwrap());
+        assert!(
+            !filter
+                .matches(new_primary_key(&["b", "x"]).as_slice())
+                .unwrap()
+        );
+
+        assert_eq!(hits.load(Ordering::Relaxed), 2);
+    }
+
+    fn new_test_filters(exprs: &[datafusion_expr::Expr]) -> Vec<SimpleFilterEvaluator> {
+        exprs
+            .iter()
+            .filter_map(SimpleFilterEvaluator::try_new)
+            .collect()
+    }
+
+    fn new_simple_filter_contexts(
+        metadata: &RegionMetadataRef,
+        exprs: &[datafusion_expr::Expr],
+    ) -> Vec<SimpleFilterContext> {
+        exprs
+            .iter()
+            .filter_map(|expr| SimpleFilterContext::new_opt(metadata, None, expr))
+            .collect()
+    }
+
+    fn new_physical_filter_contexts(
+        metadata: &RegionMetadataRef,
+        read_format: &FlatReadFormat,
+        exprs: &[datafusion_expr::Expr],
+    ) -> Vec<PhysicalFilterContext> {
+        exprs
+            .iter()
+            .filter_map(|expr| PhysicalFilterContext::new_opt(metadata, None, read_format, expr))
+            .collect()
+    }
+
+    fn parquet_schema(read_format: &FlatReadFormat) -> SchemaDescriptor {
+        ArrowSchemaConverter::new()
+            .convert(read_format.arrow_schema())
+            .unwrap()
+    }
+
+    fn new_raw_batch(primary_keys: &[&[u8]], field_values: &[u64]) -> RecordBatch {
+        assert_eq!(primary_keys.len(), field_values.len());
+
+        let metadata = Arc::new(sst_region_metadata());
+        let arrow_schema = metadata.schema.arrow_schema();
+        let field_column = arrow_schema
+            .field(arrow_schema.index_of("field_0").unwrap())
+            .clone();
+        let time_index_column = arrow_schema
+            .field(arrow_schema.index_of("ts").unwrap())
+            .clone();
+        let mut fields = vec![field_column, time_index_column];
+        fields.extend(
+            internal_fields()
+                .into_iter()
+                .map(|field| field.as_ref().clone()),
+        );
+        let schema = Arc::new(Schema::new(fields));
+
+        let mut dict_values = Vec::new();
+        let mut keys = Vec::with_capacity(primary_keys.len());
+        for pk in primary_keys {
+            let key = dict_values
+                .iter()
+                .position(|existing: &&[u8]| existing == pk)
+                .unwrap_or_else(|| {
+                    dict_values.push(*pk);
+                    dict_values.len() - 1
+                });
+            keys.push(key as u32);
+        }
+        let pk_array: ArrayRef = Arc::new(DictionaryArray::<UInt32Type>::new(
+            UInt32Array::from(keys),
+            Arc::new(BinaryArray::from_iter_values(dict_values.iter().copied())),
+        ));
+
+        RecordBatch::try_new(
+            schema,
+            vec![
+                Arc::new(UInt64Array::from(field_values.to_vec())),
+                Arc::new(TimestampMillisecondArray::from_iter_values(
+                    0..primary_keys.len() as i64,
+                )),
+                pk_array,
+                Arc::new(UInt64Array::from(vec![1; primary_keys.len()])),
+                Arc::new(UInt8Array::from(vec![1; primary_keys.len()])),
+            ],
+        )
+        .unwrap()
+    }
+
+    fn new_prefilter_batch(primary_keys: &[&[u8]], field_values: &[u64]) -> RecordBatch {
+        assert_eq!(primary_keys.len(), field_values.len());
+
+        let metadata = Arc::new(sst_region_metadata());
+        let arrow_schema = metadata.schema.arrow_schema();
+        let field_column = arrow_schema
+            .field(arrow_schema.index_of("field_0").unwrap())
+            .clone();
+        let time_index_column = arrow_schema
+            .field(arrow_schema.index_of("ts").unwrap())
+            .clone();
+        let schema = Arc::new(Schema::new(vec![
+            field_column,
+            time_index_column,
+            internal_fields()[0].as_ref().clone(),
+        ]));
+
+        let mut dict_values = Vec::new();
+        let mut keys = Vec::with_capacity(primary_keys.len());
+        for pk in primary_keys {
+            let key = dict_values
+                .iter()
+                .position(|existing: &&[u8]| existing == pk)
+                .unwrap_or_else(|| {
+                    dict_values.push(*pk);
+                    dict_values.len() - 1
+                });
+            keys.push(key as u32);
+        }
+        let pk_array: ArrayRef = Arc::new(DictionaryArray::<UInt32Type>::new(
+            UInt32Array::from(keys),
+            Arc::new(BinaryArray::from_iter_values(dict_values.iter().copied())),
+        ));
+
+        RecordBatch::try_new(
+            schema,
+            vec![
+                Arc::new(UInt64Array::from(field_values.to_vec())),
+                Arc::new(TimestampMillisecondArray::from_iter_values(
+                    0..primary_keys.len() as i64,
+                )),
+                pk_array,
+            ],
+        )
+        .unwrap()
+    }
+
+    fn field_values(batch: &RecordBatch) -> Vec<u64> {
+        batch
+            .column(0)
+            .as_any()
+            .downcast_ref::<UInt64Array>()
+            .unwrap()
+            .values()
+            .to_vec()
+    }
+
+    fn remaining_simple_filter_columns(filters: &[SimpleFilterContext]) -> Vec<&str> {
+        filters
+            .iter()
+            .map(|filter_ctx| filter_ctx.filter().as_filter().unwrap().column_name())
+            .collect()
+    }
+
+    #[test]
+    fn test_prefilter_primary_key_drops_single_dictionary_batch() {
+        let metadata = Arc::new(sst_region_metadata());
+        let filters = Arc::new(new_test_filters(&[col("tag_0").eq(lit("b"))]));
+        let mut primary_key_filter = build_primary_key_codec(metadata.as_ref())
+            .primary_key_filter(&metadata, filters, false);
+        let pk_a = new_primary_key(&["a", "x"]);
+        let batch = new_raw_batch(&[pk_a.as_slice(), pk_a.as_slice()], &[10, 11]);
+        let pk_col_idx = primary_key_column_index(batch.num_columns());
+
+        let filtered =
+            prefilter_flat_batch_by_primary_key(batch, pk_col_idx, primary_key_filter.as_mut())
+                .unwrap();
+
+        assert!(filtered.is_none());
+    }
+
+    #[test]
+    fn test_prefilter_primary_key_builds_mask_for_fragmented_matches() {
+        let metadata = Arc::new(sst_region_metadata());
+        let filters = Arc::new(new_test_filters(&[col("tag_0")
+            .eq(lit("a"))
+            .or(col("tag_0").eq(lit("c")))]));
+        let mut primary_key_filter = build_primary_key_codec(metadata.as_ref())
+            .primary_key_filter(&metadata, filters, false);
+        let pk_a = new_primary_key(&["a", "x"]);
+        let pk_b = new_primary_key(&["b", "x"]);
+        let pk_c = new_primary_key(&["c", "x"]);
+        let pk_d = new_primary_key(&["d", "x"]);
+        let batch = new_raw_batch(
+            &[
+                pk_a.as_slice(),
+                pk_a.as_slice(),
+                pk_b.as_slice(),
+                pk_b.as_slice(),
+                pk_c.as_slice(),
+                pk_c.as_slice(),
+                pk_d.as_slice(),
+                pk_d.as_slice(),
+            ],
+            &[10, 11, 12, 13, 14, 15, 16, 17],
+        );
+        let pk_col_idx = primary_key_column_index(batch.num_columns());
+
+        let filtered =
+            prefilter_flat_batch_by_primary_key(batch, pk_col_idx, primary_key_filter.as_mut())
+                .unwrap()
+                .unwrap();
+
+        assert_eq!(filtered.num_rows(), 4);
+        assert_eq!(field_values(&filtered), vec![10, 11, 14, 15]);
+    }
+
+    #[test]
+    fn test_prefilter_builder_returns_none_without_selected_filters() {
+        let metadata: RegionMetadataRef =
+            Arc::new(sst_region_metadata_with_encoding(PrimaryKeyEncoding::Dense));
+        let read_format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids(
+                metadata.column_metadatas.iter().map(|c| c.column_id),
+            ),
+            None,
+            "test",
+            false,
+        )
+        .unwrap();
+        let codec = build_primary_key_codec(metadata.as_ref());
+        let parquet_schema = parquet_schema(&read_format);
+
+        let builder = PrefilterContextBuilder::new(
+            &read_format,
+            &codec,
+            None,
+            Vec::new(),
+            Vec::new(),
+            &parquet_schema,
+        );
+        assert!(builder.is_none());
+    }
+
+    #[test]
+    fn test_should_use_prefilter() {
+        assert!(should_use_prefilter(1, 5, 6));
+        assert!(!should_use_prefilter(1, 0, 1));
+        assert!(!should_use_prefilter(1, 1, 2));
+        assert!(!should_use_prefilter(4, 3, 7));
+        assert!(should_use_prefilter(3, 3, 6));
+    }
+
+    #[test]
+    fn test_build_bulk_filter_plan_classifies_filters_across_read_paths() {
+        let metadata: RegionMetadataRef = Arc::new(sst_region_metadata_with_encoding(
+            PrimaryKeyEncoding::Sparse,
+        ));
+        let legacy_read_format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids(
+                metadata.column_metadatas.iter().map(|c| c.column_id),
+            ),
+            None,
+            "memtable",
+            false,
+        )
+        .unwrap();
+        assert!(!legacy_read_format.batch_has_raw_pk_columns());
+
+        let plan = build_bulk_filter_plan(
+            &legacy_read_format,
+            Some(&Predicate::new(vec![
+                col("tag_0").eq(lit("a")),
+                col("field_0").gt(lit(1_u64)),
+            ])),
+        );
+        assert_eq!(
+            plan.pk_filters.as_ref().map(|filters| filters.len()),
+            Some(1)
+        );
+        assert_eq!(
+            remaining_simple_filter_columns(&plan.remaining_simple_filters),
+            vec!["field_0"]
+        );
+
+        let metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
+        let raw_pk_read_format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids(
+                metadata.column_metadatas.iter().map(|c| c.column_id),
+            ),
+            None,
+            "memtable",
+            true,
+        )
+        .unwrap();
+        assert!(raw_pk_read_format.batch_has_raw_pk_columns());
+
+        let tag_only_plan = build_bulk_filter_plan(
+            &raw_pk_read_format,
+            Some(&Predicate::new(vec![col("tag_0").eq(lit("a"))])),
+        );
+        assert!(tag_only_plan.pk_filters.is_none());
+        assert_eq!(
+            remaining_simple_filter_columns(&tag_only_plan.remaining_simple_filters),
+            vec!["tag_0"]
+        );
+
+        let field_only_plan = build_bulk_filter_plan(
+            &raw_pk_read_format,
+            Some(&Predicate::new(vec![col("field_0").gt(lit(1_u64))])),
+        );
+        assert!(field_only_plan.pk_filters.is_none());
+        assert_eq!(
+            remaining_simple_filter_columns(&field_only_plan.remaining_simple_filters),
+            vec!["field_0"]
+        );
+    }
+
+    #[test]
+    fn test_build_reader_filter_plan_classifies_filters_for_prefilter_modes() {
+        let metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
+        let full_read_format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids(
+                metadata.column_metadatas.iter().map(|c| c.column_id),
+            ),
+            None,
+            "test",
+            true,
+        )
+        .unwrap();
+        let full_parquet_schema = parquet_schema(&full_read_format);
+        let codec = build_primary_key_codec(metadata.as_ref());
+
+        let skip_fields_plan = build_reader_filter_plan(
+            Some(&Predicate::new(vec![
+                col("tag_0").eq(lit("a")),
+                col("field_0").gt(lit(1_u64)),
+            ])),
+            None,
+            PreFilterMode::SkipFields,
+            &full_read_format,
+            &full_parquet_schema,
+            &codec,
+        );
+        assert!(skip_fields_plan.prefilter_builder.is_some());
+        assert_eq!(
+            remaining_simple_filter_columns(&skip_fields_plan.remaining_simple_filters),
+            vec!["field_0"]
+        );
+
+        let field_0 = metadata.column_by_name("field_0").unwrap().column_id;
+        let ts = metadata.time_index_column().column_id;
+        let projected_read_format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids([field_0, ts]),
+            None,
+            "test",
+            true,
+        )
+        .unwrap();
+        let projected_parquet_schema = parquet_schema(&projected_read_format);
+        let pk_prefilter_plan = build_reader_filter_plan(
+            Some(&Predicate::new(vec![col("tag_0").eq(lit("a"))])),
+            None,
+            PreFilterMode::All,
+            &projected_read_format,
+            &projected_parquet_schema,
+            &codec,
+        );
+        assert!(pk_prefilter_plan.prefilter_builder.is_some());
+        assert!(pk_prefilter_plan.remaining_simple_filters.is_empty());
+    }
+
+    #[test]
+    fn test_apply_filters_to_batch_uses_flat_tag_columns_directly() {
+        let metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
+        let filters = new_simple_filter_contexts(&metadata, &[col("tag_0").eq(lit("a"))]);
+        let batch = new_record_batch_with_custom_sequence(&["a", "x"], 0, 4, 1);
+
+        let mut no_pk_filter = None;
+        let mask = apply_filters_to_batch(&batch, &mut no_pk_filter, &filters, &[], "test")
+            .unwrap()
+            .unwrap();
+        assert_eq!(mask.count_set_bits(), 4);
+    }
+
+    #[test]
+    fn test_apply_filters_to_batch_errors_on_missing_selected_column() {
+        let metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
+        let filters = new_simple_filter_contexts(&metadata, &[col("tag_0").eq(lit("a"))]);
+        let pk = new_primary_key(&["a", "x"]);
+        let batch = new_raw_batch(&[pk.as_slice()], &[10]);
+
+        let mut no_pk_filter = None;
+        let err =
+            apply_filters_to_batch(&batch, &mut no_pk_filter, &filters, &[], "test").unwrap_err();
+        let err = err.to_string();
+        assert!(err.contains("Prefilter column"));
+        assert!(err.contains("tag_0"));
+    }
+
+    #[test]
+    fn test_apply_filters_to_batch_evaluates_physical_filters() {
+        let metadata: RegionMetadataRef =
+            Arc::new(sst_region_metadata_with_encoding(PrimaryKeyEncoding::Dense));
+        let read_format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids(
+                metadata.column_metadatas.iter().map(|c| c.column_id),
+            ),
+            None,
+            "test",
+            false,
+        )
+        .unwrap();
+        let expr = col("field_0").in_list(vec![lit(11_u64)], false);
+        let physical_filters = new_physical_filter_contexts(&metadata, &read_format, &[expr]);
+        let pk = new_primary_key(&["a", "x"]);
+        let batch = new_raw_batch(&[pk.as_slice(), pk.as_slice(), pk.as_slice()], &[9, 10, 11]);
+
+        let mut no_pk_filter = None;
+        let mask =
+            apply_filters_to_batch(&batch, &mut no_pk_filter, &[], &physical_filters, "test")
+                .unwrap()
+                .unwrap();
+        assert_eq!(mask.count_set_bits(), 1);
+    }
+
+    #[test]
+    fn test_apply_filters_to_batch_uses_last_projected_column_for_pk_prefilter() {
+        let metadata = Arc::new(sst_region_metadata());
+        let filters = Arc::new(new_test_filters(&[col("tag_0").eq(lit("a"))]));
+        let mut pk_filter = Some(Box::new(CachedPrimaryKeyFilter::new(
+            build_primary_key_codec(metadata.as_ref())
+                .primary_key_filter(&metadata, filters, false),
+        )) as Box<dyn PrimaryKeyFilter>);
+        let pk_a = new_primary_key(&["a", "x"]);
+        let pk_b = new_primary_key(&["b", "x"]);
+        let batch = new_prefilter_batch(
+            &[
+                pk_a.as_slice(),
+                pk_a.as_slice(),
+                pk_b.as_slice(),
+                pk_b.as_slice(),
+            ],
+            &[10, 11, 12, 13],
+        );
+
+        let mask = apply_filters_to_batch(&batch, &mut pk_filter, &[], &[], "test")
+            .unwrap()
+            .unwrap();
+
+        assert_eq!(mask.count_set_bits(), 2);
+    }
+}
diff --git a/src/mito2/src/sst/parquet/read_columns.rs b/src/mito2/src/sst/parquet/read_columns.rs
new file mode 100644
index 0000000000..bca1873060
--- /dev/null
+++ b/src/mito2/src/sst/parquet/read_columns.rs
@@ -0,0 +1,465 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::{HashMap, HashSet};
+
+use parquet::arrow::ProjectionMask;
+use parquet::schema::types::SchemaDescriptor;
+
+/// A nested field access path inside one parquet root column.
+pub type ParquetNestedPath = Vec<String>;
+
+/// The parquet columns to read.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ParquetReadColumns {
+    /// Root parquet column indices in the same order as `cols`.
+    ///
+    /// Most readers need these indices as a borrowed slice for Arrow schema
+    /// projection or parquet root-column projection. Keeping them here avoids
+    /// repeatedly collecting `cols.iter().map(|col| col.root_index)`.
+    root_indices: Vec<usize>,
+    cols: Vec<ParquetReadColumn>,
+    has_nested: bool,
+}
+
+impl ParquetReadColumns {
+    /// Builds parquet read columns from deduplicated, normalized input.
+    ///
+    /// `cols` must not contain duplicate root indices, and nested paths must
+    /// already be merged. Empty `nested_paths` means reading the whole root column.
+    ///
+    /// This constructor does not validate or merge input.
+    pub fn from_deduped(cols: Vec<ParquetReadColumn>) -> Self {
+        let has_nested = cols.iter().any(|col| !col.nested_paths.is_empty());
+        let root_indices = cols.iter().map(|col| col.root_index).collect();
+        Self {
+            root_indices,
+            cols,
+            has_nested,
+        }
+    }
+
+    /// Builds root-column projections from root indices that are already
+    /// deduplicated.
+    ///
+    /// Note: this constructor does not check for duplicates.
+    pub fn from_deduped_root_indices(root_indices: impl IntoIterator<Item = usize>) -> Self {
+        let root_indices = root_indices.into_iter().collect::<Vec<_>>();
+        let cols = root_indices
+            .iter()
+            .copied()
+            .map(ParquetReadColumn::new)
+            .collect();
+        Self {
+            root_indices,
+            cols,
+            has_nested: false,
+        }
+    }
+
+    pub fn columns(&self) -> &[ParquetReadColumn] {
+        &self.cols
+    }
+
+    pub fn has_nested(&self) -> bool {
+        self.has_nested
+    }
+
+    pub fn root_indices_iter(&self) -> impl Iterator<Item = usize> + '_ {
+        self.root_indices.iter().copied()
+    }
+
+    /// Returns root parquet column indices.
+    pub fn root_indices(&self) -> &[usize] {
+        &self.root_indices
+    }
+}
+
+/// Read requirement for a single parquet root column.
+///
+/// `root_index` identifies the root column in the parquet schema.
+///
+/// If `nested_paths` is empty, the whole root column is read. Otherwise, only
+/// leaves under the specified nested paths are read.
+///
+/// To construct a [`ParquetReadColumn`]:
+/// - `ParquetReadColumn::new(0)` reads the whole root column at index `0`.
+/// - `ParquetReadColumn::new(0).with_nested_paths(vec![vec!["j".into(), "b".into()]])`
+///   reads only leaves under `j.b`.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ParquetReadColumn {
+    /// Root field index in the parquet schema.
+    root_index: usize,
+    /// Nested paths to read under this root column.
+    ///
+    /// Each path includes the root column itself. For example, for a root
+    /// column `j`, path `["j", "a", "b"]` refers to `j.a.b`.
+    ///
+    /// If empty, the whole root column is read.
+    nested_paths: Vec<ParquetNestedPath>,
+}
+
+impl ParquetReadColumn {
+    pub fn new(root_index: usize) -> Self {
+        Self {
+            root_index,
+            nested_paths: vec![],
+        }
+    }
+
+    pub fn with_nested_paths(self, nested_paths: Vec<ParquetNestedPath>) -> Self {
+        Self {
+            nested_paths,
+            ..self
+        }
+    }
+
+    /// Merges additional nested paths into this root column.
+    pub fn merge_nested_paths(&mut self, nested_paths: Vec<ParquetNestedPath>) {
+        let reads_whole_root = self.nested_paths.is_empty() || nested_paths.is_empty();
+        if reads_whole_root {
+            // Empty nested paths means reading the whole root column.
+            self.nested_paths = vec![];
+        } else {
+            self.nested_paths.extend(nested_paths);
+        }
+    }
+
+    pub fn root_index(&self) -> usize {
+        self.root_index
+    }
+
+    pub fn nested_paths(&self) -> &[ParquetNestedPath] {
+        &self.nested_paths
+    }
+}
+
+/// Projection plan built for a parquet file.
+#[derive(Clone)]
+pub struct ProjectionMaskPlan {
+    /// `mask` is the projection mask applied to the parquet reader.
+    pub mask: ProjectionMask,
+    /// A boolean mask in output schema order indicating whether each
+    /// projected root column is physically present in the parquet
+    /// read result.
+    ///
+    /// - `true`: the column exists in the input `RecordBatch`.
+    /// - `false`: the column is missing (e.g., due to unmatched nested
+    ///   paths) and must be synthesized during post-processing (typically
+    ///   filled with null/default values).
+    ///
+    /// The length of `projected_root_presence` is always equal to the
+    /// number of fields in the output schema.
+    pub projected_root_presence: Vec<bool>,
+}
+
+/// Builds a projection mask plan for reading a parquet file.
+///
+/// `parquet_read_cols` defines the requested root columns and optional
+/// nested paths to read.
+///
+/// `parquet_schema_desc` is the schema descriptor of the current parquet
+/// file. It is used to resolve requested nested paths to actual leaf
+/// column indices.
+///
+/// See [`ProjectionMaskPlan`] for the returned value.
+///
+/// For example, if the query requests `j.a` and `k`, but the current
+/// parquet file only contains leaves under `j.b` and `k`, then the
+/// returned plan keeps `k` in the projection mask and marks `j` as
+/// not present in the output, so it can be synthesized during
+/// post-processing.
+pub fn build_projection_plan(
+    parquet_read_cols: &ParquetReadColumns,
+    parquet_schema_desc: &SchemaDescriptor,
+) -> ProjectionMaskPlan {
+    if !parquet_read_cols.has_nested() {
+        let mask =
+            ProjectionMask::roots(parquet_schema_desc, parquet_read_cols.root_indices_iter());
+        return ProjectionMaskPlan {
+            mask,
+            projected_root_presence: vec![true; parquet_read_cols.columns().len()],
+        };
+    }
+
+    let (leaf_indices, matched_roots) =
+        build_parquet_leaves_indices(parquet_schema_desc, parquet_read_cols);
+
+    let projected_root_presence = parquet_read_cols
+        .columns()
+        .iter()
+        .map(|col| matched_roots.contains(&col.root_index()))
+        .collect();
+
+    let mask = ProjectionMask::leaves(parquet_schema_desc, leaf_indices);
+    ProjectionMaskPlan {
+        mask,
+        projected_root_presence,
+    }
+}
+
+/// Builds parquet leaf-column indices for reading a parquet file.
+///
+/// Returns `(leaf_indices, matched_roots)`:
+/// - `leaf_indices`: matched parquet leaf column indices
+/// - `matched_roots`: root column indices that match at least one leaf in the
+///   current parquet schema.
+fn build_parquet_leaves_indices(
+    parquet_schema_desc: &SchemaDescriptor,
+    projection: &ParquetReadColumns,
+) -> (Vec<usize>, HashSet<usize>) {
+    let mut map = HashMap::with_capacity(projection.cols.len());
+    for col in &projection.cols {
+        map.insert(col.root_index, &col.nested_paths);
+    }
+
+    let mut leaf_indices = Vec::new();
+    let mut matched_roots = HashSet::with_capacity(projection.cols.len());
+    for (leaf_idx, leaf_col) in parquet_schema_desc.columns().iter().enumerate() {
+        let root_idx = parquet_schema_desc.get_column_root_idx(leaf_idx);
+        let Some(nested_paths) = map.get(&root_idx) else {
+            continue;
+        };
+        if nested_paths.is_empty() {
+            leaf_indices.push(leaf_idx);
+            matched_roots.insert(root_idx);
+            continue;
+        }
+
+        let leaf_path = leaf_col.path().parts();
+        if nested_paths
+            .iter()
+            .any(|nested_path| leaf_path.starts_with(nested_path))
+        {
+            leaf_indices.push(leaf_idx);
+            matched_roots.insert(root_idx);
+        }
+    }
+    (leaf_indices, matched_roots)
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use parquet::basic::Repetition;
+    use parquet::schema::types::Type;
+
+    use super::*;
+
+    #[test]
+    fn test_build_projection_mask_without_nested_paths() {
+        let parquet_schema_desc = build_test_nested_parquet_schema();
+        let projection = ParquetReadColumns::from_deduped_root_indices([0, 1]);
+
+        let plan = build_projection_plan(&projection, &parquet_schema_desc);
+
+        assert_eq!(vec![true, true], plan.projected_root_presence);
+        assert_eq!(
+            ProjectionMask::roots(&parquet_schema_desc, [0, 1]),
+            plan.mask
+        );
+    }
+
+    #[test]
+    fn test_reads_whole_root() {
+        let parquet_schema_desc = build_test_nested_parquet_schema();
+
+        let projection = ParquetReadColumns::from_deduped(vec![ParquetReadColumn {
+            root_index: 0,
+            nested_paths: vec![],
+        }]);
+
+        let (leaf_indices, matched_roots) =
+            build_parquet_leaves_indices(&parquet_schema_desc, &projection);
+        assert_eq!(vec![0, 1, 2], leaf_indices);
+        assert_eq!(HashSet::from([0]), matched_roots);
+    }
+
+    #[test]
+    fn test_filters_nested_paths() {
+        let parquet_schema_desc = build_test_nested_parquet_schema();
+
+        let projection = ParquetReadColumns::from_deduped(vec![
+            ParquetReadColumn {
+                root_index: 0,
+                nested_paths: vec![vec!["j".to_string(), "b".to_string()]],
+            },
+            ParquetReadColumn {
+                root_index: 1,
+                nested_paths: vec![],
+            },
+        ]);
+
+        let (leaf_indices, matched_roots) =
+            build_parquet_leaves_indices(&parquet_schema_desc, &projection);
+        assert_eq!(vec![1, 2, 3], leaf_indices);
+        assert_eq!(HashSet::from([0, 1]), matched_roots);
+    }
+
+    #[test]
+    fn test_reads_middle_level_path() {
+        let parquet_schema_desc = build_test_nested_parquet_schema();
+
+        let projection = ParquetReadColumns::from_deduped(vec![ParquetReadColumn {
+            root_index: 0,
+            nested_paths: vec![vec!["j".to_string(), "b".to_string()]],
+        }]);
+
+        let (leaf_indices, matched_roots) =
+            build_parquet_leaves_indices(&parquet_schema_desc, &projection);
+        assert_eq!(vec![1, 2], leaf_indices);
+        assert_eq!(HashSet::from([0]), matched_roots);
+    }
+
+    #[test]
+    fn test_reads_leaf_level_path() {
+        let parquet_schema_desc = build_test_nested_parquet_schema();
+
+        let projection = ParquetReadColumns::from_deduped(vec![ParquetReadColumn {
+            root_index: 0,
+            nested_paths: vec![vec!["j".to_string(), "b".to_string(), "c".to_string()]],
+        }]);
+
+        let (leaf_indices, matched_roots) =
+            build_parquet_leaves_indices(&parquet_schema_desc, &projection);
+        assert_eq!(vec![1], leaf_indices);
+        assert_eq!(HashSet::from([0]), matched_roots);
+    }
+
+    #[test]
+    fn test_build_projection_mask_with_unmatched_roots() {
+        let parquet_schema_desc = build_test_nested_parquet_schema();
+
+        let projection = ParquetReadColumns::from_deduped(vec![
+            ParquetReadColumn {
+                root_index: 0,
+                nested_paths: vec![vec!["j".to_string(), "missing".to_string()]],
+            },
+            ParquetReadColumn {
+                root_index: 1,
+                nested_paths: vec![],
+            },
+        ]);
+
+        let plan = build_projection_plan(&projection, &parquet_schema_desc);
+
+        assert_eq!(vec![false, true], plan.projected_root_presence);
+        assert_eq!(
+            ProjectionMask::leaves(&parquet_schema_desc, vec![3]),
+            plan.mask
+        );
+    }
+
+    #[test]
+    fn test_merges_mixed_paths() {
+        let parquet_schema_desc = build_test_nested_parquet_schema();
+
+        let projection = ParquetReadColumns::from_deduped(vec![ParquetReadColumn {
+            root_index: 0,
+            nested_paths: vec![
+                vec!["j".to_string(), "a".to_string()],
+                vec!["j".to_string(), "b".to_string(), "d".to_string()],
+            ],
+        }]);
+
+        let (leaf_indices, matched_roots) =
+            build_parquet_leaves_indices(&parquet_schema_desc, &projection);
+        assert_eq!(vec![0, 2], leaf_indices);
+        assert_eq!(HashSet::from([0]), matched_roots);
+    }
+
+    #[test]
+    fn test_merge_nested_paths_extends_paths() {
+        let mut col = ParquetReadColumn::new(0)
+            .with_nested_paths(vec![vec!["j".to_string(), "a".to_string()]]);
+
+        col.merge_nested_paths(vec![vec!["j".to_string(), "b".to_string()]]);
+
+        assert_eq!(
+            &[
+                vec!["j".to_string(), "a".to_string()],
+                vec!["j".to_string(), "b".to_string()],
+            ],
+            col.nested_paths()
+        );
+    }
+
+    #[test]
+    fn test_merge_nested_paths_with_whole_root() {
+        let mut col = ParquetReadColumn::new(0)
+            .with_nested_paths(vec![vec!["j".to_string(), "a".to_string()]]);
+
+        col.merge_nested_paths(vec![]);
+
+        assert!(col.nested_paths().is_empty());
+    }
+
+    // Test schema:
+    // schema
+    // |- j
+    // |  |- a: INT64
+    // |  `- b
+    // |     |- c: INT64
+    // |     `- d: INT64
+    // `- k: INT64
+    fn build_test_nested_parquet_schema() -> SchemaDescriptor {
+        let leaf_a = Arc::new(
+            Type::primitive_type_builder("a", parquet::basic::Type::INT64)
+                .with_repetition(Repetition::REQUIRED)
+                .build()
+                .unwrap(),
+        );
+        let leaf_c = Arc::new(
+            Type::primitive_type_builder("c", parquet::basic::Type::INT64)
+                .with_repetition(Repetition::REQUIRED)
+                .build()
+                .unwrap(),
+        );
+        let leaf_d = Arc::new(
+            Type::primitive_type_builder("d", parquet::basic::Type::INT64)
+                .with_repetition(Repetition::REQUIRED)
+                .build()
+                .unwrap(),
+        );
+        let group_b = Arc::new(
+            Type::group_type_builder("b")
+                .with_repetition(Repetition::REQUIRED)
+                .with_fields(vec![leaf_c, leaf_d])
+                .build()
+                .unwrap(),
+        );
+        let root_j = Arc::new(
+            Type::group_type_builder("j")
+                .with_repetition(Repetition::REQUIRED)
+                .with_fields(vec![leaf_a, group_b])
+                .build()
+                .unwrap(),
+        );
+        let root_k = Arc::new(
+            Type::primitive_type_builder("k", parquet::basic::Type::INT64)
+                .with_repetition(Repetition::REQUIRED)
+                .build()
+                .unwrap(),
+        );
+        let schema = Arc::new(
+            Type::group_type_builder("schema")
+                .with_fields(vec![root_j, root_k])
+                .build()
+                .unwrap(),
+        );
+
+        SchemaDescriptor::new(schema)
+    }
+}
diff --git a/src/mito2/src/sst/parquet/reader.rs b/src/mito2/src/sst/parquet/reader.rs
index 500f32ae91..f03856a521 100644
--- a/src/mito2/src/sst/parquet/reader.rs
+++ b/src/mito2/src/sst/parquet/reader.rs
@@ -14,51 +14,54 @@
 
 //! Parquet reader.
 
+mod stream;
+
 #[cfg(feature = "vector_index")]
 use std::collections::BTreeSet;
-use std::collections::{HashSet, VecDeque};
+use std::collections::HashSet;
 use std::sync::Arc;
 use std::time::{Duration, Instant};
 
 use api::v1::SemanticType;
-use async_trait::async_trait;
 use common_recordbatch::filter::SimpleFilterEvaluator;
-use common_telemetry::{debug, tracing, warn};
+use common_telemetry::{error, tracing, warn};
+use datafusion::physical_plan::PhysicalExpr;
 use datafusion_expr::Expr;
+use datafusion_expr::utils::expr_to_columns;
 use datatypes::arrow::array::ArrayRef;
-use datatypes::arrow::datatypes::Field;
-use datatypes::arrow::error::ArrowError;
+use datatypes::arrow::datatypes::{Field, Schema as ArrowSchema, SchemaRef};
 use datatypes::arrow::record_batch::RecordBatch;
 use datatypes::data_type::ConcreteDataType;
 use datatypes::prelude::DataType;
+use datatypes::schema::ext::ArrowSchemaExt;
+use futures::StreamExt;
 use mito_codec::row_converter::build_primary_key_codec;
 use object_store::ObjectStore;
-use parquet::arrow::arrow_reader::{ParquetRecordBatchReader, RowSelection};
-use parquet::arrow::{FieldLevels, ProjectionMask, parquet_to_arrow_field_levels};
-use parquet::file::metadata::{KeyValue, PageIndexPolicy, ParquetMetaData};
+use parquet::arrow::ProjectionMask;
+use parquet::arrow::arrow_reader::{ArrowReaderMetadata, ArrowReaderOptions, RowSelection};
+use parquet::arrow::async_reader::{ParquetRecordBatchStream, ParquetRecordBatchStreamBuilder};
+use parquet::file::metadata::{PageIndexPolicy, ParquetMetaData};
 use partition::expr::PartitionExpr;
-use snafu::{OptionExt, ResultExt};
+use snafu::ResultExt;
 use store_api::codec::PrimaryKeyEncoding;
 use store_api::metadata::{ColumnMetadata, RegionMetadata, RegionMetadataRef};
 use store_api::region_request::PathType;
 use store_api::storage::{ColumnId, FileId};
 use table::predicate::Predicate;
 
-use crate::cache::CacheStrategy;
+use self::stream::{NestedSchemaAligner, ParquetErrorAdapter, ProjectedRecordBatchStream};
 use crate::cache::index::result_cache::PredicateKey;
+use crate::cache::{CacheStrategy, CachedSstMeta};
 #[cfg(feature = "vector_index")]
 use crate::error::ApplyVectorIndexSnafu;
-use crate::error::{
-    ArrowReaderSnafu, InvalidMetadataSnafu, InvalidParquetSnafu, ReadDataPartSnafu,
-    ReadParquetSnafu, Result, SerializePartitionExprSnafu,
-};
+use crate::error::{ReadDataPartSnafu, ReadParquetSnafu, Result, SerializePartitionExprSnafu};
 use crate::metrics::{
     PRECISE_FILTER_ROWS_TOTAL, READ_ROW_GROUPS_TOTAL, READ_ROWS_IN_ROW_GROUP_TOTAL,
     READ_ROWS_TOTAL, READ_STAGE_ELAPSED,
 };
 use crate::read::flat_projection::CompactionProjectionMapper;
-use crate::read::prune::{PruneReader, Source};
-use crate::read::{Batch, BatchReader};
+use crate::read::prune::FlatPruneReader;
+use crate::read::read_columns::ReadColumns;
 use crate::sst::file::FileHandle;
 use crate::sst::index::bloom_filter::applier::{
     BloomFilterIndexApplierRef, BloomFilterIndexApplyMetrics,
@@ -71,16 +74,21 @@ use crate::sst::index::inverted_index::applier::{
 };
 #[cfg(feature = "vector_index")]
 use crate::sst::index::vector_index::applier::VectorIndexApplierRef;
+use crate::sst::parquet::DEFAULT_READ_BATCH_SIZE;
+use crate::sst::parquet::async_reader::SstAsyncFileReader;
 use crate::sst::parquet::file_range::{
     FileRangeContext, FileRangeContextRef, PartitionFilterContext, PreFilterMode, RangeBase,
-    row_group_contains_delete,
 };
-use crate::sst::parquet::format::{ReadFormat, need_override_sequence};
+use crate::sst::parquet::flat_format::FlatReadFormat;
+use crate::sst::parquet::format::need_override_sequence;
 use crate::sst::parquet::metadata::MetadataLoader;
-use crate::sst::parquet::row_group::{InMemoryRowGroup, ParquetFetchMetrics};
+use crate::sst::parquet::prefilter::{
+    PrefilterContextBuilder, build_reader_filter_plan, execute_prefilter,
+};
+use crate::sst::parquet::read_columns::{ProjectionMaskPlan, build_projection_plan};
+use crate::sst::parquet::row_group::ParquetFetchMetrics;
 use crate::sst::parquet::row_selection::RowGroupSelection;
 use crate::sst::parquet::stats::RowGroupPruningStats;
-use crate::sst::parquet::{DEFAULT_READ_BATCH_SIZE, PARQUET_METADATA_KEY};
 use crate::sst::tag_maybe_to_dictionary_field;
 
 const INDEX_TYPE_FULLTEXT: &str = "fulltext";
@@ -119,11 +127,11 @@ pub struct ParquetReaderBuilder {
     object_store: ObjectStore,
     /// Predicate to push down.
     predicate: Option<Predicate>,
-    /// Metadata of columns to read.
+    /// The columns to read.
     ///
     /// `None` reads all columns. Due to schema change, the projection
     /// can contain columns not in the parquet file.
-    projection: Option<Vec<ColumnId>>,
+    read_cols: Option<ReadColumns>,
     /// Strategy to cache SST data.
     cache_strategy: CacheStrategy,
     /// Index appliers.
@@ -140,8 +148,6 @@ pub struct ParquetReaderBuilder {
     /// This is usually the latest metadata of the region. The reader use
     /// it get the correct column id of a column by name.
     expected_metadata: Option<RegionMetadataRef>,
-    /// Whether to use flat format for reading.
-    flat_format: bool,
     /// Whether this reader is for compaction.
     compaction: bool,
     /// Mode to pre-filter columns.
@@ -165,7 +171,7 @@ impl ParquetReaderBuilder {
             file_handle,
             object_store,
             predicate: None,
-            projection: None,
+            read_cols: None,
             cache_strategy: CacheStrategy::Disabled,
             inverted_index_appliers: [None, None],
             bloom_filter_index_appliers: [None, None],
@@ -175,7 +181,6 @@ impl ParquetReaderBuilder {
             #[cfg(feature = "vector_index")]
             vector_index_k: None,
             expected_metadata: None,
-            flat_format: false,
             compaction: false,
             pre_filter_mode: PreFilterMode::All,
             decode_primary_key_values: false,
@@ -194,8 +199,8 @@ impl ParquetReaderBuilder {
     ///
     /// The reader only applies the projection to fields.
     #[must_use]
-    pub fn projection(mut self, projection: Option<Vec<ColumnId>>) -> ParquetReaderBuilder {
-        self.projection = projection;
+    pub fn projection(mut self, read_cols: Option<ReadColumns>) -> ParquetReaderBuilder {
+        self.read_cols = read_cols;
         self
     }
 
@@ -256,13 +261,6 @@ impl ParquetReaderBuilder {
         self
     }
 
-    /// Sets the flat format flag.
-    #[must_use]
-    pub fn flat_format(mut self, flat_format: bool) -> Self {
-        self.flat_format = flat_format;
-        self
-    }
-
     /// Sets the compaction flag.
     #[must_use]
     pub fn compaction(mut self, compaction: bool) -> Self {
@@ -303,7 +301,7 @@ impl ParquetReaderBuilder {
     pub async fn build(&self) -> Result<Option<ParquetReader>> {
         let mut metrics = ReaderMetrics::default();
 
-        let Some((context, selection)) = self.build_reader_input(&mut metrics).await? else {
+        let Some((context, selection)) = self.build_reader_input_inner(&mut metrics).await? else {
             return Ok(None);
         };
         ParquetReader::new(Arc::new(context), selection)
@@ -338,7 +336,7 @@ impl ParquetReaderBuilder {
         let file_size = self.file_handle.meta_ref().file_size;
 
         // Loads parquet metadata of the file.
-        let (parquet_meta, cache_miss) = self
+        let (sst_meta, cache_miss) = self
             .read_parquet_metadata(
                 &file_path,
                 file_size,
@@ -346,9 +344,8 @@ impl ParquetReaderBuilder {
                 self.page_index_policy,
             )
             .await?;
-        // Decodes region metadata.
-        let key_value_meta = parquet_meta.file_metadata().key_value_metadata();
-        let region_meta = Arc::new(Self::get_region_metadata(&file_path, key_value_meta)?);
+        let parquet_meta = sst_meta.parquet_metadata();
+        let region_meta = sst_meta.region_metadata();
         let region_partition_expr_str = self
             .expected_metadata
             .as_ref()
@@ -373,7 +370,6 @@ impl ParquetReaderBuilder {
         // before compat handling.
         let compaction_projection_mapper = if self.compaction
             && !is_same_region_partition
-            && self.flat_format
             && region_meta.primary_key_encoding == PrimaryKeyEncoding::Sparse
         {
             Some(CompactionProjectionMapper::try_new(&region_meta)?)
@@ -381,40 +377,34 @@ impl ParquetReaderBuilder {
             None
         };
 
-        let mut read_format = if let Some(column_ids) = &self.projection {
-            ReadFormat::new(
-                region_meta.clone(),
-                Some(column_ids),
-                self.flat_format,
-                Some(parquet_meta.file_metadata().schema_descr().num_columns()),
-                &file_path,
-                skip_auto_convert,
-            )?
+        let read_cols = if let Some(read_cols) = &self.read_cols {
+            read_cols.clone()
         } else {
-            // Lists all column ids to read, we always use the expected metadata if possible.
             let expected_meta = self.expected_metadata.as_ref().unwrap_or(&region_meta);
-            let column_ids: Vec<_> = expected_meta
-                .column_metadatas
-                .iter()
-                .map(|col| col.column_id)
-                .collect();
-            ReadFormat::new(
-                region_meta.clone(),
-                Some(&column_ids),
-                self.flat_format,
-                Some(parquet_meta.file_metadata().schema_descr().num_columns()),
-                &file_path,
-                skip_auto_convert,
-            )?
+            // Lists all column ids to read, we always use the expected metadata if possible.
+            ReadColumns::from_deduped_column_ids(
+                expected_meta
+                    .column_metadatas
+                    .iter()
+                    .map(|col| col.column_id),
+            )
         };
-        if self.decode_primary_key_values {
-            read_format.set_decode_primary_key_values(true);
-        }
+        let mut read_format = FlatReadFormat::new(
+            region_meta.clone(),
+            read_cols,
+            Some(parquet_meta.file_metadata().schema_descr().num_columns()),
+            &file_path,
+            skip_auto_convert,
+        )?;
         if need_override_sequence(&parquet_meta) {
             read_format
                 .set_override_sequence(self.file_handle.meta_ref().sequence.map(|x| x.get()));
         }
 
+        // Computes the projection mask.
+        let parquet_schema_desc = parquet_meta.file_metadata().schema_descr();
+        let parquet_read_cols = read_format.parquet_read_columns();
+        let projection_plan = build_projection_plan(parquet_read_cols, parquet_schema_desc);
         let selection = self
             .row_groups_to_read(&read_format, &parquet_meta, &mut metrics.filter_metrics)
             .await;
@@ -446,45 +436,16 @@ impl ParquetReaderBuilder {
             .map(|meta| meta.schema.clone())
             .unwrap_or_else(|| region_meta.schema.clone());
 
-        // Computes the projection mask.
-        let parquet_schema_desc = parquet_meta.file_metadata().schema_descr();
-        let indices = read_format.projection_indices();
-        // Now we assumes we don't have nested schemas.
-        // TODO(yingwen): Revisit this if we introduce nested types such as JSON type.
-        let projection_mask = ProjectionMask::roots(parquet_schema_desc, indices.iter().copied());
-
-        // Computes the field levels.
-        let hint = Some(read_format.arrow_schema().fields());
-        let field_levels =
-            parquet_to_arrow_field_levels(parquet_schema_desc, projection_mask.clone(), hint)
+        // Create ArrowReaderMetadata for async stream building.
+        let mut arrow_reader_options = ArrowReaderOptions::new();
+        if !read_format.arrow_schema().has_json_extension_field() {
+            arrow_reader_options =
+                arrow_reader_options.with_schema(read_format.arrow_schema().clone());
+        }
+        let arrow_metadata =
+            ArrowReaderMetadata::try_new(parquet_meta.clone(), arrow_reader_options)
                 .context(ReadDataPartSnafu)?;
 
-        let reader_builder = RowGroupReaderBuilder {
-            file_handle: self.file_handle.clone(),
-            file_path,
-            parquet_meta,
-            object_store: self.object_store.clone(),
-            projection: projection_mask,
-            field_levels,
-            cache_strategy: self.cache_strategy.clone(),
-        };
-
-        let filters = if let Some(predicate) = &self.predicate {
-            predicate
-                .exprs()
-                .iter()
-                .filter_map(|expr| {
-                    SimpleFilterContext::new_opt(
-                        &region_meta,
-                        self.expected_metadata.as_deref(),
-                        expr,
-                    )
-                })
-                .collect::<Vec<_>>()
-        } else {
-            vec![]
-        };
-
         let dyn_filters = if let Some(predicate) = &self.predicate {
             predicate.dyn_filters().as_ref().clone()
         } else {
@@ -493,12 +454,36 @@ impl ParquetReaderBuilder {
 
         let codec = build_primary_key_codec(read_format.metadata());
 
+        let filter_plan = build_reader_filter_plan(
+            self.predicate.as_ref(),
+            self.expected_metadata.as_deref(),
+            self.pre_filter_mode,
+            &read_format,
+            parquet_meta.file_metadata().schema_descr(),
+            &codec,
+        );
+
+        let output_schema = read_format.output_arrow_schema()?;
+
+        let reader_builder = RowGroupReaderBuilder {
+            file_handle: self.file_handle.clone(),
+            file_path,
+            parquet_meta,
+            arrow_metadata,
+            output_schema,
+            object_store: self.object_store.clone(),
+            projection: projection_plan,
+            has_nested_projection: parquet_read_cols.has_nested(),
+            cache_strategy: self.cache_strategy.clone(),
+            prefilter_builder: filter_plan.prefilter_builder,
+        };
+
         let partition_filter = self.build_partition_filter(&read_format, &prune_schema)?;
 
         let context = FileRangeContext::new(
             reader_builder,
             RangeBase {
-                filters,
+                filters: filter_plan.remaining_simple_filters,
                 dyn_filters,
                 read_format,
                 expected_metadata: self.expected_metadata.clone(),
@@ -533,7 +518,7 @@ impl ParquetReaderBuilder {
     /// and build a partition filter if they differ.
     fn build_partition_filter(
         &self,
-        read_format: &ReadFormat,
+        read_format: &FlatReadFormat,
         prune_schema: &Arc<datatypes::schema::Schema>,
     ) -> Result<Option<PartitionFilterContext>> {
         let region_partition_expr_str = self
@@ -560,15 +545,13 @@ impl ParquetReaderBuilder {
         region_partition_expr.collect_column_names(&mut referenced_columns);
 
         // Build a partition_schema containing only referenced columns.
-        let is_flat = read_format.as_flat().is_some();
         let partition_schema = Arc::new(datatypes::schema::Schema::new(
             prune_schema
                 .column_schemas()
                 .iter()
                 .filter(|col| referenced_columns.contains(&col.name))
                 .map(|col| {
-                    if is_flat
-                        && let Some(column_meta) = read_format.metadata().column_by_name(&col.name)
+                    if let Some(column_meta) = read_format.metadata().column_by_name(&col.name)
                         && column_meta.semantic_type == SemanticType::Tag
                         && col.data_type.is_string()
                     {
@@ -599,42 +582,15 @@ impl ParquetReaderBuilder {
         }))
     }
 
-    /// Decodes region metadata from key value.
-    fn get_region_metadata(
-        file_path: &str,
-        key_value_meta: Option<&Vec<KeyValue>>,
-    ) -> Result<RegionMetadata> {
-        let key_values = key_value_meta.context(InvalidParquetSnafu {
-            file: file_path,
-            reason: "missing key value meta",
-        })?;
-        let meta_value = key_values
-            .iter()
-            .find(|kv| kv.key == PARQUET_METADATA_KEY)
-            .with_context(|| InvalidParquetSnafu {
-                file: file_path,
-                reason: format!("key {} not found", PARQUET_METADATA_KEY),
-            })?;
-        let json = meta_value
-            .value
-            .as_ref()
-            .with_context(|| InvalidParquetSnafu {
-                file: file_path,
-                reason: format!("No value for key {}", PARQUET_METADATA_KEY),
-            })?;
-
-        RegionMetadata::from_json(json).context(InvalidMetadataSnafu)
-    }
-
     /// Reads parquet metadata of specific file.
-    /// Returns (metadata, cache_miss_flag).
+    /// Returns (fused metadata, cache_miss_flag).
     async fn read_parquet_metadata(
         &self,
         file_path: &str,
         file_size: u64,
         cache_metrics: &mut MetadataCacheMetrics,
         page_index_policy: PageIndexPolicy,
-    ) -> Result<(Arc<ParquetMetaData>, bool)> {
+    ) -> Result<(Arc<CachedSstMeta>, bool)> {
         let start = Instant::now();
         let _t = READ_STAGE_ELAPSED
             .with_label_values(&["read_parquet_metadata"])
@@ -644,7 +600,7 @@ impl ParquetReaderBuilder {
         // Tries to get from cache with metrics tracking.
         if let Some(metadata) = self
             .cache_strategy
-            .get_parquet_meta_data(file_id, cache_metrics, page_index_policy)
+            .get_sst_meta_data(file_id, cache_metrics, page_index_policy)
             .await
         {
             cache_metrics.metadata_load_cost += start.elapsed();
@@ -657,10 +613,10 @@ impl ParquetReaderBuilder {
         metadata_loader.with_page_index_policy(page_index_policy);
         let metadata = metadata_loader.load(cache_metrics).await?;
 
-        let metadata = Arc::new(metadata);
+        let metadata = Arc::new(CachedSstMeta::try_new(file_path, metadata)?);
         // Cache the metadata.
         self.cache_strategy
-            .put_parquet_meta_data(file_id, metadata.clone());
+            .put_sst_meta_data(file_id, metadata.clone());
 
         cache_metrics.metadata_load_cost += start.elapsed();
         Ok((metadata, true))
@@ -676,7 +632,7 @@ impl ParquetReaderBuilder {
     )]
     async fn row_groups_to_read(
         &self,
-        read_format: &ReadFormat,
+        read_format: &FlatReadFormat,
         parquet_meta: &ParquetMetaData,
         metrics: &mut ReaderFilterMetrics,
     ) -> RowGroupSelection {
@@ -697,7 +653,7 @@ impl ParquetReaderBuilder {
         metrics.rows_total += num_rows as usize;
 
         // Compute skip_fields once for all pruning operations
-        let skip_fields = self.compute_skip_fields(parquet_meta);
+        let skip_fields = self.pre_filter_mode.skip_fields();
 
         let mut output = self.row_groups_by_minmax(
             read_format,
@@ -725,6 +681,7 @@ impl ParquetReaderBuilder {
         }
 
         self.prune_row_groups_by_inverted_index(
+            read_format.metadata(),
             row_group_size,
             num_row_groups,
             &mut output,
@@ -851,6 +808,7 @@ impl ParquetReaderBuilder {
     /// the correctness of the index.
     async fn prune_row_groups_by_inverted_index(
         &self,
+        sst_metadata: &RegionMetadataRef,
         row_group_size: usize,
         num_row_groups: usize,
         output: &mut RowGroupSelection,
@@ -869,12 +827,19 @@ impl ParquetReaderBuilder {
             &self.inverted_index_appliers[..]
         };
         for index_applier in appliers.iter().flatten() {
-            let predicate_key = index_applier.predicate_key();
+            let Ok(Some(plan)) = index_applier
+                .plan_for_sst(sst_metadata)
+                .inspect_err(|e| warn!(e; "failed to build compatible plan for sst"))
+            else {
+                continue;
+            };
+
             // Fast path: return early if the result is in the cache.
-            let cached = self
-                .cache_strategy
-                .index_result_cache()
-                .and_then(|cache| cache.get(predicate_key, self.file_handle.file_id().file_id()));
+            let cached = self.cache_strategy.index_result_cache().and_then(|cache| {
+                let file_id = self.file_handle.file_id().file_id();
+                cache.get(&plan.predicate_key, file_id)
+            });
+
             if let Some(result) = cached.as_ref()
                 && all_required_row_groups_searched(output, result)
             {
@@ -891,9 +856,11 @@ impl ParquetReaderBuilder {
                 .apply(
                     self.file_handle.index_id(),
                     Some(file_size_hint),
+                    &plan.index_applier,
                     metrics.inverted_index_apply_metrics.as_mut(),
                 )
                 .await;
+
             let selection = match apply_res {
                 Ok(apply_output) => RowGroupSelection::from_inverted_index_apply_output(
                     row_group_size,
@@ -907,7 +874,7 @@ impl ParquetReaderBuilder {
             };
 
             self.apply_index_result_and_update_cache(
-                predicate_key,
+                &plan.predicate_key,
                 self.file_handle.file_id().file_id(),
                 selection,
                 output,
@@ -1133,29 +1100,10 @@ impl ParquetReaderBuilder {
         pruned
     }
 
-    /// Computes whether to skip field columns when building statistics based on PreFilterMode.
-    fn compute_skip_fields(&self, parquet_meta: &ParquetMetaData) -> bool {
-        match self.pre_filter_mode {
-            PreFilterMode::All => false,
-            PreFilterMode::SkipFields => true,
-            PreFilterMode::SkipFieldsOnDelete => {
-                // Check if any row group contains delete op
-                let file_path = self.file_handle.file_path(&self.table_dir, self.path_type);
-                (0..parquet_meta.num_row_groups()).any(|rg_idx| {
-                    row_group_contains_delete(parquet_meta, rg_idx, &file_path)
-                        .inspect_err(|e| {
-                            warn!(e; "Failed to decode min value of op_type, fallback to not skipping fields");
-                        })
-                        .unwrap_or(false)
-                })
-            }
-        }
-    }
-
     /// Computes row groups selection after min-max pruning.
     fn row_groups_by_minmax(
         &self,
-        read_format: &ReadFormat,
+        read_format: &FlatReadFormat,
         parquet_meta: &ParquetMetaData,
         row_group_size: usize,
         total_row_count: usize,
@@ -1667,7 +1615,7 @@ impl ReaderMetrics {
     }
 }
 
-/// Builder to build a [ParquetRecordBatchReader] for a row group.
+/// Builder to build a [ParquetRecordBatchStream] for a row group.
 pub(crate) struct RowGroupReaderBuilder {
     /// SST file to read.
     ///
@@ -1677,14 +1625,31 @@ pub(crate) struct RowGroupReaderBuilder {
     file_path: String,
     /// Metadata of the parquet file.
     parquet_meta: Arc<ParquetMetaData>,
+    /// Arrow reader metadata for building async stream.
+    arrow_metadata: ArrowReaderMetadata,
+    /// Projected output schema aligned with `projection.projected_root_presence`.
+    output_schema: SchemaRef,
     /// Object store as an Operator.
     object_store: ObjectStore,
     /// Projection mask.
-    projection: ProjectionMask,
-    /// Field levels to read.
-    field_levels: FieldLevels,
+    projection: ProjectionMaskPlan,
+    /// Whether projected read columns include nested paths.
+    has_nested_projection: bool,
     /// Cache.
     cache_strategy: CacheStrategy,
+    /// Pre-built prefilter state. `None` if prefiltering is not applicable.
+    prefilter_builder: Option<PrefilterContextBuilder>,
+}
+
+/// Context passed to [RowGroupReaderBuilder::build()] carrying all information
+/// needed for prefiltering decisions.
+pub(crate) struct RowGroupBuildContext<'a> {
+    /// Index of the row group to read.
+    pub(crate) row_group_idx: usize,
+    /// Row selection for the row group. `None` means all rows.
+    pub(crate) row_selection: Option<RowSelection>,
+    /// Metrics for tracking fetch operations.
+    pub(crate) fetch_metrics: Option<&'a ParquetFetchMetrics>,
 }
 
 impl RowGroupReaderBuilder {
@@ -1706,69 +1671,127 @@ impl RowGroupReaderBuilder {
         &self.cache_strategy
     }
 
-    /// Builds a [ParquetRecordBatchReader] to read the row group at `row_group_idx`.
+    pub(crate) fn has_predicate_prefilter(&self) -> bool {
+        self.prefilter_builder.is_some()
+    }
+
+    /// Builds a [ParquetRecordBatchStream] to read the row group at `row_group_idx`.
+    ///
+    /// If prefiltering is applicable (based on `build_ctx`), this performs a two-phase read:
+    /// 1. Reads only the prefilter columns (e.g. PK column), applies filters to get a refined row selection
+    /// 2. Reads the full projection with the refined row selection
+    ///
+    /// The prefilter pass is *best-effort pruning*, not the precise filter for the query.
+    /// Predicates that cannot be lowered to prefilter columns (column not projected,
+    /// expression not supported, etc.) are silently skipped. Correctness rests on the
+    /// DataFusion `FilterExec` above this reader, which always re-applies the original
+    /// predicate. Tag and timestamp predicates that flow through [`SimpleFilterEvaluator`]
+    /// are an exception — the engine enforces them precisely, so the prefilter pass is the
+    /// only place they execute. See [`build_reader_filter_plan`] for the bucketing rules.
+    ///
+    /// When the prefilter result selects no rows, the second read still issues but
+    /// parquet-rs short-circuits before any column-chunk IO: the row-group state machine
+    /// jumps to `Finished` once it sees `num_rows_selected() == 0`, so no fast path is
+    /// added here.
     pub(crate) async fn build(
+        &self,
+        build_ctx: RowGroupBuildContext<'_>,
+    ) -> Result<ProjectedRecordBatchStream> {
+        let prefilter_ctx = self.prefilter_builder.as_ref().map(|b| b.build());
+
+        let Some(mut prefilter_ctx) = prefilter_ctx else {
+            // No prefilter applicable, build stream with full projection.
+            let stream = self
+                .build_with_projection(
+                    build_ctx.row_group_idx,
+                    build_ctx.row_selection,
+                    self.projection.mask.clone(),
+                    build_ctx.fetch_metrics,
+                )
+                .await?;
+            return self.make_projected_stream(stream);
+        };
+
+        let prefilter_start = Instant::now();
+        let prefilter_result = execute_prefilter(&mut prefilter_ctx, self, &build_ctx).await?;
+        if let Some(metrics) = build_ctx.fetch_metrics {
+            let mut data = metrics.data.lock().unwrap();
+            data.prefilter_cost += prefilter_start.elapsed();
+            data.prefilter_filtered_rows += prefilter_result.filtered_rows;
+        }
+
+        let refined_selection = Some(prefilter_result.refined_selection);
+
+        let stream = self
+            .build_with_projection(
+                build_ctx.row_group_idx,
+                refined_selection,
+                self.projection.mask.clone(),
+                build_ctx.fetch_metrics,
+            )
+            .await?;
+        self.make_projected_stream(stream)
+    }
+
+    fn make_projected_stream(
+        &self,
+        stream: ParquetRecordBatchStream<SstAsyncFileReader>,
+    ) -> Result<ProjectedRecordBatchStream> {
+        let stream = ParquetErrorAdapter::new(stream, self.file_path.clone());
+        if !self.has_nested_projection {
+            return Ok(stream.boxed());
+        }
+
+        Ok(NestedSchemaAligner::new(
+            stream,
+            self.projection.projected_root_presence.clone(),
+            self.output_schema.clone(),
+        )?
+        .boxed())
+    }
+
+    /// Builds a [ParquetRecordBatchStream] with a custom projection mask.
+    pub(crate) async fn build_with_projection(
         &self,
         row_group_idx: usize,
         row_selection: Option<RowSelection>,
+        projection: ProjectionMask,
         fetch_metrics: Option<&ParquetFetchMetrics>,
-    ) -> Result<ParquetRecordBatchReader> {
-        let fetch_start = Instant::now();
-
-        let mut row_group = InMemoryRowGroup::create(
-            self.file_handle.region_id(),
-            self.file_handle.file_id().file_id(),
-            &self.parquet_meta,
-            row_group_idx,
-            self.cache_strategy.clone(),
-            &self.file_path,
+    ) -> Result<ParquetRecordBatchStream<SstAsyncFileReader>> {
+        // Create async file reader with caching support.
+        let async_reader = SstAsyncFileReader::new(
+            self.file_handle.file_id(),
+            self.file_path.clone(),
             self.object_store.clone(),
-        );
-        // Fetches data into memory.
-        row_group
-            .fetch(&self.projection, row_selection.as_ref(), fetch_metrics)
-            .await
-            .context(ReadParquetSnafu {
-                path: &self.file_path,
-            })?;
-
-        // Record total fetch elapsed time.
-        if let Some(metrics) = fetch_metrics {
-            metrics.data.lock().unwrap().total_fetch_elapsed += fetch_start.elapsed();
-        }
-
-        // Builds the parquet reader.
-        // Now the row selection is None.
-        ParquetRecordBatchReader::try_new_with_row_groups(
-            &self.field_levels,
-            &row_group,
-            DEFAULT_READ_BATCH_SIZE,
-            row_selection,
+            self.cache_strategy.clone(),
+            self.parquet_meta.clone(),
+            row_group_idx,
         )
-        .context(ReadParquetSnafu {
-            path: &self.file_path,
-        })
-    }
-}
+        .with_fetch_metrics(fetch_metrics.cloned());
 
-/// The state of a [ParquetReader].
-enum ReaderState {
-    /// The reader is reading a row group.
-    Readable(PruneReader),
-    /// The reader is exhausted.
-    Exhausted(ReaderMetrics),
-}
+        // Build the async stream using ArrowReaderBuilder API.
+        let mut builder = ParquetRecordBatchStreamBuilder::new_with_metadata(
+            async_reader,
+            self.arrow_metadata.clone(),
+        );
+        builder = builder
+            .with_row_groups(vec![row_group_idx])
+            .with_projection(projection)
+            .with_batch_size(DEFAULT_READ_BATCH_SIZE);
 
-impl ReaderState {
-    /// Returns the metrics of the reader.
-    fn metrics(&self) -> ReaderMetrics {
-        match self {
-            ReaderState::Readable(reader) => reader.metrics(),
-            ReaderState::Exhausted(m) => m.clone(),
+        if let Some(selection) = row_selection {
+            builder = builder.with_row_selection(selection);
         }
+
+        let stream = builder.build().context(ReadParquetSnafu {
+            path: &self.file_path,
+        })?;
+
+        Ok(stream)
     }
 }
 
+#[derive(Clone)]
 /// The filter to evaluate or the prune result of the default value.
 pub(crate) enum MaybeFilter {
     /// The filter to evaluate.
@@ -1779,6 +1802,17 @@ pub(crate) enum MaybeFilter {
     Pruned,
 }
 
+impl MaybeFilter {
+    /// Returns the inner filter when it is available.
+    pub(crate) fn as_filter(&self) -> Option<&SimpleFilterEvaluator> {
+        match self {
+            MaybeFilter::Filter(filter) => Some(filter),
+            MaybeFilter::Matched | MaybeFilter::Pruned => None,
+        }
+    }
+}
+
+#[derive(Clone)]
 /// Context to evaluate the column filter for a parquet file.
 pub(crate) struct SimpleFilterContext {
     /// Filter to evaluate.
@@ -1787,8 +1821,6 @@ pub(crate) struct SimpleFilterContext {
     column_id: ColumnId,
     /// Semantic type of the column.
     semantic_type: SemanticType,
-    /// The data type of the column.
-    data_type: ConcreteDataType,
 }
 
 impl SimpleFilterContext {
@@ -1811,8 +1843,21 @@ impl SimpleFilterContext {
                 match sst_meta.column_by_id(column.column_id) {
                     Some(sst_column) => {
                         debug_assert_eq!(column.semantic_type, sst_column.semantic_type);
-
-                        (column, MaybeFilter::Filter(filter))
+                        // Schema evolution can make field columns with the same id have
+                        // different concrete data types across SSTs. In that case,
+                        // evaluating this simple filter against current SST column may
+                        // raise an invalid cross-type comparison error (e.g. Float64 == Utf8).
+                        let maybe_filter = if sst_column.column_schema.data_type
+                            == column.column_schema.data_type
+                        {
+                            MaybeFilter::Filter(filter)
+                        } else {
+                            // Altering tag or timestamp column types is not allowed,
+                            // so only field columns can reach this branch.
+                            debug_assert_eq!(column.semantic_type, SemanticType::Field);
+                            return None;
+                        };
+                        (column, maybe_filter)
                     }
                     None => {
                         // If the column is not present in the SST metadata, we evaluate the filter
@@ -1836,7 +1881,6 @@ impl SimpleFilterContext {
             filter: maybe_filter,
             column_id: column_metadata.column_id,
             semantic_type: column_metadata.semantic_type,
-            data_type: column_metadata.column_schema.data_type.clone(),
         })
     }
 
@@ -1854,10 +1898,116 @@ impl SimpleFilterContext {
     pub(crate) fn semantic_type(&self) -> SemanticType {
         self.semantic_type
     }
+}
 
-    /// Returns the data type of the column.
-    pub(crate) fn data_type(&self) -> &ConcreteDataType {
-        &self.data_type
+/// Context to evaluate a physical expression for a parquet file.
+#[derive(Clone)]
+pub(crate) struct PhysicalFilterContext {
+    /// Filter to evaluate.
+    filter: Arc<dyn PhysicalExpr>,
+    /// Id of the column to evaluate.
+    column_id: ColumnId,
+    /// Name of the column to evaluate.
+    column_name: String,
+    /// Semantic type of the column.
+    semantic_type: SemanticType,
+    /// Schema containing only the referenced column.
+    schema: SchemaRef,
+}
+
+impl PhysicalFilterContext {
+    /// Creates a context for the `expr`.
+    ///
+    /// Returns None if the expression doesn't reference exactly one column or the
+    /// column to filter doesn't exist in the SST metadata or the expected metadata.
+    pub(crate) fn new_opt(
+        sst_meta: &RegionMetadataRef,
+        expected_meta: Option<&RegionMetadata>,
+        read_format: &FlatReadFormat,
+        expr: &Expr,
+    ) -> Option<Self> {
+        if !Self::is_prefilter_candidate(expr) {
+            return None;
+        }
+        let column_name = Self::single_column_name(expr)?;
+        let column_metadata = match expected_meta {
+            Some(meta) => {
+                let column = meta.column_by_name(&column_name)?;
+                let sst_column = sst_meta.column_by_id(column.column_id)?;
+                // Physical expr requires the column name to match the SST column name.
+                if sst_column.column_schema.name != column_name {
+                    return None;
+                }
+                column
+            }
+            None => sst_meta.column_by_name(&column_name)?,
+        };
+
+        // The column must be present in the projected arrow schema for the
+        // prefilter to be able to read it.
+        let (_, field) = read_format.arrow_schema().column_with_name(&column_name)?;
+        let field = field.clone();
+        let schema = Arc::new(ArrowSchema::new(vec![field]));
+        let physical_expr = Predicate::to_physical_expr(expr, &schema)
+            .inspect_err(|e| {
+                error!(e; "Unable to build physical filter for {expr}, schema: {schema:?}");
+            })
+            .ok()?;
+
+        Some(Self {
+            filter: physical_expr,
+            column_id: column_metadata.column_id,
+            column_name,
+            semantic_type: column_metadata.semantic_type,
+            schema,
+        })
+    }
+
+    /// Returns true if the expression is a variant we want to evaluate as a
+    /// physical prefilter. Binary exprs are intentionally excluded because
+    /// [`SimpleFilterEvaluator`] already handles them.
+    // TODO(yingwen): extend more expressions if necessary. For example, allow some cheap scalar functions (e.g. `lower`, `length`, date truncations)
+    fn is_prefilter_candidate(expr: &Expr) -> bool {
+        matches!(
+            expr,
+            Expr::InList(_) | Expr::IsNull(_) | Expr::IsNotNull(_) | Expr::Between(_)
+        )
+    }
+
+    fn single_column_name(expr: &Expr) -> Option<String> {
+        let mut columns = HashSet::new();
+        if expr_to_columns(expr, &mut columns).is_err() {
+            return None;
+        }
+        if columns.len() != 1 {
+            return None;
+        }
+        columns.iter().next().map(|column| column.name.clone())
+    }
+
+    /// Returns the filter to evaluate.
+    pub(crate) fn filter(&self) -> &Arc<dyn PhysicalExpr> {
+        &self.filter
+    }
+
+    /// Returns the column id.
+    pub(crate) fn column_id(&self) -> ColumnId {
+        self.column_id
+    }
+
+    /// Returns the column name.
+    pub(crate) fn column_name(&self) -> &str {
+        &self.column_name
+    }
+
+    /// Returns the semantic type of the column.
+    pub(crate) fn semantic_type(&self) -> SemanticType {
+        self.semantic_type
+    }
+
+    /// Returns the schema containing only the referenced column.
+    pub(crate) fn schema(&self) -> &SchemaRef {
+        &self.schema
     }
 }
 
@@ -1879,13 +2029,12 @@ pub struct ParquetReader {
     /// Row group selection to read.
     selection: RowGroupSelection,
     /// Reader of current row group.
-    reader_state: ReaderState,
+    reader: Option<FlatPruneReader>,
     /// Metrics for tracking row group fetch operations.
     fetch_metrics: ParquetFetchMetrics,
 }
 
-#[async_trait]
-impl BatchReader for ParquetReader {
+impl ParquetReader {
     #[tracing::instrument(
         skip_all,
         fields(
@@ -1893,76 +2042,37 @@ impl BatchReader for ParquetReader {
             file_id = %self.context.reader_builder().file_handle.file_id()
         )
     )]
-    async fn next_batch(&mut self) -> Result<Option<Batch>> {
-        let ReaderState::Readable(reader) = &mut self.reader_state else {
-            return Ok(None);
-        };
+    pub async fn next_record_batch(&mut self) -> Result<Option<RecordBatch>> {
+        loop {
+            if let Some(reader) = &mut self.reader {
+                if let Some(batch) = reader.next_batch().await? {
+                    return Ok(Some(batch));
+                }
+                self.reader = None;
+                continue;
+            }
 
-        // We don't collect the elapsed time if the reader returns an error.
-        if let Some(batch) = reader.next_batch().await? {
-            return Ok(Some(batch));
-        }
+            let Some((row_group_idx, row_selection)) = self.selection.pop_first() else {
+                return Ok(None);
+            };
 
-        // No more items in current row group, reads next row group.
-        while let Some((row_group_idx, row_selection)) = self.selection.pop_first() {
+            let skip_fields = self.context.pre_filter_mode().skip_fields();
             let parquet_reader = self
                 .context
                 .reader_builder()
-                .build(
+                .build(self.context.build_context(
                     row_group_idx,
                     Some(row_selection),
                     Some(&self.fetch_metrics),
-                )
+                ))
                 .await?;
-
-            // Resets the parquet reader.
-            // Compute skip_fields for this row group
-            let skip_fields = self.context.should_skip_fields(row_group_idx);
-            reader.reset_source(
-                Source::RowGroup(RowGroupReader::new(self.context.clone(), parquet_reader)),
+            self.reader = Some(FlatPruneReader::new_with_row_group_reader(
+                self.context.clone(),
+                FlatRowGroupReader::new(self.context.clone(), parquet_reader),
                 skip_fields,
-            );
-            if let Some(batch) = reader.next_batch().await? {
-                return Ok(Some(batch));
-            }
+            ));
         }
-
-        // The reader is exhausted.
-        self.reader_state = ReaderState::Exhausted(reader.metrics().clone());
-        Ok(None)
     }
-}
-
-impl Drop for ParquetReader {
-    fn drop(&mut self) {
-        let metrics = self.reader_state.metrics();
-        debug!(
-            "Read parquet {} {}, range: {:?}, {}/{} row groups, metrics: {:?}",
-            self.context.reader_builder().file_handle.region_id(),
-            self.context.reader_builder().file_handle.file_id(),
-            self.context.reader_builder().file_handle.time_range(),
-            metrics.filter_metrics.rg_total
-                - metrics.filter_metrics.rg_inverted_filtered
-                - metrics.filter_metrics.rg_minmax_filtered
-                - metrics.filter_metrics.rg_fulltext_filtered
-                - metrics.filter_metrics.rg_bloom_filtered,
-            metrics.filter_metrics.rg_total,
-            metrics
-        );
-
-        // Report metrics.
-        READ_STAGE_ELAPSED
-            .with_label_values(&["build_parquet_reader"])
-            .observe(metrics.build_cost.as_secs_f64());
-        READ_STAGE_ELAPSED
-            .with_label_values(&["scan_row_groups"])
-            .observe(metrics.scan_cost.as_secs_f64());
-        metrics.observe_rows("parquet_reader");
-        metrics.filter_metrics.observe();
-    }
-}
-
-impl ParquetReader {
     /// Creates a new reader.
     #[tracing::instrument(
         skip_all,
@@ -1976,27 +2086,29 @@ impl ParquetReader {
         mut selection: RowGroupSelection,
     ) -> Result<Self> {
         let fetch_metrics = ParquetFetchMetrics::default();
-        // No more items in current row group, reads next row group.
-        let reader_state = if let Some((row_group_idx, row_selection)) = selection.pop_first() {
+        let reader = if let Some((row_group_idx, row_selection)) = selection.pop_first() {
+            let skip_fields = context.pre_filter_mode().skip_fields();
             let parquet_reader = context
                 .reader_builder()
-                .build(row_group_idx, Some(row_selection), Some(&fetch_metrics))
+                .build(context.build_context(
+                    row_group_idx,
+                    Some(row_selection),
+                    Some(&fetch_metrics),
+                ))
                 .await?;
-            // Compute skip_fields once for this row group
-            let skip_fields = context.should_skip_fields(row_group_idx);
-            ReaderState::Readable(PruneReader::new_with_row_group_reader(
+            Some(FlatPruneReader::new_with_row_group_reader(
                 context.clone(),
-                RowGroupReader::new(context.clone(), parquet_reader),
+                FlatRowGroupReader::new(context.clone(), parquet_reader),
                 skip_fields,
             ))
         } else {
-            ReaderState::Exhausted(ReaderMetrics::default())
+            None
         };
 
         Ok(ParquetReader {
             context,
             selection,
-            reader_state,
+            reader,
             fetch_metrics,
         })
     }
@@ -2011,151 +2123,19 @@ impl ParquetReader {
     }
 }
 
-/// RowGroupReaderContext represents the fields that cannot be shared
-/// between different `RowGroupReader`s.
-pub(crate) trait RowGroupReaderContext: Send {
-    fn map_result(
-        &self,
-        result: std::result::Result<Option<RecordBatch>, ArrowError>,
-    ) -> Result<Option<RecordBatch>>;
-
-    fn read_format(&self) -> &ReadFormat;
-}
-
-impl RowGroupReaderContext for FileRangeContextRef {
-    fn map_result(
-        &self,
-        result: std::result::Result<Option<RecordBatch>, ArrowError>,
-    ) -> Result<Option<RecordBatch>> {
-        result.context(ArrowReaderSnafu {
-            path: self.file_path(),
-        })
-    }
-
-    fn read_format(&self) -> &ReadFormat {
-        self.as_ref().read_format()
-    }
-}
-
-/// [RowGroupReader] that reads from [FileRange].
-pub(crate) type RowGroupReader = RowGroupReaderBase<FileRangeContextRef>;
-
-impl RowGroupReader {
-    /// Creates a new reader from file range.
-    pub(crate) fn new(context: FileRangeContextRef, reader: ParquetRecordBatchReader) -> Self {
-        Self::create(context, reader)
-    }
-}
-
-/// Reader to read a row group of a parquet file.
-pub(crate) struct RowGroupReaderBase<T> {
-    /// Context of [RowGroupReader] so adapts to different underlying implementation.
-    context: T,
-    /// Inner parquet reader.
-    reader: ParquetRecordBatchReader,
-    /// Buffered batches to return.
-    batches: VecDeque<Batch>,
-    /// Local scan metrics.
-    metrics: ReaderMetrics,
-    /// Cached sequence array to override sequences.
-    override_sequence: Option<ArrayRef>,
-}
-
-impl<T> RowGroupReaderBase<T>
-where
-    T: RowGroupReaderContext,
-{
-    /// Creates a new reader to read the primary key format.
-    pub(crate) fn create(context: T, reader: ParquetRecordBatchReader) -> Self {
-        // The batch length from the reader should be less than or equal to DEFAULT_READ_BATCH_SIZE.
-        let override_sequence = context
-            .read_format()
-            .new_override_sequence_array(DEFAULT_READ_BATCH_SIZE);
-        assert!(context.read_format().as_primary_key().is_some());
-
-        Self {
-            context,
-            reader,
-            batches: VecDeque::new(),
-            metrics: ReaderMetrics::default(),
-            override_sequence,
-        }
-    }
-
-    /// Gets the metrics.
-    pub(crate) fn metrics(&self) -> &ReaderMetrics {
-        &self.metrics
-    }
-
-    /// Gets [ReadFormat] of underlying reader.
-    pub(crate) fn read_format(&self) -> &ReadFormat {
-        self.context.read_format()
-    }
-
-    /// Tries to fetch next [RecordBatch] from the reader.
-    fn fetch_next_record_batch(&mut self) -> Result<Option<RecordBatch>> {
-        self.context.map_result(self.reader.next().transpose())
-    }
-
-    /// Returns the next [Batch].
-    pub(crate) fn next_inner(&mut self) -> Result<Option<Batch>> {
-        let scan_start = Instant::now();
-        if let Some(batch) = self.batches.pop_front() {
-            self.metrics.num_rows += batch.num_rows();
-            self.metrics.scan_cost += scan_start.elapsed();
-            return Ok(Some(batch));
-        }
-
-        // We need to fetch next record batch and convert it to batches.
-        while self.batches.is_empty() {
-            let Some(record_batch) = self.fetch_next_record_batch()? else {
-                self.metrics.scan_cost += scan_start.elapsed();
-                return Ok(None);
-            };
-            self.metrics.num_record_batches += 1;
-
-            // Safety: We ensures the format is primary key in the RowGroupReaderBase::create().
-            self.context
-                .read_format()
-                .as_primary_key()
-                .unwrap()
-                .convert_record_batch(
-                    &record_batch,
-                    self.override_sequence.as_ref(),
-                    &mut self.batches,
-                )?;
-            self.metrics.num_batches += self.batches.len();
-        }
-        let batch = self.batches.pop_front();
-        self.metrics.num_rows += batch.as_ref().map(|b| b.num_rows()).unwrap_or(0);
-        self.metrics.scan_cost += scan_start.elapsed();
-        Ok(batch)
-    }
-}
-
-#[async_trait::async_trait]
-impl<T> BatchReader for RowGroupReaderBase<T>
-where
-    T: RowGroupReaderContext,
-{
-    async fn next_batch(&mut self) -> Result<Option<Batch>> {
-        self.next_inner()
-    }
-}
-
 /// Reader to read a row group of a parquet file in flat format, returning RecordBatch.
 pub(crate) struct FlatRowGroupReader {
     /// Context for file ranges.
     context: FileRangeContextRef,
-    /// Inner parquet reader.
-    reader: ParquetRecordBatchReader,
+    /// Inner parquet record batch stream.
+    stream: ProjectedRecordBatchStream,
     /// Cached sequence array to override sequences.
     override_sequence: Option<ArrayRef>,
 }
 
 impl FlatRowGroupReader {
     /// Creates a new flat reader from file range.
-    pub(crate) fn new(context: FileRangeContextRef, reader: ParquetRecordBatchReader) -> Self {
+    pub(crate) fn new(context: FileRangeContextRef, stream: ProjectedRecordBatchStream) -> Self {
         // The batch length from the reader should be less than or equal to DEFAULT_READ_BATCH_SIZE.
         let override_sequence = context
             .read_format()
@@ -2163,23 +2143,21 @@ impl FlatRowGroupReader {
 
         Self {
             context,
-            reader,
+            stream,
             override_sequence,
         }
     }
 
     /// Returns the next RecordBatch.
-    pub(crate) fn next_batch(&mut self) -> Result<Option<RecordBatch>> {
-        match self.reader.next() {
+    pub(crate) async fn next_batch(&mut self) -> Result<Option<RecordBatch>> {
+        match self.stream.next().await {
             Some(batch_result) => {
-                let record_batch = batch_result.context(ArrowReaderSnafu {
-                    path: self.context.file_path(),
-                })?;
+                let record_batch = batch_result?;
 
-                // Safety: Only flat format use FlatRowGroupReader.
-                let flat_format = self.context.read_format().as_flat().unwrap();
-                let record_batch =
-                    flat_format.convert_batch(record_batch, self.override_sequence.as_ref())?;
+                let record_batch = self
+                    .context
+                    .read_format()
+                    .convert_batch(record_batch, self.override_sequence.as_ref())?;
                 Ok(Some(record_batch))
             }
             None => Ok(None),
@@ -2198,12 +2176,17 @@ mod tests {
     use datafusion_expr::expr::ScalarFunction;
     use datafusion_expr::{
         ColumnarValue, Expr, ScalarFunctionArgs, ScalarUDF, ScalarUDFImpl, Signature, Volatility,
+        col, lit,
     };
     use datatypes::arrow::array::{ArrayRef, Int64Array};
     use datatypes::arrow::record_batch::RecordBatch;
+    use datatypes::prelude::ConcreteDataType;
+    use datatypes::schema::ColumnSchema;
     use object_store::services::Memory;
     use parquet::arrow::ArrowWriter;
+    use store_api::metadata::{ColumnMetadata, RegionMetadata, RegionMetadataBuilder};
     use store_api::region_request::PathType;
+    use store_api::storage::RegionId;
     use table::predicate::Predicate;
 
     use super::*;
@@ -2264,8 +2247,19 @@ mod tests {
         object_store.write(&file_path, parquet_bytes).await.unwrap();
 
         let region_metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
-        let read_format =
-            ReadFormat::new(region_metadata, None, false, None, &file_path, false).unwrap();
+        let read_format = FlatReadFormat::new(
+            region_metadata.clone(),
+            ReadColumns::from_deduped_column_ids(
+                region_metadata
+                    .column_metadatas
+                    .iter()
+                    .map(|column| column.column_id),
+            ),
+            None,
+            &file_path,
+            false,
+        )
+        .unwrap();
 
         let mut cache_metrics = MetadataCacheMetrics::default();
         let loader = MetadataLoader::new(object_store.clone(), &file_path, file_size);
@@ -2294,4 +2288,188 @@ mod tests {
 
         assert!(!selection.is_empty());
     }
+
+    fn expected_metadata_with_reused_tag_name(
+        old_metadata: &RegionMetadata,
+    ) -> Arc<RegionMetadata> {
+        let mut builder = RegionMetadataBuilder::new(old_metadata.region_id);
+        builder
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "tag_0".to_string(),
+                    ConcreteDataType::string_datatype(),
+                    true,
+                ),
+                semantic_type: SemanticType::Tag,
+                column_id: 10,
+            })
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "tag_1".to_string(),
+                    ConcreteDataType::string_datatype(),
+                    true,
+                ),
+                semantic_type: SemanticType::Tag,
+                column_id: 1,
+            })
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "field_0".to_string(),
+                    ConcreteDataType::uint64_datatype(),
+                    true,
+                ),
+                semantic_type: SemanticType::Field,
+                column_id: 2,
+            })
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "ts".to_string(),
+                    ConcreteDataType::timestamp_millisecond_datatype(),
+                    false,
+                ),
+                semantic_type: SemanticType::Timestamp,
+                column_id: 3,
+            })
+            .primary_key(vec![10, 1]);
+
+        Arc::new(builder.build().unwrap())
+    }
+
+    #[test]
+    fn test_simple_filter_context_uses_default_value_for_mismatched_expected_metadata() {
+        let metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
+        let expected_metadata = expected_metadata_with_reused_tag_name(metadata.as_ref());
+        let ctx = SimpleFilterContext::new_opt(
+            &metadata,
+            Some(expected_metadata.as_ref()),
+            &col("tag_0").eq(lit("a")),
+        )
+        .unwrap();
+        assert!(matches!(
+            ctx.filter(),
+            MaybeFilter::Matched | MaybeFilter::Pruned
+        ));
+    }
+
+    #[test]
+    fn test_simple_filter_context_drops_mismatched_field_filter() {
+        let (sst_metadata, latest_metadata) = mock_metadata();
+        let ctx = SimpleFilterContext::new_opt(
+            &sst_metadata,
+            Some(latest_metadata.as_ref()),
+            &col("field_0").eq(lit(1_i64)),
+        );
+
+        assert!(ctx.is_none());
+    }
+
+    fn mock_metadata() -> (RegionMetadataRef, RegionMetadataRef) {
+        let region_id = RegionId::new(1, 1);
+        let make_tag_0 = || ColumnMetadata {
+            column_schema: ColumnSchema::new(
+                "tag_0".to_string(),
+                ConcreteDataType::string_datatype(),
+                true,
+            ),
+            semantic_type: SemanticType::Tag,
+            column_id: 0,
+        };
+        let make_ts = || ColumnMetadata {
+            column_schema: ColumnSchema::new(
+                "ts".to_string(),
+                ConcreteDataType::timestamp_millisecond_datatype(),
+                false,
+            ),
+            semantic_type: SemanticType::Timestamp,
+            column_id: 2,
+        };
+        let make_field_0 = |data_type| ColumnMetadata {
+            column_schema: ColumnSchema::new("field_0".to_string(), data_type, true),
+            semantic_type: SemanticType::Field,
+            column_id: 1,
+        };
+
+        let mut sst_builder = RegionMetadataBuilder::new(region_id);
+        sst_builder
+            .push_column_metadata(make_tag_0())
+            .push_column_metadata(make_field_0(ConcreteDataType::uint64_datatype()))
+            .push_column_metadata(make_ts())
+            .primary_key(vec![0]);
+        let sst_metadata = Arc::new(sst_builder.build().unwrap());
+
+        let mut expected_builder = RegionMetadataBuilder::new(region_id);
+        expected_builder
+            .push_column_metadata(make_tag_0())
+            .push_column_metadata(make_field_0(ConcreteDataType::int64_datatype()))
+            .push_column_metadata(make_ts())
+            .primary_key(vec![0]);
+
+        let expected_metadata = Arc::new(expected_builder.build().unwrap());
+
+        (sst_metadata, expected_metadata)
+    }
+
+    #[test]
+    fn test_physical_filter_context_skips_renamed_column() {
+        let metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
+        let expected_metadata = expected_metadata_with_reused_tag_name(metadata.as_ref());
+        let read_format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids(
+                metadata.column_metadatas.iter().map(|c| c.column_id),
+            ),
+            None,
+            "test",
+            true,
+        )
+        .unwrap();
+
+        let ctx = PhysicalFilterContext::new_opt(
+            &metadata,
+            Some(expected_metadata.as_ref()),
+            &read_format,
+            &col("tag_0").in_list(vec![lit("a"), lit("b")], false),
+        );
+
+        assert!(ctx.is_none());
+    }
+
+    #[test]
+    fn test_physical_filter_context_only_accepts_prefilter_candidates() {
+        let metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
+        let read_format = FlatReadFormat::new(
+            metadata.clone(),
+            ReadColumns::from_deduped_column_ids(
+                metadata.column_metadatas.iter().map(|c| c.column_id),
+            ),
+            None,
+            "test",
+            true,
+        )
+        .unwrap();
+
+        // InList is on the allowlist — should build a context.
+        let in_list = col("tag_0").in_list(vec![lit("a"), lit("b")], false);
+        assert!(PhysicalFilterContext::new_opt(&metadata, None, &read_format, &in_list).is_some());
+
+        // NOT IN uses the same variant with `negated: true` — also accepted.
+        let not_in = col("tag_0").in_list(vec![lit("a"), lit("b")], true);
+        assert!(PhysicalFilterContext::new_opt(&metadata, None, &read_format, &not_in).is_some());
+
+        // IS NULL / IS NOT NULL are accepted.
+        let is_null = col("tag_0").is_null();
+        assert!(PhysicalFilterContext::new_opt(&metadata, None, &read_format, &is_null).is_some());
+        let is_not_null = col("tag_0").is_not_null();
+        assert!(
+            PhysicalFilterContext::new_opt(&metadata, None, &read_format, &is_not_null).is_some()
+        );
+
+        // BETWEEN is accepted.
+        let between = col("field_0").between(lit(1_u64), lit(10_u64));
+        assert!(PhysicalFilterContext::new_opt(&metadata, None, &read_format, &between).is_some());
+
+        // Binary expr is handled by SimpleFilterEvaluator — rejected here.
+        let binary = col("tag_0").eq(lit("a"));
+        assert!(PhysicalFilterContext::new_opt(&metadata, None, &read_format, &binary).is_none());
+    }
 }
diff --git a/src/mito2/src/sst/parquet/reader/stream.rs b/src/mito2/src/sst/parquet/reader/stream.rs
new file mode 100644
index 0000000000..f6211df475
--- /dev/null
+++ b/src/mito2/src/sst/parquet/reader/stream.rs
@@ -0,0 +1,418 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::pin::Pin;
+use std::task::{Context, Poll};
+
+use datafusion_common::cast_column;
+use datafusion_common::format::DEFAULT_CAST_OPTIONS;
+use datatypes::arrow::array::{ArrayRef, new_null_array};
+use datatypes::arrow::datatypes::{DataType, FieldRef, SchemaRef};
+use datatypes::arrow::record_batch::RecordBatch;
+use futures::Stream;
+use futures::stream::BoxStream;
+use parquet::arrow::async_reader::ParquetRecordBatchStream;
+use snafu::{IntoError, ResultExt, ensure};
+
+use crate::error::{
+    CastColumnSnafu, NewRecordBatchSnafu, ReadParquetSnafu, Result, UnexpectedSnafu,
+};
+use crate::sst::parquet::async_reader::SstAsyncFileReader;
+
+/// Aligns projected batches to the expected output schema for nested projections.
+///
+/// Background
+/// ----------
+/// Nested projection may ask parquet to read leaves under a root column. If none
+/// of the requested leaves exists in the current parquet file, parquet decoding
+/// omits the whole root from the physical [`RecordBatch`].
+///
+/// In addition, after nested-path filtering, returned struct arrays may contain
+/// only a subset of fields. The current output schema is not pruned by nested
+/// paths, so physical struct fields can be a subset of the expected struct
+/// fields, and their nested schema can differ from the expected output schema.
+///
+/// To keep projected batches schema-consistent before entering upper readers:
+/// - Root-column presence alignment restores missing projected root columns by
+///   inserting root-level null arrays.
+/// - Nested struct alignment aligns struct arrays to the expected nested field
+///   layout.
+pub struct NestedSchemaAligner<S> {
+    inner: S,
+    /// Output schema expected by the upper reader.
+    output_schema: SchemaRef,
+    /// Whether each projected root exists in the physical batch returned by
+    /// parquet.
+    projected_root_presence: Vec<bool>,
+    /// Number of columns expected from the physical batch returned by parquet.
+    expected_input_col_num: usize,
+    /// Whether all projected roots are present and the stream can pass batches
+    /// through.
+    all_roots_present: bool,
+    /// The cache for whether incoming batches already match output schema.
+    is_schema_matched: Option<bool>,
+}
+
+pub(crate) type ProjectedRecordBatchStream = BoxStream<'static, Result<RecordBatch>>;
+
+impl<S> NestedSchemaAligner<S>
+where
+    S: Stream<Item = Result<RecordBatch>>,
+{
+    pub fn new(
+        inner: S,
+        projected_root_presence: Vec<bool>,
+        output_schema: SchemaRef,
+    ) -> Result<NestedSchemaAligner<S>> {
+        ensure!(
+            projected_root_presence.len() == output_schema.fields().len(),
+            UnexpectedSnafu {
+                reason: format!(
+                    "NestedSchemaAligner projected root presence len {} does not match output schema columns {}",
+                    projected_root_presence.len(),
+                    output_schema.fields().len()
+                ),
+            }
+        );
+
+        let expected_input_col_num = projected_root_presence
+            .iter()
+            .filter(|matched| **matched)
+            .count();
+        let all_roots_present = projected_root_presence.iter().all(|&m| m);
+
+        Ok(NestedSchemaAligner {
+            inner,
+            output_schema,
+            projected_root_presence,
+            expected_input_col_num,
+            all_roots_present,
+            is_schema_matched: None,
+        })
+    }
+}
+
+impl<S> Stream for NestedSchemaAligner<S>
+where
+    S: Stream<Item = Result<RecordBatch>> + Unpin,
+{
+    type Item = Result<RecordBatch>;
+
+    fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        let this = self.get_mut();
+
+        match Pin::new(&mut this.inner).poll_next(cx) {
+            Poll::Ready(Some(Ok(rb))) => {
+                let rb = if this.all_roots_present {
+                    rb
+                } else {
+                    fill_missing_cols(
+                        rb,
+                        &this.output_schema,
+                        &this.projected_root_presence,
+                        this.expected_input_col_num,
+                    )?
+                };
+
+                let is_schema_matched = *this
+                    .is_schema_matched
+                    .get_or_insert_with(|| rb.schema() == this.output_schema);
+
+                if is_schema_matched {
+                    Poll::Ready(Some(Ok(rb)))
+                } else {
+                    Poll::Ready(Some(align_batch_to_schema(rb, &this.output_schema)))
+                }
+            }
+            Poll::Ready(Some(Err(err))) => Poll::Ready(Some(Err(err))),
+            Poll::Ready(None) => Poll::Ready(None),
+            Poll::Pending => Poll::Pending,
+        }
+    }
+}
+
+fn fill_missing_cols(
+    rb: RecordBatch,
+    output_schema: &SchemaRef,
+    projected_root_matches: &[bool],
+    expected_input_col_num: usize,
+) -> Result<RecordBatch> {
+    ensure!(
+        rb.columns().len() == expected_input_col_num,
+        UnexpectedSnafu {
+            reason: format!(
+                "NestedSchemaAligner expected {} input columns but got {}",
+                expected_input_col_num,
+                rb.columns().len()
+            ),
+        }
+    );
+
+    let mut cols = Vec::with_capacity(projected_root_matches.len());
+    let mut idx = 0;
+
+    for (field, matched) in output_schema.fields().iter().zip(projected_root_matches) {
+        if *matched {
+            cols.push(rb.column(idx).clone());
+            idx += 1;
+        } else {
+            cols.push(new_null_array(field.data_type(), rb.num_rows()));
+        }
+    }
+
+    RecordBatch::try_new(output_schema.clone(), cols).context(NewRecordBatchSnafu)
+}
+
+fn align_batch_to_schema(rb: RecordBatch, output_schema: &SchemaRef) -> Result<RecordBatch> {
+    ensure!(
+        rb.num_columns() == output_schema.fields().len(),
+        UnexpectedSnafu {
+            reason: format!(
+                "NestedSchemaAligner expected {} columns but got {}",
+                output_schema.fields().len(),
+                rb.num_columns()
+            ),
+        }
+    );
+
+    let columns = rb
+        .columns()
+        .iter()
+        .zip(output_schema.fields())
+        .map(|(array, field)| align_array(array, field))
+        .collect::<Result<Vec<_>>>()?;
+
+    RecordBatch::try_new(output_schema.clone(), columns).context(NewRecordBatchSnafu)
+}
+
+fn align_array(array: &ArrayRef, field: &FieldRef) -> Result<ArrayRef> {
+    if array.data_type() == field.data_type() {
+        return Ok(array.clone());
+    }
+
+    if !matches!(field.data_type(), DataType::Struct(_)) {
+        return Ok(array.clone());
+    }
+
+    cast_column(array, field.as_ref(), &DEFAULT_CAST_OPTIONS).context(CastColumnSnafu)
+}
+
+/// Maps parquet stream errors into mito errors before batches enter the filler.
+pub(crate) struct ParquetErrorAdapter {
+    inner: ParquetRecordBatchStream<SstAsyncFileReader>,
+    path: String,
+}
+
+impl ParquetErrorAdapter {
+    pub(crate) fn new(inner: ParquetRecordBatchStream<SstAsyncFileReader>, path: String) -> Self {
+        Self { inner, path }
+    }
+}
+
+impl Stream for ParquetErrorAdapter {
+    type Item = Result<RecordBatch>;
+
+    fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        let this = self.get_mut();
+
+        match Pin::new(&mut this.inner).poll_next(cx) {
+            Poll::Ready(Some(Ok(rb))) => Poll::Ready(Some(Ok(rb))),
+            Poll::Ready(Some(Err(err))) => {
+                Poll::Ready(Some(Err(
+                    ReadParquetSnafu { path: &this.path }.into_error(err)
+                )))
+            }
+            Poll::Ready(None) => Poll::Ready(None),
+            Poll::Pending => Poll::Pending,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use datatypes::arrow::array::{Array, ArrayRef, Int64Array, StringArray, StructArray};
+    use datatypes::arrow::datatypes::{DataType, Field, Fields, Schema};
+    use futures::{StreamExt, stream};
+
+    use super::*;
+
+    #[tokio::test]
+    async fn test_aligner_with_all_projected_roots_match() {
+        let output_schema = schema([
+            Field::new("a", DataType::Int64, true),
+            Field::new("b", DataType::Utf8, true),
+        ]);
+        let input = RecordBatch::try_new(
+            output_schema.clone(),
+            vec![int_array([1, 2, 3]), string_array(["x", "y", "z"])],
+        )
+        .unwrap();
+        let stream = stream::iter([Ok(input.clone())]);
+
+        let mut aligner =
+            NestedSchemaAligner::new(stream, vec![true, true], output_schema.clone()).unwrap();
+        let output = aligner.next().await.unwrap().unwrap();
+
+        assert_eq!(input, output);
+        assert!(aligner.next().await.is_none());
+    }
+
+    #[tokio::test]
+    async fn test_aligner_with_fills_null_root_columns() {
+        let input_schema = schema([Field::new("a", DataType::Int64, true)]);
+        let output_schema = schema([
+            Field::new("a", DataType::Int64, true),
+            Field::new("missing", DataType::Utf8, true),
+            Field::new("c", DataType::Int64, true),
+        ]);
+        let input = RecordBatch::try_new(input_schema, vec![int_array([10, 20])]).unwrap();
+        let stream = stream::iter([Ok(input)]);
+
+        let mut aligner =
+            NestedSchemaAligner::new(stream, vec![true, false, false], output_schema.clone())
+                .unwrap();
+        let output = aligner.next().await.unwrap().unwrap();
+
+        assert_eq!(output_schema, output.schema());
+        assert_eq!(3, output.num_columns());
+        assert_eq!(
+            &[Some(10), Some(20)],
+            output
+                .column(0)
+                .as_any()
+                .downcast_ref::<Int64Array>()
+                .unwrap()
+                .iter()
+                .collect::<Vec<_>>()
+                .as_slice()
+        );
+        assert_eq!(DataType::Utf8, *output.column(1).data_type());
+        assert_eq!(output.num_rows(), output.column(1).null_count());
+        assert_eq!(DataType::Int64, *output.column(2).data_type());
+        assert_eq!(output.num_rows(), output.column(2).null_count());
+    }
+
+    #[tokio::test]
+    async fn test_aligner_with_fills_missing_struct_root_column() {
+        let input_schema = schema([Field::new("a", DataType::Int64, true)]);
+        let struct_type = DataType::Struct(Fields::from(vec![
+            Field::new("x", DataType::Int64, true),
+            Field::new("y", DataType::Utf8, true),
+        ]));
+        let output_schema = schema([
+            Field::new("a", DataType::Int64, true),
+            Field::new("missing_struct", struct_type.clone(), true),
+        ]);
+        let input = RecordBatch::try_new(input_schema, vec![int_array([10, 20])]).unwrap();
+        let stream = stream::iter([Ok(input)]);
+
+        let mut aligner =
+            NestedSchemaAligner::new(stream, vec![true, false], output_schema.clone()).unwrap();
+        let output = aligner.next().await.unwrap().unwrap();
+
+        assert_eq!(output_schema, output.schema());
+        assert_eq!(2, output.num_columns());
+        assert_eq!(struct_type, output.column(1).data_type().clone());
+        assert_eq!(output.num_rows(), output.column(1).null_count());
+    }
+
+    #[tokio::test]
+    async fn test_aligner_reject_projection_len_mismatch() {
+        let output_schema = schema([Field::new("a", DataType::Int64, true)]);
+        let stream = stream::iter([]);
+
+        let err = match NestedSchemaAligner::new(stream, vec![true, false], output_schema) {
+            Ok(_) => panic!("NestedSchemaAligner should reject projection length mismatch"),
+            Err(err) => err,
+        };
+
+        assert!(
+            err.to_string()
+                .contains("projected root presence len 2 does not match output schema columns 1")
+        );
+    }
+
+    #[tokio::test]
+    async fn test_aligner_reject_with_input_column_mismatch() {
+        let input_schema = schema([Field::new("a", DataType::Int64, true)]);
+        let output_schema = schema([
+            Field::new("a", DataType::Int64, true),
+            Field::new("b", DataType::Int64, true),
+            Field::new("missing", DataType::Int64, true),
+        ]);
+        let input = RecordBatch::try_new(input_schema, vec![int_array([1, 2])]).unwrap();
+        let stream = stream::iter([Ok(input)]);
+
+        let mut aligner =
+            NestedSchemaAligner::new(stream, vec![true, true, false], output_schema).unwrap();
+        let err = aligner.next().await.unwrap().unwrap_err();
+
+        assert!(
+            err.to_string()
+                .contains("expected 2 input columns but got 1")
+        );
+    }
+
+    #[tokio::test]
+    async fn test_nested_schema_aligner_aligns_struct_field() {
+        let output_schema = schema([Field::new(
+            "nested",
+            DataType::Struct(Fields::from(vec![
+                Field::new("x", DataType::Int64, true),
+                Field::new("y", DataType::Utf8, true),
+            ])),
+            true,
+        )]);
+        let input = RecordBatch::try_new(
+            schema([Field::new(
+                "nested",
+                DataType::Struct(Fields::from(vec![Field::new("x", DataType::Int64, true)])),
+                true,
+            )]),
+            vec![Arc::new(StructArray::from(vec![(
+                Arc::new(Field::new("x", DataType::Int64, true)),
+                int_array([1, 2]),
+            )]))],
+        )
+        .unwrap();
+
+        let mut aligner =
+            NestedSchemaAligner::new(stream::iter([Ok(input)]), vec![true], output_schema.clone())
+                .unwrap();
+        let output = aligner.next().await.unwrap().unwrap();
+
+        assert_eq!(output_schema, output.schema());
+        let nested = output
+            .column(0)
+            .as_any()
+            .downcast_ref::<StructArray>()
+            .unwrap();
+        assert_eq!(2, nested.columns().len());
+        assert_eq!(2, nested.column(1).null_count());
+    }
+
+    fn schema(fields: impl IntoIterator<Item = Field>) -> SchemaRef {
+        Arc::new(Schema::new(fields.into_iter().collect::<Vec<_>>()))
+    }
+
+    fn int_array(values: impl IntoIterator<Item = i64>) -> ArrayRef {
+        Arc::new(Int64Array::from_iter_values(values))
+    }
+
+    fn string_array(values: impl IntoIterator<Item = &'static str>) -> ArrayRef {
+        Arc::new(StringArray::from_iter_values(values))
+    }
+}
diff --git a/src/mito2/src/sst/parquet/row_group.rs b/src/mito2/src/sst/parquet/row_group.rs
index 8f3f6c5f62..d8a44dfff5 100644
--- a/src/mito2/src/sst/parquet/row_group.rs
+++ b/src/mito2/src/sst/parquet/row_group.rs
@@ -12,28 +12,12 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-//! Ports private structs from [parquet crate](https://github.com/apache/arrow-rs/blob/7e134f4d277c0b62c27529fc15a4739de3ad0afd/parquet/src/arrow/async_reader/mod.rs#L644-L650).
+//! Parquet row group reading utilities.
 
 use std::ops::Range;
 use std::sync::Arc;
 
-use bytes::{Buf, Bytes};
-use object_store::ObjectStore;
-use parquet::arrow::ProjectionMask;
-use parquet::arrow::arrow_reader::{RowGroups, RowSelection};
-use parquet::column::page::{PageIterator, PageReader};
-use parquet::errors::{ParquetError, Result};
-use parquet::file::metadata::{ParquetMetaData, RowGroupMetaData};
-use parquet::file::page_index::offset_index::OffsetIndexMetaData;
-use parquet::file::reader::{ChunkReader, Length};
-use parquet::file::serialized_reader::SerializedPageReader;
-use store_api::storage::{FileId, RegionId};
-use tokio::task::yield_now;
-
-use crate::cache::file_cache::{FileType, IndexKey};
-use crate::cache::{CacheStrategy, PageKey, PageValue};
-use crate::metrics::{READ_STAGE_ELAPSED, READ_STAGE_FETCH_PAGES};
-use crate::sst::parquet::helper::{MERGE_GAP, fetch_byte_ranges};
+use crate::sst::parquet::helper::MERGE_GAP;
 
 /// Inner data for ParquetFetchMetrics.
 #[derive(Default, Debug, Clone)]
@@ -64,19 +48,23 @@ pub struct ParquetFetchMetricsData {
     pub store_fetch_elapsed: std::time::Duration,
     /// Total elapsed time for fetching row groups.
     pub total_fetch_elapsed: std::time::Duration,
+    /// Elapsed time for prefilter execution.
+    pub prefilter_cost: std::time::Duration,
+    /// Number of rows filtered out by prefiltering.
+    pub prefilter_filtered_rows: usize,
 }
 
 impl ParquetFetchMetricsData {
     /// Returns true if the metrics are empty (contain no meaningful data).
     fn is_empty(&self) -> bool {
-        self.total_fetch_elapsed.is_zero()
+        self.total_fetch_elapsed.is_zero() && self.prefilter_cost.is_zero()
     }
 }
 
 /// Metrics for tracking page/row group fetch operations.
-#[derive(Default)]
+#[derive(Default, Clone)]
 pub struct ParquetFetchMetrics {
-    pub data: std::sync::Mutex<ParquetFetchMetricsData>,
+    pub data: Arc<std::sync::Mutex<ParquetFetchMetricsData>>,
 }
 
 impl std::fmt::Debug for ParquetFetchMetrics {
@@ -100,6 +88,8 @@ impl std::fmt::Debug for ParquetFetchMetrics {
             write_cache_fetch_elapsed,
             store_fetch_elapsed,
             total_fetch_elapsed,
+            prefilter_cost,
+            prefilter_filtered_rows,
         } = *data;
 
         write!(f, "{{")?;
@@ -158,6 +148,16 @@ impl std::fmt::Debug for ParquetFetchMetrics {
         if !store_fetch_elapsed.is_zero() {
             write!(f, ", \"store_fetch_elapsed\":\"{:?}\"", store_fetch_elapsed)?;
         }
+        if !prefilter_cost.is_zero() {
+            write!(f, ", \"prefilter_cost\":\"{:?}\"", prefilter_cost)?;
+        }
+        if prefilter_filtered_rows > 0 {
+            write!(
+                f,
+                ", \"prefilter_filtered_rows\":{}",
+                prefilter_filtered_rows
+            )?;
+        }
 
         write!(f, "}}")
     }
@@ -185,6 +185,8 @@ impl ParquetFetchMetrics {
             write_cache_fetch_elapsed,
             store_fetch_elapsed,
             total_fetch_elapsed,
+            prefilter_cost,
+            prefilter_filtered_rows,
         } = *other.data.lock().unwrap();
 
         let mut data = self.data.lock().unwrap();
@@ -201,357 +203,8 @@ impl ParquetFetchMetrics {
         data.write_cache_fetch_elapsed += write_cache_fetch_elapsed;
         data.store_fetch_elapsed += store_fetch_elapsed;
         data.total_fetch_elapsed += total_fetch_elapsed;
-    }
-}
-
-pub(crate) struct RowGroupBase<'a> {
-    parquet_metadata: &'a ParquetMetaData,
-    row_group_idx: usize,
-    pub(crate) offset_index: Option<&'a [OffsetIndexMetaData]>,
-    /// Compressed page of each column.
-    column_chunks: Vec<Option<Arc<ColumnChunkData>>>,
-    pub(crate) row_count: usize,
-}
-
-impl<'a> RowGroupBase<'a> {
-    pub(crate) fn new(parquet_meta: &'a ParquetMetaData, row_group_idx: usize) -> Self {
-        let metadata = parquet_meta.row_group(row_group_idx);
-        // `offset_index` is always `None` if we don't set
-        // [with_page_index()](https://docs.rs/parquet/latest/parquet/arrow/arrow_reader/struct.ArrowReaderOptions.html#method.with_page_index)
-        // to `true`.
-        let offset_index = parquet_meta
-            .offset_index()
-            // filter out empty offset indexes (old versions specified Some(vec![]) when no present)
-            .filter(|index| !index.is_empty())
-            .map(|x| x[row_group_idx].as_slice());
-
-        Self {
-            parquet_metadata: parquet_meta,
-            row_group_idx,
-            offset_index,
-            column_chunks: vec![None; metadata.columns().len()],
-            row_count: metadata.num_rows() as usize,
-        }
-    }
-
-    pub(crate) fn calc_sparse_read_ranges(
-        &self,
-        projection: &ProjectionMask,
-        offset_index: &[OffsetIndexMetaData],
-        selection: &RowSelection,
-    ) -> (Vec<Range<u64>>, Vec<Vec<usize>>) {
-        // If we have a `RowSelection` and an `OffsetIndex` then only fetch pages required for the
-        // `RowSelection`
-        let mut page_start_offsets: Vec<Vec<usize>> = vec![];
-        let ranges = self
-            .column_chunks
-            .iter()
-            .zip(self.row_group_metadata().columns())
-            .enumerate()
-            .filter(|&(idx, (chunk, _chunk_meta))| chunk.is_none() && projection.leaf_included(idx))
-            .flat_map(|(idx, (_chunk, chunk_meta))| {
-                // If the first page does not start at the beginning of the column,
-                // then we need to also fetch a dictionary page.
-                let mut ranges = vec![];
-                let (start, _len) = chunk_meta.byte_range();
-                match offset_index[idx].page_locations.first() {
-                    Some(first) if first.offset as u64 != start => {
-                        ranges.push(start..first.offset as u64);
-                    }
-                    _ => (),
-                }
-
-                ranges.extend(
-                    selection
-                        .scan_ranges(&offset_index[idx].page_locations)
-                        .iter()
-                        .map(|range| range.start..range.end),
-                );
-                page_start_offsets.push(ranges.iter().map(|range| range.start as usize).collect());
-
-                ranges
-            })
-            .collect::<Vec<_>>();
-        (ranges, page_start_offsets)
-    }
-
-    pub(crate) fn assign_sparse_chunk(
-        &mut self,
-        projection: &ProjectionMask,
-        data: Vec<Bytes>,
-        page_start_offsets: Vec<Vec<usize>>,
-    ) {
-        let mut page_start_offsets = page_start_offsets.into_iter();
-        let mut chunk_data = data.into_iter();
-
-        for (idx, chunk) in self.column_chunks.iter_mut().enumerate() {
-            if chunk.is_some() || !projection.leaf_included(idx) {
-                continue;
-            }
-
-            if let Some(offsets) = page_start_offsets.next() {
-                let mut chunks = Vec::with_capacity(offsets.len());
-                for _ in 0..offsets.len() {
-                    chunks.push(chunk_data.next().unwrap());
-                }
-
-                let column = self
-                    .parquet_metadata
-                    .row_group(self.row_group_idx)
-                    .column(idx);
-                *chunk = Some(Arc::new(ColumnChunkData::Sparse {
-                    length: column.byte_range().1 as usize,
-                    data: offsets.into_iter().zip(chunks).collect(),
-                }))
-            }
-        }
-    }
-
-    pub(crate) fn calc_dense_read_ranges(&self, projection: &ProjectionMask) -> Vec<Range<u64>> {
-        self.column_chunks
-            .iter()
-            .enumerate()
-            .filter(|&(idx, chunk)| chunk.is_none() && projection.leaf_included(idx))
-            .map(|(idx, _chunk)| {
-                let column = self.row_group_metadata().column(idx);
-                let (start, length) = column.byte_range();
-                start..(start + length)
-            })
-            .collect::<Vec<_>>()
-    }
-
-    /// Assigns compressed chunk binary data to [RowGroupBase::column_chunks]
-    /// and returns the chunk offset and binary data assigned.
-    pub(crate) fn assign_dense_chunk(
-        &mut self,
-        projection: &ProjectionMask,
-        chunk_data: Vec<Bytes>,
-    ) {
-        let mut chunk_data = chunk_data.into_iter();
-
-        for (idx, chunk) in self.column_chunks.iter_mut().enumerate() {
-            if chunk.is_some() || !projection.leaf_included(idx) {
-                continue;
-            }
-
-            // Get the fetched page.
-            let Some(data) = chunk_data.next() else {
-                continue;
-            };
-
-            let column = self
-                .parquet_metadata
-                .row_group(self.row_group_idx)
-                .column(idx);
-            *chunk = Some(Arc::new(ColumnChunkData::Dense {
-                offset: column.byte_range().0 as usize,
-                data,
-            }));
-        }
-    }
-
-    /// Create [PageReader] from [RowGroupBase::column_chunks]
-    pub(crate) fn column_reader(
-        &self,
-        col_idx: usize,
-    ) -> Result<SerializedPageReader<ColumnChunkData>> {
-        let page_reader = match &self.column_chunks[col_idx] {
-            None => {
-                return Err(ParquetError::General(format!(
-                    "Invalid column index {col_idx}, column was not fetched"
-                )));
-            }
-            Some(data) => {
-                let page_locations = self
-                    .offset_index
-                    // filter out empty offset indexes (old versions specified Some(vec![]) when no present)
-                    .filter(|index| !index.is_empty())
-                    .map(|index| index[col_idx].page_locations.clone());
-                SerializedPageReader::new(
-                    data.clone(),
-                    self.row_group_metadata().column(col_idx),
-                    self.row_count,
-                    page_locations,
-                )?
-            }
-        };
-
-        Ok(page_reader)
-    }
-
-    pub(crate) fn parquet_metadata(&self) -> &ParquetMetaData {
-        self.parquet_metadata
-    }
-
-    pub(crate) fn row_group_metadata(&self) -> &RowGroupMetaData {
-        self.parquet_metadata().row_group(self.row_group_idx)
-    }
-}
-
-/// An in-memory collection of column chunks
-pub struct InMemoryRowGroup<'a> {
-    region_id: RegionId,
-    file_id: FileId,
-    row_group_idx: usize,
-    cache_strategy: CacheStrategy,
-    file_path: &'a str,
-    /// Object store.
-    object_store: ObjectStore,
-    base: RowGroupBase<'a>,
-}
-
-impl<'a> InMemoryRowGroup<'a> {
-    /// Creates a new [InMemoryRowGroup] by `row_group_idx`.
-    ///
-    /// # Panics
-    /// Panics if the `row_group_idx` is invalid.
-    pub fn create(
-        region_id: RegionId,
-        file_id: FileId,
-        parquet_meta: &'a ParquetMetaData,
-        row_group_idx: usize,
-        cache_strategy: CacheStrategy,
-        file_path: &'a str,
-        object_store: ObjectStore,
-    ) -> Self {
-        Self {
-            region_id,
-            file_id,
-            row_group_idx,
-            cache_strategy,
-            file_path,
-            object_store,
-            base: RowGroupBase::new(parquet_meta, row_group_idx),
-        }
-    }
-
-    /// Fetches the necessary column data into memory
-    pub async fn fetch(
-        &mut self,
-        projection: &ProjectionMask,
-        selection: Option<&RowSelection>,
-        metrics: Option<&ParquetFetchMetrics>,
-    ) -> Result<()> {
-        if let Some((selection, offset_index)) = selection.zip(self.base.offset_index) {
-            let (fetch_ranges, page_start_offsets) =
-                self.base
-                    .calc_sparse_read_ranges(projection, offset_index, selection);
-
-            let chunk_data = self.fetch_bytes(&fetch_ranges, metrics).await?;
-            // Assign sparse chunk data to base.
-            self.base
-                .assign_sparse_chunk(projection, chunk_data, page_start_offsets);
-        } else {
-            // Release the CPU to avoid blocking the runtime. Since `fetch_pages_from_cache`
-            // is a synchronous, CPU-bound operation.
-            yield_now().await;
-
-            // Calculate ranges to read.
-            let fetch_ranges = self.base.calc_dense_read_ranges(projection);
-
-            if fetch_ranges.is_empty() {
-                // Nothing to fetch.
-                return Ok(());
-            }
-
-            // Fetch data with ranges
-            let chunk_data = self.fetch_bytes(&fetch_ranges, metrics).await?;
-
-            // Assigns fetched data to base.
-            self.base.assign_dense_chunk(projection, chunk_data);
-        }
-
-        Ok(())
-    }
-
-    /// Try to fetch data from the memory cache or the WriteCache,
-    /// if not in WriteCache, fetch data from object store directly.
-    async fn fetch_bytes(
-        &self,
-        ranges: &[Range<u64>],
-        metrics: Option<&ParquetFetchMetrics>,
-    ) -> Result<Vec<Bytes>> {
-        // Now fetch page timer includes the whole time to read pages.
-        let _timer = READ_STAGE_FETCH_PAGES.start_timer();
-
-        let page_key = PageKey::new(self.file_id, self.row_group_idx, ranges.to_vec());
-        if let Some(pages) = self.cache_strategy.get_pages(&page_key) {
-            if let Some(metrics) = metrics {
-                let total_size: u64 = ranges.iter().map(|r| r.end - r.start).sum();
-                let mut metrics_data = metrics.data.lock().unwrap();
-                metrics_data.page_cache_hit += 1;
-                metrics_data.pages_to_fetch_mem += ranges.len();
-                metrics_data.page_size_to_fetch_mem += total_size;
-                metrics_data.page_size_needed += total_size;
-            }
-            return Ok(pages.compressed.clone());
-        }
-
-        // Calculate total range size for metrics.
-        let (total_range_size, unaligned_size) = compute_total_range_size(ranges);
-
-        let key = IndexKey::new(self.region_id, self.file_id, FileType::Parquet);
-        let fetch_write_cache_start = metrics.map(|_| std::time::Instant::now());
-        let write_cache_result = self.fetch_ranges_from_write_cache(key, ranges).await;
-        let pages = match write_cache_result {
-            Some(data) => {
-                if let Some(metrics) = metrics {
-                    let elapsed = fetch_write_cache_start
-                        .map(|start| start.elapsed())
-                        .unwrap_or_default();
-                    let range_size_needed: u64 = ranges.iter().map(|r| r.end - r.start).sum();
-                    let mut metrics_data = metrics.data.lock().unwrap();
-                    metrics_data.write_cache_fetch_elapsed += elapsed;
-                    metrics_data.write_cache_hit += 1;
-                    metrics_data.pages_to_fetch_write_cache += ranges.len();
-                    metrics_data.page_size_to_fetch_write_cache += unaligned_size;
-                    metrics_data.page_size_needed += range_size_needed;
-                }
-                data
-            }
-            None => {
-                // Fetch data from object store.
-                let _timer = READ_STAGE_ELAPSED
-                    .with_label_values(&["cache_miss_read"])
-                    .start_timer();
-
-                let start = metrics.map(|_| std::time::Instant::now());
-                let data = fetch_byte_ranges(self.file_path, self.object_store.clone(), ranges)
-                    .await
-                    .map_err(|e| ParquetError::External(Box::new(e)))?;
-                if let Some(metrics) = metrics {
-                    let elapsed = start.map(|start| start.elapsed()).unwrap_or_default();
-                    let range_size_needed: u64 = ranges.iter().map(|r| r.end - r.start).sum();
-                    let mut metrics_data = metrics.data.lock().unwrap();
-                    metrics_data.store_fetch_elapsed += elapsed;
-                    metrics_data.cache_miss += 1;
-                    metrics_data.pages_to_fetch_store += ranges.len();
-                    metrics_data.page_size_to_fetch_store += unaligned_size;
-                    metrics_data.page_size_needed += range_size_needed;
-                }
-                data
-            }
-        };
-
-        // Put pages back to the cache.
-        let page_value = PageValue::new(pages.clone(), total_range_size);
-        self.cache_strategy
-            .put_pages(page_key, Arc::new(page_value));
-
-        Ok(pages)
-    }
-
-    /// Fetches data from write cache.
-    /// Returns `None` if the data is not in the cache.
-    async fn fetch_ranges_from_write_cache(
-        &self,
-        key: IndexKey,
-        ranges: &[Range<u64>],
-    ) -> Option<Vec<Bytes>> {
-        if let Some(cache) = self.cache_strategy.write_cache() {
-            return cache.file_cache().read_ranges(key, ranges).await;
-        }
-        None
+        data.prefilter_cost += prefilter_cost;
+        data.prefilter_filtered_rows += prefilter_filtered_rows;
     }
 }
 
@@ -560,14 +213,14 @@ impl<'a> InMemoryRowGroup<'a> {
 /// - aligned_size: total size aligned to pooled buffer size
 /// - unaligned_size: actual total size without alignment
 // See https://github.com/apache/opendal/blob/v0.54.0/core/src/types/read/reader.rs#L166-L192
-fn compute_total_range_size(ranges: &[Range<u64>]) -> (u64, u64) {
+pub(crate) fn compute_total_range_size(ranges: &[Range<u64>]) -> (u64, u64) {
     if ranges.is_empty() {
         return (0, 0);
     }
 
     let gap = MERGE_GAP as u64;
     let mut sorted_ranges = ranges.to_vec();
-    sorted_ranges.sort_unstable_by(|a, b| a.start.cmp(&b.start));
+    sorted_ranges.sort_unstable_by_key(|a| a.start);
 
     let mut total_size_aligned = 0;
     let mut total_size_unaligned = 0;
@@ -602,96 +255,3 @@ fn align_to_pooled_buf_size(size: u64) -> u64 {
     const POOLED_BUF_SIZE: u64 = 2 * 1024 * 1024;
     size.div_ceil(POOLED_BUF_SIZE) * POOLED_BUF_SIZE
 }
-
-impl RowGroups for InMemoryRowGroup<'_> {
-    fn num_rows(&self) -> usize {
-        self.base.row_count
-    }
-
-    fn column_chunks(&self, i: usize) -> Result<Box<dyn PageIterator>> {
-        // Creates a page reader to read column at `i`.
-        let page_reader = self.base.column_reader(i)?;
-
-        Ok(Box::new(ColumnChunkIterator {
-            reader: Some(Ok(Box::new(page_reader))),
-        }))
-    }
-
-    fn row_groups(&self) -> Box<dyn Iterator<Item = &RowGroupMetaData> + '_> {
-        Box::new(std::iter::once(self.base.row_group_metadata()))
-    }
-
-    fn metadata(&self) -> &ParquetMetaData {
-        self.base.parquet_metadata()
-    }
-}
-
-/// An in-memory column chunk
-#[derive(Clone)]
-pub(crate) enum ColumnChunkData {
-    /// Column chunk data representing only a subset of data pages
-    Sparse {
-        /// Length of the full column chunk
-        length: usize,
-        /// Set of data pages included in this sparse chunk. Each element is a tuple
-        /// of (page offset, page data)
-        data: Vec<(usize, Bytes)>,
-    },
-    /// Full column chunk and its offset
-    Dense { offset: usize, data: Bytes },
-}
-
-impl ColumnChunkData {
-    fn get(&self, start: u64) -> Result<Bytes> {
-        match &self {
-            ColumnChunkData::Sparse { data, .. } => data
-                .binary_search_by_key(&start, |(offset, _)| *offset as u64)
-                .map(|idx| data[idx].1.clone())
-                .map_err(|_| {
-                    ParquetError::General(format!(
-                        "Invalid offset in sparse column chunk data: {start}"
-                    ))
-                }),
-            ColumnChunkData::Dense { offset, data } => {
-                let start = start as usize - *offset;
-                Ok(data.slice(start..))
-            }
-        }
-    }
-}
-
-impl Length for ColumnChunkData {
-    fn len(&self) -> u64 {
-        match &self {
-            ColumnChunkData::Sparse { length, .. } => *length as u64,
-            ColumnChunkData::Dense { data, .. } => data.len() as u64,
-        }
-    }
-}
-
-impl ChunkReader for ColumnChunkData {
-    type T = bytes::buf::Reader<Bytes>;
-
-    fn get_read(&self, start: u64) -> Result<Self::T> {
-        Ok(self.get(start)?.reader())
-    }
-
-    fn get_bytes(&self, start: u64, length: usize) -> Result<Bytes> {
-        Ok(self.get(start)?.slice(..length))
-    }
-}
-
-/// Implements [`PageIterator`] for a single column chunk, yielding a single [`PageReader`]
-pub(crate) struct ColumnChunkIterator {
-    pub(crate) reader: Option<Result<Box<dyn PageReader>>>,
-}
-
-impl Iterator for ColumnChunkIterator {
-    type Item = Result<Box<dyn PageReader>>;
-
-    fn next(&mut self) -> Option<Self::Item> {
-        self.reader.take()
-    }
-}
-
-impl PageIterator for ColumnChunkIterator {}
diff --git a/src/mito2/src/sst/parquet/row_selection.rs b/src/mito2/src/sst/parquet/row_selection.rs
index 595f1d352a..cf28734ffc 100644
--- a/src/mito2/src/sst/parquet/row_selection.rs
+++ b/src/mito2/src/sst/parquet/row_selection.rs
@@ -554,11 +554,23 @@ fn intersect_row_selections(left: &RowSelection, right: &RowSelection) -> RowSel
 /// or if there's a gap that requires skipping rows. It handles both "select" and "skip" actions,
 /// optimizing the list of selectors by merging contiguous actions of the same type.
 ///
+/// The returned selection intentionally stops at the end of the last matched range and may omit a
+/// trailing `skip` that would extend it to `total_row_count`. That is fine when the selection is
+/// used directly by the parquet reader, which simply stops once the selectors are exhausted.
+///
 /// Note: overlapping ranges are not supported and will result in an incorrect selection.
 pub(crate) fn row_selection_from_row_ranges(
     row_ranges: impl Iterator<Item = Range<usize>>,
     total_row_count: usize,
 ) -> RowSelection {
+    let (selectors, _) = build_selectors_from_row_ranges(row_ranges, total_row_count);
+    RowSelection::from(selectors)
+}
+
+fn build_selectors_from_row_ranges(
+    row_ranges: impl Iterator<Item = Range<usize>>,
+    total_row_count: usize,
+) -> (Vec<RowSelector>, usize) {
     let mut selectors: Vec<RowSelector> = Vec::new();
     let mut last_processed_end = 0;
 
@@ -572,7 +584,7 @@ pub(crate) fn row_selection_from_row_ranges(
         last_processed_end = end;
     }
 
-    RowSelection::from(selectors)
+    (selectors, last_processed_end)
 }
 
 /// Converts an iterator of sorted row IDs into a `RowSelection`.
diff --git a/src/mito2/src/sst/parquet/stats.rs b/src/mito2/src/sst/parquet/stats.rs
index 7c5da69d4b..cdedde94e2 100644
--- a/src/mito2/src/sst/parquet/stats.rs
+++ b/src/mito2/src/sst/parquet/stats.rs
@@ -26,14 +26,15 @@ use parquet::file::metadata::RowGroupMetaData;
 use store_api::metadata::RegionMetadataRef;
 use store_api::storage::ColumnId;
 
-use crate::sst::parquet::format::{ReadFormat, StatValues};
+use crate::sst::parquet::flat_format::FlatReadFormat;
+use crate::sst::parquet::format::StatValues;
 
 /// Statistics for pruning row groups.
 pub(crate) struct RowGroupPruningStats<'a, T> {
     /// Metadata of SST row groups.
     row_groups: &'a [T],
     /// Helper to read the SST.
-    read_format: &'a ReadFormat,
+    read_format: &'a FlatReadFormat,
     /// The metadata of the region.
     /// It contains the schema a query expects to read. If it is not None, we use it instead
     /// of the metadata in the SST to get the column id of a column as the SST may have
@@ -47,7 +48,7 @@ impl<'a, T> RowGroupPruningStats<'a, T> {
     /// Creates a new statistics to prune specific `row_groups`.
     pub(crate) fn new(
         row_groups: &'a [T],
-        read_format: &'a ReadFormat,
+        read_format: &'a FlatReadFormat,
         expected_metadata: Option<RegionMetadataRef>,
         skip_fields: bool,
     ) -> Self {
diff --git a/src/mito2/src/sst/parquet/writer.rs b/src/mito2/src/sst/parquet/writer.rs
index b207f11ef8..13005ff9fc 100644
--- a/src/mito2/src/sst/parquet/writer.rs
+++ b/src/mito2/src/sst/parquet/writer.rs
@@ -50,7 +50,7 @@ use crate::config::{IndexBuildMode, IndexConfig};
 use crate::error::{
     InvalidMetadataSnafu, OpenDalSnafu, Result, UnexpectedSnafu, WriteParquetSnafu,
 };
-use crate::read::{Batch, FlatSource, Source};
+use crate::read::FlatSource;
 use crate::sst::file::RegionFileId;
 use crate::sst::index::{IndexOutput, Indexer, IndexerBuilder};
 use crate::sst::parquet::flat_format::{FlatWriteFormat, time_index_column_index};
@@ -60,6 +60,28 @@ use crate::sst::{
     DEFAULT_WRITE_BUFFER_SIZE, DEFAULT_WRITE_CONCURRENCY, FlatSchemaOptions, SeriesEstimator,
 };
 
+/// Converts a flat RecordBatch for writing to parquet.
+enum FlatBatchConverter {
+    /// Write as-is in flat format.
+    Flat(FlatWriteFormat),
+    /// Convert flat batch to primary-key format by stripping tag columns.
+    PrimaryKey {
+        format: PrimaryKeyWriteFormat,
+        num_fields: usize,
+    },
+}
+
+impl FlatBatchConverter {
+    fn convert_batch(&self, batch: &RecordBatch) -> Result<RecordBatch> {
+        match self {
+            FlatBatchConverter::Flat(f) => f.convert_batch(batch),
+            FlatBatchConverter::PrimaryKey { format, num_fields } => {
+                format.convert_flat_batch(batch, *num_fields)
+            }
+        }
+    }
+}
+
 /// Parquet SST writer.
 pub struct ParquetWriter<'a, F: WriterFactory, I: IndexerBuilder, P: FilePathProvider> {
     /// Path provider that creates SST and index file paths according to file id.
@@ -240,81 +262,6 @@ where
         Ok(())
     }
 
-    /// Iterates source and writes all rows to Parquet file.
-    ///
-    /// Returns the [SstInfo] if the SST is written.
-    pub async fn write_all(
-        &mut self,
-        source: Source,
-        override_sequence: Option<SequenceNumber>, // override the `sequence` field from `Source`
-        opts: &WriteOptions,
-    ) -> Result<SstInfoArray> {
-        let res = self
-            .write_all_without_cleaning(source, override_sequence, opts)
-            .await;
-        if res.is_err() {
-            // Clean tmp files explicitly on failure.
-            let file_id = self.current_file;
-            if let Some(cleaner) = &self.file_cleaner {
-                cleaner.clean_by_file_id(file_id).await;
-            }
-        }
-        res
-    }
-
-    async fn write_all_without_cleaning(
-        &mut self,
-        mut source: Source,
-        override_sequence: Option<SequenceNumber>, // override the `sequence` field from `Source`
-        opts: &WriteOptions,
-    ) -> Result<SstInfoArray> {
-        let mut results = smallvec![];
-        let write_format = PrimaryKeyWriteFormat::new(self.metadata.clone())
-            .with_override_sequence(override_sequence);
-        let mut stats = SourceStats::default();
-
-        while let Some(res) = self
-            .write_next_batch(&mut source, &write_format, opts)
-            .await
-            .transpose()
-        {
-            match res {
-                Ok(mut batch) => {
-                    stats.update(&batch);
-                    let start = Instant::now();
-                    // safety: self.current_indexer must be set when first batch has been written.
-                    match self.index_config.build_mode {
-                        IndexBuildMode::Sync => {
-                            self.current_indexer
-                                .as_mut()
-                                .unwrap()
-                                .update(&mut batch)
-                                .await;
-                        }
-                        IndexBuildMode::Async => {}
-                    }
-                    self.metrics.update_index += start.elapsed();
-                    if let Some(max_file_size) = opts.max_file_size
-                        && self.bytes_written.load(Ordering::Relaxed) > max_file_size
-                    {
-                        self.finish_current_file(&mut results, &mut stats).await?;
-                    }
-                }
-                Err(e) => {
-                    if let Some(indexer) = &mut self.current_indexer {
-                        indexer.abort().await;
-                    }
-                    return Err(e);
-                }
-            }
-        }
-
-        self.finish_current_file(&mut results, &mut stats).await?;
-
-        // object_store.write will make sure all bytes are written or an error is raised.
-        Ok(results)
-    }
-
     /// Iterates FlatSource and writes all RecordBatch in flat format to Parquet file.
     ///
     /// Returns the [SstInfo] if the SST is written.
@@ -324,11 +271,15 @@ where
         override_sequence: Option<SequenceNumber>,
         opts: &WriteOptions,
     ) -> Result<SstInfoArray> {
-        let res = self
-            .write_all_flat_without_cleaning(source, override_sequence, opts)
-            .await;
+        let converter = FlatBatchConverter::Flat(
+            FlatWriteFormat::new(
+                self.metadata.clone(),
+                &FlatSchemaOptions::from_encoding(self.metadata.primary_key_encoding),
+            )
+            .with_override_sequence(override_sequence),
+        );
+        let res = self.write_all_flat_inner(source, &converter, opts).await;
         if res.is_err() {
-            // Clean tmp files explicitly on failure.
             let file_id = self.current_file;
             if let Some(cleaner) = &self.file_cleaner {
                 cleaner.clean_by_file_id(file_id).await;
@@ -337,36 +288,58 @@ where
         res
     }
 
-    async fn write_all_flat_without_cleaning(
+    /// Iterates FlatSource and writes all RecordBatch in primary-key format to Parquet file.
+    ///
+    /// Returns the [SstInfo] if the SST is written.
+    pub async fn write_all_flat_as_primary_key(
         &mut self,
-        mut source: FlatSource,
+        source: FlatSource,
         override_sequence: Option<SequenceNumber>,
         opts: &WriteOptions,
+    ) -> Result<SstInfoArray> {
+        let num_fields = self.metadata.field_columns().count();
+        let converter = FlatBatchConverter::PrimaryKey {
+            format: PrimaryKeyWriteFormat::new(self.metadata.clone())
+                .with_override_sequence(override_sequence),
+            num_fields,
+        };
+        let res = self.write_all_flat_inner(source, &converter, opts).await;
+        if res.is_err() {
+            let file_id = self.current_file;
+            if let Some(cleaner) = &self.file_cleaner {
+                cleaner.clean_by_file_id(file_id).await;
+            }
+        }
+        res
+    }
+
+    async fn write_all_flat_inner(
+        &mut self,
+        mut source: FlatSource,
+        converter: &FlatBatchConverter,
+        opts: &WriteOptions,
     ) -> Result<SstInfoArray> {
         let mut results = smallvec![];
-        let flat_format = FlatWriteFormat::new(
-            self.metadata.clone(),
-            &FlatSchemaOptions::from_encoding(self.metadata.primary_key_encoding),
-        )
-        .with_override_sequence(override_sequence);
         let mut stats = SourceStats::default();
 
         while let Some(record_batch) = self
-            .write_next_flat_batch(&mut source, &flat_format, opts)
+            .write_next_flat_batch(&mut source, converter, opts)
             .await
             .transpose()
         {
             match record_batch {
                 Ok(batch) => {
                     stats.update_flat(&batch)?;
-                    let start = Instant::now();
-                    // safety: self.current_indexer must be set when first batch has been written.
-                    self.current_indexer
-                        .as_mut()
-                        .unwrap()
-                        .update_flat(&batch)
-                        .await;
-                    self.metrics.update_index += start.elapsed();
+                    if matches!(self.index_config.build_mode, IndexBuildMode::Sync) {
+                        let start = Instant::now();
+                        // safety: self.current_indexer must be set when first batch has been written.
+                        self.current_indexer
+                            .as_mut()
+                            .unwrap()
+                            .update_flat(&batch)
+                            .await;
+                        self.metrics.update_index += start.elapsed();
+                    }
                     if let Some(max_file_size) = opts.max_file_size
                         && self.bytes_written.load(Ordering::Relaxed) > max_file_size
                     {
@@ -411,34 +384,10 @@ where
             .set_column_compression(op_type_col, Compression::UNCOMPRESSED)
     }
 
-    async fn write_next_batch(
-        &mut self,
-        source: &mut Source,
-        write_format: &PrimaryKeyWriteFormat,
-        opts: &WriteOptions,
-    ) -> Result<Option<Batch>> {
-        let start = Instant::now();
-        let Some(batch) = source.next_batch().await? else {
-            return Ok(None);
-        };
-        self.metrics.iter_source += start.elapsed();
-
-        let arrow_batch = write_format.convert_batch(&batch)?;
-
-        let start = Instant::now();
-        self.maybe_init_writer(write_format.arrow_schema(), opts)
-            .await?
-            .write(&arrow_batch)
-            .await
-            .context(WriteParquetSnafu)?;
-        self.metrics.write_batch += start.elapsed();
-        Ok(Some(batch))
-    }
-
     async fn write_next_flat_batch(
         &mut self,
         source: &mut FlatSource,
-        flat_format: &FlatWriteFormat,
+        converter: &FlatBatchConverter,
         opts: &WriteOptions,
     ) -> Result<Option<RecordBatch>> {
         let start = Instant::now();
@@ -447,15 +396,16 @@ where
         };
         self.metrics.iter_source += start.elapsed();
 
-        let arrow_batch = flat_format.convert_batch(&record_batch)?;
+        let arrow_batch = converter.convert_batch(&record_batch)?;
 
         let start = Instant::now();
-        self.maybe_init_writer(flat_format.arrow_schema(), opts)
+        self.maybe_init_writer(arrow_batch.schema_ref(), opts)
             .await?
             .write(&arrow_batch)
             .await
             .context(WriteParquetSnafu)?;
         self.metrics.write_batch += start.elapsed();
+        // Return original flat batch for stats/indexer which use flat layout.
         Ok(Some(record_batch))
     }
 
@@ -515,26 +465,6 @@ struct SourceStats {
 }
 
 impl SourceStats {
-    fn update(&mut self, batch: &Batch) {
-        if batch.is_empty() {
-            return;
-        }
-
-        self.num_rows += batch.num_rows();
-        self.series_estimator.update(batch);
-        // Safety: batch is not empty.
-        let (min_in_batch, max_in_batch) = (
-            batch.first_timestamp().unwrap(),
-            batch.last_timestamp().unwrap(),
-        );
-        if let Some(time_range) = &mut self.time_range {
-            time_range.0 = time_range.0.min(min_in_batch);
-            time_range.1 = time_range.1.max(max_in_batch);
-        } else {
-            self.time_range = Some((min_in_batch, max_in_batch));
-        }
-    }
-
     fn update_flat(&mut self, record_batch: &RecordBatch) -> Result<()> {
         if record_batch.num_rows() == 0 {
             return Ok(());
diff --git a/src/mito2/src/sst/version.rs b/src/mito2/src/sst/version.rs
index a9e71eb5d9..5958cf7513 100644
--- a/src/mito2/src/sst/version.rs
+++ b/src/mito2/src/sst/version.rs
@@ -78,7 +78,9 @@ impl SstVersion {
                         *f = FileHandle::new(file.clone(), file_purger.clone());
                     }
                 })
-                .or_insert_with(|| FileHandle::new(file.clone(), file_purger.clone()));
+                .or_insert_with(|| {
+                    FileHandle::new(file.clone(), file_purger.clone())
+                });
         }
     }
 
diff --git a/src/mito2/src/test_util.rs b/src/mito2/src/test_util.rs
index 842689bba6..c4ca7cc7b3 100644
--- a/src/mito2/src/test_util.rs
+++ b/src/mito2/src/test_util.rs
@@ -15,6 +15,7 @@
 //! Utilities for testing.
 
 pub mod batch_util;
+pub mod bench_util;
 pub mod memtable_util;
 pub mod scheduler_util;
 pub mod sst_util;
@@ -609,9 +610,16 @@ impl TestEnv {
         let manifest_dir = data_home.join("manifest").as_path().display().to_string();
 
         let builder = Fs::default();
-        let object_store = ObjectStore::new(builder.root(&manifest_dir))
-            .unwrap()
-            .finish();
+        let object_store = if let Some(mock_layer) = self.object_store_mock_layer.as_ref() {
+            ObjectStore::new(builder.root(&manifest_dir))
+                .unwrap()
+                .layer(mock_layer.clone())
+                .finish()
+        } else {
+            ObjectStore::new(builder.root(&manifest_dir))
+                .unwrap()
+                .finish()
+        };
 
         // The "manifest_dir" here should be the relative path from the `object_store`'s root.
         // Otherwise the OpenDal's list operation would fail with "StripPrefixError". This is
@@ -1239,7 +1247,10 @@ pub async fn flush_region(engine: &MitoEngine, region_id: RegionId, row_group_si
     let result = engine
         .handle_request(
             region_id,
-            RegionRequest::Flush(RegionFlushRequest { row_group_size }),
+            RegionRequest::Flush(RegionFlushRequest {
+                row_group_size,
+                reason: None,
+            }),
         )
         .await
         .unwrap();
diff --git a/src/mito2/src/test_util/bench_util.rs b/src/mito2/src/test_util/bench_util.rs
new file mode 100644
index 0000000000..8f182e4157
--- /dev/null
+++ b/src/mito2/src/test_util/bench_util.rs
@@ -0,0 +1,259 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Shared utilities for mito2 benchmarks.
+//!
+//! Provides a TSBS cpu-like data generator ([`CpuDataGenerator`]) and schema
+//! ([`cpu_metadata`]) used by multiple benchmark binaries in this directory.
+
+use api::v1::value::ValueData;
+use api::v1::{Row, Rows, SemanticType};
+use datafusion_common::Column;
+use datafusion_expr::{Expr, lit};
+use datatypes::data_type::ConcreteDataType;
+use datatypes::schema::ColumnSchema;
+use rand::Rng;
+use rand::rngs::ThreadRng;
+use rand::seq::IndexedRandom;
+use store_api::metadata::{
+    ColumnMetadata, RegionMetadata, RegionMetadataBuilder, RegionMetadataRef,
+};
+use store_api::storage::RegionId;
+use table::predicate::Predicate;
+
+use crate::memtable::KeyValues;
+use crate::test_util::memtable_util::region_metadata_to_row_schema;
+
+pub struct Host {
+    pub hostname: String,
+    pub region: String,
+    pub datacenter: String,
+    pub rack: String,
+    pub os: String,
+    pub arch: String,
+    pub team: String,
+    pub service: String,
+    pub service_version: String,
+    pub service_environment: String,
+}
+
+impl Host {
+    pub fn random_with_id(id: usize) -> Host {
+        let mut rng = rand::rng();
+        let region = format!("ap-southeast-{}", rng.random_range(0..10));
+        let datacenter = format!(
+            "{}{}",
+            region,
+            ['a', 'b', 'c', 'd', 'e'].choose(&mut rng).unwrap()
+        );
+        Host {
+            hostname: format!("host_{id}"),
+            region,
+            datacenter,
+            rack: rng.random_range(0..100).to_string(),
+            os: "Ubuntu16.04LTS".to_string(),
+            arch: "x86".to_string(),
+            team: "CHI".to_string(),
+            service: rng.random_range(0..100).to_string(),
+            service_version: rng.random_range(0..10).to_string(),
+            service_environment: "test".to_string(),
+        }
+    }
+
+    pub fn fill_values(&self, values: &mut Vec<api::v1::Value>) {
+        let tags = [
+            api::v1::Value {
+                value_data: Some(ValueData::StringValue(self.hostname.clone())),
+            },
+            api::v1::Value {
+                value_data: Some(ValueData::StringValue(self.region.clone())),
+            },
+            api::v1::Value {
+                value_data: Some(ValueData::StringValue(self.datacenter.clone())),
+            },
+            api::v1::Value {
+                value_data: Some(ValueData::StringValue(self.rack.clone())),
+            },
+            api::v1::Value {
+                value_data: Some(ValueData::StringValue(self.os.clone())),
+            },
+            api::v1::Value {
+                value_data: Some(ValueData::StringValue(self.arch.clone())),
+            },
+            api::v1::Value {
+                value_data: Some(ValueData::StringValue(self.team.clone())),
+            },
+            api::v1::Value {
+                value_data: Some(ValueData::StringValue(self.service.clone())),
+            },
+            api::v1::Value {
+                value_data: Some(ValueData::StringValue(self.service_version.clone())),
+            },
+            api::v1::Value {
+                value_data: Some(ValueData::StringValue(self.service_environment.clone())),
+            },
+        ];
+        for tag in tags {
+            values.push(tag);
+        }
+    }
+}
+
+pub struct CpuDataGenerator {
+    pub metadata: RegionMetadataRef,
+    column_schemas: Vec<api::v1::ColumnSchema>,
+    hosts: Vec<Host>,
+    start_sec: i64,
+    end_sec: i64,
+}
+
+impl CpuDataGenerator {
+    pub fn new(
+        metadata: RegionMetadataRef,
+        num_hosts: usize,
+        start_sec: i64,
+        end_sec: i64,
+    ) -> Self {
+        let column_schemas = region_metadata_to_row_schema(&metadata);
+        Self {
+            metadata,
+            column_schemas,
+            hosts: Self::generate_hosts(num_hosts),
+            start_sec,
+            end_sec,
+        }
+    }
+
+    pub fn iter(&self) -> impl Iterator<Item = KeyValues> + '_ {
+        // point per 10s.
+        (self.start_sec..self.end_sec)
+            .step_by(10)
+            .enumerate()
+            .map(|(seq, ts)| self.build_key_values(seq, ts))
+    }
+
+    pub fn build_key_values(&self, seq: usize, current_sec: i64) -> KeyValues {
+        let rows = self
+            .hosts
+            .iter()
+            .map(|host| {
+                let mut rng = rand::rng();
+                let mut values = Vec::with_capacity(21);
+                values.push(api::v1::Value {
+                    value_data: Some(ValueData::TimestampMillisecondValue(current_sec * 1000)),
+                });
+                host.fill_values(&mut values);
+                for _ in 0..10 {
+                    values.push(api::v1::Value {
+                        value_data: Some(ValueData::F64Value(Self::random_f64(&mut rng))),
+                    });
+                }
+                Row { values }
+            })
+            .collect();
+        let mutation = api::v1::Mutation {
+            op_type: api::v1::OpType::Put as i32,
+            sequence: seq as u64,
+            rows: Some(Rows {
+                schema: self.column_schemas.clone(),
+                rows,
+            }),
+            write_hint: None,
+        };
+
+        KeyValues::new(&self.metadata, mutation).unwrap()
+    }
+
+    pub fn random_host_filter(&self) -> Predicate {
+        let host = self.random_hostname();
+        let expr = Expr::Column(Column::from_name("hostname")).eq(lit(host));
+        Predicate::new(vec![expr])
+    }
+
+    pub fn random_host_filter_exprs(&self) -> Vec<Expr> {
+        let host = self.random_hostname();
+        vec![Expr::Column(Column::from_name("hostname")).eq(lit(host))]
+    }
+
+    pub fn random_hostname(&self) -> String {
+        let mut rng = rand::rng();
+        self.hosts.choose(&mut rng).unwrap().hostname.clone()
+    }
+
+    pub fn random_f64(rng: &mut ThreadRng) -> f64 {
+        let base: u32 = rng.random_range(30..95);
+        base as f64
+    }
+
+    pub fn generate_hosts(num_hosts: usize) -> Vec<Host> {
+        (0..num_hosts).map(Host::random_with_id).collect()
+    }
+}
+
+/// Creates a metadata for TSBS cpu-like table.
+pub fn cpu_metadata() -> RegionMetadata {
+    let mut builder = RegionMetadataBuilder::new(RegionId::new(1, 1));
+    builder.push_column_metadata(ColumnMetadata {
+        column_schema: ColumnSchema::new(
+            "ts",
+            ConcreteDataType::timestamp_millisecond_datatype(),
+            false,
+        ),
+        semantic_type: SemanticType::Timestamp,
+        column_id: 0,
+    });
+    let mut column_id = 1;
+    let tags = [
+        "hostname",
+        "region",
+        "datacenter",
+        "rack",
+        "os",
+        "arch",
+        "team",
+        "service",
+        "service_version",
+        "service_environment",
+    ];
+    for tag in tags {
+        builder.push_column_metadata(ColumnMetadata {
+            column_schema: ColumnSchema::new(tag, ConcreteDataType::string_datatype(), true),
+            semantic_type: SemanticType::Tag,
+            column_id,
+        });
+        column_id += 1;
+    }
+    let fields = [
+        "usage_user",
+        "usage_system",
+        "usage_idle",
+        "usage_nice",
+        "usage_iowait",
+        "usage_irq",
+        "usage_softirq",
+        "usage_steal",
+        "usage_guest",
+        "usage_guest_nice",
+    ];
+    for field in fields {
+        builder.push_column_metadata(ColumnMetadata {
+            column_schema: ColumnSchema::new(field, ConcreteDataType::float64_datatype(), true),
+            semantic_type: SemanticType::Field,
+            column_id,
+        });
+        column_id += 1;
+    }
+    builder.primary_key(vec![1, 2, 3, 4, 5, 6, 7, 8, 9, 10]);
+    builder.build().unwrap()
+}
diff --git a/src/mito2/src/test_util/memtable_util.rs b/src/mito2/src/test_util/memtable_util.rs
index 7ddac4ee0d..25ab9bb8b4 100644
--- a/src/mito2/src/test_util/memtable_util.rs
+++ b/src/mito2/src/test_util/memtable_util.rs
@@ -30,8 +30,7 @@ use mito_codec::row_converter::{DensePrimaryKeyCodec, PrimaryKeyCodecExt, SortFi
 use store_api::metadata::{
     ColumnMetadata, RegionMetadata, RegionMetadataBuilder, RegionMetadataRef,
 };
-use store_api::storage::{ColumnId, RegionId, SequenceNumber, SequenceRange};
-use table::predicate::Predicate;
+use store_api::storage::{ColumnId, RegionId, SequenceNumber};
 
 use crate::error::Result;
 use crate::memtable::bulk::part::BulkPart;
@@ -83,16 +82,6 @@ impl Memtable for EmptyMemtable {
         Ok(())
     }
 
-    #[cfg(any(test, feature = "test"))]
-    fn iter(
-        &self,
-        _projection: Option<&[ColumnId]>,
-        _filters: Option<Predicate>,
-        _sequence: Option<SequenceRange>,
-    ) -> Result<BoxedBatchIterator> {
-        Ok(Box::new(std::iter::empty()))
-    }
-
     fn ranges(
         &self,
         _projection: Option<&[ColumnId]>,
diff --git a/src/mito2/src/test_util/sst_util.rs b/src/mito2/src/test_util/sst_util.rs
index 389d9bf107..3f635c6e94 100644
--- a/src/mito2/src/test_util/sst_util.rs
+++ b/src/mito2/src/test_util/sst_util.rs
@@ -17,8 +17,13 @@
 use std::sync::Arc;
 
 use api::v1::{OpType, SemanticType};
+use arrow_schema::Schema;
 use common_time::Timestamp;
-use datatypes::arrow::array::{BinaryArray, TimestampMillisecondArray, UInt8Array, UInt64Array};
+use datatypes::arrow::array::{
+    ArrayRef, BinaryDictionaryBuilder, RecordBatch, StringDictionaryBuilder,
+    TimestampMillisecondArray, UInt8Array, UInt64Array,
+};
+use datatypes::arrow::datatypes::UInt32Type;
 use datatypes::prelude::ConcreteDataType;
 use datatypes::schema::{ColumnSchema, SkippingIndexOptions};
 use datatypes::value::ValueRef;
@@ -32,9 +37,10 @@ use store_api::metric_engine_consts::{
 use store_api::storage::consts::ReservedColumnId;
 use store_api::storage::{FileId, RegionId};
 
-use crate::read::{Batch, BatchBuilder, Source};
+use crate::read::{Batch, FlatSource};
 use crate::sst::file::{FileHandle, FileMeta};
-use crate::test_util::{VecBatchReader, new_batch_builder, new_noop_file_purger};
+use crate::sst::{FlatSchemaOptions, to_flat_sst_arrow_schema};
+use crate::test_util::{new_batch_builder, new_noop_file_purger};
 
 /// Test region id.
 const REGION_ID: RegionId = RegionId::new(0, 0);
@@ -185,12 +191,6 @@ pub fn new_sparse_primary_key(
     buffer
 }
 
-/// Creates a [Source] from `batches`.
-pub fn new_source(batches: &[Batch]) -> Source {
-    let reader = VecBatchReader::new(batches);
-    Source::Reader(Box::new(reader))
-}
-
 /// Creates a SST file handle with provided file id
 pub fn sst_file_handle_with_file_id(file_id: FileId, start_ms: i64, end_ms: i64) -> FileHandle {
     let file_purger = new_noop_file_purger();
@@ -214,6 +214,7 @@ pub fn sst_file_handle_with_file_id(file_id: FileId, start_ms: i64, end_ms: i64)
             num_series: 0,
             sequence: None,
             partition_expr: None,
+            ..Default::default()
         },
         file_purger,
     )
@@ -246,34 +247,74 @@ pub fn new_batch_by_range(tags: &[&str], start: usize, end: usize) -> Batch {
     new_batch_with_custom_sequence(tags, start, end, 1000)
 }
 
-pub fn new_batch_with_binary(tags: &[&str], start: usize, end: usize) -> Batch {
+/// Creates a flat format RecordBatch for testing.
+/// Similar to `new_batch_by_range` but returns a RecordBatch in flat format.
+pub fn new_record_batch_by_range(tags: &[&str], start: usize, end: usize) -> RecordBatch {
+    new_record_batch_with_custom_sequence(tags, start, end, 1000)
+}
+
+/// Creates a flat format RecordBatch for testing with a custom sequence.
+pub fn new_record_batch_with_custom_sequence(
+    tags: &[&str],
+    start: usize,
+    end: usize,
+    sequence: u64,
+) -> RecordBatch {
     assert!(end >= start);
+    let metadata = Arc::new(sst_region_metadata());
+    let flat_schema = to_flat_sst_arrow_schema(&metadata, &FlatSchemaOptions::default());
+
+    let num_rows = end - start;
+    let mut columns = Vec::new();
+
+    // Add primary key columns (tag_0, tag_1) as dictionary arrays
+    let mut tag_0_builder = StringDictionaryBuilder::<UInt32Type>::new();
+    let mut tag_1_builder = StringDictionaryBuilder::<UInt32Type>::new();
+
+    for _ in 0..num_rows {
+        tag_0_builder.append_value(tags[0]);
+        tag_1_builder.append_value(tags[1]);
+    }
+
+    columns.push(Arc::new(tag_0_builder.finish()) as ArrayRef);
+    columns.push(Arc::new(tag_1_builder.finish()) as ArrayRef);
+
+    // Add field column (field_0)
+    let field_values: Vec<u64> = (start..end).map(|v| v as u64).collect();
+    columns.push(Arc::new(UInt64Array::from(field_values)));
+
+    // Add time index column (ts)
+    let timestamps: Vec<i64> = (start..end).map(|v| v as i64).collect();
+    columns.push(Arc::new(TimestampMillisecondArray::from(timestamps)));
+
+    // Add encoded primary key column
     let pk = new_primary_key(tags);
-    let timestamps: Vec<_> = (start..end).map(|v| v as i64).collect();
-    let sequences = vec![1000; end - start];
-    let op_types = vec![OpType::Put; end - start];
+    let mut pk_builder = BinaryDictionaryBuilder::<UInt32Type>::new();
+    for _ in 0..num_rows {
+        pk_builder.append(&pk).unwrap();
+    }
+    columns.push(Arc::new(pk_builder.finish()));
 
-    let field: Vec<_> = (start..end)
-        .map(|_v| "some data".as_bytes().to_vec())
-        .collect();
+    // Add sequence column
+    columns.push(Arc::new(UInt64Array::from_value(sequence, num_rows)));
 
-    let mut builder = BatchBuilder::new(pk);
-    builder
-        .timestamps_array(Arc::new(TimestampMillisecondArray::from_iter_values(
-            timestamps.iter().copied(),
-        )))
-        .unwrap()
-        .sequences_array(Arc::new(UInt64Array::from_iter_values(
-            sequences.iter().copied(),
-        )))
-        .unwrap()
-        .op_types_array(Arc::new(UInt8Array::from_iter_values(
-            op_types.iter().map(|v| *v as u8),
-        )))
-        .unwrap()
-        .push_field_array(1, Arc::new(BinaryArray::from_iter_values(field)))
-        .unwrap();
-    builder.build().unwrap()
+    // Add op_type column
+    columns.push(Arc::new(UInt8Array::from_value(
+        OpType::Put as u8,
+        num_rows,
+    )));
+    RecordBatch::try_new(flat_schema, columns).unwrap()
+}
+
+/// Creates a FlatSource from flat format RecordBatches.
+pub(crate) fn new_flat_source_from_record_batches(batches: Vec<RecordBatch>) -> FlatSource {
+    FlatSource::new_iter(
+        batches
+            .first()
+            .map(|x| x.schema())
+            .unwrap_or_else(|| Arc::new(Schema::empty())),
+        Box::new(batches.into_iter().map(Ok)),
+    )
 }
 
 /// Creates a new region metadata for testing SSTs with binary datatype.
diff --git a/src/mito2/src/test_util/version_util.rs b/src/mito2/src/test_util/version_util.rs
index 7a407ece4f..b14eea0958 100644
--- a/src/mito2/src/test_util/version_util.rs
+++ b/src/mito2/src/test_util/version_util.rs
@@ -115,6 +115,7 @@ impl VersionControlBuilder {
                         .expect("partition expression should be valid JSON"),
                     None => None,
                 },
+                ..Default::default()
             },
         );
         self
@@ -207,6 +208,7 @@ pub(crate) fn apply_edit(
                         .expect("partition expression should be valid JSON"),
                     None => None,
                 },
+                ..Default::default()
             }
         })
         .collect();
diff --git a/src/mito2/src/worker.rs b/src/mito2/src/worker.rs
index 71896b3d5d..51b3594b95 100644
--- a/src/mito2/src/worker.rs
+++ b/src/mito2/src/worker.rs
@@ -73,8 +73,8 @@ use crate::region::{
     RegionMapRef,
 };
 use crate::request::{
-    BackgroundNotify, DdlRequest, SenderBulkRequest, SenderDdlRequest, SenderWriteRequest,
-    WorkerRequest, WorkerRequestWithTime,
+    BackgroundNotify, BulkInsertRequest, DdlRequest, SenderBulkRequest, SenderDdlRequest,
+    SenderWriteRequest, WorkerRequest, WorkerRequestWithTime,
 };
 use crate::schedule::scheduler::{LocalScheduler, SchedulerRef};
 use crate::sst::file::RegionFileId;
@@ -207,6 +207,7 @@ impl WorkerGroup {
                 .vector_cache_size(config.vector_cache_size.as_bytes())
                 .page_cache_size(config.page_cache_size.as_bytes())
                 .selector_result_cache_size(config.selector_result_cache_size.as_bytes())
+                .range_result_cache_size(config.range_result_cache_size.as_bytes())
                 .index_metadata_size(config.index.metadata_cache_size.as_bytes())
                 .index_content_size(config.index.content_cache_size.as_bytes())
                 .index_content_page_size(config.index.content_cache_page_size.as_bytes())
@@ -421,6 +422,7 @@ impl WorkerGroup {
                 .vector_cache_size(config.vector_cache_size.as_bytes())
                 .page_cache_size(config.page_cache_size.as_bytes())
                 .selector_result_cache_size(config.selector_result_cache_size.as_bytes())
+                .range_result_cache_size(config.range_result_cache_size.as_bytes())
                 .write_cache(write_cache)
                 .build(),
         );
@@ -917,6 +919,7 @@ impl<S: LogStore> RegionWorkerLoop<S> {
             write_req_buffer.clear();
             ddl_req_buffer.clear();
             general_req_buffer.clear();
+            let mut bulk_insert_req_num = 0;
 
             let max_wait_time = self.time_provider.wait_duration(CHECK_REGION_INTERVAL);
             let sleep = tokio::time::sleep(max_wait_time);
@@ -933,6 +936,11 @@ impl<S: LogStore> RegionWorkerLoop<S> {
                             match request_with_time.request {
                                 WorkerRequest::Write(sender_req) => write_req_buffer.push(sender_req),
                                 WorkerRequest::Ddl(sender_req) => ddl_req_buffer.push(sender_req),
+                                WorkerRequest::BulkInserts(bulk_insert) => {
+                                    bulk_insert_req_num += 1;
+                                    self.buffer_bulk_insert_request(bulk_insert, &mut bulk_req_buffer)
+                                        .await;
+                                }
                                 req => general_req_buffer.push(req),
                             }
                         },
@@ -981,6 +989,11 @@ impl<S: LogStore> RegionWorkerLoop<S> {
                         match request_with_time.request {
                             WorkerRequest::Write(sender_req) => write_req_buffer.push(sender_req),
                             WorkerRequest::Ddl(sender_req) => ddl_req_buffer.push(sender_req),
+                            WorkerRequest::BulkInserts(bulk_insert) => {
+                                bulk_insert_req_num += 1;
+                                self.buffer_bulk_insert_request(bulk_insert, &mut bulk_req_buffer)
+                                    .await
+                            }
                             req => general_req_buffer.push(req),
                         }
                     }
@@ -990,7 +1003,10 @@ impl<S: LogStore> RegionWorkerLoop<S> {
             }
 
             self.listener.on_recv_requests(
-                write_req_buffer.len() + ddl_req_buffer.len() + general_req_buffer.len(),
+                write_req_buffer.len()
+                    + ddl_req_buffer.len()
+                    + general_req_buffer.len()
+                    + bulk_insert_req_num,
             );
 
             self.handle_requests(
@@ -1009,9 +1025,34 @@ impl<S: LogStore> RegionWorkerLoop<S> {
         info!("Exit region worker thread {}", self.id);
     }
 
+    async fn buffer_bulk_insert_request(
+        &mut self,
+        bulk_insert: BulkInsertRequest,
+        bulk_requests: &mut Vec<SenderBulkRequest>,
+    ) {
+        let BulkInsertRequest {
+            metadata,
+            request,
+            sender,
+        } = bulk_insert;
+
+        if let Some(region_metadata) = metadata {
+            self.handle_bulk_insert_batch(region_metadata, request, bulk_requests, sender)
+                .await;
+        } else {
+            error!("Cannot find region metadata for {}", request.region_id);
+            sender.send(
+                error::RegionNotFoundSnafu {
+                    region_id: request.region_id,
+                }
+                .fail(),
+            );
+        }
+    }
+
     /// Dispatches and processes requests.
     ///
-    /// `buffer` should be empty.
+    /// `general_requests` should not contain categorized write, ddl, or bulk insert requests.
     async fn handle_requests(
         &mut self,
         write_requests: &mut Vec<SenderWriteRequest>,
@@ -1022,10 +1063,31 @@ impl<S: LogStore> RegionWorkerLoop<S> {
         for worker_req in general_requests.drain(..) {
             match worker_req {
                 WorkerRequest::Write(_) | WorkerRequest::Ddl(_) => {
-                    // These requests are categorized into write_requests and ddl_requests.
+                    // These requests are categorized before dispatching general requests.
                     continue;
                 }
+                WorkerRequest::BulkInserts(_) => unreachable!("bulk inserts are buffered"),
                 WorkerRequest::Background { region_id, notify } => {
+                    if matches!(
+                        &notify,
+                        BackgroundNotify::RegionEdit(edit_result)
+                            if edit_result.update_region_state
+                    ) {
+                        // Region state must be Editing when reach here.
+                        // This call only moves write/bulk write request into stall queue. When region edit result
+                        // is processed inside handle_background_notify and region state is switched back to Writable,
+                        // stalled request will be processed before the next region edit is dequeued from
+                        // RegionEditQueue immediately in handle_region_edit_result. It not only ensured pending writes
+                        // are processed in time, but also prevents them from starvation.
+                        // TODO(hl): maybe we need to merge those queues for pending requests like pending_ddl,
+                        // region edits and stalled request, so we can simplify the coordination between these queues.
+                        self.handle_buffered_region_write_requests(
+                            &region_id,
+                            write_requests,
+                            bulk_requests,
+                        )
+                        .await;
+                    }
                     // For background notify, we handle it directly.
                     self.handle_background_notify(region_id, notify).await;
                 }
@@ -1046,29 +1108,6 @@ impl<S: LogStore> RegionWorkerLoop<S> {
                 WorkerRequest::SyncRegion(req) => {
                     self.handle_region_sync(req).await;
                 }
-                WorkerRequest::BulkInserts {
-                    metadata,
-                    request,
-                    sender,
-                } => {
-                    if let Some(region_metadata) = metadata {
-                        self.handle_bulk_insert_batch(
-                            region_metadata,
-                            request,
-                            bulk_requests,
-                            sender,
-                        )
-                        .await;
-                    } else {
-                        error!("Cannot find region metadata for {}", request.region_id);
-                        sender.send(
-                            error::RegionNotFoundSnafu {
-                                region_id: request.region_id,
-                            }
-                            .fail(),
-                        );
-                    }
-                }
                 WorkerRequest::RemapManifests(req) => {
                     self.handle_remap_manifests_request(req);
                 }
@@ -1194,6 +1233,9 @@ impl<S: LogStore> RegionWorkerLoop<S> {
             BackgroundNotify::CompactionFinished(req) => {
                 self.handle_compaction_finished(region_id, req).await
             }
+            BackgroundNotify::CompactionCancelled(req) => {
+                self.handle_compaction_cancelled(region_id, req).await
+            }
             BackgroundNotify::CompactionFailed(req) => self.handle_compaction_failure(req).await,
             BackgroundNotify::Truncate(req) => self.handle_truncate_result(req).await,
             BackgroundNotify::RegionChange(req) => {
diff --git a/src/mito2/src/worker/handle_alter.rs b/src/mito2/src/worker/handle_alter.rs
index 459aa8dd32..6fa560e90c 100644
--- a/src/mito2/src/worker/handle_alter.rs
+++ b/src/mito2/src/worker/handle_alter.rs
@@ -216,15 +216,6 @@ impl<S: LogStore> RegionWorkerLoop<S> {
                     // If the format is unchanged, we also consider the option is altered.
                     if new_format != current_options.sst_format.unwrap_or_default() {
                         all_options_altered = false;
-
-                        // Validates the format type.
-                        ensure!(
-                            new_format == FormatType::Flat,
-                            store_api::metadata::InvalidRegionRequestSnafu {
-                                region_id: region.region_id,
-                                err: "Only allow changing format type to flat",
-                            }
-                        );
                     }
                 }
                 SetRegionOption::AppendMode(new_append_mode) => {
@@ -274,8 +265,6 @@ fn new_region_options_on_empty_memtable(
             SetRegionOption::Format(format_str) => {
                 // Safety: handle_alter_region_options_fast() has validated this.
                 let new_format = format_str.parse::<FormatType>().unwrap();
-                assert_eq!(FormatType::Flat, new_format);
-
                 current_options.sst_format = Some(new_format);
             }
             SetRegionOption::AppendMode(new_append_mode) => {
diff --git a/src/mito2/src/worker/handle_apply_staging.rs b/src/mito2/src/worker/handle_apply_staging.rs
index e773150356..876d5c3c31 100644
--- a/src/mito2/src/worker/handle_apply_staging.rs
+++ b/src/mito2/src/worker/handle_apply_staging.rs
@@ -75,7 +75,7 @@ impl<S: LogStore> RegionWorkerLoop<S> {
             return;
         }
 
-        let staging_partition_info = region.staging_partition_info.lock().unwrap().clone();
+        let staging_partition_info = region.manifest_ctx.staging_partition_info();
 
         let staging_partition_expr = staging_partition_info
             .as_ref()
diff --git a/src/mito2/src/worker/handle_close.rs b/src/mito2/src/worker/handle_close.rs
index 84efc283de..1abd772f9f 100644
--- a/src/mito2/src/worker/handle_close.rs
+++ b/src/mito2/src/worker/handle_close.rs
@@ -50,9 +50,7 @@ impl<S: LogStore> RegionWorkerLoop<S> {
             info!("Region {} has pending data, waiting for flush", region_id);
             self.handle_flush_request(
                 region_id,
-                RegionFlushRequest {
-                    row_group_size: None,
-                },
+                RegionFlushRequest::default(),
                 Some(FlushReason::Closing),
                 sender,
             );
@@ -71,6 +69,8 @@ impl<S: LogStore> RegionWorkerLoop<S> {
             return;
         };
         region.stop().await;
+        self.fail_region_stalled_requests_as_not_found(&region_id);
+        self.reject_region_edit_queue_as_not_found(region_id);
         // Clean flush status.
         self.flush_scheduler.on_region_closed(region_id);
         // Clean compaction status.
diff --git a/src/mito2/src/worker/handle_compaction.rs b/src/mito2/src/worker/handle_compaction.rs
index 841d10025d..fd021771b1 100644
--- a/src/mito2/src/worker/handle_compaction.rs
+++ b/src/mito2/src/worker/handle_compaction.rs
@@ -23,7 +23,8 @@ use crate::error::RegionNotFoundSnafu;
 use crate::metrics::COMPACTION_REQUEST_COUNT;
 use crate::region::MitoRegionRef;
 use crate::request::{
-    BuildIndexRequest, CompactionFailed, CompactionFinished, OnFailure, OptionOutputTx,
+    BuildIndexRequest, CompactionCancelled, CompactionFailed, CompactionFinished, OnFailure,
+    OptionOutputTx,
 };
 use crate::sst::index::IndexBuildType;
 use crate::worker::RegionWorkerLoop;
@@ -119,6 +120,28 @@ impl<S> RegionWorkerLoop<S> {
         self.handle_ddl_requests(&mut pending_ddls).await;
     }
 
+    pub(crate) async fn handle_compaction_cancelled(
+        &mut self,
+        region_id: RegionId,
+        request: CompactionCancelled,
+    ) where
+        S: LogStore,
+    {
+        request.on_success();
+
+        // Reuse the scheduler's finish path to wake pending DDLs after a cooperative stop.
+        let mut pending_ddls = match self.regions.get_region(region_id) {
+            Some(_) => {
+                self.compaction_scheduler
+                    .on_compaction_cancelled(region_id)
+                    .await
+            }
+            None => Vec::new(),
+        };
+
+        self.handle_ddl_requests(&mut pending_ddls).await;
+    }
+
     /// When compaction fails, we simply log the error.
     pub(crate) async fn handle_compaction_failure(&mut self, req: CompactionFailed) {
         error!(req.err; "Failed to compact region: {}", req.region_id);
diff --git a/src/mito2/src/worker/handle_drop.rs b/src/mito2/src/worker/handle_drop.rs
index 5d7149768c..00a96b0d16 100644
--- a/src/mito2/src/worker/handle_drop.rs
+++ b/src/mito2/src/worker/handle_drop.rs
@@ -290,7 +290,8 @@ pub(crate) async fn remove_region_dir_once(
             .context(OpenDalSnafu)?;
         // then remove the marker with this dir
         object_store
-            .remove_all(region_path)
+            .delete_with(region_path)
+            .recursive(true)
             .await
             .context(OpenDalSnafu)?;
         Ok(true)
diff --git a/src/mito2/src/worker/handle_enter_staging.rs b/src/mito2/src/worker/handle_enter_staging.rs
index 8b75fdd24f..54bca8e4d8 100644
--- a/src/mito2/src/worker/handle_enter_staging.rs
+++ b/src/mito2/src/worker/handle_enter_staging.rs
@@ -19,6 +19,7 @@ use store_api::logstore::LogStore;
 use store_api::region_request::{EnterStagingRequest, StagingPartitionDirective};
 use store_api::storage::RegionId;
 
+use crate::compaction::RequestCancelResult;
 use crate::error::{RegionNotFoundSnafu, Result, StagingPartitionExprMismatchSnafu};
 use crate::flush::FlushReason;
 use crate::manifest::action::{RegionMetaAction, RegionMetaActionList, RegionPartitionExprChange};
@@ -42,7 +43,7 @@ impl<S: LogStore> RegionWorkerLoop<S> {
 
         // If the region is already in staging mode, verify the partition directive matches.
         if region.is_staging() {
-            let staging_partition_info = region.staging_partition_info.lock().unwrap().clone();
+            let staging_partition_info = region.manifest_ctx.staging_partition_info();
             // If the partition directive mismatches, return error.
             if staging_partition_info
                 .as_ref()
@@ -98,18 +99,24 @@ impl<S: LogStore> RegionWorkerLoop<S> {
             return;
         }
 
-        if self.compaction_scheduler.is_compacting(region_id) {
-            // Safety: region is compacting, add ddl request to pending queue.
-            self.compaction_scheduler
-                .add_ddl_request_to_pending(SenderDdlRequest {
-                    region_id,
-                    sender,
-                    request: DdlRequest::EnterStaging(EnterStagingRequest {
-                        partition_directive,
-                    }),
-                });
+        match self.compaction_scheduler.request_cancel(region_id) {
+            RequestCancelResult::CancelIssued
+            | RequestCancelResult::AlreadyCancelling
+            | RequestCancelResult::TooLateToCancel => {
+                // Safety: region is compacting or has entered the non-cancellable publish stage,
+                // keep the DDL pending until the current task finishes or acknowledges cancellation.
+                self.compaction_scheduler
+                    .add_ddl_request_to_pending(SenderDdlRequest {
+                        region_id,
+                        sender,
+                        request: DdlRequest::EnterStaging(EnterStagingRequest {
+                            partition_directive,
+                        }),
+                    });
 
-            return;
+                return;
+            }
+            RequestCancelResult::NotRunning => {}
         }
 
         self.handle_enter_staging(region, partition_directive, sender);
@@ -271,7 +278,7 @@ impl<S: LogStore> RegionWorkerLoop<S> {
             .sender
             .send(enter_staging_result.result.map(|_| 0));
         // Handles the stalled requests.
-        self.handle_region_stalled_requests(&enter_staging_result.region_id)
+        self.handle_region_stalled_requests(&enter_staging_result.region_id, true)
             .await;
     }
 
@@ -279,10 +286,8 @@ impl<S: LogStore> RegionWorkerLoop<S> {
         region: &MitoRegionRef,
         partition_directive: StagingPartitionDirective,
     ) {
-        let mut staging_partition_info = region.staging_partition_info.lock().unwrap();
-        debug_assert!(staging_partition_info.is_none());
-        *staging_partition_info = Some(StagingPartitionInfo::from_partition_directive(
-            partition_directive,
-        ));
+        region.manifest_ctx.set_staging_partition_info(
+            StagingPartitionInfo::from_partition_directive(partition_directive),
+        );
     }
 }
diff --git a/src/mito2/src/worker/handle_flush.rs b/src/mito2/src/worker/handle_flush.rs
index a8d88ed846..fce995134d 100644
--- a/src/mito2/src/worker/handle_flush.rs
+++ b/src/mito2/src/worker/handle_flush.rs
@@ -19,7 +19,7 @@ use std::sync::atomic::Ordering;
 
 use common_telemetry::{debug, error, info};
 use store_api::logstore::LogStore;
-use store_api::region_request::RegionFlushRequest;
+use store_api::region_request::{RegionFlushReason, RegionFlushRequest};
 use store_api::storage::RegionId;
 
 use crate::config::{IndexBuildMode, MitoConfig};
@@ -30,6 +30,22 @@ use crate::request::{BuildIndexRequest, FlushFailed, FlushFinished, OnFailure, O
 use crate::sst::index::IndexBuildType;
 use crate::worker::RegionWorkerLoop;
 
+fn resolve_flush_reason(
+    explicit_reason: Option<FlushReason>,
+    request_reason: Option<RegionFlushReason>,
+    is_downgrading: bool,
+) -> FlushReason {
+    explicit_reason
+        .or_else(|| request_reason.map(FlushReason::from))
+        .unwrap_or({
+            if is_downgrading {
+                FlushReason::Downgrading
+            } else {
+                FlushReason::Manual
+            }
+        })
+}
+
 impl<S: LogStore> RegionWorkerLoop<S> {
     /// On region flush job failed.
     pub(crate) async fn handle_flush_failed(&mut self, region_id: RegionId, request: FlushFailed) {
@@ -176,13 +192,7 @@ impl<S: LogStore> RegionWorkerLoop<S> {
         // when handling flush request instead of in `schedule_flush` or `flush_finished`.
         self.update_topic_latest_entry_id(&region);
 
-        let reason = reason.unwrap_or_else(|| {
-            if region.is_downgrading() {
-                FlushReason::Downgrading
-            } else {
-                FlushReason::Manual
-            }
-        });
+        let reason = resolve_flush_reason(reason, request.reason, region.is_downgrading());
         let mut task =
             self.new_flush_task(&region, reason, request.row_group_size, self.config.clone());
         task.push_sender(sender);
@@ -283,9 +293,6 @@ impl<S: LogStore> RegionWorkerLoop<S> {
         let flush_on_close = request.flush_reason == FlushReason::Closing;
         let index_build_file_metas = std::mem::take(&mut request.edit.files_to_add);
 
-        // Notifies waiters and observes the flush timer.
-        request.on_success();
-
         // In async mode, create indexes after flush.
         if self.config.index.build_mode == IndexBuildMode::Async {
             self.handle_rebuild_index(
@@ -304,7 +311,10 @@ impl<S: LogStore> RegionWorkerLoop<S> {
             // no need to handle requests and schedule compactions.
             self.remove_region(region_id).await;
             info!("Region {} closed after flush", region_id);
+            request.on_success();
         } else {
+            // Notifies waiters and observes the flush timer.
+            request.on_success();
             // Handle pending requests for the region.
             if let Some((mut ddl_requests, mut write_requests, mut bulk_writes)) =
                 self.flush_scheduler.on_flush_success(region_id)
@@ -349,3 +359,43 @@ impl<S: LogStore> RegionWorkerLoop<S> {
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_resolve_flush_reason_prefers_explicit_reason() {
+        let reason = resolve_flush_reason(
+            Some(FlushReason::Closing),
+            Some(RegionFlushReason::RemoteWalPrune),
+            true,
+        );
+        assert_eq!(reason, FlushReason::Closing);
+    }
+
+    #[test]
+    fn test_resolve_flush_reason_uses_request_reason() {
+        assert_eq!(
+            resolve_flush_reason(None, Some(RegionFlushReason::RegionMigration), true),
+            FlushReason::RegionMigration
+        );
+        assert_eq!(
+            resolve_flush_reason(None, Some(RegionFlushReason::Repartition), false),
+            FlushReason::Repartition
+        );
+        assert_eq!(
+            resolve_flush_reason(None, Some(RegionFlushReason::RemoteWalPrune), false),
+            FlushReason::RemoteWalPrune
+        );
+    }
+
+    #[test]
+    fn test_resolve_flush_reason_fallback_unchanged() {
+        assert_eq!(
+            resolve_flush_reason(None, None, true),
+            FlushReason::Downgrading
+        );
+        assert_eq!(resolve_flush_reason(None, None, false), FlushReason::Manual);
+    }
+}
diff --git a/src/mito2/src/worker/handle_manifest.rs b/src/mito2/src/worker/handle_manifest.rs
index 1be6586430..40366f9ccd 100644
--- a/src/mito2/src/worker/handle_manifest.rs
+++ b/src/mito2/src/worker/handle_manifest.rs
@@ -49,9 +49,12 @@ use crate::worker::{RegionWorkerLoop, WorkerListener};
 
 pub(crate) type RegionEditQueues = HashMap<RegionId, RegionEditQueue>;
 
-/// A queue for temporary store region edit requests, if the region is in the "Editing" state.
-/// When the current region edit request is completed, the next (if there exists) request in the
-/// queue will be processed.
+/// A queue for region edit requests received while the region is already `Editing`.
+///
+/// Normal writes and bulk inserts that arrive during `Editing` use the stalled-write queue instead.
+/// When an edit completes, those writes are handled before the next queued edit starts, preserving
+/// sequence ordering between direct SST edits and WAL/memtable writes unless global reject
+/// backpressure rejects them first.
 /// Everything is done in the region worker loop.
 pub(crate) struct RegionEditQueue {
     region_id: RegionId,
@@ -84,9 +87,31 @@ impl RegionEditQueue {
     fn dequeue(&mut self) -> Option<RegionEditRequest> {
         self.requests.pop_front()
     }
+
+    fn is_empty(&self) -> bool {
+        self.requests.is_empty()
+    }
+
+    fn reject_all_as_not_found(mut self) {
+        while let Some(request) = self.requests.pop_front() {
+            let _ = request.tx.send(
+                RegionNotFoundSnafu {
+                    region_id: self.region_id,
+                }
+                .fail(),
+            );
+        }
+    }
 }
 
 impl<S: LogStore> RegionWorkerLoop<S> {
+    /// Rejects queued region edit requests as region not found.
+    pub(crate) fn reject_region_edit_queue_as_not_found(&mut self, region_id: RegionId) {
+        if let Some(edit_queue) = self.region_edit_queues.remove(&region_id) {
+            edit_queue.reject_all_as_not_found();
+        }
+    }
+
     /// Handles region change result.
     pub(crate) async fn handle_manifest_region_change_result(
         &mut self,
@@ -134,7 +159,7 @@ impl<S: LogStore> RegionWorkerLoop<S> {
             .await;
         }
         // Handles the stalled requests.
-        self.handle_region_stalled_requests(&change_result.region_id)
+        self.handle_region_stalled_requests(&change_result.region_id, true)
             .await;
     }
 
@@ -213,7 +238,7 @@ impl<S: LogStore> RegionWorkerLoop<S> {
     }
 }
 
-impl<S> RegionWorkerLoop<S> {
+impl<S: LogStore> RegionWorkerLoop<S> {
     /// Handles region edit request.
     pub(crate) fn handle_region_edit(&mut self, request: RegionEditRequest) {
         let region_id = request.region_id;
@@ -299,6 +324,10 @@ impl<S> RegionWorkerLoop<S> {
         let region = match self.regions.get_region(edit_result.region_id) {
             Some(region) => region,
             None => {
+                // Fail writes stalled behind this edit if the region was removed before the
+                // edit-completion notification reached the worker.
+                self.fail_region_stalled_requests_as_not_found(&edit_result.region_id);
+                self.reject_region_edit_queue_as_not_found(edit_result.region_id);
                 let _ = edit_result.sender.send(
                     RegionNotFoundSnafu {
                         region_id: edit_result.region_id,
@@ -338,9 +367,24 @@ impl<S> RegionWorkerLoop<S> {
 
         let _ = edit_result.sender.send(edit_result.result);
 
-        if let Some(edit_queue) = self.region_edit_queues.get_mut(&edit_result.region_id)
-            && let Some(request) = edit_queue.dequeue()
-        {
+        if edit_result.update_region_state {
+            // Writes stalled specifically by this edit are handled before the next queued edit.
+            // Otherwise the next edit could reserve a committed sequence before those writes.
+            self.handle_region_stalled_requests(&edit_result.region_id, false)
+                .await;
+        }
+
+        let next_request =
+            if let Some(edit_queue) = self.region_edit_queues.get_mut(&edit_result.region_id) {
+                let request = edit_queue.dequeue();
+                if edit_queue.is_empty() {
+                    self.region_edit_queues.remove(&edit_result.region_id);
+                }
+                request
+            } else {
+                None
+            };
+        if let Some(request) = next_request {
             self.handle_region_edit(request);
         }
 
diff --git a/src/mito2/src/worker/handle_write.rs b/src/mito2/src/worker/handle_write.rs
index 802073b816..721415411e 100644
--- a/src/mito2/src/worker/handle_write.rs
+++ b/src/mito2/src/worker/handle_write.rs
@@ -26,8 +26,8 @@ use store_api::logstore::LogStore;
 use store_api::storage::RegionId;
 
 use crate::error::{
-    InvalidRequestSnafu, PartitionExprVersionMismatchSnafu, RegionStateSnafu, RejectWriteSnafu,
-    Result,
+    InvalidRequestSnafu, PartitionExprVersionMismatchSnafu, RegionNotFoundSnafu, RegionStateSnafu,
+    RejectWriteSnafu, Result,
 };
 use crate::metrics;
 use crate::metrics::{
@@ -193,15 +193,79 @@ impl<S: LogStore> RegionWorkerLoop<S> {
         reject_write_requests(&mut requests, &mut bulk);
     }
 
+    /// Fails a specific region's stalled requests if the region no longer exists.
+    pub(crate) fn fail_region_stalled_requests_as_not_found(&mut self, region_id: &RegionId) {
+        debug!(
+            "Fails stalled requests for region {} as region not found",
+            region_id
+        );
+        let (requests, bulk) = self.stalled_requests.remove(region_id);
+        self.stalling_count
+            .sub((requests.len() + bulk.len()) as i64);
+
+        for req in requests {
+            req.sender.send(
+                RegionNotFoundSnafu {
+                    region_id: req.request.region_id,
+                }
+                .fail(),
+            );
+        }
+        for req in bulk {
+            req.sender.send(
+                RegionNotFoundSnafu {
+                    region_id: req.region_id,
+                }
+                .fail(),
+            );
+        }
+    }
+
     /// Handles a specific region's stalled requests.
-    pub(crate) async fn handle_region_stalled_requests(&mut self, region_id: &RegionId) {
+    ///
+    /// `allow_stall` should be false for backpressure retry paths to avoid stalling the same
+    /// requests again. It should remain true for non-backpressure retries, such as requests stalled
+    /// by alter, staging, and region editing. Global reject backpressure still applies before the
+    /// stall check.
+    pub(crate) async fn handle_region_stalled_requests(
+        &mut self,
+        region_id: &RegionId,
+        allow_stall: bool,
+    ) {
         debug!("Handles stalled requests for region {}", region_id);
         let (mut requests, mut bulk) = self.stalled_requests.remove(region_id);
         self.stalling_count
             .sub((requests.len() + bulk.len()) as i64);
-        self.handle_write_requests(&mut requests, &mut bulk, true)
+        self.handle_write_requests(&mut requests, &mut bulk, allow_stall)
             .await;
     }
+
+    /// Processes same-batch writes for a region before handling its edit-completion notification.
+    ///
+    /// The worker dispatch loop handles background notifications before the current batch's write
+    /// buffer. Without this step, writes that arrived during edit N could be classified only after
+    /// edit N+1 is started, placing them behind that next edit.
+    pub(crate) async fn handle_buffered_region_write_requests(
+        &mut self,
+        region_id: &RegionId,
+        write_requests: &mut Vec<SenderWriteRequest>,
+        bulk_requests: &mut Vec<SenderBulkRequest>,
+    ) {
+        let mut current_region_write_requests = write_requests
+            .extract_if(.., |r| r.request.region_id == *region_id)
+            .collect::<Vec<_>>();
+
+        let mut current_region_bulk_requests = bulk_requests
+            .extract_if(.., |r| r.region_id == *region_id)
+            .collect::<Vec<_>>();
+
+        self.handle_write_requests(
+            &mut current_region_write_requests,
+            &mut current_region_bulk_requests,
+            true,
+        )
+        .await;
+    }
 }
 
 impl<S> RegionWorkerLoop<S> {
@@ -269,10 +333,14 @@ impl<S> RegionWorkerLoop<S> {
 
                         e.insert(region_ctx);
                     }
-                    RegionRoleState::Leader(RegionLeaderState::Altering) => {
+                    RegionRoleState::Leader(RegionLeaderState::Altering)
+                    | RegionRoleState::Leader(RegionLeaderState::Editing) => {
+                        // Editing is transient: queue the write so edit completion can drain it
+                        // before starting the next queued edit.
                         debug!(
-                            "Region {} is altering, add request to pending writes",
-                            region.region_id
+                            "Region {} is {:?}, add request to pending writes",
+                            region.region_id,
+                            region.state()
                         );
                         self.stalling_count.add(1);
                         WRITE_STALL_TOTAL.inc();
@@ -408,10 +476,14 @@ impl<S> RegionWorkerLoop<S> {
 
                         e.insert(region_ctx);
                     }
-                    RegionRoleState::Leader(RegionLeaderState::Altering) => {
+                    RegionRoleState::Leader(RegionLeaderState::Altering)
+                    | RegionRoleState::Leader(RegionLeaderState::Editing) => {
+                        // Editing is transient: queue the bulk write so edit completion can drain
+                        // it before starting the next queued edit.
                         debug!(
-                            "Region {} is altering, add request to pending writes",
-                            region.region_id
+                            "Region {} is {:?}, add request to pending writes",
+                            region.region_id,
+                            region.state()
                         );
                         self.stalling_count.add(1);
                         WRITE_STALL_TOTAL.inc();
diff --git a/src/object-store/Cargo.toml b/src/object-store/Cargo.toml
index 2ef251d04d..5334d0f30a 100644
--- a/src/object-store/Cargo.toml
+++ b/src/object-store/Cargo.toml
@@ -12,17 +12,19 @@ services-memory = ["opendal/services-memory"]
 testing = ["derive_builder"]
 
 [dependencies]
+async-trait.workspace = true
 bytes.workspace = true
+chrono.workspace = true
 common-base.workspace = true
 common-error.workspace = true
 common-macro.workspace = true
 common-telemetry.workspace = true
+datafusion_object_store.workspace = true
 derive_builder = { workspace = true, optional = true }
 futures.workspace = true
 humantime-serde.workspace = true
 lazy_static.workspace = true
-moka = { workspace = true, features = ["future"] }
-opendal = { version = "0.54", features = [
+opendal = { version = "0.56", features = [
     "layers-tracing",
     "layers-prometheus",
     "services-azblob",
@@ -41,6 +43,10 @@ uuid.workspace = true
 
 [dev-dependencies]
 anyhow = "1.0"
+arrow_object_store.workspace = true
 common-telemetry.workspace = true
 common-test-util.workspace = true
+object_store_opendal.workspace = true
+rand.workspace = true
+tempfile.workspace = true
 tokio.workspace = true
diff --git a/src/object-store/src/compat.rs b/src/object-store/src/compat.rs
new file mode 100644
index 0000000000..3e46355bd1
--- /dev/null
+++ b/src/object-store/src/compat.rs
@@ -0,0 +1,1053 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::HashMap;
+use std::fmt::{self, Debug, Display, Formatter};
+use std::future::IntoFuture;
+use std::io;
+use std::ops::Range;
+use std::sync::Arc;
+
+use async_trait::async_trait;
+use bytes::Bytes;
+use datafusion_object_store::path::Path;
+use datafusion_object_store::{
+    Attribute, Attributes, GetOptions, GetRange, GetResult, GetResultPayload, ListResult,
+    MultipartUpload, ObjectMeta, ObjectStore as ArrowObjectStore, PutMode, PutMultipartOptions,
+    PutOptions, PutPayload, PutResult, UploadPart,
+};
+use futures::stream::BoxStream;
+use futures::{FutureExt, StreamExt, TryStreamExt};
+use opendal::options::CopyOptions;
+use opendal::raw::percent_decode_path;
+use opendal::{Buffer, Operator, OperatorInfo, Writer};
+use tokio::sync::{Mutex, oneshot};
+
+/// OpendalStore implements ObjectStore trait by using opendal.
+///
+/// This allows users to use opendal as an object store without extra cost.
+///
+/// Visit [`opendal::services`] for more information about supported services.
+///
+/// ```no_run
+/// use std::sync::Arc;
+///
+/// use bytes::Bytes;
+/// use object_store::path::Path;
+/// use object_store::ObjectStore;
+/// use object_store_opendal::OpendalStore;
+/// use opendal::services::S3;
+/// use opendal::{Builder, Operator};
+///
+/// #[tokio::main]
+/// async fn main() {
+///    let builder = S3::default()
+///     .access_key_id("my_access_key")
+///     .secret_access_key("my_secret_key")
+///     .endpoint("my_endpoint")
+///     .region("my_region");
+///
+///     // Create a new operator
+///     let operator = Operator::new(builder).unwrap().finish();
+///
+///     // Create a new object store
+///     let object_store = Arc::new(OpendalStore::new(operator));
+///
+///     let path = Path::from("data/nested/test.txt");
+///     let bytes = Bytes::from_static(b"hello, world! I am nested.");
+///
+///     object_store.put(&path, bytes.clone().into()).await.unwrap();
+///
+///     let content = object_store
+///         .get(&path)
+///         .await
+///         .unwrap()
+///         .bytes()
+///         .await
+///         .unwrap();
+///
+///     assert_eq!(content, bytes);
+/// }
+/// ```
+#[derive(Clone)]
+pub struct OpendalStore {
+    info: Arc<OperatorInfo>,
+    inner: Operator,
+}
+
+impl OpendalStore {
+    /// Create OpendalStore by given Operator.
+    pub fn new(op: Operator) -> Self {
+        Self {
+            info: op.info().into(),
+            inner: op,
+        }
+    }
+
+    /// Get the Operator info.
+    pub fn info(&self) -> &OperatorInfo {
+        self.info.as_ref()
+    }
+
+    /// Copy a file from one location to another.
+    async fn copy_request(
+        &self,
+        from: &Path,
+        to: &Path,
+        if_not_exists: bool,
+    ) -> datafusion_object_store::Result<()> {
+        let mut copy_options = CopyOptions::default();
+        if if_not_exists {
+            copy_options.if_not_exists = true;
+        }
+
+        // Perform the copy operation
+        self.inner
+            .copy_options(
+                &percent_decode_path(from.as_ref()),
+                &percent_decode_path(to.as_ref()),
+                copy_options,
+            )
+            .await
+            .map_err(|err| {
+                if if_not_exists && err.kind() == opendal::ErrorKind::AlreadyExists {
+                    datafusion_object_store::Error::AlreadyExists {
+                        path: to.to_string(),
+                        source: Box::new(err),
+                    }
+                } else {
+                    format_object_store_error(err, from.as_ref())
+                }
+            })?;
+
+        Ok(())
+    }
+}
+
+impl Debug for OpendalStore {
+    fn fmt(&self, f: &mut Formatter<'_>) -> fmt::Result {
+        f.debug_struct("OpendalStore")
+            .field("scheme", &self.info.scheme())
+            .field("name", &self.info.name())
+            .field("root", &self.info.root())
+            .field("capability", &self.info.full_capability())
+            .finish()
+    }
+}
+
+impl Display for OpendalStore {
+    fn fmt(&self, f: &mut Formatter<'_>) -> fmt::Result {
+        let info = self.inner.info();
+        write!(
+            f,
+            "Opendal({}, bucket={}, root={})",
+            info.scheme(),
+            info.name(),
+            info.root()
+        )
+    }
+}
+
+impl From<Operator> for OpendalStore {
+    fn from(value: Operator) -> Self {
+        Self::new(value)
+    }
+}
+
+#[async_trait]
+impl ArrowObjectStore for OpendalStore {
+    async fn put_opts(
+        &self,
+        location: &Path,
+        bytes: PutPayload,
+        opts: PutOptions,
+    ) -> datafusion_object_store::Result<PutResult> {
+        let decoded_location = percent_decode_path(location.as_ref());
+        let mut future_write = self
+            .inner
+            .write_with(&decoded_location, Buffer::from_iter(bytes));
+        let opts_mode = opts.mode.clone();
+        match opts.mode {
+            PutMode::Overwrite => {}
+            PutMode::Create => {
+                future_write = future_write.if_not_exists(true);
+            }
+            PutMode::Update(update_version) => {
+                let Some(etag) = update_version.e_tag else {
+                    return Err(datafusion_object_store::Error::NotSupported {
+                        source: Box::new(opendal::Error::new(
+                            opendal::ErrorKind::Unsupported,
+                            "etag is required for conditional put",
+                        )),
+                    });
+                };
+                future_write = future_write.if_match(etag.as_str());
+            }
+        }
+        let rp = future_write.await.map_err(|err| {
+            match format_object_store_error(err, location.as_ref()) {
+                datafusion_object_store::Error::Precondition { path, source }
+                    if opts_mode == PutMode::Create =>
+                {
+                    datafusion_object_store::Error::AlreadyExists { path, source }
+                }
+                e => e,
+            }
+        })?;
+
+        let e_tag = rp.etag().map(|s| s.to_string());
+        let version = rp.version().map(|s| s.to_string());
+
+        Ok(PutResult { e_tag, version })
+    }
+
+    async fn put_multipart(
+        &self,
+        location: &Path,
+    ) -> datafusion_object_store::Result<Box<dyn MultipartUpload>> {
+        let decoded_location = percent_decode_path(location.as_ref());
+        let writer = self
+            .inner
+            .writer_with(&decoded_location)
+            .concurrent(8)
+            .await
+            .map_err(|err| format_object_store_error(err, location.as_ref()))?;
+        let upload = OpendalMultipartUpload::new(writer, location.clone());
+
+        Ok(Box::new(upload))
+    }
+
+    async fn put_multipart_opts(
+        &self,
+        location: &Path,
+        opts: PutMultipartOptions,
+    ) -> datafusion_object_store::Result<Box<dyn MultipartUpload>> {
+        const DEFAULT_CONCURRENT: usize = 8;
+
+        let mut options = opendal::options::WriteOptions {
+            concurrent: DEFAULT_CONCURRENT,
+            ..Default::default()
+        };
+
+        let mut user_metadata = HashMap::new();
+
+        for (key, value) in opts.attributes.iter() {
+            match key {
+                Attribute::CacheControl => {
+                    options.cache_control = Some(value.to_string());
+                }
+                Attribute::ContentDisposition => {
+                    options.content_disposition = Some(value.to_string());
+                }
+                Attribute::ContentEncoding => {
+                    options.content_encoding = Some(value.to_string());
+                }
+                Attribute::ContentLanguage => continue,
+                Attribute::ContentType => {
+                    options.content_type = Some(value.to_string());
+                }
+                Attribute::Metadata(k) => {
+                    user_metadata.insert(k.to_string(), value.to_string());
+                }
+                _ => {}
+            }
+        }
+
+        if !user_metadata.is_empty() {
+            options.user_metadata = Some(user_metadata);
+        }
+
+        let decoded_location = percent_decode_path(location.as_ref());
+        let writer = self
+            .inner
+            .writer_options(&decoded_location, options)
+            .await
+            .map_err(|err| format_object_store_error(err, location.as_ref()))?;
+        let upload = OpendalMultipartUpload::new(writer, location.clone());
+
+        Ok(Box::new(upload))
+    }
+
+    async fn get_opts(
+        &self,
+        location: &Path,
+        options: GetOptions,
+    ) -> datafusion_object_store::Result<GetResult> {
+        let raw_location = percent_decode_path(location.as_ref());
+        let meta = {
+            let mut s = self.inner.stat_with(&raw_location);
+            if let Some(version) = &options.version {
+                s = s.version(version.as_str())
+            }
+            if let Some(if_match) = &options.if_match {
+                s = s.if_match(if_match.as_str());
+            }
+            if let Some(if_none_match) = &options.if_none_match {
+                s = s.if_none_match(if_none_match.as_str());
+            }
+            if let Some(if_modified_since) =
+                options.if_modified_since.and_then(datetime_to_timestamp)
+            {
+                s = s.if_modified_since(if_modified_since);
+            }
+            if let Some(if_unmodified_since) =
+                options.if_unmodified_since.and_then(datetime_to_timestamp)
+            {
+                s = s.if_unmodified_since(if_unmodified_since);
+            }
+            s.await
+                .map_err(|err| format_object_store_error(err, location.as_ref()))?
+        };
+
+        let mut attributes = Attributes::new();
+        if let Some(user_meta) = meta.user_metadata() {
+            for (key, value) in user_meta {
+                attributes.insert(
+                    Attribute::Metadata(key.clone().into()),
+                    value.clone().into(),
+                );
+            }
+        }
+
+        let meta = ObjectMeta {
+            location: location.clone(),
+            last_modified: meta
+                .last_modified()
+                .and_then(timestamp_to_datetime)
+                .unwrap_or_default(),
+            size: meta.content_length(),
+            e_tag: meta.etag().map(|x| x.to_string()),
+            version: meta.version().map(|x| x.to_string()),
+        };
+
+        if options.head {
+            return Ok(GetResult {
+                payload: GetResultPayload::Stream(Box::pin(futures::stream::empty())),
+                range: 0..0,
+                meta,
+                attributes,
+            });
+        }
+
+        let reader = {
+            let mut r = self.inner.reader_with(raw_location.as_ref());
+            if let Some(version) = options.version {
+                r = r.version(version.as_str());
+            }
+            if let Some(if_match) = options.if_match {
+                r = r.if_match(if_match.as_str());
+            }
+            if let Some(if_none_match) = options.if_none_match {
+                r = r.if_none_match(if_none_match.as_str());
+            }
+            if let Some(if_modified_since) =
+                options.if_modified_since.and_then(datetime_to_timestamp)
+            {
+                r = r.if_modified_since(if_modified_since);
+            }
+            if let Some(if_unmodified_since) =
+                options.if_unmodified_since.and_then(datetime_to_timestamp)
+            {
+                r = r.if_unmodified_since(if_unmodified_since);
+            }
+            r.await
+                .map_err(|err| format_object_store_error(err, location.as_ref()))?
+        };
+
+        let read_range = match options.range {
+            Some(GetRange::Bounded(r)) => {
+                if r.start >= r.end || r.start >= meta.size {
+                    0..0
+                } else {
+                    let end = r.end.min(meta.size);
+                    r.start..end
+                }
+            }
+            Some(GetRange::Offset(r)) => {
+                if r < meta.size {
+                    r..meta.size
+                } else {
+                    0..0
+                }
+            }
+            Some(GetRange::Suffix(r)) if r < meta.size => (meta.size - r)..meta.size,
+            _ => 0..meta.size,
+        };
+
+        let stream = reader
+            .into_bytes_stream(read_range.start..read_range.end)
+            .await
+            .map_err(|err| format_object_store_error(err, location.as_ref()))?
+            .map_ok(|buf| buf)
+            .map_err(|err: io::Error| datafusion_object_store::Error::Generic {
+                store: "IoError",
+                source: Box::new(err),
+            });
+
+        Ok(GetResult {
+            payload: GetResultPayload::Stream(Box::pin(stream)),
+            range: read_range.start..read_range.end,
+            meta,
+            attributes,
+        })
+    }
+
+    async fn get_range(
+        &self,
+        location: &Path,
+        range: Range<u64>,
+    ) -> datafusion_object_store::Result<Bytes> {
+        let raw_location = percent_decode_path(location.as_ref());
+        let reader = self
+            .inner
+            .reader_with(&raw_location)
+            .await
+            .map_err(|err| format_object_store_error(err, location.as_ref()))?;
+
+        reader
+            .read(range.start..range.end)
+            .await
+            .map(|buf| buf.to_bytes())
+            .map_err(|err| format_object_store_error(err, location.as_ref()))
+    }
+
+    async fn delete(&self, location: &Path) -> datafusion_object_store::Result<()> {
+        let decoded_location = percent_decode_path(location.as_ref());
+        self.inner
+            .delete(&decoded_location)
+            .await
+            .map_err(|err| format_object_store_error(err, location.as_ref()))?;
+
+        Ok(())
+    }
+
+    fn list(
+        &self,
+        prefix: Option<&Path>,
+    ) -> BoxStream<'static, datafusion_object_store::Result<ObjectMeta>> {
+        // object_store `Path` always removes trailing slash
+        // need to add it back
+        let path = prefix.map_or("".into(), |x| {
+            format!("{}/", percent_decode_path(x.as_ref()))
+        });
+
+        let this = self.clone();
+        let fut = async move {
+            let stream = this
+                .inner
+                .lister_with(&path)
+                .recursive(true)
+                .await
+                .map_err(|err| format_object_store_error(err, &path))?;
+
+            let stream = stream.then(|res| async {
+                let entry = res.map_err(|err| format_object_store_error(err, ""))?;
+                let meta = entry.metadata();
+
+                Ok(format_object_meta(entry.path(), meta))
+            });
+            Ok::<_, datafusion_object_store::Error>(stream)
+        };
+
+        fut.into_stream().try_flatten().boxed()
+    }
+
+    fn list_with_offset(
+        &self,
+        prefix: Option<&Path>,
+        offset: &Path,
+    ) -> BoxStream<'static, datafusion_object_store::Result<ObjectMeta>> {
+        let path = prefix.map_or("".into(), |x| {
+            format!("{}/", percent_decode_path(x.as_ref()))
+        });
+        let offset = offset.clone();
+
+        // clone self for 'static lifetime
+        // clone self is cheap
+        let this = self.clone();
+
+        let fut = async move {
+            let list_with_start_after = this.inner.info().full_capability().list_with_start_after;
+            let mut fut = this.inner.lister_with(&path).recursive(true);
+
+            // Use native start_after support if possible.
+            if list_with_start_after {
+                fut = fut.start_after(offset.as_ref());
+            }
+
+            let lister = fut
+                .await
+                .map_err(|err| format_object_store_error(err, &path))?
+                .then(move |entry| {
+                    let path = path.clone();
+                    let this = this.clone();
+                    async move {
+                        let entry = entry.map_err(|err| format_object_store_error(err, &path))?;
+                        let (path, metadata) = entry.into_parts();
+
+                        // If it's a dir or last_modified is present, we can use it directly.
+                        if metadata.is_dir() || metadata.last_modified().is_some() {
+                            let object_meta = format_object_meta(&path, &metadata);
+                            return Ok(object_meta);
+                        }
+
+                        let metadata = this
+                            .inner
+                            .stat(&path)
+                            .await
+                            .map_err(|err| format_object_store_error(err, &path))?;
+                        let object_meta = format_object_meta(&path, &metadata);
+                        Ok::<_, datafusion_object_store::Error>(object_meta)
+                    }
+                })
+                .boxed();
+
+            let stream = if list_with_start_after {
+                lister
+            } else {
+                lister
+                    .try_filter(move |entry| futures::future::ready(entry.location > offset))
+                    .boxed()
+            };
+
+            Ok::<_, datafusion_object_store::Error>(stream)
+        };
+
+        fut.into_stream().try_flatten().boxed()
+    }
+
+    async fn list_with_delimiter(
+        &self,
+        prefix: Option<&Path>,
+    ) -> datafusion_object_store::Result<ListResult> {
+        let path = prefix.map_or("".into(), |x| {
+            format!("{}/", percent_decode_path(x.as_ref()))
+        });
+        let mut stream = self
+            .inner
+            .lister_with(&path)
+            .into_future()
+            .await
+            .map_err(|err| format_object_store_error(err, &path))?;
+
+        let mut common_prefixes = Vec::new();
+        let mut objects = Vec::new();
+
+        while let Some(res) = stream.next().await {
+            let entry = res.map_err(|err| format_object_store_error(err, ""))?;
+            let meta = entry.metadata();
+
+            if meta.is_dir() {
+                common_prefixes.push(entry.path().into());
+            } else if meta.last_modified().is_some() {
+                objects.push(format_object_meta(entry.path(), meta));
+            } else {
+                let meta = self
+                    .inner
+                    .stat(entry.path())
+                    .await
+                    .map_err(|err| format_object_store_error(err, entry.path()))?;
+                objects.push(format_object_meta(entry.path(), &meta));
+            }
+        }
+
+        Ok(ListResult {
+            common_prefixes,
+            objects,
+        })
+    }
+
+    async fn copy(&self, from: &Path, to: &Path) -> datafusion_object_store::Result<()> {
+        self.copy_request(from, to, false).await
+    }
+
+    async fn rename(&self, from: &Path, to: &Path) -> datafusion_object_store::Result<()> {
+        self.inner
+            .rename(
+                &percent_decode_path(from.as_ref()),
+                &percent_decode_path(to.as_ref()),
+            )
+            .await
+            .map_err(|err| format_object_store_error(err, from.as_ref()))?;
+
+        Ok(())
+    }
+
+    async fn copy_if_not_exists(
+        &self,
+        from: &Path,
+        to: &Path,
+    ) -> datafusion_object_store::Result<()> {
+        self.copy_request(from, to, true).await
+    }
+}
+
+/// `MultipartUpload` implementation based on `Writer` in opendal.
+///
+/// # Notes
+///
+/// OpenDAL writer can handle concurrent internally we don't generate real `UploadPart` like existing
+/// implementation do. Instead, we just write the part and notify the next task to be written.
+///
+/// The lock here doesn't really involve the write process, it's just for the notify mechanism.
+struct OpendalMultipartUpload {
+    writer: Arc<Mutex<Writer>>,
+    location: Path,
+    next_notify: oneshot::Receiver<()>,
+}
+
+impl OpendalMultipartUpload {
+    fn new(writer: Writer, location: Path) -> Self {
+        // an immediately dropped sender for the first part to write without waiting
+        let (_, rx) = oneshot::channel();
+
+        Self {
+            writer: Arc::new(Mutex::new(writer)),
+            location,
+            next_notify: rx,
+        }
+    }
+}
+
+#[async_trait]
+impl MultipartUpload for OpendalMultipartUpload {
+    fn put_part(&mut self, data: PutPayload) -> UploadPart {
+        let writer = self.writer.clone();
+        let location = self.location.clone();
+
+        // Generate next notify which will be notified after the current part is written.
+        let (tx, rx) = oneshot::channel();
+        // Fetch the notify for current part to wait for it to be written.
+        let last_rx = std::mem::replace(&mut self.next_notify, rx);
+
+        async move {
+            // Wait for the previous part to be written
+            let _ = last_rx.await;
+
+            let mut writer = writer.lock().await;
+            let result = writer
+                .write(Buffer::from_iter(data))
+                .await
+                .map_err(|err| format_object_store_error(err, location.as_ref()));
+
+            // Notify the next part to be written
+            drop(tx);
+
+            result
+        }
+        .boxed()
+    }
+
+    async fn complete(&mut self) -> datafusion_object_store::Result<PutResult> {
+        let mut writer = self.writer.lock().await;
+        let metadata = writer
+            .close()
+            .await
+            .map_err(|err| format_object_store_error(err, self.location.as_ref()))?;
+
+        let e_tag = metadata.etag().map(|s| s.to_string());
+        let version = metadata.version().map(|s| s.to_string());
+
+        Ok(PutResult { e_tag, version })
+    }
+
+    async fn abort(&mut self) -> datafusion_object_store::Result<()> {
+        let mut writer = self.writer.lock().await;
+        writer
+            .abort()
+            .await
+            .map_err(|err| format_object_store_error(err, self.location.as_ref()))
+    }
+}
+
+impl Debug for OpendalMultipartUpload {
+    fn fmt(&self, f: &mut Formatter<'_>) -> fmt::Result {
+        f.debug_struct("OpendalMultipartUpload")
+            .field("location", &self.location)
+            .finish()
+    }
+}
+
+fn format_object_store_error(err: opendal::Error, path: &str) -> datafusion_object_store::Error {
+    match err.kind() {
+        opendal::ErrorKind::NotFound => datafusion_object_store::Error::NotFound {
+            path: path.to_string(),
+            source: Box::new(err),
+        },
+        opendal::ErrorKind::Unsupported => datafusion_object_store::Error::NotSupported {
+            source: Box::new(err),
+        },
+        opendal::ErrorKind::AlreadyExists => datafusion_object_store::Error::AlreadyExists {
+            path: path.to_string(),
+            source: Box::new(err),
+        },
+        opendal::ErrorKind::ConditionNotMatch => datafusion_object_store::Error::Precondition {
+            path: path.to_string(),
+            source: Box::new(err),
+        },
+        kind => datafusion_object_store::Error::Generic {
+            store: kind.into_static(),
+            source: Box::new(err),
+        },
+    }
+}
+
+fn format_object_meta(path: &str, meta: &opendal::Metadata) -> ObjectMeta {
+    ObjectMeta {
+        location: path.into(),
+        last_modified: meta
+            .last_modified()
+            .and_then(timestamp_to_datetime)
+            .unwrap_or_default(),
+        size: meta.content_length(),
+        e_tag: meta.etag().map(|x| x.to_string()),
+        version: meta.version().map(|x| x.to_string()),
+    }
+}
+
+fn timestamp_to_datetime(ts: opendal::raw::Timestamp) -> Option<chrono::DateTime<chrono::Utc>> {
+    let ts = ts.into_inner();
+    chrono::DateTime::<chrono::Utc>::from_timestamp(ts.as_second(), ts.subsec_nanosecond() as u32)
+}
+
+fn datetime_to_timestamp(dt: chrono::DateTime<chrono::Utc>) -> Option<opendal::raw::Timestamp> {
+    opendal::raw::Timestamp::new(dt.timestamp(), dt.timestamp_subsec_nanos() as i32).ok()
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use bytes::Bytes;
+    use datafusion_object_store::path::Path;
+    use datafusion_object_store::{ObjectStore as ArrowObjectStore, WriteMultipart};
+    use opendal::{Operator, services};
+    use rand::{Rng, RngCore};
+
+    use super::*;
+
+    async fn create_test_object_store() -> Arc<dyn ArrowObjectStore> {
+        let op = Operator::new(services::Memory::default()).unwrap().finish();
+        let object_store = Arc::new(OpendalStore::new(op));
+
+        let path: Path = "data/test.txt".into();
+        let bytes = Bytes::from_static(b"hello, world!");
+        object_store.put(&path, bytes.into()).await.unwrap();
+
+        let path: Path = "data/nested/test.txt".into();
+        let bytes = Bytes::from_static(b"hello, world! I am nested.");
+        object_store.put(&path, bytes.into()).await.unwrap();
+
+        object_store
+    }
+
+    #[tokio::test]
+    async fn test_basic() {
+        let op = Operator::new(services::Memory::default()).unwrap().finish();
+        let object_store: Arc<dyn ArrowObjectStore> = Arc::new(OpendalStore::new(op));
+
+        // Retrieve a specific file
+        let path: Path = "data/test.txt".into();
+
+        let bytes = Bytes::from_static(b"hello, world!");
+        object_store.put(&path, bytes.clone().into()).await.unwrap();
+
+        let meta = object_store.head(&path).await.unwrap();
+
+        assert_eq!(meta.size, 13);
+
+        assert_eq!(
+            object_store
+                .get(&path)
+                .await
+                .unwrap()
+                .bytes()
+                .await
+                .unwrap(),
+            bytes
+        );
+    }
+
+    #[tokio::test]
+    async fn test_put_multipart() {
+        let op = Operator::new(services::Memory::default()).unwrap().finish();
+        let object_store: Arc<dyn ArrowObjectStore> = Arc::new(OpendalStore::new(op));
+
+        let mut rng = rand::rng();
+
+        // Case complete
+        let path: Path = "data/test_complete.txt".into();
+        let upload = object_store.put_multipart(&path).await.unwrap();
+
+        let mut write = WriteMultipart::new(upload);
+
+        let mut all_bytes = vec![];
+        let round = rng.random_range(1..=1024);
+        for _ in 0..round {
+            let size = rng.random_range(1..=1024);
+            let mut bytes = vec![0; size];
+            rng.fill_bytes(&mut bytes);
+
+            all_bytes.extend_from_slice(&bytes);
+            write.put(bytes.into());
+        }
+
+        let _ = write.finish().await.unwrap();
+
+        let meta = object_store.head(&path).await.unwrap();
+
+        assert_eq!(meta.size, all_bytes.len() as u64);
+
+        assert_eq!(
+            object_store
+                .get(&path)
+                .await
+                .unwrap()
+                .bytes()
+                .await
+                .unwrap(),
+            Bytes::from(all_bytes)
+        );
+
+        // Case abort
+        let path: Path = "data/test_abort.txt".into();
+        let mut upload = object_store.put_multipart(&path).await.unwrap();
+        upload.put_part(vec![1; 1024].into()).await.unwrap();
+        upload.abort().await.unwrap();
+
+        let res = object_store.head(&path).await;
+        let err = res.unwrap_err();
+
+        assert!(matches!(
+            err,
+            datafusion_object_store::Error::NotFound { .. }
+        ))
+    }
+
+    #[tokio::test]
+    async fn test_list() {
+        let object_store = create_test_object_store().await;
+        let path: Path = "data/".into();
+        let results = object_store.list(Some(&path)).collect::<Vec<_>>().await;
+        assert_eq!(results.len(), 2);
+        let mut locations = results
+            .iter()
+            .map(|x| x.as_ref().unwrap().location.as_ref())
+            .collect::<Vec<_>>();
+
+        let expected_files = vec![
+            (
+                "data/nested/test.txt",
+                Bytes::from_static(b"hello, world! I am nested."),
+            ),
+            ("data/test.txt", Bytes::from_static(b"hello, world!")),
+        ];
+
+        let expected_locations = expected_files.iter().map(|x| x.0).collect::<Vec<&str>>();
+
+        locations.sort();
+        assert_eq!(locations, expected_locations);
+
+        for (location, bytes) in expected_files {
+            let path: Path = location.into();
+            assert_eq!(
+                object_store
+                    .get(&path)
+                    .await
+                    .unwrap()
+                    .bytes()
+                    .await
+                    .unwrap(),
+                bytes
+            );
+        }
+    }
+
+    #[tokio::test]
+    async fn test_list_with_delimiter() {
+        let object_store = create_test_object_store().await;
+        let path: Path = "data/".into();
+        let result = object_store.list_with_delimiter(Some(&path)).await.unwrap();
+        assert_eq!(result.objects.len(), 1);
+        assert_eq!(result.common_prefixes.len(), 1);
+        assert_eq!(result.objects[0].location.as_ref(), "data/test.txt");
+        assert_eq!(result.common_prefixes[0].as_ref(), "data/nested");
+    }
+
+    #[tokio::test]
+    async fn test_list_with_offset() {
+        let object_store = create_test_object_store().await;
+        let path: Path = "data/".into();
+        let offset: Path = "data/nested/test.txt".into();
+        let result = object_store
+            .list_with_offset(Some(&path), &offset)
+            .collect::<Vec<_>>()
+            .await;
+        assert_eq!(result.len(), 1);
+        assert_eq!(
+            result[0].as_ref().unwrap().location.as_ref(),
+            "data/test.txt"
+        );
+    }
+
+    mod stat_counter {
+        use std::sync::atomic::{AtomicUsize, Ordering};
+
+        use super::*;
+
+        #[derive(Debug, Clone)]
+        pub struct StatCounterLayer {
+            count: Arc<AtomicUsize>,
+        }
+
+        impl StatCounterLayer {
+            pub fn new(count: Arc<AtomicUsize>) -> Self {
+                Self { count }
+            }
+        }
+
+        impl<A: opendal::raw::Access> opendal::raw::Layer<A> for StatCounterLayer {
+            type LayeredAccess = StatCounterAccessor<A>;
+
+            fn layer(&self, inner: A) -> Self::LayeredAccess {
+                StatCounterAccessor {
+                    inner,
+                    count: self.count.clone(),
+                }
+            }
+        }
+
+        #[derive(Debug, Clone)]
+        pub struct StatCounterAccessor<A> {
+            inner: A,
+            count: Arc<AtomicUsize>,
+        }
+
+        impl<A: opendal::raw::Access> opendal::raw::LayeredAccess for StatCounterAccessor<A> {
+            type Inner = A;
+            type Reader = A::Reader;
+            type Writer = A::Writer;
+            type Lister = A::Lister;
+            type Deleter = A::Deleter;
+
+            fn inner(&self) -> &Self::Inner {
+                &self.inner
+            }
+
+            async fn stat(
+                &self,
+                path: &str,
+                args: opendal::raw::OpStat,
+            ) -> opendal::Result<opendal::raw::RpStat> {
+                self.count.fetch_add(1, Ordering::SeqCst);
+                self.inner.stat(path, args).await
+            }
+
+            async fn read(
+                &self,
+                path: &str,
+                args: opendal::raw::OpRead,
+            ) -> opendal::Result<(opendal::raw::RpRead, Self::Reader)> {
+                self.inner.read(path, args).await
+            }
+
+            async fn write(
+                &self,
+                path: &str,
+                args: opendal::raw::OpWrite,
+            ) -> opendal::Result<(opendal::raw::RpWrite, Self::Writer)> {
+                self.inner.write(path, args).await
+            }
+
+            async fn delete(&self) -> opendal::Result<(opendal::raw::RpDelete, Self::Deleter)> {
+                self.inner.delete().await
+            }
+
+            async fn list(
+                &self,
+                path: &str,
+                args: opendal::raw::OpList,
+            ) -> opendal::Result<(opendal::raw::RpList, Self::Lister)> {
+                self.inner.list(path, args).await
+            }
+
+            async fn copy(
+                &self,
+                from: &str,
+                to: &str,
+                args: opendal::raw::OpCopy,
+            ) -> opendal::Result<opendal::raw::RpCopy> {
+                self.inner.copy(from, to, args).await
+            }
+
+            async fn rename(
+                &self,
+                from: &str,
+                to: &str,
+                args: opendal::raw::OpRename,
+            ) -> opendal::Result<opendal::raw::RpRename> {
+                self.inner.rename(from, to, args).await
+            }
+        }
+    }
+
+    #[tokio::test]
+    async fn test_get_range_no_stat() {
+        use std::sync::atomic::{AtomicUsize, Ordering};
+
+        // Create a stat counter and operator with tracking layer
+        let stat_count = Arc::new(AtomicUsize::new(0));
+        let op = Operator::new(opendal::services::Memory::default())
+            .unwrap()
+            .layer(stat_counter::StatCounterLayer::new(stat_count.clone()))
+            .finish();
+        let store = OpendalStore::new(op);
+
+        // Create a test file
+        let location = "test_get_range.txt".into();
+        let value = Bytes::from_static(b"Hello, world!");
+        store.put(&location, value.clone().into()).await.unwrap();
+
+        // Reset counter after put
+        stat_count.store(0, Ordering::SeqCst);
+
+        // Test 1: get_range should NOT call stat()
+        let ret = store.get_range(&location, 0..5).await.unwrap();
+        assert_eq!(Bytes::from_static(b"Hello"), ret);
+        assert_eq!(
+            stat_count.load(Ordering::SeqCst),
+            0,
+            "get_range should not call stat()"
+        );
+
+        // Reset counter
+        stat_count.store(0, Ordering::SeqCst);
+
+        // Test 2: get_opts SHOULD call stat() to get metadata
+        let opts = datafusion_object_store::GetOptions {
+            range: Some(datafusion_object_store::GetRange::Bounded(0..5)),
+            ..Default::default()
+        };
+        let ret = store.get_opts(&location, opts).await.unwrap();
+        let data = ret.bytes().await.unwrap();
+        assert_eq!(Bytes::from_static(b"Hello"), data);
+        assert!(
+            stat_count.load(Ordering::SeqCst) > 0,
+            "get_opts should call stat() to get metadata"
+        );
+
+        // Cleanup
+        store.delete(&location).await.unwrap();
+    }
+}
diff --git a/src/object-store/src/layers/mock.rs b/src/object-store/src/layers/mock.rs
index e55af3bfe0..3df8aae535 100644
--- a/src/object-store/src/layers/mock.rs
+++ b/src/object-store/src/layers/mock.rs
@@ -131,12 +131,12 @@ pub struct MockDeleter {
 }
 
 impl oio::Delete for MockDeleter {
-    fn delete(&mut self, path: &str, args: OpDelete) -> Result<()> {
-        self.inner.delete(path, args)
+    async fn delete(&mut self, path: &str, args: OpDelete) -> Result<()> {
+        self.inner.delete(path, args).await
     }
 
-    async fn flush(&mut self) -> Result<usize> {
-        self.inner.flush().await
+    async fn close(&mut self) -> Result<()> {
+        self.inner.close().await
     }
 }
 
diff --git a/src/object-store/src/lib.rs b/src/object-store/src/lib.rs
index e7ee6afcdb..3a5f72c5ce 100644
--- a/src/object-store/src/lib.rs
+++ b/src/object-store/src/lib.rs
@@ -15,9 +15,10 @@
 pub use opendal::raw::{Access, HttpClient};
 pub use opendal::{
     Buffer, Builder as ObjectStoreBuilder, Entry, EntryMode, Error, ErrorKind, FuturesAsyncReader,
-    FuturesAsyncWriter, Lister, Operator as ObjectStore, Reader, Result, Scheme, Writer, services,
+    FuturesAsyncWriter, Lister, Operator as ObjectStore, Reader, Result, Writer, services,
 };
 
+pub mod compat;
 pub mod config;
 pub mod error;
 pub mod factory;
diff --git a/src/object-store/src/test_util.rs b/src/object-store/src/test_util.rs
index 3279db7b7c..af68978b0e 100644
--- a/src/object-store/src/test_util.rs
+++ b/src/object-store/src/test_util.rs
@@ -32,7 +32,7 @@ impl TempFolder {
     }
 
     pub async fn remove_all(&self) -> Result<()> {
-        self.store.remove_all(&self.path).await
+        self.store.delete_with(&self.path).recursive(true).await
     }
 }
 
diff --git a/src/object-store/src/util.rs b/src/object-store/src/util.rs
index a402b8237c..f43e5a8d83 100644
--- a/src/object-store/src/util.rs
+++ b/src/object-store/src/util.rs
@@ -131,7 +131,7 @@ pub fn normalize_path(path: &str) -> String {
 pub fn with_instrument_layers(object_store: ObjectStore, path_label: bool) -> ObjectStore {
     object_store
         .layer(LoggingLayer::new(DefaultLoggingInterceptor))
-        .layer(TracingLayer)
+        .layer(TracingLayer::new())
         .layer(crate::layers::build_prometheus_metrics_layer(path_label))
 }
 
diff --git a/src/object-store/tests/object_store_test.rs b/src/object-store/tests/object_store_test.rs
index 5df6ab04bd..4ade4fc383 100644
--- a/src/object-store/tests/object_store_test.rs
+++ b/src/object-store/tests/object_store_test.rs
@@ -13,16 +13,22 @@
 // limitations under the License.
 
 use std::env;
+use std::sync::Arc;
 
 use anyhow::Result;
+use arrow_object_store::path::Path;
+use arrow_object_store::{ObjectStore as ArrowObjectStore, ObjectStoreExt};
+use bytes::Bytes;
 use common_telemetry::info;
-use common_test_util::temp_dir::create_temp_dir;
+use futures::TryStreamExt;
 use object_store::ObjectStore;
 use object_store::services::{Fs, S3};
 use object_store::test_util::TempFolder;
+use object_store_opendal::OpendalStore;
 use opendal::EntryMode;
-use opendal::services::{Azblob, Gcs, Oss};
+use opendal::services::{Azblob, Gcs, Memory, Oss};
 use prometheus::{Encoder, TextEncoder};
+use tempfile::TempDir;
 
 async fn test_object_crud(store: &ObjectStore) -> Result<()> {
     // Create object handler.
@@ -103,6 +109,109 @@ async fn test_object_list(store: &ObjectStore) -> Result<()> {
     Ok(())
 }
 
+async fn test_object_list_start_after(store: &ObjectStore) -> Result<()> {
+    let scheme = store.info().scheme();
+    // `start_after` is a service-level capability. Skip the checks when the
+    // backend (e.g. the local Fs service) doesn't honor it natively — the
+    // bound would be silently ignored and the full listing returned.
+    if !store.info().native_capability().list_with_start_after {
+        info!("Skip test_object_list_start_after: backend {scheme} lacks start_after support");
+        return Ok(());
+    }
+    info!("Run test_object_list_start_after on backend {scheme}");
+
+    let files = [
+        "00000000000000000001.json",
+        "00000000000000000002.json",
+        "00000000000000000003.checkpoint",
+        "00000000000000000003.json",
+        "00000000000000000004.json",
+    ];
+    for name in files {
+        store.write(name, "x").await?;
+    }
+
+    // Bare 20-digit bound: versions 1..=2 are skipped; version-3 deltas and
+    // checkpoint are kept (their `.` suffix sorts after the bound).
+    let lister = store
+        .lister_with("/")
+        .start_after("00000000000000000003")
+        .await?;
+    let mut got: Vec<String> = lister
+        .try_collect::<Vec<_>>()
+        .await?
+        .into_iter()
+        .filter(|e| e.metadata().mode() == EntryMode::FILE)
+        .map(|e| e.name().to_string())
+        .collect();
+    got.sort();
+    let mut expected = vec![
+        "00000000000000000003.checkpoint".to_string(),
+        "00000000000000000003.json".to_string(),
+        "00000000000000000004.json".to_string(),
+    ];
+    expected.sort();
+    assert_eq!(expected, got);
+
+    // A bound that matches an existing name exactly excludes that name.
+    let lister = store
+        .lister_with("/")
+        .start_after("00000000000000000003.json")
+        .await?;
+    let got: Vec<String> = lister
+        .try_collect::<Vec<_>>()
+        .await?
+        .into_iter()
+        .filter(|e| e.metadata().mode() == EntryMode::FILE)
+        .map(|e| e.name().to_string())
+        .collect();
+    assert_eq!(vec!["00000000000000000004.json".to_string()], got);
+
+    for name in files {
+        store.delete(name).await?;
+    }
+
+    // OpenDAL resolves `start_after` against the operator root, not the
+    // `lister_with` path. For a nested prefix like `manifest/`, the bound
+    // must also embed that prefix — passing only the bare 20-digit name is
+    // silently a no-op because the full keys start with `m...` > `0...`.
+    let nested_files = [
+        "manifest/00000000000000000001.json",
+        "manifest/00000000000000000002.json",
+        "manifest/00000000000000000003.checkpoint",
+        "manifest/00000000000000000003.json",
+        "manifest/00000000000000000004.json",
+    ];
+    for name in nested_files {
+        store.write(name, "x").await?;
+    }
+
+    let lister = store
+        .lister_with("manifest/")
+        .start_after("manifest/00000000000000000003")
+        .await?;
+    let mut got: Vec<String> = lister
+        .try_collect::<Vec<_>>()
+        .await?
+        .into_iter()
+        .filter(|e| e.metadata().mode() == EntryMode::FILE)
+        .map(|e| e.name().to_string())
+        .collect();
+    got.sort();
+    let mut expected = vec![
+        "00000000000000000003.checkpoint".to_string(),
+        "00000000000000000003.json".to_string(),
+        "00000000000000000004.json".to_string(),
+    ];
+    expected.sort();
+    assert_eq!(expected, got);
+
+    for name in nested_files {
+        store.delete(name).await?;
+    }
+    Ok(())
+}
+
 fn assert_opendal_metrics() {
     let metric_families = prometheus::gather();
     let mut buffer = Vec::new();
@@ -116,10 +225,39 @@ fn assert_opendal_metrics() {
     );
 }
 
+fn create_temp_dir(prefix: &str) -> Result<TempDir> {
+    Ok(tempfile::Builder::new().prefix(prefix).tempdir()?)
+}
+
+#[tokio::test]
+async fn test_opendal_memory_smoke() -> Result<()> {
+    let op = opendal::Operator::new(Memory::default())?.finish();
+    let store: OpendalStore = OpendalStore::new(op);
+    assert_eq!("memory", store.info().scheme());
+    assert!(format!("{store}").contains("memory"));
+    let store: Arc<dyn ArrowObjectStore> = Arc::new(store);
+    let location = Path::from("smoke/test.txt");
+    store
+        .put(&location, Bytes::from_static(b"hello, memory").into())
+        .await?;
+
+    let content = store.get(&location).await?.bytes().await?;
+    assert_eq!(content, Bytes::from_static(b"hello, memory"));
+
+    let listed = store
+        .list(Some(&Path::from("smoke")))
+        .try_collect::<Vec<_>>()
+        .await?;
+    assert_eq!(listed.len(), 1);
+    assert_eq!(listed[0].location, location);
+
+    Ok(())
+}
+
 #[tokio::test]
 async fn test_fs_backend() -> Result<()> {
-    let data_dir = create_temp_dir("test_fs_backend");
-    let tmp_dir = create_temp_dir("test_fs_backend");
+    let data_dir = create_temp_dir("test_fs_backend")?;
+    let tmp_dir = create_temp_dir("test_fs_backend")?;
     let builder = Fs::default()
         .root(&data_dir.path().to_string_lossy())
         .atomic_write_dir(&tmp_dir.path().to_string_lossy());
@@ -129,6 +267,7 @@ async fn test_fs_backend() -> Result<()> {
 
     test_object_crud(&store).await?;
     test_object_list(&store).await?;
+    test_object_list_start_after(&store).await?;
 
     assert_opendal_metrics();
 
@@ -158,6 +297,7 @@ async fn test_s3_backend() -> Result<()> {
         let guard = TempFolder::new(&store, "/");
         test_object_crud(&store).await?;
         test_object_list(&store).await?;
+        test_object_list_start_after(&store).await?;
         assert_opendal_metrics();
         guard.remove_all().await?;
     }
@@ -187,6 +327,7 @@ async fn test_oss_backend() -> Result<()> {
         let guard = TempFolder::new(&store, "/");
         test_object_crud(&store).await?;
         test_object_list(&store).await?;
+        test_object_list_start_after(&store).await?;
         assert_opendal_metrics();
         guard.remove_all().await?;
     }
@@ -216,6 +357,7 @@ async fn test_azblob_backend() -> Result<()> {
         let guard = TempFolder::new(&store, "/");
         test_object_crud(&store).await?;
         test_object_list(&store).await?;
+        test_object_list_start_after(&store).await?;
         assert_opendal_metrics();
         guard.remove_all().await?;
     }
@@ -244,6 +386,7 @@ async fn test_gcs_backend() -> Result<()> {
         let guard = TempFolder::new(&store, "/");
         test_object_crud(&store).await?;
         test_object_list(&store).await?;
+        test_object_list_start_after(&store).await?;
         assert_opendal_metrics();
         guard.remove_all().await?;
     }
diff --git a/src/operator/Cargo.toml b/src/operator/Cargo.toml
index f3de3a9cd7..ea353f4ced 100644
--- a/src/operator/Cargo.toml
+++ b/src/operator/Cargo.toml
@@ -53,9 +53,8 @@ lazy_static.workspace = true
 meta-client.workspace = true
 meter-core.workspace = true
 meter-macros.workspace = true
-moka.workspace = true
+moka = { workspace = true, features = ["future"] }
 object-store.workspace = true
-object_store_opendal.workspace = true
 partition.workspace = true
 prometheus.workspace = true
 prost.workspace = true
diff --git a/src/operator/src/error.rs b/src/operator/src/error.rs
index a5a3e9b0be..da65cc8c2f 100644
--- a/src/operator/src/error.rs
+++ b/src/operator/src/error.rs
@@ -55,7 +55,7 @@ pub enum Error {
     #[snafu(display("Failed to build admin function args: {msg}"))]
     BuildAdminFunctionArgs { msg: String },
 
-    #[snafu(display("Failed to execute admin function: {msg}, error: {error}"))]
+    #[snafu(display("Failed to execute admin function {msg}"))]
     ExecuteAdminFunction {
         msg: String,
         #[snafu(source)]
diff --git a/src/operator/src/expr_helper.rs b/src/operator/src/expr_helper.rs
index b2ec96a76b..378122030c 100644
--- a/src/operator/src/expr_helper.rs
+++ b/src/operator/src/expr_helper.rs
@@ -1047,10 +1047,22 @@ pub fn to_create_flow_task_expr(
         eval_interval: eval_interval.map(|seconds| api::v1::EvalInterval { seconds }),
         comment: create_flow.comment.unwrap_or_default(),
         sql: create_flow.query.to_string(),
-        flow_options: Default::default(),
+        flow_options: stringify_flow_options(create_flow.flow_options)?,
     })
 }
 
+fn stringify_flow_options(flow_options: OptionMap) -> Result<HashMap<String, String>> {
+    let options_len = flow_options.len();
+    let flow_options = flow_options.into_map();
+    ensure!(
+        flow_options.len() == options_len,
+        InvalidSqlSnafu {
+            err_msg: "flow options only support scalar string-compatible values".to_string(),
+        }
+    );
+    Ok(flow_options)
+}
+
 /// sanitize the flow name, remove possible quotes
 fn sanitize_flow_name(mut flow_name: ObjectName) -> Result<String> {
     ensure!(
@@ -1065,6 +1077,8 @@ fn sanitize_flow_name(mut flow_name: ObjectName) -> Result<String> {
 
 #[cfg(test)]
 mod tests {
+    use std::collections::HashMap;
+
     use api::v1::{SetDatabaseOptions, UnsetDatabaseOptions};
     use datatypes::value::Value;
     use session::context::{QueryContext, QueryContextBuilder};
@@ -1327,6 +1341,75 @@ SELECT max(c1), min(c2) FROM schema_2.table_2;";
             to_dot_sep(expr.source_table_names[0].clone())
         );
         assert_eq!("SELECT max(c1), min(c2) FROM schema_2.table_2", expr.sql);
+        assert!(expr.flow_options.is_empty());
+
+        let sql = r"
+CREATE FLOW task_3
+SINK TO schema_1.table_1
+WITH (defer_on_missing_source = 'true', foo = 'bar')
+AS
+SELECT max(c1), min(c2) FROM schema_2.table_2;";
+        let stmt =
+            ParserContext::create_with_dialect(sql, &GreptimeDbDialect {}, ParseOptions::default())
+                .unwrap()
+                .pop()
+                .unwrap();
+
+        let Statement::CreateFlow(create_flow) = stmt else {
+            unreachable!()
+        };
+        let expr = to_create_flow_task_expr(create_flow, &QueryContext::arc()).unwrap();
+        assert_eq!(
+            expr.flow_options,
+            HashMap::from([
+                ("defer_on_missing_source".to_string(), "true".to_string()),
+                ("foo".to_string(), "bar".to_string()),
+            ])
+        );
+
+        let sql = r"
+CREATE FLOW task_4
+SINK TO schema_1.table_1
+WITH (defer_on_missing_source = true)
+AS
+SELECT max(c1), min(c2) FROM schema_2.table_2;";
+        let stmt =
+            ParserContext::create_with_dialect(sql, &GreptimeDbDialect {}, ParseOptions::default())
+                .unwrap()
+                .pop()
+                .unwrap();
+
+        let Statement::CreateFlow(create_flow) = stmt else {
+            unreachable!()
+        };
+        let expr = to_create_flow_task_expr(create_flow, &QueryContext::arc()).unwrap();
+        assert_eq!(
+            expr.flow_options,
+            HashMap::from([("defer_on_missing_source".to_string(), "true".to_string(),)])
+        );
+
+        let sql = r"
+CREATE FLOW task_5
+SINK TO schema_1.table_1
+WITH (defer_on_missing_source = [true])
+AS
+SELECT max(c1), min(c2) FROM schema_2.table_2;";
+        let stmt =
+            ParserContext::create_with_dialect(sql, &GreptimeDbDialect {}, ParseOptions::default())
+                .unwrap()
+                .pop()
+                .unwrap();
+
+        let Statement::CreateFlow(create_flow) = stmt else {
+            unreachable!()
+        };
+        let res = to_create_flow_task_expr(create_flow, &QueryContext::arc());
+        assert!(res.is_err());
+        assert!(
+            res.unwrap_err()
+                .to_string()
+                .contains("flow options only support scalar string-compatible values")
+        );
 
         let sql = r"
 CREATE FLOW abc.`task_2`
diff --git a/src/operator/src/insert.rs b/src/operator/src/insert.rs
index 2366db7897..81df09e4bc 100644
--- a/src/operator/src/insert.rs
+++ b/src/operator/src/insert.rs
@@ -62,7 +62,7 @@ use table::TableRef;
 use table::metadata::TableInfo;
 use table::requests::{
     AUTO_CREATE_TABLE_KEY, InsertRequest as TableInsertRequest, TABLE_DATA_MODEL,
-    TABLE_DATA_MODEL_TRACE_V1, VALID_TABLE_OPTION_KEYS,
+    TABLE_DATA_MODEL_TRACE_V1, TRACE_TABLE_PARTITIONS_HINT_KEY, VALID_TABLE_OPTION_KEYS,
 };
 use table::table_reference::TableReference;
 
@@ -103,7 +103,7 @@ pub enum AutoCreateTableType {
 }
 
 impl AutoCreateTableType {
-    fn as_str(&self) -> &'static str {
+    pub fn as_str(&self) -> &'static str {
         match self {
             AutoCreateTableType::Logical(_) => "logical",
             AutoCreateTableType::Physical => "physical",
@@ -343,8 +343,7 @@ impl Inserter {
             .convert(request)
             .await?;
 
-        let table_infos =
-            HashMap::from_iter([(table_info.table_id(), table_info.clone())].into_iter());
+        let table_infos = HashMap::from_iter([(table_info.table_id(), table_info.clone())]);
 
         self.do_request(inserts, &table_infos, &ctx).await
     }
@@ -353,15 +352,13 @@ impl Inserter {
         &self,
         insert: &Insert,
         ctx: &QueryContextRef,
-        statement_executor: &StatementExecutor,
     ) -> Result<Output> {
         let (inserts, table_info) =
             StatementToRegion::new(self.catalog_manager.as_ref(), &self.partition_manager, ctx)
-                .convert(insert, ctx, statement_executor)
+                .convert(insert, ctx)
                 .await?;
 
-        let table_infos =
-            HashMap::from_iter([(table_info.table_id(), table_info.clone())].into_iter());
+        let table_infos = HashMap::from_iter([(table_info.table_id(), table_info.clone())]);
 
         self.do_request(inserts, &table_infos, ctx).await
     }
@@ -625,6 +622,23 @@ impl Inserter {
                     .extension(TRACE_TABLE_NAME_SESSION_KEY)
                     .unwrap_or(TRACE_TABLE_NAME);
 
+                let trace_table_partitions = if let Some(trace_table_partitions) =
+                    ctx.extension(TRACE_TABLE_PARTITIONS_HINT_KEY)
+                {
+                    let p = trace_table_partitions.parse::<u32>().map_err(|_| {
+                        InvalidInsertRequestSnafu {
+                            reason: format!(
+                                "Failed to parse trace_table_partitions: {}",
+                                trace_table_partitions
+                            ),
+                        }
+                        .build()
+                    })?;
+                    Some(p)
+                } else {
+                    None
+                };
+
                 // note that auto create table shouldn't be ttl instant table
                 // for it's a very unexpected behavior and should be set by user explicitly
                 for mut create_table in create_tables {
@@ -649,8 +663,18 @@ impl Inserter {
                     } else {
                         // prebuilt partition rules for uuid data: see the function
                         // for more information
-                        let partitions = partition_rule_for_hexstring(TRACE_ID_COLUMN)
+                        let partitions = if matches!(trace_table_partitions, Some(0) | Some(1)) {
+                            // disable partitions
+                            None
+                        } else {
+                            let p = partition_rule_for_hexstring(
+                                TRACE_ID_COLUMN,
+                                trace_table_partitions,
+                            )
                             .context(CreatePartitionRulesSnafu)?;
+                            Some(p)
+                        };
+
                         // add skip index to
                         // - trace_id: when searching by trace id
                         // - parent_span_id: when searching root span
@@ -683,7 +707,7 @@ impl Inserter {
                         let table = self
                             .create_physical_table(
                                 create_table,
-                                Some(partitions),
+                                partitions,
                                 ctx,
                                 statement_executor,
                             )
@@ -1066,7 +1090,15 @@ pub fn fill_table_options_for_create(
             table_options.insert(APPEND_MODE_KEY.to_string(), "true".to_string());
         }
         AutoCreateTableType::LastNonNull => {
-            table_options.insert(MERGE_MODE_KEY.to_string(), "last_non_null".to_string());
+            if ctx
+                .extension(APPEND_MODE_KEY)
+                .is_some_and(|value| value.eq_ignore_ascii_case("true"))
+            {
+                table_options.insert(APPEND_MODE_KEY.to_string(), "true".to_string());
+                table_options.insert(MERGE_MODE_KEY.to_string(), "last_row".to_string());
+            } else {
+                table_options.insert(MERGE_MODE_KEY.to_string(), "last_non_null".to_string());
+            }
         }
         AutoCreateTableType::Trace => {
             table_options.insert(APPEND_MODE_KEY.to_string(), "true".to_string());
@@ -1310,4 +1342,56 @@ mod tests {
         assert_eq!(req_schema[0].column_name, ts_name);
         assert_eq!(req_schema[1].column_name, field_name);
     }
+
+    #[test]
+    fn test_last_non_null_create_options_preserve_default_without_append_mode() {
+        let ctx = Arc::new(QueryContext::with(
+            DEFAULT_CATALOG_NAME,
+            DEFAULT_SCHEMA_NAME,
+        ));
+        let mut table_options = Default::default();
+
+        fill_table_options_for_create(&mut table_options, &AutoCreateTableType::LastNonNull, &ctx);
+
+        assert_eq!(
+            Some("last_non_null"),
+            table_options.get(MERGE_MODE_KEY).map(String::as_str)
+        );
+        assert!(!table_options.contains_key(APPEND_MODE_KEY));
+    }
+
+    #[test]
+    fn test_last_non_null_create_options_preserve_default_with_append_mode_false() {
+        let mut ctx = QueryContext::with(DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME);
+        ctx.set_extension(APPEND_MODE_KEY, "false");
+        let ctx = Arc::new(ctx);
+        let mut table_options = Default::default();
+
+        fill_table_options_for_create(&mut table_options, &AutoCreateTableType::LastNonNull, &ctx);
+
+        assert!(!table_options.contains_key(APPEND_MODE_KEY));
+        assert_eq!(
+            Some("last_non_null"),
+            table_options.get(MERGE_MODE_KEY).map(String::as_str)
+        );
+    }
+
+    #[test]
+    fn test_last_non_null_create_options_use_last_row_with_append_mode_true() {
+        let mut ctx = QueryContext::with(DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME);
+        ctx.set_extension(APPEND_MODE_KEY, "true");
+        let ctx = Arc::new(ctx);
+        let mut table_options = Default::default();
+
+        fill_table_options_for_create(&mut table_options, &AutoCreateTableType::LastNonNull, &ctx);
+
+        assert_eq!(
+            Some("true"),
+            table_options.get(APPEND_MODE_KEY).map(String::as_str)
+        );
+        assert_eq!(
+            Some("last_row"),
+            table_options.get(MERGE_MODE_KEY).map(String::as_str)
+        );
+    }
 }
diff --git a/src/operator/src/lib.rs b/src/operator/src/lib.rs
index e31ad87ca1..5e723faeb5 100644
--- a/src/operator/src/lib.rs
+++ b/src/operator/src/lib.rs
@@ -12,9 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
-#![feature(if_let_guard)]
-
 mod bulk_insert;
 pub mod delete;
 pub mod error;
diff --git a/src/operator/src/req_convert/insert/stmt_to_region.rs b/src/operator/src/req_convert/insert/stmt_to_region.rs
index e2e0969035..39a30c61b1 100644
--- a/src/operator/src/req_convert/insert/stmt_to_region.rs
+++ b/src/operator/src/req_convert/insert/stmt_to_region.rs
@@ -12,23 +12,13 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::cell::LazyCell;
-use std::collections::HashMap;
-
 use api::helper::{ColumnDataTypeWrapper, to_grpc_value};
-use api::v1::alter_table_expr::Kind;
 use api::v1::column_def::options_from_column_schema;
 use api::v1::region::InsertRequests as RegionInsertRequests;
-use api::v1::{
-    AlterTableExpr, ColumnSchema as GrpcColumnSchema, ModifyColumnType, ModifyColumnTypes, Row,
-    Rows,
-};
+use api::v1::{ColumnSchema as GrpcColumnSchema, Row, Rows};
 use catalog::CatalogManager;
-use common_telemetry::info;
 use common_time::Timezone;
-use datatypes::data_type::ConcreteDataType;
 use datatypes::schema::{ColumnSchema, SchemaRef};
-use datatypes::types::JsonType;
 use datatypes::value::Value;
 use partition::manager::PartitionRuleManager;
 use session::context::{QueryContext, QueryContextRef};
@@ -47,7 +37,6 @@ use crate::error::{
 use crate::insert::InstantAndNormalInsertRequests;
 use crate::req_convert::common::partitioner::Partitioner;
 use crate::req_convert::insert::semantic_type;
-use crate::statement::StatementExecutor;
 
 const DEFAULT_PLACEHOLDER_VALUE: &str = "default";
 
@@ -74,11 +63,10 @@ impl<'a> StatementToRegion<'a> {
         &self,
         stmt: &Insert,
         query_ctx: &QueryContextRef,
-        statement_executor: &StatementExecutor,
     ) -> Result<(InstantAndNormalInsertRequests, TableInfoRef)> {
         let name = stmt.table_name().context(ParseSqlSnafu)?;
         let (catalog, schema, table_name) = self.get_full_name(name)?;
-        let mut table = self.get_table(&catalog, &schema, &table_name).await?;
+        let table = self.get_table(&catalog, &schema, &table_name).await?;
         let table_schema = table.schema();
 
         ensure!(
@@ -128,26 +116,9 @@ impl<'a> StatementToRegion<'a> {
                 .collect::<Result<Vec<_>>>()
         }
 
-        let mut insert_columns = find_insert_columns(&table, &column_names)?;
+        let insert_columns = find_insert_columns(&table, &column_names)?;
         let converter = SqlRowConverter::new(&insert_columns, query_ctx);
-
-        // Convert the SQL values to GreptimeDB values, and merge a "largest" JSON types of all
-        // values on the way by `JsonColumnTypeUpdater`.
-        let mut updater = JsonColumnTypeUpdater::new(statement_executor, query_ctx);
-        let value_rows = converter.convert(&mut updater, &sql_rows)?;
-
-        // If the JSON values have a "larger" json type than the one in the table schema, modify
-        // the column's json type first, by executing an "alter table" DDL.
-        if updater
-            .maybe_update_column_type(&catalog, &schema, &table_name, &insert_columns)
-            .await?
-        {
-            // Update with the latest schema, if changed.
-            table = self.get_table(&catalog, &schema, &table_name).await?;
-            insert_columns = find_insert_columns(&table, &column_names)?;
-        }
-
-        // Finally convert GreptimeDB values to GRPC values, ready to do insertion on Datanode.
+        let value_rows = converter.convert(&sql_rows)?;
         for (i, row) in value_rows.into_iter().enumerate() {
             for value in row {
                 let grpc_value = to_grpc_value(value);
@@ -248,11 +219,7 @@ impl<'a, 'b> SqlRowConverter<'a, 'b> {
         }
     }
 
-    fn convert(
-        &self,
-        updater: &mut JsonColumnTypeUpdater<'_, 'a>,
-        sql_rows: &[Vec<SqlValue>],
-    ) -> Result<Vec<Vec<Value>>> {
+    fn convert(&self, sql_rows: &[Vec<SqlValue>]) -> Result<Vec<Vec<Value>>> {
         let timezone = Some(&self.query_context.timezone());
         let auto_string_to_numeric = self.query_context.auto_string_to_numeric();
 
@@ -263,9 +230,6 @@ impl<'a, 'b> SqlRowConverter<'a, 'b> {
             for (insert_column, sql_value) in self.insert_columns.iter().zip(sql_row) {
                 let value =
                     sql_value_to_value(insert_column, sql_value, timezone, auto_string_to_numeric)?;
-
-                updater.merge_types(insert_column, &value)?;
-
                 value_row.push(value);
             }
             value_rows.push(value_row);
@@ -274,108 +238,6 @@ impl<'a, 'b> SqlRowConverter<'a, 'b> {
     }
 }
 
-struct JsonColumnTypeUpdater<'a, 'b> {
-    statement_executor: &'a StatementExecutor,
-    query_context: &'a QueryContextRef,
-    merged_value_types: LazyCell<HashMap<&'b str, JsonType>>,
-}
-
-impl<'a, 'b> JsonColumnTypeUpdater<'a, 'b> {
-    fn new(statement_executor: &'a StatementExecutor, query_context: &'a QueryContextRef) -> Self {
-        Self {
-            statement_executor,
-            query_context,
-            merged_value_types: LazyCell::new(Default::default),
-        }
-    }
-
-    fn merge_types(&mut self, column_schema: &'b ColumnSchema, value: &Value) -> Result<()> {
-        if !matches!(value, Value::Json(_)) {
-            return Ok(());
-        }
-
-        if let ConcreteDataType::Json(value_type) = value.data_type() {
-            let merged_type = self
-                .merged_value_types
-                .entry(&column_schema.name)
-                .or_insert_with(|| value_type.clone());
-
-            if !merged_type.is_include(&value_type) {
-                merged_type.merge(&value_type).map_err(|e| {
-                    InvalidInsertRequestSnafu {
-                        reason: format!(r#"cannot merge "{value_type}" into "{merged_type}": {e}"#),
-                    }
-                    .build()
-                })?;
-            }
-        }
-        Ok(())
-    }
-
-    async fn maybe_update_column_type(
-        self,
-        catalog: &str,
-        schema: &str,
-        table: &str,
-        insert_columns: &[&ColumnSchema],
-    ) -> Result<bool> {
-        let mut has_update = false;
-        for (column_name, merged_type) in self.merged_value_types.iter() {
-            let Some(column_type) = insert_columns
-                .iter()
-                .find_map(|x| (&x.name == column_name).then(|| x.data_type.as_json()))
-                .flatten()
-            else {
-                continue;
-            };
-            if column_type.is_include(merged_type) {
-                continue;
-            }
-
-            let new_column_type = {
-                let mut x = column_type.clone();
-                x.merge(merged_type)
-                    .map_err(|e| {
-                        InvalidInsertRequestSnafu {
-                            reason: format!(
-                                r#"cannot merge "{merged_type}" into "{column_type}": {e}"#
-                            ),
-                        }
-                        .build()
-                    })
-                    .map(|()| x)
-            }?;
-            info!(
-                "updating table {}.{}.{} column {} json type: {} => {}",
-                catalog, schema, table, column_name, column_type, new_column_type,
-            );
-
-            let (target_type, target_type_extension) =
-                ColumnDataTypeWrapper::try_from(ConcreteDataType::Json(new_column_type))
-                    .context(ColumnDataTypeSnafu)?
-                    .into_parts();
-            let alter_expr = AlterTableExpr {
-                catalog_name: catalog.to_string(),
-                schema_name: schema.to_string(),
-                table_name: table.to_string(),
-                kind: Some(Kind::ModifyColumnTypes(ModifyColumnTypes {
-                    modify_column_types: vec![ModifyColumnType {
-                        column_name: column_name.to_string(),
-                        target_type: target_type as i32,
-                        target_type_extension,
-                    }],
-                })),
-            };
-            self.statement_executor
-                .alter_table_inner(alter_expr, self.query_context.clone())
-                .await?;
-
-            has_update = true;
-        }
-        Ok(has_update)
-    }
-}
-
 fn column_names<'a>(stmt: &'a Insert, table_schema: &'a SchemaRef) -> Vec<&'a String> {
     if !stmt.columns().is_empty() {
         stmt.columns()
diff --git a/src/operator/src/statement.rs b/src/operator/src/statement.rs
index b87be42ff4..f3931255d0 100644
--- a/src/operator/src/statement.rs
+++ b/src/operator/src/statement.rs
@@ -933,7 +933,7 @@ impl Inserter for InserterImpl {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
 
     use common_time::range::TimestampRange;
diff --git a/src/operator/src/statement/admin.rs b/src/operator/src/statement/admin.rs
index e606f8a939..d51e739dc2 100644
--- a/src/operator/src/statement/admin.rs
+++ b/src/operator/src/statement/admin.rs
@@ -133,7 +133,7 @@ impl StatementExecutor {
             .invoke_async_with_args(func_args)
             .await
             .with_context(|_| ExecuteAdminFunctionSnafu {
-                msg: format!("Failed to execute admin function {}", fn_name),
+                msg: fn_name.to_string(),
             })?;
 
         // Convert result back to VectorRef
diff --git a/src/operator/src/statement/ddl.rs b/src/operator/src/statement/ddl.rs
index 0a24eb3713..6fd6e4adb4 100644
--- a/src/operator/src/statement/ddl.rs
+++ b/src/operator/src/statement/ddl.rs
@@ -113,6 +113,9 @@ struct DdlSubmitOptions {
     timeout: Duration,
 }
 
+const DEFER_ON_MISSING_SOURCE_KEY: &str = "defer_on_missing_source";
+const ALLOWED_FLOW_OPTIONS: [&str; 1] = [DEFER_ON_MISSING_SOURCE_KEY];
+
 fn build_procedure_id_output(procedure_id: Vec<u8>) -> Result<Output> {
     let procedure_id = String::from_utf8_lossy(&procedure_id).to_string();
     let vector: VectorRef = Arc::new(StringVector::from(vec![procedure_id]));
@@ -152,6 +155,55 @@ fn parse_ddl_options(options: &OptionMap) -> Result<DdlSubmitOptions> {
     Ok(DdlSubmitOptions { wait, timeout })
 }
 
+fn supported_flow_options() -> String {
+    ALLOWED_FLOW_OPTIONS.join(", ")
+}
+
+fn normalize_flow_bool_option(key: &str, value: &str) -> Result<String> {
+    value
+        .trim()
+        .to_ascii_lowercase()
+        .parse::<bool>()
+        .map(|value| value.to_string())
+        .map_err(|_| {
+            InvalidSqlSnafu {
+                err_msg: format!("invalid flow option '{key}': '{value}'"),
+            }
+            .build()
+        })
+}
+
+fn validate_and_normalize_flow_options(
+    options: HashMap<String, String>,
+) -> Result<HashMap<String, String>> {
+    options
+        .into_iter()
+        .map(|(key, value)| {
+            if key == FlowType::FLOW_TYPE_KEY {
+                return InvalidSqlSnafu {
+                    err_msg: format!("flow option '{key}' is reserved for internal use"),
+                }
+                .fail();
+            }
+
+            let normalized_value = match key.as_str() {
+                DEFER_ON_MISSING_SOURCE_KEY => normalize_flow_bool_option(&key, &value)?,
+                _ => {
+                    return InvalidSqlSnafu {
+                        err_msg: format!(
+                            "unknown flow option '{key}', supported options: {}",
+                            supported_flow_options()
+                        ),
+                    }
+                    .fail();
+                }
+            };
+
+            Ok((key, normalized_value))
+        })
+        .collect()
+}
+
 impl StatementExecutor {
     pub fn catalog_manager(&self) -> CatalogManagerRef {
         self.catalog_manager.clone()
@@ -626,20 +678,18 @@ impl StatementExecutor {
 
     async fn create_flow_procedure(
         &self,
-        expr: CreateFlowExpr,
+        mut expr: CreateFlowExpr,
         query_context: QueryContextRef,
     ) -> Result<SubmitDdlTaskResponse> {
+        expr.flow_options = validate_and_normalize_flow_options(expr.flow_options)?;
+
         let flow_type = self
             .determine_flow_type(&expr, query_context.clone())
             .await?;
         info!("determined flow={} type: {:#?}", expr.flow_name, flow_type);
 
-        let expr = {
-            let mut expr = expr;
-            expr.flow_options
-                .insert(FlowType::FLOW_TYPE_KEY.to_string(), flow_type.to_string());
-            expr
-        };
+        expr.flow_options
+            .insert(FlowType::FLOW_TYPE_KEY.to_string(), flow_type.to_string());
 
         let task = CreateFlowTask::try_from(PbCreateFlowTask {
             create_flow: Some(expr),
@@ -2334,6 +2384,94 @@ mod test {
         assert_eq!(Duration::from_secs(300), ddl_options.timeout);
     }
 
+    #[test]
+    fn test_validate_and_normalize_flow_options_empty() {
+        assert!(
+            validate_and_normalize_flow_options(HashMap::new())
+                .unwrap()
+                .is_empty()
+        );
+    }
+
+    #[test]
+    fn test_validate_and_normalize_flow_options_valid() {
+        let options =
+            HashMap::from([(DEFER_ON_MISSING_SOURCE_KEY.to_string(), "TRUE".to_string())]);
+
+        assert_eq!(
+            validate_and_normalize_flow_options(options).unwrap(),
+            HashMap::from([(DEFER_ON_MISSING_SOURCE_KEY.to_string(), "true".to_string(),)])
+        );
+    }
+
+    #[test]
+    fn test_validate_and_normalize_flow_options_unknown_option() {
+        let err = validate_and_normalize_flow_options(HashMap::from([(
+            "foo".to_string(),
+            "bar".to_string(),
+        )]))
+        .unwrap_err();
+
+        assert!(
+            err.to_string()
+                .contains("unknown flow option 'foo', supported options: defer_on_missing_source")
+        );
+    }
+
+    #[test]
+    fn test_validate_and_normalize_flow_options_reserved_option() {
+        let err = validate_and_normalize_flow_options(HashMap::from([(
+            FlowType::FLOW_TYPE_KEY.to_string(),
+            FlowType::BATCHING.to_string(),
+        )]))
+        .unwrap_err();
+
+        assert!(
+            err.to_string()
+                .contains("flow option 'flow_type' is reserved for internal use")
+        );
+    }
+
+    #[test]
+    fn test_validate_and_normalize_flow_options_invalid_bool() {
+        let err = validate_and_normalize_flow_options(HashMap::from([(
+            DEFER_ON_MISSING_SOURCE_KEY.to_string(),
+            "not-a-bool".to_string(),
+        )]))
+        .unwrap_err();
+
+        assert!(
+            err.to_string()
+                .contains("invalid flow option 'defer_on_missing_source': 'not-a-bool'")
+        );
+    }
+
+    #[test]
+    fn test_validate_and_normalize_flow_options_rejects_redacted_invalid_input() {
+        let sql = r"
+CREATE FLOW task_6
+SINK TO schema_1.table_1
+WITH (access_key_id = [true])
+AS
+SELECT max(c1), min(c2) FROM schema_2.table_2;";
+        let stmt =
+            ParserContext::create_with_dialect(sql, &GreptimeDbDialect {}, ParseOptions::default())
+                .unwrap()
+                .pop()
+                .unwrap();
+
+        let Statement::CreateFlow(create_flow) = stmt else {
+            unreachable!()
+        };
+        let expr =
+            expr_helper::to_create_flow_task_expr(create_flow, &QueryContext::arc()).unwrap();
+        let err = validate_and_normalize_flow_options(expr.flow_options).unwrap_err();
+
+        assert!(err.to_string().contains(
+            "unknown flow option 'access_key_id', supported options: defer_on_missing_source"
+        ));
+    }
+
     #[test]
     fn test_name_is_match() {
         assert!(!NAME_PATTERN_REG.is_match("/adaf"));
diff --git a/src/operator/src/statement/dml.rs b/src/operator/src/statement/dml.rs
index 41169398ab..827bfd8b66 100644
--- a/src/operator/src/statement/dml.rs
+++ b/src/operator/src/statement/dml.rs
@@ -28,7 +28,7 @@ impl StatementExecutor {
         if insert.can_extract_values() {
             // Fast path: plain insert ("insert with literal values") is executed directly
             self.inserter
-                .handle_statement_insert(insert.as_ref(), &query_ctx, self)
+                .handle_statement_insert(insert.as_ref(), &query_ctx)
                 .await
         } else {
             // Slow path: insert with subquery. Execute using query engine.
diff --git a/src/operator/src/utils.rs b/src/operator/src/utils.rs
index 93da5f028e..6e9386b3fa 100644
--- a/src/operator/src/utils.rs
+++ b/src/operator/src/utils.rs
@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+use std::sync::{Arc, RwLock};
+
 use common_time::Timezone;
 use session::context::{QueryContextBuilder, QueryContextRef};
 use snafu::ResultExt;
@@ -27,6 +29,8 @@ pub fn to_meta_query_context(
         timezone: query_context.timezone().to_string(),
         extensions: query_context.extensions(),
         channel: query_context.channel() as u8,
+        snapshot_seqs: query_context.snapshots(),
+        sst_min_sequences: query_context.sst_min_sequences(),
     }
 }
 
@@ -43,5 +47,41 @@ pub fn try_to_session_query_context(
         )
         .extensions(value.extensions)
         .channel((value.channel as u32).into())
+        .snapshot_seqs(Arc::new(RwLock::new(value.snapshot_seqs)))
+        .sst_min_sequences(Arc::new(RwLock::new(value.sst_min_sequences)))
         .build())
 }
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+    use std::sync::{Arc, RwLock};
+
+    use common_time::Timezone;
+    use session::context::QueryContextBuilder;
+
+    use super::{to_meta_query_context, try_to_session_query_context};
+
+    #[test]
+    fn test_query_context_meta_roundtrip_with_sequences() {
+        let session_ctx = Arc::new(
+            QueryContextBuilder::default()
+                .current_catalog("c1".to_string())
+                .current_schema("s1".to_string())
+                .timezone(Timezone::from_tz_string("UTC").unwrap())
+                .set_extension("flow.return_region_seq".to_string(), "true".to_string())
+                .snapshot_seqs(Arc::new(RwLock::new(HashMap::from([(10, 100)]))))
+                .sst_min_sequences(Arc::new(RwLock::new(HashMap::from([(10, 90)]))))
+                .build(),
+        );
+
+        let meta_ctx = to_meta_query_context(session_ctx);
+        let roundtrip = try_to_session_query_context(meta_ctx).unwrap();
+
+        assert_eq!(roundtrip.current_catalog(), "c1");
+        assert_eq!(roundtrip.current_schema(), "s1");
+        assert_eq!(roundtrip.snapshots(), HashMap::from([(10, 100)]));
+        assert_eq!(roundtrip.sst_min_sequences(), HashMap::from([(10, 90)]));
+        assert_eq!(roundtrip.extension("flow.return_region_seq"), Some("true"));
+    }
+}
diff --git a/src/partition/Cargo.toml b/src/partition/Cargo.toml
index d498ed8c13..a8e3a8ae11 100644
--- a/src/partition/Cargo.toml
+++ b/src/partition/Cargo.toml
@@ -15,6 +15,7 @@ common-error.workspace = true
 common-macro.workspace = true
 common-meta.workspace = true
 common-query.workspace = true
+common-telemetry.workspace = true
 datafusion-common.workspace = true
 datafusion-expr.workspace = true
 datafusion-physical-expr.workspace = true
diff --git a/src/partition/src/cache.rs b/src/partition/src/cache.rs
index a886e1e08d..4066b69aa3 100644
--- a/src/partition/src/cache.rs
+++ b/src/partition/src/cache.rs
@@ -121,10 +121,12 @@ pub fn new_partition_info_cache(
     CacheContainer::new(
         name,
         cache,
-        Box::new(|cache, ident| {
+        Box::new(|cache, idents| {
             Box::pin(async move {
-                if let CacheIdent::TableId(table_id) = ident {
-                    cache.invalidate(table_id).await
+                for ident in idents {
+                    if let CacheIdent::TableId(table_id) = ident {
+                        cache.invalidate(table_id).await
+                    }
                 }
                 Ok(())
             })
diff --git a/src/partition/src/lib.rs b/src/partition/src/lib.rs
index c9257e8ee5..3bfd7297cc 100644
--- a/src/partition/src/lib.rs
+++ b/src/partition/src/lib.rs
@@ -12,7 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
 //! Structs and traits for partitioning rule.
 
 pub mod cache;
@@ -27,5 +26,6 @@ pub mod partition;
 pub mod simplify;
 pub mod splitter;
 pub mod subtask;
+pub mod utils;
 
 pub use crate::partition::{PartitionRule, PartitionRuleRef};
diff --git a/src/partition/src/multi_dim.rs b/src/partition/src/multi_dim.rs
index 7b1f7aa3dd..a68479888b 100644
--- a/src/partition/src/multi_dim.rs
+++ b/src/partition/src/multi_dim.rs
@@ -338,8 +338,8 @@ impl PartitionRule for MultiDimPartitionRule {
         self
     }
 
-    fn partition_columns(&self) -> Vec<String> {
-        self.partition_columns.clone()
+    fn partition_columns(&self) -> &[String] {
+        &self.partition_columns
     }
 
     fn find_region(&self, values: &[Value]) -> Result<RegionNumber> {
@@ -356,7 +356,7 @@ impl PartitionRule for MultiDimPartitionRule {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use super::*;
     use crate::error::{self, Error};
diff --git a/src/partition/src/partition.rs b/src/partition/src/partition.rs
index 110f61a39e..f4c585e404 100644
--- a/src/partition/src/partition.rs
+++ b/src/partition/src/partition.rs
@@ -29,7 +29,7 @@ pub type PartitionRuleRef = Arc<dyn PartitionRule>;
 pub trait PartitionRule: Sync + Send {
     fn as_any(&self) -> &dyn Any;
 
-    fn partition_columns(&self) -> Vec<String>;
+    fn partition_columns(&self) -> &[String];
 
     /// Finds the target region by the partition values.
     ///
diff --git a/src/partition/src/splitter.rs b/src/partition/src/splitter.rs
index fa19b74ad3..176422a173 100644
--- a/src/partition/src/splitter.rs
+++ b/src/partition/src/splitter.rs
@@ -66,7 +66,7 @@ impl<'a> SplitReadRowHelper<'a> {
             .collect::<HashMap<_, _>>();
         let partition_cols = partition_rule.partition_columns();
         let partition_cols_indexes = partition_cols
-            .into_iter()
+            .iter()
             .map(|col_name| col_name_to_idx.get(&col_name).cloned())
             .collect::<Vec<_>>();
 
@@ -176,15 +176,25 @@ mod tests {
     }
 
     #[derive(Debug, Serialize, Deserialize)]
-    struct MockPartitionRule;
+    struct MockPartitionRule {
+        partition_columns: Vec<String>,
+    }
+
+    impl Default for MockPartitionRule {
+        fn default() -> Self {
+            Self {
+                partition_columns: vec!["id".to_string()],
+            }
+        }
+    }
 
     impl PartitionRule for MockPartitionRule {
         fn as_any(&self) -> &dyn Any {
             self
         }
 
-        fn partition_columns(&self) -> Vec<String> {
-            vec!["id".to_string()]
+        fn partition_columns(&self) -> &[String] {
+            &self.partition_columns
         }
 
         fn find_region(&self, values: &[Value]) -> Result<RegionNumber> {
@@ -206,15 +216,25 @@ mod tests {
     }
 
     #[derive(Debug, Serialize, Deserialize)]
-    struct MockMissedColPartitionRule;
+    struct MockMissedColPartitionRule {
+        partition_columns: Vec<String>,
+    }
+
+    impl Default for MockMissedColPartitionRule {
+        fn default() -> Self {
+            Self {
+                partition_columns: vec!["missed_col".to_string()],
+            }
+        }
+    }
 
     impl PartitionRule for MockMissedColPartitionRule {
         fn as_any(&self) -> &dyn Any {
             self
         }
 
-        fn partition_columns(&self) -> Vec<String> {
-            vec!["missed_col".to_string()]
+        fn partition_columns(&self) -> &[String] {
+            &self.partition_columns
         }
 
         fn find_region(&self, values: &[Value]) -> Result<RegionNumber> {
@@ -243,8 +263,8 @@ mod tests {
             self
         }
 
-        fn partition_columns(&self) -> Vec<String> {
-            vec![]
+        fn partition_columns(&self) -> &[String] {
+            &[]
         }
 
         fn find_region(&self, _values: &[Value]) -> Result<RegionNumber> {
@@ -261,7 +281,7 @@ mod tests {
     #[test]
     fn test_writer_splitter() {
         let rows = mock_rows();
-        let rule = Arc::new(MockPartitionRule) as PartitionRuleRef;
+        let rule = Arc::new(MockPartitionRule::default()) as PartitionRuleRef;
         let splitter = RowSplitter::new(rule);
 
         let mut splits = splitter.split(rows).unwrap();
@@ -276,7 +296,7 @@ mod tests {
     #[test]
     fn test_missed_col_writer_splitter() {
         let rows = mock_rows();
-        let rule = Arc::new(MockMissedColPartitionRule) as PartitionRuleRef;
+        let rule = Arc::new(MockMissedColPartitionRule::default()) as PartitionRuleRef;
 
         let splitter = RowSplitter::new(rule);
         let mut splits = splitter.split(rows).unwrap();
diff --git a/src/partition/src/utils.rs b/src/partition/src/utils.rs
new file mode 100644
index 0000000000..de212b5589
--- /dev/null
+++ b/src/partition/src/utils.rs
@@ -0,0 +1,15 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+pub mod split;
diff --git a/src/partition/src/utils/split.rs b/src/partition/src/utils/split.rs
new file mode 100644
index 0000000000..4b1980e34e
--- /dev/null
+++ b/src/partition/src/utils/split.rs
@@ -0,0 +1,1263 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Expression split utilities for partition rules.
+//!
+//! This module provides a conservative way to split one partition expression `R`
+//! by a split expression `S` into:
+//! - `left = R AND S`
+//! - `right = R AND NOT(S)`
+//!
+//! The implementation intentionally reuses existing partition components
+//! (`Collider`, `simplify`, `PartitionChecker`) and degrades to no-split when an
+//! unsupported shape/type is encountered.
+
+use std::collections::{BTreeMap, HashSet};
+
+use datatypes::value::Value;
+use snafu::ensure;
+
+use crate::collider::Collider;
+use crate::error::{self, Result};
+use crate::expr::{Operand, PartitionExpr, RestrictedOp};
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum ExprSplitDegradeReason {
+    UnsupportedType,
+    UnsupportedNotExpansion,
+    ColliderRejected,
+    EmptyBranch,
+}
+
+/// Splits one partition expression with a split predicate.
+///
+/// Returns `(left, right)` on success, where:
+/// - `left = R AND S`
+/// - `right = R AND NOT(S)`
+///
+/// Supported shape:
+/// - `split_expr` must be a single atomic range predicate (`<`, `<=`, `>`, `>=`).
+/// - `base_expr` must be a pure `AND` tree of atomic range predicates, possibly
+///   across unrelated columns.
+///
+/// Returns [`ExprSplitDegradeReason`] when this cannot safely process the shape/type.
+pub fn split_partition_expr(
+    base_expr: PartitionExpr,
+    split_expr: PartitionExpr,
+) -> std::result::Result<(PartitionExpr, PartitionExpr), ExprSplitDegradeReason> {
+    let base = base_expr.canonicalize();
+    let split = split_expr.canonicalize();
+
+    if validate_supported_expr(&base).is_err() || validate_supported_expr(&split).is_err() {
+        return Err(ExprSplitDegradeReason::UnsupportedType);
+    }
+
+    if !validate_base_expr_shape(&base) || !validate_split_expr_shape(&split) {
+        return Err(ExprSplitDegradeReason::UnsupportedType);
+    }
+
+    let not_split = match negate_split_expr(&split) {
+        Ok(expr) => expr,
+        Err(_) => {
+            return Err(ExprSplitDegradeReason::UnsupportedNotExpansion);
+        }
+    };
+
+    let left_raw = base.clone().and(split);
+    let right_raw = base.clone().and(not_split);
+
+    if Collider::new(std::slice::from_ref(&left_raw)).is_err()
+        || Collider::new(std::slice::from_ref(&right_raw)).is_err()
+    {
+        return Err(ExprSplitDegradeReason::ColliderRejected);
+    }
+
+    let left_expr = simplify_and_bounds(left_raw);
+    let right_expr = simplify_and_bounds(right_raw);
+
+    if is_empty_and_conjunction(&left_expr) || is_empty_and_conjunction(&right_expr) {
+        return Err(ExprSplitDegradeReason::EmptyBranch);
+    }
+
+    Ok((left_expr, right_expr))
+}
+
+/// Detects whether a pure conjunction expression is definitely unsatisfiable.
+///
+/// Scope and intent:
+/// - This checker is intentionally conservative.
+/// - It only analyzes expressions that can be flattened into:
+///   `atom1 AND atom2 AND ...`
+/// - If any `OR` is present, it returns `false` (unknown / not handled here).
+///
+/// Strategy:
+/// - For each column, keep only the tightest lower bound (`>` / `>=`) and
+///   tightest upper bound (`<` / `<=`).
+/// - `=` is treated as both lower and upper bound at the same value.
+/// - `!=` is tracked per column to catch direct conflicts with `=`.
+/// - After bounds are collected, the conjunction is empty iff for any column:
+///   - lower value is greater than upper value, or
+///   - lower value equals upper value but at least one bound is exclusive.
+/// - For discrete domains (`Int*`, `UInt*`), adjacent open bounds with no
+///   representable value in between are also treated as empty.
+///
+/// Notes:
+/// - This is still a conservative fast path focused on conjunction emptiness
+///   detection for split degradation.
+/// - `split_partition_expr` currently restricts its main path to range-only
+///   conjunctions, but this helper remains slightly more general so shared
+///   bound collection and direct conflict checks stay reusable.
+fn is_empty_and_conjunction(expr: &PartitionExpr) -> bool {
+    let Some(collected) = collect_conjunction_bounds(expr) else {
+        return false;
+    };
+
+    if collected.has_conflict {
+        return true;
+    }
+
+    let CollectedConjunction {
+        lowers,
+        uppers,
+        not_equals,
+        passthrough: _,
+        has_conflict: _,
+    } = collected;
+
+    if lowers
+        .iter()
+        .any(|(col, lower)| !uppers.contains_key(col) && is_strictly_greater_than_domain_max(lower))
+    {
+        return true;
+    }
+
+    // Check for contradiction between collected lower/upper bounds per column.
+    lowers.into_iter().any(|(col, lower)| {
+        let Some(upper) = uppers.get(&col) else {
+            return false;
+        };
+
+        match lower.value.partial_cmp(&upper.value) {
+            Some(std::cmp::Ordering::Greater) => true,
+            Some(std::cmp::Ordering::Equal) => {
+                if !lower.inclusive || !upper.inclusive {
+                    true
+                } else {
+                    not_equals
+                        .get(&col)
+                        .is_some_and(|excluded| excluded.contains(&lower.value))
+                }
+            }
+            Some(std::cmp::Ordering::Less) => {
+                match (
+                    discrete_value_index(&lower.value),
+                    discrete_value_index(&upper.value),
+                ) {
+                    (Some(lower_idx), Some(upper_idx)) => {
+                        let min_candidate = if lower.inclusive {
+                            Some(lower_idx)
+                        } else {
+                            lower_idx.checked_add(1)
+                        };
+                        let max_candidate = if upper.inclusive {
+                            Some(upper_idx)
+                        } else {
+                            upper_idx.checked_sub(1)
+                        };
+                        match (min_candidate, max_candidate) {
+                            (Some(min_val), Some(max_val)) => min_val > max_val,
+                            _ => true,
+                        }
+                    }
+                    _ => false,
+                }
+            }
+            _ => false,
+        }
+    })
+}
+
+fn discrete_value_index(v: &Value) -> Option<i128> {
+    match v {
+        Value::Int8(x) => Some(*x as i128),
+        Value::Int16(x) => Some(*x as i128),
+        Value::Int32(x) => Some(*x as i128),
+        Value::Int64(x) => Some(*x as i128),
+        Value::UInt8(x) => Some(*x as i128),
+        Value::UInt16(x) => Some(*x as i128),
+        Value::UInt32(x) => Some(*x as i128),
+        Value::UInt64(x) => Some(*x as i128),
+        _ => None,
+    }
+}
+
+fn is_strictly_greater_than_domain_max(bound: &LowerBound) -> bool {
+    if bound.inclusive {
+        return false;
+    }
+
+    is_domain_max_value(&bound.value)
+}
+
+fn is_domain_max_value(v: &Value) -> bool {
+    match v {
+        Value::Float32(v) => v.0 == f32::MAX,
+        Value::Float64(v) => v.0 == f64::MAX,
+        Value::UInt8(v) => *v == u8::MAX,
+        Value::UInt16(v) => *v == u16::MAX,
+        Value::UInt32(v) => *v == u32::MAX,
+        Value::UInt64(v) => *v == u64::MAX,
+        Value::Int8(v) => *v == i8::MAX,
+        Value::Int16(v) => *v == i16::MAX,
+        Value::Int32(v) => *v == i32::MAX,
+        Value::Int64(v) => *v == i64::MAX,
+        _ => false,
+    }
+}
+
+/// Rewrites `NOT(expr)` into an equivalent `PartitionExpr` without introducing a unary NOT node.
+///
+/// Why this function exists:
+/// - `PartitionExpr` only models binary operators.
+/// - Cut logic needs `R AND NOT(S)`.
+/// - We therefore rewrite `NOT(S)` into an equivalent binary-expression tree.
+///
+/// Rewrite rules:
+/// - Atomic comparisons:
+///   - `=`  <-> `!=`
+///   - `<`  <-> `>=`
+///   - `<=` <-> `>`
+///   - `>`  <-> `<=`
+///   - `>=` <-> `<`
+/// - Boolean composition:
+///   - `NOT(A AND B)` => `NOT(A) OR NOT(B)`
+///   - `NOT(A OR B)`  => `NOT(A) AND NOT(B)`
+///
+/// Failure behavior:
+/// - For `AND/OR`, both sides must be `Operand::Expr`; otherwise returns `NoExprOperand`.
+/// - Any unsupported shape bubbles up as an error and the caller degrades to no-split.
+pub fn negate_split_expr(expr: &PartitionExpr) -> Result<PartitionExpr> {
+    match expr.op() {
+        RestrictedOp::Eq
+        | RestrictedOp::NotEq
+        | RestrictedOp::Lt
+        | RestrictedOp::LtEq
+        | RestrictedOp::Gt
+        | RestrictedOp::GtEq => {
+            // Atomic negate by operator inversion.
+            let op = match expr.op() {
+                RestrictedOp::Eq => RestrictedOp::NotEq,
+                RestrictedOp::NotEq => RestrictedOp::Eq,
+                RestrictedOp::Lt => RestrictedOp::GtEq,
+                RestrictedOp::LtEq => RestrictedOp::Gt,
+                RestrictedOp::Gt => RestrictedOp::LtEq,
+                RestrictedOp::GtEq => RestrictedOp::Lt,
+                RestrictedOp::And | RestrictedOp::Or => unreachable!(),
+            };
+            Ok(PartitionExpr::new(
+                expr.lhs().clone(),
+                op,
+                expr.rhs().clone(),
+            ))
+        }
+        RestrictedOp::And | RestrictedOp::Or => {
+            // De Morgan transform on recursive sub-expressions.
+            let lhs = match expr.lhs() {
+                Operand::Expr(lhs) => lhs,
+                other => {
+                    return error::NoExprOperandSnafu {
+                        operand: other.clone(),
+                    }
+                    .fail();
+                }
+            };
+            let rhs = match expr.rhs() {
+                Operand::Expr(rhs) => rhs,
+                other => {
+                    return error::NoExprOperandSnafu {
+                        operand: other.clone(),
+                    }
+                    .fail();
+                }
+            };
+            let not_lhs = negate_split_expr(lhs)?;
+            let not_rhs = negate_split_expr(rhs)?;
+            let op = match expr.op() {
+                // NOT(A AND B) => NOT(A) OR NOT(B)
+                RestrictedOp::And => RestrictedOp::Or,
+                // NOT(A OR B) => NOT(A) AND NOT(B)
+                RestrictedOp::Or => RestrictedOp::And,
+                _ => unreachable!(),
+            };
+            Ok(PartitionExpr::new(
+                Operand::Expr(not_lhs),
+                op,
+                Operand::Expr(not_rhs),
+            ))
+        }
+    }
+}
+
+pub fn validate_supported_expr(expr: &PartitionExpr) -> Result<()> {
+    match expr.op() {
+        RestrictedOp::And | RestrictedOp::Or => {
+            let lhs = match expr.lhs() {
+                Operand::Expr(lhs) => lhs,
+                other => {
+                    return error::NoExprOperandSnafu {
+                        operand: other.clone(),
+                    }
+                    .fail();
+                }
+            };
+            let rhs = match expr.rhs() {
+                Operand::Expr(rhs) => rhs,
+                other => {
+                    return error::NoExprOperandSnafu {
+                        operand: other.clone(),
+                    }
+                    .fail();
+                }
+            };
+            validate_supported_expr(lhs)?;
+            validate_supported_expr(rhs)?;
+            Ok(())
+        }
+        _ => validate_atomic(expr),
+    }
+}
+
+fn validate_atomic(expr: &PartitionExpr) -> Result<()> {
+    let (lhs, rhs) = (expr.lhs(), expr.rhs());
+    match (lhs, rhs) {
+        (Operand::Column(_), Operand::Value(v)) | (Operand::Value(v), Operand::Column(_)) => {
+            ensure!(
+                is_supported_value(v),
+                error::InvalidExprSnafu { expr: expr.clone() }
+            );
+            if is_nan_value(v) || is_infinite_value(v) {
+                return error::InvalidExprSnafu { expr: expr.clone() }.fail();
+            }
+            Ok(())
+        }
+        _ => error::InvalidExprSnafu { expr: expr.clone() }.fail(),
+    }
+}
+
+/// Validates that `base_expr` stays within the range-only split contract.
+///
+/// Scope and intent:
+/// - The split utility only handles interval-style partition predicates.
+/// - `base_expr` may mention multiple columns, but it must remain a pure `AND`
+///   tree of atomic range predicates.
+fn validate_base_expr_shape(expr: &PartitionExpr) -> bool {
+    let mut atoms = Vec::new();
+    if !collect_and_atoms(expr, &mut atoms) {
+        return false;
+    }
+
+    atoms
+        .into_iter()
+        .all(|atom| is_atomic_range_expr(&atom.canonicalize()))
+}
+
+/// Validates that `split_expr` is a single atomic range predicate.
+///
+/// This restriction keeps `NOT(split_expr)` in the same range-only subset so the
+/// resulting left/right branches stay within the supported contract.
+fn validate_split_expr_shape(expr: &PartitionExpr) -> bool {
+    is_atomic_range_expr(expr)
+}
+
+/// Returns whether `expr` is an atomic `column op value` range predicate.
+///
+/// Supported operators are limited to `<`, `<=`, `>`, and `>=`.
+fn is_atomic_range_expr(expr: &PartitionExpr) -> bool {
+    atom_col_op_val(expr).is_some_and(|(_, op, _)| {
+        matches!(
+            op,
+            RestrictedOp::Lt | RestrictedOp::LtEq | RestrictedOp::Gt | RestrictedOp::GtEq
+        )
+    })
+}
+
+fn is_supported_value(v: &Value) -> bool {
+    matches!(
+        v,
+        Value::Int8(_)
+            | Value::Int16(_)
+            | Value::Int32(_)
+            | Value::Int64(_)
+            | Value::UInt8(_)
+            | Value::UInt16(_)
+            | Value::UInt32(_)
+            | Value::UInt64(_)
+            | Value::Float32(_)
+            | Value::Float64(_)
+            | Value::String(_)
+    )
+}
+
+fn is_nan_value(v: &Value) -> bool {
+    match v {
+        Value::Float32(x) => x.0.is_nan(),
+        Value::Float64(x) => x.0.is_nan(),
+        _ => false,
+    }
+}
+
+fn is_infinite_value(v: &Value) -> bool {
+    match v {
+        Value::Float32(x) => x.0.is_infinite(),
+        Value::Float64(x) => x.0.is_infinite(),
+        _ => false,
+    }
+}
+
+#[derive(Debug, Clone)]
+struct LowerBound {
+    value: Value,
+    inclusive: bool,
+}
+
+#[derive(Debug, Clone)]
+struct UpperBound {
+    value: Value,
+    inclusive: bool,
+}
+
+struct CollectedConjunction {
+    lowers: BTreeMap<String, LowerBound>,
+    uppers: BTreeMap<String, UpperBound>,
+    not_equals: BTreeMap<String, HashSet<Value>>,
+    passthrough: Vec<PartitionExpr>,
+    has_conflict: bool,
+}
+
+/// Simplifies conjunction-only range predicates by keeping the tightest bounds per column.
+///
+/// This pass is intentionally conservative and only runs when the whole expression
+/// can be flattened into `atom1 AND atom2 AND ...` without any `OR` node.
+///
+/// Behavior:
+/// - For each column, collect all lower-bound predicates (`>` / `>=`) and keep the
+///   tightest one.
+/// - For each column, collect all upper-bound predicates (`<` / `<=`) and keep the
+///   tightest one.
+/// - Non-range predicates (for example `=` / `!=`) are preserved as-is.
+/// - If the expression contains `OR`, this function returns the original expression.
+///
+/// Tightness rules:
+/// - Upper bound: smaller value is tighter; if equal value, exclusive (`<`) is tighter.
+/// - Lower bound: larger value is tighter; if equal value, exclusive (`>`) is tighter.
+///
+/// Examples:
+/// - `a <= 10 AND a < 10` => `a < 10`
+/// - `a >= 10 AND a > 10` => `a > 10`
+/// - `a < 10 AND a < 5`   => `a < 5`
+fn simplify_and_bounds(expr: PartitionExpr) -> PartitionExpr {
+    let Some(collected) = collect_conjunction_bounds(&expr) else {
+        return expr;
+    };
+
+    let CollectedConjunction {
+        lowers,
+        uppers,
+        not_equals: _,
+        passthrough,
+        has_conflict: _,
+    } = collected;
+
+    let mut out = passthrough;
+    out.extend(lowers.into_iter().map(|(col, lower)| {
+        PartitionExpr::new(
+            Operand::Column(col),
+            if lower.inclusive {
+                RestrictedOp::GtEq
+            } else {
+                RestrictedOp::Gt
+            },
+            Operand::Value(lower.value),
+        )
+    }));
+    out.extend(uppers.into_iter().map(|(col, upper)| {
+        PartitionExpr::new(
+            Operand::Column(col),
+            if upper.inclusive {
+                RestrictedOp::LtEq
+            } else {
+                RestrictedOp::Lt
+            },
+            Operand::Value(upper.value),
+        )
+    }));
+
+    fold_and_exprs(out).unwrap_or(expr)
+}
+
+/// Flattens an expression into atomic terms when it is a pure conjunction tree.
+///
+/// Returns `false` if any `OR` is encountered, signaling caller to skip this
+/// simplification path.
+fn collect_and_atoms(expr: &PartitionExpr, out: &mut Vec<PartitionExpr>) -> bool {
+    match expr.op() {
+        RestrictedOp::And => {
+            let lhs = match expr.lhs() {
+                Operand::Expr(lhs) => lhs,
+                _ => return false,
+            };
+            let rhs = match expr.rhs() {
+                Operand::Expr(rhs) => rhs,
+                _ => return false,
+            };
+            collect_and_atoms(lhs, out) && collect_and_atoms(rhs, out)
+        }
+        RestrictedOp::Or => false,
+        _ => {
+            out.push(expr.clone());
+            true
+        }
+    }
+}
+
+/// Extracts `(column, op, value)` from a canonicalized atomic expression.
+fn atom_col_op_val(expr: &PartitionExpr) -> Option<(String, RestrictedOp, Value)> {
+    let lhs = expr.lhs();
+    let rhs = expr.rhs();
+    match (lhs, rhs) {
+        (Operand::Column(col), Operand::Value(v)) => {
+            Some((col.clone(), expr.op().clone(), v.clone()))
+        }
+        _ => None,
+    }
+}
+
+/// Collects per-column bounds and passthrough atoms from a pure `AND` tree.
+///
+/// Scope and intent:
+/// - This helper is shared by [`is_empty_and_conjunction`] and
+///   [`simplify_and_bounds`] so both paths interpret conjunction atoms the same
+///   way.
+/// - It only handles conjunction-only expressions. If any `OR` is present, it
+///   returns `None` and lets callers keep their conservative fallback behavior.
+///
+/// Behavior:
+/// - Tightest lower/upper bounds are recorded per column.
+/// - `=` contributes both a lower and an upper bound at the same value.
+/// - `!=` and non-range atoms are preserved in `passthrough` for callers that
+///   need to rebuild the conjunction.
+/// - `has_conflict` is set when atomic constraints already contradict each
+///   other (for example `a = 1 AND a <> 1`).
+///
+/// Notes:
+/// - This helper is intentionally a bit more general than the current
+///   `split_partition_expr` contract, which now only feeds range-only
+///   conjunctions into the main split path.
+fn collect_conjunction_bounds(expr: &PartitionExpr) -> Option<CollectedConjunction> {
+    let mut atoms = Vec::new();
+    if !collect_and_atoms(expr, &mut atoms) {
+        return None;
+    }
+
+    let mut lowers = BTreeMap::new();
+    let mut uppers = BTreeMap::new();
+    let mut equals = BTreeMap::new();
+    let mut not_equals: BTreeMap<String, HashSet<Value>> = BTreeMap::new();
+    let mut passthrough = Vec::new();
+    let mut seen = HashSet::new();
+    let mut has_conflict = false;
+
+    for atom in atoms {
+        let atom = atom.canonicalize();
+        let Some((col, op, val)) = atom_col_op_val(&atom) else {
+            push_unique_expr(&mut passthrough, &mut seen, atom);
+            continue;
+        };
+
+        match op {
+            RestrictedOp::Lt | RestrictedOp::LtEq => update_upper_bound(
+                &mut uppers,
+                col,
+                UpperBound {
+                    value: val,
+                    inclusive: matches!(op, RestrictedOp::LtEq),
+                },
+            ),
+            RestrictedOp::Gt | RestrictedOp::GtEq => update_lower_bound(
+                &mut lowers,
+                col,
+                LowerBound {
+                    value: val,
+                    inclusive: matches!(op, RestrictedOp::GtEq),
+                },
+            ),
+            RestrictedOp::Eq => {
+                if let Some(existing) = equals.get(&col)
+                    && existing != &val
+                {
+                    has_conflict = true;
+                }
+                if not_equals
+                    .get(&col)
+                    .is_some_and(|excluded| excluded.contains(&val))
+                {
+                    has_conflict = true;
+                }
+                equals.insert(col.clone(), val.clone());
+                update_lower_bound(
+                    &mut lowers,
+                    col.clone(),
+                    LowerBound {
+                        value: val.clone(),
+                        inclusive: true,
+                    },
+                );
+                update_upper_bound(
+                    &mut uppers,
+                    col,
+                    UpperBound {
+                        value: val,
+                        inclusive: true,
+                    },
+                );
+                push_unique_expr(&mut passthrough, &mut seen, atom);
+            }
+            RestrictedOp::NotEq => {
+                if equals.get(&col).is_some_and(|eq| eq == &val) {
+                    has_conflict = true;
+                }
+                not_equals.entry(col).or_default().insert(val);
+                push_unique_expr(&mut passthrough, &mut seen, atom);
+            }
+            RestrictedOp::And | RestrictedOp::Or => {
+                push_unique_expr(&mut passthrough, &mut seen, atom);
+            }
+        }
+    }
+
+    Some(CollectedConjunction {
+        lowers,
+        uppers,
+        not_equals,
+        passthrough,
+        has_conflict,
+    })
+}
+
+fn push_unique_expr(out: &mut Vec<PartitionExpr>, seen: &mut HashSet<String>, expr: PartitionExpr) {
+    let key = expr.to_string();
+    if seen.insert(key) {
+        out.push(expr);
+    }
+}
+
+fn update_upper_bound(
+    uppers: &mut BTreeMap<String, UpperBound>,
+    col: String,
+    candidate: UpperBound,
+) {
+    match uppers.get_mut(&col) {
+        Some(current) => {
+            if prefer_upper(&candidate, current) {
+                *current = candidate;
+            }
+        }
+        None => {
+            uppers.insert(col, candidate);
+        }
+    }
+}
+
+fn update_lower_bound(
+    lowers: &mut BTreeMap<String, LowerBound>,
+    col: String,
+    candidate: LowerBound,
+) {
+    match lowers.get_mut(&col) {
+        Some(current) => {
+            if prefer_lower(&candidate, current) {
+                *current = candidate;
+            }
+        }
+        None => {
+            lowers.insert(col, candidate);
+        }
+    }
+}
+
+fn prefer_upper(candidate: &UpperBound, current: &UpperBound) -> bool {
+    // "Smaller" upper bound is tighter. For equal value, exclusive is tighter.
+    match candidate.value.partial_cmp(&current.value) {
+        Some(std::cmp::Ordering::Less) => true,
+        Some(std::cmp::Ordering::Equal) => !candidate.inclusive && current.inclusive,
+        _ => false,
+    }
+}
+
+fn prefer_lower(candidate: &LowerBound, current: &LowerBound) -> bool {
+    // "Larger" lower bound is tighter. For equal value, exclusive is tighter.
+    match candidate.value.partial_cmp(&current.value) {
+        Some(std::cmp::Ordering::Greater) => true,
+        Some(std::cmp::Ordering::Equal) => !candidate.inclusive && current.inclusive,
+        _ => false,
+    }
+}
+
+/// Folds a list of expressions into a left-associated AND tree.
+/// Returns `None` if the input list is empty.
+fn fold_and_exprs(mut exprs: Vec<PartitionExpr>) -> Option<PartitionExpr> {
+    exprs.drain(..).reduce(|acc, next| acc.and(next))
+}
+
+#[cfg(test)]
+mod tests {
+    use datatypes::value::{OrderedFloat, Value};
+    use store_api::storage::RegionNumber;
+
+    use super::*;
+    use crate::checker::PartitionChecker;
+    use crate::expr::col;
+    use crate::multi_dim::MultiDimPartitionRule;
+
+    fn validate_cut_result_with_checker(
+        original_rule_exprs: &[PartitionExpr],
+        replaced_index: usize,
+        left: &Option<PartitionExpr>,
+        right: &Option<PartitionExpr>,
+        partition_columns: Vec<String>,
+        regions: Vec<RegionNumber>,
+    ) -> Result<()> {
+        ensure!(
+            replaced_index < original_rule_exprs.len(),
+            error::UnexpectedSnafu {
+                err_msg: format!(
+                    "replaced index out of bounds: {replaced_index} >= {}",
+                    original_rule_exprs.len()
+                )
+            }
+        );
+
+        let mut exprs = original_rule_exprs.to_vec();
+        exprs.remove(replaced_index);
+        exprs.extend(left.iter().cloned());
+        exprs.extend(right.iter().cloned());
+
+        ensure!(
+            !exprs.is_empty(),
+            error::UnexpectedSnafu {
+                err_msg: "empty rule exprs after split".to_string()
+            }
+        );
+
+        let final_regions = if regions.len() == exprs.len() {
+            regions
+        } else {
+            (0..exprs.len() as RegionNumber).collect()
+        };
+
+        let rule = MultiDimPartitionRule::try_new(partition_columns, final_regions, exprs, false)?;
+        let checker = PartitionChecker::try_new(&rule)?;
+        checker.check()?;
+        Ok(())
+    }
+
+    #[test]
+    fn test_split_simple_range() {
+        // R: a < 10
+        let base = col("a").lt(Value::Int64(10));
+        // S: a < 5
+        let split = col("a").lt(Value::Int64(5));
+        let (left, right) = split_partition_expr(base, split).unwrap();
+        // left = R AND S = a < 5
+        assert_eq!(left.to_string(), "a < 5");
+        // right = R AND NOT(S) = a >= 5 AND a < 10
+        assert_eq!(right.to_string(), "a >= 5 AND a < 10");
+    }
+
+    #[test]
+    fn test_split_string_interval() {
+        // R: v > 'm' AND v < 'n'
+        let base = col("v")
+            .gt(Value::String("m".into()))
+            .and(col("v").lt(Value::String("n".into())));
+        // S: v < 'm~'
+        let split = col("v").lt(Value::String("m~".into()));
+        let (left, right) = split_partition_expr(base, split).unwrap();
+        // left = (v > m AND v < n) AND (v < m~) -> v > m AND v < m~
+        assert_eq!(left.to_string(), "v > m AND v < m~");
+        // right = (v > m AND v < n) AND (v >= m~) -> v >= m~ AND v < n
+        assert_eq!(right.to_string(), "v >= m~ AND v < n");
+    }
+
+    #[test]
+    fn test_split_numeric_interval_mid_split() {
+        // R: a > 3 AND a < 10
+        let base = col("a")
+            .gt(Value::Int64(3))
+            .and(col("a").lt(Value::Int64(10)));
+        // S: a < 5
+        let split = col("a").lt(Value::Int64(5));
+
+        let (left, right) = split_partition_expr(base, split).unwrap();
+
+        // left = (a > 3 AND a < 10) AND (a < 5) -> a > 3 AND a < 5
+        assert_eq!(left.to_string(), "a > 3 AND a < 5");
+        // right = (a > 3 AND a < 10) AND (a >= 5) -> a >= 5 AND a < 10
+        assert_eq!(right.to_string(), "a >= 5 AND a < 10");
+    }
+
+    #[test]
+    fn test_split_base_expr_allows_unrelated_range_columns() {
+        // R: a > 20 AND b < 20
+        let base = col("a")
+            .gt(Value::Int64(20))
+            .and(col("b").lt(Value::Int64(20)));
+        // S: a < 30
+        let split = col("a").lt(Value::Int64(30));
+
+        let (left, right) = split_partition_expr(base, split).unwrap();
+
+        // left keeps the unrelated `b < 20` bound while splitting column `a`.
+        assert_eq!(left.to_string(), "a > 20 AND a < 30 AND b < 20");
+        // right also preserves the unrelated column bound.
+        assert_eq!(right.to_string(), "a >= 30 AND b < 20");
+    }
+
+    #[test]
+    fn test_split_degrade_on_unsupported_type() {
+        // intentionally excludes boolean from split-able value types.
+        let base = col("a").eq(Value::Boolean(true));
+        let split = col("a").eq(Value::Boolean(true));
+
+        let result = split_partition_expr(base, split);
+        assert_eq!(result.unwrap_err(), ExprSplitDegradeReason::UnsupportedType);
+    }
+
+    #[test]
+    fn test_validate_cut_result_with_checker() {
+        // Original partition set: a < 10, a >= 10
+        let original = vec![
+            col("a").lt(Value::Int64(10)),
+            col("a").gt_eq(Value::Int64(10)),
+        ];
+        let left = Some(col("a").lt(Value::Int64(5)));
+        let right = Some(
+            col("a")
+                .gt_eq(Value::Int64(5))
+                .and(col("a").lt(Value::Int64(10))),
+        );
+
+        validate_cut_result_with_checker(
+            &original,
+            0,
+            &left,
+            &right,
+            vec!["a".to_string()],
+            vec![1, 2, 3],
+        )
+        .unwrap();
+    }
+
+    #[test]
+    fn test_split_degrade_on_empty_branch() {
+        // R: a < 10
+        let base = col("a").lt(Value::Int64(10));
+        // S: a < 20
+        let split = col("a").lt(Value::Int64(20));
+
+        // right = (a < 10) AND (a >= 20) is unsatisfiable, should degrade.
+        let result = split_partition_expr(base, split);
+        assert_eq!(result.unwrap_err(), ExprSplitDegradeReason::EmptyBranch);
+    }
+
+    #[test]
+    fn test_split_rejects_eq_in_base_expr() {
+        // R: a = 5 falls outside the range-only base_expr contract.
+        let base = col("a").eq(Value::Int64(5));
+        // S: a < 6 remains a valid range split.
+        let split = col("a").lt(Value::Int64(6));
+
+        let result = split_partition_expr(base, split);
+        assert_eq!(result.unwrap_err(), ExprSplitDegradeReason::UnsupportedType);
+    }
+
+    #[test]
+    fn test_split_degrade_on_discrete_gap_int() {
+        // R: a < 5
+        let base = col("a").lt(Value::Int64(5));
+        // S: a <= 4
+        let split = col("a").lt_eq(Value::Int64(4));
+
+        // right = (a < 5) AND (a > 4) has no integer solution, should degrade.
+        let result = split_partition_expr(base, split);
+        assert_eq!(result.unwrap_err(), ExprSplitDegradeReason::EmptyBranch);
+    }
+
+    #[test]
+    fn test_split_degrade_on_unsupported_date_type() {
+        // Date is intentionally excluded from split-supported value types.
+        let base = col("d").lt(Value::Date(5.into()));
+        let split = col("d").lt_eq(Value::Date(4.into()));
+
+        let result = split_partition_expr(base, split);
+        assert_eq!(result.unwrap_err(), ExprSplitDegradeReason::UnsupportedType);
+    }
+
+    #[test]
+    fn test_split_degrade_on_unsupported_timestamp_type() {
+        // Timestamp is intentionally excluded from split-supported value types.
+        let base = col("ts").lt(Value::Timestamp(0.into()));
+        let split = col("ts").lt_eq(Value::Timestamp(1.into()));
+
+        let result = split_partition_expr(base, split);
+        assert_eq!(result.unwrap_err(), ExprSplitDegradeReason::UnsupportedType);
+    }
+
+    #[test]
+    fn test_split_rejects_not_eq_in_split_expr() {
+        // R: a >= 5 AND a <= 5
+        let base = col("a")
+            .gt_eq(Value::Int64(5))
+            .and(col("a").lt_eq(Value::Int64(5)));
+        // S: a <> 5 falls outside the range-only split_expr contract.
+        let split = col("a").not_eq(Value::Int64(5));
+
+        let result = split_partition_expr(base, split);
+        assert_eq!(result.unwrap_err(), ExprSplitDegradeReason::UnsupportedType);
+    }
+
+    #[test]
+    fn test_split_rejects_eq_in_split_expr() {
+        // R: a >= 5 AND a <= 5
+        let base = col("a")
+            .gt_eq(Value::Int64(5))
+            .and(col("a").lt_eq(Value::Int64(5)));
+        // S: a = 5 falls outside the range-only split_expr contract.
+        let split = col("a").eq(Value::Int64(5));
+
+        let result = split_partition_expr(base, split);
+        assert_eq!(result.unwrap_err(), ExprSplitDegradeReason::UnsupportedType);
+    }
+
+    #[test]
+    fn test_split_degrade_on_uint_one_sided_impossible_upper_bound() {
+        // R: a < 10 (UInt64 domain)
+        let base = col("a").lt(Value::UInt64(10));
+        // S: a < 0 is still satisfiable by NULL under null-first partition semantics.
+        // The split keeps a nullable left branch instead of degrading it as empty.
+        let split = col("a").lt(Value::UInt64(0));
+
+        let (left, right) = split_partition_expr(base, split).unwrap();
+        assert_eq!(left.to_string(), "a < 0");
+        assert_eq!(right.to_string(), "a >= 0 AND a < 10");
+    }
+
+    #[test]
+    fn test_split_degrade_on_uint_one_sided_impossible_lower_bound() {
+        // R: a < 10 (UInt64 domain)
+        let base = col("a").lt(Value::UInt64(10));
+        // S: a > u64::MAX (impossible on UInt64)
+        let split = col("a").gt(Value::UInt64(u64::MAX));
+
+        // left = (a < 10) AND (a > u64::MAX) is unsatisfiable on UInt64, should degrade.
+        let result = split_partition_expr(base, split);
+        assert_eq!(result.unwrap_err(), ExprSplitDegradeReason::EmptyBranch);
+    }
+
+    #[test]
+    fn test_split_degrade_on_int_one_sided_impossible_upper_bound() {
+        // R: a < 10 (Int64 domain)
+        let base = col("a").lt(Value::Int64(10));
+        // S: a < i64::MIN is still satisfiable by NULL under null-first partition semantics.
+        // The split keeps a nullable left branch instead of degrading it as empty.
+        let split = col("a").lt(Value::Int64(i64::MIN));
+
+        let (left, right) = split_partition_expr(base, split).unwrap();
+        assert_eq!(left.to_string(), format!("a < {}", i64::MIN));
+        assert_eq!(right.to_string(), format!("a >= {} AND a < 10", i64::MIN));
+    }
+
+    #[test]
+    fn test_split_degrade_on_int_one_sided_impossible_lower_bound() {
+        // R: a < 10 (Int64 domain)
+        let base = col("a").lt(Value::Int64(10));
+        // S: a > i64::MAX (impossible on Int64)
+        let split = col("a").gt(Value::Int64(i64::MAX));
+
+        // left = (a < 10) AND (a > i64::MAX) is unsatisfiable on Int64, should degrade.
+        let result = split_partition_expr(base, split);
+        assert_eq!(result.unwrap_err(), ExprSplitDegradeReason::EmptyBranch);
+    }
+
+    #[test]
+    fn test_split_degrade_on_string_one_sided_impossible_upper_bound() {
+        // R: s < "z" (String domain)
+        let base = col("s").lt(Value::String("z".into()));
+        // S: s < "" is still satisfiable by NULL under null-first partition semantics.
+        // The split keeps a nullable left branch instead of degrading it as empty.
+        let split = col("s").lt(Value::String("".into()));
+
+        let (left, right) = split_partition_expr(base, split).unwrap();
+        assert_eq!(left.to_string(), "s < ");
+        assert_eq!(right.to_string(), "s >=  AND s < z");
+    }
+
+    #[test]
+    fn test_split_degrade_on_float64_one_sided_impossible_upper_bound() {
+        // R: a < 10.0 (Float64 domain)
+        let base = col("a").lt(Value::Float64(OrderedFloat(10.0)));
+        // S: a < f64::MIN is still satisfiable by NULL under null-first partition semantics.
+        // The split keeps a nullable left branch instead of degrading it as empty.
+        let split = col("a").lt(Value::Float64(OrderedFloat(f64::MIN)));
+
+        let (left, right) = split_partition_expr(base, split).unwrap();
+        assert_eq!(left.to_string(), format!("a < {}", f64::MIN));
+        assert_eq!(right.to_string(), format!("a >= {} AND a < 10", f64::MIN));
+    }
+
+    #[test]
+    fn test_split_degrade_on_float64_one_sided_impossible_lower_bound() {
+        // R: a < 10.0 (Float64 domain)
+        let base = col("a").lt(Value::Float64(OrderedFloat(10.0)));
+        // S: a > f64::MAX (impossible with finite-only float policy)
+        let split = col("a").gt(Value::Float64(OrderedFloat(f64::MAX)));
+
+        // left = (a < 10.0) AND (a > f64::MAX) is unsatisfiable, should degrade.
+        let result = split_partition_expr(base, split);
+        assert_eq!(result.unwrap_err(), ExprSplitDegradeReason::EmptyBranch);
+    }
+
+    #[test]
+    fn test_split_degrade_on_float32_one_sided_impossible_upper_bound() {
+        // R: a < 10.0f32 (Float32 domain)
+        let base = col("a").lt(Value::Float32(OrderedFloat(10.0)));
+        // S: a < f32::MIN is still satisfiable by NULL under null-first partition semantics.
+        // The split keeps a nullable left branch instead of degrading it as empty.
+        let split = col("a").lt(Value::Float32(OrderedFloat(f32::MIN)));
+
+        let (left, right) = split_partition_expr(base, split).unwrap();
+        assert_eq!(left.to_string(), format!("a < {}", f32::MIN));
+        assert_eq!(right.to_string(), format!("a >= {} AND a < 10", f32::MIN));
+    }
+
+    #[test]
+    fn test_split_degrade_on_float32_one_sided_impossible_lower_bound() {
+        // R: a < 10.0f32 (Float32 domain)
+        let base = col("a").lt(Value::Float32(OrderedFloat(10.0)));
+        // S: a > f32::MAX (impossible with finite-only float policy)
+        let split = col("a").gt(Value::Float32(OrderedFloat(f32::MAX)));
+
+        // left = (a < 10.0f32) AND (a > f32::MAX) is unsatisfiable, should degrade.
+        let result = split_partition_expr(base, split);
+        assert_eq!(result.unwrap_err(), ExprSplitDegradeReason::EmptyBranch);
+    }
+
+    #[test]
+    fn test_simplify_same_upper_bound_prefers_strict() {
+        // a <= 10 AND a < 10 => a < 10
+        let expr = col("a")
+            .lt_eq(Value::Int64(10))
+            .and(col("a").lt(Value::Int64(10)));
+
+        let simplified = simplify_and_bounds(expr);
+        assert_eq!(simplified.to_string(), "a < 10");
+    }
+
+    #[test]
+    fn test_simplify_same_lower_bound_prefers_strict() {
+        // a >= 10 AND a > 10 => a > 10
+        let expr = col("a")
+            .gt_eq(Value::Int64(10))
+            .and(col("a").gt(Value::Int64(10)));
+
+        let simplified = simplify_and_bounds(expr);
+        assert_eq!(simplified.to_string(), "a > 10");
+    }
+
+    #[test]
+    fn test_negate_split_expr_demorgan_and() {
+        // expr: (a < 10) AND (a >= 3)
+        let expr = col("a")
+            .lt(Value::Int64(10))
+            .and(col("a").gt_eq(Value::Int64(3)));
+        let not_expr = negate_split_expr(&expr).unwrap();
+        // NOT(expr) => (a >= 10) OR (a < 3)
+        assert_eq!(not_expr.to_string(), "a >= 10 OR a < 3");
+    }
+
+    #[test]
+    fn test_negate_split_expr_demorgan_or() {
+        // expr: (a = 1) OR (a <> 2)
+        let expr = PartitionExpr::new(
+            Operand::Expr(col("a").eq(Value::Int64(1))),
+            RestrictedOp::Or,
+            Operand::Expr(col("a").not_eq(Value::Int64(2))),
+        );
+        let not_expr = negate_split_expr(&expr).unwrap();
+        // NOT(expr) => (a <> 1) AND (a = 2)
+        assert_eq!(not_expr.to_string(), "a <> 1 AND a = 2");
+    }
+
+    #[test]
+    fn test_negate_split_expr_invalid_and_operand() {
+        // malformed AND: rhs is a scalar value, not an Expr subtree.
+        let malformed = PartitionExpr {
+            lhs: Box::new(Operand::Expr(col("a").lt(Value::Int64(10)))),
+            op: RestrictedOp::And,
+            rhs: Box::new(Operand::Value(Value::Int64(1))),
+        };
+        assert!(negate_split_expr(&malformed).is_err());
+    }
+
+    #[test]
+    fn test_validate_supported_expr_value_column_allowed() {
+        // Canonicalization can flip to column-value; validator must accept value-column input.
+        let expr = PartitionExpr::new(
+            Operand::Value(Value::Int64(10)),
+            RestrictedOp::Lt,
+            Operand::Column("a".to_string()),
+        );
+        assert!(validate_supported_expr(&expr).is_ok());
+    }
+
+    #[test]
+    fn test_validate_supported_expr_invalid_atomic_shape() {
+        // column-column atomic comparison is out of shape.
+        let expr = PartitionExpr::new(
+            Operand::Column("a".to_string()),
+            RestrictedOp::Eq,
+            Operand::Column("b".to_string()),
+        );
+        assert!(validate_supported_expr(&expr).is_err());
+    }
+
+    #[test]
+    fn test_validate_supported_expr_nan_comparison_rejected() {
+        // NaN cannot be used in any supported comparison predicate.
+        let expr = col("a").lt(Value::Float64(OrderedFloat(f64::NAN)));
+        assert!(validate_supported_expr(&expr).is_err());
+    }
+
+    #[test]
+    fn test_validate_supported_expr_infinite_comparison_rejected() {
+        // Infinity cannot be used in any supported comparison predicate under
+        // finite-only float policy.
+        let pos_inf = col("a").gt(Value::Float64(OrderedFloat(f64::INFINITY)));
+        let neg_inf = col("a").lt(Value::Float32(OrderedFloat(f32::NEG_INFINITY)));
+        assert!(validate_supported_expr(&pos_inf).is_err());
+        assert!(validate_supported_expr(&neg_inf).is_err());
+    }
+
+    #[test]
+    fn test_validate_supported_expr_nan_eq_rejected() {
+        let expr = col("a").eq(Value::Float64(OrderedFloat(f64::NAN)));
+        assert!(validate_supported_expr(&expr).is_err());
+    }
+
+    #[test]
+    fn test_validate_supported_expr_infinite_eq_rejected() {
+        let pos_inf = col("a").eq(Value::Float64(OrderedFloat(f64::INFINITY)));
+        let neg_inf = col("a").not_eq(Value::Float32(OrderedFloat(f32::NEG_INFINITY)));
+        assert!(validate_supported_expr(&pos_inf).is_err());
+        assert!(validate_supported_expr(&neg_inf).is_err());
+    }
+
+    #[test]
+    fn test_simplify_and_bounds_or_keeps_original() {
+        // OR tree is intentionally not flattened by AND-only simplifier.
+        let expr = PartitionExpr::new(
+            Operand::Expr(col("a").lt(Value::Int64(10))),
+            RestrictedOp::Or,
+            Operand::Expr(col("a").gt_eq(Value::Int64(20))),
+        );
+        let simplified = simplify_and_bounds(expr.clone());
+        assert_eq!(simplified.to_string(), expr.to_string());
+    }
+
+    #[test]
+    fn test_simplify_and_bounds_keep_stronger_when_weaker_seen_later() {
+        // upper: stronger bound first, weaker later -> keep stronger (< 5).
+        let upper = col("a")
+            .lt(Value::Int64(5))
+            .and(col("a").lt(Value::Int64(10)));
+        assert_eq!(simplify_and_bounds(upper).to_string(), "a < 5");
+
+        // lower: stronger bound first, weaker later -> keep stronger (> 10).
+        let lower = col("a")
+            .gt(Value::Int64(10))
+            .and(col("a").gt(Value::Int64(5)));
+        assert_eq!(simplify_and_bounds(lower).to_string(), "a > 10");
+    }
+
+    #[test]
+    fn test_internal_helpers_uncovered_branches() {
+        // Empty AND fold should return None.
+        assert!(fold_and_exprs(vec![]).is_none());
+
+        // Any OR in tree disables AND-bound simplification path.
+        let mut out = Vec::new();
+        let or_expr = PartitionExpr::new(
+            Operand::Expr(col("a").lt(Value::Int64(10))),
+            RestrictedOp::Or,
+            Operand::Expr(col("a").gt_eq(Value::Int64(20))),
+        );
+        assert!(!collect_and_atoms(&or_expr, &mut out));
+
+        // value-value atom has no (column, op, value) projection.
+        let value_value = PartitionExpr::new(
+            Operand::Value(Value::Int64(1)),
+            RestrictedOp::Eq,
+            Operand::Value(Value::Int64(2)),
+        );
+        assert!(atom_col_op_val(&value_value).is_none());
+    }
+
+    #[test]
+    fn test_split_rejects_or_in_base_expr() {
+        // R: (a < 10) OR (a >= 20 AND a < 30) falls outside the AND-only base_expr contract.
+        let base = PartitionExpr::new(
+            Operand::Expr(col("a").lt(Value::Int64(10))),
+            RestrictedOp::Or,
+            Operand::Expr(
+                col("a")
+                    .gt_eq(Value::Int64(20))
+                    .and(col("a").lt(Value::Int64(30))),
+            ),
+        );
+        // S: a < 5
+        let split = col("a").lt(Value::Int64(5));
+
+        let result = split_partition_expr(base, split);
+        assert_eq!(result.unwrap_err(), ExprSplitDegradeReason::UnsupportedType);
+    }
+
+    #[test]
+    fn test_split_rejects_or_in_split_expr() {
+        // R: a < 10
+        let base = col("a").lt(Value::Int64(10));
+        // S: (a < 5) OR (a >= 8 AND a < 9) falls outside the atomic split_expr contract.
+        let split = PartitionExpr::new(
+            Operand::Expr(col("a").lt(Value::Int64(5))),
+            RestrictedOp::Or,
+            Operand::Expr(
+                col("a")
+                    .gt_eq(Value::Int64(8))
+                    .and(col("a").lt(Value::Int64(9))),
+            ),
+        );
+
+        let result = split_partition_expr(base, split);
+        assert_eq!(result.unwrap_err(), ExprSplitDegradeReason::UnsupportedType);
+    }
+}
diff --git a/src/pipeline/src/error.rs b/src/pipeline/src/error.rs
index 4130112129..2da62dda2e 100644
--- a/src/pipeline/src/error.rs
+++ b/src/pipeline/src/error.rs
@@ -389,6 +389,61 @@ pub enum Error {
         #[snafu(implicit)]
         location: Location,
     },
+    #[snafu(display("Transform index `type` must be set."))]
+    TransformIndexTypeMustBeSet {
+        #[snafu(implicit)]
+        location: Location,
+    },
+    #[snafu(display("Unsupported field in transform index config: {field}"))]
+    TransformIndexUnsupportedField {
+        field: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+    #[snafu(display(
+        "Transform index option `{field}` must be a string, boolean, integer, or real scalar"
+    ))]
+    TransformIndexOptionMustBeScalar {
+        field: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+    #[snafu(display("Index type `{index}` does not support options in pipeline config"))]
+    TransformIndexOptionsUnsupported {
+        index: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+    #[snafu(display("Unsupported option `{key}` for `{index}` index"))]
+    TransformIndexOptionUnsupported {
+        index: String,
+        key: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+    #[snafu(display("Index `{index}` only supports {expected} columns, but got {actual}"))]
+    TransformIndexTypeMismatch {
+        index: String,
+        expected: String,
+        actual: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+    #[snafu(display("Invalid options for `{index}` index"))]
+    TransformIndexOption {
+        index: String,
+        #[snafu(source)]
+        source: datatypes::error::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+    #[snafu(display("Transform index `{index}` does not match options declared for `{options}`"))]
+    TransformIndexStateMismatch {
+        index: String,
+        options: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
     #[snafu(display("Column name must be unique, but got duplicated: {duplicates}"))]
     TransformColumnNameMustBeUnique {
         duplicates: String,
@@ -897,6 +952,14 @@ impl ErrorExt for Error {
             | TransformElementMustBeMap { .. }
             | TransformFieldMustBeSet { .. }
             | TransformTypeMustBeSet { .. }
+            | TransformIndexTypeMustBeSet { .. }
+            | TransformIndexUnsupportedField { .. }
+            | TransformIndexOptionMustBeScalar { .. }
+            | TransformIndexOptionsUnsupported { .. }
+            | TransformIndexOptionUnsupported { .. }
+            | TransformIndexTypeMismatch { .. }
+            | TransformIndexOption { .. }
+            | TransformIndexStateMismatch { .. }
             | TransformColumnNameMustBeUnique { .. }
             | TransformMultipleTimestampIndex { .. }
             | TransformTimestampIndexCount { .. }
diff --git a/src/pipeline/src/etl.rs b/src/pipeline/src/etl.rs
index dd4d540376..26d3520829 100644
--- a/src/pipeline/src/etl.rs
+++ b/src/pipeline/src/etl.rs
@@ -482,6 +482,7 @@ mod tests {
     use std::sync::Arc;
 
     use api::v1::Rows;
+    use datatypes::schema::{FulltextOptions, SkippingIndexOptions};
     use greptime_proto::v1::value::ValueData;
     use greptime_proto::v1::{self, ColumnDataType, SemanticType};
     use vrl::prelude::Bytes;
@@ -1467,4 +1468,88 @@ transform:
         // Empty array should produce empty HashMap
         assert_eq!(rows_by_context.len(), 0);
     }
+
+    #[test]
+    fn test_pipeline_detailed_index_options_roundtrip() {
+        let pipeline_yaml = r#"
+transform:
+  - field: message
+    type: string
+    index:
+      type: fulltext
+      options:
+        analyzer: Chinese
+        case_sensitive: true
+        backend: tantivy
+  - field: trace_id
+    type: int64
+    index:
+      type: skipping
+      options:
+        granularity: 2048
+        false_positive_rate: 0.02
+        type: BLOOM
+  - field: ts
+    type: timestamp, ns
+    index: time
+"#;
+
+        let pipeline: Pipeline = parse(&Content::Yaml(pipeline_yaml)).unwrap();
+        let schema = pipeline.schemas().unwrap().clone();
+
+        let message = schema
+            .iter()
+            .find(|column| column.column_name == "message")
+            .unwrap();
+        let trace_id = schema
+            .iter()
+            .find(|column| column.column_name == "trace_id")
+            .unwrap();
+        let message_options = message.options.clone();
+        let trace_id_options = trace_id.options.clone();
+
+        let fulltext: FulltextOptions = serde_json::from_str(
+            message
+                .options
+                .as_ref()
+                .unwrap()
+                .options
+                .get("fulltext")
+                .unwrap(),
+        )
+        .unwrap();
+        assert!(fulltext.enable);
+        assert_eq!(fulltext.analyzer.to_string(), "Chinese");
+        assert!(fulltext.case_sensitive);
+        assert_eq!(fulltext.backend.to_string(), "tantivy");
+
+        let skipping: SkippingIndexOptions = serde_json::from_str(
+            trace_id
+                .options
+                .as_ref()
+                .unwrap()
+                .options
+                .get("skipping_index")
+                .unwrap(),
+        )
+        .unwrap();
+        assert_eq!(skipping.granularity, 2048);
+        assert_eq!(skipping.false_positive_rate(), 0.02);
+        assert_eq!(skipping.index_type.to_string(), "BLOOM");
+
+        let roundtrip_schema = SchemaInfo::from_schema_list(schema)
+            .column_schemas()
+            .unwrap();
+        let roundtrip_message = roundtrip_schema
+            .iter()
+            .find(|column| column.column_name == "message")
+            .unwrap();
+        let roundtrip_trace_id = roundtrip_schema
+            .iter()
+            .find(|column| column.column_name == "trace_id")
+            .unwrap();
+
+        assert_eq!(message_options, roundtrip_message.options);
+        assert_eq!(trace_id_options, roundtrip_trace_id.options);
+    }
 }
diff --git a/src/pipeline/src/etl/ctx_req.rs b/src/pipeline/src/etl/ctx_req.rs
index 23873cfdf1..0bfac33dac 100644
--- a/src/pipeline/src/etl/ctx_req.rs
+++ b/src/pipeline/src/etl/ctx_req.rs
@@ -223,7 +223,7 @@ impl ContextReq {
     }
 
     pub fn all_req(self) -> impl Iterator<Item = RowInsertRequest> {
-        self.req.into_iter().flat_map(|(_, req)| req)
+        self.req.into_values().flatten()
     }
 
     pub fn ref_all_req(&self) -> impl Iterator<Item = &RowInsertRequest> {
diff --git a/src/pipeline/src/etl/processor/dissect.rs b/src/pipeline/src/etl/processor/dissect.rs
index 6694049253..69bdbe0f89 100644
--- a/src/pipeline/src/etl/processor/dissect.rs
+++ b/src/pipeline/src/etl/processor/dissect.rs
@@ -527,7 +527,7 @@ impl DissectProcessor {
             };
 
             for (name, mut values) in appends {
-                values.sort_by(|a, b| a.1.cmp(&b.1));
+                values.sort_by_key(|a| a.1);
                 let value = values.into_iter().map(|(a, _)| a).join(sep);
                 map.push((name, VrlValue::Bytes(Bytes::from(value))));
             }
diff --git a/src/pipeline/src/etl/processor/filter.rs b/src/pipeline/src/etl/processor/filter.rs
index a700a8f2d3..d1820ba8ff 100644
--- a/src/pipeline/src/etl/processor/filter.rs
+++ b/src/pipeline/src/etl/processor/filter.rs
@@ -158,20 +158,18 @@ impl Processor for FilterProcessor {
         for field in self.fields.iter() {
             let val = val.as_object_mut().context(ValueMustBeMapSnafu)?;
             let index = field.input_field();
-            match val.get(index) {
-                Some(VrlValue::Bytes(b)) => {
+            if let Some(v) = val.get(index) {
+                if let VrlValue::Bytes(b) = v {
                     if self.match_target(&String::from_utf8_lossy(b)) {
                         return Ok(VrlValue::Null);
                     }
-                }
-                Some(v) => {
+                } else {
                     return ProcessorExpectStringSnafu {
                         processor: self.kind(),
                         v: v.clone(),
                     }
                     .fail();
                 }
-                None => {}
             }
         }
 
diff --git a/src/pipeline/src/etl/transform.rs b/src/pipeline/src/etl/transform.rs
index b6ce8d54cb..efcb4765ad 100644
--- a/src/pipeline/src/etl/transform.rs
+++ b/src/pipeline/src/etl/transform.rs
@@ -15,15 +15,24 @@
 pub mod index;
 pub mod transformer;
 
+use std::collections::HashMap;
+
 use api::v1::ColumnDataType;
 use api::v1::value::ValueData;
 use chrono::Utc;
-use snafu::{OptionExt, ensure};
+use datatypes::schema::{FulltextOptions, SkippingIndexOptions};
+use snafu::{OptionExt, ResultExt, ensure};
+use sql::parsers::utils::{
+    validate_column_fulltext_create_option, validate_column_skipping_index_create_option,
+};
 
 use crate::error::{
-    Error, KeyMustBeStringSnafu, Result, TransformElementMustBeMapSnafu,
-    TransformFieldMustBeSetSnafu, TransformOnFailureInvalidValueSnafu, TransformTypeMustBeSetSnafu,
-    UnsupportedTypeInPipelineSnafu,
+    Error, FieldMustBeTypeSnafu, KeyMustBeStringSnafu, Result, TransformElementMustBeMapSnafu,
+    TransformFieldMustBeSetSnafu, TransformIndexOptionMustBeScalarSnafu, TransformIndexOptionSnafu,
+    TransformIndexOptionUnsupportedSnafu, TransformIndexOptionsUnsupportedSnafu,
+    TransformIndexTypeMismatchSnafu, TransformIndexTypeMustBeSetSnafu,
+    TransformIndexUnsupportedFieldSnafu, TransformOnFailureInvalidValueSnafu,
+    TransformTypeMustBeSetSnafu, UnsupportedTypeInPipelineSnafu,
 };
 use crate::etl::field::Fields;
 use crate::etl::processor::{yaml_bool, yaml_new_field, yaml_new_fields, yaml_string};
@@ -34,6 +43,9 @@ const TRANSFORM_FIELD: &str = "field";
 const TRANSFORM_FIELDS: &str = "fields";
 const TRANSFORM_TYPE: &str = "type";
 const TRANSFORM_INDEX: &str = "index";
+const TRANSFORM_INDEX_TYPE_FIELD: &str = "index.type";
+const TRANSFORM_INDEX_OPTIONS: &str = "options";
+const TRANSFORM_INDEX_OPTIONS_FIELD: &str = "index.options";
 const TRANSFORM_TAG: &str = "tag";
 const TRANSFORM_DEFAULT: &str = "default";
 const TRANSFORM_ON_FAILURE: &str = "on_failure";
@@ -131,10 +143,42 @@ pub struct Transform {
     pub type_: ColumnDataType,
     pub default: Option<ValueData>,
     pub index: Option<Index>,
+    pub index_options: Option<TransformIndexOptions>,
     pub tag: bool,
     pub on_failure: Option<OnFailure>,
 }
 
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum TransformIndexOptions {
+    Fulltext(FulltextOptions),
+    Skipping(SkippingIndexOptions),
+}
+
+impl TransformIndexOptions {
+    pub(crate) fn index(&self) -> Index {
+        match self {
+            TransformIndexOptions::Fulltext(_) => Index::Fulltext,
+            TransformIndexOptions::Skipping(_) => Index::Skipping,
+        }
+    }
+
+    #[cfg(test)]
+    pub(crate) fn as_fulltext(&self) -> Option<&FulltextOptions> {
+        match self {
+            TransformIndexOptions::Fulltext(options) => Some(options),
+            TransformIndexOptions::Skipping(_) => None,
+        }
+    }
+
+    #[cfg(test)]
+    pub(crate) fn as_skipping(&self) -> Option<&SkippingIndexOptions> {
+        match self {
+            TransformIndexOptions::Skipping(options) => Some(options),
+            TransformIndexOptions::Fulltext(_) => None,
+        }
+    }
+}
+
 // valid types
 // ColumnDataType::Int8
 // ColumnDataType::Int16
@@ -231,13 +275,157 @@ fn get_default_for_type(ty: &ColumnDataType) -> Result<ValueData> {
     Ok(v)
 }
 
+fn parse_transform_index(
+    value: &yaml_rust::Yaml,
+) -> Result<(Index, Option<HashMap<String, String>>)> {
+    match value {
+        yaml_rust::Yaml::String(_) => {
+            let index_str = yaml_string(value, TRANSFORM_INDEX)?;
+            Ok((index_str.try_into()?, None))
+        }
+        yaml_rust::Yaml::Hash(hash) => {
+            let mut index = None;
+            let mut index_options = None;
+
+            for (k, v) in hash {
+                let key = k
+                    .as_str()
+                    .with_context(|| KeyMustBeStringSnafu { k: k.clone() })?;
+                match key {
+                    TRANSFORM_TYPE => {
+                        let index_str = yaml_string(v, TRANSFORM_INDEX_TYPE_FIELD)?;
+                        index = Some(index_str.try_into()?);
+                    }
+                    TRANSFORM_INDEX_OPTIONS => {
+                        index_options = Some(parse_transform_index_options(v)?);
+                    }
+                    _ => {
+                        return TransformIndexUnsupportedFieldSnafu {
+                            field: key.to_string(),
+                        }
+                        .fail();
+                    }
+                }
+            }
+
+            Ok((
+                index.context(TransformIndexTypeMustBeSetSnafu)?,
+                index_options,
+            ))
+        }
+        _ => FieldMustBeTypeSnafu {
+            field: TRANSFORM_INDEX,
+            ty: "string or map",
+        }
+        .fail(),
+    }
+}
+
+fn parse_transform_index_options(value: &yaml_rust::Yaml) -> Result<HashMap<String, String>> {
+    let hash = value.as_hash().context(FieldMustBeTypeSnafu {
+        field: TRANSFORM_INDEX_OPTIONS_FIELD,
+        ty: "map",
+    })?;
+    let mut options = HashMap::with_capacity(hash.len());
+
+    for (k, v) in hash {
+        let key = k
+            .as_str()
+            .with_context(|| KeyMustBeStringSnafu { k: k.clone() })?;
+
+        let field = format!("{TRANSFORM_INDEX_OPTIONS_FIELD}.{key}");
+        let value = match v {
+            yaml_rust::Yaml::String(v) => v.clone(),
+            yaml_rust::Yaml::Boolean(v) => v.to_string(),
+            yaml_rust::Yaml::Integer(v) => v.to_string(),
+            yaml_rust::Yaml::Real(v) => v.clone(),
+            _ => {
+                return TransformIndexOptionMustBeScalarSnafu { field }.fail();
+            }
+        };
+        options.insert(key.to_string(), value);
+    }
+
+    Ok(options)
+}
+
+fn lower_typed_transform_index_options<T>(
+    index: Index,
+    index_options: Option<HashMap<String, String>>,
+    validate: fn(&str) -> bool,
+    wrap: fn(T) -> TransformIndexOptions,
+) -> Result<Option<TransformIndexOptions>>
+where
+    T: TryFrom<HashMap<String, String>, Error = datatypes::error::Error>,
+{
+    index_options
+        .map(|opts| {
+            for key in opts.keys() {
+                ensure!(
+                    validate(key),
+                    TransformIndexOptionUnsupportedSnafu {
+                        index: index.to_string(),
+                        key: key.clone(),
+                    }
+                );
+            }
+
+            let options = opts.try_into().context(TransformIndexOptionSnafu {
+                index: index.to_string(),
+            })?;
+
+            Ok(wrap(options))
+        })
+        .transpose()
+}
+
+fn lower_transform_index_options(
+    index: Index,
+    column_type: &ColumnDataType,
+    index_options: Option<HashMap<String, String>>,
+) -> Result<Option<TransformIndexOptions>> {
+    match index {
+        Index::Fulltext => {
+            ensure!(
+                *column_type == ColumnDataType::String,
+                TransformIndexTypeMismatchSnafu {
+                    index: index.to_string(),
+                    expected: ColumnDataType::String.as_str_name().to_string(),
+                    actual: column_type.as_str_name().to_string(),
+                }
+            );
+
+            lower_typed_transform_index_options(
+                index,
+                index_options,
+                validate_column_fulltext_create_option,
+                TransformIndexOptions::Fulltext,
+            )
+        }
+        Index::Skipping => lower_typed_transform_index_options(
+            index,
+            index_options,
+            validate_column_skipping_index_create_option,
+            TransformIndexOptions::Skipping,
+        ),
+        Index::Inverted | Index::Time | Index::Tag => {
+            ensure!(
+                index_options.is_none(),
+                TransformIndexOptionsUnsupportedSnafu {
+                    index: index.to_string(),
+                }
+            );
+            Ok(None)
+        }
+    }
+}
 impl TryFrom<&yaml_rust::yaml::Hash> for Transform {
     type Error = Error;
 
     fn try_from(hash: &yaml_rust::yaml::Hash) -> Result<Self> {
         let mut fields = Fields::default();
         let mut default = None;
-        let mut index = None;
+        let mut index_value = None;
         let mut tag = false;
         let mut on_failure = None;
 
@@ -262,8 +450,7 @@ impl TryFrom<&yaml_rust::yaml::Hash> for Transform {
                 }
 
                 TRANSFORM_INDEX => {
-                    let index_str = yaml_string(v, TRANSFORM_INDEX)?;
-                    index = Some(index_str.try_into()?);
+                    index_value = Some(v);
                 }
 
                 TRANSFORM_TAG => {
@@ -299,6 +486,16 @@ impl TryFrom<&yaml_rust::yaml::Hash> for Transform {
             fields: format!("{:?}", fields),
         })?;
 
+        let (index, index_options) = match index_value {
+            Some(value) => {
+                let (index, raw_index_options) = parse_transform_index(value)?;
+                let index_options =
+                    lower_transform_index_options(index, &type_, raw_index_options)?;
+                (Some(index), index_options)
+            }
+            None => (None, None),
+        };
+
         let final_default = if let Some(default_value) = default {
             let target = parse_str_value(&type_, &default_value)?;
             on_failure = Some(OnFailure::Default);
@@ -312,6 +509,7 @@ impl TryFrom<&yaml_rust::yaml::Hash> for Transform {
             type_,
             default: final_default,
             index,
+            index_options,
             on_failure,
             tag,
         };
@@ -319,3 +517,214 @@ impl TryFrom<&yaml_rust::yaml::Hash> for Transform {
         Ok(builder)
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use yaml_rust::YamlLoader;
+
+    use super::*;
+
+    fn parse_transform(yaml: &str) -> Result<Transform> {
+        let docs = YamlLoader::load_from_str(yaml).unwrap();
+        docs[0].as_hash().unwrap().try_into()
+    }
+
+    #[test]
+    fn test_transform_parses_legacy_string_index() {
+        let transform = parse_transform(
+            r#"
+field: message
+type: string
+index: fulltext
+"#,
+        )
+        .unwrap();
+
+        assert_eq!(transform.index, Some(Index::Fulltext));
+        assert!(transform.index_options.is_none());
+    }
+
+    #[test]
+    fn test_transform_parses_index_object_without_options() {
+        let transform = parse_transform(
+            r#"
+field: message
+type: string
+index:
+  type: inverted
+"#,
+        )
+        .unwrap();
+
+        assert_eq!(transform.index, Some(Index::Inverted));
+        assert!(transform.index_options.is_none());
+    }
+
+    #[test]
+    fn test_transform_parses_index_object_with_scalar_options() {
+        let transform = parse_transform(
+            r#"
+field: message
+type: string
+index:
+  type: fulltext
+  options:
+    analyzer: English
+    case_sensitive: false
+    granularity: 2048
+    false_positive_rate: 0.02
+"#,
+        )
+        .unwrap();
+
+        assert_eq!(transform.index, Some(Index::Fulltext));
+        let options = transform.index_options.as_ref().unwrap();
+        let fulltext = options.as_fulltext().unwrap();
+        assert!(fulltext.enable);
+        assert_eq!(fulltext.analyzer.to_string(), "English");
+        assert!(!fulltext.case_sensitive);
+        assert_eq!(fulltext.granularity, 2048);
+        assert_eq!(fulltext.false_positive_rate(), 0.02);
+    }
+
+    #[test]
+    fn test_transform_rejects_invalid_index_options_type() {
+        let result = parse_transform(
+            r#"
+field: message
+type: string
+index:
+  type: fulltext
+  options: invalid
+"#,
+        );
+
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_transform_rejects_non_scalar_index_option_value() {
+        let result = parse_transform(
+            r#"
+field: message
+type: string
+index:
+  type: fulltext
+  options:
+    analyzer:
+      kind: English
+"#,
+        );
+
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_transform_rejects_unknown_index_field() {
+        let result = parse_transform(
+            r#"
+field: message
+type: string
+index:
+  type: fulltext
+  config: {}
+"#,
+        );
+
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_transform_rejects_unsupported_fulltext_option_key() {
+        let result = parse_transform(
+            r#"
+field: message
+type: string
+index:
+  type: fulltext
+  options:
+    tokenizer: english
+"#,
+        );
+
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_transform_rejects_options_for_inverted_index() {
+        let result = parse_transform(
+            r#"
+field: message
+type: string
+index:
+  type: inverted
+  options:
+    backend: bloom
+"#,
+        );
+
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_transform_rejects_empty_options_for_unsupported_indexes() {
+        for index in ["inverted", "time", "tag"] {
+            let yaml = format!(
+                r#"
+field: message
+type: string
+index:
+  type: {index}
+  options: {{}}
+"#
+            );
+
+            let result = parse_transform(&yaml);
+            assert!(
+                result.is_err(),
+                "expected `{index}` to reject empty options"
+            );
+        }
+    }
+
+    #[test]
+    fn test_transform_rejects_fulltext_index_on_non_string_column() {
+        let result = parse_transform(
+            r#"
+field: count
+type: int64
+index: fulltext
+"#,
+        );
+
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_transform_allows_skipping_index_on_numeric_column() {
+        let transform = parse_transform(
+            r#"
+field: count
+type: int64
+index:
+  type: skipping
+  options:
+    granularity: 2048
+    false_positive_rate: 0.02
+    type: BLOOM
+"#,
+        )
+        .unwrap();
+
+        assert_eq!(transform.index, Some(Index::Skipping));
+        let skipping = transform
+            .index_options
+            .as_ref()
+            .unwrap()
+            .as_skipping()
+            .unwrap();
+        assert_eq!(skipping.granularity, 2048);
+        assert_eq!(skipping.false_positive_rate(), 0.02);
+        assert_eq!(skipping.index_type.to_string(), "BLOOM");
+    }
+}
diff --git a/src/pipeline/src/etl/transform/transformer/greptime.rs b/src/pipeline/src/etl/transform/transformer/greptime.rs
index 92b9dee4b9..a1941aca05 100644
--- a/src/pipeline/src/etl/transform/transformer/greptime.rs
+++ b/src/pipeline/src/etl/transform/transformer/greptime.rs
@@ -151,6 +151,7 @@ impl GreptimeTransformer {
             type_,
             default,
             index: Some(Index::Time),
+            index_options: None,
             on_failure: Some(crate::etl::transform::OnFailure::Default),
             tag: false,
         };
@@ -418,7 +419,7 @@ fn resolve_schema(
         match (column_type, value_type) {
             (column_type, value_type) if column_type == value_type => Ok(()),
             (ConcreteDataType::Json(column_type), ConcreteDataType::Json(value_type))
-                if column_type.is_include(value_type) =>
+                if column_type.is_json2() && value_type.is_json2() =>
             {
                 Ok(())
             }
@@ -689,17 +690,16 @@ fn resolve_value(
         }
 
         VrlValue::Array(_) | VrlValue::Object(_) => {
-            let is_json_native_type = schema_info
+            let is_json2 = schema_info
                 .find_column_schema_in_table(&column_name)
                 .is_some_and(|x| {
-                    if let ConcreteDataType::Json(column_type) = &x.column_schema.data_type {
-                        column_type.is_native_type()
-                    } else {
-                        false
-                    }
+                    matches!(
+                        &x.column_schema.data_type,
+                        ConcreteDataType::Json(column_type) if column_type.is_json2()
+                    )
                 });
 
-            let value = if is_json_native_type {
+            let value = if is_json2 {
                 let json_extension_type: Option<JsonExtensionType> =
                     if let Some(x) = schema_info.find_column_schema_in_table(&column_name) {
                         x.column_schema.extension_type()?
diff --git a/src/pipeline/src/etl/transform/transformer/greptime/coerce.rs b/src/pipeline/src/etl/transform/transformer/greptime/coerce.rs
index a5f283881f..d92898f384 100644
--- a/src/pipeline/src/etl/transform/transformer/greptime/coerce.rs
+++ b/src/pipeline/src/etl/transform/transformer/greptime/coerce.rs
@@ -18,17 +18,18 @@ use api::v1::{ColumnDataTypeExtension, ColumnOptions, JsonTypeExtension};
 use datatypes::schema::{FulltextOptions, SkippingIndexOptions};
 use greptime_proto::v1::value::ValueData;
 use greptime_proto::v1::{ColumnDataType, ColumnSchema, SemanticType};
-use snafu::{OptionExt, ResultExt};
+use snafu::{OptionExt, ResultExt, ensure};
 use vrl::value::Value as VrlValue;
 
 use crate::error::{
     CoerceIncompatibleTypesSnafu, CoerceJsonTypeToSnafu, CoerceStringToTypeSnafu,
     CoerceTypeToJsonSnafu, CoerceUnsupportedEpochTypeSnafu, ColumnOptionsSnafu,
-    InvalidTimestampSnafu, Result, UnsupportedTypeInPipelineSnafu, VrlRegexValueSnafu,
+    InvalidTimestampSnafu, Result, TransformIndexStateMismatchSnafu,
+    UnsupportedTypeInPipelineSnafu, VrlRegexValueSnafu,
 };
 use crate::etl::transform::index::Index;
 use crate::etl::transform::transformer::greptime::vrl_value_to_jsonb_value;
-use crate::etl::transform::{OnFailure, Transform};
+use crate::etl::transform::{OnFailure, Transform, TransformIndexOptions};
 
 pub(crate) fn coerce_columns(transform: &Transform) -> Result<Vec<ColumnSchema>> {
     let mut columns = Vec::new();
@@ -73,15 +74,68 @@ fn coerce_semantic_type(transform: &Transform) -> SemanticType {
     }
 }
 
-fn coerce_options(transform: &Transform) -> Result<Option<ColumnOptions>> {
-    match transform.index {
-        Some(Index::Fulltext) => options_from_fulltext(&FulltextOptions {
+fn transform_index_label(index: Option<Index>) -> String {
+    index
+        .map(|index| index.to_string())
+        .unwrap_or_else(|| "none".to_string())
+}
+
+fn validate_transform_index_state(transform: &Transform) -> Result<()> {
+    let Some(index_options) = transform.index_options.as_ref() else {
+        return Ok(());
+    };
+
+    let options_index = index_options.index();
+    let index = transform.index;
+    ensure!(
+        index == Some(options_index),
+        TransformIndexStateMismatchSnafu {
+            index: transform_index_label(index),
+            options: options_index.to_string(),
+        }
+    );
+
+    Ok(())
+}
+
+fn build_fulltext_index_options(transform: &Transform) -> Result<FulltextOptions> {
+    match transform.index_options.as_ref() {
+        None => Ok(FulltextOptions {
             enable: true,
             ..Default::default()
-        })
-        .context(ColumnOptionsSnafu),
+        }),
+        Some(TransformIndexOptions::Fulltext(options)) => Ok(options.clone()),
+        Some(options) => TransformIndexStateMismatchSnafu {
+            index: Index::Fulltext.to_string(),
+            options: options.index().to_string(),
+        }
+        .fail(),
+    }
+}
+
+fn build_skipping_index_options(transform: &Transform) -> Result<SkippingIndexOptions> {
+    match transform.index_options.as_ref() {
+        None => Ok(SkippingIndexOptions::default()),
+        Some(TransformIndexOptions::Skipping(options)) => Ok(options.clone()),
+        Some(options) => TransformIndexStateMismatchSnafu {
+            index: Index::Skipping.to_string(),
+            options: options.index().to_string(),
+        }
+        .fail(),
+    }
+}
+
+fn coerce_options(transform: &Transform) -> Result<Option<ColumnOptions>> {
+    validate_transform_index_state(transform)?;
+
+    match transform.index {
+        Some(Index::Fulltext) => {
+            let options = build_fulltext_index_options(transform)?;
+            options_from_fulltext(&options).context(ColumnOptionsSnafu)
+        }
         Some(Index::Skipping) => {
-            options_from_skipping(&SkippingIndexOptions::default()).context(ColumnOptionsSnafu)
+            let options = build_skipping_index_options(transform)?;
+            options_from_skipping(&options).context(ColumnOptionsSnafu)
         }
         Some(Index::Inverted) => Ok(Some(options_from_inverted())),
         _ => Ok(None),
@@ -382,6 +436,7 @@ fn coerce_json_value(v: &VrlValue, transform: &Transform) -> Result<Option<Value
 #[cfg(test)]
 mod tests {
 
+    use datatypes::schema::{FulltextAnalyzer, FulltextBackend, SkippingIndexType};
     use vrl::prelude::Bytes;
 
     use super::*;
@@ -394,6 +449,7 @@ mod tests {
             type_: ColumnDataType::Int32,
             default: None,
             index: None,
+            index_options: None,
             on_failure: None,
             tag: false,
         };
@@ -420,6 +476,7 @@ mod tests {
             type_: ColumnDataType::Int32,
             default: None,
             index: None,
+            index_options: None,
             on_failure: Some(OnFailure::Ignore),
             tag: false,
         };
@@ -436,6 +493,7 @@ mod tests {
             type_: ColumnDataType::Int32,
             default: None,
             index: None,
+            index_options: None,
             on_failure: Some(OnFailure::Default),
             tag: false,
         };
@@ -455,4 +513,99 @@ mod tests {
             assert_eq!(result, Some(ValueData::I32Value(42)));
         }
     }
+
+    #[test]
+    fn test_coerce_fulltext_options_with_custom_values() {
+        let transform = Transform {
+            fields: Fields::default(),
+            type_: ColumnDataType::String,
+            default: None,
+            index: Some(Index::Fulltext),
+            index_options: Some(TransformIndexOptions::Fulltext(
+                FulltextOptions::new_unchecked(
+                    true,
+                    FulltextAnalyzer::Chinese,
+                    true,
+                    FulltextBackend::Tantivy,
+                    10240,
+                    0.01,
+                ),
+            )),
+            on_failure: None,
+            tag: false,
+        };
+
+        let options = coerce_options(&transform).unwrap().unwrap();
+        let fulltext: FulltextOptions =
+            serde_json::from_str(options.options.get("fulltext").unwrap()).unwrap();
+
+        assert!(fulltext.enable);
+        assert_eq!(fulltext.analyzer.to_string(), "Chinese");
+        assert!(fulltext.case_sensitive);
+        assert_eq!(fulltext.backend.to_string(), "tantivy");
+    }
+
+    #[test]
+    fn test_coerce_skipping_options_with_custom_values() {
+        let transform = Transform {
+            fields: Fields::default(),
+            type_: ColumnDataType::Int64,
+            default: None,
+            index: Some(Index::Skipping),
+            index_options: Some(TransformIndexOptions::Skipping(
+                SkippingIndexOptions::new_unchecked(2048, 0.02, SkippingIndexType::BloomFilter),
+            )),
+            on_failure: None,
+            tag: false,
+        };
+
+        let options = coerce_options(&transform).unwrap().unwrap();
+        let skipping: SkippingIndexOptions =
+            serde_json::from_str(options.options.get("skipping_index").unwrap()).unwrap();
+
+        assert_eq!(skipping.granularity, 2048);
+        assert_eq!(skipping.false_positive_rate(), 0.02);
+        assert_eq!(skipping.index_type.to_string(), "BLOOM");
+    }
+
+    #[test]
+    fn test_coerce_rejects_mismatched_index_options() {
+        let transform = Transform {
+            fields: Fields::default(),
+            type_: ColumnDataType::String,
+            default: None,
+            index: Some(Index::Fulltext),
+            index_options: Some(TransformIndexOptions::Skipping(
+                SkippingIndexOptions::new_unchecked(2048, 0.02, SkippingIndexType::BloomFilter),
+            )),
+            on_failure: None,
+            tag: false,
+        };
+
+        assert!(coerce_options(&transform).is_err());
+    }
+
+    #[test]
+    fn test_coerce_rejects_index_options_without_index() {
+        let transform = Transform {
+            fields: Fields::default(),
+            type_: ColumnDataType::String,
+            default: None,
+            index: None,
+            index_options: Some(TransformIndexOptions::Fulltext(
+                FulltextOptions::new_unchecked(
+                    true,
+                    FulltextAnalyzer::Chinese,
+                    true,
+                    FulltextBackend::Tantivy,
+                    10240,
+                    0.01,
+                ),
+            )),
+            on_failure: None,
+            tag: false,
+        };
+
+        assert!(coerce_options(&transform).is_err());
+    }
 }
diff --git a/src/promql/Cargo.toml b/src/promql/Cargo.toml
index f93cb8beb9..460be8ddd9 100644
--- a/src/promql/Cargo.toml
+++ b/src/promql/Cargo.toml
@@ -13,7 +13,6 @@ async-trait.workspace = true
 bytemuck.workspace = true
 common-error.workspace = true
 common-macro.workspace = true
-common-recordbatch.workspace = true
 common-telemetry.workspace = true
 datafusion.workspace = true
 datafusion-common.workspace = true
@@ -27,4 +26,13 @@ prost.workspace = true
 snafu.workspace = true
 
 [dev-dependencies]
+common-recordbatch.workspace = true
+criterion.workspace = true
+datafusion-common.workspace = true
+datafusion-expr.workspace = true
+datatypes.workspace = true
 tokio.workspace = true
+
+[[bench]]
+name = "bench_main"
+harness = false
diff --git a/src/promql/benches/bench_main.rs b/src/promql/benches/bench_main.rs
new file mode 100644
index 0000000000..2d93887041
--- /dev/null
+++ b/src/promql/benches/bench_main.rs
@@ -0,0 +1,21 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use criterion::criterion_main;
+
+mod bench_range_fn;
+
+criterion_main! {
+    bench_range_fn::benches
+}
diff --git a/src/promql/benches/bench_range_fn.rs b/src/promql/benches/bench_range_fn.rs
new file mode 100644
index 0000000000..840956b942
--- /dev/null
+++ b/src/promql/benches/bench_range_fn.rs
@@ -0,0 +1,355 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Benchmarks for PromQL range functions.
+
+use std::sync::Arc;
+
+use criterion::{BenchmarkId, Criterion, criterion_group};
+use datafusion::arrow::array::{Float64Array, TimestampMillisecondArray};
+use datafusion::physical_plan::ColumnarValue;
+use datafusion_common::ScalarValue;
+use datafusion_common::config::ConfigOptions;
+use datafusion_expr::ScalarFunctionArgs;
+use datatypes::arrow::datatypes::{DataType, Field};
+use promql::functions::{Delta, IDelta, Increase, PredictLinear, QuantileOverTime, Rate};
+use promql::range_array::RangeArray;
+
+fn build_sliding_ranges(
+    num_points: usize,
+    window_size: u32,
+    values: Vec<f64>,
+    eval_offset_ms: i64,
+) -> (RangeArray, RangeArray, Arc<TimestampMillisecondArray>) {
+    let step_ms = 1000i64;
+    let timestamps: Vec<i64> = (0..num_points as i64).map(|i| (i + 1) * step_ms).collect();
+
+    let ts_array = Arc::new(TimestampMillisecondArray::from(timestamps.clone()));
+    let val_array = Arc::new(Float64Array::from(values));
+
+    let num_windows = if num_points >= window_size as usize {
+        num_points - window_size as usize + 1
+    } else {
+        0
+    };
+
+    let ranges: Vec<(u32, u32)> = (0..num_windows).map(|i| (i as u32, window_size)).collect();
+
+    let eval_ts: Vec<i64> = (0..num_windows)
+        .map(|i| timestamps[i + window_size as usize - 1] + eval_offset_ms)
+        .collect();
+    let eval_ts_array = Arc::new(TimestampMillisecondArray::from(eval_ts));
+
+    let ts_range = RangeArray::from_ranges(ts_array, ranges.clone()).unwrap();
+    let val_range = RangeArray::from_ranges(val_array, ranges).unwrap();
+
+    (ts_range, val_range, eval_ts_array)
+}
+
+fn build_monotonic_counter_values(num_points: usize) -> Vec<f64> {
+    let mut current = 0.0;
+    (0..num_points)
+        .map(|i| {
+            current += 1.0 + (i % 7) as f64 * 0.25;
+            current
+        })
+        .collect()
+}
+
+fn build_resetting_counter_values(num_points: usize) -> Vec<f64> {
+    let mut current = 0.0;
+    (0..num_points)
+        .map(|i| {
+            if i > 0 && i % 37 == 0 {
+                current = 1.0;
+            } else {
+                current += 1.0 + (i % 5) as f64 * 0.5;
+            }
+            current
+        })
+        .collect()
+}
+
+fn build_gauge_values(num_points: usize) -> Vec<f64> {
+    (0..num_points)
+        .map(|i| ((i % 29) as f64 - 14.0) * 1.25 + (i % 3) as f64 * 0.1)
+        .collect()
+}
+
+fn build_default_values(num_points: usize) -> Vec<f64> {
+    (0..num_points).map(|i| i as f64 * 1.5 + 0.1).collect()
+}
+
+fn make_extrapolated_rate_input(
+    num_points: usize,
+    window_size: u32,
+    values: Vec<f64>,
+    eval_offset_ms: i64,
+) -> Vec<ColumnarValue> {
+    let (ts_range, val_range, eval_ts) =
+        build_sliding_ranges(num_points, window_size, values, eval_offset_ms);
+    let range_length = window_size as i64 * 1000;
+    vec![
+        ColumnarValue::Array(Arc::new(ts_range.into_dict())),
+        ColumnarValue::Array(Arc::new(val_range.into_dict())),
+        ColumnarValue::Array(eval_ts),
+        ColumnarValue::Scalar(ScalarValue::Int64(Some(range_length))),
+    ]
+}
+
+fn make_idelta_input(num_points: usize, window_size: u32) -> Vec<ColumnarValue> {
+    let (ts_range, val_range, _) =
+        build_sliding_ranges(num_points, window_size, build_default_values(num_points), 0);
+    vec![
+        ColumnarValue::Array(Arc::new(ts_range.into_dict())),
+        ColumnarValue::Array(Arc::new(val_range.into_dict())),
+    ]
+}
+
+fn make_quantile_input(num_points: usize, window_size: u32) -> Vec<ColumnarValue> {
+    let (ts_range, val_range, _) =
+        build_sliding_ranges(num_points, window_size, build_default_values(num_points), 0);
+    vec![
+        ColumnarValue::Array(Arc::new(ts_range.into_dict())),
+        ColumnarValue::Array(Arc::new(val_range.into_dict())),
+        ColumnarValue::Scalar(ScalarValue::Float64(Some(0.9))),
+    ]
+}
+
+fn make_predict_linear_input(num_points: usize, window_size: u32) -> Vec<ColumnarValue> {
+    let (ts_range, val_range, _) =
+        build_sliding_ranges(num_points, window_size, build_default_values(num_points), 0);
+    vec![
+        ColumnarValue::Array(Arc::new(ts_range.into_dict())),
+        ColumnarValue::Array(Arc::new(val_range.into_dict())),
+        // predict 60s into the future
+        ColumnarValue::Scalar(ScalarValue::Int64(Some(60))),
+    ]
+}
+
+struct PreparedUdfCall {
+    args: Vec<ColumnarValue>,
+    arg_fields: Vec<Arc<Field>>,
+    number_rows: usize,
+    return_field: Arc<Field>,
+    config_options: Arc<ConfigOptions>,
+}
+
+impl PreparedUdfCall {
+    fn new(args: Vec<ColumnarValue>) -> Self {
+        let arg_fields = args
+            .iter()
+            .enumerate()
+            .map(|(i, c)| Arc::new(Field::new(format!("c{i}"), c.data_type(), true)))
+            .collect();
+        let number_rows = args
+            .iter()
+            .find_map(|c| match c {
+                ColumnarValue::Array(a) => Some(a.len()),
+                _ => None,
+            })
+            .unwrap_or(1);
+        Self {
+            args,
+            arg_fields,
+            number_rows,
+            return_field: Arc::new(Field::new("out", DataType::Float64, true)),
+            config_options: Arc::new(ConfigOptions::default()),
+        }
+    }
+}
+
+fn invoke_prepared(udf: &datafusion::logical_expr::ScalarUDF, prepared: &PreparedUdfCall) {
+    udf.invoke_with_args(ScalarFunctionArgs {
+        args: prepared.args.clone(),
+        arg_fields: prepared.arg_fields.clone(),
+        number_rows: prepared.number_rows,
+        return_field: prepared.return_field.clone(),
+        config_options: prepared.config_options.clone(),
+    })
+    .unwrap();
+}
+
+fn bench_range_functions(c: &mut Criterion) {
+    let mut group = c.benchmark_group("range_fn");
+
+    // Benchmark parameters: (total_points, window_size)
+    let params: &[(usize, u32)] = &[
+        (1_000, 10),   // small series, small window
+        (10_000, 10),  // large series, small window
+        (10_000, 60),  // large series, typical 1-min window at 1s step
+        (10_000, 360), // large series, wide 6-min window
+    ];
+
+    // --- rate (monotonic counter) ---
+    let rate_udf = Rate::scalar_udf();
+    for &(n, w) in params {
+        let prepared = PreparedUdfCall::new(make_extrapolated_rate_input(
+            n,
+            w,
+            build_monotonic_counter_values(n),
+            500,
+        ));
+        group.bench_with_input(
+            BenchmarkId::new("rate_counter", format!("n{n}_w{w}")),
+            &(n, w),
+            |b, _| b.iter(|| invoke_prepared(&rate_udf, &prepared)),
+        );
+    }
+
+    // --- rate (periodic resets) ---
+    for &(n, w) in params {
+        let prepared = PreparedUdfCall::new(make_extrapolated_rate_input(
+            n,
+            w,
+            build_resetting_counter_values(n),
+            500,
+        ));
+        group.bench_with_input(
+            BenchmarkId::new("rate_counter_reset", format!("n{n}_w{w}")),
+            &(n, w),
+            |b, _| b.iter(|| invoke_prepared(&rate_udf, &prepared)),
+        );
+    }
+
+    // --- increase (monotonic counter) ---
+    let increase_udf = Increase::scalar_udf();
+    for &(n, w) in params {
+        let prepared = PreparedUdfCall::new(make_extrapolated_rate_input(
+            n,
+            w,
+            build_monotonic_counter_values(n),
+            500,
+        ));
+        group.bench_with_input(
+            BenchmarkId::new("increase_counter", format!("n{n}_w{w}")),
+            &(n, w),
+            |b, _| b.iter(|| invoke_prepared(&increase_udf, &prepared)),
+        );
+    }
+
+    // --- increase (periodic resets) ---
+    for &(n, w) in params {
+        let prepared = PreparedUdfCall::new(make_extrapolated_rate_input(
+            n,
+            w,
+            build_resetting_counter_values(n),
+            500,
+        ));
+        group.bench_with_input(
+            BenchmarkId::new("increase_counter_reset", format!("n{n}_w{w}")),
+            &(n, w),
+            |b, _| b.iter(|| invoke_prepared(&increase_udf, &prepared)),
+        );
+    }
+
+    // --- delta (gauge) ---
+    let delta_udf = Delta::scalar_udf();
+    for &(n, w) in params {
+        let prepared = PreparedUdfCall::new(make_extrapolated_rate_input(
+            n,
+            w,
+            build_gauge_values(n),
+            500,
+        ));
+        group.bench_with_input(
+            BenchmarkId::new("delta_gauge", format!("n{n}_w{w}")),
+            &(n, w),
+            |b, _| b.iter(|| invoke_prepared(&delta_udf, &prepared)),
+        );
+    }
+
+    // --- idelta ---
+    let idelta_udf = IDelta::<false>::scalar_udf();
+    for &(n, w) in params {
+        let prepared = PreparedUdfCall::new(make_idelta_input(n, w));
+        group.bench_with_input(
+            BenchmarkId::new("idelta", format!("n{n}_w{w}")),
+            &(n, w),
+            |b, _| b.iter(|| invoke_prepared(&idelta_udf, &prepared)),
+        );
+    }
+
+    // --- irate ---
+    let irate_udf = IDelta::<true>::scalar_udf();
+    for &(n, w) in params {
+        let prepared = PreparedUdfCall::new(make_idelta_input(n, w));
+        group.bench_with_input(
+            BenchmarkId::new("irate", format!("n{n}_w{w}")),
+            &(n, w),
+            |b, _| b.iter(|| invoke_prepared(&irate_udf, &prepared)),
+        );
+    }
+
+    // --- quantile_over_time ---
+    let quantile_udf = QuantileOverTime::scalar_udf();
+    for &(n, w) in params {
+        let prepared = PreparedUdfCall::new(make_quantile_input(n, w));
+        group.bench_with_input(
+            BenchmarkId::new("quantile_over_time", format!("n{n}_w{w}")),
+            &(n, w),
+            |b, _| b.iter(|| invoke_prepared(&quantile_udf, &prepared)),
+        );
+    }
+
+    // --- predict_linear ---
+    let predict_udf = PredictLinear::scalar_udf();
+    for &(n, w) in params {
+        let prepared = PreparedUdfCall::new(make_predict_linear_input(n, w));
+        group.bench_with_input(
+            BenchmarkId::new("predict_linear", format!("n{n}_w{w}")),
+            &(n, w),
+            |b, _| b.iter(|| invoke_prepared(&predict_udf, &prepared)),
+        );
+    }
+
+    // --- RangeArray: get vs get_offset_length micro-benchmark ---
+    // Isolates the overhead of array slicing vs offset/length lookup
+    for &(n, w) in params {
+        let step_ms = 1000i64;
+        let timestamps: Vec<i64> = (0..n as i64).map(|i| (i + 1) * step_ms).collect();
+        let ts_array = Arc::new(TimestampMillisecondArray::from(timestamps));
+        let num_windows = n - w as usize + 1;
+        let ranges: Vec<(u32, u32)> = (0..num_windows).map(|i| (i as u32, w)).collect();
+        let range_array = RangeArray::from_ranges(ts_array, ranges).unwrap();
+
+        group.bench_with_input(
+            BenchmarkId::new("range_array_get", format!("n{n}_w{w}")),
+            &(),
+            |b, _| {
+                b.iter(|| {
+                    for i in 0..range_array.len() {
+                        std::hint::black_box(range_array.get(i));
+                    }
+                })
+            },
+        );
+
+        group.bench_with_input(
+            BenchmarkId::new("range_array_get_offset_length", format!("n{n}_w{w}")),
+            &(),
+            |b, _| {
+                b.iter(|| {
+                    for i in 0..range_array.len() {
+                        std::hint::black_box(range_array.get_offset_length(i));
+                    }
+                })
+            },
+        );
+    }
+
+    group.finish();
+}
+
+criterion_group!(benches, bench_range_functions);
diff --git a/src/promql/src/extension_plan/absent.rs b/src/promql/src/extension_plan/absent.rs
index db31a3d901..71af413029 100644
--- a/src/promql/src/extension_plan/absent.rs
+++ b/src/promql/src/extension_plan/absent.rs
@@ -49,9 +49,6 @@ use snafu::ResultExt;
 use crate::error::DeserializeSnafu;
 use crate::extension_plan::{Millisecond, resolve_column_name, serialize_column_index};
 
-/// Maximum number of rows per output batch
-const ABSENT_BATCH_SIZE: usize = 8192;
-
 #[derive(Debug, PartialEq, Eq, Hash)]
 pub struct Absent {
     start: Millisecond,
@@ -390,11 +387,13 @@ impl ExecutionPlan for AbsentExec {
         context: Arc<TaskContext>,
     ) -> DataFusionResult<SendableRecordBatchStream> {
         let baseline_metric = BaselineMetrics::new(&self.metric, partition);
+        let batch_size = context.session_config().batch_size();
         let input = self.input.execute(partition, context)?;
 
         Ok(Box::pin(AbsentStream {
             end: self.end,
             step: self.step,
+            batch_size,
             time_index_column_index: self
                 .input
                 .schema()
@@ -407,6 +406,8 @@ impl ExecutionPlan for AbsentExec {
             metric: baseline_metric,
             // Buffer for streaming output timestamps
             output_timestamps: Vec::new(),
+            input_timestamps: Vec::new(),
+            input_timestamp_offset: 0,
             // Current timestamp in the output range
             output_ts_cursor: self.start,
             input_finished: false,
@@ -441,6 +442,7 @@ impl DisplayAs for AbsentExec {
 pub struct AbsentStream {
     end: Millisecond,
     step: Millisecond,
+    batch_size: usize,
     time_index_column_index: usize,
     output_schema: SchemaRef,
     fake_labels: Vec<(String, String)>,
@@ -448,6 +450,9 @@ pub struct AbsentStream {
     metric: BaselineMetrics,
     // Buffer for streaming output timestamps
     output_timestamps: Vec<Millisecond>,
+    // Current input timestamps being processed incrementally.
+    input_timestamps: Vec<Millisecond>,
+    input_timestamp_offset: usize,
     // Current timestamp in the output range
     output_ts_cursor: Millisecond,
     input_finished: bool,
@@ -464,52 +469,53 @@ impl Stream for AbsentStream {
 
     fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
         loop {
-            if !self.input_finished {
-                match ready!(self.input.poll_next_unpin(cx)) {
-                    Some(Ok(batch)) => {
-                        let timer = std::time::Instant::now();
-                        if let Err(e) = self.process_input_batch(&batch) {
-                            return Poll::Ready(Some(Err(e)));
-                        }
-                        self.metric.elapsed_compute().add_elapsed(timer);
-
-                        // If we have enough data for a batch, output it
-                        if self.output_timestamps.len() >= ABSENT_BATCH_SIZE {
-                            let timer = std::time::Instant::now();
-                            let result = self.flush_output_batch();
-                            self.metric.elapsed_compute().add_elapsed(timer);
-
-                            match result {
-                                Ok(Some(batch)) => return Poll::Ready(Some(Ok(batch))),
-                                Ok(None) => continue,
-                                Err(e) => return Poll::Ready(Some(Err(e))),
-                            }
-                        }
-                    }
-                    Some(Err(e)) => return Poll::Ready(Some(Err(e))),
-                    None => {
-                        self.input_finished = true;
-
-                        let timer = std::time::Instant::now();
-                        // Process any remaining absent timestamps
-                        if let Err(e) = self.process_remaining_absent_timestamps() {
-                            return Poll::Ready(Some(Err(e)));
-                        }
-                        let result = self.flush_output_batch();
-                        self.metric.elapsed_compute().add_elapsed(timer);
-                        return Poll::Ready(result.transpose());
-                    }
+            if self.has_pending_input_timestamps() {
+                let timer = std::time::Instant::now();
+                if let Err(e) = self.process_input_batch() {
+                    return Poll::Ready(Some(Err(e)));
+                }
+                self.metric.elapsed_compute().add_elapsed(timer);
+
+                match self.flush_output_batch() {
+                    Ok(Some(batch)) => return Poll::Ready(Some(Ok(batch))),
+                    Ok(None) => continue,
+                    Err(e) => return Poll::Ready(Some(Err(e))),
+                }
+            }
+
+            if self.input_finished {
+                let timer = std::time::Instant::now();
+                if let Err(e) = self.process_remaining_absent_timestamps() {
+                    return Poll::Ready(Some(Err(e)));
+                }
+                self.metric.elapsed_compute().add_elapsed(timer);
+
+                match self.flush_output_batch() {
+                    Ok(Some(batch)) => return Poll::Ready(Some(Ok(batch))),
+                    Ok(None) => return Poll::Ready(None),
+                    Err(e) => return Poll::Ready(Some(Err(e))),
+                }
+            }
+
+            match ready!(self.input.poll_next_unpin(cx)) {
+                Some(Ok(batch)) => {
+                    let timer = std::time::Instant::now();
+                    if let Err(e) = self.buffer_input_timestamps(&batch) {
+                        return Poll::Ready(Some(Err(e)));
+                    }
+                    self.metric.elapsed_compute().add_elapsed(timer);
+                }
+                Some(Err(e)) => return Poll::Ready(Some(Err(e))),
+                None => {
+                    self.input_finished = true;
                 }
-            } else {
-                return Poll::Ready(None);
             }
         }
     }
 }
 
 impl AbsentStream {
-    fn process_input_batch(&mut self, batch: &RecordBatch) -> DataFusionResult<()> {
-        // Extract timestamps from this batch
+    fn buffer_input_timestamps(&mut self, batch: &RecordBatch) -> DataFusionResult<()> {
         let timestamp_array = batch.column(self.time_index_column_index);
         let milli_ts_array = arrow::compute::cast(
             timestamp_array,
@@ -519,29 +525,52 @@ impl AbsentStream {
             .as_any()
             .downcast_ref::<TimestampMillisecondArray>()
             .unwrap();
+        self.input_timestamps.clear();
+        self.input_timestamps
+            .extend_from_slice(timestamp_array.values());
+        self.input_timestamp_offset = 0;
+        Ok(())
+    }
+
+    fn has_pending_input_timestamps(&self) -> bool {
+        self.input_timestamp_offset < self.input_timestamps.len()
+    }
+
+    fn process_input_batch(&mut self) -> DataFusionResult<()> {
+        while self.input_timestamp_offset < self.input_timestamps.len() {
+            let input_ts = self.input_timestamps[self.input_timestamp_offset];
 
-        // Process against current output cursor position
-        for &input_ts in timestamp_array.values() {
             // Generate absent timestamps up to this input timestamp
             while self.output_ts_cursor < input_ts && self.output_ts_cursor <= self.end {
                 self.output_timestamps.push(self.output_ts_cursor);
                 self.output_ts_cursor += self.step;
+
+                if self.output_timestamps.len() >= self.batch_size {
+                    return Ok(());
+                }
             }
 
             // Skip the input timestamp if it matches our cursor
             if self.output_ts_cursor == input_ts {
                 self.output_ts_cursor += self.step;
             }
+
+            self.input_timestamp_offset += 1;
         }
 
+        self.input_timestamps.clear();
+        self.input_timestamp_offset = 0;
         Ok(())
     }
 
     fn process_remaining_absent_timestamps(&mut self) -> DataFusionResult<()> {
-        // Generate all remaining absent timestamps (input is finished)
         while self.output_ts_cursor <= self.end {
             self.output_timestamps.push(self.output_ts_cursor);
             self.output_ts_cursor += self.step;
+
+            if self.output_timestamps.len() >= self.batch_size {
+                return Ok(());
+            }
         }
         Ok(())
     }
@@ -551,11 +580,16 @@ impl AbsentStream {
             return Ok(None);
         }
 
+        let timestamps = if self.output_timestamps.len() <= self.batch_size {
+            std::mem::take(&mut self.output_timestamps)
+        } else {
+            let remaining = self.output_timestamps.split_off(self.batch_size);
+            std::mem::replace(&mut self.output_timestamps, remaining)
+        };
+
         let mut columns: Vec<ArrayRef> = Vec::with_capacity(self.output_schema.fields().len());
-        let num_rows = self.output_timestamps.len();
-        columns.push(Arc::new(TimestampMillisecondArray::from(
-            self.output_timestamps.clone(),
-        )) as _);
+        let num_rows = timestamps.len();
+        columns.push(Arc::new(TimestampMillisecondArray::from(timestamps)) as _);
         columns.push(Arc::new(Float64Array::from(vec![1.0; num_rows])) as _);
 
         for (_, value) in self.fake_labels.iter() {
@@ -567,7 +601,6 @@ impl AbsentStream {
 
         let batch = RecordBatch::try_new(self.output_schema.clone(), columns)?;
 
-        self.output_timestamps.clear();
         Ok(Some(batch))
     }
 }
@@ -580,7 +613,7 @@ mod tests {
     use datafusion::arrow::record_batch::RecordBatch;
     use datafusion::catalog::memory::DataSourceExec;
     use datafusion::datasource::memory::MemorySourceConfig;
-    use datafusion::prelude::SessionContext;
+    use datafusion::prelude::{SessionConfig, SessionContext};
     use datatypes::arrow::array::{Float64Array, TimestampMillisecondArray};
 
     use super::*;
@@ -725,4 +758,146 @@ mod tests {
         // Should output all timestamps in range: 0, 1000, 2000
         assert_eq!(output_timestamps, vec![0, 1000, 2000]);
     }
+
+    #[tokio::test]
+    async fn test_absent_respects_session_batch_size_for_large_gap() {
+        let schema = Arc::new(Schema::new(vec![
+            Field::new(
+                "timestamp",
+                DataType::Timestamp(TimeUnit::Millisecond, None),
+                true,
+            ),
+            Field::new("value", DataType::Float64, true),
+        ]));
+
+        let timestamp_array = Arc::new(TimestampMillisecondArray::from(vec![9]));
+        let value_array = Arc::new(Float64Array::from(vec![1.0]));
+        let batch =
+            RecordBatch::try_new(schema.clone(), vec![timestamp_array, value_array]).unwrap();
+
+        let memory_exec = DataSourceExec::new(Arc::new(
+            MemorySourceConfig::try_new(&[vec![batch]], schema, None).unwrap(),
+        ));
+
+        let output_schema = Arc::new(Schema::new(vec![
+            Field::new(
+                "timestamp",
+                DataType::Timestamp(TimeUnit::Millisecond, None),
+                true,
+            ),
+            Field::new("value", DataType::Float64, true),
+        ]));
+
+        let absent_exec = AbsentExec {
+            start: 0,
+            end: 10,
+            step: 1,
+            time_index_column: "timestamp".to_string(),
+            value_column: "value".to_string(),
+            fake_labels: vec![],
+            output_schema: output_schema.clone(),
+            input: Arc::new(memory_exec),
+            properties: Arc::new(PlanProperties::new(
+                EquivalenceProperties::new(output_schema.clone()),
+                Partitioning::UnknownPartitioning(1),
+                EmissionType::Incremental,
+                Boundedness::Bounded,
+            )),
+            metric: ExecutionPlanMetricsSet::new(),
+        };
+
+        let session_ctx = SessionContext::new_with_config(SessionConfig::new().with_batch_size(3));
+        let task_ctx = session_ctx.task_ctx();
+        let mut stream = absent_exec.execute(0, task_ctx).unwrap();
+
+        let mut batch_sizes = Vec::new();
+        let mut output_timestamps = Vec::new();
+        while let Some(batch_result) = stream.next().await {
+            let batch = batch_result.unwrap();
+            batch_sizes.push(batch.num_rows());
+
+            let ts_array = batch
+                .column(0)
+                .as_any()
+                .downcast_ref::<TimestampMillisecondArray>()
+                .unwrap();
+            for i in 0..ts_array.len() {
+                if !ts_array.is_null(i) {
+                    output_timestamps.push(ts_array.value(i));
+                }
+            }
+        }
+
+        assert_eq!(batch_sizes, vec![3, 3, 3, 1]);
+        assert_eq!(output_timestamps, vec![0, 1, 2, 3, 4, 5, 6, 7, 8, 10]);
+    }
+
+    #[tokio::test]
+    async fn test_absent_resumes_same_input_timestamp_after_batch_flush() {
+        let schema = Arc::new(Schema::new(vec![
+            Field::new(
+                "timestamp",
+                DataType::Timestamp(TimeUnit::Millisecond, None),
+                true,
+            ),
+            Field::new("value", DataType::Float64, true),
+        ]));
+
+        let timestamp_array = Arc::new(TimestampMillisecondArray::from(vec![9]));
+        let value_array = Arc::new(Float64Array::from(vec![1.0]));
+        let batch =
+            RecordBatch::try_new(schema.clone(), vec![timestamp_array, value_array]).unwrap();
+
+        let memory_exec = DataSourceExec::new(Arc::new(
+            MemorySourceConfig::try_new(&[vec![batch]], schema, None).unwrap(),
+        ));
+
+        let output_schema = Arc::new(Schema::new(vec![
+            Field::new(
+                "timestamp",
+                DataType::Timestamp(TimeUnit::Millisecond, None),
+                true,
+            ),
+            Field::new("value", DataType::Float64, true),
+        ]));
+
+        let absent_exec = AbsentExec {
+            start: 0,
+            end: 9,
+            step: 1,
+            time_index_column: "timestamp".to_string(),
+            value_column: "value".to_string(),
+            fake_labels: vec![],
+            output_schema: output_schema.clone(),
+            input: Arc::new(memory_exec),
+            properties: Arc::new(PlanProperties::new(
+                EquivalenceProperties::new(output_schema.clone()),
+                Partitioning::UnknownPartitioning(1),
+                EmissionType::Incremental,
+                Boundedness::Bounded,
+            )),
+            metric: ExecutionPlanMetricsSet::new(),
+        };
+
+        let session_ctx = SessionContext::new_with_config(SessionConfig::new().with_batch_size(3));
+        let task_ctx = session_ctx.task_ctx();
+        let mut stream = absent_exec.execute(0, task_ctx).unwrap();
+
+        let mut output_timestamps = Vec::new();
+        while let Some(batch_result) = stream.next().await {
+            let batch = batch_result.unwrap();
+            let ts_array = batch
+                .column(0)
+                .as_any()
+                .downcast_ref::<TimestampMillisecondArray>()
+                .unwrap();
+            for i in 0..ts_array.len() {
+                if !ts_array.is_null(i) {
+                    output_timestamps.push(ts_array.value(i));
+                }
+            }
+        }
+
+        assert_eq!(output_timestamps, vec![0, 1, 2, 3, 4, 5, 6, 7, 8]);
+    }
 }
diff --git a/src/promql/src/extension_plan/range_manipulate.rs b/src/promql/src/extension_plan/range_manipulate.rs
index 6d57de62c6..ac03ff5e79 100644
--- a/src/promql/src/extension_plan/range_manipulate.rs
+++ b/src/promql/src/extension_plan/range_manipulate.rs
@@ -753,7 +753,7 @@ impl RangeManipulateStream {
                 if ts <= curr_ts {
                     range_end = range_end.max(cursor);
                 } else {
-                    range_start_index = range_start_index.checked_sub(1usize).unwrap_or_default();
+                    range_start_index = range_start_index.saturating_sub(1usize);
                     break;
                 }
                 cursor += 1;
diff --git a/src/promql/src/extension_plan/series_divide.rs b/src/promql/src/extension_plan/series_divide.rs
index 4a0b32f4e3..19385b7a97 100644
--- a/src/promql/src/extension_plan/series_divide.rs
+++ b/src/promql/src/extension_plan/series_divide.rs
@@ -632,7 +632,7 @@ impl SeriesDivideStream {
             let tags = TagIdentifier::try_new(batch, &self.tag_indices)?;
 
             // check if the first row is the same with last batch's last row
-            if resumed_batch_index > self.inspect_start.checked_sub(1).unwrap_or_default() {
+            if resumed_batch_index > self.inspect_start.saturating_sub(1) {
                 let last_batch = &self.buffer[resumed_batch_index - 1];
                 let last_row = last_batch.num_rows() - 1;
                 let last_tags = TagIdentifier::try_new(last_batch, &self.tag_indices)?;
diff --git a/src/promql/src/functions.rs b/src/promql/src/functions.rs
index e392d7dcf5..7c7452566a 100644
--- a/src/promql/src/functions.rs
+++ b/src/promql/src/functions.rs
@@ -31,9 +31,13 @@ pub use aggr_over_time::{
     PresentOverTime, StddevOverTime, StdvarOverTime, SumOverTime,
 };
 pub use changes::Changes;
-use datafusion::arrow::array::{ArrayRef, Float64Array, TimestampMillisecondArray};
+use datafusion::arrow::array::{
+    ArrayRef, DictionaryArray, Float64Array, TimestampMillisecondArray,
+};
 use datafusion::error::DataFusionError;
 use datafusion::physical_plan::ColumnarValue;
+use datatypes::arrow::array::Array;
+use datatypes::arrow::datatypes::Int64Type;
 pub use deriv::Deriv;
 pub use double_exponential_smoothing::DoubleExponentialSmoothing;
 pub use extrapolate_rate::{Delta, Increase, Rate};
@@ -44,6 +48,8 @@ pub use quantile_aggr::{QUANTILE_NAME, quantile_udaf};
 pub use resets::Resets;
 pub use round::Round;
 
+use crate::range_array::RangeArray;
+
 /// Extracts an array from a `ColumnarValue`.
 ///
 /// If the `ColumnarValue` is a scalar, it converts it to an array of size 1.
@@ -54,6 +60,24 @@ pub(crate) fn extract_array(columnar_value: &ColumnarValue) -> Result<ArrayRef,
     }
 }
 
+/// Extracts a validated [RangeArray] from a [ColumnarValue].
+pub(crate) fn extract_range_array(
+    columnar_value: &ColumnarValue,
+) -> Result<RangeArray, DataFusionError> {
+    let array = extract_array(columnar_value)?;
+    let dict = array
+        .as_any()
+        .downcast_ref::<DictionaryArray<Int64Type>>()
+        .ok_or_else(|| {
+            DataFusionError::Execution(format!(
+                "expected DictionaryArray<Int64>, found {}",
+                array.data_type()
+            ))
+        })?
+        .clone();
+    RangeArray::try_new(dict).map_err(DataFusionError::from)
+}
+
 /// compensation(Kahan) summation algorithm - a technique for reducing the numerical error
 /// in floating-point arithmetic. The algorithm also includes the modification ("Neumaier improvement")
 /// that reduces the numerical error further in cases
@@ -78,6 +102,29 @@ pub(crate) fn linear_regression(
     values: &Float64Array,
     intercept_time: i64,
 ) -> (Option<f64>, Option<f64>) {
+    linear_regression_slice(times.values(), values, 0, values.len(), intercept_time)
+}
+
+pub(crate) fn linear_regression_slice(
+    times: &[i64],
+    values: &Float64Array,
+    offset: usize,
+    len: usize,
+    intercept_time: i64,
+) -> (Option<f64>, Option<f64>) {
+    linear_regression_slices(times, offset, values, offset, len, intercept_time)
+}
+
+pub(crate) fn linear_regression_slices(
+    times: &[i64],
+    time_offset: usize,
+    values: &Float64Array,
+    value_offset: usize,
+    len: usize,
+    intercept_time: i64,
+) -> (Option<f64>, Option<f64>) {
+    let raw_values = values.values();
+    let has_nulls = values.null_count() > 0;
     let mut count: f64 = 0.0;
     let mut sum_x: f64 = 0.0;
     let mut sum_y: f64 = 0.0;
@@ -89,15 +136,18 @@ pub(crate) fn linear_regression(
     let mut comp_x2: f64 = 0.0;
 
     let mut const_y = true;
-    let init_y: f64 = values.value(0);
+    let mut init_y = None;
 
-    for (i, value) in values.iter().enumerate() {
-        let time = times.value(i) as f64;
-        if value.is_none() {
+    for i in 0..len {
+        let time_idx = time_offset + i;
+        let value_idx = value_offset + i;
+        if has_nulls && values.is_null(value_idx) {
             continue;
         }
-        let value = value.unwrap();
-        if const_y && i > 0 && value != init_y {
+        let value = raw_values[value_idx];
+        let time = times[time_idx] as f64;
+        let initial = init_y.get_or_insert(value);
+        if const_y && count > 0.0 && value != *initial {
             const_y = false;
         }
         count += 1.0;
@@ -113,6 +163,7 @@ pub(crate) fn linear_regression(
     }
 
     if const_y {
+        let init_y = init_y.unwrap();
         if !init_y.is_finite() {
             return (None, None);
         }
@@ -135,7 +186,14 @@ pub(crate) fn linear_regression(
 
 #[cfg(test)]
 mod test {
+    use std::sync::Arc;
+
+    use datafusion::physical_plan::ColumnarValue;
+    use datatypes::arrow::array::Int64Array;
+    use datatypes::arrow::datatypes::Int64Type;
+
     use super::*;
+    use crate::range_array::RangeArray;
 
     #[test]
     fn calculate_linear_regression_none() {
@@ -253,4 +311,26 @@ mod test {
         }
         assert_eq!(sum + c, 2.0)
     }
+
+    #[test]
+    fn extract_range_array_rejects_external_dictionary_with_null_keys() {
+        let keys = Int64Array::from_iter([Some(0), None]);
+        let values = Arc::new(Float64Array::from_iter([1.0, 2.0]));
+        let dict = DictionaryArray::<Int64Type>::try_new(keys, values).unwrap();
+
+        let err = extract_range_array(&ColumnarValue::Array(Arc::new(dict))).unwrap_err();
+        assert!(err.to_string().contains("Empty range is not expected"));
+    }
+
+    #[test]
+    fn extract_range_array_accepts_internal_packed_ranges() {
+        let values = Arc::new(Float64Array::from_iter([1.0, 2.0, 3.0]));
+        let range_array = RangeArray::from_ranges(values, [(0, 2), (1, 2)]).unwrap();
+
+        let extracted =
+            extract_range_array(&ColumnarValue::Array(Arc::new(range_array.into_dict()))).unwrap();
+
+        assert_eq!(extracted.get_offset_length(0), Some((0, 2)));
+        assert_eq!(extracted.get_offset_length(1), Some((1, 2)));
+    }
 }
diff --git a/src/promql/src/functions/extrapolate_rate.rs b/src/promql/src/functions/extrapolate_rate.rs
index 126a506bc9..726218de0f 100644
--- a/src/promql/src/functions/extrapolate_rate.rs
+++ b/src/promql/src/functions/extrapolate_rate.rs
@@ -32,18 +32,19 @@
 use std::fmt::Display;
 use std::sync::Arc;
 
-use datafusion::arrow::array::{Float64Array, TimestampMillisecondArray};
+use datafusion::arrow::array::{
+    DictionaryArray, Float64Array, Float64Builder, TimestampMillisecondArray,
+};
 use datafusion::arrow::datatypes::TimeUnit;
 use datafusion::common::{DataFusionError, Result as DfResult};
 use datafusion::logical_expr::{ScalarUDF, Volatility};
 use datafusion::physical_plan::ColumnarValue;
 use datafusion_expr::create_udf;
 use datatypes::arrow::array::{Array, Int64Array};
-use datatypes::arrow::datatypes::DataType;
+use datatypes::arrow::datatypes::{DataType, Int64Type};
 
-use crate::extension_plan::Millisecond;
 use crate::functions::extract_array;
-use crate::range_array::RangeArray;
+use crate::range_array::{RangeArray, unpack};
 
 pub type Delta = ExtrapolatedRate<false, false>;
 pub type Rate = ExtrapolatedRate<true, true>;
@@ -63,6 +64,17 @@ impl<const IS_COUNTER: bool, const IS_RATE: bool> ExtrapolatedRate<IS_COUNTER, I
         Self { range_length }
     }
 
+    fn func_name() -> &'static str {
+        match (IS_COUNTER, IS_RATE) {
+            (true, true) => "prom_rate",
+            (true, false) => "prom_increase",
+            (false, false) => "prom_delta",
+            (false, true) => {
+                unreachable!("gauge rate is not supported by ExtrapolatedRate")
+            }
+        }
+    }
+
     fn scalar_udf_with_name(name: &str) -> ScalarUDF {
         let input_types = vec![
             // timestamp range vector
@@ -92,11 +104,23 @@ impl<const IS_COUNTER: bool, const IS_RATE: bool> ExtrapolatedRate<IS_COUNTER, I
         }
 
         let range_length_array = extract_array(&inputs[3])?;
-        let range_length = range_length_array
+        let range_length_array = range_length_array
             .as_any()
             .downcast_ref::<Int64Array>()
-            .unwrap()
-            .value(0) as i64;
+            .ok_or_else(|| {
+                DataFusionError::Execution(format!(
+                    "{}: expect Int64 as range length type, found {}",
+                    Self::func_name(),
+                    range_length_array.data_type()
+                ))
+            })?;
+        if range_length_array.is_empty() || range_length_array.is_null(0) {
+            return Err(DataFusionError::Execution(format!(
+                "{}: range length must contain a non-null Int64 value",
+                Self::func_name()
+            )));
+        }
+        let range_length = range_length_array.value(0);
 
         Ok(Self::new(range_length))
     }
@@ -105,131 +129,209 @@ impl<const IS_COUNTER: bool, const IS_RATE: bool> ExtrapolatedRate<IS_COUNTER, I
     /// * 0: timestamp range vector
     /// * 1: value range vector
     /// * 2: timestamp vector
-    /// * 3: range length. Range duration in millisecond. Not used here
+    /// * 3: range length. Range duration in milliseconds
     fn calc(&self, input: &[ColumnarValue]) -> DfResult<ColumnarValue> {
-        assert_eq!(input.len(), 4);
+        if input.len() != 4 {
+            return Err(DataFusionError::Plan(
+                "ExtrapolatedRate function should have 4 inputs".to_string(),
+            ));
+        }
 
-        // construct matrix from input
-        let ts_array = extract_array(&input[0])?;
-        let ts_range = RangeArray::try_new(ts_array.to_data().into())?;
-        let value_array = extract_array(&input[1])?;
-        let value_range = RangeArray::try_new(value_array.to_data().into())?;
-        let ts = extract_array(&input[2])?;
-        let ts = ts
-            .as_any()
-            .downcast_ref::<TimestampMillisecondArray>()
-            .unwrap();
+        let ts_dict = extract_range_dict(
+            &input[0],
+            Self::func_name(),
+            "timestamp range vector",
+            &DataType::Timestamp(TimeUnit::Millisecond, None),
+        )?;
+        let value_dict = extract_range_dict(
+            &input[1],
+            Self::func_name(),
+            "value range vector",
+            &DataType::Float64,
+        )?;
+        let eval_ts_array = extract_eval_timestamps(&input[2], Self::func_name())?;
 
-        // calculation
-        let mut result_array = Vec::with_capacity(ts_range.len());
+        let keys = ts_dict.keys().values();
+        let num_windows = keys.len();
+        if value_dict.keys().len() != num_windows {
+            return Err(DataFusionError::Execution(format!(
+                "{}: timestamp and value ranges should have the same number of windows, found {} and {}",
+                Self::func_name(),
+                num_windows,
+                value_dict.keys().len()
+            )));
+        }
+        if value_dict.keys().values() != keys {
+            return Err(DataFusionError::Execution(format!(
+                "{}: timestamp and value ranges should have the same window layout",
+                Self::func_name()
+            )));
+        }
+        if eval_ts_array.len() != num_windows {
+            return Err(DataFusionError::Execution(format!(
+                "{}: evaluation timestamp vector should have the same number of rows as range inputs, found {} and {}",
+                Self::func_name(),
+                eval_ts_array.len(),
+                num_windows
+            )));
+        }
 
-        let all_timestamps = ts_range
+        let all_timestamps = ts_dict
             .values()
             .as_any()
             .downcast_ref::<TimestampMillisecondArray>()
-            .unwrap()
+            .expect("validated by extract_range_dict")
             .values();
-        let all_values = value_range
+        let all_values = value_dict
             .values()
             .as_any()
             .downcast_ref::<Float64Array>()
-            .unwrap()
+            .expect("validated by extract_range_dict")
             .values();
-        for index in 0..ts_range.len() {
-            // Safety: we are inside `ts_range`'s iterator which guarantees the index is valid.
-            let (offset, length) = ts_range.get_offset_length(index).unwrap();
+        let eval_ts = eval_ts_array.values();
 
-            let timestamps = &all_timestamps[offset..offset + length];
-            let end_ts = ts.value(index);
-            let values = &all_values[offset..offset + length];
+        let mut result_builder = Float64Builder::with_capacity(num_windows);
+        let range_length = self.range_length;
+        let range_length_secs = range_length as f64 / 1000.0;
 
-            if values.len() < 2 {
-                result_array.push(None);
+        let mut counter_correction = 0.0;
+        let mut prev_offset = usize::MAX;
+        let mut prev_length = 0usize;
+
+        for index in 0..num_windows {
+            let (raw_offset, raw_length) = unpack(keys[index]);
+            let offset = raw_offset as usize;
+            let length = raw_length as usize;
+
+            if length < 2 {
+                result_builder.append_null();
+                prev_offset = usize::MAX;
                 continue;
             }
 
-            // refer to functions.go L83-L110
-            let mut result_value = values.last().unwrap() - values.first().unwrap();
-            if IS_COUNTER {
-                for window in values.windows(2) {
-                    let prev = window[0];
-                    let curr = window[1];
-                    if curr < prev {
-                        result_value += prev
+            let end = offset + length;
+            let first_value = all_values[offset];
+            let last_value = all_values[end - 1];
+
+            let result_value = if IS_COUNTER {
+                // Adjacent normalized windows usually slide forward by one sample. Reuse the
+                // previous window's accumulated reset correction and adjust only the dropped and
+                // newly added edges, falling back to a full scan when the layout changes.
+                if prev_offset != usize::MAX && offset == prev_offset + 1 && length == prev_length {
+                    if all_values[prev_offset + 1] < all_values[prev_offset] {
+                        counter_correction -= all_values[prev_offset];
                     }
+                    if all_values[end - 1] < all_values[end - 2] {
+                        counter_correction += all_values[end - 2];
+                    }
+                } else {
+                    counter_correction = 0.0;
+                    for pair in all_values[offset..end].windows(2) {
+                        if pair[1] < pair[0] {
+                            counter_correction += pair[0];
+                        }
+                    }
+                }
+                last_value - first_value + counter_correction
+            } else {
+                last_value - first_value
+            };
+
+            prev_offset = offset;
+            prev_length = length;
+
+            let first_ts = all_timestamps[offset];
+            let last_ts = all_timestamps[end - 1];
+            let range_end = eval_ts[index];
+            let range_start = range_end - range_length;
+            let sampled_interval_ms = (last_ts - first_ts) as f64;
+            let average_interval_ms = sampled_interval_ms / (length - 1) as f64;
+            let mut duration_to_start_ms = (first_ts - range_start) as f64;
+            let duration_to_end_ms = (range_end - last_ts) as f64;
+
+            // Counters cannot be negative, so Prometheus allows the extrapolation window to snap
+            // back to the inferred zero point instead of extending into negative values.
+            if IS_COUNTER && result_value > 0.0 && first_value >= 0.0 {
+                let duration_to_zero = sampled_interval_ms * (first_value / result_value);
+                if duration_to_zero < duration_to_start_ms {
+                    duration_to_start_ms = duration_to_zero;
                 }
             }
 
-            let mut factor = Self::extrapolate_factor(
-                timestamps,
-                end_ts,
-                self.range_length,
-                *values.first().unwrap(),
-                result_value,
-            );
+            let extrapolation_threshold = average_interval_ms * 1.1;
+            let mut extrapolated_interval_ms = sampled_interval_ms;
+
+            // Mirror Prometheus extrapolation: extend to the real range boundary when a sample is
+            // close enough, otherwise add half an average sampling interval on that side.
+            if duration_to_start_ms < extrapolation_threshold {
+                extrapolated_interval_ms += duration_to_start_ms;
+            } else {
+                extrapolated_interval_ms += average_interval_ms / 2.0;
+            }
+            if duration_to_end_ms < extrapolation_threshold {
+                extrapolated_interval_ms += duration_to_end_ms;
+            } else {
+                extrapolated_interval_ms += average_interval_ms / 2.0;
+            }
+
+            let mut factor = extrapolated_interval_ms / sampled_interval_ms;
 
             if IS_RATE {
-                // safety: range_length is checked to be non-zero in the planner.
-                factor /= self.range_length as f64 / 1000.0;
+                factor /= range_length_secs;
             }
 
-            result_array.push(Some(result_value * factor));
+            result_builder.append_value(result_value * factor);
         }
 
-        let result = ColumnarValue::Array(Arc::new(Float64Array::from_iter(result_array)));
+        let result = ColumnarValue::Array(Arc::new(result_builder.finish()));
         Ok(result)
     }
+}
 
-    fn extrapolate_factor(
-        timestamps: &[Millisecond],
-        range_end: Millisecond,
-        range_length: Millisecond,
-        // the following two parameters are for counters.
-        // see functions.go L121 - L127
-        first_value: f64,
-        result_value: f64,
-    ) -> f64 {
-        // result_value
-        // refer to functions.go extrapolatedRate fn
-        // assume offset is processed (and it should be processed in normalize plan)
-        let range_start = range_end - range_length;
-        let mut duration_to_start = (timestamps.first().unwrap() - range_start) as f64 / 1000.0;
-        let duration_to_end = (range_end - timestamps.last().unwrap()) as f64 / 1000.0;
-        let sampled_interval =
-            (timestamps.last().unwrap() - timestamps.first().unwrap()) as f64 / 1000.0;
-        let average_duration_between_samples = sampled_interval / (timestamps.len() - 1) as f64;
+fn extract_range_dict(
+    columnar_value: &ColumnarValue,
+    func_name: &str,
+    arg_name: &str,
+    expected_value_type: &DataType,
+) -> DfResult<DictionaryArray<Int64Type>> {
+    let array = extract_array(columnar_value)?;
+    let dict = array
+        .as_any()
+        .downcast_ref::<DictionaryArray<Int64Type>>()
+        .ok_or_else(|| {
+            DataFusionError::Execution(format!(
+                "{func_name}: expect {arg_name} as DictionaryArray<Int64>, found {}",
+                array.data_type()
+            ))
+        })?
+        .clone();
 
-        // functions.go L122 - L134. quote:
-        // Counters cannot be negative. If we have any slope at
-        // all (i.e. resultValue went up), we can extrapolate
-        // the zero point of the counter. If the duration to the
-        // zero point is shorter than the durationToStart, we
-        // take the zero point as the start of the series,
-        // thereby avoiding extrapolation to negative counter
-        // values.
-        if IS_COUNTER && result_value > 0.0 && first_value >= 0.0 {
-            let duration_to_zero = sampled_interval * (first_value / result_value);
-            if duration_to_zero < duration_to_start {
-                duration_to_start = duration_to_zero;
-            }
-        }
-
-        let extrapolation_threshold = average_duration_between_samples * 1.1;
-        let mut extrapolate_to_interval = sampled_interval;
-
-        if duration_to_start < extrapolation_threshold {
-            extrapolate_to_interval += duration_to_start;
-        } else {
-            extrapolate_to_interval += average_duration_between_samples / 2.0;
-        }
-        if duration_to_end < extrapolation_threshold {
-            extrapolate_to_interval += duration_to_end;
-        } else {
-            extrapolate_to_interval += average_duration_between_samples / 2.0;
-        }
-
-        extrapolate_to_interval / sampled_interval
+    if &dict.value_type() != expected_value_type {
+        return Err(DataFusionError::Execution(format!(
+            "{func_name}: expect {arg_name} values of type {expected_value_type}, found {}",
+            dict.value_type()
+        )));
     }
+
+    RangeArray::try_new(dict.clone()).map_err(DataFusionError::from)?;
+    Ok(dict)
+}
+
+fn extract_eval_timestamps(
+    columnar_value: &ColumnarValue,
+    func_name: &str,
+) -> DfResult<TimestampMillisecondArray> {
+    let array = extract_array(columnar_value)?;
+    let timestamps = array
+        .as_any()
+        .downcast_ref::<TimestampMillisecondArray>()
+        .ok_or_else(|| {
+            DataFusionError::Execution(format!(
+                "{func_name}: expect evaluation timestamp vector as Timestamp(Millisecond), found {}",
+                array.data_type()
+            ))
+        })?;
+    Ok(timestamps.clone())
 }
 
 // delta
@@ -287,6 +389,7 @@ impl Display for ExtrapolatedRate<true, false> {
 mod test {
 
     use datafusion::arrow::array::ArrayRef;
+    use datafusion_common::ScalarValue;
 
     use super::*;
 
@@ -317,6 +420,80 @@ mod test {
         assert_eq!(output, expected);
     }
 
+    fn sample_range_inputs() -> (ColumnarValue, ColumnarValue, ColumnarValue) {
+        let ts_values = Arc::new(TimestampMillisecondArray::from_iter(
+            [1, 2, 3].into_iter().map(Some),
+        ));
+        let value_values = Arc::new(Float64Array::from_iter([1.0, 2.0, 3.0]));
+        let ranges = [(0, 2), (1, 2)];
+
+        let ts_range = RangeArray::from_ranges(ts_values, ranges).unwrap();
+        let value_range = RangeArray::from_ranges(value_values, ranges).unwrap();
+        let eval_ts = Arc::new(TimestampMillisecondArray::from_iter(
+            [2, 3].into_iter().map(Some),
+        )) as _;
+
+        (
+            ColumnarValue::Array(Arc::new(ts_range.into_dict())),
+            ColumnarValue::Array(Arc::new(value_range.into_dict())),
+            ColumnarValue::Array(eval_ts),
+        )
+    }
+
+    #[test]
+    fn rate_rejects_wrong_input_arity() {
+        let err = ExtrapolatedRate::<true, true>::new(5)
+            .calc(&[])
+            .unwrap_err();
+
+        assert!(err.to_string().contains("should have 4 inputs"));
+    }
+
+    #[test]
+    fn rate_rejects_non_int64_range_length() {
+        let (ts_range, value_range, eval_ts) = sample_range_inputs();
+
+        let err = ExtrapolatedRate::<true, true>::create_function(&[
+            ts_range,
+            value_range,
+            eval_ts,
+            ColumnarValue::Scalar(ScalarValue::Float64(Some(5.0))),
+        ])
+        .unwrap_err();
+
+        assert!(err.to_string().contains("range length type"));
+    }
+
+    #[test]
+    fn rate_rejects_empty_range_length() {
+        let (ts_range, value_range, eval_ts) = sample_range_inputs();
+
+        let err = ExtrapolatedRate::<true, true>::create_function(&[
+            ts_range,
+            value_range,
+            eval_ts,
+            ColumnarValue::Array(Arc::new(Int64Array::from(Vec::<i64>::new()))),
+        ])
+        .unwrap_err();
+
+        assert!(err.to_string().contains("range length must contain"));
+    }
+
+    #[test]
+    fn rate_rejects_null_range_length() {
+        let (ts_range, value_range, eval_ts) = sample_range_inputs();
+
+        let err = ExtrapolatedRate::<true, true>::create_function(&[
+            ts_range,
+            value_range,
+            eval_ts,
+            ColumnarValue::Array(Arc::new(Int64Array::from(vec![None]))),
+        ])
+        .unwrap_err();
+
+        assert!(err.to_string().contains("range length must contain"));
+    }
+
     #[test]
     fn increase_abnormal_input() {
         let ts_array = Arc::new(TimestampMillisecondArray::from_iter(
@@ -449,6 +626,169 @@ mod test {
         );
     }
 
+    #[test]
+    fn increase_counter_reset_wide_windows() {
+        let ts_array = Arc::new(TimestampMillisecondArray::from_iter(
+            [1, 2, 3, 4, 5, 6, 7].into_iter().map(Some),
+        ));
+        let values_array = Arc::new(Float64Array::from_iter([1.0, 2.0, 3.0, 1.0, 2.0, 1.0, 2.0]));
+        let ranges = [(0, 4), (1, 4), (2, 4), (3, 4)];
+        let ts_range = RangeArray::from_ranges(ts_array, ranges).unwrap();
+        let value_range = RangeArray::from_ranges(values_array, ranges).unwrap();
+        let timestamps = Arc::new(TimestampMillisecondArray::from_iter(
+            [4, 5, 6, 7].into_iter().map(Some),
+        )) as _;
+        extrapolated_rate_runner::<true, false>(
+            ts_range,
+            value_range,
+            timestamps,
+            vec![4.0, 3.5, 3.5, 4.0],
+        );
+    }
+
+    #[test]
+    fn rate_rejects_non_array_timestamp_ranges() {
+        let value_values = Arc::new(Float64Array::from_iter([1.0, 2.0]));
+        let value_range = RangeArray::from_ranges(value_values, [(0, 2)]).unwrap();
+        let eval_ts = Arc::new(TimestampMillisecondArray::from_iter([Some(2)]));
+
+        let err = ExtrapolatedRate::<true, true>::new(5)
+            .calc(&[
+                ColumnarValue::Scalar(ScalarValue::Int64(Some(0))),
+                ColumnarValue::Array(Arc::new(value_range.into_dict())),
+                ColumnarValue::Array(eval_ts),
+                ColumnarValue::Scalar(ScalarValue::Int64(Some(5))),
+            ])
+            .unwrap_err();
+
+        assert!(err.to_string().contains("timestamp range vector"));
+    }
+
+    #[test]
+    fn rate_rejects_non_timestamp_timestamp_range_values() {
+        let ts_values = Arc::new(Int64Array::from_iter([1, 2]));
+        let value_values = Arc::new(Float64Array::from_iter([1.0, 2.0]));
+        let ts_range = RangeArray::from_ranges(ts_values, [(0, 2)]).unwrap();
+        let value_range = RangeArray::from_ranges(value_values, [(0, 2)]).unwrap();
+        let eval_ts = Arc::new(TimestampMillisecondArray::from_iter([Some(2)]));
+
+        let err = ExtrapolatedRate::<true, true>::new(5)
+            .calc(&[
+                ColumnarValue::Array(Arc::new(ts_range.into_dict())),
+                ColumnarValue::Array(Arc::new(value_range.into_dict())),
+                ColumnarValue::Array(eval_ts),
+                ColumnarValue::Scalar(ScalarValue::Int64(Some(5))),
+            ])
+            .unwrap_err();
+
+        assert!(err.to_string().contains("values of type Timestamp"));
+    }
+
+    #[test]
+    fn rate_rejects_non_float_value_range_values() {
+        let ts_values = Arc::new(TimestampMillisecondArray::from_iter(
+            [1, 2].into_iter().map(Some),
+        ));
+        let value_values = Arc::new(Int64Array::from_iter([1, 2]));
+        let ts_range = RangeArray::from_ranges(ts_values, [(0, 2)]).unwrap();
+        let value_range = RangeArray::from_ranges(value_values, [(0, 2)]).unwrap();
+        let eval_ts = Arc::new(TimestampMillisecondArray::from_iter([Some(2)]));
+
+        let err = ExtrapolatedRate::<true, true>::new(5)
+            .calc(&[
+                ColumnarValue::Array(Arc::new(ts_range.into_dict())),
+                ColumnarValue::Array(Arc::new(value_range.into_dict())),
+                ColumnarValue::Array(eval_ts),
+                ColumnarValue::Scalar(ScalarValue::Int64(Some(5))),
+            ])
+            .unwrap_err();
+
+        assert!(
+            err.to_string()
+                .contains("value range vector values of type Float64")
+        );
+    }
+
+    #[test]
+    fn rate_rejects_mismatched_range_counts() {
+        let ts_values = Arc::new(TimestampMillisecondArray::from_iter(
+            [1, 2, 3].into_iter().map(Some),
+        ));
+        let value_values = Arc::new(Float64Array::from_iter([1.0, 2.0, 3.0]));
+        let ts_range = RangeArray::from_ranges(ts_values, [(0, 2), (1, 2)]).unwrap();
+        let value_range = RangeArray::from_ranges(value_values, [(0, 2)]).unwrap();
+        let eval_ts = Arc::new(TimestampMillisecondArray::from_iter(
+            [2, 3].into_iter().map(Some),
+        ));
+
+        let err = ExtrapolatedRate::<true, true>::new(5)
+            .calc(&[
+                ColumnarValue::Array(Arc::new(ts_range.into_dict())),
+                ColumnarValue::Array(Arc::new(value_range.into_dict())),
+                ColumnarValue::Array(eval_ts),
+                ColumnarValue::Scalar(ScalarValue::Int64(Some(5))),
+            ])
+            .unwrap_err();
+
+        assert!(err.to_string().contains("same number of windows"));
+    }
+
+    #[test]
+    fn rate_rejects_mismatched_range_layouts() {
+        let ts_values = Arc::new(TimestampMillisecondArray::from_iter(
+            [1, 2, 3, 4].into_iter().map(Some),
+        ));
+        let value_values = Arc::new(Float64Array::from_iter([1.0, 2.0, 3.0, 4.0]));
+        let ts_range = RangeArray::from_ranges(ts_values, [(0, 2), (1, 2)]).unwrap();
+        let value_range = RangeArray::from_ranges(value_values, [(0, 2), (2, 2)]).unwrap();
+        let eval_ts = Arc::new(TimestampMillisecondArray::from_iter(
+            [2, 4].into_iter().map(Some),
+        ));
+
+        let err = ExtrapolatedRate::<true, true>::new(5)
+            .calc(&[
+                ColumnarValue::Array(Arc::new(ts_range.into_dict())),
+                ColumnarValue::Array(Arc::new(value_range.into_dict())),
+                ColumnarValue::Array(eval_ts),
+                ColumnarValue::Scalar(ScalarValue::Int64(Some(5))),
+            ])
+            .unwrap_err();
+
+        assert!(err.to_string().contains("same window layout"));
+    }
+
+    #[test]
+    fn rate_rejects_non_timestamp_eval_vector() {
+        let (ts_range, value_range, _) = sample_range_inputs();
+
+        let err = ExtrapolatedRate::<true, true>::new(5)
+            .calc(&[
+                ts_range,
+                value_range,
+                ColumnarValue::Array(Arc::new(Float64Array::from_iter([2.0, 3.0]))),
+                ColumnarValue::Scalar(ScalarValue::Int64(Some(5))),
+            ])
+            .unwrap_err();
+
+        assert!(err.to_string().contains("evaluation timestamp vector"));
+    }
+
+    #[test]
+    fn rate_rejects_mismatched_eval_timestamp_rows() {
+        let (ts_range, value_range, _) = sample_range_inputs();
+
+        let err = ExtrapolatedRate::<true, true>::new(5)
+            .calc(&[
+                ts_range,
+                value_range,
+                ColumnarValue::Array(Arc::new(TimestampMillisecondArray::from_iter([Some(2)]))),
+                ColumnarValue::Scalar(ScalarValue::Int64(Some(5))),
+            ])
+            .unwrap_err();
+
+        assert!(err.to_string().contains("same number of rows"));
+    }
+
     #[test]
     fn rate_counter_reset() {
         let ts_array = Arc::new(TimestampMillisecondArray::from_iter(
diff --git a/src/promql/src/functions/idelta.rs b/src/promql/src/functions/idelta.rs
index eeec9a4be9..0772b0bf1e 100644
--- a/src/promql/src/functions/idelta.rs
+++ b/src/promql/src/functions/idelta.rs
@@ -15,7 +15,7 @@
 use std::fmt::Display;
 use std::sync::Arc;
 
-use datafusion::arrow::array::{Float64Array, TimestampMillisecondArray};
+use datafusion::arrow::array::{Float64Array, Float64Builder, TimestampMillisecondArray};
 use datafusion::arrow::datatypes::TimeUnit;
 use datafusion::common::DataFusionError;
 use datafusion::logical_expr::{ScalarUDF, Volatility};
@@ -94,49 +94,54 @@ impl<const IS_RATE: bool> IDelta<IS_RATE> {
             )),
         )?;
 
-        // calculation
-        let mut result_array = Vec::with_capacity(ts_range.len());
+        let ts_values = ts_range.values();
+        let ts_values = ts_values
+            .as_any()
+            .downcast_ref::<TimestampMillisecondArray>()
+            .unwrap()
+            .values();
+
+        let value_values = value_range.values();
+        let value_values = value_values
+            .as_any()
+            .downcast_ref::<Float64Array>()
+            .unwrap()
+            .values();
+
+        let mut result_builder = Float64Builder::with_capacity(ts_range.len());
 
         for index in 0..ts_range.len() {
-            let timestamps = ts_range.get(index).unwrap();
-            let timestamps = timestamps
-                .as_any()
-                .downcast_ref::<TimestampMillisecondArray>()
-                .unwrap()
-                .values();
-
-            let values = value_range.get(index).unwrap();
-            let values = values
-                .as_any()
-                .downcast_ref::<Float64Array>()
-                .unwrap()
-                .values();
+            let (ts_offset, len) = ts_range.get_offset_length(index).unwrap();
+            let (value_offset, value_len) = value_range.get_offset_length(index).unwrap();
             error::ensure(
-                timestamps.len() == values.len(),
+                len == value_len,
                 DataFusionError::Execution(format!(
                     "{}: input arrays should have the same length, found {} and {}",
                     Self::name(),
-                    timestamps.len(),
-                    values.len()
+                    len,
+                    value_len
                 )),
             )?;
-
-            let len = timestamps.len();
             if len < 2 {
-                result_array.push(None);
+                result_builder.append_null();
                 continue;
             }
 
-            // if is delta
+            let last_offset = ts_offset + len - 1;
+            let prev_offset = last_offset - 1;
+            let sampled_interval =
+                (ts_values[last_offset] - ts_values[prev_offset]) as f64 / 1000.0;
+
+            let last_value_offset = value_offset + len - 1;
+            let prev_value_offset = last_value_offset - 1;
+            let last_value = value_values[last_value_offset];
+            let prev_value = value_values[prev_value_offset];
+
             if !IS_RATE {
-                result_array.push(Some(values[len - 1] - values[len - 2]));
+                result_builder.append_value(last_value - prev_value);
                 continue;
             }
 
-            // else is rate
-            let sampled_interval = (timestamps[len - 1] - timestamps[len - 2]) as f64 / 1000.0;
-            let last_value = values[len - 1];
-            let prev_value = values[len - 2];
             let result_value = if last_value < prev_value {
                 // counter reset
                 last_value
@@ -144,10 +149,10 @@ impl<const IS_RATE: bool> IDelta<IS_RATE> {
                 last_value - prev_value
             };
 
-            result_array.push(Some(result_value / sampled_interval as f64));
+            result_builder.append_value(result_value / sampled_interval);
         }
 
-        let result = ColumnarValue::Array(Arc::new(Float64Array::from_iter(result_array)));
+        let result = ColumnarValue::Array(Arc::new(result_builder.finish()));
         Ok(result)
     }
 }
diff --git a/src/promql/src/functions/predict_linear.rs b/src/promql/src/functions/predict_linear.rs
index 09a46ed48f..dc49ec5d9f 100644
--- a/src/promql/src/functions/predict_linear.rs
+++ b/src/promql/src/functions/predict_linear.rs
@@ -17,7 +17,7 @@
 
 use std::sync::Arc;
 
-use datafusion::arrow::array::{Float64Array, TimestampMillisecondArray};
+use datafusion::arrow::array::{Float64Array, Float64Builder, TimestampMillisecondArray};
 use datafusion::arrow::datatypes::TimeUnit;
 use datafusion::common::DataFusionError;
 use datafusion::logical_expr::{ScalarUDF, Volatility};
@@ -28,7 +28,7 @@ use datatypes::arrow::array::Array;
 use datatypes::arrow::datatypes::DataType;
 
 use crate::error;
-use crate::functions::{extract_array, linear_regression};
+use crate::functions::{extract_range_array, linear_regression_slices};
 use crate::range_array::RangeArray;
 
 pub struct PredictLinear;
@@ -62,12 +62,10 @@ impl PredictLinear {
             DataFusionError::Plan("prom_predict_linear function should have 3 inputs".to_string()),
         )?;
 
-        let ts_array = extract_array(&input[0])?;
-        let value_array = extract_array(&input[1])?;
         let t_col = &input[2];
 
-        let ts_range: RangeArray = RangeArray::try_new(ts_array.to_data().into())?;
-        let value_range: RangeArray = RangeArray::try_new(value_array.to_data().into())?;
+        let ts_range = extract_range_array(&input[0])?;
+        let value_range = extract_range_array(&input[1])?;
         error::ensure(
             ts_range.len() == value_range.len(),
             DataFusionError::Execution(format!(
@@ -130,74 +128,85 @@ impl PredictLinear {
                 Box::new(t_array.iter())
             }
         };
-        let mut result_array = Vec::with_capacity(ts_range.len());
+        let all_timestamps = ts_range
+            .values()
+            .as_any()
+            .downcast_ref::<TimestampMillisecondArray>()
+            .unwrap()
+            .values();
+        let all_values = value_range
+            .values()
+            .as_any()
+            .downcast_ref::<Float64Array>()
+            .unwrap();
+        let mut result_builder = Float64Builder::with_capacity(ts_range.len());
         for (index, t) in t_iter.enumerate() {
-            let (timestamps, values) = get_ts_values(&ts_range, &value_range, index, Self::name())?;
-            let ret = predict_linear_impl(&timestamps, &values, t.unwrap());
-            result_array.push(ret);
+            match predict_linear_impl(
+                &ts_range,
+                &value_range,
+                all_timestamps,
+                all_values,
+                index,
+                t.unwrap(),
+                Self::name(),
+            )? {
+                Some(value) => result_builder.append_value(value),
+                None => result_builder.append_null(),
+            }
         }
 
-        let result = ColumnarValue::Array(Arc::new(Float64Array::from_iter(result_array)));
+        let result = ColumnarValue::Array(Arc::new(result_builder.finish()));
         Ok(result)
     }
 }
 
-fn get_ts_values(
+fn predict_linear_impl(
     ts_range: &RangeArray,
     value_range: &RangeArray,
+    all_timestamps: &[i64],
+    all_values: &Float64Array,
     index: usize,
+    t: i64,
     func_name: &str,
-) -> Result<(TimestampMillisecondArray, Float64Array), DataFusionError> {
-    let timestamps = ts_range
-        .get(index)
-        .unwrap()
-        .as_any()
-        .downcast_ref::<TimestampMillisecondArray>()
-        .unwrap()
-        .clone();
-    let values = value_range
-        .get(index)
-        .unwrap()
-        .as_any()
-        .downcast_ref::<Float64Array>()
-        .unwrap()
-        .clone();
+) -> Result<Option<f64>, DataFusionError> {
+    let (ts_offset, ts_len) = ts_range.get_offset_length(index).unwrap();
+    let (value_offset, value_len) = value_range.get_offset_length(index).unwrap();
     error::ensure(
-        timestamps.len() == values.len(),
+        ts_len == value_len,
         DataFusionError::Execution(format!(
             "{}: time and value arrays in a group should have the same length, found {} and {}",
-            func_name,
-            timestamps.len(),
-            values.len()
+            func_name, ts_len, value_len
         )),
     )?;
-    Ok((timestamps, values))
-}
-
-fn predict_linear_impl(
-    timestamps: &TimestampMillisecondArray,
-    values: &Float64Array,
-    t: i64,
-) -> Option<f64> {
-    if timestamps.len() < 2 {
-        return None;
+    if ts_len < 2 {
+        return Ok(None);
     }
 
     // last timestamp is evaluation timestamp
-    let evaluate_ts = timestamps.value(timestamps.len() - 1);
-    let (slope, intercept) = linear_regression(timestamps, values, evaluate_ts);
+    let evaluate_ts = all_timestamps[ts_offset + ts_len - 1];
+    let (slope, intercept) = linear_regression_slices(
+        all_timestamps,
+        ts_offset,
+        all_values,
+        value_offset,
+        value_len,
+        evaluate_ts,
+    );
 
     if slope.is_none() || intercept.is_none() {
-        return None;
+        return Ok(None);
     }
 
-    Some(slope.unwrap() * t as f64 + intercept.unwrap())
+    Ok(Some(slope.unwrap() * t as f64 + intercept.unwrap()))
 }
 
 #[cfg(test)]
 mod test {
     use std::vec;
 
+    use datafusion::arrow::array::{DictionaryArray, Int64Array};
+    use datatypes::arrow::datatypes::Int64Type;
+
     use super::*;
     use crate::functions::test_util::simple_range_udf_runner;
 
@@ -304,4 +313,44 @@ mod test {
             vec![Some(82765.9090909091)],
         );
     }
+
+    #[test]
+    fn calculate_predict_linear_with_misaligned_offsets() {
+        let ts_values = Arc::new(TimestampMillisecondArray::from_iter(
+            [0i64, 1000, 2000, 3000].into_iter().map(Some),
+        ));
+        let value_values = Arc::new(Float64Array::from_iter([10.0, 20.0, 30.0]));
+        let ts_array = RangeArray::from_ranges(ts_values, [(1, 3)]).unwrap();
+        let value_array = RangeArray::from_ranges(value_values, [(0, 3)]).unwrap();
+
+        simple_range_udf_runner(
+            PredictLinear::scalar_udf(),
+            ts_array,
+            value_array,
+            vec![ScalarValue::Int64(Some(0))],
+            vec![Some(30.0)],
+        );
+    }
+
+    #[test]
+    fn predict_linear_rejects_external_dictionary_with_null_keys() {
+        let ts_values = Arc::new(TimestampMillisecondArray::from_iter(
+            [0i64, 1000].into_iter().map(Some),
+        ));
+        let ts_keys = Int64Array::from_iter([Some(0), None]);
+        let ts_dict = DictionaryArray::<Int64Type>::try_new(ts_keys, ts_values).unwrap();
+
+        let value_values = Arc::new(Float64Array::from_iter([1.0, 2.0]));
+        let value_keys = Int64Array::from_iter([Some(0), Some(1)]);
+        let value_dict = DictionaryArray::<Int64Type>::try_new(value_keys, value_values).unwrap();
+
+        let err = PredictLinear::predict_linear(&[
+            ColumnarValue::Array(Arc::new(ts_dict)),
+            ColumnarValue::Array(Arc::new(value_dict)),
+            ColumnarValue::Scalar(ScalarValue::Int64(Some(0))),
+        ])
+        .unwrap_err();
+
+        assert!(err.to_string().contains("Empty range is not expected"));
+    }
 }
diff --git a/src/promql/src/functions/quantile.rs b/src/promql/src/functions/quantile.rs
index f368d5908c..35f0c1aa81 100644
--- a/src/promql/src/functions/quantile.rs
+++ b/src/promql/src/functions/quantile.rs
@@ -14,7 +14,7 @@
 
 use std::sync::Arc;
 
-use datafusion::arrow::array::Float64Array;
+use datafusion::arrow::array::{Float64Array, Float64Builder};
 use datafusion::arrow::datatypes::TimeUnit;
 use datafusion::common::DataFusionError;
 use datafusion::logical_expr::{ScalarUDF, Volatility};
@@ -93,8 +93,14 @@ impl QuantileOverTime {
             )),
         )?;
 
-        // calculation
-        let mut result_array = Vec::with_capacity(ts_range.len());
+        let all_values = value_range
+            .values()
+            .as_any()
+            .downcast_ref::<Float64Array>()
+            .unwrap()
+            .values();
+        let mut result_builder = Float64Builder::with_capacity(ts_range.len());
+        let mut scratch = Vec::new();
 
         match quantile_col {
             ColumnarValue::Scalar(quantile_scalar) => {
@@ -107,25 +113,26 @@ impl QuantileOverTime {
                 };
 
                 for index in 0..ts_range.len() {
-                    let timestamps = ts_range.get(index).unwrap();
-                    let values = value_range.get(index).unwrap();
-                    let values = values
-                        .as_any()
-                        .downcast_ref::<Float64Array>()
-                        .unwrap()
-                        .values();
+                    let (_, ts_len) = ts_range.get_offset_length(index).unwrap();
+                    let (value_offset, value_len) = value_range.get_offset_length(index).unwrap();
                     error::ensure(
-                        timestamps.len() == values.len(),
+                        ts_len == value_len,
                         DataFusionError::Execution(format!(
                             "{}: time and value arrays in a group should have the same length, found {} and {}",
                             Self::name(),
-                            timestamps.len(),
-                            values.len()
+                            ts_len,
+                            value_len
                         )),
                     )?;
 
-                    let result = quantile_impl(values, quantile);
-                    result_array.push(result);
+                    match quantile_with_scratch(
+                        &all_values[value_offset..value_offset + value_len],
+                        quantile,
+                        &mut scratch,
+                    ) {
+                        Some(value) => result_builder.append_value(value),
+                        None => result_builder.append_null(),
+                    }
                 }
             }
             ColumnarValue::Array(quantile_array) => {
@@ -150,20 +157,15 @@ impl QuantileOverTime {
                     )),
                 )?;
                 for index in 0..ts_range.len() {
-                    let timestamps = ts_range.get(index).unwrap();
-                    let values = value_range.get(index).unwrap();
-                    let values = values
-                        .as_any()
-                        .downcast_ref::<Float64Array>()
-                        .unwrap()
-                        .values();
+                    let (_, ts_len) = ts_range.get_offset_length(index).unwrap();
+                    let (value_offset, value_len) = value_range.get_offset_length(index).unwrap();
                     error::ensure(
-                        timestamps.len() == values.len(),
+                        ts_len == value_len,
                         DataFusionError::Execution(format!(
                             "{}: time and value arrays in a group should have the same length, found {} and {}",
                             Self::name(),
-                            timestamps.len(),
-                            values.len()
+                            ts_len,
+                            value_len
                         )),
                     )?;
                     let quantile = if quantile_array.is_null(index) {
@@ -171,19 +173,32 @@ impl QuantileOverTime {
                     } else {
                         quantile_array.value(index)
                     };
-                    let result = quantile_impl(values, quantile);
-                    result_array.push(result);
+                    match quantile_with_scratch(
+                        &all_values[value_offset..value_offset + value_len],
+                        quantile,
+                        &mut scratch,
+                    ) {
+                        Some(value) => result_builder.append_value(value),
+                        None => result_builder.append_null(),
+                    }
                 }
             }
         }
 
-        let result = ColumnarValue::Array(Arc::new(Float64Array::from_iter(result_array)));
+        let result = ColumnarValue::Array(Arc::new(result_builder.finish()));
         Ok(result)
     }
 }
 
 /// Refer to <https://github.com/prometheus/prometheus/blob/6e2905a4d4ff9b47b1f6d201333f5bd53633f921/promql/quantile.go#L357-L386>
 pub(crate) fn quantile_impl(values: &[f64], quantile: f64) -> Option<f64> {
+    let mut scratch = Vec::new();
+    quantile_with_scratch(values, quantile, &mut scratch)
+}
+
+/// Same as [quantile_impl] but reuses a caller-provided scratch buffer to avoid
+/// per-call allocation.
+fn quantile_with_scratch(values: &[f64], quantile: f64, scratch: &mut Vec<f64>) -> Option<f64> {
     if quantile.is_nan() || values.is_empty() {
         return Some(f64::NAN);
     }
@@ -194,17 +209,18 @@ pub(crate) fn quantile_impl(values: &[f64], quantile: f64) -> Option<f64> {
         return Some(f64::INFINITY);
     }
 
-    let mut values = values.to_vec();
-    values.sort_unstable_by(f64::total_cmp);
+    scratch.clear();
+    scratch.extend_from_slice(values);
+    scratch.sort_unstable_by(f64::total_cmp);
 
-    let length = values.len();
+    let length = scratch.len();
     let rank = quantile * (length - 1) as f64;
 
-    let lower_index = 0.max(rank.floor() as usize);
+    let lower_index = rank.floor() as usize;
     let upper_index = (length - 1).min(lower_index + 1);
     let weight = rank - rank.floor();
 
-    let result = values[lower_index] * (1.0 - weight) + values[upper_index] * weight;
+    let result = scratch[lower_index] * (1.0 - weight) + scratch[upper_index] * weight;
     Some(result)
 }
 
diff --git a/src/promql/src/range_array.rs b/src/promql/src/range_array.rs
index 4af5217a76..dccc199181 100644
--- a/src/promql/src/range_array.rs
+++ b/src/promql/src/range_array.rs
@@ -248,7 +248,7 @@ fn pack(offset: u32, length: u32) -> i64 {
     bytemuck::cast::<[u32; 2], i64>([offset, length])
 }
 
-fn unpack(compound: i64) -> (u32, u32) {
+pub(crate) fn unpack(compound: i64) -> (u32, u32) {
     let [offset, length] = bytemuck::cast::<i64, [u32; 2]>(compound);
     (offset, length)
 }
diff --git a/src/puffin/src/puffin_manager/fs_puffin_manager/reader.rs b/src/puffin/src/puffin_manager/fs_puffin_manager/reader.rs
index c660d1e19a..b35b03875e 100644
--- a/src/puffin/src/puffin_manager/fs_puffin_manager/reader.rs
+++ b/src/puffin/src/puffin_manager/fs_puffin_manager/reader.rs
@@ -262,7 +262,7 @@ where
             tasks.push(task);
         }
 
-        let size = futures::future::try_join_all(tasks.into_iter())
+        let size = futures::future::try_join_all(tasks)
             .await
             .into_iter()
             .flatten()
diff --git a/src/query/Cargo.toml b/src/query/Cargo.toml
index af81325aa9..b92921d29b 100644
--- a/src/query/Cargo.toml
+++ b/src/query/Cargo.toml
@@ -40,11 +40,13 @@ common-time.workspace = true
 datafusion.workspace = true
 datafusion-common.workspace = true
 datafusion-expr.workspace = true
+datafusion-expr-common.workspace = true
 datafusion-functions.workspace = true
 datafusion-optimizer.workspace = true
 datafusion-physical-expr.workspace = true
 datafusion-sql.workspace = true
 datatypes.workspace = true
+either.workspace = true
 futures.workspace = true
 futures-util.workspace = true
 greptime-proto.workspace = true
diff --git a/src/query/src/datafusion.rs b/src/query/src/datafusion.rs
index dc84c4afac..a173f2aee8 100644
--- a/src/query/src/datafusion.rs
+++ b/src/query/src/datafusion.rs
@@ -15,6 +15,7 @@
 //! Planner, QueryEngine implementations based on DataFusion.
 
 mod error;
+mod json_expr_planner;
 mod planner;
 
 use std::any::Any;
@@ -54,12 +55,15 @@ use crate::analyze::DistAnalyzeExec;
 pub use crate::datafusion::planner::DfContextProviderAdapter;
 use crate::dist_plan::{DistPlannerOptions, MergeScanLogicalPlan};
 use crate::error::{
-    CatalogSnafu, ConvertSchemaSnafu, CreateRecordBatchSnafu, MissingTableMutationHandlerSnafu,
+    CatalogSnafu, CreateRecordBatchSnafu, MissingTableMutationHandlerSnafu,
     MissingTimestampColumnSnafu, QueryExecutionSnafu, Result, TableMutationSnafu,
     TableNotFoundSnafu, TableReadOnlySnafu, UnsupportedExprSnafu,
 };
 use crate::executor::QueryExecutor;
-use crate::metrics::{OnDone, QUERY_STAGE_ELAPSED};
+use crate::metrics::{
+    OnDone, QUERY_STAGE_ELAPSED, maybe_attach_region_watermark_metrics,
+    should_collect_region_watermark_from_query_ctx,
+};
 use crate::physical_wrapper::PhysicalPlanWrapperRef;
 use crate::planner::{DfLogicalPlanner, LogicalPlanner};
 use crate::query_engine::{DescribeResult, QueryEngineContext, QueryEngineState};
@@ -100,8 +104,10 @@ impl DatafusionQueryEngine {
             optimized_physical_plan
         };
 
+        let stream = self.execute_stream(&ctx, &physical_plan)?;
+
         Ok(Output::new(
-            OutputData::Stream(self.execute_stream(&ctx, &physical_plan)?),
+            OutputData::Stream(stream),
             OutputMeta::new_with_plan(physical_plan),
         ))
     }
@@ -128,10 +134,10 @@ impl DatafusionQueryEngine {
         let table_name = dml.table_name.resolve(default_catalog, default_schema);
         let table = self.find_table(&table_name, &query_ctx).await?;
 
-        let output = self
+        let Output { data, meta } = self
             .exec_query_plan((*dml.input).clone(), query_ctx.clone())
             .await?;
-        let mut stream = match output.data {
+        let mut stream = match data {
             OutputData::RecordBatches(batches) => batches.as_stream(),
             OutputData::Stream(stream) => stream,
             _ => unreachable!(),
@@ -167,7 +173,7 @@ impl DatafusionQueryEngine {
         }
         Ok(Output::new(
             OutputData::AffectedRows(affected_rows),
-            OutputMeta::new_with_cost(insert_cost),
+            OutputMeta::new(meta.plan, insert_cost),
         ))
     }
 
@@ -354,25 +360,6 @@ impl DatafusionQueryEngine {
         Ok(physical_plan)
     }
 
-    #[tracing::instrument(skip_all)]
-    pub fn optimize(
-        &self,
-        context: &QueryEngineContext,
-        plan: &LogicalPlan,
-    ) -> Result<LogicalPlan> {
-        let _timer = metrics::OPTIMIZE_LOGICAL_ELAPSED.start_timer();
-
-        // Optimized by extension rules
-        let optimized_plan = self
-            .state
-            .optimize_by_extension_rules(plan.clone(), context)?;
-
-        // Optimized by datafusion optimizer
-        let optimized_plan = self.state.session_state().optimize(&optimized_plan)?;
-
-        Ok(optimized_plan)
-    }
-
     #[tracing::instrument(skip_all)]
     fn optimize_physical_plan(
         &self,
@@ -444,32 +431,9 @@ impl QueryEngine for DatafusionQueryEngine {
     async fn describe(
         &self,
         plan: LogicalPlan,
-        query_ctx: QueryContextRef,
+        _query_ctx: QueryContextRef,
     ) -> Result<DescribeResult> {
-        let ctx = self.engine_context(query_ctx);
-        if let Ok(optimised_plan) = self.optimize(&ctx, &plan) {
-            let schema = optimised_plan
-                .schema()
-                .clone()
-                .try_into()
-                .context(ConvertSchemaSnafu)?;
-            Ok(DescribeResult {
-                schema,
-                logical_plan: optimised_plan,
-            })
-        } else {
-            // Table's like those in information_schema cannot be optimized when
-            // it contains parameters. So we fallback to original plans.
-            let schema = plan
-                .schema()
-                .clone()
-                .try_into()
-                .context(ConvertSchemaSnafu)?;
-            Ok(DescribeResult {
-                schema,
-                logical_plan: plan,
-            })
-        }
+        Ok(DescribeResult { logical_plan: plan })
     }
 
     async fn execute(&self, plan: LogicalPlan, query_ctx: QueryContextRef) -> Result<Output> {
@@ -585,7 +549,10 @@ impl QueryExecutor for DatafusionQueryEngine {
         ctx: &QueryEngineContext,
         plan: &Arc<dyn ExecutionPlan>,
     ) -> Result<SendableRecordBatchStream> {
-        let explain_verbose = ctx.query_ctx().explain_verbose();
+        let query_ctx = ctx.query_ctx();
+        let explain_verbose = query_ctx.explain_verbose();
+        let should_collect_region_watermark =
+            should_collect_region_watermark_from_query_ctx(&query_ctx)?;
         let output_partitions = plan.properties().output_partitioning().partition_count();
         if explain_verbose {
             common_telemetry::info!("Executing query plan, output_partitions: {output_partitions}");
@@ -621,7 +588,11 @@ impl QueryExecutor for DatafusionQueryEngine {
                         );
                     }
                 });
-                Ok(Box::pin(stream))
+                Ok(maybe_attach_region_watermark_metrics(
+                    Box::pin(stream),
+                    plan.clone(),
+                    should_collect_region_watermark,
+                ))
             }
             _ => {
                 // merge into a single partition
@@ -640,7 +611,7 @@ impl QueryExecutor for DatafusionQueryEngine {
                     .map_err(BoxedError::new)
                     .context(QueryExecutionSnafu)?;
                 stream.set_metrics2(plan.clone());
-                stream.set_explain_verbose(ctx.query_ctx().explain_verbose());
+                stream.set_explain_verbose(explain_verbose);
                 let stream = OnDone::new(Box::pin(stream), move || {
                     let exec_cost = exec_timer.stop_and_record();
                     if explain_verbose {
@@ -650,7 +621,11 @@ impl QueryExecutor for DatafusionQueryEngine {
                         );
                     }
                 });
-                Ok(Box::pin(stream))
+                Ok(maybe_attach_region_watermark_metrics(
+                    Box::pin(stream),
+                    plan.clone(),
+                    should_collect_region_watermark,
+                ))
             }
         }
     }
@@ -910,10 +885,10 @@ mod tests {
             .await
             .unwrap();
 
-        let DescribeResult {
-            schema,
-            logical_plan,
-        } = engine.describe(plan, QueryContext::arc()).await.unwrap();
+        let DescribeResult { logical_plan } =
+            engine.describe(plan, QueryContext::arc()).await.unwrap();
+
+        let schema: Schema = logical_plan.schema().clone().try_into().unwrap();
 
         assert_eq!(
             schema.column_schemas()[0],
@@ -924,7 +899,7 @@ mod tests {
             )
         );
         assert_eq!(
-            "Limit: skip=0, fetch=20\n  Aggregate: groupBy=[[]], aggr=[[sum(CAST(numbers.number AS UInt64))]]\n    TableScan: numbers projection=[number]",
+            "Limit: skip=0, fetch=20\n  Projection: sum(numbers.number)\n    Aggregate: groupBy=[[]], aggr=[[sum(numbers.number)]]\n      TableScan: numbers",
             format!("{}", logical_plan.display_indent())
         );
     }
diff --git a/src/query/src/datafusion/json_expr_planner.rs b/src/query/src/datafusion/json_expr_planner.rs
new file mode 100644
index 0000000000..4546c0a84d
--- /dev/null
+++ b/src/query/src/datafusion/json_expr_planner.rs
@@ -0,0 +1,286 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::sync::{Arc, LazyLock};
+
+use arrow_schema::Field;
+use arrow_schema::extension::ExtensionType;
+use common_function::scalars::json::json_get::JsonGetWithType;
+use common_function::scalars::udf::create_udf;
+use datafusion_common::arrow::datatypes::DataType;
+use datafusion_common::{Column, DFSchema, Result, ScalarValue, TableReference};
+use datafusion_expr::expr::{BinaryExpr, ScalarFunction};
+use datafusion_expr::planner::{ExprPlanner, PlannerResult, RawBinaryExpr};
+use datafusion_expr::{Expr, ExprSchemable, Operator, ScalarUDF};
+use datatypes::extension::json::JsonExtensionType;
+use either::Either;
+use sqlparser::ast::BinaryOperator;
+
+/// Rewrites JSON-aware SQL expressions into DataFusion expressions.
+///
+/// This planner handles two cases:
+/// - Rewrites compound identifiers on JSON extension columns into `json_get` function.
+///   For example, `select a.b.c` => `select json_get(a, "b.c")`.
+/// - Pushes an "expected type" argument into the `json_get` function when it participates in a
+///   binary operator. So that `json_get` knows the wanted data type when dealing with variant
+///   JSON values.
+///   For example, `select json_get(a, "b.c") + 1` => `select json_get(a, "b.c", NULL::Int64) + 1`.
+#[derive(Debug)]
+pub(crate) struct JsonExprPlanner;
+
+impl ExprPlanner for JsonExprPlanner {
+    fn plan_binary_op(
+        &self,
+        expr: RawBinaryExpr,
+        schema: &DFSchema,
+    ) -> Result<PlannerResult<RawBinaryExpr>> {
+        let RawBinaryExpr {
+            op,
+            mut left,
+            mut right,
+        } = expr;
+
+        if extract_untyped_json_get(&mut left).is_none()
+            && extract_untyped_json_get(&mut right).is_none()
+        {
+            return Ok(PlannerResult::Original(RawBinaryExpr { op, left, right }));
+        }
+
+        let Some(expr_op) = parse_sql_op(&op) else {
+            return Ok(PlannerResult::Original(RawBinaryExpr { op, left, right }));
+        };
+
+        let left_type = left.get_type(schema)?;
+        let right_type = right.get_type(schema)?;
+        let left = push_json_get_type_arg(left, right_type)?;
+        let right = push_json_get_type_arg(right, left_type)?;
+        match (left, right) {
+            (Either::Left(left), Either::Left(right)) => {
+                Ok(PlannerResult::Original(RawBinaryExpr { op, left, right }))
+            }
+            (left, right) => Ok(PlannerResult::Planned(Expr::BinaryExpr(BinaryExpr::new(
+                Box::new(left.into_inner()),
+                expr_op,
+                Box::new(right.into_inner()),
+            )))),
+        }
+    }
+
+    fn plan_compound_identifier(
+        &self,
+        field: &Field,
+        qualifier: Option<&TableReference>,
+        nested_names: &[String],
+    ) -> Result<PlannerResult<Vec<Expr>>> {
+        if field.extension_type_name() != Some(JsonExtensionType::NAME) {
+            return Ok(PlannerResult::Original(Vec::new()));
+        }
+
+        static JSON_GET_UDF: LazyLock<Arc<ScalarUDF>> =
+            LazyLock::new(|| Arc::new(create_udf(Arc::new(JsonGetWithType::default()))));
+
+        let json_get = JSON_GET_UDF.clone();
+        let path = nested_names.join(".");
+        Ok(PlannerResult::Planned(Expr::ScalarFunction(
+            ScalarFunction::new_udf(
+                json_get,
+                vec![
+                    Expr::Column(Column::from((qualifier, field))),
+                    Expr::Literal(ScalarValue::Utf8(Some(path)), None),
+                ],
+            ),
+        )))
+    }
+}
+
+fn extract_untyped_json_get(expr: &mut Expr) -> Option<&mut ScalarFunction> {
+    match expr {
+        Expr::ScalarFunction(f)
+            if f.func.name().eq_ignore_ascii_case(JsonGetWithType::NAME) && f.args.len() == 2 =>
+        {
+            Some(f)
+        }
+        _ => None,
+    }
+}
+
+fn push_json_get_type_arg(mut expr: Expr, data_type: DataType) -> Result<Either<Expr, Expr>> {
+    let Some(json_get) = extract_untyped_json_get(&mut expr) else {
+        return Ok(Either::Left(expr));
+    };
+
+    let with_type = ScalarValue::try_new_null(&data_type).map(|x| Expr::Literal(x, None))?;
+    json_get.args.push(with_type);
+
+    Ok(Either::Right(expr))
+}
+
+fn parse_sql_op(op: &BinaryOperator) -> Option<Operator> {
+    match *op {
+        BinaryOperator::Plus => Some(Operator::Plus),
+        BinaryOperator::Minus => Some(Operator::Minus),
+        BinaryOperator::Multiply => Some(Operator::Multiply),
+        BinaryOperator::Divide => Some(Operator::Divide),
+        BinaryOperator::Modulo => Some(Operator::Modulo),
+        BinaryOperator::Gt => Some(Operator::Gt),
+        BinaryOperator::GtEq => Some(Operator::GtEq),
+        BinaryOperator::Lt => Some(Operator::Lt),
+        BinaryOperator::LtEq => Some(Operator::LtEq),
+        BinaryOperator::Eq => Some(Operator::Eq),
+        BinaryOperator::NotEq => Some(Operator::NotEq),
+        BinaryOperator::And => Some(Operator::And),
+        BinaryOperator::Or => Some(Operator::Or),
+        BinaryOperator::BitwiseAnd => Some(Operator::BitwiseAnd),
+        BinaryOperator::BitwiseOr => Some(Operator::BitwiseOr),
+        BinaryOperator::BitwiseXor => Some(Operator::BitwiseXor),
+        _ => None,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use arrow_schema::Fields;
+    use datatypes::extension::json::JsonMetadata;
+
+    use super::*;
+
+    fn json_get_expr(base: Expr, path: &str) -> Expr {
+        let json_get = Arc::new(create_udf(Arc::new(JsonGetWithType::default())));
+        Expr::ScalarFunction(ScalarFunction::new_udf(
+            json_get,
+            vec![
+                base,
+                Expr::Literal(ScalarValue::Utf8(Some(path.to_string())), None),
+            ],
+        ))
+    }
+
+    fn json_field(name: &str) -> Field {
+        Field::new(name, DataType::Binary, true)
+            .with_extension_type(JsonExtensionType::new(Arc::new(JsonMetadata::default())))
+    }
+
+    #[test]
+    fn test_plan_binary_op() -> Result<()> {
+        let planner = JsonExprPlanner;
+        let schema = DFSchema::from_unqualified_fields(
+            Fields::from(vec![Field::new("value", DataType::Int64, true)]),
+            Default::default(),
+        )?;
+
+        let planned = planner.plan_binary_op(
+            RawBinaryExpr {
+                op: BinaryOperator::Eq,
+                left: json_get_expr(
+                    Expr::Literal(ScalarValue::Binary(Some(b"{\"a\": 1}".to_vec())), None),
+                    "a",
+                ),
+                right: Expr::Column(Column::new_unqualified("value")),
+            },
+            &schema,
+        )?;
+
+        match planned {
+            PlannerResult::Planned(Expr::BinaryExpr(expr)) => {
+                assert_eq!(expr.op, Operator::Eq);
+
+                match expr.left.as_ref() {
+                    Expr::ScalarFunction(func) => {
+                        assert_eq!(func.func.name(), JsonGetWithType::NAME);
+                        assert_eq!(func.args.len(), 3);
+                        assert_eq!(func.args[2], Expr::Literal(ScalarValue::Int64(None), None));
+                    }
+                    other => panic!("expected json_get on left side, got {other:?}"),
+                }
+
+                assert_eq!(
+                    expr.right.as_ref(),
+                    &Expr::Column(Column::new_unqualified("value"))
+                );
+            }
+            other => panic!("expected planned binary expression, got {other:?}"),
+        }
+
+        let original = planner.plan_binary_op(
+            RawBinaryExpr {
+                op: BinaryOperator::StringConcat,
+                left: Expr::Column(Column::new_unqualified("value")),
+                right: Expr::Literal(ScalarValue::Utf8(Some("x".to_string())), None),
+            },
+            &schema,
+        )?;
+
+        match original {
+            PlannerResult::Original(expr) => {
+                assert!(matches!(expr.op, BinaryOperator::StringConcat));
+                assert_eq!(expr.left, Expr::Column(Column::new_unqualified("value")));
+                assert_eq!(
+                    expr.right,
+                    Expr::Literal(ScalarValue::Utf8(Some("x".to_string())), None)
+                );
+            }
+            other => panic!(
+                "expected original expression for unsupported operator, got {:?}",
+                other,
+            ),
+        }
+
+        Ok(())
+    }
+
+    #[test]
+    fn test_plan_compound_identifier() -> Result<()> {
+        let planner = JsonExprPlanner;
+        let qualifier = TableReference::bare("events");
+        let nested_names = vec!["payload".to_string(), "cpu".to_string()];
+
+        let planned = planner.plan_compound_identifier(
+            &json_field("labels"),
+            Some(&qualifier),
+            &nested_names,
+        )?;
+
+        match planned {
+            PlannerResult::Planned(Expr::ScalarFunction(func)) => {
+                assert_eq!(func.func.name(), JsonGetWithType::NAME);
+                assert_eq!(func.args.len(), 2);
+                assert_eq!(
+                    func.args[0],
+                    Expr::Column(Column::new(Some(qualifier.clone()), "labels"))
+                );
+                assert_eq!(
+                    func.args[1],
+                    Expr::Literal(ScalarValue::Utf8(Some("payload.cpu".to_string())), None)
+                );
+            }
+            other => panic!("expected json_get scalar function, got {other:?}"),
+        }
+
+        let original = planner.plan_compound_identifier(
+            &Field::new("plain", DataType::Utf8, true),
+            Some(&qualifier),
+            &nested_names,
+        )?;
+
+        match original {
+            PlannerResult::Original(exprs) => assert!(exprs.is_empty()),
+            other => panic!(
+                "expected original empty result for non-json field, got {:?}",
+                other,
+            ),
+        }
+
+        Ok(())
+    }
+}
diff --git a/src/query/src/datafusion/planner.rs b/src/query/src/datafusion/planner.rs
index d9c74b9d5a..856f2189d2 100644
--- a/src/query/src/datafusion/planner.rs
+++ b/src/query/src/datafusion/planner.rs
@@ -38,6 +38,7 @@ use datafusion_sql::parser::Statement as DfStatement;
 use session::context::QueryContextRef;
 use snafu::{Location, ResultExt};
 
+use crate::datafusion::json_expr_planner::JsonExprPlanner;
 use crate::error::{CatalogSnafu, Result};
 use crate::query_engine::{DefaultPlanDecoder, QueryEngineState};
 
@@ -87,6 +88,9 @@ impl DfContextProviderAdapter {
             .map(|format| (format.get_ext().to_lowercase(), format))
             .collect();
 
+        let mut expr_planners = SessionStateDefaults::default_expr_planners();
+        expr_planners.insert(0, Arc::new(JsonExprPlanner));
+
         Ok(Self {
             engine_state,
             session_state,
@@ -94,7 +98,7 @@ impl DfContextProviderAdapter {
             table_provider,
             query_ctx,
             file_formats,
-            expr_planners: SessionStateDefaults::default_expr_planners(),
+            expr_planners,
         })
     }
 }
diff --git a/src/query/src/dist_plan/analyzer/test.rs b/src/query/src/dist_plan/analyzer/test.rs
index 47848c9ceb..9f72df55a5 100644
--- a/src/query/src/dist_plan/analyzer/test.rs
+++ b/src/query/src/dist_plan/analyzer/test.rs
@@ -126,7 +126,7 @@ impl TestDataSource {
 
 impl DataSource for TestDataSource {
     fn get_stream(&self, request: ScanRequest) -> Result<SendableRecordBatchStream, BoxedError> {
-        let projected_schema = match &request.projection {
+        let projected_schema = match request.projection_indices() {
             Some(projection) => Arc::new(self.schema.try_project(projection).unwrap()),
             None => self.schema.clone(),
         };
diff --git a/src/query/src/dist_plan/analyzer/utils.rs b/src/query/src/dist_plan/analyzer/utils.rs
index e9205e33f4..dd1ad867e5 100644
--- a/src/query/src/dist_plan/analyzer/utils.rs
+++ b/src/query/src/dist_plan/analyzer/utils.rs
@@ -89,7 +89,7 @@ pub fn patch_batch_timezone(
     let patched_columns: Vec<ArrayRef> = expected_schema
         .fields()
         .iter()
-        .zip(columns.into_iter())
+        .zip(columns)
         .map(|(expected_field, column)| {
             let expected_type = expected_field.data_type();
             let actual_type = column.data_type();
diff --git a/src/query/src/dist_plan/merge_scan.rs b/src/query/src/dist_plan/merge_scan.rs
index 2d32ce16b3..470b4d325f 100644
--- a/src/query/src/dist_plan/merge_scan.rs
+++ b/src/query/src/dist_plan/merge_scan.rs
@@ -433,14 +433,15 @@ impl MergeScanExec {
             .values()
             .flat_map(|aliases| aliases.iter().map(|c| c.name()))
             .collect();
-        let mut overlaps = vec![];
-        for expr in &hash_exprs {
-            if let Some(col_expr) = expr.as_any().downcast_ref::<Column>()
-                && all_partition_col_aliases.contains(col_expr.name())
-            {
-                overlaps.push(expr.clone());
-            }
-        }
+        let overlaps: Vec<_> = hash_exprs
+            .iter()
+            .filter(|expr| {
+                expr.as_any()
+                    .downcast_ref::<Column>()
+                    .is_some_and(|col_expr| all_partition_col_aliases.contains(col_expr.name()))
+            })
+            .cloned()
+            .collect();
 
         if overlaps.is_empty() {
             return None;
@@ -476,6 +477,10 @@ impl MergeScanExec {
             .collect()
     }
 
+    pub fn regions(&self) -> &[RegionId] {
+        &self.regions
+    }
+
     pub fn partition_count(&self) -> usize {
         self.target_partition
     }
diff --git a/src/query/src/dummy_catalog.rs b/src/query/src/dummy_catalog.rs
index 239cf7cea8..0ff985ea00 100644
--- a/src/query/src/dummy_catalog.rs
+++ b/src/query/src/dummy_catalog.rs
@@ -15,6 +15,7 @@
 //! Dummy catalog for region server.
 
 use std::any::Any;
+use std::collections::HashMap;
 use std::fmt;
 use std::sync::{Arc, Mutex};
 
@@ -30,6 +31,7 @@ use datafusion::physical_plan::ExecutionPlan;
 use datafusion_common::DataFusionError;
 use datafusion_expr::{Expr, TableProviderFilterPushDown, TableType};
 use datatypes::arrow::datatypes::SchemaRef;
+use datatypes::types::json_type::JsonNativeType;
 use futures::stream::BoxStream;
 use session::context::{QueryContext, QueryContextRef};
 use snafu::ResultExt;
@@ -43,6 +45,7 @@ use table::metadata::{TableId, TableInfoRef};
 use table::table::scan::RegionScanExec;
 
 use crate::error::{GetRegionMetadataSnafu, Result};
+use crate::options::FlowQueryExtensions;
 
 /// Resolve to the given region (specified by [RegionId]) unconditionally.
 #[derive(Clone, Debug)]
@@ -178,17 +181,32 @@ impl TableProvider for DummyTableProvider {
         limit: Option<usize>,
     ) -> datafusion::error::Result<Arc<dyn ExecutionPlan>> {
         let mut request = self.scan_request.lock().unwrap().clone();
-        request.projection = projection.cloned();
+        request.projection_input = projection.map(|p| p.clone().into());
         request.filters = filters.to_vec();
         request.limit = limit;
 
+        if let Some(query_ctx) = &self.query_ctx {
+            let is_sink_scan = is_sink_scan(query_ctx, self.region_id)
+                .map_err(|e| DataFusionError::External(Box::new(e)))?;
+            apply_cached_snapshot_to_request(query_ctx, self.region_id, is_sink_scan, &mut request);
+        }
+
         let scanner = self
             .engine
             .handle_query(self.region_id, request.clone())
             .await
             .map_err(|e| DataFusionError::External(Box::new(e)))?;
-        let query_memory_permit = self.engine.register_query_memory_permit();
-        let mut scan_exec = RegionScanExec::new(scanner, request, query_memory_permit)?;
+
+        if request.snapshot_on_scan
+            && let Some(query_ctx) = &self.query_ctx
+            && let Some(snapshot_sequence) = scanner.snapshot_sequence()
+        {
+            bind_snapshot_bound_region_seq(query_ctx, self.region_id, snapshot_sequence)
+                .map_err(|e| DataFusionError::External(Box::new(e)))?;
+        }
+
+        let query_memory_tracker = self.engine.query_memory_tracker();
+        let mut scan_exec = RegionScanExec::new(scanner, request, query_memory_tracker)?;
         if let Some(query_ctx) = &self.query_ctx {
             scan_exec.set_explain_verbose(query_ctx.explain_verbose());
         }
@@ -270,6 +288,10 @@ impl DummyTableProvider {
         self.scan_request.lock().unwrap().memtable_max_sequence = Some(sequence);
     }
 
+    pub(crate) fn with_json_type_hint(&self, hint: HashMap<String, JsonNativeType>) {
+        self.scan_request.lock().unwrap().json_type_hint = hint;
+    }
+
     /// Gets the scan request of the provider.
     #[cfg(test)]
     pub fn scan_request(&self) -> ScanRequest {
@@ -295,14 +317,11 @@ impl DummyTableProviderFactory {
                     region_id,
                 })?;
 
-        let scan_request = query_ctx
-            .as_ref()
-            .map(|ctx| ScanRequest {
-                memtable_max_sequence: ctx.get_snapshot(region_id.as_u64()),
-                sst_min_sequence: ctx.sst_min_sequence(region_id.as_u64()),
-                ..Default::default()
-            })
-            .unwrap_or_default();
+        let scan_request = if let Some(ctx) = query_ctx.as_ref() {
+            scan_request_from_query_context(region_id, ctx)?
+        } else {
+            ScanRequest::default()
+        };
 
         Ok(DummyTableProvider {
             region_id,
@@ -314,6 +333,152 @@ impl DummyTableProviderFactory {
     }
 }
 
+fn scan_request_from_query_context(
+    region_id: RegionId,
+    query_ctx: &QueryContext,
+) -> Result<ScanRequest> {
+    let decision = decide_flow_scan(query_ctx, region_id)?;
+    Ok(build_scan_request(query_ctx, region_id, &decision))
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct FlowScanDecision {
+    /// Whether this region is the flow sink-table scan.
+    /// Sink scans intentionally bypass incremental and snapshot-binding semantics.
+    is_sink_scan: bool,
+    /// Whether this scan should bind a memtable upper bound when opening the scan.
+    /// This is only the initial intent; if a cached bound already exists in `query_ctx`,
+    /// we reuse that cached bound instead and clear this flag.
+    snapshot_on_scan: bool,
+    /// Optional lower exclusive memtable sequence bound for incremental reads.
+    /// When set, only rows with sequence strictly greater than this bound are read from memtables.
+    memtable_min_sequence: Option<u64>,
+    /// Optional cached per-region snapshot already bound in `query_ctx`.
+    /// When present, this becomes the effective memtable upper bound and suppresses
+    /// binding a new snapshot on scan open.
+    memtable_max_sequence: Option<u64>,
+}
+
+impl FlowScanDecision {
+    fn plain_scan() -> Self {
+        Self {
+            is_sink_scan: true,
+            snapshot_on_scan: false,
+            memtable_min_sequence: None,
+            memtable_max_sequence: None,
+        }
+    }
+}
+
+fn decide_flow_scan(query_ctx: &QueryContext, region_id: RegionId) -> Result<FlowScanDecision> {
+    let Some(flow_extensions) =
+        FlowQueryExtensions::parse_flow_extensions(&query_ctx.extensions())?
+    else {
+        return Ok(FlowScanDecision {
+            is_sink_scan: false,
+            snapshot_on_scan: false,
+            memtable_min_sequence: None,
+            memtable_max_sequence: query_ctx.get_snapshot(region_id.as_u64()),
+        });
+    };
+
+    // Sink-table scans intentionally bypass all flow scan semantics. They should
+    // behave like plain reads and must not participate in incremental lower bounds
+    // or per-region snapshot binding/reuse.
+    if flow_extensions.sink_table_id == Some(region_id.table_id()) {
+        return Ok(FlowScanDecision::plain_scan());
+    }
+
+    let apply_incremental = flow_extensions.validate_for_scan(region_id)?;
+
+    let memtable_min_sequence = if apply_incremental {
+        flow_extensions
+            .incremental_after_seqs
+            .as_ref()
+            .and_then(|seqs| seqs.get(&region_id.as_u64()))
+            .copied()
+    } else {
+        None
+    };
+
+    let memtable_max_sequence = query_ctx.get_snapshot(region_id.as_u64());
+
+    Ok(FlowScanDecision {
+        is_sink_scan: false,
+        snapshot_on_scan: memtable_max_sequence.is_none()
+            && flow_extensions.should_collect_region_watermark(),
+        memtable_min_sequence,
+        memtable_max_sequence,
+    })
+}
+
+fn build_scan_request(
+    query_ctx: &QueryContext,
+    region_id: RegionId,
+    decision: &FlowScanDecision,
+) -> ScanRequest {
+    // Build the initial scan request from the final decision known at provider creation
+    // time. A later scan may still refresh `memtable_max_sequence` if another source scan
+    // has bound a snapshot into `query_ctx` after this provider was created.
+    ScanRequest {
+        sst_min_sequence: (!decision.is_sink_scan)
+            .then(|| query_ctx.sst_min_sequence(region_id.as_u64()))
+            .flatten(),
+        snapshot_on_scan: decision.snapshot_on_scan,
+        memtable_min_sequence: decision.memtable_min_sequence,
+        memtable_max_sequence: decision.memtable_max_sequence,
+        ..Default::default()
+    }
+}
+
+fn is_sink_scan(query_ctx: &QueryContext, region_id: RegionId) -> Result<bool> {
+    Ok(
+        FlowQueryExtensions::parse_flow_extensions(&query_ctx.extensions())?
+            .is_some_and(|exts| exts.sink_table_id == Some(region_id.table_id())),
+    )
+}
+
+fn apply_cached_snapshot_to_request(
+    query_ctx: &QueryContext,
+    region_id: RegionId,
+    is_sink_scan: bool,
+    scan_request: &mut ScanRequest,
+) {
+    if is_sink_scan {
+        return;
+    }
+
+    if let Some(snapshot_sequence) = query_ctx.get_snapshot(region_id.as_u64()) {
+        // Reuse the previously bound per-region snapshot instead of rebinding a new
+        // upper bound on scan open. This refresh is still needed at scan time because
+        // the provider's cached request may have been built before another source scan
+        // bound the shared query-level snapshot into `query_ctx`.
+        scan_request.memtable_max_sequence = Some(snapshot_sequence);
+        scan_request.snapshot_on_scan = false;
+    }
+}
+
+fn bind_snapshot_bound_region_seq(
+    query_ctx: &QueryContext,
+    region_id: RegionId,
+    snapshot_sequence: u64,
+) -> Result<u64> {
+    if let Some(existing) = query_ctx.get_snapshot(region_id.as_u64()) {
+        if existing != snapshot_sequence {
+            return crate::error::ConflictingSnapshotSequenceSnafu {
+                region_id,
+                existing,
+                new: snapshot_sequence,
+            }
+            .fail();
+        }
+        Ok(existing)
+    } else {
+        query_ctx.set_snapshot(region_id.as_u64(), snapshot_sequence);
+        Ok(snapshot_sequence)
+    }
+}
+
 #[async_trait]
 impl TableProviderFactory for DummyTableProviderFactory {
     async fn create(
@@ -443,3 +608,277 @@ impl CatalogManager for DummyCatalogManager {
         Box::pin(futures::stream::empty())
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+    use std::sync::{Arc, RwLock};
+
+    use common_error::ext::ErrorExt;
+    use common_error::status_code::StatusCode;
+    use session::context::QueryContextBuilder;
+
+    use super::*;
+    use crate::error::Error;
+    use crate::options::{FLOW_INCREMENTAL_AFTER_SEQS, FLOW_INCREMENTAL_MODE, FLOW_SINK_TABLE_ID};
+
+    fn test_region_id() -> RegionId {
+        RegionId::new(1024, 1)
+    }
+
+    #[test]
+    fn test_scan_request_from_query_context_uses_snapshot_bound_intent() {
+        let region_id = test_region_id();
+        let query_ctx = QueryContextBuilder::default()
+            .extensions(HashMap::from([(
+                "flow.return_region_seq".to_string(),
+                "true".to_string(),
+            )]))
+            .snapshot_seqs(Arc::new(RwLock::new(HashMap::from([(
+                region_id.as_u64(),
+                42_u64,
+            )]))))
+            .sst_min_sequences(Arc::new(RwLock::new(HashMap::from([(
+                region_id.as_u64(),
+                7_u64,
+            )]))))
+            .build();
+
+        let request = scan_request_from_query_context(region_id, &query_ctx).unwrap();
+
+        assert!(!request.snapshot_on_scan);
+        assert_eq!(request.memtable_max_sequence, Some(42));
+        assert_eq!(request.sst_min_sequence, Some(7));
+    }
+
+    #[test]
+    fn test_scan_request_from_incremental_context_uses_snapshot_bound_intent() {
+        let region_id = test_region_id();
+        let query_ctx = QueryContextBuilder::default()
+            .extensions(HashMap::from([(
+                "flow.incremental_after_seqs".to_string(),
+                format!(r#"{{"{}":10}}"#, region_id.as_u64()),
+            )]))
+            .build();
+
+        let request = scan_request_from_query_context(region_id, &query_ctx).unwrap();
+
+        assert!(request.snapshot_on_scan);
+        assert_eq!(request.memtable_min_sequence, Some(10));
+        assert_eq!(request.memtable_max_sequence, None);
+    }
+
+    #[test]
+    fn test_scan_request_from_query_context_keeps_snapshot_fields() {
+        let region_id = test_region_id();
+        let query_ctx = QueryContextBuilder::default()
+            .snapshot_seqs(Arc::new(RwLock::new(HashMap::from([(
+                region_id.as_u64(),
+                100,
+            )]))))
+            .sst_min_sequences(Arc::new(RwLock::new(HashMap::from([(
+                region_id.as_u64(),
+                90,
+            )]))))
+            .build();
+
+        let request = scan_request_from_query_context(region_id, &query_ctx).unwrap();
+        assert_eq!(request.memtable_max_sequence, Some(100));
+        assert_eq!(request.sst_min_sequence, Some(90));
+        assert_eq!(request.memtable_min_sequence, None);
+        assert!(!request.snapshot_on_scan);
+    }
+
+    #[test]
+    fn test_scan_request_from_query_context_reuses_existing_snapshot_for_incremental_scan() {
+        let region_id = test_region_id();
+        let query_ctx = QueryContextBuilder::default()
+            .extensions(HashMap::from([(
+                FLOW_INCREMENTAL_AFTER_SEQS.to_string(),
+                format!(r#"{{"{}":10}}"#, region_id.as_u64()),
+            )]))
+            .snapshot_seqs(Arc::new(RwLock::new(HashMap::from([(
+                region_id.as_u64(),
+                42_u64,
+            )]))))
+            .build();
+
+        let request = scan_request_from_query_context(region_id, &query_ctx).unwrap();
+
+        assert_eq!(request.memtable_min_sequence, Some(10));
+        assert_eq!(request.memtable_max_sequence, Some(42));
+        assert!(!request.snapshot_on_scan);
+    }
+
+    #[test]
+    fn test_apply_cached_snapshot_to_request_updates_cached_scan_request() {
+        let region_id = test_region_id();
+        let query_ctx = QueryContextBuilder::default()
+            .snapshot_seqs(Arc::new(RwLock::new(HashMap::from([(
+                region_id.as_u64(),
+                88_u64,
+            )]))))
+            .build();
+        let mut request = ScanRequest {
+            snapshot_on_scan: true,
+            ..Default::default()
+        };
+
+        apply_cached_snapshot_to_request(&query_ctx, region_id, false, &mut request);
+
+        assert_eq!(request.memtable_max_sequence, Some(88));
+        assert!(!request.snapshot_on_scan);
+    }
+
+    #[test]
+    fn test_apply_cached_snapshot_to_request_skips_sink_scan() {
+        let region_id = test_region_id();
+        let query_ctx = QueryContextBuilder::default()
+            .snapshot_seqs(Arc::new(RwLock::new(HashMap::from([(
+                region_id.as_u64(),
+                88_u64,
+            )]))))
+            .build();
+        let mut request = ScanRequest {
+            snapshot_on_scan: true,
+            ..Default::default()
+        };
+
+        apply_cached_snapshot_to_request(&query_ctx, region_id, true, &mut request);
+
+        assert_eq!(request.memtable_max_sequence, None);
+        assert!(request.snapshot_on_scan);
+    }
+
+    #[test]
+    fn test_bind_snapshot_bound_region_seq_reuses_existing_snapshot() {
+        let region_id = test_region_id();
+        let query_ctx = QueryContextBuilder::default()
+            .snapshot_seqs(Arc::new(RwLock::new(HashMap::from([(
+                region_id.as_u64(),
+                42_u64,
+            )]))))
+            .build();
+
+        let err = bind_snapshot_bound_region_seq(&query_ctx, region_id, 99).unwrap_err();
+
+        assert!(matches!(err, Error::ConflictingSnapshotSequence { .. }));
+        assert_eq!(query_ctx.get_snapshot(region_id.as_u64()), Some(42));
+    }
+
+    #[test]
+    fn test_bind_snapshot_bound_region_seq_sets_snapshot_once() {
+        let region_id = test_region_id();
+        let query_ctx = QueryContextBuilder::default().build();
+
+        let seq = bind_snapshot_bound_region_seq(&query_ctx, region_id, 99).unwrap();
+
+        assert_eq!(seq, 99);
+        assert_eq!(query_ctx.get_snapshot(region_id.as_u64()), Some(99));
+    }
+
+    #[test]
+    fn test_scan_request_from_query_context_applies_incremental_after_seq_for_source_scan() {
+        let region_id = test_region_id();
+        let query_ctx = QueryContextBuilder::default()
+            .extensions(HashMap::from([
+                (
+                    FLOW_INCREMENTAL_MODE.to_string(),
+                    "memtable_only".to_string(),
+                ),
+                (
+                    FLOW_INCREMENTAL_AFTER_SEQS.to_string(),
+                    format!(r#"{{"{}":55}}"#, region_id.as_u64()),
+                ),
+            ]))
+            .build();
+
+        let request = scan_request_from_query_context(region_id, &query_ctx).unwrap();
+        assert_eq!(request.memtable_min_sequence, Some(55));
+    }
+
+    #[test]
+    fn test_scan_request_from_query_context_does_not_apply_incremental_for_sink_table() {
+        let region_id = test_region_id();
+        let query_ctx = QueryContextBuilder::default()
+            .extensions(HashMap::from([
+                (
+                    FLOW_INCREMENTAL_MODE.to_string(),
+                    "memtable_only".to_string(),
+                ),
+                (
+                    FLOW_INCREMENTAL_AFTER_SEQS.to_string(),
+                    format!(r#"{{"{}":55}}"#, region_id.as_u64()),
+                ),
+                (
+                    FLOW_SINK_TABLE_ID.to_string(),
+                    region_id.table_id().to_string(),
+                ),
+            ]))
+            .snapshot_seqs(Arc::new(RwLock::new(HashMap::from([(
+                region_id.as_u64(),
+                88_u64,
+            )]))))
+            .sst_min_sequences(Arc::new(RwLock::new(HashMap::from([(
+                region_id.as_u64(),
+                77_u64,
+            )]))))
+            .build();
+
+        let request = scan_request_from_query_context(region_id, &query_ctx).unwrap();
+        assert_eq!(request.memtable_min_sequence, None);
+        assert_eq!(request.memtable_max_sequence, None);
+        assert_eq!(request.sst_min_sequence, None);
+        assert!(!request.snapshot_on_scan);
+    }
+
+    #[test]
+    fn test_scan_request_from_query_context_rejects_missing_memtable_only_region() {
+        let region_id = test_region_id();
+        let query_ctx = QueryContextBuilder::default()
+            .extensions(HashMap::from([
+                (
+                    FLOW_INCREMENTAL_MODE.to_string(),
+                    "memtable_only".to_string(),
+                ),
+                (
+                    FLOW_INCREMENTAL_AFTER_SEQS.to_string(),
+                    r#"{"9":55}"#.to_string(),
+                ),
+            ]))
+            .build();
+
+        let err = scan_request_from_query_context(region_id, &query_ctx).unwrap_err();
+        assert!(matches!(err, Error::InvalidQueryContextExtension { .. }));
+    }
+
+    #[test]
+    fn test_scan_request_from_query_context_rejects_invalid_incremental_json() {
+        let region_id = test_region_id();
+        let query_ctx = QueryContextBuilder::default()
+            .extensions(HashMap::from([(
+                FLOW_INCREMENTAL_AFTER_SEQS.to_string(),
+                "not-json".to_string(),
+            )]))
+            .build();
+
+        let err = scan_request_from_query_context(region_id, &query_ctx).unwrap_err();
+        assert!(matches!(err, Error::InvalidQueryContextExtension { .. }));
+        assert_eq!(err.status_code(), StatusCode::InvalidArguments);
+    }
+
+    #[test]
+    fn test_scan_request_from_query_context_rejects_invalid_sink_table_id() {
+        let region_id = test_region_id();
+        let query_ctx = QueryContextBuilder::default()
+            .extensions(HashMap::from([(
+                FLOW_SINK_TABLE_ID.to_string(),
+                "abc".to_string(),
+            )]))
+            .build();
+
+        let err = scan_request_from_query_context(region_id, &query_ctx).unwrap_err();
+        assert!(matches!(err, Error::InvalidQueryContextExtension { .. }));
+        assert_eq!(err.status_code(), StatusCode::InvalidArguments);
+    }
+}
diff --git a/src/query/src/error.rs b/src/query/src/error.rs
index f863a26c4a..6d2a81bd73 100644
--- a/src/query/src/error.rs
+++ b/src/query/src/error.rs
@@ -368,6 +368,27 @@ pub enum Error {
         location: Location,
     },
 
+    #[snafu(display("Invalid query context extension: {}", reason))]
+    InvalidQueryContextExtension {
+        reason: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display(
+        "Conflicting snapshot sequence observed for region {} in a single query (existing={}, new={})",
+        region_id,
+        existing,
+        new
+    ))]
+    ConflictingSnapshotSequence {
+        region_id: RegionId,
+        existing: u64,
+        new: u64,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
     #[snafu(transparent)]
     Datatypes {
         source: datatypes::error::Error,
@@ -399,7 +420,9 @@ impl ErrorExt for Error {
             | ColumnSchemaNoDefault { .. }
             | CteColumnSchemaMismatch { .. }
             | ConvertValue { .. }
-            | TryIntoDuration { .. } => StatusCode::InvalidArguments,
+            | TryIntoDuration { .. }
+            | InvalidQueryContextExtension { .. }
+            | ConflictingSnapshotSequence { .. } => StatusCode::InvalidArguments,
 
             BuildBackend { .. } | ListObjects { .. } => StatusCode::StorageUnavailable,
 
diff --git a/src/query/src/metrics.rs b/src/query/src/metrics.rs
index e0d02e9a3d..7541b191fa 100644
--- a/src/query/src/metrics.rs
+++ b/src/query/src/metrics.rs
@@ -12,16 +12,44 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+use std::collections::BTreeMap;
 use std::pin::Pin;
+use std::sync::Arc;
 use std::task::{Context, Poll};
 
-use common_recordbatch::adapter::RecordBatchMetrics;
+use common_recordbatch::adapter::{RecordBatchMetrics, RegionWatermarkEntry};
 use common_recordbatch::{OrderOption, RecordBatch, RecordBatchStream, SendableRecordBatchStream};
+use common_telemetry::warn;
+use datafusion::physical_plan::ExecutionPlan;
 use datatypes::schema::SchemaRef;
 use futures::Stream;
 use futures_util::ready;
 use lazy_static::lazy_static;
 use prometheus::*;
+use session::context::QueryContextRef;
+
+use crate::dist_plan::MergeScanExec;
+use crate::error::Result;
+use crate::options::FlowQueryExtensions;
+
+/// Intermediate merge state for one participating region while collecting
+/// terminal correctness watermarks across merge-scan sub-stages.
+enum MergeState {
+    /// The region participated, but no explicit watermark result has been seen
+    /// yet for this merge.
+    Participated,
+    /// At least one branch reported that this region cannot prove a safe
+    /// checkpoint watermark for the current query round.
+    Unproved,
+    /// All seen branches agree the region can advance safely to this sequence.
+    Proved(u64),
+    /// Different proved sequences were reported for the same region. The final
+    /// result is degraded to `None`, and the collected values are logged.
+    Conflict {
+        /// Distinct proved watermark candidates reported for the region.
+        watermarks: Vec<u64>,
+    },
+}
 
 lazy_static! {
     /// Timer of different stages in query.
@@ -129,3 +157,350 @@ impl<F: FnOnce() + Unpin> Stream for OnDone<F> {
         self.stream.size_hint()
     }
 }
+
+pub struct RegionWatermarkMetricsStream {
+    stream: SendableRecordBatchStream,
+    plan: Arc<dyn ExecutionPlan>,
+}
+
+impl RegionWatermarkMetricsStream {
+    pub fn new(stream: SendableRecordBatchStream, plan: Arc<dyn ExecutionPlan>) -> Self {
+        Self { stream, plan }
+    }
+}
+
+impl RecordBatchStream for RegionWatermarkMetricsStream {
+    fn name(&self) -> &str {
+        self.stream.name()
+    }
+
+    fn schema(&self) -> SchemaRef {
+        self.stream.schema()
+    }
+
+    fn output_ordering(&self) -> Option<&[OrderOption]> {
+        self.stream.output_ordering()
+    }
+
+    fn metrics(&self) -> Option<RecordBatchMetrics> {
+        let mut metrics = self.stream.metrics()?;
+        let region_watermarks = collect_region_watermarks(self.plan.clone());
+        if !region_watermarks.is_empty() {
+            metrics.region_watermarks = region_watermarks;
+        }
+        Some(metrics)
+    }
+}
+
+impl Stream for RegionWatermarkMetricsStream {
+    type Item = common_recordbatch::error::Result<RecordBatch>;
+
+    fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        Pin::new(&mut self.stream).poll_next(cx)
+    }
+
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        self.stream.size_hint()
+    }
+}
+
+/// Returns whether terminal region watermark metrics should be collected for the query context.
+pub fn should_collect_region_watermark_from_query_ctx(query_ctx: &QueryContextRef) -> Result<bool> {
+    Ok(
+        FlowQueryExtensions::parse_flow_extensions(&query_ctx.extensions())?
+            .is_some_and(|extensions| extensions.should_collect_region_watermark()),
+    )
+}
+
+/// Attaches terminal region watermark metrics to `stream` when collection is requested.
+pub fn maybe_attach_region_watermark_metrics(
+    stream: SendableRecordBatchStream,
+    plan: Arc<dyn ExecutionPlan>,
+    should_collect_region_watermark: bool,
+) -> SendableRecordBatchStream {
+    if should_collect_region_watermark {
+        Box::pin(RegionWatermarkMetricsStream::new(stream, plan))
+    } else {
+        stream
+    }
+}
+
+pub fn terminal_recordbatch_metrics_from_plan(
+    plan: Arc<dyn ExecutionPlan>,
+) -> Option<RecordBatchMetrics> {
+    let region_watermarks = collect_region_watermarks(plan);
+    if region_watermarks.is_empty() {
+        None
+    } else {
+        Some(RecordBatchMetrics {
+            region_watermarks,
+            ..Default::default()
+        })
+    }
+}
+
+/// Collects terminal record-batch metrics from `plan` only when requested.
+pub fn terminal_recordbatch_metrics_from_plan_if_requested(
+    plan: Option<Arc<dyn ExecutionPlan>>,
+    should_collect_region_watermark: bool,
+) -> Option<RecordBatchMetrics> {
+    if should_collect_region_watermark {
+        plan.and_then(terminal_recordbatch_metrics_from_plan)
+    } else {
+        None
+    }
+}
+
+fn collect_region_watermarks(plan: Arc<dyn ExecutionPlan>) -> Vec<RegionWatermarkEntry> {
+    let mut merged = BTreeMap::<u64, MergeState>::new();
+    let mut stack = vec![plan];
+
+    while let Some(plan) = stack.pop() {
+        if let Some(merge_scan) = plan.as_any().downcast_ref::<MergeScanExec>() {
+            merge_merge_scan_region_watermarks(
+                &mut merged,
+                merge_scan
+                    .regions()
+                    .iter()
+                    .map(|region_id| region_id.as_u64()),
+                merge_scan.sub_stage_metrics(),
+            );
+        }
+        stack.extend(plan.children().into_iter().cloned());
+    }
+
+    finalize_region_watermarks(merged)
+}
+
+/// Merge a batch of per-region watermark entries into the global merged state.
+///
+/// # Merge strategy: correctness over maximum
+///
+/// Flow checkpoint advancement requires provable watermarks so that incremental
+/// queries never miss rows. This merge uses correctness-first semantics:
+///
+/// | Current state  | New entry       | Result            | Rationale |
+/// |---------------|-----------------|-------------------|-----------|
+/// | Participated  | Proved(seq)     | Proved(seq)       | First proof for the region |
+/// | Participated  | Unproved         | Unproved          | One branch cannot prove → region is unsafe |
+/// | Proved(old)   | Proved(same)    | Proved(old)       | Convergent proof, keep |
+/// | Proved(old)   | Proved(diff)    | Conflict([old,diff]) | Ambiguous → degrade to unproved |
+/// | Unproved      | _anything_      | Unproved          | Already unsafe, stays unsafe |
+/// | Conflict{..}  | Proved(seq)     | Conflict[...seq]  | Record for diagnostics |
+///
+/// Using `max(old, new)` would be incorrect because it could advance a
+/// checkpoint past rows that a competing MergeScan sub-stage has not yet
+/// scanned, causing Flow to skip data.
+fn merge_region_watermark_entries(
+    merged: &mut BTreeMap<u64, MergeState>,
+    entries: impl IntoIterator<Item = RegionWatermarkEntry>,
+) {
+    for entry in entries {
+        merged
+            .entry(entry.region_id)
+            .and_modify(|existing| match entry.watermark {
+                None => match existing {
+                    MergeState::Participated | MergeState::Proved(_) => {
+                        *existing = MergeState::Unproved;
+                    }
+                    MergeState::Unproved | MergeState::Conflict { .. } => {}
+                },
+                Some(seq) => match existing {
+                    MergeState::Participated => {
+                        *existing = MergeState::Proved(seq);
+                    }
+                    MergeState::Unproved => {}
+                    MergeState::Proved(existing_seq) if *existing_seq == seq => {}
+                    MergeState::Proved(existing_seq) => {
+                        let old_seq = *existing_seq;
+                        *existing = MergeState::Conflict {
+                            watermarks: vec![old_seq, seq],
+                        };
+                    }
+                    MergeState::Conflict { watermarks } => {
+                        if !watermarks.contains(&seq) {
+                            watermarks.push(seq);
+                        }
+                    }
+                },
+            })
+            .or_insert(match entry.watermark {
+                Some(seq) => MergeState::Proved(seq),
+                None => MergeState::Unproved,
+            });
+    }
+}
+
+fn merge_merge_scan_region_watermarks(
+    merged: &mut BTreeMap<u64, MergeState>,
+    regions: impl IntoIterator<Item = u64>,
+    sub_stage_metrics: impl IntoIterator<Item = RecordBatchMetrics>,
+) {
+    // Regions listed by MergeScanExec participated even when no sub-stage can
+    // prove a watermark. Keep them as explicit `None` entries so callers can
+    // distinguish unproved participation from non-participation.
+    for region_id in regions {
+        merged.entry(region_id).or_insert(MergeState::Participated);
+    }
+
+    for metrics in sub_stage_metrics {
+        merge_region_watermark_entries(merged, metrics.region_watermarks);
+    }
+}
+
+fn finalize_region_watermarks(merged: BTreeMap<u64, MergeState>) -> Vec<RegionWatermarkEntry> {
+    merged
+        .into_iter()
+        .map(|(region_id, state)| RegionWatermarkEntry {
+            region_id,
+            watermark: match state {
+                MergeState::Participated => None,
+                MergeState::Unproved => None,
+                MergeState::Proved(seq) => Some(seq),
+                MergeState::Conflict { watermarks } => {
+                    warn!(
+                        "Conflicting proved watermarks for region {}: {:?}; degrading to unproved",
+                        region_id, watermarks
+                    );
+                    None
+                }
+            },
+        })
+        .collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use datafusion::arrow::datatypes::Schema as ArrowSchema;
+    use datafusion::physical_plan::empty::EmptyExec;
+
+    use super::*;
+
+    fn metrics_with_region_watermarks(entries: &[(u64, Option<u64>)]) -> RecordBatchMetrics {
+        RecordBatchMetrics {
+            region_watermarks: entries
+                .iter()
+                .map(|(region_id, watermark)| RegionWatermarkEntry {
+                    region_id: *region_id,
+                    watermark: *watermark,
+                })
+                .collect(),
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn terminal_metrics_returns_none_without_merge_scan() {
+        let plan: Arc<dyn ExecutionPlan> = Arc::new(EmptyExec::new(Arc::new(ArrowSchema::empty())));
+        assert!(terminal_recordbatch_metrics_from_plan(plan).is_none());
+    }
+
+    #[test]
+    fn merge_merge_scan_region_watermarks_marks_missing_watermarks_unproved() {
+        let mut merged = BTreeMap::new();
+
+        merge_merge_scan_region_watermarks(&mut merged, [1, 2], std::iter::empty());
+
+        assert_eq!(
+            finalize_region_watermarks(merged),
+            vec![
+                RegionWatermarkEntry {
+                    region_id: 1,
+                    watermark: None,
+                },
+                RegionWatermarkEntry {
+                    region_id: 2,
+                    watermark: None,
+                },
+            ]
+        );
+    }
+
+    #[test]
+    fn merge_merge_scan_region_watermarks_keeps_matching_proved_values() {
+        let mut merged = BTreeMap::new();
+
+        merge_merge_scan_region_watermarks(
+            &mut merged,
+            [42],
+            [
+                metrics_with_region_watermarks(&[(42, Some(7))]),
+                metrics_with_region_watermarks(&[(42, Some(7))]),
+            ],
+        );
+
+        assert_eq!(
+            finalize_region_watermarks(merged),
+            vec![RegionWatermarkEntry {
+                region_id: 42,
+                watermark: Some(7),
+            }]
+        );
+    }
+
+    #[test]
+    fn merge_merge_scan_region_watermarks_degrades_conflicting_proved_values() {
+        let mut merged = BTreeMap::new();
+
+        merge_merge_scan_region_watermarks(
+            &mut merged,
+            [7],
+            [
+                metrics_with_region_watermarks(&[(7, Some(11))]),
+                metrics_with_region_watermarks(&[(7, Some(13))]),
+            ],
+        );
+
+        assert_eq!(
+            finalize_region_watermarks(merged),
+            vec![RegionWatermarkEntry {
+                region_id: 7,
+                watermark: None,
+            }]
+        );
+    }
+
+    #[test]
+    fn merge_merge_scan_region_watermarks_none_vetoes_proved_value() {
+        let mut merged = BTreeMap::new();
+
+        merge_merge_scan_region_watermarks(
+            &mut merged,
+            [9],
+            [
+                metrics_with_region_watermarks(&[(9, Some(21))]),
+                metrics_with_region_watermarks(&[(9, None)]),
+            ],
+        );
+
+        assert_eq!(
+            finalize_region_watermarks(merged),
+            vec![RegionWatermarkEntry {
+                region_id: 9,
+                watermark: None,
+            }]
+        );
+    }
+
+    #[test]
+    fn merge_merge_scan_region_watermarks_none_vetoes_proved_value_regardless_of_order() {
+        let mut merged = BTreeMap::new();
+
+        merge_merge_scan_region_watermarks(
+            &mut merged,
+            [9],
+            [
+                metrics_with_region_watermarks(&[(9, None)]),
+                metrics_with_region_watermarks(&[(9, Some(21))]),
+            ],
+        );
+
+        assert_eq!(
+            finalize_region_watermarks(merged),
+            vec![RegionWatermarkEntry {
+                region_id: 9,
+                watermark: None,
+            }]
+        );
+    }
+}
diff --git a/src/query/src/optimizer.rs b/src/query/src/optimizer.rs
index 4259b587ba..b85320a495 100644
--- a/src/query/src/optimizer.rs
+++ b/src/query/src/optimizer.rs
@@ -12,8 +12,11 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+pub mod const_normalization;
 pub mod constant_term;
+pub mod count_nest_aggr;
 pub mod count_wildcard;
+pub(crate) mod json_type_concretize;
 pub mod parallelize_scan;
 pub mod pass_distribution;
 pub mod remove_duplicate;
diff --git a/src/query/src/optimizer/const_normalization.rs b/src/query/src/optimizer/const_normalization.rs
new file mode 100644
index 0000000000..fac0e90538
--- /dev/null
+++ b/src/query/src/optimizer/const_normalization.rs
@@ -0,0 +1,1133 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::sync::Arc;
+
+use arrow_schema::{DataType, TimeUnit as ArrowTimeUnit};
+use datafusion::config::ConfigOptions;
+use datafusion_common::tree_node::{Transformed, TreeNode, TreeNodeRecursion, TreeNodeRewriter};
+use datafusion_common::{DFSchemaRef, Result, ScalarValue};
+use datafusion_expr::expr::{Cast, InList, Like, TryCast};
+use datafusion_expr::{Between, BinaryExpr, Expr, ExprSchemable, LogicalPlan, Operator, lit};
+use datafusion_expr_common::casts::try_cast_literal_to_type;
+use datafusion_optimizer::analyzer::AnalyzerRule;
+
+use crate::plan::ExtractExpr;
+
+/// ConstNormalizationRule rewrites castable constants against their
+/// non-constant comparison operand ahead of filter pushdown.
+#[derive(Debug)]
+pub struct ConstNormalizationRule;
+
+impl AnalyzerRule for ConstNormalizationRule {
+    fn analyze(&self, plan: LogicalPlan, _config: &ConfigOptions) -> Result<LogicalPlan> {
+        plan.transform(|plan| match plan {
+            LogicalPlan::Filter(filter) => {
+                let schema = filter.input.schema().clone();
+                rewrite_plan_exprs(LogicalPlan::Filter(filter), schema)
+            }
+            LogicalPlan::TableScan(scan) => {
+                let schema = scan.projected_schema.clone();
+                rewrite_plan_exprs(LogicalPlan::TableScan(scan), schema)
+            }
+            _ => Ok(Transformed::no(plan)),
+        })
+        .map(|x| x.data)
+    }
+
+    fn name(&self) -> &str {
+        "ConstNormalizationRule"
+    }
+}
+
+fn rewrite_plan_exprs(plan: LogicalPlan, schema: DFSchemaRef) -> Result<Transformed<LogicalPlan>> {
+    let mut rewriter = ConstNormalizationRewriter {
+        schema,
+        transformed: false,
+    };
+    let exprs = plan
+        .expressions_consider_join()
+        .into_iter()
+        .map(|expr| expr.rewrite(&mut rewriter).map(|rewritten| rewritten.data))
+        .collect::<Result<Vec<_>>>()?;
+    if !rewriter.transformed {
+        return Ok(Transformed::no(plan));
+    }
+
+    let inputs = plan.inputs().into_iter().cloned().collect::<Vec<_>>();
+    plan.with_new_exprs(exprs, inputs).map(Transformed::yes)
+}
+
+struct ConstNormalizationRewriter {
+    schema: DFSchemaRef,
+    transformed: bool,
+}
+
+impl TreeNodeRewriter for ConstNormalizationRewriter {
+    type Node = Expr;
+
+    fn f_down(&mut self, expr: Expr) -> Result<Transformed<Expr>> {
+        let recursion = if matches!(
+            expr,
+            Expr::Exists(_) | Expr::InSubquery(_) | Expr::ScalarSubquery(_)
+        ) {
+            TreeNodeRecursion::Jump
+        } else {
+            TreeNodeRecursion::Continue
+        };
+
+        Ok(Transformed::new(expr, false, recursion))
+    }
+
+    fn f_up(&mut self, expr: Expr) -> Result<Transformed<Expr>> {
+        let rewritten = rewrite_expr_node(expr, &self.schema)?;
+        self.transformed |= rewritten.transformed;
+        Ok(rewritten)
+    }
+}
+
+fn rewrite_expr_node(expr: Expr, schema: &DFSchemaRef) -> Result<Transformed<Expr>> {
+    match expr {
+        Expr::BinaryExpr(binary) => match rewrite_binary_expr(binary.clone(), schema)? {
+            Some(expr) => Ok(Transformed::yes(expr)),
+            None => Ok(Transformed::no(Expr::BinaryExpr(binary))),
+        },
+        Expr::Between(between) => match rewrite_between_expr(between.clone(), schema)? {
+            Some(expr) => Ok(Transformed::yes(expr)),
+            None => Ok(Transformed::no(Expr::Between(between))),
+        },
+        Expr::InList(in_list) => match rewrite_in_list_expr(in_list.clone(), schema)? {
+            Some(expr) => Ok(Transformed::yes(expr)),
+            None => Ok(Transformed::no(Expr::InList(in_list))),
+        },
+        Expr::Like(like) => rewrite_like_expr(like, PatternMatchKind::Like, schema),
+        Expr::SimilarTo(like) => rewrite_like_expr(like, PatternMatchKind::SimilarTo, schema),
+        expr => Ok(Transformed::no(expr)),
+    }
+}
+
+fn rewrite_between_expr(between: Between, schema: &DFSchemaRef) -> Result<Option<Expr>> {
+    let Between {
+        expr,
+        negated,
+        low,
+        high,
+    } = between;
+    let expr = *expr;
+    let low_expr = *low;
+    let high_expr = *high;
+    let Some((target, constants)) =
+        extract_rewrite_operands(&expr, &[low_expr.clone(), high_expr.clone()], schema)?
+    else {
+        return Ok(None);
+    };
+
+    if let Some(mut constants) = target.normalize_constants(&constants) {
+        let high = constants
+            .pop()
+            .expect("between normalization expects high constant");
+        let low = constants
+            .pop()
+            .expect("between normalization expects low constant");
+        return Ok(Some(Expr::Between(Between {
+            expr: Box::new(target.expr.clone()),
+            negated,
+            low: Box::new(lit(low)),
+            high: Box::new(lit(high)),
+        })));
+    }
+
+    Ok((!negated)
+        .then(|| target.normalize_timestamp_between(&constants[0], &constants[1]))
+        .flatten())
+}
+
+fn rewrite_in_list_expr(in_list: InList, schema: &DFSchemaRef) -> Result<Option<Expr>> {
+    let InList {
+        expr,
+        list,
+        negated,
+    } = in_list;
+    let expr = *expr;
+    let Some((target, constants)) = extract_rewrite_operands(&expr, &list, schema)? else {
+        return Ok(None);
+    };
+
+    Ok(target.normalize_constants(&constants).map(|constants| {
+        target
+            .expr
+            .clone()
+            .in_list(constants.into_iter().map(lit).collect(), negated)
+    }))
+}
+
+fn rewrite_like_expr(
+    like: Like,
+    kind: PatternMatchKind,
+    schema: &DFSchemaRef,
+) -> Result<Transformed<Expr>> {
+    let original = match kind {
+        PatternMatchKind::Like => Expr::Like(like.clone()),
+        PatternMatchKind::SimilarTo => Expr::SimilarTo(like.clone()),
+    };
+    let Like {
+        negated,
+        expr,
+        pattern,
+        escape_char,
+        case_insensitive,
+    } = like;
+    let expr = *expr;
+    let pattern = *pattern;
+    let Some((target, constants)) =
+        extract_rewrite_operands(&expr, std::slice::from_ref(&pattern), schema)?
+    else {
+        return Ok(Transformed::no(original));
+    };
+    let Some(mut constants) = target.normalize_constants(&constants) else {
+        return Ok(Transformed::no(original));
+    };
+
+    let pattern = lit(constants
+        .pop()
+        .expect("pattern normalization expects one constant"));
+    let like = Like::new(
+        negated,
+        Box::new(target.expr.clone()),
+        Box::new(pattern),
+        escape_char,
+        case_insensitive,
+    );
+    let rewritten = match kind {
+        PatternMatchKind::Like => Expr::Like(like),
+        PatternMatchKind::SimilarTo => Expr::SimilarTo(like),
+    };
+    Ok(Transformed::yes(rewritten))
+}
+
+fn rewrite_binary_expr(binary: BinaryExpr, schema: &DFSchemaRef) -> Result<Option<Expr>> {
+    if !binary.op.supports_propagation() {
+        return Ok(None);
+    }
+
+    let BinaryExpr { left, op, right } = binary;
+    let left = *left;
+    let right = *right;
+    if let Some(expr) = rewrite_binary_side(left.clone(), op, right.clone(), schema)? {
+        return Ok(Some(expr));
+    }
+
+    let Some(swapped_op) = op.swap() else {
+        return Ok(None);
+    };
+
+    rewrite_binary_side(right, swapped_op, left, schema)
+}
+
+fn rewrite_binary_side(
+    target_expr: Expr,
+    op: Operator,
+    constant_expr: Expr,
+    schema: &DFSchemaRef,
+) -> Result<Option<Expr>> {
+    let Some((target, constants)) =
+        extract_rewrite_operands(&target_expr, std::slice::from_ref(&constant_expr), schema)?
+    else {
+        return Ok(None);
+    };
+
+    if let Some(mut constants) = target.normalize_constants(&constants) {
+        let constant = constants
+            .pop()
+            .expect("binary normalization expects one constant");
+        return Ok(Some(Expr::BinaryExpr(BinaryExpr {
+            left: Box::new(target.expr.clone()),
+            op,
+            right: Box::new(lit(constant)),
+        })));
+    }
+
+    Ok(target.normalize_timestamp_binary(op, &constants[0]))
+}
+
+fn extract_rewrite_operands(
+    target_expr: &Expr,
+    constant_exprs: &[Expr],
+    schema: &DFSchemaRef,
+) -> Result<Option<(NormalizationTarget, Vec<ScalarValue>)>> {
+    let Some(target) = extract_normalization_target(target_expr, schema)? else {
+        return Ok(None);
+    };
+
+    extract_constant_scalars(constant_exprs)
+        .map(|constants| constants.map(|constants| (target, constants)))
+}
+
+#[derive(Clone)]
+struct NormalizationTarget {
+    expr: Expr,
+    data_type: DataType,
+    kind: NormalizationKind,
+}
+
+#[derive(Clone)]
+enum NormalizationKind {
+    /// The cast preserves every source value exactly, so literals can be cast directly.
+    Lossless,
+    /// The cast drops timestamp precision and must widen predicate bounds to preserve semantics.
+    TimestampDowncast {
+        source_unit: ArrowTimeUnit,
+        target_unit: ArrowTimeUnit,
+        timezone: Option<Arc<str>>,
+    },
+}
+
+impl NormalizationTarget {
+    /// Normalizes constants for rewrites that can preserve the original predicate with a direct
+    /// literal cast. Timestamp precision-changing casts are handled by timestamp-specific helpers.
+    fn normalize_constants(&self, constants: &[ScalarValue]) -> Option<Vec<ScalarValue>> {
+        constants
+            .iter()
+            .map(|constant| self.normalize_constant(constant))
+            .collect()
+    }
+
+    fn normalize_constant(&self, constant: &ScalarValue) -> Option<ScalarValue> {
+        match self.kind {
+            NormalizationKind::TimestampDowncast { .. } => None,
+            NormalizationKind::Lossless => try_cast_literal_to_type(constant, &self.data_type),
+        }
+    }
+
+    /// Rewrites predicates over timestamp downcasts into source-side half-open bounds.
+    fn normalize_timestamp_binary(&self, op: Operator, constant: &ScalarValue) -> Option<Expr> {
+        let NormalizationKind::TimestampDowncast {
+            source_unit,
+            target_unit,
+            timezone,
+        } = &self.kind
+        else {
+            return None;
+        };
+
+        let constant = constant
+            .cast_to(&DataType::Timestamp(*target_unit, timezone.clone()))
+            .ok()?;
+        let value = timestamp_scalar_value(&constant)?;
+        let bound = match op {
+            Operator::GtEq => lower_bound_for_ge(value, *source_unit, *target_unit)?,
+            Operator::Gt => lower_bound_for_ge(value.checked_add(1)?, *source_unit, *target_unit)?,
+            Operator::Lt => lower_bound_for_ge(value, *source_unit, *target_unit)?,
+            Operator::LtEq => {
+                lower_bound_for_ge(value.checked_add(1)?, *source_unit, *target_unit)?
+            }
+            _ => return None,
+        };
+
+        let normalized_op = match op {
+            Operator::GtEq | Operator::Gt => Operator::GtEq,
+            Operator::Lt | Operator::LtEq => Operator::Lt,
+            _ => return None,
+        };
+
+        Some(match normalized_op {
+            Operator::GtEq => self.expr.clone().gt_eq(lit(timestamp_scalar(
+                *source_unit,
+                timezone.clone(),
+                bound,
+            ))),
+            Operator::Lt => {
+                self.expr
+                    .clone()
+                    .lt(lit(timestamp_scalar(*source_unit, timezone.clone(), bound)))
+            }
+            _ => unreachable!("timestamp normalization only rewrites to >= or <"),
+        })
+    }
+
+    /// Rewrites `BETWEEN` over timestamp downcasts into an inclusive lower bound and exclusive
+    /// upper bound over the source timestamp unit.
+    fn normalize_timestamp_between(&self, low: &ScalarValue, high: &ScalarValue) -> Option<Expr> {
+        let NormalizationKind::TimestampDowncast {
+            source_unit,
+            target_unit,
+            timezone,
+        } = &self.kind
+        else {
+            return None;
+        };
+
+        let target_type = DataType::Timestamp(*target_unit, timezone.clone());
+        let low = low.cast_to(&target_type).ok()?;
+        let high = high.cast_to(&target_type).ok()?;
+        let low = timestamp_scalar_value(&low)?;
+        let high = timestamp_scalar_value(&high)?;
+
+        let lower = lower_bound_for_ge(low, *source_unit, *target_unit)?;
+        let upper = lower_bound_for_ge(high.checked_add(1)?, *source_unit, *target_unit)?;
+
+        Some(
+            self.expr
+                .clone()
+                .gt_eq(lit(timestamp_scalar(*source_unit, timezone.clone(), lower)))
+                .and(self.expr.clone().lt(lit(timestamp_scalar(
+                    *source_unit,
+                    timezone.clone(),
+                    upper,
+                )))),
+        )
+    }
+}
+
+/// Returns the non-constant side we should normalize against.
+///
+/// Plain expressions normalize literals to their own type. Cast expressions only participate when
+/// the cast is lossless or when timestamp downcasts can be rewritten as wider source-side bounds.
+fn extract_normalization_target(
+    expr: &Expr,
+    schema: &DFSchemaRef,
+) -> Result<Option<NormalizationTarget>> {
+    if extract_constant_scalar(expr)?.is_some() {
+        return Ok(None);
+    }
+
+    let Some((_, source_expr, target_type)) = extract_cast_input(expr) else {
+        return Ok(Some(NormalizationTarget {
+            expr: expr.clone(),
+            data_type: expr.get_type(schema)?,
+            kind: NormalizationKind::Lossless,
+        }));
+    };
+
+    let data_type = source_expr.get_type(schema)?;
+    let Some(kind) = classify_normalization_kind(&data_type, target_type) else {
+        return Ok(None);
+    };
+
+    Ok(Some(NormalizationTarget {
+        expr: source_expr.clone(),
+        data_type,
+        kind,
+    }))
+}
+
+fn classify_normalization_kind(
+    source_type: &DataType,
+    target_type: &DataType,
+) -> Option<NormalizationKind> {
+    // Timestamp casts that change precision need boundary-aware rewrites. A finer target literal
+    // may not map exactly back to the coarser source unit, so the generic lossless path is only
+    // safe for timestamp casts that keep the same unit.
+    if is_lossless_cast(source_type, target_type) {
+        return Some(NormalizationKind::Lossless);
+    }
+
+    match (source_type, target_type) {
+        (
+            DataType::Timestamp(source_unit, source_tz),
+            DataType::Timestamp(target_unit, target_tz),
+        ) if source_tz == target_tz
+            && time_unit_rank(*source_unit) > time_unit_rank(*target_unit) =>
+        {
+            Some(NormalizationKind::TimestampDowncast {
+                source_unit: *source_unit,
+                target_unit: *target_unit,
+                timezone: source_tz.clone(),
+            })
+        }
+        _ => None,
+    }
+}
+
+/// Returns whether every value of `source_type` is representable in `target_type`.
+fn is_lossless_cast(source_type: &DataType, target_type: &DataType) -> bool {
+    match (source_type, target_type) {
+        (DataType::Int8, DataType::Int16 | DataType::Int32 | DataType::Int64)
+        | (DataType::Int16, DataType::Int32 | DataType::Int64)
+        | (DataType::Int32, DataType::Int64)
+        | (DataType::UInt8, DataType::UInt16 | DataType::UInt32 | DataType::UInt64)
+        | (DataType::UInt8, DataType::Int16 | DataType::Int32 | DataType::Int64)
+        | (DataType::UInt16, DataType::UInt32 | DataType::UInt64)
+        | (DataType::UInt16, DataType::Int32 | DataType::Int64)
+        | (DataType::UInt32, DataType::UInt64 | DataType::Int64)
+        | (DataType::Utf8, DataType::Utf8View | DataType::LargeUtf8) => true,
+        (
+            DataType::Timestamp(source_unit, source_tz),
+            DataType::Timestamp(target_unit, target_tz),
+        ) => source_tz == target_tz && source_unit == target_unit,
+        _ => false,
+    }
+}
+
+#[derive(Clone, Copy)]
+enum PatternMatchKind {
+    Like,
+    SimilarTo,
+}
+
+fn extract_constant_scalars(exprs: &[Expr]) -> Result<Option<Vec<ScalarValue>>> {
+    let mut values = Vec::with_capacity(exprs.len());
+    for expr in exprs {
+        let Some(value) = extract_constant_scalar(expr)? else {
+            return Ok(None);
+        };
+        values.push(value);
+    }
+
+    Ok(Some(values))
+}
+
+/// Extracts a literal scalar from an expression, folding constant `CAST` and `TRY_CAST` nodes.
+fn extract_constant_scalar(expr: &Expr) -> Result<Option<ScalarValue>> {
+    if let Some(value) = expr.as_literal() {
+        return Ok(Some(value.clone()));
+    }
+
+    let Some((kind, expr, data_type)) = extract_cast_input(expr) else {
+        return Ok(None);
+    };
+
+    match kind {
+        CastInputKind::Cast => extract_constant_scalar(expr)?
+            .map(|value| value.cast_to(data_type))
+            .transpose(),
+        CastInputKind::TryCast => {
+            Ok(extract_constant_scalar(expr)?.and_then(|value| value.cast_to(data_type).ok()))
+        }
+    }
+}
+
+#[derive(Clone, Copy)]
+enum CastInputKind {
+    Cast,
+    TryCast,
+}
+
+/// Returns the input expression and target type for `CAST` and `TRY_CAST` expressions.
+fn extract_cast_input(expr: &Expr) -> Option<(CastInputKind, &Expr, &DataType)> {
+    match expr {
+        Expr::Cast(Cast { expr, data_type }) => {
+            Some((CastInputKind::Cast, expr.as_ref(), data_type))
+        }
+        Expr::TryCast(TryCast { expr, data_type }) => {
+            Some((CastInputKind::TryCast, expr.as_ref(), data_type))
+        }
+        _ => None,
+    }
+}
+
+fn time_unit_rank(unit: ArrowTimeUnit) -> usize {
+    match unit {
+        ArrowTimeUnit::Second => 0,
+        ArrowTimeUnit::Millisecond => 1,
+        ArrowTimeUnit::Microsecond => 2,
+        ArrowTimeUnit::Nanosecond => 3,
+    }
+}
+
+fn time_unit_scale(unit: ArrowTimeUnit) -> i64 {
+    match unit {
+        ArrowTimeUnit::Second => 1,
+        ArrowTimeUnit::Millisecond => 1_000,
+        ArrowTimeUnit::Microsecond => 1_000_000,
+        ArrowTimeUnit::Nanosecond => 1_000_000_000,
+    }
+}
+
+/// Returns the number of source-unit ticks in one target-unit tick for finer-to-coarser casts.
+fn finer_to_coarser_ratio(source_unit: ArrowTimeUnit, target_unit: ArrowTimeUnit) -> Option<i64> {
+    let source_scale = time_unit_scale(source_unit);
+    let target_scale = time_unit_scale(target_unit);
+    (source_scale >= target_scale).then_some(source_scale / target_scale)
+}
+
+/// Returns the smallest source-unit timestamp whose downcast is greater than or equal to
+/// `target_value`.
+///
+/// DataFusion timestamp downcasts truncate toward zero. For non-positive buckets that means the
+/// bucket starts before `target_value * ratio`, so `<= x` can be rewritten as `< lower_bound(x+1)`
+/// without dropping rows near zero or across negative boundaries.
+fn lower_bound_for_ge(
+    target_value: i64,
+    source_unit: ArrowTimeUnit,
+    target_unit: ArrowTimeUnit,
+) -> Option<i64> {
+    let ratio = finer_to_coarser_ratio(source_unit, target_unit)?;
+    let base = target_value.checked_mul(ratio)?;
+    if target_value <= 0 {
+        base.checked_sub(ratio - 1)
+    } else {
+        Some(base)
+    }
+}
+
+fn timestamp_scalar_value(value: &ScalarValue) -> Option<i64> {
+    match value {
+        ScalarValue::TimestampSecond(Some(value), _)
+        | ScalarValue::TimestampMillisecond(Some(value), _)
+        | ScalarValue::TimestampMicrosecond(Some(value), _)
+        | ScalarValue::TimestampNanosecond(Some(value), _) => Some(*value),
+        _ => None,
+    }
+}
+
+fn timestamp_scalar(unit: ArrowTimeUnit, timezone: Option<Arc<str>>, value: i64) -> ScalarValue {
+    match unit {
+        ArrowTimeUnit::Second => ScalarValue::TimestampSecond(Some(value), timezone),
+        ArrowTimeUnit::Millisecond => ScalarValue::TimestampMillisecond(Some(value), timezone),
+        ArrowTimeUnit::Microsecond => ScalarValue::TimestampMicrosecond(Some(value), timezone),
+        ArrowTimeUnit::Nanosecond => ScalarValue::TimestampNanosecond(Some(value), timezone),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use arrow_schema::{DataType, TimeUnit as ArrowTimeUnit};
+    use async_trait::async_trait;
+    use common_time::Timestamp;
+    use common_time::range::TimestampRange;
+    use common_time::timestamp::TimeUnit;
+    use datafusion::catalog::Session;
+    use datafusion::config::ConfigOptions;
+    use datafusion::datasource::{TableProvider, provider_as_source};
+    use datafusion::physical_plan::ExecutionPlan;
+    use datafusion_common::arrow::datatypes::Field;
+    use datafusion_common::{DFSchema, ScalarValue, ToDFSchema};
+    use datafusion_expr::expr::{Between, Like};
+    use datafusion_expr::expr_fn::{cast, col, try_cast};
+    use datafusion_expr::{
+        Expr, LogicalPlan, LogicalPlanBuilder, TableProviderFilterPushDown, TableScan, TableSource,
+        TableType, lit,
+    };
+    use datafusion_optimizer::analyzer::AnalyzerRule;
+    use datafusion_optimizer::optimizer::{Optimizer, OptimizerContext};
+    use datafusion_optimizer::push_down_filter::PushDownFilter;
+    use table::predicate::build_time_range_predicate;
+
+    use super::{
+        ConstNormalizationRule, PatternMatchKind, lower_bound_for_ge, try_cast_literal_to_type,
+    };
+
+    #[test]
+    fn test_normalize_direct_integer_cast_comparison() {
+        assert_filter_plan(
+            vec![Field::new("v", DataType::Int32, false)],
+            cast(col("v"), DataType::Int64).gt_eq(lit(42_i64)),
+            "Filter: t.v >= Int32(42)\n  TableScan: t",
+        );
+    }
+
+    #[test]
+    fn test_normalize_non_column_operand() {
+        assert_filter_plan(
+            vec![Field::new("v", DataType::Int32, false)],
+            cast(col("v") + lit(1_i32), DataType::Int64).gt_eq(lit(42_i64)),
+            "Filter: t.v + Int32(1) >= Int32(42)\n  TableScan: t",
+        );
+    }
+
+    #[test]
+    fn test_normalize_swapped_binary_comparison() {
+        assert_filter_plan(
+            vec![Field::new("v", DataType::Int16, false)],
+            lit(42_i64).lt_eq(cast(col("v"), DataType::Int64)),
+            "Filter: t.v >= Int16(42)\n  TableScan: t",
+        );
+    }
+
+    #[test]
+    fn test_normalize_try_cast_target() {
+        assert_filter_plan(
+            vec![Field::new("v", DataType::Int16, false)],
+            try_cast(col("v"), DataType::Int64).gt_eq(lit(42_i64)),
+            "Filter: t.v >= Int16(42)\n  TableScan: t",
+        );
+    }
+
+    #[test]
+    fn test_normalize_casted_constants() {
+        let fields = vec![Field::new("v", DataType::Int16, false)];
+        let cases = [
+            (
+                col("v").gt_eq(cast(lit(42_i8), DataType::Int64)),
+                "Filter: t.v >= Int16(42)\n  TableScan: t",
+            ),
+            (
+                col("v").in_list(
+                    vec![
+                        cast(lit(1_i8), DataType::Int64),
+                        try_cast(lit(2_i8), DataType::Int64),
+                    ],
+                    false,
+                ),
+                "Filter: t.v IN ([Int16(1), Int16(2)])\n  TableScan: t",
+            ),
+        ];
+
+        for (predicate, expected) in cases {
+            assert_filter_plan(fields.clone(), predicate, expected);
+        }
+    }
+
+    #[test]
+    fn test_normalize_plain_integer_literals() {
+        let fields = vec![Field::new("v", DataType::Int16, false)];
+        let cases = [
+            (
+                col("v").gt_eq(lit(42_i64)),
+                "Filter: t.v >= Int16(42)\n  TableScan: t",
+            ),
+            (
+                col("v").in_list(vec![lit(1_i64), lit(2_i64)], false),
+                "Filter: t.v IN ([Int16(1), Int16(2)])\n  TableScan: t",
+            ),
+            (
+                col("v").between(lit(3_i64), lit(5_i64)),
+                "Filter: t.v BETWEEN Int16(3) AND Int16(5)\n  TableScan: t",
+            ),
+        ];
+
+        for (predicate, expected) in cases {
+            assert_filter_plan(fields.clone(), predicate, expected);
+        }
+    }
+
+    #[test]
+    fn test_normalize_unsigned_to_signed_literals() {
+        let cases = [
+            (
+                vec![Field::new("v", DataType::UInt8, false)],
+                cast(col("v"), DataType::Int16).lt_eq(lit(255_i16)),
+                "Filter: t.v <= UInt8(255)\n  TableScan: t",
+            ),
+            (
+                vec![Field::new("v", DataType::UInt16, false)],
+                cast(col("v"), DataType::Int32).gt_eq(lit(42_i32)),
+                "Filter: t.v >= UInt16(42)\n  TableScan: t",
+            ),
+            (
+                vec![Field::new("v", DataType::UInt32, false)],
+                cast(col("v"), DataType::Int64).between(lit(3_i64), lit(5_i64)),
+                "Filter: t.v BETWEEN UInt32(3) AND UInt32(5)\n  TableScan: t",
+            ),
+        ];
+
+        for (fields, predicate, expected) in cases {
+            assert_filter_plan(fields, predicate, expected);
+        }
+    }
+
+    #[test]
+    fn test_normalize_in_list_and_between() {
+        let fields = vec![Field::new("v", DataType::Int16, false)];
+        let cases = [
+            (
+                cast(col("v"), DataType::Int64).in_list(vec![lit(1_i64), lit(2_i64)], false),
+                "Filter: t.v IN ([Int16(1), Int16(2)])\n  TableScan: t",
+            ),
+            (
+                cast(col("v"), DataType::Int64).between(lit(3_i64), lit(5_i64)),
+                "Filter: t.v BETWEEN Int16(3) AND Int16(5)\n  TableScan: t",
+            ),
+        ];
+
+        for (predicate, expected) in cases {
+            assert_filter_plan(fields.clone(), predicate, expected);
+        }
+    }
+
+    #[test]
+    fn test_keep_non_lossless_literal_unchanged() {
+        assert_filter_plan(
+            vec![Field::new("v", DataType::Int16, false)],
+            col("v").gt_eq(lit(100_000_i64)),
+            "Filter: t.v >= Int64(100000)\n  TableScan: t",
+        );
+    }
+
+    #[test]
+    fn test_normalize_scan_filters() {
+        let scan = build_scan_plan(test_schema(vec![Field::new("v", DataType::Int16, false)]));
+        let LogicalPlan::TableScan(scan) = scan else {
+            panic!("expected table scan");
+        };
+        let plan = LogicalPlan::TableScan(TableScan {
+            filters: vec![cast(col("v"), DataType::Int64).gt_eq(lit(42_i64))],
+            ..scan
+        });
+
+        let analyzed = analyze_plan(plan);
+
+        assert_eq!(
+            vec![col("v").gt_eq(lit(42_i16))],
+            extract_scan_filters(&analyzed)
+        );
+    }
+
+    #[test]
+    fn test_normalize_negated_between() {
+        assert_filter_plan(
+            vec![Field::new("v", DataType::Int16, false)],
+            Expr::Between(Between {
+                expr: Box::new(cast(col("v"), DataType::Int64)),
+                negated: true,
+                low: Box::new(lit(3_i64)),
+                high: Box::new(lit(5_i64)),
+            }),
+            "Filter: t.v NOT BETWEEN Int16(3) AND Int16(5)\n  TableScan: t",
+        );
+    }
+
+    #[test]
+    fn test_normalize_like_literal() {
+        assert_pattern_match_plan(
+            PatternMatchKind::Like,
+            ScalarValue::LargeUtf8(Some("api%".to_string())),
+            "Filter: t.s LIKE Utf8(\"api%\")\n  TableScan: t",
+        );
+    }
+
+    #[test]
+    fn test_normalize_similar_to_literal() {
+        assert_pattern_match_plan(
+            PatternMatchKind::SimilarTo,
+            ScalarValue::LargeUtf8(Some("api.*".to_string())),
+            "Filter: t.s SIMILAR TO Utf8(\"api.*\")\n  TableScan: t",
+        );
+    }
+
+    #[test]
+    fn test_normalize_direct_timestamp_filter() {
+        assert_timestamp_pushdown(
+            vec![
+                Field::new(
+                    "ts",
+                    DataType::Timestamp(ArrowTimeUnit::Nanosecond, None),
+                    false,
+                ),
+                Field::new("tag", DataType::Utf8, true),
+            ],
+            ts_cast_to_ms()
+                .gt_eq(ts_ms_literal(-299_999))
+                .and(ts_cast_to_ms().lt_eq(ts_ms_literal(10_000)))
+                .and(col("tag").eq(lit("api"))),
+            "Filter: t.ts >= TimestampNanosecond(-299999999999, None) AND t.ts < TimestampNanosecond(10001000000, None) AND t.tag = Utf8(\"api\")\n  TableScan: t",
+            "TableScan: t, full_filters=[t.ts >= TimestampNanosecond(-299999999999, None), t.ts < TimestampNanosecond(10001000000, None), t.tag = Utf8(\"api\")]",
+            TimestampRange::new_inclusive(
+                Some(Timestamp::new_nanosecond(-299_999_999_999)),
+                Some(Timestamp::new_nanosecond(10_000_999_999)),
+            ),
+        );
+    }
+
+    #[test]
+    fn test_normalize_timestamp_between_filter() {
+        assert_timestamp_pushdown(
+            vec![Field::new(
+                "ts",
+                DataType::Timestamp(ArrowTimeUnit::Nanosecond, None),
+                false,
+            )],
+            ts_cast_to_ms().between(ts_ms_literal(-299_999), ts_ms_literal(10_000)),
+            "Filter: t.ts >= TimestampNanosecond(-299999999999, None) AND t.ts < TimestampNanosecond(10001000000, None)\n  TableScan: t",
+            "TableScan: t, full_filters=[t.ts >= TimestampNanosecond(-299999999999, None), t.ts < TimestampNanosecond(10001000000, None)]",
+            TimestampRange::new_inclusive(
+                Some(Timestamp::new_nanosecond(-299_999_999_999)),
+                Some(Timestamp::new_nanosecond(10_000_999_999)),
+            ),
+        );
+    }
+
+    #[test]
+    fn test_normalize_strict_timestamp_filter() {
+        assert_timestamp_pushdown(
+            vec![Field::new(
+                "ts",
+                DataType::Timestamp(ArrowTimeUnit::Nanosecond, None),
+                false,
+            )],
+            ts_cast_to_ms()
+                .gt(ts_ms_literal(10_000))
+                .and(ts_cast_to_ms().lt(ts_ms_literal(20_000))),
+            "Filter: t.ts >= TimestampNanosecond(10001000000, None) AND t.ts < TimestampNanosecond(20000000000, None)\n  TableScan: t",
+            "TableScan: t, full_filters=[t.ts >= TimestampNanosecond(10001000000, None), t.ts < TimestampNanosecond(20000000000, None)]",
+            TimestampRange::new_inclusive(
+                Some(Timestamp::new_nanosecond(10_001_000_000)),
+                Some(Timestamp::new_nanosecond(19_999_999_999)),
+            ),
+        );
+    }
+
+    #[test]
+    fn test_normalize_zero_boundary_timestamp_filter() {
+        let fields = vec![Field::new(
+            "ts",
+            DataType::Timestamp(ArrowTimeUnit::Nanosecond, None),
+            false,
+        )];
+
+        assert_timestamp_pushdown(
+            fields.clone(),
+            ts_cast_to_ms().gt_eq(ts_ms_literal(0)),
+            "Filter: t.ts >= TimestampNanosecond(-999999, None)\n  TableScan: t",
+            "TableScan: t, full_filters=[t.ts >= TimestampNanosecond(-999999, None)]",
+            TimestampRange::from_start(Timestamp::new_nanosecond(-999_999)),
+        );
+
+        assert_timestamp_pushdown(
+            fields.clone(),
+            ts_cast_to_ms().lt(ts_ms_literal(0)),
+            "Filter: t.ts < TimestampNanosecond(-999999, None)\n  TableScan: t",
+            "TableScan: t, full_filters=[t.ts < TimestampNanosecond(-999999, None)]",
+            TimestampRange::until_end(Timestamp::new_nanosecond(-999_999), false),
+        );
+
+        assert_timestamp_pushdown(
+            fields,
+            ts_cast_to_ms().between(ts_ms_literal(0), ts_ms_literal(0)),
+            "Filter: t.ts >= TimestampNanosecond(-999999, None) AND t.ts < TimestampNanosecond(1000000, None)\n  TableScan: t",
+            "TableScan: t, full_filters=[t.ts >= TimestampNanosecond(-999999, None), t.ts < TimestampNanosecond(1000000, None)]",
+            TimestampRange::new_inclusive(
+                Some(Timestamp::new_nanosecond(-999_999)),
+                Some(Timestamp::new_nanosecond(999_999)),
+            ),
+        );
+    }
+
+    #[test]
+    fn test_timestamp_downcast_contract_matches_datafusion_casts() {
+        let cases = [
+            (-1_000_001, -1),
+            (-1_000_000, -1),
+            (-999_999, 0),
+            (-1, 0),
+            (0, 0),
+            (999_999, 0),
+            (1_000_000, 1),
+        ];
+
+        for (source, expected) in cases {
+            let casted = try_cast_literal_to_type(
+                &ScalarValue::TimestampNanosecond(Some(source), None),
+                &DataType::Timestamp(ArrowTimeUnit::Millisecond, None),
+            )
+            .unwrap();
+            assert_eq!(
+                ScalarValue::TimestampMillisecond(Some(expected), None),
+                casted
+            );
+        }
+
+        assert_eq!(
+            Some(-1_999_999),
+            lower_bound_for_ge(-1, ArrowTimeUnit::Nanosecond, ArrowTimeUnit::Millisecond)
+        );
+        assert_eq!(
+            Some(-999_999),
+            lower_bound_for_ge(0, ArrowTimeUnit::Nanosecond, ArrowTimeUnit::Millisecond)
+        );
+        assert_eq!(
+            Some(1_000_000),
+            lower_bound_for_ge(1, ArrowTimeUnit::Nanosecond, ArrowTimeUnit::Millisecond)
+        );
+    }
+
+    #[test]
+    fn test_normalize_plain_timestamp_literals() {
+        assert_timestamp_pushdown(
+            vec![Field::new(
+                "ts",
+                DataType::Timestamp(ArrowTimeUnit::Nanosecond, None),
+                false,
+            )],
+            col("ts")
+                .gt_eq(ts_ms_literal(-299_999))
+                .and(col("ts").lt_eq(ts_ms_literal(10_000))),
+            "Filter: t.ts >= TimestampNanosecond(-299999000000, None) AND t.ts <= TimestampNanosecond(10000000000, None)\n  TableScan: t",
+            "TableScan: t, full_filters=[t.ts >= TimestampNanosecond(-299999000000, None), t.ts <= TimestampNanosecond(10000000000, None)]",
+            TimestampRange::new_inclusive(
+                Some(Timestamp::new_nanosecond(-299_999_000_000)),
+                Some(Timestamp::new_nanosecond(10_000_000_000)),
+            ),
+        );
+    }
+
+    #[test]
+    fn test_keep_timestamp_upcast_filter_unchanged() {
+        assert_filter_plan(
+            vec![Field::new(
+                "ts",
+                DataType::Timestamp(ArrowTimeUnit::Millisecond, None),
+                false,
+            )],
+            cast(
+                col("ts"),
+                DataType::Timestamp(ArrowTimeUnit::Nanosecond, None),
+            )
+            .gt_eq(lit(ScalarValue::TimestampNanosecond(Some(1), None))),
+            "Filter: CAST(t.ts AS Timestamp(ns)) >= TimestampNanosecond(1, None)\n  TableScan: t",
+        );
+    }
+
+    fn assert_pattern_match_plan(kind: PatternMatchKind, pattern: ScalarValue, expected: &str) {
+        let predicate = match kind {
+            PatternMatchKind::Like => Expr::Like(Like::new(
+                false,
+                Box::new(cast(col("s"), DataType::LargeUtf8)),
+                Box::new(lit(pattern)),
+                None,
+                false,
+            )),
+            PatternMatchKind::SimilarTo => Expr::SimilarTo(Like::new(
+                false,
+                Box::new(cast(col("s"), DataType::LargeUtf8)),
+                Box::new(lit(pattern)),
+                None,
+                false,
+            )),
+        };
+
+        assert_filter_plan(
+            vec![Field::new("s", DataType::Utf8, false)],
+            predicate,
+            expected,
+        );
+    }
+
+    fn assert_filter_plan(fields: Vec<Field>, predicate: Expr, expected: &str) {
+        assert_eq!(expected, analyze_filter(fields, predicate).to_string());
+    }
+
+    fn assert_timestamp_pushdown(
+        fields: Vec<Field>,
+        predicate: Expr,
+        expected_analyzed: &str,
+        expected_pushed: &str,
+        expected_range: TimestampRange,
+    ) {
+        let analyzed = analyze_filter(fields, predicate);
+        assert_eq!(expected_analyzed, analyzed.to_string());
+
+        let pushed = push_down_filters(analyzed);
+        assert_eq!(expected_pushed, pushed.to_string());
+
+        let range =
+            build_time_range_predicate("ts", TimeUnit::Nanosecond, &extract_scan_filters(&pushed));
+        assert_eq!(expected_range, range);
+    }
+
+    fn analyze_filter(fields: Vec<Field>, predicate: Expr) -> LogicalPlan {
+        analyze_plan(build_filter_plan(test_schema(fields), predicate))
+    }
+
+    fn analyze_plan(plan: LogicalPlan) -> LogicalPlan {
+        ConstNormalizationRule
+            .analyze(plan, &ConfigOptions::default())
+            .unwrap()
+    }
+
+    fn build_filter_plan(schema: Arc<DFSchema>, predicate: Expr) -> LogicalPlan {
+        LogicalPlanBuilder::scan("t", test_source(schema), None)
+            .unwrap()
+            .filter(predicate)
+            .unwrap()
+            .build()
+            .unwrap()
+    }
+
+    fn build_scan_plan(schema: Arc<DFSchema>) -> LogicalPlan {
+        LogicalPlanBuilder::scan("t", test_source(schema), None)
+            .unwrap()
+            .build()
+            .unwrap()
+    }
+
+    fn push_down_filters(plan: LogicalPlan) -> LogicalPlan {
+        Optimizer::with_rules(vec![Arc::new(PushDownFilter::new())])
+            .optimize(plan, &OptimizerContext::new(), |_, _| {})
+            .unwrap()
+    }
+
+    fn ts_cast_to_ms() -> Expr {
+        cast(
+            col("ts"),
+            DataType::Timestamp(ArrowTimeUnit::Millisecond, None),
+        )
+    }
+
+    fn ts_ms_literal(value: i64) -> Expr {
+        lit(ScalarValue::TimestampMillisecond(Some(value), None))
+    }
+
+    fn extract_scan_filters(plan: &LogicalPlan) -> Vec<Expr> {
+        match plan {
+            LogicalPlan::TableScan(scan) => scan.filters.clone(),
+            _ => plan
+                .inputs()
+                .into_iter()
+                .flat_map(extract_scan_filters)
+                .collect(),
+        }
+    }
+
+    fn test_schema(fields: Vec<Field>) -> Arc<DFSchema> {
+        arrow_schema::Schema::new(fields).to_dfschema_ref().unwrap()
+    }
+
+    fn test_source(schema: Arc<DFSchema>) -> Arc<dyn TableSource> {
+        let table = ExactPushdownProvider {
+            schema: Arc::new(schema.as_ref().as_arrow().clone()),
+        };
+        provider_as_source(Arc::new(table))
+    }
+
+    #[derive(Debug)]
+    struct ExactPushdownProvider {
+        schema: arrow_schema::SchemaRef,
+    }
+
+    #[async_trait]
+    impl TableProvider for ExactPushdownProvider {
+        fn as_any(&self) -> &dyn std::any::Any {
+            self
+        }
+
+        fn schema(&self) -> arrow_schema::SchemaRef {
+            self.schema.clone()
+        }
+
+        fn table_type(&self) -> TableType {
+            TableType::Base
+        }
+
+        async fn scan(
+            &self,
+            _state: &dyn Session,
+            _projection: Option<&Vec<usize>>,
+            _filters: &[Expr],
+            _limit: Option<usize>,
+        ) -> datafusion::error::Result<Arc<dyn ExecutionPlan>> {
+            unreachable!("scan should not be called in const_normalization tests")
+        }
+
+        fn supports_filters_pushdown(
+            &self,
+            filters: &[&Expr],
+        ) -> datafusion::error::Result<Vec<TableProviderFilterPushDown>> {
+            Ok(vec![TableProviderFilterPushDown::Exact; filters.len()])
+        }
+    }
+}
diff --git a/src/query/src/optimizer/count_nest_aggr.rs b/src/query/src/optimizer/count_nest_aggr.rs
new file mode 100644
index 0000000000..89ba426074
--- /dev/null
+++ b/src/query/src/optimizer/count_nest_aggr.rs
@@ -0,0 +1,346 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::HashSet;
+use std::sync::Arc;
+
+use datafusion::config::ConfigOptions;
+use datafusion::functions_aggregate::count::count_udaf;
+use datafusion::logical_expr::{Extension, LogicalPlan, LogicalPlanBuilder, Sort};
+use datafusion_common::Result;
+use datafusion_common::tree_node::{Transformed, TreeNode};
+use datafusion_expr::{Expr, UserDefinedLogicalNodeCore, lit};
+use promql::extension_plan::{InstantManipulate, SeriesDivide, SeriesNormalize};
+use store_api::metric_engine_consts::DATA_SCHEMA_TSID_COLUMN_NAME;
+
+use crate::QueryEngineContext;
+use crate::optimizer::ExtensionAnalyzerRule;
+
+/// Rewrites `count(<presence-preserving-agg>(<vector_selector>) by (...))` into a presence-based
+/// group count.
+///
+/// This stays intentionally narrow:
+/// - the outer aggregate must be plain `count`
+/// - the inner aggregate must be a plain aggregate whose result existence is equivalent to input
+///   group existence
+/// - the inner input must be the direct instant-vector-selector plan
+/// - the outer count must only group by the evaluation timestamp
+#[derive(Debug)]
+pub struct CountNestAggrRule;
+
+impl ExtensionAnalyzerRule for CountNestAggrRule {
+    fn analyze(
+        &self,
+        plan: LogicalPlan,
+        _ctx: &QueryEngineContext,
+        _config: &ConfigOptions,
+    ) -> Result<LogicalPlan> {
+        plan.transform_down(&Self::rewrite_plan).map(|x| x.data)
+    }
+}
+
+impl CountNestAggrRule {
+    fn rewrite_plan(plan: LogicalPlan) -> Result<Transformed<LogicalPlan>> {
+        let LogicalPlan::Sort(sort) = plan else {
+            return Ok(Transformed::no(plan));
+        };
+
+        if let Some(rewritten) = Self::try_rewrite_sort(&sort)? {
+            Ok(Transformed::yes(rewritten))
+        } else {
+            Ok(Transformed::no(LogicalPlan::Sort(sort)))
+        }
+    }
+
+    fn try_rewrite_sort(sort: &Sort) -> Result<Option<LogicalPlan>> {
+        if sort.fetch.is_some() {
+            return Ok(None);
+        }
+
+        let LogicalPlan::Aggregate(outer_agg) = sort.input.as_ref() else {
+            return Ok(None);
+        };
+        if outer_agg.group_expr.len() != 1 || outer_agg.aggr_expr.len() != 1 {
+            return Ok(None);
+        }
+        let outer_time_expr = outer_agg.group_expr[0].clone();
+        let outer_count_arg =
+            match Self::aggregate_if(&outer_agg.aggr_expr[0], |name| name == "count") {
+                Some((_, arg)) => arg,
+                None => return Ok(None),
+            };
+
+        let LogicalPlan::Sort(inner_sort) = outer_agg.input.as_ref() else {
+            return Ok(None);
+        };
+        if inner_sort.fetch.is_some() {
+            return Ok(None);
+        }
+
+        let LogicalPlan::Aggregate(inner_agg) = inner_sort.input.as_ref() else {
+            return Ok(None);
+        };
+        if inner_agg.aggr_expr.len() != 1 || inner_agg.group_expr.is_empty() {
+            return Ok(None);
+        }
+        let (inner_is_count, inner_value_expr) =
+            match Self::aggregate_if(&inner_agg.aggr_expr[0], |name| {
+                Self::is_supported_inner_aggregate(name)
+            }) {
+                Some((name, arg)) => (name == "count", arg),
+                None => return Ok(None),
+            };
+        let Expr::Column(_) = inner_value_expr else {
+            return Ok(None);
+        };
+
+        let Expr::Column(outer_count_column) = outer_count_arg else {
+            return Ok(None);
+        };
+        let inner_output_field = inner_agg.schema.field(inner_agg.group_expr.len());
+        if outer_count_column.name != *inner_output_field.name() {
+            return Ok(None);
+        }
+
+        if !Self::is_projection_chain_to_instant(inner_agg.input.as_ref()) {
+            return Ok(None);
+        }
+
+        if !inner_agg
+            .group_expr
+            .iter()
+            .all(|expr| matches!(expr, Expr::Column(_)))
+        {
+            return Ok(None);
+        }
+
+        let Some(time_expr_pos) = inner_agg
+            .group_expr
+            .iter()
+            .position(|expr| expr == &outer_time_expr)
+        else {
+            return Ok(None);
+        };
+
+        let mut presence_group_exprs = Vec::with_capacity(inner_agg.group_expr.len());
+        presence_group_exprs.push(outer_time_expr.clone());
+        presence_group_exprs.extend(
+            inner_agg
+                .group_expr
+                .iter()
+                .enumerate()
+                .filter(|(idx, _)| *idx != time_expr_pos)
+                .map(|(_, expr)| expr.clone()),
+        );
+
+        let mut required_input_columns =
+            Self::collect_required_input_columns(&presence_group_exprs, inner_value_expr);
+        required_input_columns.extend(Self::collect_required_instant_columns(
+            inner_agg.input.as_ref(),
+        ));
+        let presence_source = Self::rebuild_projection_chain_to_instant(
+            inner_agg.input.as_ref(),
+            &required_input_columns,
+        )?;
+
+        let outer_value_name = outer_agg
+            .schema
+            .field(outer_agg.group_expr.len())
+            .name()
+            .clone();
+        let mut presence_input = LogicalPlanBuilder::from(presence_source);
+        if !inner_is_count {
+            presence_input = presence_input.filter(inner_value_expr.clone().is_not_null())?;
+        }
+        let presence_input = presence_input
+            .project(presence_group_exprs.clone())?
+            .distinct()?
+            .build()?;
+
+        let rewritten = LogicalPlanBuilder::from(presence_input)
+            .aggregate(
+                outer_agg.group_expr.clone(),
+                vec![count_udaf().call(vec![lit(1_i64)]).alias(outer_value_name)],
+            )?
+            .sort(sort.expr.clone())?
+            .build()?;
+
+        Ok(Some(rewritten))
+    }
+
+    fn collect_required_input_columns(group_exprs: &[Expr], value_expr: &Expr) -> HashSet<String> {
+        let mut required = HashSet::new();
+
+        for expr in group_exprs {
+            if let Expr::Column(column) = expr {
+                required.insert(column.name.clone());
+            }
+        }
+        if let Expr::Column(column) = value_expr {
+            // Keep the value column in the pruned instant input so `InstantManipulate`
+            // can still perform stale-NaN filtering before we project down to keys.
+            required.insert(column.name.clone());
+        }
+
+        required
+    }
+
+    fn collect_required_instant_columns(plan: &LogicalPlan) -> HashSet<String> {
+        let mut required = HashSet::new();
+        Self::collect_required_instant_columns_into(plan, &mut required);
+        required
+    }
+
+    fn collect_required_instant_columns_into(plan: &LogicalPlan, required: &mut HashSet<String>) {
+        match plan {
+            LogicalPlan::Projection(projection) => {
+                Self::collect_required_instant_columns_into(projection.input.as_ref(), required);
+            }
+            LogicalPlan::Extension(extension) => {
+                for expr in extension.node.expressions() {
+                    if let Expr::Column(column) = expr {
+                        required.insert(column.name);
+                    }
+                }
+
+                if extension.node.as_any().is::<SeriesDivide>()
+                    && extension.node.inputs()[0]
+                        .schema()
+                        .fields()
+                        .iter()
+                        .any(|field| field.name() == DATA_SCHEMA_TSID_COLUMN_NAME)
+                {
+                    required.insert(DATA_SCHEMA_TSID_COLUMN_NAME.to_string());
+                }
+
+                if let Some(input) = extension.node.inputs().into_iter().next() {
+                    Self::collect_required_instant_columns_into(input, required);
+                }
+            }
+            _ => {}
+        }
+    }
+
+    fn aggregate_if<F>(expr: &Expr, accept_name: F) -> Option<(&str, &Expr)>
+    where
+        F: FnOnce(&str) -> bool,
+    {
+        let Expr::AggregateFunction(func) = expr else {
+            return None;
+        };
+        let name = func.func.name();
+        if !accept_name(name)
+            || func.params.filter.is_some()
+            || func.params.distinct
+            || !func.params.order_by.is_empty()
+            || func.params.args.len() != 1
+        {
+            return None;
+        }
+
+        Some((name, &func.params.args[0]))
+    }
+
+    fn is_supported_inner_aggregate(name: &str) -> bool {
+        matches!(
+            name,
+            "count" | "sum" | "avg" | "min" | "max" | "stddev_pop" | "var_pop"
+        )
+    }
+
+    fn is_projection_chain_to_instant(plan: &LogicalPlan) -> bool {
+        let mut current = plan;
+        loop {
+            match current {
+                LogicalPlan::Projection(projection) => current = projection.input.as_ref(),
+                LogicalPlan::Extension(ext) => {
+                    return ext.node.as_any().is::<InstantManipulate>();
+                }
+                _ => return false,
+            }
+        }
+    }
+
+    fn rebuild_projection_chain_to_instant(
+        plan: &LogicalPlan,
+        required_columns: &HashSet<String>,
+    ) -> Result<LogicalPlan> {
+        match plan {
+            LogicalPlan::Projection(projection) => {
+                let input = Self::rebuild_projection_chain_to_instant(
+                    projection.input.as_ref(),
+                    required_columns,
+                )?;
+                LogicalPlanBuilder::from(input)
+                    .project(projection.expr.clone())?
+                    .build()
+            }
+            LogicalPlan::Extension(extension) => {
+                if let Some(instant) = extension.node.as_any().downcast_ref::<InstantManipulate>() {
+                    let input =
+                        Self::prune_instant_input(extension.node.inputs()[0], required_columns)?;
+                    return Ok(LogicalPlan::Extension(Extension {
+                        node: Arc::new(instant.with_exprs_and_inputs(vec![], vec![input])?),
+                    }));
+                }
+
+                Ok(plan.clone())
+            }
+            _ => Ok(plan.clone()),
+        }
+    }
+
+    fn prune_instant_input(
+        plan: &LogicalPlan,
+        required_columns: &HashSet<String>,
+    ) -> Result<LogicalPlan> {
+        match plan {
+            LogicalPlan::Extension(extension) => {
+                if let Some(normalize) = extension.node.as_any().downcast_ref::<SeriesNormalize>() {
+                    let input =
+                        Self::prune_instant_input(extension.node.inputs()[0], required_columns)?;
+                    return Ok(LogicalPlan::Extension(Extension {
+                        node: Arc::new(normalize.with_exprs_and_inputs(vec![], vec![input])?),
+                    }));
+                }
+
+                if let Some(divide) = extension.node.as_any().downcast_ref::<SeriesDivide>() {
+                    let divide_input = extension.node.inputs()[0].clone();
+
+                    let projection_exprs = divide_input
+                        .schema()
+                        .fields()
+                        .iter()
+                        .filter(|field| required_columns.contains(field.name()))
+                        .map(|field| {
+                            Expr::Column(datafusion_common::Column::from_name(field.name().clone()))
+                        })
+                        .collect::<Vec<_>>();
+                    let projected_input = LogicalPlanBuilder::from(divide_input)
+                        .project(projection_exprs)?
+                        .build()?;
+
+                    return Ok(LogicalPlan::Extension(Extension {
+                        node: Arc::new(
+                            divide.with_exprs_and_inputs(vec![], vec![projected_input])?,
+                        ),
+                    }));
+                }
+
+                Ok(plan.clone())
+            }
+            _ => Ok(plan.clone()),
+        }
+    }
+}
diff --git a/src/query/src/optimizer/json_type_concretize.rs b/src/query/src/optimizer/json_type_concretize.rs
new file mode 100644
index 0000000000..9c7e764efe
--- /dev/null
+++ b/src/query/src/optimizer/json_type_concretize.rs
@@ -0,0 +1,308 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::HashMap;
+
+use arrow_schema::DataType;
+use common_function::scalars::json::json_get::JsonGetWithType;
+use datafusion::datasource::DefaultTableSource;
+use datafusion_common::tree_node::{Transformed, TreeNode, TreeNodeRecursion};
+use datafusion_common::{Result, plan_datafusion_err, plan_err};
+use datafusion_expr::{Expr, LogicalPlan};
+use datafusion_optimizer::{OptimizerConfig, OptimizerRule};
+use datatypes::types::json_type::{JsonNativeType, JsonObjectType};
+
+use crate::dummy_catalog::DummyTableProvider;
+
+/// Concretize (deduce) the expected JSON type from query.
+/// For example, we can concretize a JSON type of `{ a: { b: Number } }` from `select j.a.b::Int64`.
+/// The JSON type will be later set into the scan request, for converting the JSON arrays.
+#[derive(Debug)]
+pub(crate) struct JsonTypeConcretizeRule;
+
+impl OptimizerRule for JsonTypeConcretizeRule {
+    fn name(&self) -> &str {
+        "JsonTypeConcretizeRule"
+    }
+
+    fn rewrite(
+        &self,
+        plan: LogicalPlan,
+        _config: &dyn OptimizerConfig,
+    ) -> Result<Transformed<LogicalPlan>> {
+        let json_types = deduce_json_types(&plan)?;
+        if json_types.is_empty() {
+            return Ok(Transformed::no(plan));
+        }
+
+        plan.transform_down(|plan| match &plan {
+            LogicalPlan::TableScan(table_scan) => {
+                let Some(source) = table_scan
+                    .source
+                    .as_any()
+                    .downcast_ref::<DefaultTableSource>()
+                else {
+                    return Ok(Transformed::no(plan));
+                };
+
+                let Some(adapter) = source
+                    .table_provider
+                    .as_any()
+                    .downcast_ref::<DummyTableProvider>()
+                else {
+                    return Ok(Transformed::no(plan));
+                };
+
+                adapter.with_json_type_hint(json_types.clone());
+                Ok(Transformed::yes(plan))
+            }
+            _ => Ok(Transformed::no(plan)),
+        })
+    }
+}
+
+fn deduce_json_types(plan: &LogicalPlan) -> Result<HashMap<String, JsonNativeType>> {
+    let mut json_types = HashMap::<String, JsonNativeType>::new();
+
+    plan.apply(|plan| {
+        for expr in plan.expressions() {
+            expr.apply(|expr| {
+                if let Some((column, json_type)) = deduce_json_type(expr)? {
+                    json_types.entry(column).or_default().merge(&json_type);
+                }
+                Ok(TreeNodeRecursion::Continue)
+            })?;
+        }
+        Ok(TreeNodeRecursion::Continue)
+    })?;
+    Ok(json_types)
+}
+
+fn deduce_json_type(expr: &Expr) -> Result<Option<(String, JsonNativeType)>> {
+    let f = match expr {
+        Expr::ScalarFunction(f) if f.name().eq_ignore_ascii_case(JsonGetWithType::NAME) => f,
+        _ => return Ok(None),
+    };
+
+    let Some(Expr::Column(column)) = f.args.first() else {
+        return plan_err!(
+            "First argument of {} is expected to be a column expr, actual: {:?}",
+            JsonGetWithType::NAME,
+            f.args.first()
+        );
+    };
+
+    let Some(path) = f
+        .args
+        .get(1)
+        .and_then(|expr| expr.as_literal())
+        .and_then(|x| x.try_as_str())
+        .flatten()
+    else {
+        return plan_err!(
+            "Second argument of {} is expected to be a string literal, actual: {:?}",
+            JsonGetWithType::NAME,
+            f.args.get(1)
+        );
+    };
+
+    let with_type = f
+        .args
+        .get(2)
+        .and_then(|expr| expr.as_literal())
+        .map(|x| x.data_type())
+        .unwrap_or(DataType::Utf8View);
+    let with_type =
+        JsonNativeType::try_from(&with_type).map_err(|e| plan_datafusion_err!("{e:?}"))?;
+
+    let mut split = path.rsplit(".");
+    let Some(leaf) = split.next() else {
+        return Ok(Some((column.name.clone(), JsonNativeType::String)));
+    };
+
+    let mut object = JsonObjectType::new();
+    object.insert(leaf.to_string(), with_type);
+    let mut root = JsonNativeType::Object(object);
+
+    for s in split {
+        let mut object = JsonObjectType::new();
+        object.insert(s.to_string(), root);
+        root = JsonNativeType::Object(object);
+    }
+
+    Ok(Some((column.name.clone(), root)))
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use common_function::scalars::udf::create_udf;
+    use datafusion::datasource::provider_as_source;
+    use datafusion_common::{Column, ScalarValue};
+    use datafusion_expr::expr::ScalarFunction;
+    use datafusion_expr::{LogicalPlanBuilder, col};
+    use datafusion_optimizer::OptimizerContext;
+    use store_api::storage::RegionId;
+
+    use super::*;
+    use crate::optimizer::test_util::mock_table_provider;
+
+    fn json_get_expr(base: Expr, path: Expr, with_type: Option<DataType>) -> Result<Expr> {
+        let json_get = Arc::new(create_udf(Arc::new(JsonGetWithType::default())));
+        let mut args = vec![base, path];
+        if let Some(with_type) = with_type {
+            let with_type = ScalarValue::try_new_null(&with_type)?;
+            args.push(Expr::Literal(with_type, None));
+        }
+        Ok(Expr::ScalarFunction(ScalarFunction::new_udf(
+            json_get, args,
+        )))
+    }
+
+    fn path_expr(path: &str) -> Expr {
+        Expr::Literal(ScalarValue::Utf8(Some(path.to_string())), None)
+    }
+
+    fn build_plan(exprs: Vec<Expr>) -> Result<(Arc<DummyTableProvider>, LogicalPlan)> {
+        let provider = Arc::new(mock_table_provider(RegionId::new(1024, 1)));
+        let plan = LogicalPlanBuilder::scan("t", provider_as_source(provider.clone()), None)?
+            .project(exprs)?
+            .build()?;
+        Ok((provider, plan))
+    }
+
+    #[test]
+    fn test_json_type_concretize_rule_rewrite() -> Result<()> {
+        let exprs = vec![
+            json_get_expr(col("k0"), path_expr("a.b"), Some(DataType::Int64))?.alias("ab"),
+            json_get_expr(col("k0"), path_expr("a.c"), None)?.alias("ac"),
+            json_get_expr(col("k0"), path_expr("d"), Some(DataType::Boolean))?.alias("d"),
+        ];
+        let (provider, plan) = build_plan(exprs)?;
+
+        assert!(
+            JsonTypeConcretizeRule
+                .rewrite(plan, &OptimizerContext::default())?
+                .transformed
+        );
+
+        let expected = JsonNativeType::Object(JsonObjectType::from([
+            (
+                "a".to_string(),
+                JsonNativeType::Object(JsonObjectType::from([
+                    ("b".to_string(), JsonNativeType::i64()),
+                    ("c".to_string(), JsonNativeType::String),
+                ])),
+            ),
+            ("d".to_string(), JsonNativeType::Bool),
+        ]));
+
+        let request = provider.scan_request();
+        assert_eq!(1, request.json_type_hint.len());
+        assert_eq!(Some(&expected), request.json_type_hint.get("k0"));
+        Ok(())
+    }
+
+    #[test]
+    fn test_json_type_concretize_rule_conflict_to_variant() -> Result<()> {
+        let exprs = vec![
+            json_get_expr(col("k0"), path_expr("a"), Some(DataType::Int64))?.alias("a_num"),
+            json_get_expr(col("k0"), path_expr("a.b"), Some(DataType::Boolean))?.alias("a_obj"),
+        ];
+        let (provider, plan) = build_plan(exprs)?;
+
+        assert!(
+            JsonTypeConcretizeRule
+                .rewrite(plan, &OptimizerContext::default())?
+                .transformed
+        );
+
+        let expected = JsonNativeType::Object(JsonObjectType::from([(
+            "a".to_string(),
+            JsonNativeType::Variant,
+        )]));
+        assert_eq!(
+            Some(&expected),
+            provider.scan_request().json_type_hint.get("k0")
+        );
+        Ok(())
+    }
+
+    #[test]
+    fn test_json_type_concretize_rule_no_json_get() -> Result<()> {
+        let (provider, plan) = build_plan(vec![col("k0"), col("v0")])?;
+
+        assert!(
+            !JsonTypeConcretizeRule
+                .rewrite(plan, &OptimizerContext::default())?
+                .transformed
+        );
+        assert!(provider.scan_request().json_type_hint.is_empty());
+        Ok(())
+    }
+
+    #[test]
+    fn test_deduce_json_type_with_non_column_base() -> Result<()> {
+        let expr = json_get_expr(
+            Expr::Literal(ScalarValue::Utf8(Some("{}".to_string())), None),
+            path_expr("a"),
+            Some(DataType::Int64),
+        )?;
+
+        let err = deduce_json_type(&expr).unwrap_err();
+        assert!(
+            err.to_string()
+                .contains("First argument of json_get is expected to be a column expr")
+        );
+        Ok(())
+    }
+
+    #[test]
+    fn test_deduce_json_type_with_non_literal_path() -> Result<()> {
+        let expr = json_get_expr(
+            Expr::Column(Column::new_unqualified("k0")),
+            Expr::Column(Column::new_unqualified("path_col")),
+            Some(DataType::Int64),
+        )?;
+
+        let err = deduce_json_type(&expr).unwrap_err();
+        assert!(
+            err.to_string()
+                .contains("Second argument of json_get is expected to be a string literal")
+        );
+        Ok(())
+    }
+
+    #[test]
+    fn test_deduce_json_type_default_string() -> Result<()> {
+        let expr = json_get_expr(
+            Expr::Column(Column::new_unqualified("k0")),
+            path_expr("a.b"),
+            None,
+        )?;
+
+        let deduced = deduce_json_type(&expr)?;
+        let expected = JsonNativeType::Object(JsonObjectType::from([(
+            "a".to_string(),
+            JsonNativeType::Object(JsonObjectType::from([(
+                "b".to_string(),
+                JsonNativeType::String,
+            )])),
+        )]));
+
+        assert_eq!(Some(("k0".to_string(), expected)), deduced);
+        Ok(())
+    }
+}
diff --git a/src/query/src/optimizer/parallelize_scan.rs b/src/query/src/optimizer/parallelize_scan.rs
index 171fc6e919..dd2ba06290 100644
--- a/src/query/src/optimizer/parallelize_scan.rs
+++ b/src/query/src/optimizer/parallelize_scan.rs
@@ -139,7 +139,7 @@ impl ParallelizeScan {
         }
 
         // Sort ranges by number of rows in descending order.
-        ranges.sort_by(|a, b| b.num_rows.cmp(&a.num_rows));
+        ranges.sort_by_key(|b| std::cmp::Reverse(b.num_rows));
         let mut partition_ranges = vec![vec![]; expected_partition_num];
 
         #[derive(Eq, PartialEq)]
diff --git a/src/query/src/optimizer/pass_distribution.rs b/src/query/src/optimizer/pass_distribution.rs
index 7ce2ffd752..1004252636 100644
--- a/src/query/src/optimizer/pass_distribution.rs
+++ b/src/query/src/optimizer/pass_distribution.rs
@@ -17,8 +17,10 @@ use std::sync::Arc;
 use datafusion::config::ConfigOptions;
 use datafusion::physical_optimizer::PhysicalOptimizerRule;
 use datafusion::physical_plan::ExecutionPlan;
+use datafusion::physical_plan::projection::ProjectionExec;
 use datafusion_common::Result as DfResult;
 use datafusion_physical_expr::Distribution;
+use datafusion_physical_expr::utils::map_columns_before_projection;
 
 use crate::dist_plan::MergeScanExec;
 
@@ -83,6 +85,9 @@ impl PassDistribution {
         let mut new_children = Vec::with_capacity(children.len());
         for (idx, child) in children.into_iter().enumerate() {
             let child_req = match required.get(idx) {
+                Some(Distribution::UnspecifiedDistribution) if idx == 0 => {
+                    Self::map_hash_requirement_through_projection(plan.as_ref(), &current_req)
+                }
                 Some(Distribution::UnspecifiedDistribution) => None,
                 None => current_req.clone(),
                 Some(req) => Some(req.clone()),
@@ -103,4 +108,200 @@ impl PassDistribution {
             plan.with_new_children(new_children)
         }
     }
+
+    fn map_hash_requirement_through_projection(
+        plan: &dyn ExecutionPlan,
+        current_req: &Option<Distribution>,
+    ) -> Option<Distribution> {
+        let Some(Distribution::HashPartitioned(required_exprs)) = current_req else {
+            return None;
+        };
+
+        let projection = plan.as_any().downcast_ref::<ProjectionExec>()?;
+        let proj_exprs = projection
+            .expr()
+            .iter()
+            .map(|expr| (Arc::clone(&expr.expr), expr.alias.clone()))
+            .collect::<Vec<_>>();
+        let mapped = map_columns_before_projection(required_exprs, &proj_exprs);
+
+        (mapped.len() == required_exprs.len()).then_some(Distribution::HashPartitioned(mapped))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::{BTreeMap, BTreeSet};
+
+    use arrow_schema::{DataType, Field, Schema, SchemaRef, TimeUnit};
+    use async_trait::async_trait;
+    use common_query::request::QueryRequest;
+    use common_recordbatch::SendableRecordBatchStream;
+    use datafusion::common::NullEquality;
+    use datafusion::execution::SessionStateBuilder;
+    use datafusion::physical_optimizer::PhysicalOptimizerRule;
+    use datafusion::physical_plan::joins::{HashJoinExec, PartitionMode};
+    use datafusion::physical_plan::projection::{ProjectionExec, ProjectionExpr};
+    use datafusion::physical_plan::{ExecutionPlanProperties, Partitioning};
+    use datafusion_expr::{JoinType, LogicalPlanBuilder};
+    use datafusion_physical_expr::PhysicalExpr;
+    use datafusion_physical_expr::expressions::Column as PhysicalColumn;
+    use session::ReadPreference;
+    use session::context::QueryContext;
+    use store_api::metric_engine_consts::DATA_SCHEMA_TSID_COLUMN_NAME;
+    use store_api::storage::RegionId;
+    use table::table_name::TableName;
+
+    use super::*;
+    use crate::error::Result as QueryResult;
+    use crate::region_query::RegionQueryHandler;
+
+    struct NoopRegionQueryHandler;
+
+    #[async_trait]
+    impl RegionQueryHandler for NoopRegionQueryHandler {
+        async fn do_get(
+            &self,
+            _read_preference: ReadPreference,
+            _request: QueryRequest,
+        ) -> QueryResult<SendableRecordBatchStream> {
+            unreachable!("pass distribution tests should not execute remote queries")
+        }
+    }
+
+    #[test]
+    fn passes_hash_requirement_through_projection_to_merge_scan() {
+        let schema = test_schema();
+        let left_merge_scan = Arc::new(test_merge_scan_exec(schema.clone()));
+        let right_merge_scan = Arc::new(test_merge_scan_exec(schema.clone()));
+        let left_projection = Arc::new(
+            ProjectionExec::try_new(
+                vec![
+                    ProjectionExpr::new(partition_column("greptime_value", 3), "greptime_value"),
+                    ProjectionExpr::new(
+                        partition_column(DATA_SCHEMA_TSID_COLUMN_NAME, 1),
+                        DATA_SCHEMA_TSID_COLUMN_NAME,
+                    ),
+                    ProjectionExpr::new(
+                        partition_column("greptime_timestamp", 2),
+                        "greptime_timestamp",
+                    ),
+                ],
+                left_merge_scan,
+            )
+            .unwrap(),
+        ) as Arc<dyn datafusion::physical_plan::ExecutionPlan>;
+        let join = Arc::new(
+            HashJoinExec::try_new(
+                left_projection,
+                right_merge_scan,
+                vec![
+                    (
+                        partition_column(DATA_SCHEMA_TSID_COLUMN_NAME, 1),
+                        partition_column(DATA_SCHEMA_TSID_COLUMN_NAME, 1),
+                    ),
+                    (
+                        partition_column("greptime_timestamp", 2),
+                        partition_column("greptime_timestamp", 2),
+                    ),
+                ],
+                None,
+                &JoinType::Inner,
+                None,
+                PartitionMode::Partitioned,
+                NullEquality::NullEqualsNull,
+                false,
+            )
+            .unwrap(),
+        ) as Arc<dyn datafusion::physical_plan::ExecutionPlan>;
+
+        let optimized = PassDistribution
+            .optimize(join, &ConfigOptions::default())
+            .unwrap();
+        let hash_join = optimized.as_any().downcast_ref::<HashJoinExec>().unwrap();
+        let left_projection = hash_join
+            .left()
+            .as_any()
+            .downcast_ref::<ProjectionExec>()
+            .unwrap();
+        let left_partitioning = left_projection.input().output_partitioning();
+        let right_partitioning = hash_join.right().output_partitioning();
+
+        let Partitioning::Hash(left_exprs, left_count) = left_partitioning else {
+            panic!("expected left merge scan hash partitioning");
+        };
+        let Partitioning::Hash(right_exprs, right_count) = right_partitioning else {
+            panic!("expected right merge scan hash partitioning");
+        };
+
+        assert_eq!(*left_count, 32);
+        assert_eq!(*right_count, 32);
+        assert_eq!(
+            column_names(left_exprs),
+            vec![DATA_SCHEMA_TSID_COLUMN_NAME, "greptime_timestamp"]
+        );
+        assert_eq!(
+            column_names(right_exprs),
+            vec![DATA_SCHEMA_TSID_COLUMN_NAME, "greptime_timestamp"]
+        );
+    }
+
+    fn test_merge_scan_exec(schema: SchemaRef) -> MergeScanExec {
+        let session_state = SessionStateBuilder::new().with_default_features().build();
+        let partition_cols = BTreeMap::from([
+            (
+                DATA_SCHEMA_TSID_COLUMN_NAME.to_string(),
+                BTreeSet::from([datafusion_common::Column::from_name(
+                    DATA_SCHEMA_TSID_COLUMN_NAME,
+                )]),
+            ),
+            (
+                "greptime_timestamp".to_string(),
+                BTreeSet::from([datafusion_common::Column::from_name("greptime_timestamp")]),
+            ),
+        ]);
+        let plan = LogicalPlanBuilder::empty(false).build().unwrap();
+
+        MergeScanExec::new(
+            &session_state,
+            TableName::new("greptime", "public", "test"),
+            vec![RegionId::new(1, 0), RegionId::new(1, 1)],
+            plan,
+            schema.as_ref(),
+            Arc::new(NoopRegionQueryHandler),
+            QueryContext::arc(),
+            32,
+            partition_cols,
+        )
+        .unwrap()
+    }
+
+    fn test_schema() -> SchemaRef {
+        Arc::new(Schema::new(vec![
+            Field::new("host", DataType::Utf8, true),
+            Field::new(DATA_SCHEMA_TSID_COLUMN_NAME, DataType::UInt64, false),
+            Field::new(
+                "greptime_timestamp",
+                DataType::Timestamp(TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new("greptime_value", DataType::Float64, true),
+        ]))
+    }
+
+    fn partition_column(name: &str, index: usize) -> Arc<dyn PhysicalExpr> {
+        Arc::new(PhysicalColumn::new(name, index))
+    }
+
+    fn column_names(exprs: &[Arc<dyn PhysicalExpr>]) -> Vec<&str> {
+        exprs
+            .iter()
+            .map(|expr| {
+                expr.as_any()
+                    .downcast_ref::<PhysicalColumn>()
+                    .unwrap()
+                    .name()
+            })
+            .collect()
+    }
 }
diff --git a/src/query/src/optimizer/scan_hint.rs b/src/query/src/optimizer/scan_hint.rs
index da70813404..89e3afaed0 100644
--- a/src/query/src/optimizer/scan_hint.rs
+++ b/src/query/src/optimizer/scan_hint.rs
@@ -375,10 +375,10 @@ impl TreeNodeVisitor<'_> for ScanHintVisitor {
             LogicalPlan::Filter(_) => {
                 self.vector_search.on_filter_exit();
             }
-            LogicalPlan::Subquery(_) | LogicalPlan::SubqueryAlias(_) => {
-                if is_branching_for_vector(_node) {
-                    self.vector_search.on_branching_exit();
-                }
+            LogicalPlan::Subquery(_) | LogicalPlan::SubqueryAlias(_)
+                if is_branching_for_vector(_node) =>
+            {
+                self.vector_search.on_branching_exit();
             }
             _ if _node.inputs().len() > 1 => {
                 self.vector_search.on_branching_exit();
diff --git a/src/query/src/optimizer/windowed_sort.rs b/src/query/src/optimizer/windowed_sort.rs
index 9365c8e1e8..55a48c1869 100644
--- a/src/query/src/optimizer/windowed_sort.rs
+++ b/src/query/src/optimizer/windowed_sort.rs
@@ -12,9 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::collections::HashSet;
 use std::sync::Arc;
 
+use arrow_schema::DataType;
 use datafusion::physical_optimizer::PhysicalOptimizerRule;
 use datafusion::physical_plan::ExecutionPlan;
 use datafusion::physical_plan::coalesce_partitions::CoalescePartitionsExec;
@@ -26,7 +26,8 @@ use datafusion::physical_plan::sorts::sort::SortExec;
 use datafusion::physical_plan::sorts::sort_preserving_merge::SortPreservingMergeExec;
 use datafusion_common::Result as DataFusionResult;
 use datafusion_common::tree_node::{Transformed, TreeNode};
-use datafusion_physical_expr::expressions::Column as PhysicalColumn;
+use datafusion_physical_expr::expressions::{CastExpr, Column as PhysicalColumn};
+use datafusion_physical_expr::{PhysicalExpr, ScalarFunctionExpr};
 use store_api::region_engine::PartitionRange;
 use table::table::scan::RegionScanExec;
 
@@ -91,9 +92,13 @@ impl WindowedSortPhysicalRule {
                         .expr
                         .as_any()
                         .downcast_ref::<PhysicalColumn>()
-                        && scanner_info
-                            .time_index
-                            .contains(input_schema.field(column_expr.index()).name())
+                        && matches!(
+                            input_schema.field(column_expr.index()).data_type(),
+                            DataType::Timestamp(_, _)
+                        )
+                        && is_time_index_expr(&sort_input, &first_sort_expr.expr)?
+                        && sort_exec.fetch().is_none()
+                    // skip if there is a limit, as dyn filter along is good enough in this case
                     {
                     } else {
                         return Ok(Transformed::no(plan));
@@ -152,14 +157,11 @@ impl WindowedSortPhysicalRule {
 #[derive(Debug)]
 struct ScannerInfo {
     partition_ranges: Vec<Vec<PartitionRange>>,
-    time_index: HashSet<String>,
     tag_columns: Vec<String>,
 }
 
 fn fetch_partition_range(input: Arc<dyn ExecutionPlan>) -> DataFusionResult<Option<ScannerInfo>> {
     let mut partition_ranges = None;
-    let mut time_index = HashSet::new();
-    let mut alias_map = Vec::new();
     let mut tag_columns = None;
 
     input.transform_up(|plan| {
@@ -182,18 +184,6 @@ fn fetch_partition_range(input: Arc<dyn ExecutionPlan>) -> DataFusionResult<Opti
             partition_ranges = None;
         }
 
-        // TODO(discord9): do this in logical plan instead as it's lessy bugy there
-        // Collects alias of the time index column.
-        if let Some(projection) = plan.as_any().downcast_ref::<ProjectionExec>() {
-            for expr in projection.expr() {
-                if let Some(column_expr) = expr.expr.as_any().downcast_ref::<PhysicalColumn>() {
-                    alias_map.push((column_expr.name().to_string(), expr.alias.clone()));
-                }
-            }
-            // resolve alias properly
-            time_index = resolve_alias(&alias_map, &time_index);
-        }
-
         if let Some(region_scan_exec) = plan.as_any().downcast_ref::<RegionScanExec>() {
             // `PerSeries` distribution is not supported in windowed sort.
             if region_scan_exec.distribution()
@@ -204,8 +194,6 @@ fn fetch_partition_range(input: Arc<dyn ExecutionPlan>) -> DataFusionResult<Opti
             }
 
             partition_ranges = Some(region_scan_exec.get_uncollapsed_partition_ranges());
-            // Reset time index column.
-            time_index = HashSet::from([region_scan_exec.time_index()]);
             tag_columns = Some(region_scan_exec.tag_columns());
 
             region_scan_exec.with_distinguish_partition_range(true);
@@ -217,7 +205,6 @@ fn fetch_partition_range(input: Arc<dyn ExecutionPlan>) -> DataFusionResult<Opti
     let result = try {
         ScannerInfo {
             partition_ranges: partition_ranges?,
-            time_index,
             tag_columns: tag_columns?,
         }
     };
@@ -225,6 +212,103 @@ fn fetch_partition_range(input: Arc<dyn ExecutionPlan>) -> DataFusionResult<Opti
     Ok(result)
 }
 
+fn is_time_index_expr(
+    plan: &Arc<dyn ExecutionPlan>,
+    expr: &Arc<dyn PhysicalExpr>,
+) -> DataFusionResult<bool> {
+    if let Some(column_expr) = expr.as_any().downcast_ref::<PhysicalColumn>() {
+        return is_time_index_column(plan, column_expr);
+    }
+
+    if let Some(cast_expr) = expr.as_any().downcast_ref::<CastExpr>() {
+        return if matches!(cast_expr.cast_type(), DataType::Timestamp(_, _)) {
+            is_time_index_expr(plan, cast_expr.expr())
+        } else {
+            Ok(false)
+        };
+    }
+
+    if let Some(scalar_function_expr) = expr.as_any().downcast_ref::<ScalarFunctionExpr>() {
+        return if is_supported_time_index_wrapper(scalar_function_expr)
+            && scalar_function_expr.args().len() == 1
+        {
+            is_time_index_expr(plan, &scalar_function_expr.args()[0])
+        } else {
+            Ok(false)
+        };
+    }
+
+    Ok(false)
+}
+
+fn is_time_index_column(
+    plan: &Arc<dyn ExecutionPlan>,
+    column_expr: &PhysicalColumn,
+) -> DataFusionResult<bool> {
+    if let Some(projection) = plan.as_any().downcast_ref::<ProjectionExec>() {
+        let Some(projection_expr) = projection.expr().get(column_expr.index()) else {
+            return Ok(false);
+        };
+        return is_time_index_expr(projection.input(), &projection_expr.expr);
+    }
+
+    if let Some(filter) = plan.as_any().downcast_ref::<FilterExec>() {
+        let child_column_expr = filter
+            .projection()
+            .as_ref()
+            .and_then(|projection| projection.get(column_expr.index()).copied())
+            .map(|input_index| {
+                PhysicalColumn::new(
+                    filter.input().schema().field(input_index).name(),
+                    input_index,
+                )
+            })
+            .unwrap_or_else(|| column_expr.clone());
+        let child_expr = Arc::new(child_column_expr) as Arc<dyn PhysicalExpr>;
+        return is_time_index_expr(filter.input(), &child_expr);
+    }
+
+    if let Some(region_scan_exec) = plan.as_any().downcast_ref::<RegionScanExec>() {
+        let schema = plan.schema();
+        let column_field = schema.field(column_expr.index());
+        return Ok(
+            matches!(column_field.data_type(), DataType::Timestamp(_, _))
+                && *column_field.name() == region_scan_exec.time_index(),
+        );
+    }
+
+    let Some(child) = passthrough_child(plan.as_ref()) else {
+        return Ok(false);
+    };
+    let child_expr = Arc::new(column_expr.clone()) as Arc<dyn PhysicalExpr>;
+    is_time_index_expr(&child, &child_expr)
+}
+
+fn passthrough_child(plan: &dyn ExecutionPlan) -> Option<Arc<dyn ExecutionPlan>> {
+    if plan.as_any().is::<CoalescePartitionsExec>()
+        || plan.as_any().is::<RepartitionExec>()
+        || plan.as_any().is::<CooperativeExec>()
+    {
+        return schema_preserving_child(plan);
+    }
+
+    None
+}
+
+fn schema_preserving_child(plan: &dyn ExecutionPlan) -> Option<Arc<dyn ExecutionPlan>> {
+    let child = plan.children().first().cloned().cloned()?;
+    (plan.schema().as_ref() == child.schema().as_ref()).then_some(child)
+}
+
+fn is_supported_time_index_wrapper(expr: &ScalarFunctionExpr) -> bool {
+    (expr.name().eq_ignore_ascii_case("to_timestamp")
+        || expr.name().eq_ignore_ascii_case("to_timestamp_seconds")
+        || expr.name().eq_ignore_ascii_case("to_timestamp_millis")
+        || expr.name().eq_ignore_ascii_case("to_timestamp_micros")
+        || expr.name().eq_ignore_ascii_case("to_timestamp_nanos"))
+        && matches!(expr.return_type(), DataType::Timestamp(_, _))
+}
+
 /// Removes the repartition plan between the filter and region scan.
 fn remove_repartition(
     plan: Arc<dyn ExecutionPlan>,
@@ -246,80 +330,321 @@ fn remove_repartition(
     })
 }
 
-/// Resolves alias of the time index column.
-///
-/// i.e if a is time index, alias= {a:b, b:c}, then result should be {a, b}(not {a, c}) because projection is not transitive
-/// if alias={b:a} and a is time index, then return empty
-fn resolve_alias(alias_map: &[(String, String)], time_index: &HashSet<String>) -> HashSet<String> {
-    // available old name for time index
-    let mut avail_old_name = time_index.clone();
-    let mut new_time_index = HashSet::new();
-    for (old, new) in alias_map {
-        if time_index.contains(old) {
-            new_time_index.insert(new.clone());
-        } else if time_index.contains(new) && old != new {
-            // other alias to time index, remove the old name
-            avail_old_name.remove(new);
-            continue;
-        }
-    }
-    // add the remaining time index that is not in alias map
-    new_time_index.extend(avail_old_name);
-    new_time_index
-}
-
 #[cfg(test)]
 mod test {
-    use itertools::Itertools;
+    use std::sync::Arc;
+
+    use api::v1::SemanticType;
+    use arrow_schema::{Field, TimeUnit};
+    use common_recordbatch::RecordBatches;
+    use datafusion::config::ConfigOptions;
+    use datafusion::physical_plan::filter::FilterExecBuilder;
+    use datafusion_common::ScalarValue;
+    use datafusion_functions::datetime::to_timestamp_millis;
+    use datafusion_physical_expr::expressions::{CastExpr, Literal};
+    use datatypes::data_type::ConcreteDataType;
+    use datatypes::schema::{ColumnSchema, Schema};
+    use store_api::metadata::{ColumnMetadata, RegionMetadataBuilder};
+    use store_api::region_engine::SinglePartitionScanner;
+    use store_api::storage::{RegionId, ScanRequest};
 
     use super::*;
 
     #[test]
-    fn test_alias() {
-        let testcases = [
-            // notice the old name is still in the result
-            (
-                vec![("a", "b"), ("b", "c")],
-                HashSet::from(["a"]),
-                HashSet::from(["a", "b"]),
-            ),
-            // alias swap
-            (
-                vec![("b", "a"), ("a", "b")],
-                HashSet::from(["a"]),
-                HashSet::from(["b"]),
-            ),
-            (
-                vec![("b", "a"), ("b", "c")],
-                HashSet::from(["a"]),
-                HashSet::from([]),
-            ),
-            // not in alias map
-            (
-                vec![("c", "d"), ("d", "c")],
-                HashSet::from(["a"]),
-                HashSet::from(["a"]),
-            ),
-            // no alias
-            (vec![], HashSet::from(["a"]), HashSet::from(["a"])),
-            // empty time index
-            (vec![], HashSet::from([]), HashSet::from([])),
-        ];
-        for (alias_map, time_index, expected) in testcases {
-            let alias_map = alias_map
-                .into_iter()
-                .map(|(k, v)| (k.to_string(), v.to_string()))
-                .collect_vec();
-            let time_index = time_index.into_iter().map(|i| i.to_string()).collect();
-            let expected: HashSet<String> = expected.into_iter().map(|i| i.to_string()).collect();
+    fn test_is_time_index_expr_tracks_aliases_through_projection() {
+        let scan = new_region_scan();
+        let projection = Arc::new(
+            ProjectionExec::try_new(
+                vec![(
+                    Arc::new(PhysicalColumn::new("ts", 1)) as Arc<dyn PhysicalExpr>,
+                    "alias_ts".to_string(),
+                )],
+                scan,
+            )
+            .unwrap(),
+        ) as Arc<dyn ExecutionPlan>;
+        let expr = Arc::new(PhysicalColumn::new("alias_ts", 0)) as Arc<dyn PhysicalExpr>;
 
-            assert_eq!(
-                expected,
-                resolve_alias(&alias_map, &time_index),
-                "alias_map={:?}, time_index={:?}",
-                alias_map,
-                time_index
-            );
-        }
+        assert!(is_time_index_expr(&projection, &expr).unwrap());
+    }
+
+    #[test]
+    fn test_is_time_index_expr_tracks_multi_level_aliases() {
+        let scan = new_region_scan();
+        let first_projection = Arc::new(
+            ProjectionExec::try_new(
+                vec![(
+                    Arc::new(PhysicalColumn::new("ts", 1)) as Arc<dyn PhysicalExpr>,
+                    "alias_1".to_string(),
+                )],
+                scan,
+            )
+            .unwrap(),
+        ) as Arc<dyn ExecutionPlan>;
+        let second_projection = Arc::new(
+            ProjectionExec::try_new(
+                vec![(
+                    Arc::new(PhysicalColumn::new("alias_1", 0)) as Arc<dyn PhysicalExpr>,
+                    "alias_2".to_string(),
+                )],
+                first_projection,
+            )
+            .unwrap(),
+        ) as Arc<dyn ExecutionPlan>;
+        let expr = Arc::new(PhysicalColumn::new("alias_2", 0)) as Arc<dyn PhysicalExpr>;
+
+        assert!(is_time_index_expr(&second_projection, &expr).unwrap());
+    }
+
+    #[test]
+    fn test_is_time_index_expr_tracks_wrapped_aliases_through_projection() {
+        let scan = new_region_scan();
+        let config = Arc::new(ConfigOptions::default());
+        let return_field = Arc::new(Field::new(
+            "ts",
+            DataType::Timestamp(TimeUnit::Millisecond, None),
+            false,
+        ));
+        let projection = Arc::new(
+            ProjectionExec::try_new(
+                vec![(
+                    Arc::new(ScalarFunctionExpr::new(
+                        "to_timestamp_millis",
+                        to_timestamp_millis(config.as_ref()),
+                        vec![Arc::new(PhysicalColumn::new("ts", 1))],
+                        return_field,
+                        config,
+                    )) as Arc<dyn PhysicalExpr>,
+                    "ts".to_string(),
+                )],
+                scan,
+            )
+            .unwrap(),
+        ) as Arc<dyn ExecutionPlan>;
+        let expr = Arc::new(PhysicalColumn::new("ts", 0)) as Arc<dyn PhysicalExpr>;
+
+        assert!(is_time_index_expr(&projection, &expr).unwrap());
+    }
+
+    #[test]
+    fn test_is_time_index_expr_tracks_cast_aliases_through_projection() {
+        let scan = new_region_scan();
+        let projection = Arc::new(
+            ProjectionExec::try_new(
+                vec![(
+                    Arc::new(CastExpr::new(
+                        Arc::new(PhysicalColumn::new("ts", 1)),
+                        DataType::Timestamp(TimeUnit::Millisecond, None),
+                        None,
+                    )) as Arc<dyn PhysicalExpr>,
+                    "ts_ms".to_string(),
+                )],
+                scan,
+            )
+            .unwrap(),
+        ) as Arc<dyn ExecutionPlan>;
+        let expr = Arc::new(PhysicalColumn::new("ts_ms", 0)) as Arc<dyn PhysicalExpr>;
+
+        assert!(is_time_index_expr(&projection, &expr).unwrap());
+    }
+
+    #[test]
+    fn test_is_time_index_expr_rejects_unsupported_wrappers() {
+        let scan = new_region_scan();
+        let config = Arc::new(ConfigOptions::default());
+        let return_field = Arc::new(Field::new(
+            "ts",
+            DataType::Timestamp(TimeUnit::Millisecond, None),
+            false,
+        ));
+        let projection = Arc::new(
+            ProjectionExec::try_new(
+                vec![(
+                    Arc::new(ScalarFunctionExpr::new(
+                        "date_trunc",
+                        to_timestamp_millis(config.as_ref()),
+                        vec![Arc::new(PhysicalColumn::new("ts", 1))],
+                        return_field,
+                        config,
+                    )) as Arc<dyn PhysicalExpr>,
+                    "ts".to_string(),
+                )],
+                scan,
+            )
+            .unwrap(),
+        ) as Arc<dyn ExecutionPlan>;
+        let expr = Arc::new(PhysicalColumn::new("ts", 0)) as Arc<dyn PhysicalExpr>;
+
+        assert!(!is_time_index_expr(&projection, &expr).unwrap());
+    }
+
+    #[test]
+    fn test_is_supported_time_index_wrapper_ignores_function_name_case() {
+        let config = Arc::new(ConfigOptions::default());
+        let return_field = Arc::new(Field::new(
+            "ts",
+            DataType::Timestamp(TimeUnit::Millisecond, None),
+            false,
+        ));
+        let expr = ScalarFunctionExpr::new(
+            "To_Timestamp_Millis",
+            to_timestamp_millis(config.as_ref()),
+            vec![Arc::new(PhysicalColumn::new("ts", 1))],
+            return_field,
+            config,
+        );
+
+        assert!(is_supported_time_index_wrapper(&expr));
+    }
+
+    #[test]
+    fn test_is_time_index_expr_rejects_non_timestamp_casts() {
+        let scan = new_region_scan();
+        let cast_expr = Arc::new(CastExpr::new(
+            Arc::new(PhysicalColumn::new("ts", 1)),
+            DataType::Timestamp(TimeUnit::Millisecond, None),
+            None,
+        )) as Arc<dyn PhysicalExpr>;
+        assert!(is_time_index_expr(&scan, &cast_expr).unwrap());
+
+        let non_timestamp_cast = Arc::new(CastExpr::new(
+            Arc::new(PhysicalColumn::new("ts", 1)),
+            DataType::Int64,
+            None,
+        )) as Arc<dyn PhysicalExpr>;
+        assert!(!is_time_index_expr(&scan, &non_timestamp_cast).unwrap());
+    }
+
+    #[test]
+    fn test_is_time_index_expr_tracks_time_index_through_filter() {
+        let scan = new_region_scan();
+        let filter = Arc::new(
+            FilterExec::try_new(
+                Arc::new(Literal::new(ScalarValue::Boolean(Some(true)))),
+                scan,
+            )
+            .unwrap(),
+        ) as Arc<dyn ExecutionPlan>;
+        let expr = Arc::new(PhysicalColumn::new("ts", 1)) as Arc<dyn PhysicalExpr>;
+
+        assert!(is_time_index_expr(&filter, &expr).unwrap());
+    }
+
+    #[test]
+    fn test_is_time_index_expr_tracks_time_index_through_passthrough_wrapper_and_filter_projection()
+    {
+        let scan = new_region_scan();
+        let projected_filter = Arc::new(
+            FilterExecBuilder::new(
+                Arc::new(Literal::new(ScalarValue::Boolean(Some(true)))),
+                scan,
+            )
+            .apply_projection(Some(vec![1]))
+            .unwrap()
+            .build()
+            .unwrap(),
+        ) as Arc<dyn ExecutionPlan>;
+        let cooperative =
+            Arc::new(CooperativeExec::new(projected_filter)) as Arc<dyn ExecutionPlan>;
+        let expr = Arc::new(PhysicalColumn::new("ts", 0)) as Arc<dyn PhysicalExpr>;
+
+        assert!(is_time_index_expr(&cooperative, &expr).unwrap());
+    }
+
+    #[test]
+    fn test_schema_preserving_child_rejects_schema_changing_projection() {
+        let scan = new_region_scan();
+        let projection = ProjectionExec::try_new(
+            vec![(
+                Arc::new(PhysicalColumn::new("ts", 1)) as Arc<dyn PhysicalExpr>,
+                "ts".to_string(),
+            )],
+            scan,
+        )
+        .unwrap();
+
+        assert!(schema_preserving_child(&projection).is_none());
+    }
+
+    #[test]
+    fn test_cooperative_exec_satisfies_passthrough_schema_contract() {
+        let child = new_region_scan();
+        let plan = Arc::new(CooperativeExec::new(child.clone())) as Arc<dyn ExecutionPlan>;
+
+        assert_passthrough_schema_contract(plan, child);
+    }
+
+    #[test]
+    fn test_repartition_exec_satisfies_passthrough_schema_contract() {
+        let child = new_region_scan();
+        let plan = Arc::new(
+            RepartitionExec::try_new(
+                child.clone(),
+                datafusion_physical_expr::Partitioning::RoundRobinBatch(2),
+            )
+            .unwrap(),
+        ) as Arc<dyn ExecutionPlan>;
+
+        assert_passthrough_schema_contract(plan, child);
+    }
+
+    #[test]
+    fn test_coalesce_partitions_exec_satisfies_passthrough_schema_contract() {
+        let child = new_region_scan();
+        let plan = Arc::new(CoalescePartitionsExec::new(child.clone())) as Arc<dyn ExecutionPlan>;
+
+        assert_passthrough_schema_contract(plan, child);
+    }
+
+    fn assert_passthrough_schema_contract(
+        plan: Arc<dyn ExecutionPlan>,
+        child: Arc<dyn ExecutionPlan>,
+    ) {
+        assert_eq!(plan.schema().as_ref(), child.schema().as_ref());
+
+        let passthrough = passthrough_child(plan.as_ref()).expect("wrapper should preserve schema");
+        assert_eq!(passthrough.schema().as_ref(), child.schema().as_ref());
+    }
+
+    fn new_region_scan() -> Arc<dyn ExecutionPlan> {
+        let schema = Arc::new(Schema::new(vec![
+            ColumnSchema::new("value", ConcreteDataType::int32_datatype(), false),
+            ColumnSchema::new(
+                "ts",
+                ConcreteDataType::timestamp_nanosecond_datatype(),
+                false,
+            ),
+        ]));
+        let recordbatches = RecordBatches::try_new(schema.clone(), vec![]).unwrap();
+        let stream = recordbatches.as_stream();
+
+        let mut builder = RegionMetadataBuilder::new(RegionId::new(1, 1));
+        builder
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "value",
+                    ConcreteDataType::int32_datatype(),
+                    false,
+                ),
+                semantic_type: SemanticType::Field,
+                column_id: 1,
+            })
+            .push_column_metadata(ColumnMetadata {
+                column_schema: ColumnSchema::new(
+                    "ts",
+                    ConcreteDataType::timestamp_nanosecond_datatype(),
+                    false,
+                ),
+                semantic_type: SemanticType::Timestamp,
+                column_id: 2,
+            });
+
+        let scanner = Box::new(SinglePartitionScanner::new(
+            stream,
+            false,
+            Arc::new(builder.build().unwrap()),
+            None,
+        ));
+        Arc::new(RegionScanExec::new(scanner, ScanRequest::default(), None).unwrap())
     }
 }
diff --git a/src/query/src/options.rs b/src/query/src/options.rs
index 50ca1177a5..688d6315b0 100644
--- a/src/query/src/options.rs
+++ b/src/query/src/options.rs
@@ -12,8 +12,21 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+use std::collections::HashMap;
+
 use common_base::memory_limit::MemoryLimit;
 use serde::{Deserialize, Serialize};
+use store_api::storage::RegionId;
+use table::metadata::TableId;
+
+use crate::error::{Error, InvalidQueryContextExtensionSnafu, Result};
+
+pub const FLOW_INCREMENTAL_AFTER_SEQS: &str = "flow.incremental_after_seqs";
+pub const FLOW_INCREMENTAL_MODE: &str = "flow.incremental_mode";
+pub const FLOW_RETURN_REGION_SEQ: &str = "flow.return_region_seq";
+pub const FLOW_SINK_TABLE_ID: &str = "flow.sink_table_id";
+
+pub const FLOW_INCREMENTAL_MODE_MEMTABLE_ONLY: &str = "memtable_only";
 
 /// Query engine config
 #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
@@ -39,3 +52,453 @@ impl Default for QueryOptions {
         }
     }
 }
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum FlowIncrementalMode {
+    MemtableOnly,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Default)]
+pub struct FlowQueryExtensions {
+    /// Maps region id -> lower exclusive sequence bound for incremental reads.
+    pub incremental_after_seqs: Option<HashMap<u64, u64>>,
+    /// Incremental read mode requested by the caller.
+    pub incremental_mode: Option<FlowIncrementalMode>,
+    /// Whether the caller expects per-region watermark metadata in terminal metrics.
+    pub return_region_seq: bool,
+    /// Optional sink table id used to distinguish source scans from sink reads.
+    pub sink_table_id: Option<TableId>,
+}
+
+impl FlowQueryExtensions {
+    /// Parses flow-specific query extensions when any flow key is present.
+    ///
+    /// Returns `Ok(None)` for ordinary queries with no flow-related extensions,
+    /// `Ok(Some(_))` when flow context is present and valid, and `Err(_)` when a
+    /// flow-related extension is present but malformed or incomplete.
+    pub fn parse_flow_extensions(extensions: &HashMap<String, String>) -> Result<Option<Self>> {
+        let has_flow_context = extensions.contains_key(FLOW_INCREMENTAL_AFTER_SEQS)
+            || extensions.contains_key(FLOW_INCREMENTAL_MODE)
+            || extensions.contains_key(FLOW_RETURN_REGION_SEQ)
+            || extensions.contains_key(FLOW_SINK_TABLE_ID);
+
+        if !has_flow_context {
+            return Ok(None);
+        }
+
+        let incremental_mode = extensions
+            .get(FLOW_INCREMENTAL_MODE)
+            .map(|value| match value.as_str() {
+                v if v.eq_ignore_ascii_case(FLOW_INCREMENTAL_MODE_MEMTABLE_ONLY) => {
+                    Ok(FlowIncrementalMode::MemtableOnly)
+                }
+                _ => Err(invalid_query_context_extension(format!(
+                    "Invalid value for {}: {}",
+                    FLOW_INCREMENTAL_MODE, value
+                ))),
+            })
+            .transpose()?;
+
+        let incremental_after_seqs = extensions
+            .get(FLOW_INCREMENTAL_AFTER_SEQS)
+            .map(|value| parse_incremental_after_seqs(value.as_str()))
+            .transpose()?;
+
+        let return_region_seq = extensions
+            .get(FLOW_RETURN_REGION_SEQ)
+            .map(|value| parse_bool(value.as_str()))
+            .transpose()?
+            .unwrap_or(false);
+
+        let sink_table_id = extensions
+            .get(FLOW_SINK_TABLE_ID)
+            .map(|value| {
+                value.parse::<TableId>().map_err(|_| {
+                    invalid_query_context_extension(format!(
+                        "Invalid value for {}: {}",
+                        FLOW_SINK_TABLE_ID, value
+                    ))
+                })
+            })
+            .transpose()?;
+
+        if matches!(incremental_mode, Some(FlowIncrementalMode::MemtableOnly)) {
+            let after_seqs = incremental_after_seqs.as_ref().ok_or_else(|| {
+                invalid_query_context_extension(format!(
+                    "{} is required when {}={}.",
+                    FLOW_INCREMENTAL_AFTER_SEQS,
+                    FLOW_INCREMENTAL_MODE,
+                    FLOW_INCREMENTAL_MODE_MEMTABLE_ONLY
+                ))
+            })?;
+            if after_seqs.is_empty() {
+                return Err(invalid_query_context_extension(format!(
+                    "{} must not be empty when {}={}.",
+                    FLOW_INCREMENTAL_AFTER_SEQS,
+                    FLOW_INCREMENTAL_MODE,
+                    FLOW_INCREMENTAL_MODE_MEMTABLE_ONLY
+                )));
+            }
+        }
+
+        Ok(Some(Self {
+            incremental_after_seqs,
+            incremental_mode,
+            return_region_seq,
+            sink_table_id,
+        }))
+    }
+
+    pub fn validate_for_scan(&self, source_region_id: RegionId) -> Result<bool> {
+        if self.sink_table_id.is_some() && self.sink_table_id == Some(source_region_id.table_id()) {
+            return Ok(false);
+        }
+
+        if matches!(
+            self.incremental_mode,
+            Some(FlowIncrementalMode::MemtableOnly)
+        ) {
+            let after_seqs = self.incremental_after_seqs.as_ref().ok_or_else(|| {
+                invalid_query_context_extension(format!(
+                    "{} is required when {}=memtable_only.",
+                    FLOW_INCREMENTAL_AFTER_SEQS, FLOW_INCREMENTAL_MODE
+                ))
+            })?;
+
+            if !after_seqs.contains_key(&source_region_id.as_u64()) {
+                return Err(invalid_query_context_extension(format!(
+                    "Missing region {} in {} when {}=memtable_only.",
+                    source_region_id, FLOW_INCREMENTAL_AFTER_SEQS, FLOW_INCREMENTAL_MODE
+                )));
+            }
+        }
+
+        Ok(self.incremental_after_seqs.is_some())
+    }
+
+    pub fn should_collect_region_watermark(&self) -> bool {
+        should_collect_region_watermark(
+            self.return_region_seq,
+            self.incremental_after_seqs.is_some(),
+        )
+    }
+}
+
+/// Returns whether raw Flow query extensions request terminal region watermark collection.
+///
+/// This is only an intent/presence check for transport/scan plumbing; callers that need
+/// validated Flow options must still use [`FlowQueryExtensions::parse_flow_extensions`].
+pub fn should_collect_region_watermark_from_extensions(
+    extensions: &HashMap<String, String>,
+) -> bool {
+    let return_region_seq = extensions
+        .get(FLOW_RETURN_REGION_SEQ)
+        .is_some_and(|value| value.eq_ignore_ascii_case("true"));
+    let has_incremental_after_seqs = extensions.contains_key(FLOW_INCREMENTAL_AFTER_SEQS);
+
+    should_collect_region_watermark(return_region_seq, has_incremental_after_seqs)
+}
+
+fn should_collect_region_watermark(
+    return_region_seq: bool,
+    has_incremental_after_seqs: bool,
+) -> bool {
+    return_region_seq || has_incremental_after_seqs
+}
+
+fn parse_incremental_after_seqs(value: &str) -> Result<HashMap<u64, u64>> {
+    let raw = serde_json::from_str::<HashMap<String, serde_json::Value>>(value).map_err(|e| {
+        invalid_query_context_extension(format!(
+            "Invalid JSON for {}: {} ({})",
+            FLOW_INCREMENTAL_AFTER_SEQS, value, e
+        ))
+    })?;
+
+    raw.into_iter()
+        .map(|(region_id, raw_seq)| {
+            let region_id = region_id.parse::<u64>().map_err(|_| {
+                invalid_query_context_extension(format!(
+                    "Invalid region id in {}: {}",
+                    FLOW_INCREMENTAL_AFTER_SEQS, region_id
+                ))
+            })?;
+
+            let seq = match raw_seq {
+                serde_json::Value::Number(num) => num.as_u64().ok_or_else(|| {
+                    invalid_query_context_extension(format!(
+                        "Invalid sequence value in {} for region {}: {}",
+                        FLOW_INCREMENTAL_AFTER_SEQS, region_id, num
+                    ))
+                })?,
+                serde_json::Value::String(s) => s.parse::<u64>().map_err(|_| {
+                    invalid_query_context_extension(format!(
+                        "Invalid sequence string in {} for region {}: {}",
+                        FLOW_INCREMENTAL_AFTER_SEQS, region_id, s
+                    ))
+                })?,
+                _ => {
+                    return Err(invalid_query_context_extension(format!(
+                        "Invalid sequence value type in {} for region {}",
+                        FLOW_INCREMENTAL_AFTER_SEQS, region_id
+                    )));
+                }
+            };
+
+            Ok((region_id, seq))
+        })
+        .collect()
+}
+
+fn parse_bool(value: &str) -> Result<bool> {
+    match value {
+        v if v.eq_ignore_ascii_case("true") => Ok(true),
+        v if v.eq_ignore_ascii_case("false") => Ok(false),
+        _ => Err(invalid_query_context_extension(format!(
+            "Invalid value for {}: {}",
+            FLOW_RETURN_REGION_SEQ, value
+        ))),
+    }
+}
+
+fn invalid_query_context_extension(reason: String) -> Error {
+    InvalidQueryContextExtensionSnafu { reason }.build()
+}
+
+#[cfg(test)]
+mod flow_extension_tests {
+    use super::*;
+
+    #[test]
+    fn test_parse_flow_extensions_returns_none_for_non_flow_query() {
+        let exts = HashMap::new();
+        let parsed = FlowQueryExtensions::parse_flow_extensions(&exts).unwrap();
+
+        assert_eq!(parsed, None);
+    }
+
+    #[test]
+    fn test_parse_flow_extensions_memtable_only_success() {
+        let exts = HashMap::from([
+            (
+                FLOW_INCREMENTAL_MODE.to_string(),
+                FLOW_INCREMENTAL_MODE_MEMTABLE_ONLY.to_string(),
+            ),
+            (
+                FLOW_INCREMENTAL_AFTER_SEQS.to_string(),
+                r#"{"1":10,"2":20}"#.to_string(),
+            ),
+            (FLOW_RETURN_REGION_SEQ.to_string(), "true".to_string()),
+            (FLOW_SINK_TABLE_ID.to_string(), "1024".to_string()),
+        ]);
+
+        let parsed = FlowQueryExtensions::parse_flow_extensions(&exts)
+            .unwrap()
+            .unwrap();
+        assert_eq!(
+            parsed.incremental_mode,
+            Some(FlowIncrementalMode::MemtableOnly)
+        );
+        assert_eq!(
+            parsed.incremental_after_seqs.unwrap(),
+            HashMap::from([(1, 10), (2, 20)])
+        );
+        assert!(parsed.return_region_seq);
+        assert_eq!(parsed.sink_table_id, Some(1024));
+    }
+
+    #[test]
+    fn test_parse_flow_extensions_mode_requires_after_seqs() {
+        let exts = HashMap::from([(
+            FLOW_INCREMENTAL_MODE.to_string(),
+            FLOW_INCREMENTAL_MODE_MEMTABLE_ONLY.to_string(),
+        )]);
+
+        let err = FlowQueryExtensions::parse_flow_extensions(&exts).unwrap_err();
+        assert!(format!("{err}").contains(FLOW_INCREMENTAL_AFTER_SEQS));
+    }
+
+    #[test]
+    fn test_parse_flow_extensions_invalid_mode() {
+        let exts = HashMap::from([(FLOW_INCREMENTAL_MODE.to_string(), "foo".to_string())]);
+
+        let err = FlowQueryExtensions::parse_flow_extensions(&exts).unwrap_err();
+        assert!(format!("{err}").contains(FLOW_INCREMENTAL_MODE));
+    }
+
+    #[test]
+    fn test_parse_flow_extensions_invalid_after_seqs_json() {
+        let exts = HashMap::from([
+            (
+                FLOW_INCREMENTAL_MODE.to_string(),
+                FLOW_INCREMENTAL_MODE_MEMTABLE_ONLY.to_string(),
+            ),
+            (
+                FLOW_INCREMENTAL_AFTER_SEQS.to_string(),
+                "not-json".to_string(),
+            ),
+        ]);
+
+        let err = FlowQueryExtensions::parse_flow_extensions(&exts).unwrap_err();
+        assert!(format!("{err}").contains(FLOW_INCREMENTAL_AFTER_SEQS));
+    }
+
+    #[test]
+    fn test_parse_flow_extensions_after_seqs_string_values() {
+        let exts = HashMap::from([(
+            FLOW_INCREMENTAL_AFTER_SEQS.to_string(),
+            r#"{"1":"10","2":"20"}"#.to_string(),
+        )]);
+
+        let parsed = FlowQueryExtensions::parse_flow_extensions(&exts)
+            .unwrap()
+            .unwrap();
+        assert_eq!(
+            parsed.incremental_after_seqs.unwrap(),
+            HashMap::from([(1, 10), (2, 20)])
+        );
+    }
+
+    #[test]
+    fn test_parse_flow_extensions_after_seqs_invalid_value_type() {
+        let exts = HashMap::from([(
+            FLOW_INCREMENTAL_AFTER_SEQS.to_string(),
+            r#"{"1":true}"#.to_string(),
+        )]);
+
+        let err = FlowQueryExtensions::parse_flow_extensions(&exts).unwrap_err();
+        assert!(format!("{err}").contains(FLOW_INCREMENTAL_AFTER_SEQS));
+    }
+
+    #[test]
+    fn test_parse_flow_extensions_invalid_sink_table_id() {
+        let exts = HashMap::from([(FLOW_SINK_TABLE_ID.to_string(), "x".to_string())]);
+
+        let err = FlowQueryExtensions::parse_flow_extensions(&exts).unwrap_err();
+        assert!(format!("{err}").contains(FLOW_SINK_TABLE_ID));
+    }
+
+    #[test]
+    fn test_validate_for_scan_missing_source_region() {
+        let source_region_id = RegionId::new(100, 2);
+        let existing_region_id = RegionId::new(100, 1);
+        let exts = HashMap::from([
+            (
+                FLOW_INCREMENTAL_MODE.to_string(),
+                FLOW_INCREMENTAL_MODE_MEMTABLE_ONLY.to_string(),
+            ),
+            (
+                FLOW_INCREMENTAL_AFTER_SEQS.to_string(),
+                format!(r#"{{"{}":10}}"#, existing_region_id.as_u64()),
+            ),
+        ]);
+
+        let parsed = FlowQueryExtensions::parse_flow_extensions(&exts)
+            .unwrap()
+            .unwrap();
+        let err = parsed.validate_for_scan(source_region_id).unwrap_err();
+        assert!(format!("{err}").contains("Missing region"));
+    }
+
+    #[test]
+    fn test_validate_for_scan_sink_table_excluded() {
+        let source_region_id = RegionId::new(1024, 1);
+        let exts = HashMap::from([
+            (
+                FLOW_INCREMENTAL_MODE.to_string(),
+                FLOW_INCREMENTAL_MODE_MEMTABLE_ONLY.to_string(),
+            ),
+            (
+                FLOW_INCREMENTAL_AFTER_SEQS.to_string(),
+                format!(r#"{{"{}":10}}"#, source_region_id.as_u64()),
+            ),
+            (FLOW_SINK_TABLE_ID.to_string(), "1024".to_string()),
+        ]);
+
+        let parsed = FlowQueryExtensions::parse_flow_extensions(&exts)
+            .unwrap()
+            .unwrap();
+        let apply_incremental = parsed.validate_for_scan(source_region_id).unwrap();
+        assert!(!apply_incremental);
+    }
+
+    #[test]
+    fn test_should_collect_region_watermark_defaults_false() {
+        let parsed = FlowQueryExtensions::default();
+        assert!(!parsed.should_collect_region_watermark());
+    }
+
+    #[test]
+    fn test_should_collect_region_watermark_true_for_return_region_seq() {
+        let parsed = FlowQueryExtensions {
+            return_region_seq: true,
+            ..Default::default()
+        };
+        assert!(parsed.should_collect_region_watermark());
+    }
+
+    #[test]
+    fn test_should_collect_region_watermark_true_for_incremental_query() {
+        let parsed = FlowQueryExtensions {
+            incremental_after_seqs: Some(HashMap::from([(1, 10)])),
+            ..Default::default()
+        };
+        assert!(parsed.should_collect_region_watermark());
+    }
+
+    #[test]
+    fn test_should_collect_region_watermark_from_extensions() {
+        let exts = HashMap::from([(FLOW_RETURN_REGION_SEQ.to_string(), "true".to_string())]);
+        assert!(should_collect_region_watermark_from_extensions(&exts));
+
+        let exts = HashMap::from([(
+            FLOW_INCREMENTAL_AFTER_SEQS.to_string(),
+            r#"{"1":10}"#.to_string(),
+        )]);
+        assert!(should_collect_region_watermark_from_extensions(&exts));
+
+        let exts = HashMap::from([(FLOW_RETURN_REGION_SEQ.to_string(), "false".to_string())]);
+        assert!(!should_collect_region_watermark_from_extensions(&exts));
+        assert!(!should_collect_region_watermark_from_extensions(
+            &HashMap::new()
+        ));
+    }
+
+    #[test]
+    fn test_parse_flow_extensions_return_region_seq_only_returns_some() {
+        let exts = HashMap::from([(FLOW_RETURN_REGION_SEQ.to_string(), "true".to_string())]);
+
+        let parsed = FlowQueryExtensions::parse_flow_extensions(&exts)
+            .unwrap()
+            .unwrap();
+
+        assert!(parsed.return_region_seq);
+    }
+
+    #[test]
+    fn test_parse_flow_extensions_sink_table_only_returns_some() {
+        let exts = HashMap::from([(FLOW_SINK_TABLE_ID.to_string(), "1024".to_string())]);
+
+        let parsed = FlowQueryExtensions::parse_flow_extensions(&exts)
+            .unwrap()
+            .unwrap();
+
+        assert_eq!(parsed.sink_table_id, Some(1024));
+    }
+
+    #[test]
+    fn test_parse_flow_extensions_incremental_after_seqs_only_returns_some() {
+        let exts = HashMap::from([(
+            FLOW_INCREMENTAL_AFTER_SEQS.to_string(),
+            r#"{"1":10}"#.to_string(),
+        )]);
+
+        let parsed = FlowQueryExtensions::parse_flow_extensions(&exts)
+            .unwrap()
+            .unwrap();
+
+        assert_eq!(
+            parsed.incremental_after_seqs,
+            Some(HashMap::from([(1, 10)]))
+        );
+    }
+}
diff --git a/src/query/src/part_sort.rs b/src/query/src/part_sort.rs
index e12479cc5a..933f0e47cf 100644
--- a/src/query/src/part_sort.rs
+++ b/src/query/src/part_sort.rs
@@ -56,7 +56,7 @@ use snafu::location;
 use store_api::region_engine::PartitionRange;
 
 use crate::error::Result;
-use crate::window_sort::check_partition_range_monotonicity;
+use crate::window_sort::{check_partition_range_monotonicity, project_partition_range_for_sort};
 use crate::{array_iter_helper, downcast_ts_array};
 
 /// Get the primary end of a `PartitionRange` based on sort direction.
@@ -237,14 +237,10 @@ impl ExecutionPlan for PartSortExec {
         } else {
             internal_err!("No children found")?
         };
-        // create a new dynamic filter when with_new_children, as the old filter is bound to the old input and cannot be reused
-        let new = Self::try_new(
-            self.expression.clone(),
-            self.limit,
-            self.partition_ranges.clone(),
-            new_input.clone(),
-        )?;
-        Ok(Arc::new(new))
+        let mut new_exec = self.as_ref().clone();
+        new_exec.input = new_input.clone();
+        new_exec.properties = new_input.properties().clone();
+        Ok(Arc::new(new_exec))
     }
 
     fn execute(
@@ -477,6 +473,7 @@ impl PartSortStream {
                 snafu::location!()
             )?
         };
+        let cur_range = project_partition_range_for_sort(cur_range, sort_column.data_type())?;
 
         downcast_ts_array!(
             sort_column.data_type() => (array_check_helper, sort_column, cur_range, min_max_idx),
@@ -510,7 +507,10 @@ impl PartSortStream {
                 snafu::location!()
             )?;
         }
-        let cur_range = self.partition_ranges[self.cur_part_idx];
+        let cur_range = project_partition_range_for_sort(
+            self.partition_ranges[self.cur_part_idx],
+            sort_column.data_type(),
+        )?;
 
         let sort_column_iter = downcast_ts_array!(
             sort_column.data_type() => (array_iter_helper, sort_column),
diff --git a/src/query/src/planner.rs b/src/query/src/planner.rs
index 44c9bc3956..b95803a52b 100644
--- a/src/query/src/planner.rs
+++ b/src/query/src/planner.rs
@@ -26,8 +26,9 @@ use common_telemetry::tracing;
 use datafusion::common::{DFSchema, plan_err};
 use datafusion::execution::context::SessionState;
 use datafusion::sql::planner::PlannerContext;
-use datafusion_common::ToDFSchema;
 use datafusion_common::tree_node::{TreeNode, TreeNodeRecursion};
+use datafusion_common::{ScalarValue, ToDFSchema};
+use datafusion_expr::expr::{Exists, InSubquery};
 use datafusion_expr::{
     Analyze, Explain, ExplainFormat, Expr as DfExpr, LogicalPlan, LogicalPlanBuilder, PlanType,
     ToStringifiedPlan, col,
@@ -277,17 +278,22 @@ impl DfLogicalPlanner {
         let table_provider = DfTableSourceProvider::new(
             self.engine_state.catalog_manager().clone(),
             self.engine_state.disallow_cross_catalog_query(),
-            query_ctx,
+            query_ctx.clone(),
             plan_decoder,
             self.session_state
                 .config_options()
                 .sql_parser
                 .enable_ident_normalization,
         );
-        PromPlanner::stmt_to_plan(table_provider, stmt, &self.engine_state)
+        let plan = PromPlanner::stmt_to_plan(table_provider, stmt, &self.engine_state)
             .await
             .map_err(BoxedError::new)
-            .context(QueryPlanSnafu)
+            .context(QueryPlanSnafu)?;
+
+        let context = QueryEngineContext::new(self.session_state.clone(), query_ctx);
+        Ok(self
+            .engine_state
+            .optimize_by_extension_rules(plan, &context)?)
     }
 
     #[tracing::instrument(skip_all)]
@@ -424,9 +430,20 @@ impl DfLogicalPlanner {
         let mut placeholder_types = HashMap::new();
         let mut casted_placeholders = HashSet::new();
 
+        Self::extract_from_plan(plan, &mut placeholder_types, &mut casted_placeholders)?;
+
+        Ok(placeholder_types)
+    }
+
+    fn extract_from_plan(
+        plan: &LogicalPlan,
+        placeholder_types: &mut HashMap<String, Option<DataType>>,
+        casted_placeholders: &mut HashSet<String>,
+    ) -> Result<()> {
         plan.apply(|node| {
             for expr in node.expressions() {
                 let _ = expr.apply(|e| {
+                    // Handle casted placeholders
                     if let DfExpr::Cast(cast) = e
                         && let DfExpr::Placeholder(ph) = &*cast.expr
                     {
@@ -434,6 +451,20 @@ impl DfLogicalPlanner {
                         casted_placeholders.insert(ph.id.clone());
                     }
 
+                    // Handle arrow_cast(Placeholder, 'type_string') generated by SQL rewriter
+                    if let DfExpr::ScalarFunction(scalar_func) = e
+                        && scalar_func.name() == "arrow_cast"
+                        && scalar_func.args.len() == 2
+                        && let DfExpr::Placeholder(ph) = &scalar_func.args[0]
+                        && let DfExpr::Literal(ScalarValue::Utf8(Some(type_str)), _) =
+                            &scalar_func.args[1]
+                        && let Ok(data_type) = type_str.parse::<DataType>()
+                    {
+                        placeholder_types.insert(ph.id.clone(), Some(data_type));
+                        casted_placeholders.insert(ph.id.clone());
+                    }
+
+                    // Handle bare (non-casted) placeholders
                     if let DfExpr::Placeholder(ph) = e
                         && !casted_placeholders.contains(&ph.id)
                         && !placeholder_types.contains_key(&ph.id)
@@ -441,13 +472,26 @@ impl DfLogicalPlanner {
                         placeholder_types.insert(ph.id.clone(), None);
                     }
 
+                    // Recurse into subquery plans embedded in expressions
+                    match e {
+                        DfExpr::Exists(Exists { subquery, .. })
+                        | DfExpr::InSubquery(InSubquery { subquery, .. })
+                        | DfExpr::ScalarSubquery(subquery) => {
+                            Self::extract_from_plan(
+                                &subquery.subquery,
+                                placeholder_types,
+                                casted_placeholders,
+                            )?;
+                        }
+                        _ => {}
+                    }
+
                     Ok(TreeNodeRecursion::Continue)
                 });
             }
             Ok(TreeNodeRecursion::Continue)
         })?;
-
-        Ok(placeholder_types)
+        Ok(())
     }
 
     /// Gets inferred parameter types from a logical plan.
@@ -545,15 +589,22 @@ mod tests {
     use std::sync::Arc;
 
     use arrow_schema::DataType;
+    use catalog::RegisterTableRequest;
+    use catalog::memory::MemoryCatalogManager;
+    use common_catalog::consts::{DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME};
     use datatypes::prelude::ConcreteDataType;
     use datatypes::schema::{ColumnSchema, Schema};
     use session::context::QueryContext;
+    use store_api::metric_engine_consts::{
+        DATA_SCHEMA_TABLE_ID_COLUMN_NAME, DATA_SCHEMA_TSID_COLUMN_NAME, LOGICAL_TABLE_METADATA_KEY,
+        METRIC_ENGINE_NAME,
+    };
     use table::metadata::{TableInfoBuilder, TableMetaBuilder};
     use table::test_util::EmptyTable;
 
     use super::*;
-    use crate::QueryEngineRef;
-    use crate::parser::QueryLanguageParser;
+    use crate::parser::{PromQuery, QueryLanguageParser};
+    use crate::{QueryEngineFactory, QueryEngineRef};
 
     async fn create_test_engine() -> QueryEngineRef {
         let columns = vec![
@@ -574,6 +625,109 @@ mod tests {
         crate::tests::new_query_engine_with_table(table)
     }
 
+    fn create_promql_test_engine() -> QueryEngineRef {
+        let catalog_manager = MemoryCatalogManager::with_default_setup();
+        let physical_table_name = "phy";
+        let physical_table_id = 999u32;
+
+        let physical_schema = Arc::new(Schema::new(vec![
+            ColumnSchema::new(
+                DATA_SCHEMA_TABLE_ID_COLUMN_NAME.to_string(),
+                ConcreteDataType::uint32_datatype(),
+                false,
+            ),
+            ColumnSchema::new(
+                DATA_SCHEMA_TSID_COLUMN_NAME.to_string(),
+                ConcreteDataType::uint64_datatype(),
+                false,
+            ),
+            ColumnSchema::new("tag_0", ConcreteDataType::string_datatype(), false),
+            ColumnSchema::new("tag_1", ConcreteDataType::string_datatype(), false),
+            ColumnSchema::new(
+                "timestamp",
+                ConcreteDataType::timestamp_millisecond_datatype(),
+                false,
+            )
+            .with_time_index(true),
+            ColumnSchema::new("field_0", ConcreteDataType::float64_datatype(), true),
+        ]));
+        let physical_meta = TableMetaBuilder::empty()
+            .schema(physical_schema)
+            .primary_key_indices(vec![0, 1, 2, 3])
+            .value_indices(vec![4, 5])
+            .engine(METRIC_ENGINE_NAME.to_string())
+            .next_column_id(1024)
+            .build()
+            .unwrap();
+        let physical_info = TableInfoBuilder::default()
+            .table_id(physical_table_id)
+            .name(physical_table_name)
+            .meta(physical_meta)
+            .build()
+            .unwrap();
+        catalog_manager
+            .register_table_sync(RegisterTableRequest {
+                catalog: DEFAULT_CATALOG_NAME.to_string(),
+                schema: DEFAULT_SCHEMA_NAME.to_string(),
+                table_name: physical_table_name.to_string(),
+                table_id: physical_table_id,
+                table: EmptyTable::from_table_info(&physical_info),
+            })
+            .unwrap();
+
+        let mut options = table::requests::TableOptions::default();
+        options.extra_options.insert(
+            LOGICAL_TABLE_METADATA_KEY.to_string(),
+            physical_table_name.to_string(),
+        );
+        let logical_schema = Arc::new(Schema::new(vec![
+            ColumnSchema::new("tag_0", ConcreteDataType::string_datatype(), false),
+            ColumnSchema::new("tag_1", ConcreteDataType::string_datatype(), false),
+            ColumnSchema::new(
+                "timestamp",
+                ConcreteDataType::timestamp_millisecond_datatype(),
+                false,
+            )
+            .with_time_index(true),
+            ColumnSchema::new("field_0", ConcreteDataType::float64_datatype(), true),
+        ]));
+        let logical_meta = TableMetaBuilder::empty()
+            .schema(logical_schema)
+            .primary_key_indices(vec![0, 1])
+            .value_indices(vec![3])
+            .engine(METRIC_ENGINE_NAME.to_string())
+            .options(options)
+            .next_column_id(1024)
+            .build()
+            .unwrap();
+        let logical_info = TableInfoBuilder::default()
+            .table_id(1024)
+            .name("some_metric")
+            .meta(logical_meta)
+            .build()
+            .unwrap();
+        catalog_manager
+            .register_table_sync(RegisterTableRequest {
+                catalog: DEFAULT_CATALOG_NAME.to_string(),
+                schema: DEFAULT_SCHEMA_NAME.to_string(),
+                table_name: "some_metric".to_string(),
+                table_id: 1024,
+                table: EmptyTable::from_table_info(&logical_info),
+            })
+            .unwrap();
+
+        QueryEngineFactory::new(
+            catalog_manager,
+            None,
+            None,
+            None,
+            None,
+            false,
+            crate::options::QueryOptions::default(),
+        )
+        .query_engine()
+    }
+
     async fn parse_sql_to_plan(sql: &str) -> LogicalPlan {
         let stmt = QueryLanguageParser::parse_sql(sql, &QueryContext::arc()).unwrap();
         let engine = create_test_engine().await;
@@ -584,6 +738,25 @@ mod tests {
             .unwrap()
     }
 
+    async fn parse_promql_to_plan(query: &str) -> LogicalPlan {
+        let engine = create_promql_test_engine();
+        let query_ctx = QueryContext::arc();
+        let stmt = QueryLanguageParser::parse_promql(
+            &PromQuery {
+                query: query.to_string(),
+                start: "0".to_string(),
+                end: "10".to_string(),
+                step: "5s".to_string(),
+                lookback: "300s".to_string(),
+                alias: None,
+            },
+            &query_ctx,
+        )
+        .unwrap();
+
+        engine.planner().plan(&stmt, query_ctx).await.unwrap()
+    }
+
     #[tokio::test]
     async fn test_extract_placeholder_cast_types_multiple() {
         let plan = parse_sql_to_plan(
@@ -619,4 +792,115 @@ mod tests {
         assert_eq!(type_2, &Some(DataType::Utf8));
         assert_eq!(type_3, &Some(DataType::Int32));
     }
+
+    #[tokio::test]
+    async fn test_plan_pql_applies_extension_rules() {
+        for inner_agg in ["count", "sum", "avg", "min", "max", "stddev", "stdvar"] {
+            let plan = parse_promql_to_plan(&format!(
+                "sum(irate(some_metric[1h])) / scalar(count({inner_agg}(some_metric) by (tag_0)))"
+            ))
+            .await;
+            let plan_str = plan.display_indent_schema().to_string();
+            assert!(plan_str.contains("Distinct:"), "{inner_agg}: {plan_str}");
+        }
+    }
+
+    #[tokio::test]
+    async fn test_plan_pql_filters_null_only_groups_for_non_count_inner_aggs() {
+        let count_plan = parse_promql_to_plan("scalar(count(count(some_metric) by (tag_0)))").await;
+        let count_plan_str = count_plan.display_indent_schema().to_string();
+        assert!(
+            !count_plan_str.contains("field_0 IS NOT NULL"),
+            "{count_plan_str}"
+        );
+
+        for inner_agg in ["sum", "avg", "min", "max", "stddev", "stdvar"] {
+            let plan = parse_promql_to_plan(&format!(
+                "scalar(count({inner_agg}(some_metric) by (tag_0)))"
+            ))
+            .await;
+            let plan_str = plan.display_indent_schema().to_string();
+            assert!(
+                plan_str.contains("field_0 IS NOT NULL"),
+                "{inner_agg}: {plan_str}"
+            );
+        }
+    }
+
+    #[tokio::test]
+    async fn test_plan_pql_skips_extension_rules_for_non_direct_or_unsupported_inner_agg() {
+        for query in [
+            "sum(irate(some_metric[1h])) / scalar(count(sum(irate(some_metric[1h])) by (tag_0)))",
+            "sum(irate(some_metric[1h])) / scalar(count(group(some_metric) by (tag_0)))",
+        ] {
+            let plan = parse_promql_to_plan(query).await;
+            let plan_str = plan.display_indent_schema().to_string();
+            assert!(!plan_str.contains("Distinct:"), "{query}: {plan_str}");
+        }
+    }
+
+    #[tokio::test]
+    async fn test_plan_sql_does_not_apply_nested_count_rule() {
+        let plan = parse_sql_to_plan(
+            "SELECT id, count(inner_count) \
+             FROM ( \
+                 SELECT id, count(name) AS inner_count \
+                 FROM test \
+                 GROUP BY id \
+                 ORDER BY id \
+                 LIMIT 1000000 \
+             ) t \
+             GROUP BY id \
+             ORDER BY id",
+        )
+        .await;
+
+        let plan_str = plan.display_indent_schema().to_string();
+        assert!(!plan_str.contains("Distinct:"), "{plan_str}");
+    }
+
+    #[tokio::test]
+    async fn test_get_inferred_parameter_types_subquery() {
+        let plan = parse_sql_to_plan(
+            r#"SELECT * FROM test WHERE id = (SELECT id FROM test CROSS JOIN (SELECT parse_ident($1::TEXT) AS parts) p LIMIT 1)"#,
+        ).await;
+        let types = DfLogicalPlanner::get_inferred_parameter_types(&plan).unwrap();
+
+        assert_eq!(types.len(), 1);
+        let type_1 = types.get("$1").unwrap();
+        assert_eq!(type_1, &Some(DataType::Utf8));
+    }
+
+    #[tokio::test]
+    async fn test_get_inferred_parameter_types_insert() {
+        let plan = parse_sql_to_plan("INSERT INTO test (id, name) VALUES ($1, $2), ($3, $4)").await;
+        let types = DfLogicalPlanner::get_inferred_parameter_types(&plan).unwrap();
+
+        assert_eq!(types.len(), 4);
+        assert_eq!(types.get("$1"), Some(&Some(DataType::Int32)));
+        assert_eq!(types.get("$2"), Some(&Some(DataType::Utf8)));
+        assert_eq!(types.get("$3"), Some(&Some(DataType::Int32)));
+        assert_eq!(types.get("$4"), Some(&Some(DataType::Utf8)));
+    }
+
+    #[tokio::test]
+    async fn test_get_inferred_parameter_types_arrow_cast() {
+        let plan = parse_sql_to_plan("SELECT $1::INT64, $2::FLOAT64, $3::INT16, $4::INT32, $5::UINT8, $6::UINT16, $7::UINT32").await;
+        let types = DfLogicalPlanner::get_inferred_parameter_types(&plan).unwrap();
+
+        assert_eq!(types.get("$1"), Some(&Some(DataType::Int64)));
+        assert_eq!(types.get("$2"), Some(&Some(DataType::Float64)));
+        assert_eq!(types.get("$3"), Some(&Some(DataType::Int16)));
+        assert_eq!(types.get("$4"), Some(&Some(DataType::Int32)));
+        assert_eq!(types.get("$5"), Some(&Some(DataType::UInt8)));
+        assert_eq!(types.get("$6"), Some(&Some(DataType::UInt16)));
+        assert_eq!(types.get("$7"), Some(&Some(DataType::UInt32)));
+
+        let plan = parse_sql_to_plan("SELECT $1::INT8, $2::FLOAT8, $3::INT2, $4::INT8").await;
+        let types = DfLogicalPlanner::get_inferred_parameter_types(&plan).unwrap();
+
+        assert_eq!(types.get("$1"), Some(&Some(DataType::Int64)));
+        assert_eq!(types.get("$2"), Some(&Some(DataType::Float64)));
+        assert_eq!(types.get("$3"), Some(&Some(DataType::Int16)));
+    }
 }
diff --git a/src/query/src/promql/planner.rs b/src/query/src/promql/planner.rs
index 256b9d5abf..477bd7f5b3 100644
--- a/src/query/src/promql/planner.rs
+++ b/src/query/src/promql/planner.rs
@@ -523,8 +523,7 @@ impl PromPlanner {
                     .use_tsid
                     .then_some(DfExpr::Column(Column::from_name(
                         DATA_SCHEMA_TSID_COLUMN_NAME,
-                    )))
-                    .into_iter(),
+                    ))),
             )
             .chain(Some(self.create_time_index_column_expr()?));
 
@@ -710,7 +709,13 @@ impl PromPlanner {
                         self.ctx.table_name = Some("rhs".to_string());
                     }
                 }
-                let mut field_columns = left_field_columns.iter().zip(right_field_columns.iter());
+                let (output_field_columns, field_columns) =
+                    Self::align_binary_field_columns(&left_field_columns, &right_field_columns);
+                // PromQL binary arithmetic only combines the shared prefix of value columns.
+                // Keep the output field count aligned with that zipped prefix so planning
+                // remains stable even when the two sides have uneven multi-field schemas.
+                self.ctx.field_columns = output_field_columns;
+                let mut field_columns = field_columns.into_iter();
 
                 let join_plan = self.join_on_non_field_columns(
                     left_input,
@@ -811,11 +816,10 @@ impl PromPlanner {
 
         // Preserve `__tsid` if present, so it can still be used internally downstream. It's
         // stripped from the final output anyway.
-        if context.use_tsid
-            && let Ok(tsid_col) =
-                schema.qualified_field_with_name(Some(table_ref), DATA_SCHEMA_TSID_COLUMN_NAME)
+        if let Some(tsid_col) =
+            Self::optional_tsid_projection(schema, Some(table_ref), context.use_tsid)
         {
-            project_exprs.push(DfExpr::Column(tsid_col.into()));
+            project_exprs.push(tsid_col);
         }
 
         let plan = LogicalPlanBuilder::from(input)
@@ -1318,8 +1322,7 @@ impl PromPlanner {
                         .use_tsid
                         .then_some(DfExpr::Column(Column::new_unqualified(
                             DATA_SCHEMA_TSID_COLUMN_NAME,
-                        )))
-                        .into_iter(),
+                        ))),
                 )
                 .chain(Some(self.create_time_index_column_expr()?))
                 .collect::<Vec<_>>();
@@ -1833,15 +1836,10 @@ impl PromPlanner {
                         .iter()
                         .map(|tag| DfExpr::Column(Column::from_name(tag))),
                 )
-                .chain(
-                    self.ctx
-                        .use_tsid
-                        .then_some(DfExpr::Column(Column::new(
-                            Some(table_ref.clone()),
-                            DATA_SCHEMA_TSID_COLUMN_NAME.to_string(),
-                        )))
-                        .into_iter(),
-                )
+                .chain(self.ctx.use_tsid.then_some(DfExpr::Column(Column::new(
+                    Some(table_ref.clone()),
+                    DATA_SCHEMA_TSID_COLUMN_NAME.to_string(),
+                ))))
                 .chain(Some(DfExpr::Alias(Alias {
                     expr: Box::new(DfExpr::Cast(Cast {
                         expr: Box::new(self.create_time_index_column_expr()?),
@@ -1879,8 +1877,7 @@ impl PromPlanner {
                         .use_tsid
                         .then_some(DfExpr::Column(Column::from_name(
                             DATA_SCHEMA_TSID_COLUMN_NAME,
-                        )))
-                        .into_iter(),
+                        ))),
                 )
                 .chain(Some(self.create_time_index_column_expr()?))
                 .collect::<Vec<_>>();
@@ -3328,28 +3325,55 @@ impl PromPlanner {
     fn prom_token_to_binary_expr_builder(
         token: TokenType,
     ) -> Result<Box<dyn Fn(DfExpr, DfExpr) -> Result<DfExpr>>> {
+        let cast_float = |expr| {
+            if matches!(
+                &expr,
+                DfExpr::Cast(Cast {
+                    data_type: ArrowDataType::Float64,
+                    ..
+                })
+            ) || matches!(&expr, DfExpr::Literal(ScalarValue::Float64(_), _))
+            {
+                expr
+            } else {
+                DfExpr::Cast(Cast {
+                    expr: Box::new(expr),
+                    data_type: ArrowDataType::Float64,
+                })
+            }
+        };
         match token.id() {
-            token::T_ADD => Ok(Box::new(|lhs, rhs| Ok(lhs + rhs))),
-            token::T_SUB => Ok(Box::new(|lhs, rhs| Ok(lhs - rhs))),
-            token::T_MUL => Ok(Box::new(|lhs, rhs| Ok(lhs * rhs))),
-            token::T_DIV => Ok(Box::new(|lhs, rhs| Ok(lhs / rhs))),
-            token::T_MOD => Ok(Box::new(|lhs: DfExpr, rhs| Ok(lhs % rhs))),
+            token::T_ADD => Ok(Box::new(move |lhs, rhs| {
+                Ok(cast_float(lhs) + cast_float(rhs))
+            })),
+            token::T_SUB => Ok(Box::new(move |lhs, rhs| {
+                Ok(cast_float(lhs) - cast_float(rhs))
+            })),
+            token::T_MUL => Ok(Box::new(move |lhs, rhs| {
+                Ok(cast_float(lhs) * cast_float(rhs))
+            })),
+            token::T_DIV => Ok(Box::new(move |lhs, rhs| {
+                Ok(cast_float(lhs) / cast_float(rhs))
+            })),
+            token::T_MOD => Ok(Box::new(move |lhs: DfExpr, rhs| {
+                Ok(cast_float(lhs) % cast_float(rhs))
+            })),
             token::T_EQLC => Ok(Box::new(|lhs, rhs| Ok(lhs.eq(rhs)))),
             token::T_NEQ => Ok(Box::new(|lhs, rhs| Ok(lhs.not_eq(rhs)))),
             token::T_GTR => Ok(Box::new(|lhs, rhs| Ok(lhs.gt(rhs)))),
             token::T_LSS => Ok(Box::new(|lhs, rhs| Ok(lhs.lt(rhs)))),
             token::T_GTE => Ok(Box::new(|lhs, rhs| Ok(lhs.gt_eq(rhs)))),
             token::T_LTE => Ok(Box::new(|lhs, rhs| Ok(lhs.lt_eq(rhs)))),
-            token::T_POW => Ok(Box::new(|lhs, rhs| {
+            token::T_POW => Ok(Box::new(move |lhs, rhs| {
                 Ok(DfExpr::ScalarFunction(ScalarFunction {
                     func: datafusion_functions::math::power(),
-                    args: vec![lhs, rhs],
+                    args: vec![cast_float(lhs), cast_float(rhs)],
                 }))
             })),
-            token::T_ATAN2 => Ok(Box::new(|lhs, rhs| {
+            token::T_ATAN2 => Ok(Box::new(move |lhs, rhs| {
                 Ok(DfExpr::ScalarFunction(ScalarFunction {
                     func: datafusion_functions::math::atan2(),
-                    args: vec![lhs, rhs],
+                    args: vec![cast_float(lhs), cast_float(rhs)],
                 }))
             })),
             _ => UnexpectedTokenSnafu { token }.fail(),
@@ -3379,6 +3403,97 @@ impl PromPlanner {
         )
     }
 
+    fn align_binary_field_columns<'a>(
+        left_field_columns: &'a [String],
+        right_field_columns: &'a [String],
+    ) -> (Vec<String>, Vec<(&'a String, &'a String)>) {
+        let field_pairs = left_field_columns
+            .iter()
+            .zip(right_field_columns.iter())
+            .collect::<Vec<_>>();
+        let output_field_columns = field_pairs
+            .iter()
+            .map(|(left_col_name, _)| (*left_col_name).clone())
+            .collect();
+        (output_field_columns, field_pairs)
+    }
+
+    fn plan_has_tsid_column(plan: &LogicalPlan) -> bool {
+        plan.schema()
+            .fields()
+            .iter()
+            .any(|field| field.name() == DATA_SCHEMA_TSID_COLUMN_NAME)
+    }
+
+    fn optional_tsid_projection(
+        schema: &DFSchemaRef,
+        table_ref: Option<&TableReference>,
+        keep_tsid: bool,
+    ) -> Option<DfExpr> {
+        keep_tsid.then_some(()).and_then(|_| {
+            schema
+                .qualified_field_with_name(table_ref, DATA_SCHEMA_TSID_COLUMN_NAME)
+                .ok()
+                .map(|field| DfExpr::Column(field.into()))
+        })
+    }
+
+    fn binary_join_key_columns(
+        &self,
+        left: &LogicalPlan,
+        right: &LogicalPlan,
+        only_join_time_index: bool,
+        modifier: &Option<BinModifier>,
+    ) -> (BTreeSet<String>, BTreeSet<String>) {
+        let use_tsid_join = !only_join_time_index
+            && modifier.as_ref().is_none_or(|modifier| {
+                modifier.matching.is_none()
+                    && matches!(modifier.card, VectorMatchCardinality::OneToOne)
+            })
+            && Self::plan_has_tsid_column(left)
+            && Self::plan_has_tsid_column(right);
+
+        let (mut left_tag_columns, mut right_tag_columns) = if use_tsid_join {
+            (
+                BTreeSet::from([DATA_SCHEMA_TSID_COLUMN_NAME.to_string()]),
+                BTreeSet::from([DATA_SCHEMA_TSID_COLUMN_NAME.to_string()]),
+            )
+        } else {
+            let left_tag_columns = if only_join_time_index {
+                BTreeSet::new()
+            } else {
+                self.ctx
+                    .tag_columns
+                    .iter()
+                    .cloned()
+                    .collect::<BTreeSet<_>>()
+            };
+            let right_tag_columns = left_tag_columns.clone();
+            (left_tag_columns, right_tag_columns)
+        };
+
+        if !use_tsid_join
+            && let Some(modifier) = modifier
+            && let Some(matching) = &modifier.matching
+        {
+            match matching {
+                LabelModifier::Include(on) => {
+                    let mask = on.labels.iter().cloned().collect::<BTreeSet<_>>();
+                    left_tag_columns = left_tag_columns.intersection(&mask).cloned().collect();
+                    right_tag_columns = right_tag_columns.intersection(&mask).cloned().collect();
+                }
+                LabelModifier::Exclude(ignoring) => {
+                    for label in &ignoring.labels {
+                        let _ = left_tag_columns.remove(label);
+                        let _ = right_tag_columns.remove(label);
+                    }
+                }
+            }
+        }
+
+        (left_tag_columns, right_tag_columns)
+    }
+
     /// Build a inner join on time index column and tag columns to concat two logical plans.
     /// When `only_join_time_index == true` we only join on the time index, because these two plan may not have the same tag columns
     #[allow(clippy::too_many_arguments)]
@@ -3393,40 +3508,8 @@ impl PromPlanner {
         only_join_time_index: bool,
         modifier: &Option<BinModifier>,
     ) -> Result<LogicalPlan> {
-        let mut left_tag_columns = if only_join_time_index {
-            BTreeSet::new()
-        } else {
-            self.ctx
-                .tag_columns
-                .iter()
-                .cloned()
-                .collect::<BTreeSet<_>>()
-        };
-        let mut right_tag_columns = left_tag_columns.clone();
-
-        // apply modifier
-        if let Some(modifier) = modifier {
-            // apply label modifier
-            if let Some(matching) = &modifier.matching {
-                match matching {
-                    // keeps columns mentioned in `on`
-                    LabelModifier::Include(on) => {
-                        let mask = on.labels.iter().cloned().collect::<BTreeSet<_>>();
-                        left_tag_columns = left_tag_columns.intersection(&mask).cloned().collect();
-                        right_tag_columns =
-                            right_tag_columns.intersection(&mask).cloned().collect();
-                    }
-                    // removes columns memtioned in `ignoring`
-                    LabelModifier::Exclude(ignoring) => {
-                        // doesn't check existence of label
-                        for label in &ignoring.labels {
-                            let _ = left_tag_columns.remove(label);
-                            let _ = right_tag_columns.remove(label);
-                        }
-                    }
-                }
-            }
-        }
+        let (mut left_tag_columns, mut right_tag_columns) =
+            self.binary_join_key_columns(&left, &right, only_join_time_index, modifier);
 
         // push time index column if it exists
         if let (Some(left_time_index_column), Some(right_time_index_column)) =
@@ -3842,17 +3925,17 @@ impl PromPlanner {
     where
         F: FnMut(&String) -> Result<DfExpr>,
     {
+        let table_ref = self.ctx.table_name.clone().map(TableReference::bare);
         let non_field_columns_iter = self
             .ctx
             .tag_columns
             .iter()
             .chain(self.ctx.time_index_column.iter())
-            .map(|col| {
-                Ok(DfExpr::Column(Column::new(
-                    self.ctx.table_name.clone().map(TableReference::bare),
-                    col,
-                )))
-            });
+            .map(|col| Ok(DfExpr::Column(Column::new(table_ref.clone(), col))));
+        let tsid_iter =
+            Self::optional_tsid_projection(input.schema(), table_ref.as_ref(), self.ctx.use_tsid)
+                .into_iter()
+                .map(Ok);
 
         // build computation exprs
         let result_field_columns = self
@@ -3874,6 +3957,7 @@ impl PromPlanner {
 
         // chain non-field columns (unchanged) and field columns (applied computation then alias)
         let project_fields = non_field_columns_iter
+            .chain(tsid_iter)
             .chain(field_columns_iter)
             .collect::<Result<Vec<_>>>()?;
 
@@ -4038,6 +4122,7 @@ mod test {
     use table::test_util::EmptyTable;
 
     use super::*;
+    use crate::QueryEngineContext;
     use crate::options::QueryOptions;
     use crate::parser::QueryLanguageParser;
 
@@ -4065,6 +4150,76 @@ mod test {
         )
     }
 
+    async fn build_optimized_promql_plan(
+        table_provider: DfTableSourceProvider,
+        eval_stmt: &EvalStmt,
+    ) -> LogicalPlan {
+        let state = build_query_engine_state();
+        let raw_plan = PromPlanner::stmt_to_plan(table_provider, eval_stmt, &state)
+            .await
+            .unwrap();
+        let context = QueryEngineContext::new(state.session_state(), QueryContext::arc());
+        state
+            .optimize_by_extension_rules(raw_plan, &context)
+            .unwrap()
+    }
+
+    async fn build_optimized_tsid_plan(
+        query: &str,
+        num_tag: usize,
+        num_field: usize,
+        end_secs: u64,
+        lookback_secs: u64,
+    ) -> String {
+        let eval_stmt = EvalStmt {
+            expr: parser::parse(query).unwrap(),
+            start: UNIX_EPOCH,
+            end: UNIX_EPOCH
+                .checked_add(Duration::from_secs(end_secs))
+                .unwrap(),
+            interval: Duration::from_secs(5),
+            lookback_delta: Duration::from_secs(lookback_secs),
+        };
+        let table_provider = build_test_table_provider_with_tsid(
+            &[(DEFAULT_SCHEMA_NAME.to_string(), "some_metric".to_string())],
+            num_tag,
+            num_field,
+        )
+        .await;
+
+        build_optimized_promql_plan(table_provider, &eval_stmt)
+            .await
+            .display_indent_schema()
+            .to_string()
+    }
+
+    async fn assert_nested_count_rewrite_applies(query: &str, expected_outer_agg: &str) {
+        let plan_str = build_optimized_tsid_plan(query, 2, 1, 100_000, 1).await;
+
+        assert!(plan_str.contains("PromSeriesDivide: tags=[\"__tsid\"]"));
+        assert!(plan_str.contains("Projection: some_metric.timestamp, some_metric.tag_0"));
+        assert!(plan_str.contains("Distinct:"));
+        assert!(plan_str.contains(expected_outer_agg), "{plan_str}");
+        assert!(!plan_str.contains("PromSeriesDivide: tags=[\"tag_0\"]"));
+    }
+
+    async fn assert_nested_count_rewrite_missing(query: &str, num_tag: usize, lookback_secs: u64) {
+        let plan_str = build_optimized_tsid_plan(query, num_tag, 1, 100_000, lookback_secs).await;
+        assert!(!plan_str.contains("Distinct:"), "{plan_str}");
+    }
+
+    fn build_eval_stmt(expr: &str) -> EvalStmt {
+        EvalStmt {
+            expr: parser::parse(expr).unwrap(),
+            start: UNIX_EPOCH,
+            end: UNIX_EPOCH
+                .checked_add(Duration::from_secs(100_000))
+                .unwrap(),
+            interval: Duration::from_secs(5),
+            lookback_delta: Duration::from_secs(1),
+        }
+    }
+
     async fn build_test_table_provider(
         table_name_tuples: &[(String, String)],
         num_tag: usize,
@@ -4136,11 +4291,27 @@ mod test {
         table_name_tuples: &[(String, String)],
         num_tag: usize,
         num_field: usize,
+    ) -> DfTableSourceProvider {
+        let table_specs = table_name_tuples
+            .iter()
+            .map(|(schema_name, table_name)| ((schema_name.clone(), table_name.clone()), num_field))
+            .collect::<Vec<_>>();
+        build_test_table_provider_with_tsid_fields(&table_specs, num_tag).await
+    }
+
+    async fn build_test_table_provider_with_tsid_fields(
+        table_specs: &[((String, String), usize)],
+        num_tag: usize,
     ) -> DfTableSourceProvider {
         let catalog_list = MemoryCatalogManager::with_default_setup();
 
         let physical_table_name = "phy";
         let physical_table_id = 999u32;
+        let physical_num_field = table_specs
+            .iter()
+            .map(|(_, num_field)| *num_field)
+            .max()
+            .unwrap_or(0);
 
         // Register a metric engine physical table with internal columns.
         {
@@ -4171,7 +4342,7 @@ mod test {
                 )
                 .with_time_index(true),
             );
-            for i in 0..num_field {
+            for i in 0..physical_num_field {
                 columns.push(ColumnSchema::new(
                     format!("field_{i}"),
                     ConcreteDataType::float64_datatype(),
@@ -4184,7 +4355,7 @@ mod test {
             let table_meta = TableMetaBuilder::empty()
                 .schema(schema)
                 .primary_key_indices(primary_key_indices)
-                .value_indices((2 + num_tag..2 + num_tag + 1 + num_field).collect())
+                .value_indices((2 + num_tag..2 + num_tag + 1 + physical_num_field).collect())
                 .engine(METRIC_ENGINE_NAME.to_string())
                 .next_column_id(1024)
                 .build()
@@ -4211,7 +4382,7 @@ mod test {
         }
 
         // Register metric engine logical tables without `__tsid`, referencing the physical table.
-        for (idx, (schema_name, table_name)) in table_name_tuples.iter().enumerate() {
+        for (idx, ((schema_name, table_name), num_field)) in table_specs.iter().enumerate() {
             let mut columns = vec![];
             for i in 0..num_tag {
                 columns.push(ColumnSchema::new(
@@ -4228,7 +4399,7 @@ mod test {
                 )
                 .with_time_index(true),
             );
-            for i in 0..num_field {
+            for i in 0..*num_field {
                 columns.push(ColumnSchema::new(
                     format!("field_{i}"),
                     ConcreteDataType::float64_datatype(),
@@ -4246,7 +4417,7 @@ mod test {
             let table_meta = TableMetaBuilder::empty()
                 .schema(schema)
                 .primary_key_indices((0..num_tag).collect())
-                .value_indices((num_tag + 1..num_tag + 1 + num_field).collect())
+                .value_indices((num_tag + 1..num_tag + 1 + *num_field).collect())
                 .engine(METRIC_ENGINE_NAME.to_string())
                 .options(options)
                 .next_column_id(1024)
@@ -4683,6 +4854,396 @@ mod test {
         assert!(format!("{exec:?}").contains("reuse_all_non_sample_columns: true"));
     }
 
+    #[tokio::test]
+    async fn default_binary_join_uses_tsid_when_available() {
+        let eval_stmt = build_eval_stmt("some_metric / some_alt_metric");
+
+        let table_provider = build_test_table_provider_with_tsid(
+            &[
+                (DEFAULT_SCHEMA_NAME.to_string(), "some_metric".to_string()),
+                (
+                    DEFAULT_SCHEMA_NAME.to_string(),
+                    "some_alt_metric".to_string(),
+                ),
+            ],
+            1,
+            1,
+        )
+        .await;
+        let plan =
+            PromPlanner::stmt_to_plan(table_provider, &eval_stmt, &build_query_engine_state())
+                .await
+                .unwrap();
+
+        let plan_str = plan.display_indent_schema().to_string();
+        assert!(
+            plan_str.contains("some_metric.__tsid = some_alt_metric.__tsid"),
+            "{plan_str}"
+        );
+        assert!(
+            !plan_str.contains("some_metric.tag_0 = some_alt_metric.tag_0"),
+            "{plan_str}"
+        );
+    }
+
+    #[tokio::test]
+    async fn tsid_is_preserved_for_nested_default_binary_joins() {
+        let eval_stmt = build_eval_stmt("(some_metric - some_alt_metric) / some_third_metric");
+
+        let table_provider = build_test_table_provider_with_tsid(
+            &[
+                (DEFAULT_SCHEMA_NAME.to_string(), "some_metric".to_string()),
+                (
+                    DEFAULT_SCHEMA_NAME.to_string(),
+                    "some_alt_metric".to_string(),
+                ),
+                (
+                    DEFAULT_SCHEMA_NAME.to_string(),
+                    "some_third_metric".to_string(),
+                ),
+            ],
+            1,
+            1,
+        )
+        .await;
+        let plan =
+            PromPlanner::stmt_to_plan(table_provider, &eval_stmt, &build_query_engine_state())
+                .await
+                .unwrap();
+
+        let plan_str = plan.display_indent_schema().to_string();
+        assert_eq!(plan_str.matches("__tsid =").count(), 2, "{plan_str}");
+        assert!(!plan_str.contains("tag_0 ="), "{plan_str}");
+    }
+
+    #[tokio::test]
+    async fn repeated_tsid_binary_operand_keeps_tsid_join_keys() {
+        let eval_stmt = build_eval_stmt("((some_metric - some_alt_metric) / some_metric) * 100");
+
+        let table_provider = build_test_table_provider_with_tsid(
+            &[
+                (DEFAULT_SCHEMA_NAME.to_string(), "some_metric".to_string()),
+                (
+                    DEFAULT_SCHEMA_NAME.to_string(),
+                    "some_alt_metric".to_string(),
+                ),
+            ],
+            1,
+            1,
+        )
+        .await;
+        let plan =
+            PromPlanner::stmt_to_plan(table_provider, &eval_stmt, &build_query_engine_state())
+                .await
+                .unwrap();
+
+        let plan_str = plan.display_indent_schema().to_string();
+        assert_eq!(plan_str.matches("__tsid =").count(), 2, "{plan_str}");
+        assert!(!plan_str.contains("tag_0 ="), "{plan_str}");
+    }
+
+    #[tokio::test]
+    async fn repeated_tsid_binary_operand_keeps_shorter_field_side() {
+        let eval_stmt =
+            build_eval_stmt("((two_field_metric - one_field_metric) / one_field_metric) * 100");
+
+        let table_provider = build_test_table_provider_with_tsid_fields(
+            &[
+                (
+                    (
+                        DEFAULT_SCHEMA_NAME.to_string(),
+                        "two_field_metric".to_string(),
+                    ),
+                    2,
+                ),
+                (
+                    (
+                        DEFAULT_SCHEMA_NAME.to_string(),
+                        "one_field_metric".to_string(),
+                    ),
+                    1,
+                ),
+            ],
+            1,
+        )
+        .await;
+        let plan =
+            PromPlanner::stmt_to_plan(table_provider, &eval_stmt, &build_query_engine_state())
+                .await
+                .unwrap();
+
+        let field_names = plan
+            .schema()
+            .fields()
+            .iter()
+            .map(|field| field.name().clone())
+            .collect::<Vec<_>>();
+        let value_columns = field_names
+            .iter()
+            .filter(|name| {
+                *name != "tag_0" && *name != "timestamp" && *name != DATA_SCHEMA_TSID_COLUMN_NAME
+            })
+            .count();
+        assert_eq!(value_columns, 1, "{field_names:?}");
+        let plan_str = plan.display_indent_schema().to_string();
+        assert_eq!(plan_str.matches("__tsid =").count(), 2, "{plan_str}");
+        assert!(!plan_str.contains("tag_0 ="), "{plan_str}");
+    }
+
+    #[tokio::test]
+    async fn tsid_binary_join_uses_shorter_field_side() {
+        let eval_stmt = build_eval_stmt("one_field_metric / two_field_metric");
+
+        let table_provider = build_test_table_provider_with_tsid_fields(
+            &[
+                (
+                    (
+                        DEFAULT_SCHEMA_NAME.to_string(),
+                        "one_field_metric".to_string(),
+                    ),
+                    1,
+                ),
+                (
+                    (
+                        DEFAULT_SCHEMA_NAME.to_string(),
+                        "two_field_metric".to_string(),
+                    ),
+                    2,
+                ),
+            ],
+            1,
+        )
+        .await;
+        let plan =
+            PromPlanner::stmt_to_plan(table_provider, &eval_stmt, &build_query_engine_state())
+                .await
+                .unwrap();
+
+        let field_names = plan
+            .schema()
+            .fields()
+            .iter()
+            .map(|field| field.name().clone())
+            .collect::<Vec<_>>();
+        let value_columns = field_names
+            .iter()
+            .filter(|name| {
+                *name != "tag_0" && *name != "timestamp" && *name != DATA_SCHEMA_TSID_COLUMN_NAME
+            })
+            .count();
+        assert_eq!(value_columns, 1, "{field_names:?}");
+    }
+
+    #[tokio::test]
+    async fn label_matching_modifier_disables_tsid_binary_join() {
+        let eval_stmt = build_eval_stmt("some_metric / ignoring(tag_0) some_alt_metric");
+
+        let table_provider = build_test_table_provider_with_tsid(
+            &[
+                (DEFAULT_SCHEMA_NAME.to_string(), "some_metric".to_string()),
+                (
+                    DEFAULT_SCHEMA_NAME.to_string(),
+                    "some_alt_metric".to_string(),
+                ),
+            ],
+            2,
+            1,
+        )
+        .await;
+        let plan =
+            PromPlanner::stmt_to_plan(table_provider, &eval_stmt, &build_query_engine_state())
+                .await
+                .unwrap();
+
+        let plan_str = plan.display_indent_schema().to_string();
+        assert!(!plan_str.contains("__tsid ="), "{plan_str}");
+        assert!(
+            plan_str.contains("some_metric.tag_1 = some_alt_metric.tag_1"),
+            "{plan_str}"
+        );
+    }
+
+    #[tokio::test]
+    async fn comparison_binary_join_uses_tsid_and_keeps_it_in_filtered_result() {
+        let eval_stmt = build_eval_stmt("some_metric > some_alt_metric");
+
+        let table_provider = build_test_table_provider_with_tsid(
+            &[
+                (DEFAULT_SCHEMA_NAME.to_string(), "some_metric".to_string()),
+                (
+                    DEFAULT_SCHEMA_NAME.to_string(),
+                    "some_alt_metric".to_string(),
+                ),
+            ],
+            2,
+            1,
+        )
+        .await;
+        let mut planner = PromPlanner {
+            table_provider,
+            ctx: PromPlannerContext::from_eval_stmt(&eval_stmt),
+        };
+        let plan = planner
+            .prom_expr_to_plan(&eval_stmt.expr, &build_query_engine_state())
+            .await
+            .unwrap();
+
+        let plan_str = plan.display_indent_schema().to_string();
+        assert!(
+            plan_str.contains("some_metric.__tsid = some_alt_metric.__tsid"),
+            "{plan_str}"
+        );
+        assert!(
+            plan.schema()
+                .fields()
+                .iter()
+                .any(|field| field.name() == DATA_SCHEMA_TSID_COLUMN_NAME),
+            "{plan_str}"
+        );
+        assert!(planner.ctx.use_tsid, "{plan_str}");
+    }
+
+    #[tokio::test]
+    async fn comparison_bool_binary_join_uses_tsid_when_available() {
+        let eval_stmt = build_eval_stmt("some_metric > bool some_alt_metric");
+
+        let table_provider = build_test_table_provider_with_tsid(
+            &[
+                (DEFAULT_SCHEMA_NAME.to_string(), "some_metric".to_string()),
+                (
+                    DEFAULT_SCHEMA_NAME.to_string(),
+                    "some_alt_metric".to_string(),
+                ),
+            ],
+            2,
+            1,
+        )
+        .await;
+        let plan =
+            PromPlanner::stmt_to_plan(table_provider, &eval_stmt, &build_query_engine_state())
+                .await
+                .unwrap();
+
+        let plan_str = plan.display_indent_schema().to_string();
+        assert!(
+            plan_str.contains("some_metric.__tsid = some_alt_metric.__tsid"),
+            "{plan_str}"
+        );
+        assert!(!plan_str.contains("tag_0 ="), "{plan_str}");
+        assert!(!plan_str.contains("tag_1 ="), "{plan_str}");
+    }
+
+    #[tokio::test]
+    async fn scalar_count_count_range_keeps_full_window() {
+        let plan_str = build_optimized_tsid_plan(
+            "scalar(count(count(some_metric) by (tag_0)))",
+            1,
+            1,
+            100_000,
+            1,
+        )
+        .await;
+        assert!(plan_str.contains("ScalarCalculate: tags=[]"));
+        assert!(plan_str.contains("PromInstantManipulate: range=[0..100000000]"));
+        assert!(!plan_str.contains("PromInstantManipulate: range=[99999000..99999000]"));
+    }
+
+    #[tokio::test]
+    async fn scalar_count_count_rewrite_applies_inside_binary_expr_for_tsid_input() {
+        let plan_str = build_optimized_tsid_plan(
+            "sum(irate(some_metric[1h])) / scalar(count(count(some_metric) by (tag_0)))",
+            2,
+            1,
+            10,
+            300,
+        )
+        .await;
+        assert!(plan_str.contains("Distinct:"), "{plan_str}");
+    }
+
+    #[tokio::test]
+    async fn nested_count_rewrite_keeps_full_series_key_with_tsid_input() {
+        assert_nested_count_rewrite_applies(
+            "count(count(some_metric) by (tag_0))",
+            "Aggregate: groupBy=[[some_metric.timestamp]], aggr=[[count(Int64(1)) AS count(count(some_metric.field_0))]]"
+        )
+        .await;
+    }
+
+    #[tokio::test]
+    async fn nested_sum_count_rewrite_keeps_full_series_key_with_tsid_input() {
+        assert_nested_count_rewrite_applies(
+            "count(sum(some_metric) by (tag_0))",
+            "Aggregate: groupBy=[[some_metric.timestamp]], aggr=[[count(Int64(1)) AS count(sum(some_metric.field_0))]]"
+        )
+        .await;
+    }
+
+    #[tokio::test]
+    async fn nested_supported_inner_aggs_rewrite_apply_for_tsid_input() {
+        for (query, expected_outer_agg) in [
+            (
+                "count(avg(some_metric) by (tag_0))",
+                "Aggregate: groupBy=[[some_metric.timestamp]], aggr=[[count(Int64(1)) AS count(avg(some_metric.field_0))]]",
+            ),
+            (
+                "count(min(some_metric) by (tag_0))",
+                "Aggregate: groupBy=[[some_metric.timestamp]], aggr=[[count(Int64(1)) AS count(min(some_metric.field_0))]]",
+            ),
+            (
+                "count(max(some_metric) by (tag_0))",
+                "Aggregate: groupBy=[[some_metric.timestamp]], aggr=[[count(Int64(1)) AS count(max(some_metric.field_0))]]",
+            ),
+            (
+                "count(stddev(some_metric) by (tag_0))",
+                "Aggregate: groupBy=[[some_metric.timestamp]], aggr=[[count(Int64(1)) AS count(stddev_pop(some_metric.field_0))]]",
+            ),
+            (
+                "count(stdvar(some_metric) by (tag_0))",
+                "Aggregate: groupBy=[[some_metric.timestamp]], aggr=[[count(Int64(1)) AS count(var_pop(some_metric.field_0))]]",
+            ),
+        ] {
+            assert_nested_count_rewrite_applies(query, expected_outer_agg).await;
+        }
+    }
+
+    #[tokio::test]
+    async fn nested_non_count_inner_aggs_rewrite_filter_null_values_for_tsid_input() {
+        let count_plan =
+            build_optimized_tsid_plan("count(count(some_metric) by (tag_0))", 2, 1, 100_000, 1)
+                .await;
+        assert!(
+            !count_plan.contains("some_metric.field_0 IS NOT NULL"),
+            "{count_plan}"
+        );
+
+        for query in [
+            "count(sum(some_metric) by (tag_0))",
+            "count(avg(some_metric) by (tag_0))",
+            "count(min(some_metric) by (tag_0))",
+            "count(max(some_metric) by (tag_0))",
+            "count(stddev(some_metric) by (tag_0))",
+            "count(stdvar(some_metric) by (tag_0))",
+        ] {
+            let plan_str = build_optimized_tsid_plan(query, 2, 1, 100_000, 1).await;
+            assert!(
+                plan_str.contains("Filter: some_metric.field_0 IS NOT NULL"),
+                "{query}: {plan_str}"
+            );
+        }
+    }
+
+    #[tokio::test]
+    async fn nested_unsupported_or_non_direct_inner_aggs_do_not_rewrite() {
+        assert_nested_count_rewrite_missing("count(group(some_metric) by (tag_0))", 2, 1).await;
+        assert_nested_count_rewrite_missing(
+            "count(sum(irate(some_metric[1h])) by (tag_0))",
+            2,
+            300,
+        )
+        .await;
+    }
+
     #[tokio::test]
     async fn physical_table_name_is_not_leaked_in_plan() {
         let prom_expr = parser::parse("some_metric").unwrap();
@@ -5194,7 +5755,7 @@ mod test {
                 .unwrap();
 
         let expected = String::from(
-            "Projection: rhs.tag_0, rhs.timestamp, lhs.field_0 + rhs.field_0 AS lhs.field_0 + rhs.field_0 [tag_0:Utf8, timestamp:Timestamp(ms), lhs.field_0 + rhs.field_0:Float64;N]\
+            "Projection: rhs.tag_0, rhs.timestamp, CAST(lhs.field_0 AS Float64) + CAST(rhs.field_0 AS Float64) AS lhs.field_0 + rhs.field_0 [tag_0:Utf8, timestamp:Timestamp(ms), lhs.field_0 + rhs.field_0:Float64;N]\
             \n  Inner Join: lhs.tag_0 = rhs.tag_0, lhs.timestamp = rhs.timestamp [tag_0:Utf8, timestamp:Timestamp(ms), field_0:Float64;N, tag_0:Utf8, timestamp:Timestamp(ms), field_0:Float64;N]\
             \n    SubqueryAlias: lhs [tag_0:Utf8, timestamp:Timestamp(ms), field_0:Float64;N]\
             \n      PromInstantManipulate: range=[0..100000000], lookback=[1000], interval=[5000], time index=[timestamp] [tag_0:Utf8, timestamp:Timestamp(ms), field_0:Float64;N]\
@@ -5249,7 +5810,7 @@ mod test {
     async fn binary_op_literal_column() {
         let query = r#"1 + some_metric{tag_0="bar"}"#;
         let expected = String::from(
-            "Projection: some_metric.tag_0, some_metric.timestamp, Float64(1) + some_metric.field_0 AS Float64(1) + field_0 [tag_0:Utf8, timestamp:Timestamp(ms), Float64(1) + field_0:Float64;N]\
+            "Projection: some_metric.tag_0, some_metric.timestamp, Float64(1) + CAST(some_metric.field_0 AS Float64) AS Float64(1) + field_0 [tag_0:Utf8, timestamp:Timestamp(ms), Float64(1) + field_0:Float64;N]\
             \n  PromInstantManipulate: range=[0..100000000], lookback=[1000], interval=[5000], time index=[timestamp] [tag_0:Utf8, timestamp:Timestamp(ms), field_0:Float64;N]\
             \n    PromSeriesDivide: tags=[\"tag_0\"] [tag_0:Utf8, timestamp:Timestamp(ms), field_0:Float64;N]\
             \n      Sort: some_metric.tag_0 ASC NULLS FIRST, some_metric.timestamp ASC NULLS FIRST [tag_0:Utf8, timestamp:Timestamp(ms), field_0:Float64;N]\
@@ -5287,7 +5848,7 @@ mod test {
     async fn bool_with_additional_arithmetic() {
         let query = "some_metric + (1 == bool 2)";
         let expected = String::from(
-            "Projection: some_metric.tag_0, some_metric.timestamp, some_metric.field_0 + CAST(Float64(1) = Float64(2) AS Float64) AS field_0 + Float64(1) = Float64(2) [tag_0:Utf8, timestamp:Timestamp(ms), field_0 + Float64(1) = Float64(2):Float64;N]\
+            "Projection: some_metric.tag_0, some_metric.timestamp, CAST(some_metric.field_0 AS Float64) + CAST(Float64(1) = Float64(2) AS Float64) AS field_0 + Float64(1) = Float64(2) [tag_0:Utf8, timestamp:Timestamp(ms), field_0 + Float64(1) = Float64(2):Float64;N]\
             \n  PromInstantManipulate: range=[0..100000000], lookback=[1000], interval=[5000], time index=[timestamp] [tag_0:Utf8, timestamp:Timestamp(ms), field_0:Float64;N]\
             \n    PromSeriesDivide: tags=[\"tag_0\"] [tag_0:Utf8, timestamp:Timestamp(ms), field_0:Float64;N]\
             \n      Sort: some_metric.tag_0 ASC NULLS FIRST, some_metric.timestamp ASC NULLS FIRST [tag_0:Utf8, timestamp:Timestamp(ms), field_0:Float64;N]\
@@ -5397,7 +5958,7 @@ mod test {
             PromPlanner::stmt_to_plan(table_provider, &eval_stmt, &build_query_engine_state())
                 .await
                 .unwrap();
-        let expected = "Projection: http_server_requests_seconds_count.uri, http_server_requests_seconds_count.kubernetes_namespace, http_server_requests_seconds_count.kubernetes_pod_name, http_server_requests_seconds_count.greptime_timestamp, http_server_requests_seconds_sum.greptime_value / http_server_requests_seconds_count.greptime_value AS http_server_requests_seconds_sum.greptime_value / http_server_requests_seconds_count.greptime_value\
+        let expected = "Projection: http_server_requests_seconds_count.uri, http_server_requests_seconds_count.kubernetes_namespace, http_server_requests_seconds_count.kubernetes_pod_name, http_server_requests_seconds_count.greptime_timestamp, CAST(http_server_requests_seconds_sum.greptime_value AS Float64) / CAST(http_server_requests_seconds_count.greptime_value AS Float64) AS http_server_requests_seconds_sum.greptime_value / http_server_requests_seconds_count.greptime_value\
             \n  Inner Join: http_server_requests_seconds_sum.greptime_timestamp = http_server_requests_seconds_count.greptime_timestamp, http_server_requests_seconds_sum.uri = http_server_requests_seconds_count.uri\
             \n    SubqueryAlias: http_server_requests_seconds_sum\
             \n      PromInstantManipulate: range=[0..100000000], lookback=[1000], interval=[5000], time index=[greptime_timestamp]\
@@ -5788,7 +6349,7 @@ mod test {
 
         let query = "some_alt_metric{__schema__=\"greptime_private\"} / some_metric";
         let expected = String::from(
-            "Projection: some_metric.tag_0, some_metric.timestamp, greptime_private.some_alt_metric.field_0 / some_metric.field_0 AS greptime_private.some_alt_metric.field_0 / some_metric.field_0 [tag_0:Utf8, timestamp:Timestamp(ms), greptime_private.some_alt_metric.field_0 / some_metric.field_0:Float64;N]\
+            "Projection: some_metric.tag_0, some_metric.timestamp, CAST(greptime_private.some_alt_metric.field_0 AS Float64) / CAST(some_metric.field_0 AS Float64) AS greptime_private.some_alt_metric.field_0 / some_metric.field_0 [tag_0:Utf8, timestamp:Timestamp(ms), greptime_private.some_alt_metric.field_0 / some_metric.field_0:Float64;N]\
             \n  Inner Join: greptime_private.some_alt_metric.tag_0 = some_metric.tag_0, greptime_private.some_alt_metric.timestamp = some_metric.timestamp [tag_0:Utf8, timestamp:Timestamp(ms), field_0:Float64;N, tag_0:Utf8, timestamp:Timestamp(ms), field_0:Float64;N]\
             \n    SubqueryAlias: greptime_private.some_alt_metric [tag_0:Utf8, timestamp:Timestamp(ms), field_0:Float64;N]\
             \n      PromInstantManipulate: range=[0..100000000], lookback=[1000], interval=[5000], time index=[timestamp] [tag_0:Utf8, timestamp:Timestamp(ms), field_0:Float64;N]\
diff --git a/src/query/src/query_engine.rs b/src/query/src/query_engine.rs
index c6199e872f..1415998486 100644
--- a/src/query/src/query_engine.rs
+++ b/src/query/src/query_engine.rs
@@ -31,7 +31,6 @@ use common_query::Output;
 use datafusion::catalog::TableFunction;
 use datafusion::dataframe::DataFrame;
 use datafusion_expr::{AggregateUDF, LogicalPlan, WindowUDF};
-use datatypes::schema::Schema;
 pub use default_serializer::{DefaultPlanDecoder, DefaultSerializer};
 use partition::manager::PartitionRuleManagerRef;
 use session::context::QueryContextRef;
@@ -48,8 +47,6 @@ use crate::region_query::RegionQueryHandlerRef;
 /// Describe statement result
 #[derive(Debug)]
 pub struct DescribeResult {
-    /// The schema of statement
-    pub schema: Schema,
     /// The logical plan for statement
     pub logical_plan: LogicalPlan,
 }
diff --git a/src/query/src/query_engine/state.rs b/src/query/src/query_engine/state.rs
index a45fc4c896..45a5700781 100644
--- a/src/query/src/query_engine/state.rs
+++ b/src/query/src/query_engine/state.rs
@@ -59,8 +59,11 @@ use crate::dist_plan::{
 };
 use crate::metrics::{QUERY_MEMORY_POOL_REJECTED_TOTAL, QUERY_MEMORY_POOL_USAGE_BYTES};
 use crate::optimizer::ExtensionAnalyzerRule;
+use crate::optimizer::const_normalization::ConstNormalizationRule;
 use crate::optimizer::constant_term::MatchesConstantTermOptimizer;
+use crate::optimizer::count_nest_aggr::CountNestAggrRule;
 use crate::optimizer::count_wildcard::CountWildcardToTimeIndexRule;
+use crate::optimizer::json_type_concretize::JsonTypeConcretizeRule;
 use crate::optimizer::parallelize_scan::ParallelizeScan;
 use crate::optimizer::pass_distribution::PassDistribution;
 use crate::optimizer::remove_duplicate::RemoveDuplicate;
@@ -146,6 +149,7 @@ impl QueryEngineState {
 
         // The [`TypeConversionRule`] must be at first
         extension_rules.insert(0, Arc::new(TypeConversionRule) as _);
+        extension_rules.push(Arc::new(CountNestAggrRule) as _);
 
         // Apply the datafusion rules
         let mut analyzer = Analyzer::new();
@@ -154,6 +158,7 @@ impl QueryEngineState {
         analyzer
             .rules
             .insert(0, Arc::new(CountWildcardToTimeIndexRule));
+        analyzer.rules.push(Arc::new(ConstNormalizationRule));
 
         // Add ApplyFunctionRewrites rule,
         // Note we cannot use `analyzer.add_function_rewrite`
@@ -172,6 +177,7 @@ impl QueryEngineState {
 
         let mut optimizer = Optimizer::new();
         optimizer.rules.push(Arc::new(ScanHintRule));
+        optimizer.rules.push(Arc::new(JsonTypeConcretizeRule));
 
         // add physical optimizer
         let mut physical_optimizer = PhysicalOptimizer::new();
diff --git a/src/query/src/range_select/plan.rs b/src/query/src/range_select/plan.rs
index e863aaced0..d83514f435 100644
--- a/src/query/src/range_select/plan.rs
+++ b/src/query/src/range_select/plan.rs
@@ -836,6 +836,7 @@ impl ExecutionPlan for RangeSelectExec {
         context: Arc<TaskContext>,
     ) -> DfResult<DfSendableRecordBatchStream> {
         let baseline_metric = BaselineMetrics::new(&self.metric, partition);
+        let batch_size = context.session_config().batch_size();
         let input = self.input.execute(partition, context)?;
         let schema = input.schema();
         let time_index = schema
@@ -852,6 +853,7 @@ impl ExecutionPlan for RangeSelectExec {
                 .collect(),
         )?;
         Ok(Box::pin(RangeSelectStream {
+            batch_size,
             schema: self.schema.clone(),
             range_exec: self.range_exec.clone(),
             input,
@@ -868,6 +870,8 @@ impl ExecutionPlan for RangeSelectExec {
             metric: baseline_metric,
             schema_project: self.schema_project.clone(),
             schema_before_project: self.schema_before_project.clone(),
+            output_batch: None,
+            output_batch_offset: 0,
         }))
     }
 
@@ -881,6 +885,7 @@ impl ExecutionPlan for RangeSelectExec {
 }
 
 struct RangeSelectStream {
+    batch_size: usize,
     /// the schema of output column
     schema: SchemaRef,
     range_exec: Vec<RangeFnExec>,
@@ -907,6 +912,8 @@ struct RangeSelectStream {
     metric: BaselineMetrics,
     schema_project: Option<Vec<usize>>,
     schema_before_project: SchemaRef,
+    output_batch: Option<RecordBatch>,
+    output_batch_offset: usize,
 }
 
 #[derive(Debug)]
@@ -1149,6 +1156,36 @@ impl RangeSelectStream {
         };
         Ok(project_output)
     }
+
+    fn next_output_batch(&mut self) -> DfResult<Option<RecordBatch>> {
+        if self.output_batch.is_none() {
+            self.output_batch = Some(self.generate_output()?);
+            self.output_batch_offset = 0;
+        }
+
+        let num_rows = self.output_batch.as_ref().unwrap().num_rows();
+        if num_rows == 0 {
+            self.output_batch = None;
+            self.output_batch_offset = 0;
+            return Ok(None);
+        }
+
+        if self.output_batch_offset == 0 && num_rows <= self.batch_size {
+            return Ok(self.output_batch.take());
+        }
+
+        let offset = self.output_batch_offset;
+        let len = (num_rows - offset).min(self.batch_size);
+        let batch = self.output_batch.as_ref().unwrap().slice(offset, len);
+        self.output_batch_offset += len;
+
+        if self.output_batch_offset >= num_rows {
+            self.output_batch = None;
+            self.output_batch_offset = 0;
+        }
+
+        Ok(Some(batch))
+    }
 }
 
 enum ExecutionState {
@@ -1191,13 +1228,19 @@ impl Stream for RangeSelectStream {
                     }
                 }
                 ExecutionState::ProducingOutput => {
-                    let result = self.generate_output();
+                    let result = self.next_output_batch();
                     return match result {
                         // made output
-                        Ok(batch) => {
-                            self.exec_state = ExecutionState::Done;
+                        Ok(Some(batch)) => {
+                            if self.output_batch.is_none() {
+                                self.exec_state = ExecutionState::Done;
+                            }
                             Poll::Ready(Some(Ok(batch)))
                         }
+                        Ok(None) => {
+                            self.exec_state = ExecutionState::Done;
+                            Poll::Ready(None)
+                        }
                         // error making output
                         Err(error) => Poll::Ready(Some(Err(error))),
                     };
@@ -1251,7 +1294,7 @@ mod test {
     use datafusion::prelude::SessionContext;
     use datafusion_physical_expr::PhysicalSortExpr;
     use datafusion_physical_expr::expressions::Column;
-    use datatypes::arrow::array::TimestampMillisecondArray;
+    use datatypes::arrow::array::{Float64Array, Int64Array, TimestampMillisecondArray};
     use datatypes::arrow_array::StringArray;
 
     use super::*;
@@ -1313,15 +1356,49 @@ mod test {
         ))
     }
 
-    async fn do_range_select_test(
+    fn prepare_empty_test_data(is_float: bool) -> DataSourceExec {
+        let schema = Arc::new(Schema::new(vec![
+            Field::new(TIME_INDEX_COLUMN, TimestampMillisecondType::DATA_TYPE, true),
+            Field::new(
+                "value",
+                if is_float {
+                    DataType::Float64
+                } else {
+                    DataType::Int64
+                },
+                true,
+            ),
+            Field::new("host", DataType::Utf8, true),
+        ]));
+        let timestamp_column: Arc<dyn Array> =
+            Arc::new(TimestampMillisecondArray::from(Vec::<i64>::new())) as _;
+        let value_column: Arc<dyn Array> = if is_float {
+            Arc::new(Float64Array::from(Vec::<Option<f64>>::new())) as _
+        } else {
+            Arc::new(Int64Array::from(Vec::<Option<i64>>::new())) as _
+        };
+        let host_column: Arc<dyn Array> =
+            Arc::new(StringArray::from(Vec::<Option<&str>>::new())) as _;
+        let data = RecordBatch::try_new(
+            schema.clone(),
+            vec![timestamp_column, value_column, host_column],
+        )
+        .unwrap();
+
+        DataSourceExec::new(Arc::new(
+            MemorySourceConfig::try_new(&[vec![data]], schema, None).unwrap(),
+        ))
+    }
+
+    async fn collect_range_select_test(
         range1: Millisecond,
         range2: Millisecond,
         align: Millisecond,
         fill: Option<Fill>,
         is_float: bool,
         is_gap: bool,
-        expected: String,
-    ) {
+        batch_size: usize,
+    ) -> Vec<RecordBatch> {
         let data_type = if is_float {
             DataType::Float64
         } else {
@@ -1412,11 +1489,25 @@ mod test {
             .into(),
             range_select_exec,
         );
-        let session_context = SessionContext::default();
+        let session_context = SessionContext::new_with_config(
+            datafusion::execution::config::SessionConfig::new().with_batch_size(batch_size),
+        );
+        datafusion::physical_plan::collect(Arc::new(sort_exec), session_context.task_ctx())
+            .await
+            .unwrap()
+    }
+
+    async fn do_range_select_test(
+        range1: Millisecond,
+        range2: Millisecond,
+        align: Millisecond,
+        fill: Option<Fill>,
+        is_float: bool,
+        is_gap: bool,
+        expected: String,
+    ) {
         let result =
-            datafusion::physical_plan::collect(Arc::new(sort_exec), session_context.task_ctx())
-                .await
-                .unwrap();
+            collect_range_select_test(range1, range2, align, fill, is_float, is_gap, 8192).await;
 
         let result_literal = arrow::util::pretty::pretty_format_batches(&result)
             .unwrap()
@@ -1700,6 +1791,88 @@ mod test {
         .await;
     }
 
+    #[tokio::test]
+    async fn range_select_respects_session_batch_size() {
+        let result =
+            collect_range_select_test(10_000, 5_000, 5_000, Some(Fill::Null), true, false, 3).await;
+
+        let row_counts = result
+            .iter()
+            .map(|batch| batch.num_rows())
+            .collect::<Vec<_>>();
+        assert_eq!(vec![3, 3, 3, 3], row_counts);
+    }
+
+    #[tokio::test]
+    async fn range_select_skips_empty_output_batch() {
+        let memory_exec = Arc::new(prepare_empty_test_data(true));
+        let schema = Arc::new(Schema::new(vec![
+            Field::new("MIN(value)", DataType::Float64, true),
+            Field::new("MAX(value)", DataType::Float64, true),
+            Field::new(TIME_INDEX_COLUMN, TimestampMillisecondType::DATA_TYPE, true),
+            Field::new("host", DataType::Utf8, true),
+        ]));
+        let cache = Arc::new(PlanProperties::new(
+            EquivalenceProperties::new(schema.clone()),
+            Partitioning::UnknownPartitioning(1),
+            EmissionType::Incremental,
+            Boundedness::Bounded,
+        ));
+        let input_schema = memory_exec.schema().clone();
+        let range_select_exec = Arc::new(RangeSelectExec {
+            input: memory_exec,
+            range_exec: vec![
+                RangeFnExec {
+                    expr: Arc::new(
+                        AggregateExprBuilder::new(
+                            min_max::min_udaf(),
+                            vec![Arc::new(Column::new("value", 1))],
+                        )
+                        .schema(input_schema.clone())
+                        .alias("MIN(value)")
+                        .build()
+                        .unwrap(),
+                    ),
+                    range: 10_000,
+                    fill: Some(Fill::Null),
+                    need_cast: None,
+                },
+                RangeFnExec {
+                    expr: Arc::new(
+                        AggregateExprBuilder::new(
+                            min_max::max_udaf(),
+                            vec![Arc::new(Column::new("value", 1))],
+                        )
+                        .schema(input_schema)
+                        .alias("MAX(value)")
+                        .build()
+                        .unwrap(),
+                    ),
+                    range: 5_000,
+                    fill: Some(Fill::Null),
+                    need_cast: None,
+                },
+            ],
+            align: 5_000,
+            align_to: 0,
+            by: vec![Arc::new(Column::new("host", 2))],
+            time_index: TIME_INDEX_COLUMN.to_string(),
+            schema: schema.clone(),
+            schema_before_project: schema.clone(),
+            schema_project: None,
+            by_schema: Arc::new(Schema::new(vec![Field::new("host", DataType::Utf8, true)])),
+            metric: ExecutionPlanMetricsSet::new(),
+            cache,
+        });
+        let session_context = SessionContext::new();
+        let result =
+            datafusion::physical_plan::collect(range_select_exec, session_context.task_ctx())
+                .await
+                .unwrap();
+
+        assert!(result.is_empty());
+    }
+
     #[test]
     fn fill_test() {
         assert!(Fill::try_from_str("", &DataType::UInt8).unwrap().is_none());
diff --git a/src/query/src/range_select/plan_rewrite.rs b/src/query/src/range_select/plan_rewrite.rs
index 7cf8ef98b5..4d795565f0 100644
--- a/src/query/src/range_select/plan_rewrite.rs
+++ b/src/query/src/range_select/plan_rewrite.rs
@@ -35,6 +35,7 @@ use datafusion_expr::{
 };
 use datafusion_optimizer::simplify_expressions::ExprSimplifier;
 use datatypes::prelude::ConcreteDataType;
+use datatypes::schema::TIME_INDEX_KEY;
 use promql_parser::util::parse_duration;
 use session::context::QueryContextRef;
 use snafu::{OptionExt, ResultExt, ensure};
@@ -376,7 +377,7 @@ impl RangePlanRewriter {
                     }
                     .fail();
                 };
-                let (time_index, default_by) = self.get_index_by(input.schema()).await?;
+                let query_ctx = self.query_ctx.clone();
                 let mut range_rewriter = RangeExprRewriter {
                     input_plan: &input,
                     align: Duration::default(),
@@ -384,13 +385,16 @@ impl RangePlanRewriter {
                     by: vec![],
                     range_fn: BTreeSet::new(),
                     sub_aggr: aggr_plan,
-                    query_ctx: &self.query_ctx,
+                    query_ctx: &query_ctx,
                 };
                 let new_expr = expr
                     .iter()
                     .map(|expr| expr.clone().rewrite(&mut range_rewriter).map(|x| x.data))
                     .collect::<DFResult<Vec<_>>>()?;
-                if range_rewriter.by.is_empty() {
+                let need_default_by = range_rewriter.by.is_empty();
+                let (time_index, default_by) =
+                    self.get_index_by(input.schema(), need_default_by).await?;
+                if need_default_by {
                     range_rewriter.by = default_by;
                 }
                 let range_select = RangeSelect::try_new(
@@ -481,25 +485,58 @@ impl RangePlanRewriter {
         }
     }
 
-    /// this function use to find the time_index column and row columns from input schema,
-    /// return `(time_index, [row_columns])` to the rewriter.
-    /// If the user does not explicitly use the `by` keyword to indicate time series,
-    /// `[row_columns]` will be use as default time series
-    async fn get_index_by(&mut self, schema: &Arc<DFSchema>) -> Result<(Expr, Vec<Expr>)> {
+    /// Finds the time index column and default row-key grouping from the input schema.
+    ///
+    /// Returns `(time_index, [row_columns])` to the rewriter. If the user omits `BY`,
+    /// `[row_columns]` is used as the default time-series grouping.
+    ///
+    /// For derived inputs such as subqueries, joins, or set operations, the source table
+    /// qualifier may no longer resolve back to a table provider. In that case we can still
+    /// recover the time index from column metadata, but we cannot safely reconstruct the
+    /// original row-key columns, so omitted `BY` must be rejected by the caller.
+    async fn get_index_by(
+        &mut self,
+        schema: &Arc<DFSchema>,
+        need_default_by: bool,
+    ) -> Result<(Expr, Vec<Expr>)> {
         #[allow(deprecated)]
         let mut time_index_expr = Expr::Wildcard {
             qualifier: None,
             options: Box::new(WildcardOptions::default()),
         };
         let mut default_by = vec![];
+        let metadata_time_index_expr = (0..schema.fields().len()).find_map(|i| {
+            let (qualifier, field) = schema.qualified_field(i);
+            if field.metadata().contains_key(TIME_INDEX_KEY)
+                && matches!(field.data_type(), DataType::Timestamp(_, _))
+            {
+                Some(Expr::Column(Column::new(
+                    qualifier.cloned(),
+                    field.name().clone(),
+                )))
+            } else {
+                None
+            }
+        });
         for i in 0..schema.fields().len() {
             let (qualifier, _) = schema.qualified_field(i);
             if let Some(table_ref) = qualifier {
-                let table = self
-                    .table_provider
-                    .resolve_table(table_ref.clone())
-                    .await
-                    .context(CatalogSnafu)?
+                let table_source = match self.table_provider.resolve_table(table_ref.clone()).await
+                {
+                    Ok(table_source) => table_source,
+                    Err(error) => {
+                        // `TableNotExist` here usually means the qualifier now refers to a derived
+                        // input instead of a base table. We can still salvage the time index from
+                        // field metadata, but only when such metadata is present.
+                        if matches!(&error, catalog::error::Error::TableNotExist { .. })
+                            && metadata_time_index_expr.is_some()
+                        {
+                            continue;
+                        }
+                        return Err(error).context(CatalogSnafu);
+                    }
+                };
+                let table = table_source
                     .as_any()
                     .downcast_ref::<DefaultTableSource>()
                     .context(UnknownTableSnafu)?
@@ -537,6 +574,22 @@ impl RangePlanRewriter {
             }
         }
         #[allow(deprecated)]
+        if matches!(time_index_expr, Expr::Wildcard { .. })
+            && let Some(expr) = metadata_time_index_expr
+        {
+            common_telemetry::debug!(
+                "Range query falling back to time-index metadata for derived input schema: {}",
+                schema
+            );
+            ensure!(
+                !need_default_by,
+                RangeQuerySnafu {
+                    msg: "Cannot infer default BY columns from derived range query input"
+                }
+            );
+            time_index_expr = expr;
+        }
+        #[allow(deprecated)]
         if matches!(time_index_expr, Expr::Wildcard { .. }) {
             TimeIndexNotFoundSnafu {
                 table: schema.to_string(),
@@ -614,6 +667,7 @@ mod test {
     use datatypes::schema::{ColumnSchema, Schema};
     use session::context::QueryContext;
     use table::metadata::{TableInfoBuilder, TableMetaBuilder};
+    use table::table::TableRef;
     use table::test_util::EmptyTable;
 
     use super::*;
@@ -622,7 +676,45 @@ mod test {
     use crate::{QueryEngineFactory, QueryEngineRef};
 
     async fn create_test_engine() -> QueryEngineRef {
-        let table_name = "test".to_string();
+        create_test_engine_with_tables(&["test"], false).await
+    }
+
+    async fn create_union_test_engine() -> QueryEngineRef {
+        create_test_engine_with_tables(&["test_0", "test_1"], true).await
+    }
+
+    async fn create_test_engine_with_tables(
+        table_names: &[&str],
+        with_extra_timestamp: bool,
+    ) -> QueryEngineRef {
+        let catalog_list = MemoryCatalogManager::with_default_setup();
+        for (i, table_name) in table_names.iter().enumerate() {
+            let table = create_test_table(table_name, with_extra_timestamp);
+            assert!(
+                catalog_list
+                    .register_table_sync(RegisterTableRequest {
+                        catalog: DEFAULT_CATALOG_NAME.to_string(),
+                        schema: DEFAULT_SCHEMA_NAME.to_string(),
+                        table_name: (*table_name).to_string(),
+                        table_id: 1024 + i as u32,
+                        table,
+                    })
+                    .is_ok()
+            );
+        }
+        QueryEngineFactory::new(
+            catalog_list,
+            None,
+            None,
+            None,
+            None,
+            false,
+            QueryOptions::default(),
+        )
+        .query_engine()
+    }
+
+    fn create_test_table(table_name: &str, with_extra_timestamp: bool) -> TableRef {
         let mut columns = vec![];
         for i in 0..5 {
             columns.push(ColumnSchema::new(
@@ -639,6 +731,13 @@ mod test {
             )
             .with_time_index(true),
         );
+        if with_extra_timestamp {
+            columns.push(ColumnSchema::new(
+                "timestamp_2".to_string(),
+                ConcreteDataType::timestamp_millisecond_datatype(),
+                true,
+            ));
+        }
         for i in 0..5 {
             columns.push(ColumnSchema::new(
                 format!("field_{i}"),
@@ -650,38 +749,20 @@ mod test {
         let table_meta = TableMetaBuilder::empty()
             .schema(schema)
             .primary_key_indices((0..5).collect())
-            .value_indices((6..11).collect())
+            .value_indices(if with_extra_timestamp {
+                (6..12).collect()
+            } else {
+                (6..11).collect()
+            })
             .next_column_id(1024)
             .build()
             .unwrap();
         let table_info = TableInfoBuilder::default()
-            .name(&table_name)
+            .name(table_name)
             .meta(table_meta)
             .build()
             .unwrap();
-        let table = EmptyTable::from_table_info(&table_info);
-        let catalog_list = MemoryCatalogManager::with_default_setup();
-        assert!(
-            catalog_list
-                .register_table_sync(RegisterTableRequest {
-                    catalog: DEFAULT_CATALOG_NAME.to_string(),
-                    schema: DEFAULT_SCHEMA_NAME.to_string(),
-                    table_name,
-                    table_id: 1024,
-                    table,
-                })
-                .is_ok()
-        );
-        QueryEngineFactory::new(
-            catalog_list,
-            None,
-            None,
-            None,
-            None,
-            false,
-            QueryOptions::default(),
-        )
-        .query_engine()
+        EmptyTable::from_table_info(&table_info)
     }
 
     async fn do_query(sql: &str) -> Result<LogicalPlan> {
@@ -690,6 +771,12 @@ mod test {
         engine.planner().plan(&stmt, QueryContext::arc()).await
     }
 
+    async fn do_union_query(sql: &str) -> Result<LogicalPlan> {
+        let stmt = QueryLanguageParser::parse_sql(sql, &QueryContext::arc()).unwrap();
+        let engine = create_union_test_engine().await;
+        engine.planner().plan(&stmt, QueryContext::arc()).await
+    }
+
     async fn query_plan_compare(sql: &str, expected: String) {
         let plan = do_query(sql).await.unwrap();
         assert_eq!(plan.display_indent_schema().to_string(), expected);
@@ -765,6 +852,69 @@ mod test {
         query_plan_compare(query, expected).await;
     }
 
+    #[tokio::test]
+    async fn range_from_union_query() {
+        let queries = [
+            r#"SELECT timestamp, tag_0, avg(field_0) RANGE '5m'
+            FROM (
+                SELECT timestamp, tag_0, field_0, timestamp_2 FROM test_0
+                UNION ALL
+                SELECT timestamp, tag_0, field_0, timestamp_2 FROM test_1
+            )
+            WHERE timestamp >= '1970-01-01 00:00:00'
+            ALIGN '1h' by (tag_0)"#,
+            r#"SELECT tmp.timestamp, tmp.tag_0, avg(tmp.field_0) RANGE '5m'
+            FROM (
+                SELECT timestamp, tag_0, field_0, timestamp_2 FROM test_0
+                UNION ALL
+                SELECT timestamp, tag_0, field_0, timestamp_2 FROM test_1
+            ) AS tmp
+            WHERE tmp.timestamp >= '1970-01-01 00:00:00'
+            ALIGN '1h' by (tmp.tag_0)"#,
+        ];
+
+        for query in queries {
+            let plan = do_union_query(query)
+                .await
+                .unwrap()
+                .display_indent_schema()
+                .to_string();
+
+            assert!(plan.contains("RangeSelect"));
+            assert!(plan.contains("Union"));
+            assert!(plan.contains("time_index=timestamp"));
+        }
+    }
+
+    #[tokio::test]
+    async fn range_from_derived_query_without_by_err() {
+        let queries = [
+            r#"SELECT timestamp, tag_0, avg(field_0) RANGE '5m'
+            FROM (
+                SELECT timestamp, tag_0, field_0, timestamp_2 FROM test_0
+                UNION ALL
+                SELECT timestamp, tag_0, field_0, timestamp_2 FROM test_1
+            )
+            WHERE timestamp >= '1970-01-01 00:00:00'
+            ALIGN '1h'"#,
+            r#"SELECT tmp.timestamp, tmp.tag_0, avg(tmp.field_0) RANGE '5m'
+            FROM (
+                SELECT timestamp, tag_0, field_0, timestamp_2 FROM test_0
+                UNION ALL
+                SELECT timestamp, tag_0, field_0, timestamp_2 FROM test_1
+            ) AS tmp
+            WHERE tmp.timestamp >= '1970-01-01 00:00:00'
+            ALIGN '1h'"#,
+        ];
+
+        for query in queries {
+            assert_eq!(
+                do_union_query(query).await.unwrap_err().to_string(),
+                "Range Query: Cannot infer default BY columns from derived range query input"
+            );
+        }
+    }
+
     #[tokio::test]
     async fn range_in_expr() {
         let query = r#"SELECT sin(avg(field_0 + field_1) RANGE '5m' + 1) FROM test ALIGN '1h' by (tag_0,tag_1);"#;
diff --git a/src/query/src/sql.rs b/src/query/src/sql.rs
index 49e26c92ca..d7d7399a5a 100644
--- a/src/query/src/sql.rs
+++ b/src/query/src/sql.rs
@@ -33,6 +33,7 @@ use common_datasource::lister::{Lister, Source};
 use common_datasource::object_store::build_backend;
 use common_datasource::util::find_dir_and_filename;
 use common_meta::SchemaOptions;
+use common_meta::ddl::create_flow::FlowType;
 use common_meta::key::flow::flow_info::FlowInfoValue;
 use common_query::Output;
 use common_query::prelude::greptime_timestamp;
@@ -78,6 +79,7 @@ const VIEWS_COLUMN: &str = "Views";
 const FLOWS_COLUMN: &str = "Flows";
 const FIELD_COLUMN: &str = "Field";
 const TABLE_TYPE_COLUMN: &str = "Table_type";
+
 const COLUMN_NAME_COLUMN: &str = "Column";
 const COLUMN_GREPTIME_TYPE_COLUMN: &str = "Greptime_type";
 const COLUMN_TYPE_COLUMN: &str = "Type";
@@ -215,10 +217,10 @@ pub async fn show_databases(
 fn replace_column_in_expr(expr: &mut sqlparser::ast::Expr, from_column: &str, to_column: &str) {
     let _ = visit_expressions_mut(expr, |e| {
         match e {
-            sqlparser::ast::Expr::Identifier(ident) => {
-                if ident.value.eq_ignore_ascii_case(from_column) {
-                    ident.value = to_column.to_string();
-                }
+            sqlparser::ast::Expr::Identifier(ident)
+                if ident.value.eq_ignore_ascii_case(from_column) =>
+            {
+                ident.value = to_column.to_string();
             }
             sqlparser::ast::Expr::CompoundIdentifier(idents) => {
                 if let Some(last) = idents.last_mut()
@@ -748,23 +750,17 @@ pub fn show_variable(stmt: ShowVariables, query_ctx: QueryContextRef) -> Result<
                 .pg_intervalstyle_format();
             style.to_string()
         }
-        "MAX_EXECUTION_TIME" => {
-            if query_ctx.channel() == Channel::Mysql {
+        "MAX_EXECUTION_TIME"
+            if query_ctx.channel() == Channel::Mysql => {
                 query_ctx.query_timeout_as_millis().to_string()
-            } else {
-                return UnsupportedVariableSnafu { name: variable }.fail();
             }
-        }
-        "STATEMENT_TIMEOUT" => {
+        "STATEMENT_TIMEOUT"
             // Add time units to postgres query timeout display.
-            if query_ctx.channel() == Channel::Postgres {
+            if query_ctx.channel() == Channel::Postgres => {
                 let mut timeout = query_ctx.query_timeout_as_millis().to_string();
                 timeout.push_str("ms");
                 timeout
-            } else {
-                return UnsupportedVariableSnafu { name: variable }.fail();
             }
-        }
         _ => return UnsupportedVariableSnafu { name: variable }.fail(),
     };
     let schema = Arc::new(Schema::new(vec![ColumnSchema::new(
@@ -1062,6 +1058,10 @@ pub fn show_create_flow(
         expire_after: flow_val.expire_after(),
         eval_interval: flow_val.eval_interval(),
         comment,
+        flow_options: OptionMap::from_filtered_string_map(
+            flow_val.options(),
+            &[FlowType::FLOW_TYPE_KEY],
+        ),
         query,
     };
 
diff --git a/src/query/src/window_sort.rs b/src/query/src/window_sort.rs
index b667b9c3b0..7267de3dab 100644
--- a/src/query/src/window_sort.rs
+++ b/src/query/src/window_sort.rs
@@ -30,6 +30,7 @@ use common_error::status_code::StatusCode;
 use common_recordbatch::{DfRecordBatch, DfSendableRecordBatchStream};
 use common_telemetry::error;
 use common_time::Timestamp;
+use common_time::timestamp::TimeUnit as TimestampUnit;
 use datafusion::execution::memory_pool::{MemoryConsumer, MemoryPool};
 use datafusion::execution::{RecordBatchStream, TaskContext};
 use datafusion::physical_plan::memory::MemoryStream;
@@ -778,18 +779,11 @@ fn find_slice_from_range(
     sort_column: &SortColumn,
     range: &TimeRange,
 ) -> datafusion_common::Result<(usize, usize)> {
-    let ty = sort_column.values.data_type();
-    let time_unit = if let DataType::Timestamp(unit, _) = ty {
-        unit
-    } else {
-        return Err(DataFusionError::Internal(format!(
-            "Unsupported sort column type: {}",
-            sort_column.values.data_type()
-        )));
-    };
+    let time_unit = sort_timestamp_unit(sort_column.values.data_type())?;
     let array = &sort_column.values;
     let opt = &sort_column.options.unwrap_or_default();
     let descending = opt.descending;
+    let range = convert_time_range_for_sort(range, time_unit)?;
 
     let typed_sorted_range = [range.start, range.end]
         .iter()
@@ -836,6 +830,72 @@ fn find_slice_from_range(
     Ok((start, end - start))
 }
 
+fn sort_timestamp_unit(data_type: &DataType) -> datafusion_common::Result<arrow_schema::TimeUnit> {
+    if let DataType::Timestamp(unit, _) = data_type {
+        Ok(*unit)
+    } else {
+        Err(DataFusionError::Internal(format!(
+            "Unsupported sort column type: {data_type}"
+        )))
+    }
+}
+
+#[derive(Debug, Clone, Copy)]
+enum RangeBoundKind {
+    InclusiveStart,
+    ExclusiveEnd,
+}
+
+fn convert_time_range_for_sort(
+    range: &TimeRange,
+    time_unit: arrow_schema::TimeUnit,
+) -> datafusion_common::Result<TimeRange> {
+    let target_unit = time_unit.into();
+    Ok(TimeRange::new(
+        convert_timestamp_range_bound(range.start, target_unit, RangeBoundKind::InclusiveStart)?,
+        convert_timestamp_range_bound(range.end, target_unit, RangeBoundKind::ExclusiveEnd)?,
+    ))
+}
+
+fn convert_timestamp_range_bound(
+    timestamp: Timestamp,
+    target_unit: TimestampUnit,
+    bound_kind: RangeBoundKind,
+) -> datafusion_common::Result<Timestamp> {
+    let converted = match bound_kind {
+        RangeBoundKind::InclusiveStart => timestamp.convert_to(target_unit),
+        RangeBoundKind::ExclusiveEnd => timestamp.convert_to_ceil(target_unit),
+    };
+
+    converted.ok_or_else(|| {
+        DataFusionError::Internal(format!(
+            "Failed to convert timestamp from {:?} to {:?}",
+            timestamp.unit(),
+            target_unit
+        ))
+    })
+}
+
+pub(crate) fn project_partition_range_for_sort(
+    range: PartitionRange,
+    sort_data_type: &DataType,
+) -> datafusion_common::Result<PartitionRange> {
+    let target_unit = sort_timestamp_unit(sort_data_type)?.into();
+    Ok(PartitionRange {
+        start: convert_timestamp_range_bound(
+            range.start,
+            target_unit,
+            RangeBoundKind::InclusiveStart,
+        )?,
+        end: convert_timestamp_range_bound(range.end, target_unit, RangeBoundKind::ExclusiveEnd)?,
+        ..range
+    })
+}
+
+fn discrete_exclusive_end(timestamp: Timestamp) -> Timestamp {
+    Timestamp::new(timestamp.value() + 1, timestamp.unit())
+}
+
 /// Get an iterator from a primitive array.
 ///
 /// Used with `downcast_ts_array`. The returned iter is wrapped with `.enumerate()`.
@@ -905,7 +965,7 @@ impl SucRun<Timestamp> {
         let end = self
             .first_val
             .max(self.last_val)
-            .map(|i| Timestamp::new(i.value() + 1, i.unit()));
+            .map(discrete_exclusive_end);
         start.zip(end).map(|(s, e)| TimeRange::new(s, e))
     }
 }
@@ -1198,7 +1258,7 @@ fn split_overlapping_ranges(ranges: &[PartitionRange]) -> BTreeMap<TimeRange, Ve
         for (range, parts) in forward_iter.chain(backward_iter) {
             untouched = untouched.iter().flat_map(|r| r.difference(range)).collect();
             let act = split_range_by(range, parts, &key, idx);
-            actions.extend(act.into_iter());
+            actions.extend(act);
         }
 
         for action in actions {
@@ -1498,6 +1558,49 @@ mod test {
         }
     }
 
+    #[test]
+    fn test_project_partition_range_for_sort_uses_ceil_on_exclusive_end() {
+        let range = PartitionRange {
+            start: Timestamp::new_nanosecond(1_000_000),
+            end: Timestamp::new_nanosecond(1_000_001),
+            num_rows: 1,
+            identifier: 0,
+        };
+
+        let projected = project_partition_range_for_sort(
+            range,
+            &DataType::Timestamp(TimeUnit::Millisecond, None),
+        )
+        .unwrap();
+
+        assert_eq!(Timestamp::new_millisecond(1), projected.start);
+        assert_eq!(Timestamp::new_millisecond(2), projected.end);
+    }
+
+    #[test]
+    fn test_find_slice_from_range_preserves_last_row_after_precision_drop() {
+        let sort_column = SortColumn {
+            values: Arc::new(TimestampMillisecondArray::from_iter_values([1])) as ArrayRef,
+            options: Some(SortOptions::default()),
+        };
+        let range = TimeRange::new(
+            Timestamp::new_nanosecond(1_000_000),
+            Timestamp::new_nanosecond(1_000_001),
+        );
+
+        assert_eq!((0, 1), find_slice_from_range(&sort_column, &range).unwrap());
+    }
+
+    #[test]
+    fn test_discrete_exclusive_end_creates_half_open_upper_bound() {
+        let timestamp = Timestamp::new_millisecond(42);
+
+        assert_eq!(
+            Timestamp::new_millisecond(43),
+            discrete_exclusive_end(timestamp)
+        );
+    }
+
     #[allow(clippy::type_complexity)]
     fn run_compute_working_ranges_test(
         testcases: Vec<(
diff --git a/src/servers/Cargo.toml b/src/servers/Cargo.toml
index e75192c9ba..93d3689b1e 100644
--- a/src/servers/Cargo.toml
+++ b/src/servers/Cargo.toml
@@ -7,6 +7,7 @@ license.workspace = true
 [features]
 default = []
 dashboard = ["dep:rust-embed"]
+enterprise = ["sql/enterprise"]
 mem-prof = ["dep:common-mem-prof"]
 pprof = ["dep:common-pprof"]
 testing = []
@@ -63,6 +64,7 @@ datafusion-expr.workspace = true
 datafusion-pg-catalog.workspace = true
 datatypes.workspace = true
 derive_builder.workspace = true
+either.workspace = true
 futures.workspace = true
 futures-util.workspace = true
 headers = "0.4"
@@ -78,19 +80,22 @@ jsonb.workspace = true
 lazy_static.workspace = true
 log-query.workspace = true
 loki-proto.workspace = true
+metric-engine.workspace = true
 mime_guess = "2.0"
 notify.workspace = true
 object-pool = "0.5"
 once_cell.workspace = true
 openmetrics-parser = "0.4"
-opensrv-mysql = { git = "https://github.com/datafuselabs/opensrv", tag = "v0.10.0" }
+# Wait for https://github.com/databendlabs/opensrv/pull/81
+opensrv-mysql = { git = "https://github.com/GreptimeTeam/opensrv", rev = "6c5a451544194b7bb60a8318d155d4f892b49f2c" }
 opentelemetry-proto.workspace = true
 operator.workspace = true
 otel-arrow-rust.workspace = true
 parking_lot.workspace = true
+partition.workspace = true
 pg_interval = { version = "0.5.2", package = "pg_interval_2" }
-pgwire = { version = "0.38", default-features = false, features = [
-    "server-api-ring",
+pgwire = { version = "0.38.3", default-features = false, features = [
+    "server-api-aws-lc-rs",
     "pg-ext-types",
 ] }
 pin-project = "1.0"
@@ -105,7 +110,8 @@ rand.workspace = true
 regex.workspace = true
 reqwest.workspace = true
 rust-embed = { version = "6.6", optional = true, features = ["debug-embed"] }
-rustls = { workspace = true, default-features = false, features = ["ring", "logging", "std", "tls12"] }
+rust_decimal = { workspace = true, features = ["db-postgres"] }
+rustls = { workspace = true, default-features = false, features = ["aws_lc_rs", "logging", "std", "tls12"] }
 rustls-pemfile = "2.0"
 rustls-pki-types = "1.0"
 serde.workspace = true
@@ -113,6 +119,7 @@ serde_json.workspace = true
 session.workspace = true
 simd-json.workspace = true
 simdutf8 = "0.1"
+smallvec.workspace = true
 snafu.workspace = true
 snap = "1"
 socket2 = "0.5"
@@ -126,6 +133,7 @@ tokio-stream = { workspace = true, features = ["net"] }
 tokio-util.workspace = true
 tonic.workspace = true
 tonic-reflection = "0.14"
+tonic-web = "0.14"
 tower = { workspace = true, features = ["full"] }
 tower-http = { version = "0.6", features = ["full"] }
 tracing.workspace = true
@@ -147,7 +155,7 @@ common-test-util.workspace = true
 criterion = "0.5"
 json5 = "0.4"
 mysql_async = { version = "0.35", default-features = false, features = [
-    "default-rustls-ring",
+    "default-rustls",
 ] }
 permutation = "0.4"
 rand.workspace = true
@@ -175,3 +183,7 @@ harness = false
 [[bench]]
 name = "loki_labels"
 harness = false
+
+[[bench]]
+name = "flush_batch_physical"
+harness = false
diff --git a/src/servers/benches/flush_batch_physical.rs b/src/servers/benches/flush_batch_physical.rs
new file mode 100644
index 0000000000..a3d190adf2
--- /dev/null
+++ b/src/servers/benches/flush_batch_physical.rs
@@ -0,0 +1,289 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::HashMap;
+use std::sync::Arc;
+
+use api::region::RegionResponse;
+use api::v1::meta::Peer;
+use api::v1::region::RegionRequest;
+use arrow::array::{Float64Array, StringArray, TimestampMillisecondArray};
+use arrow::datatypes::{DataType as ArrowDataType, Field, Schema as ArrowSchema, TimeUnit};
+use arrow::record_batch::RecordBatch;
+use async_trait::async_trait;
+use catalog::error::Result as CatalogResult;
+use criterion::{BenchmarkId, Criterion, criterion_group, criterion_main};
+use datatypes::prelude::ConcreteDataType;
+use datatypes::schema::{ColumnSchema as DtColumnSchema, Schema as DtSchema};
+use partition::error::Result as PartitionResult;
+use partition::partition::{PartitionRule, PartitionRuleRef, RegionMask};
+use servers::error::{self, Result};
+use servers::pending_rows_batcher::{
+    PhysicalFlushCatalogProvider, PhysicalFlushNodeRequester, PhysicalFlushPartitionProvider,
+    PhysicalTableMetadata, TableBatch, flush_batch_physical,
+};
+use store_api::storage::RegionId;
+use table::test_util::table_info::test_table_info;
+use tokio::runtime::Runtime;
+
+// ---------------------------------------------------------------------------
+// Mock implementations (memory-backed, no I/O)
+// ---------------------------------------------------------------------------
+
+struct BenchCatalogProvider {
+    table: PhysicalTableMetadata,
+}
+
+#[async_trait]
+impl PhysicalFlushCatalogProvider for BenchCatalogProvider {
+    async fn physical_table(
+        &self,
+        _catalog: &str,
+        _schema: &str,
+        _table_name: &str,
+        _query_ctx: &session::context::QueryContext,
+    ) -> CatalogResult<Option<PhysicalTableMetadata>> {
+        Ok(Some(self.table.clone()))
+    }
+}
+
+struct BenchPartitionProvider;
+
+struct SingleRegionPartitionRule;
+
+impl PartitionRule for SingleRegionPartitionRule {
+    fn as_any(&self) -> &dyn std::any::Any {
+        self
+    }
+
+    fn partition_columns(&self) -> &[String] {
+        &[]
+    }
+
+    fn find_region(
+        &self,
+        _values: &[datatypes::prelude::Value],
+    ) -> PartitionResult<store_api::storage::RegionNumber> {
+        Ok(1)
+    }
+
+    fn split_record_batch(
+        &self,
+        record_batch: &RecordBatch,
+    ) -> PartitionResult<HashMap<store_api::storage::RegionNumber, RegionMask>> {
+        let n = record_batch.num_rows();
+        Ok(HashMap::from([(
+            1,
+            RegionMask::new(arrow::array::BooleanArray::from(vec![true; n]), n),
+        )]))
+    }
+}
+
+#[async_trait]
+impl PhysicalFlushPartitionProvider for BenchPartitionProvider {
+    async fn find_table_partition_rule(
+        &self,
+        _table_info: &table::metadata::TableInfo,
+    ) -> PartitionResult<PartitionRuleRef> {
+        Ok(Arc::new(SingleRegionPartitionRule))
+    }
+
+    async fn find_region_leader(&self, _region_id: RegionId) -> Result<Peer> {
+        Ok(Peer {
+            id: 1,
+            addr: "bench-node".to_string(),
+        })
+    }
+}
+
+struct BenchNodeRequester;
+
+#[async_trait]
+impl PhysicalFlushNodeRequester for BenchNodeRequester {
+    async fn handle(&self, _peer: &Peer, _request: RegionRequest) -> error::Result<RegionResponse> {
+        Ok(RegionResponse::new(0))
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+fn make_physical_table_metadata(num_tags: usize) -> PhysicalTableMetadata {
+    let mut columns = vec![
+        DtColumnSchema::new("__primary_key", ConcreteDataType::binary_datatype(), false),
+        DtColumnSchema::new(
+            "greptime_timestamp",
+            ConcreteDataType::timestamp_millisecond_datatype(),
+            false,
+        ),
+        DtColumnSchema::new("greptime_value", ConcreteDataType::float64_datatype(), true),
+    ];
+
+    let mut name_to_ids = HashMap::new();
+    let mut column_ids = vec![0u32, 1, 2];
+
+    for i in 0..num_tags {
+        let tag_name = format!("tag{}", i);
+        let col_id = (i + 3) as u32;
+        columns.push(DtColumnSchema::new(
+            &tag_name,
+            ConcreteDataType::string_datatype(),
+            true,
+        ));
+        name_to_ids.insert(tag_name, col_id);
+        column_ids.push(col_id);
+    }
+
+    let schema = Arc::new(DtSchema::try_new(columns).unwrap());
+    let mut table_info = test_table_info(1, "phy", "public", "greptime", schema);
+    table_info.meta.column_ids = column_ids;
+
+    PhysicalTableMetadata {
+        table_info: Arc::new(table_info),
+        col_name_to_ids: Some(name_to_ids),
+    }
+}
+
+fn make_tag_batch(tag_names: &[&str], num_rows: usize) -> RecordBatch {
+    let mut fields = vec![
+        Field::new(
+            "greptime_timestamp",
+            ArrowDataType::Timestamp(TimeUnit::Millisecond, None),
+            false,
+        ),
+        Field::new("greptime_value", ArrowDataType::Float64, true),
+    ];
+    for tag in tag_names {
+        fields.push(Field::new(*tag, ArrowDataType::Utf8, true));
+    }
+
+    let schema = Arc::new(ArrowSchema::new(fields));
+
+    let ts: Vec<i64> = (0..num_rows as i64).collect();
+    let vals: Vec<f64> = (0..num_rows).map(|i| i as f64).collect();
+
+    let mut arrays: Vec<Arc<dyn arrow::array::Array>> = vec![
+        Arc::new(TimestampMillisecondArray::from(ts)),
+        Arc::new(Float64Array::from(vals)),
+    ];
+
+    for (tag_idx, _tag) in tag_names.iter().enumerate() {
+        let values: Vec<String> = (0..num_rows)
+            .map(|i| format!("val-{}-{}", tag_idx, i))
+            .collect();
+        arrays.push(Arc::new(StringArray::from(values)));
+    }
+
+    RecordBatch::try_new(schema, arrays).unwrap()
+}
+
+fn make_table_batches(
+    num_logical_tables: usize,
+    rows_per_table: usize,
+    tag_names: &[&str],
+) -> Vec<TableBatch> {
+    (0..num_logical_tables)
+        .map(|i| {
+            let batch = make_tag_batch(tag_names, rows_per_table);
+            let row_count = batch.num_rows();
+            TableBatch {
+                table_name: format!("logical_{}", i),
+                table_id: (100 + i) as u32,
+                batches: vec![batch],
+                row_count,
+            }
+        })
+        .collect()
+}
+
+// ---------------------------------------------------------------------------
+// Benchmarks
+// ---------------------------------------------------------------------------
+
+fn bench_flush_batch_physical(c: &mut Criterion) {
+    let rt = Runtime::new().unwrap();
+    let ctx = session::context::QueryContext::arc();
+
+    let num_tags = 5;
+    let tag_names: Vec<String> = (0..num_tags).map(|i| format!("tag{}", i)).collect();
+    let tag_refs: Vec<&str> = tag_names.iter().map(|s| s.as_str()).collect();
+
+    let catalog_provider = BenchCatalogProvider {
+        table: make_physical_table_metadata(num_tags),
+    };
+    let partition_provider = BenchPartitionProvider;
+    let node_requester = BenchNodeRequester;
+
+    let mut group = c.benchmark_group("flush_batch_physical");
+
+    // Vary the number of logical tables
+    for num_tables in [1, 10, 50, 100] {
+        let rows_per_table = 100;
+        let table_batches = make_table_batches(num_tables, rows_per_table, &tag_refs);
+
+        group.bench_with_input(
+            BenchmarkId::new("tables", num_tables),
+            &table_batches,
+            |b, batches| {
+                b.iter(|| {
+                    rt.block_on(async {
+                        flush_batch_physical(
+                            batches,
+                            "phy",
+                            &ctx,
+                            &partition_provider,
+                            &node_requester,
+                            &catalog_provider,
+                        )
+                        .await
+                        .unwrap();
+                    });
+                });
+            },
+        );
+    }
+
+    // Vary the number of rows per table
+    for rows_per_table in [10, 100, 1000, 5000] {
+        let num_tables = 10;
+        let table_batches = make_table_batches(num_tables, rows_per_table, &tag_refs);
+
+        group.bench_with_input(
+            BenchmarkId::new("rows_per_table", rows_per_table),
+            &table_batches,
+            |b, batches| {
+                b.iter(|| {
+                    rt.block_on(async {
+                        flush_batch_physical(
+                            batches,
+                            "phy",
+                            &ctx,
+                            &partition_provider,
+                            &node_requester,
+                            &catalog_provider,
+                        )
+                        .await
+                        .unwrap();
+                    });
+                });
+            },
+        );
+    }
+
+    group.finish();
+}
+
+criterion_group!(benches, bench_flush_batch_physical);
+criterion_main!(benches);
diff --git a/src/servers/build.rs b/src/servers/build.rs
index 5f7081d024..0e38a7f83f 100644
--- a/src/servers/build.rs
+++ b/src/servers/build.rs
@@ -19,6 +19,7 @@ fn main() {
 
 #[cfg(feature = "dashboard")]
 fn fetch_dashboard_assets() {
+    use std::path::PathBuf;
     use std::process::{Command, Stdio};
 
     let message = "Failed to fetch dashboard assets";
@@ -30,7 +31,16 @@ or it's a network error, just try again or enable/disable some proxy."#;
     let mut dir = std::env::current_dir().unwrap();
     dir.pop();
     dir.pop();
-    dir.push("scripts");
+    let scripts_dir = dir.join("scripts");
+    let dashboard_dist = dir.join(PathBuf::from("src/servers/dashboard/dist"));
+
+    if dashboard_dist.join("index.html").exists() {
+        println!("cargo:rerun-if-changed=dashboard/VERSION");
+        println!("cargo:rerun-if-changed=dashboard/dist");
+        return;
+    }
+
+    dir = scripts_dir;
 
     let out_dir = std::env::var("OUT_DIR").unwrap();
 
diff --git a/src/servers/dashboard/VERSION b/src/servers/dashboard/VERSION
index 03ee1a5314..60d68b2300 100644
--- a/src/servers/dashboard/VERSION
+++ b/src/servers/dashboard/VERSION
@@ -1 +1 @@
-v0.11.13
+v0.12.2
diff --git a/src/servers/src/error.rs b/src/servers/src/error.rs
index 18ac964f05..8a3c554058 100644
--- a/src/servers/src/error.rs
+++ b/src/servers/src/error.rs
@@ -15,6 +15,7 @@
 use std::any::Any;
 use std::net::SocketAddr;
 use std::string::FromUtf8Error;
+use std::sync::Arc;
 
 use axum::http::StatusCode as HttpStatusCode;
 use axum::response::{IntoResponse, Response};
@@ -51,11 +52,16 @@ pub enum Error {
     Arrow {
         #[snafu(source)]
         error: arrow_schema::ArrowError,
+        #[snafu(implicit)]
+        location: Location,
     },
 
     #[snafu(display("Internal error: {}", err_msg))]
     Internal { err_msg: String },
 
+    #[snafu(display("Pending rows batcher channel closed"))]
+    BatcherChannelClosed,
+
     #[snafu(display("Unsupported data type: {}, reason: {}", data_type, reason))]
     UnsupportedDataType {
         data_type: ConcreteDataType,
@@ -389,7 +395,7 @@ pub enum Error {
         location: Location,
     },
 
-    #[snafu(display("Error accessing catalog"))]
+    #[snafu(transparent)]
     Catalog {
         source: catalog::error::Error,
         #[snafu(implicit)]
@@ -675,6 +681,30 @@ pub enum Error {
         #[snafu(implicit)]
         location: Location,
     },
+
+    #[snafu(transparent)]
+    DataTypes {
+        source: datatypes::error::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(transparent)]
+    Partition {
+        source: partition::error::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(transparent)]
+    MetricEngine {
+        source: metric_engine::error::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Failed to submit batch: {}", source))]
+    SubmitBatch { source: Arc<Error> },
 }
 
 pub type Result<T, E = Error> = std::result::Result<T, E>;
@@ -684,6 +714,7 @@ impl ErrorExt for Error {
         use Error::*;
         match self {
             Internal { .. }
+            | BatcherChannelClosed
             | InternalIo { .. }
             | TokioIo { .. }
             | StartHttp { .. }
@@ -752,6 +783,7 @@ impl ErrorExt for Error {
 
             Catalog { source, .. } => source.status_code(),
             RowWriter { source, .. } => source.status_code(),
+            DataTypes { source, .. } => source.status_code(),
 
             TlsRequired { .. } => StatusCode::Unknown,
             Auth { source, .. } => source.status_code(),
@@ -806,6 +838,9 @@ impl ErrorExt for Error {
             MemoryLimitExceeded { .. } => StatusCode::RateLimited,
 
             GreptimeProto { source, .. } => source.status_code(),
+            Partition { source, .. } => source.status_code(),
+            MetricEngine { source, .. } => source.status_code(),
+            SubmitBatch { source, .. } => source.status_code(),
         }
     }
 
diff --git a/src/servers/src/grpc.rs b/src/servers/src/grpc.rs
index ab4caafc34..3adfd24945 100644
--- a/src/servers/src/grpc.rs
+++ b/src/servers/src/grpc.rs
@@ -50,6 +50,7 @@ use tonic::{Request, Response, Status};
 use tonic_reflection::server::v1::{ServerReflection, ServerReflectionServer};
 
 use crate::error::{AlreadyStartedSnafu, InternalSnafu, Result, StartGrpcSnafu, TcpBindSnafu};
+use crate::install_default_crypto_provider;
 use crate::metrics::MetricsMiddlewareLayer;
 use crate::otel_arrow::{HeaderInterceptor, OtelArrowServiceHandler};
 use crate::query_handler::OpenTelemetryProtocolHandlerRef;
@@ -355,8 +356,17 @@ impl Server for GrpcServer {
             .layer(MetricsMiddlewareLayer)
             .into_inner();
 
-        let mut builder = tonic::transport::Server::builder().layer(metrics_layer);
+        let mut builder = tonic::transport::Server::builder()
+            .accept_http1(true)
+            .layer(metrics_layer)
+            .layer(tonic_web::GrpcWebLayer::new());
+
         if let Some(tls_config) = self.tls_config.clone() {
+            // tonic builds the underlying rustls server config here, which requires a
+            // process-level crypto provider to be installed first.
+            if let Err(err) = install_default_crypto_provider() {
+                warn!("Failed to install default rustls crypto provider: {err}");
+            }
             builder = builder.tls_config(tls_config).context(StartGrpcSnafu)?;
         }
 
diff --git a/src/servers/src/grpc/flight.rs b/src/servers/src/grpc/flight.rs
index 02755fcfd0..21a03a41d3 100644
--- a/src/servers/src/grpc/flight.rs
+++ b/src/servers/src/grpc/flight.rs
@@ -25,10 +25,12 @@ use arrow_flight::{
     HandshakeRequest, HandshakeResponse, PollInfo, PutResult, SchemaResult, Ticket,
 };
 use async_trait::async_trait;
-use bytes;
-use bytes::Bytes;
+use bytes::{self, Bytes};
+use common_error::ext::ErrorExt;
 use common_grpc::flight::do_put::{DoPutMetadata, DoPutResponse};
-use common_grpc::flight::{FlightDecoder, FlightEncoder, FlightMessage};
+use common_grpc::flight::{
+    FLOW_EXTENSIONS_METADATA_KEY, FlightDecoder, FlightEncoder, FlightMessage,
+};
 use common_memory_manager::MemoryGuard;
 use common_query::{Output, OutputData};
 use common_recordbatch::DfRecordBatch;
@@ -39,7 +41,9 @@ use datatypes::arrow::datatypes::SchemaRef;
 use futures::{Stream, future, ready};
 use futures_util::{StreamExt, TryStreamExt};
 use prost::Message;
-use session::context::{QueryContext, QueryContextRef};
+use query::metrics::terminal_recordbatch_metrics_from_plan_if_requested;
+use query::options::FlowQueryExtensions;
+use session::context::{Channel, QueryContextRef};
 use snafu::{IntoError, ResultExt, ensure};
 use table::table_name::TableName;
 use tokio::sync::mpsc;
@@ -48,7 +52,9 @@ use tonic::{Request, Response, Status, Streaming};
 
 use crate::error::{InvalidParameterSnafu, Result, ToJsonSnafu};
 pub use crate::grpc::flight::stream::FlightRecordBatchStream;
-use crate::grpc::greptime_handler::{GreptimeRequestHandler, get_request_type};
+use crate::grpc::greptime_handler::{
+    GreptimeRequestHandler, create_query_context, get_request_type,
+};
 use crate::grpc::{FlightCompression, TonicResult, context_auth};
 use crate::request_memory_limiter::ServerMemoryLimiter;
 use crate::request_memory_metrics::RequestMemoryMetrics;
@@ -187,11 +193,22 @@ impl FlightCraft for GreptimeRequestHandler {
         &self,
         request: Request<Ticket>,
     ) -> TonicResult<Response<TonicStream<FlightData>>> {
-        let hints = hint_headers::extract_hints(request.metadata());
+        let mut hints = hint_headers::extract_hints(request.metadata());
+        hints.extend(extract_flow_extensions(request.metadata())?);
 
         let ticket = request.into_inner().ticket;
         let request =
             GreptimeRequest::decode(ticket.as_ref()).context(error::InvalidFlightTicketSnafu)?;
+        let query_ctx =
+            create_query_context(Channel::Grpc, request.header.as_ref(), hints.clone())?;
+        // Validate flow hint syntax at the transport boundary before dispatching the request.
+        // This does not authorize or execute anything; `handle_request()` below still performs
+        // the normal frontend handling and auth checks before query execution.
+        let flow_extensions = FlowQueryExtensions::parse_flow_extensions(&query_ctx.extensions())
+            .map_err(|e| Status::invalid_argument(e.output_msg()))?;
+        let should_emit_terminal_metrics = flow_extensions
+            .as_ref()
+            .is_some_and(|extensions| extensions.should_collect_region_watermark());
 
         // The Grpc protocol pass query by Flight. It needs to be wrapped under a span, in order to record stream
         let span = info_span!(
@@ -206,7 +223,8 @@ impl FlightCraft for GreptimeRequestHandler {
                 output,
                 TracingContext::from_current_span(),
                 flight_compression,
-                QueryContext::arc(),
+                query_ctx,
+                should_emit_terminal_metrics,
             );
             Ok(Response::new(stream))
         }
@@ -513,11 +531,32 @@ impl Stream for PutRecordBatchRequestStream {
     }
 }
 
+fn extract_flow_extensions(
+    metadata: &tonic::metadata::MetadataMap,
+) -> TonicResult<Vec<(String, String)>> {
+    let Some(value) = metadata.get(FLOW_EXTENSIONS_METADATA_KEY) else {
+        return Ok(vec![]);
+    };
+
+    let value = value.to_str().map_err(|e| {
+        Status::invalid_argument(format!(
+            "Invalid {FLOW_EXTENSIONS_METADATA_KEY} metadata value: {e}"
+        ))
+    })?;
+
+    serde_json::from_str::<Vec<(String, String)>>(value).map_err(|e| {
+        Status::invalid_argument(format!(
+            "Invalid {FLOW_EXTENSIONS_METADATA_KEY} metadata JSON: {e}"
+        ))
+    })
+}
+
 fn to_flight_data_stream(
     output: Output,
     tracing_context: TracingContext,
     flight_compression: FlightCompression,
     query_ctx: QueryContextRef,
+    should_emit_terminal_metrics: bool,
 ) -> TonicStream<FlightData> {
     match output.data {
         OutputData::Stream(stream) => {
@@ -539,13 +578,70 @@ fn to_flight_data_stream(
             Box::pin(stream) as _
         }
         OutputData::AffectedRows(rows) => {
-            let stream = tokio_stream::iter(
-                FlightEncoder::default()
-                    .encode(FlightMessage::AffectedRows(rows))
-                    .into_iter()
-                    .map(Ok),
-            );
+            let terminal_metrics = match terminal_recordbatch_metrics_from_plan_if_requested(
+                output.meta.plan,
+                should_emit_terminal_metrics,
+            ) {
+                Some(metrics) => match serde_json::to_string(&metrics) {
+                    Ok(metrics) => Some(metrics),
+                    Err(e) => {
+                        let stream = tokio_stream::once(Err(Status::internal(format!(
+                            "Failed to serialize terminal metrics: {e}"
+                        ))));
+                        return Box::pin(stream) as _;
+                    }
+                },
+                None => None,
+            };
+            let affected_rows = FlightEncoder::default().encode(FlightMessage::AffectedRows {
+                rows,
+                metrics: terminal_metrics,
+            });
+            let stream = tokio_stream::iter(affected_rows.into_iter().map(Ok));
             Box::pin(stream) as _
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use tonic::metadata::{AsciiMetadataValue, MetadataMap};
+
+    use super::*;
+
+    #[test]
+    fn test_extract_flow_extensions_preserves_comma_bearing_values() {
+        let mut metadata = MetadataMap::new();
+        metadata.insert(
+            FLOW_EXTENSIONS_METADATA_KEY,
+            AsciiMetadataValue::try_from(
+                r#"[["flow.return_region_seq","true"],["flow.incremental_after_seqs","{\"1\":10,\"2\":20}"]]"#,
+            )
+            .unwrap(),
+        );
+
+        let extensions = extract_flow_extensions(&metadata).unwrap();
+        assert_eq!(
+            extensions,
+            vec![
+                ("flow.return_region_seq".to_string(), "true".to_string()),
+                (
+                    "flow.incremental_after_seqs".to_string(),
+                    r#"{"1":10,"2":20}"#.to_string()
+                ),
+            ]
+        );
+    }
+
+    #[test]
+    fn test_extract_flow_extensions_rejects_invalid_json() {
+        let mut metadata = MetadataMap::new();
+        metadata.insert(
+            FLOW_EXTENSIONS_METADATA_KEY,
+            AsciiMetadataValue::try_from("not-json").unwrap(),
+        );
+
+        let err = extract_flow_extensions(&metadata).unwrap_err();
+        assert_eq!(err.code(), tonic::Code::InvalidArgument);
+    }
+}
diff --git a/src/servers/src/http.rs b/src/servers/src/http.rs
index ca6a77a077..e5dc5380d1 100644
--- a/src/servers/src/http.rs
+++ b/src/servers/src/http.rs
@@ -16,7 +16,7 @@ use std::collections::HashMap;
 use std::convert::Infallible;
 use std::fmt::Display;
 use std::net::SocketAddr;
-use std::sync::Mutex as StdMutex;
+use std::sync::{Arc, Mutex as StdMutex};
 use std::time::Duration;
 
 use async_trait::async_trait;
@@ -75,10 +75,11 @@ use crate::http::result::null_result::NullResponse;
 use crate::interceptor::LogIngestInterceptorRef;
 use crate::metrics::http_metrics_layer;
 use crate::metrics_handler::MetricsHandler;
+use crate::pending_rows_batcher::PendingRowsBatcher;
 use crate::prometheus_handler::PrometheusHandlerRef;
 use crate::query_handler::sql::ServerSqlQueryHandlerRef;
 use crate::query_handler::{
-    InfluxdbLineProtocolHandlerRef, JaegerQueryHandlerRef, LogQueryHandlerRef,
+    DashboardHandlerRef, InfluxdbLineProtocolHandlerRef, JaegerQueryHandlerRef, LogQueryHandlerRef,
     OpenTelemetryProtocolHandlerRef, OpentsdbProtocolHandlerRef, PipelineHandlerRef,
     PromStoreProtocolHandlerRef,
 };
@@ -112,8 +113,8 @@ pub mod utils;
 use result::HttpOutputWriter;
 pub(crate) use timeout::DynamicTimeoutLayer;
 
+mod client_ip;
 use crate::prom_remote_write::validation::PromValidationMode;
-
 mod hints;
 mod read_preference;
 #[cfg(any(test, feature = "testing"))]
@@ -128,7 +129,8 @@ const DEFAULT_BODY_LIMIT: ReadableSize = ReadableSize::mb(64);
 pub const AUTHORIZATION_HEADER: &str = "x-greptime-auth";
 
 // TODO(fys): This is a temporary workaround, it will be improved later
-pub static PUBLIC_APIS: [&str; 3] = ["/v1/influxdb/ping", "/v1/influxdb/health", "/v1/health"];
+pub static PUBLIC_API_PREFIX: [&str; 3] =
+    ["/v1/influxdb/ping", "/v1/influxdb/health", "/v1/health"];
 
 #[derive(Default)]
 pub struct HttpServer {
@@ -507,6 +509,11 @@ pub struct GreptimeOptionsConfigState {
     pub greptime_config_options: String,
 }
 
+#[derive(Clone)]
+pub struct DashboardState {
+    pub handler: DashboardHandlerRef,
+}
+
 pub struct HttpServerBuilder {
     options: HttpOptions,
     plugins: Plugins,
@@ -580,12 +587,14 @@ impl HttpServerBuilder {
         pipeline_handler: Option<PipelineHandlerRef>,
         prom_store_with_metric_engine: bool,
         prom_validation_mode: PromValidationMode,
+        pending_rows_batcher: Option<Arc<PendingRowsBatcher>>,
     ) -> Self {
         let state = PromStoreState {
             prom_store_handler: handler,
             pipeline_handler,
             prom_store_with_metric_engine,
             prom_validation_mode,
+            pending_rows_batcher,
         };
 
         Self {
@@ -703,6 +712,16 @@ impl HttpServerBuilder {
         }
     }
 
+    pub fn with_dashboard_handler(self, handler: DashboardHandlerRef) -> Self {
+        Self {
+            router: self.router.nest(
+                &format!("/{HTTP_API_VERSION}/dashboards"),
+                HttpServer::route_dashboard(handler),
+            ),
+            ..self
+        }
+    }
+
     pub fn with_extra_router(self, router: Router) -> Self {
         Self {
             router: self.router.merge(router),
@@ -746,6 +765,7 @@ impl HttpServer {
         };
 
         router = router
+            .route("/", routing::get(handler::index))
             .route(
                 "/health",
                 routing::get(handler::health).post(handler::health),
@@ -868,6 +888,7 @@ impl HttpServer {
                         authorize::check_http_auth,
                     ))
                     .layer(middleware::from_fn(hints::extract_hints))
+                    .layer(middleware::from_fn(client_ip::log_error_with_client_ip))
                     .layer(middleware::from_fn(
                         read_preference::extract_read_preference,
                     )),
@@ -1169,6 +1190,26 @@ impl HttpServer {
             )
             .with_state(handler)
     }
+
+    #[cfg(feature = "dashboard")]
+    fn route_dashboard<S>(handler: DashboardHandlerRef) -> Router<S> {
+        use crate::http::dashboard::{add_dashboard, delete_dashboard, list_dashboards};
+
+        Router::new()
+            .route("/", routing::get(list_dashboards))
+            .route("/{dashboard_name}", routing::post(add_dashboard))
+            .route("/{dashboard_name}", routing::delete(delete_dashboard))
+            .layer(
+                ServiceBuilder::new()
+                    .layer(RequestDecompressionLayer::new().pass_through_unaccepted(true)),
+            )
+            .with_state(DashboardState { handler })
+    }
+
+    #[cfg(not(feature = "dashboard"))]
+    fn route_dashboard<S>(handler: DashboardHandlerRef) -> Router<S> {
+        Router::new().with_state(DashboardState { handler })
+    }
 }
 
 pub const HTTP_SERVER: &str = "HTTP_SERVER";
@@ -1212,7 +1253,10 @@ impl Server for HttpServer {
                         error!(e; "Failed to set TCP_NODELAY on incoming connection");
                     }
                 });
-            let serve = axum::serve(listener, app.into_make_service());
+            let serve = axum::serve(
+                listener,
+                app.into_make_service_with_connect_info::<SocketAddr>(),
+            );
 
             // FIXME(yingwen): Support keepalive.
             // See:
@@ -1270,7 +1314,7 @@ mod test {
     use std::io::Cursor;
     use std::sync::Arc;
 
-    use arrow_ipc::reader::FileReader;
+    use arrow_ipc::reader::StreamReader;
     use arrow_schema::DataType;
     use axum::handler::Handler;
     use axum::http::StatusCode;
@@ -1310,8 +1354,8 @@ mod test {
 
         async fn do_exec_plan(
             &self,
-            _stmt: Option<Statement>,
             _plan: LogicalPlan,
+            _stmt: Option<Statement>,
             _query_ctx: QueryContextRef,
         ) -> Result<Output> {
             unimplemented!()
@@ -1642,8 +1686,8 @@ mod test {
 
                 HttpResponse::Arrow(resp) => {
                     let output = resp.data;
-                    let mut reader =
-                        FileReader::try_new(Cursor::new(output), None).expect("Arrow reader error");
+                    let mut reader = StreamReader::try_new(Cursor::new(output), None)
+                        .expect("Arrow reader error");
                     let schema = reader.schema();
                     assert_eq!(schema.fields[0].name(), "numbers");
                     assert_eq!(schema.fields[0].data_type(), &DataType::UInt32);
diff --git a/src/servers/src/http/authorize.rs b/src/servers/src/http/authorize.rs
index 3b73482961..226f824acb 100644
--- a/src/servers/src/http/authorize.rs
+++ b/src/servers/src/http/authorize.rs
@@ -37,7 +37,7 @@ use crate::error::{
 };
 use crate::http::header::{GREPTIME_TIMEZONE_HEADER_NAME, GreptimeDbName};
 use crate::http::result::error_result::ErrorResponse;
-use crate::http::{AUTHORIZATION_HEADER, HTTP_API_PREFIX, PUBLIC_APIS};
+use crate::http::{AUTHORIZATION_HEADER, HTTP_API_PREFIX, PUBLIC_API_PREFIX};
 use crate::influxdb::{is_influxdb_request, is_influxdb_v2_request};
 
 /// AuthState is a holder state for [`UserProviderRef`]
@@ -285,7 +285,7 @@ fn decode_basic(credential: Credential) -> Result<(Username, Password)> {
 fn need_auth<B>(req: &Request<B>) -> bool {
     let path = req.uri().path();
 
-    for api in PUBLIC_APIS {
+    for api in PUBLIC_API_PREFIX {
         if path.starts_with(api) {
             return false;
         }
@@ -330,7 +330,7 @@ fn extract_influxdb_user_from_query(query: &str) -> (Option<&str>, Option<&str>)
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_base::secrets::ExposeSecret;
 
@@ -362,14 +362,13 @@ mod tests {
 
     #[test]
     fn test_decode_basic() {
-        // base64encode("username:password") == "dXNlcm5hbWU6cGFzc3dvcmQ="
-        let credential = "dXNlcm5hbWU6cGFzc3dvcmQ=";
-        let (username, pwd) = decode_basic(credential).unwrap();
+        let credential = basic_auth_credentials("username", "password");
+        let (username, pwd) = decode_basic(&credential).unwrap();
         assert_eq!("username", username);
         assert_eq!("password", pwd.expose_secret());
 
-        let wrong_credential = "dXNlcm5hbWU6cG Fzc3dvcmQ=";
-        let result = decode_basic(wrong_credential);
+        let wrong_credential = credential.replacen('c', "c ", 1);
+        let result = decode_basic(&wrong_credential);
         assert_matches!(result.err(), Some(error::Error::InvalidBase64Value { .. }));
     }
 
@@ -379,8 +378,8 @@ mod tests {
         let re: Result<AuthScheme> = auth_scheme_str.try_into();
         assert!(re.is_err());
 
-        let auth_scheme_str = "basic dGVzdDp0ZXN0";
-        let scheme: AuthScheme = auth_scheme_str.try_into().unwrap();
+        let auth_scheme_str = basic_auth("test", "test");
+        let scheme: AuthScheme = auth_scheme_str.as_str().try_into().unwrap();
         assert_matches!(scheme, AuthScheme::Basic(username, pwd) if username == "test" && pwd.expose_secret() == "test");
 
         let unsupported = "digest";
@@ -390,21 +389,37 @@ mod tests {
 
     #[test]
     fn test_auth_header() {
-        // base64encode("username:password") == "dXNlcm5hbWU6cGFzc3dvcmQ="
-        let req = mock_http_request(Some("Basic dXNlcm5hbWU6cGFzc3dvcmQ="), None).unwrap();
+        let header_value = basic_auth("username", "password");
+        let req = mock_http_request(Some(&header_value), None).unwrap();
 
         let auth_scheme = auth_header(&req).unwrap();
         assert_matches!(auth_scheme, AuthScheme::Basic(username, pwd) if username == "username" && pwd.expose_secret() == "password");
 
-        let wrong_req = mock_http_request(Some("Basic dXNlcm5hbWU6 cGFzc3dvcmQ="), None).unwrap();
+        let wrong_auth_header = header_value.replacen('c', "c ", 1);
+        let wrong_req = mock_http_request(Some(&wrong_auth_header), None).unwrap();
         let res = auth_header(&wrong_req);
         assert_matches!(res.err(), Some(error::Error::InvalidAuthHeader { .. }));
 
-        let wrong_req = mock_http_request(Some("Digest dXNlcm5hbWU6cGFzc3dvcmQ="), None).unwrap();
+        let wrong_req = mock_http_request(
+            Some(&format!(
+                "Digest {}",
+                basic_auth_credentials("username", "password")
+            )),
+            None,
+        )
+        .unwrap();
         let res = auth_header(&wrong_req);
         assert_matches!(res.err(), Some(error::Error::UnsupportedAuthScheme { .. }));
     }
 
+    fn basic_auth(username: &str, password: &str) -> String {
+        format!("Basic {}", basic_auth_credentials(username, password))
+    }
+
+    fn basic_auth_credentials(username: &str, password: &str) -> String {
+        BASE64_STANDARD.encode(format!("{username}:{password}"))
+    }
+
     fn mock_http_request(auth_header: Option<&str>, uri: Option<&str>) -> Result<Request<()>> {
         let http_api_version = crate::http::HTTP_API_VERSION;
         let mut req = Request::builder()
diff --git a/src/servers/src/http/client_ip.rs b/src/servers/src/http/client_ip.rs
new file mode 100644
index 0000000000..70df554ebb
--- /dev/null
+++ b/src/servers/src/http/client_ip.rs
@@ -0,0 +1,109 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::net::SocketAddr;
+
+use axum::body::Body;
+use axum::extract::{ConnectInfo, MatchedPath};
+use axum::http::Request;
+use axum::middleware::Next;
+use axum::response::Response;
+use common_telemetry::warn;
+
+/// Middleware that logs HTTP error responses (4xx/5xx) with client IP address.
+///
+/// Extracts client address from [`ConnectInfo`] if available.
+pub async fn log_error_with_client_ip(req: Request<Body>, next: Next) -> Response {
+    let request_info = req
+        .extensions()
+        .get::<ConnectInfo<SocketAddr>>()
+        .map(|c| c.0)
+        .map(|addr| {
+            let method = req.method().clone();
+            let uri = req.uri().clone();
+            let matched_path = req.extensions().get::<MatchedPath>().cloned();
+            (addr, method, uri, matched_path)
+        });
+
+    let response = next.run(req).await;
+
+    if (response.status().is_client_error() || response.status().is_server_error())
+        && let Some((addr, method, uri, matched_path)) = request_info
+    {
+        warn!(
+            "HTTP error response {} for {} {} (matched: {}) from client {}",
+            response.status(),
+            method,
+            uri,
+            matched_path
+                .as_ref()
+                .map(|p| p.as_str())
+                .unwrap_or("<unknown>"),
+            addr
+        );
+    }
+
+    response
+}
+
+#[cfg(test)]
+mod tests {
+    use axum::Router;
+    use axum::routing::get;
+    use http::StatusCode;
+    use tower::ServiceExt;
+
+    use super::*;
+
+    #[tokio::test]
+    async fn test_middleware_passes_error_response() {
+        async fn not_found_handler() -> StatusCode {
+            StatusCode::NOT_FOUND
+        }
+
+        let app = Router::new()
+            .route("/not-found", get(not_found_handler))
+            .layer(axum::middleware::from_fn(log_error_with_client_ip));
+
+        let response = app
+            .oneshot(
+                Request::builder()
+                    .uri("/not-found")
+                    .body(Body::empty())
+                    .unwrap(),
+            )
+            .await
+            .unwrap();
+
+        assert_eq!(response.status(), StatusCode::NOT_FOUND);
+    }
+
+    #[tokio::test]
+    async fn test_middleware_passes_success_response() {
+        async fn ok_handler() -> StatusCode {
+            StatusCode::OK
+        }
+
+        let app = Router::new()
+            .route("/ok", get(ok_handler))
+            .layer(axum::middleware::from_fn(log_error_with_client_ip));
+
+        let response = app
+            .oneshot(Request::builder().uri("/ok").body(Body::empty()).unwrap())
+            .await
+            .unwrap();
+
+        assert_eq!(response.status(), StatusCode::OK);
+    }
+}
diff --git a/src/servers/src/http/dashboard.rs b/src/servers/src/http/dashboard.rs
index bdb98490f0..ea894ca7d0 100644
--- a/src/servers/src/http/dashboard.rs
+++ b/src/servers/src/http/dashboard.rs
@@ -12,14 +12,21 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use axum::body::Body;
+use std::sync::Arc;
+use std::time::Instant;
+
+use axum::body::{Body, Bytes};
+use axum::extract::{Extension, Path, State};
 use axum::http::{StatusCode, Uri, header};
 use axum::response::Response;
-use common_telemetry::debug;
+use common_telemetry::{debug, error};
 use rust_embed::RustEmbed;
-use snafu::ResultExt;
+use session::context::{Channel, QueryContext};
+use snafu::{ResultExt, ensure};
 
-use crate::error::{BuildHttpResponseSnafu, Result};
+use crate::error::{BuildHttpResponseSnafu, InvalidParameterSnafu, Result};
+use crate::http::DashboardState;
+use crate::http::result::greptime_manage_resp::{DashboardOutput, GreptimedbManageResponse};
 
 #[derive(RustEmbed)]
 #[folder = "dashboard/dist/"]
@@ -61,3 +68,102 @@ fn get_assets(path: &str) -> Result<Response> {
     }
     .context(BuildHttpResponseSnafu)
 }
+
+#[axum_macros::debug_handler]
+pub async fn add_dashboard(
+    State(state): State<DashboardState>,
+    Path(dashboard_name): Path<String>,
+    Extension(mut query_ctx): Extension<QueryContext>,
+    payload: Bytes,
+) -> Result<GreptimedbManageResponse> {
+    let start = Instant::now();
+    let handler = state.handler;
+    ensure!(
+        !dashboard_name.is_empty(),
+        InvalidParameterSnafu {
+            reason: "dashboard_name is required in path",
+        }
+    );
+
+    let definition = String::from_utf8_lossy(&payload).to_string();
+
+    query_ctx.set_channel(Channel::HttpSql);
+    let query_ctx = Arc::new(query_ctx);
+
+    handler
+        .save(&dashboard_name, &definition, query_ctx)
+        .await
+        .map(|_| {
+            GreptimedbManageResponse::from_dashboard(
+                dashboard_name,
+                start.elapsed().as_millis() as u64,
+            )
+        })
+        .map_err(|e| {
+            error!(e; "failed to save dashboard");
+            e
+        })
+}
+
+#[axum_macros::debug_handler]
+pub async fn list_dashboards(
+    State(state): State<DashboardState>,
+    Extension(mut query_ctx): Extension<QueryContext>,
+) -> Result<GreptimedbManageResponse> {
+    let start = Instant::now();
+    let handler = state.handler;
+
+    query_ctx.set_channel(Channel::HttpSql);
+    let query_ctx = Arc::new(query_ctx);
+
+    handler
+        .list(query_ctx)
+        .await
+        .map(|dashboards| {
+            let outputs: Vec<DashboardOutput> = dashboards
+                .into_iter()
+                .map(|d| DashboardOutput {
+                    name: d.name,
+                    definition: d.definition,
+                })
+                .collect();
+            GreptimedbManageResponse::from_dashboards(outputs, start.elapsed().as_millis() as u64)
+        })
+        .map_err(|e| {
+            error!(e; "failed to list dashboards");
+            e
+        })
+}
+
+#[axum_macros::debug_handler]
+pub async fn delete_dashboard(
+    State(state): State<DashboardState>,
+    Extension(mut query_ctx): Extension<QueryContext>,
+    Path(dashboard_name): Path<String>,
+) -> Result<GreptimedbManageResponse> {
+    let start = Instant::now();
+    let handler = state.handler;
+    ensure!(
+        !dashboard_name.is_empty(),
+        InvalidParameterSnafu {
+            reason: "dashboard_name is required",
+        }
+    );
+
+    query_ctx.set_channel(Channel::HttpSql);
+    let query_ctx = Arc::new(query_ctx);
+
+    handler
+        .delete(&dashboard_name, query_ctx)
+        .await
+        .map(|_| {
+            GreptimedbManageResponse::from_dashboard(
+                dashboard_name,
+                start.elapsed().as_millis() as u64,
+            )
+        })
+        .map_err(|e| {
+            error!(e; "failed to delete dashboard");
+            e
+        })
+}
diff --git a/src/servers/src/http/event.rs b/src/servers/src/http/event.rs
index 24bb844dc7..dc468a9b75 100644
--- a/src/servers/src/http/event.rs
+++ b/src/servers/src/http/event.rs
@@ -49,7 +49,7 @@ use table::table_reference::TableReference;
 use vrl::value::{KeyString, Value as VrlValue};
 
 use crate::error::{
-    CatalogSnafu, Error, InvalidParameterSnafu, OtherSnafu, ParseJsonSnafu, PipelineSnafu, Result,
+    Error, InvalidParameterSnafu, OtherSnafu, ParseJsonSnafu, PipelineSnafu, Result,
     status_code_to_http_status,
 };
 use crate::http::HttpResponse;
@@ -265,12 +265,7 @@ pub async fn query_pipeline_ddl(
         .map_err(BoxedError::new)
         .context(OtherSnafu)?;
 
-    let message = if handler
-        .get_table(&table_name, &query_ctx)
-        .await
-        .context(CatalogSnafu)?
-        .is_some()
-    {
+    let message = if handler.get_table(&table_name, &query_ctx).await?.is_some() {
         Some(CREATE_TABLE_SQL_TABLE_EXISTS.to_string())
     } else if pipeline.is_variant_table_name() {
         Some(CREATE_TABLE_SQL_SUFFIX_EXISTS.to_string())
diff --git a/src/servers/src/http/handler.rs b/src/servers/src/http/handler.rs
index 7ba54e3ad1..4cfbb57691 100644
--- a/src/servers/src/http/handler.rs
+++ b/src/servers/src/http/handler.rs
@@ -478,3 +478,25 @@ async fn validate_schema(
         )),
     }
 }
+
+pub async fn index() -> axum::response::Html<String> {
+    let name = common_version::product_name();
+    let version = common_version::version();
+    axum::response::Html(format!(
+        r#"<!DOCTYPE html>
+<html>
+<head><title>{name}</title></head>
+<body>
+<h1>{name}</h1>
+<p>Version: {version}</p>
+<ul>
+<li><a href="/dashboard">Dashboard UI</a></li>
+<li><a href="/v1/health">Health</a> (JSON)</li>
+<li><a href="/status">Status</a> (JSON)</li>
+<li><a href="/metrics">Metrics</a> (For Prometheus Scrape)</li>
+<li><a href="/config">Config</a> (TXT)</li>
+</ul>
+</body>
+</html>"#,
+    ))
+}
diff --git a/src/servers/src/http/header.rs b/src/servers/src/http/header.rs
index ce58e15e98..785901ad7b 100644
--- a/src/servers/src/http/header.rs
+++ b/src/servers/src/http/header.rs
@@ -170,13 +170,11 @@ pub fn collect_plan_metrics(plan: &Arc<dyn ExecutionPlan>, maps: &mut [&mut Hash
             MetricValue::Gauge { name, gauge } => {
                 collect_into_maps(name, gauge.value() as u64, maps);
             }
-            MetricValue::Time { name, time } => {
-                if name.starts_with(GREPTIME_EXEC_PREFIX) {
-                    // override
-                    maps.iter_mut().for_each(|map| {
-                        map.insert(name.to_string(), time.value() as u64);
-                    });
-                }
+            MetricValue::Time { name, time } if name.starts_with(GREPTIME_EXEC_PREFIX) => {
+                // override
+                maps.iter_mut().for_each(|map| {
+                    map.insert(name.to_string(), time.value() as u64);
+                });
             }
             _ => {}
         });
diff --git a/src/servers/src/http/otlp.rs b/src/servers/src/http/otlp.rs
index 4fd2d42122..ba0007bc06 100644
--- a/src/servers/src/http/otlp.rs
+++ b/src/servers/src/http/otlp.rs
@@ -16,7 +16,7 @@ use std::sync::Arc;
 
 use axum::Extension;
 use axum::extract::State;
-use axum::http::header;
+use axum::http::{StatusCode, header};
 use axum::response::IntoResponse;
 use axum_extra::TypedHeader;
 use bytes::Bytes;
@@ -29,7 +29,7 @@ use opentelemetry_proto::tonic::collector::logs::v1::{
 };
 use opentelemetry_proto::tonic::collector::metrics::v1::ExportMetricsServiceResponse;
 use opentelemetry_proto::tonic::collector::trace::v1::{
-    ExportTraceServiceRequest, ExportTraceServiceResponse,
+    ExportTracePartialSuccess, ExportTraceServiceRequest, ExportTraceServiceResponse,
 };
 use otel_arrow_rust::proto::opentelemetry::collector::metrics::v1::ExportMetricsServiceRequest;
 use pipeline::PipelineWay;
@@ -44,7 +44,15 @@ use crate::http::extractor::{
 };
 use crate::http::header::{CONTENT_TYPE_PROTOBUF, write_cost_header_map};
 use crate::metrics::METRIC_HTTP_OPENTELEMETRY_LOGS_ELAPSED;
-use crate::query_handler::{OpenTelemetryProtocolHandlerRef, PipelineHandler};
+use crate::query_handler::{OpenTelemetryProtocolHandlerRef, PipelineHandler, TraceIngestOutcome};
+
+#[derive(Clone, prost::Message)]
+pub struct GoogleRpcStatus {
+    #[prost(int32, tag = "1")]
+    pub code: i32,
+    #[prost(string, tag = "2")]
+    pub message: String,
+}
 
 fn is_json_content_type(content_type: Option<&ContentType>) -> bool {
     match content_type {
@@ -129,7 +137,7 @@ pub async fn traces(
     Extension(mut query_ctx): Extension<QueryContext>,
     content_type: Option<TypedHeader<ContentType>>,
     bytes: Bytes,
-) -> Result<OtlpResponse<ExportTraceServiceResponse>> {
+) -> Result<OtlpTraceResponse> {
     if is_json_content_type(content_type.as_ref().map(|h| &h.0)) {
         return error::UnsupportedJsonContentTypeSnafu {}.fail();
     }
@@ -175,11 +183,14 @@ pub async fn traces(
             query_ctx,
         )
         .await
-        .map(|o| OtlpResponse {
-            resp_body: ExportTraceServiceResponse {
-                partial_success: None,
-            },
-            write_cost: o.meta.cost,
+        .map(|outcome| {
+            if outcome.accepted_spans == 0 && outcome.rejected_spans > 0 {
+                OtlpTraceResponse::Failure(outcome)
+            } else if outcome.rejected_spans > 0 || outcome.error_message.is_some() {
+                OtlpTraceResponse::PartialSuccess(outcome)
+            } else {
+                OtlpTraceResponse::FullSuccess(outcome)
+            }
         })
 }
 
@@ -255,3 +266,49 @@ impl<T: Message> IntoResponse for OtlpResponse<T> {
         (header_map, self.resp_body.encode_to_vec()).into_response()
     }
 }
+
+pub enum OtlpTraceResponse {
+    FullSuccess(TraceIngestOutcome),
+    PartialSuccess(TraceIngestOutcome),
+    Failure(TraceIngestOutcome),
+}
+
+impl IntoResponse for OtlpTraceResponse {
+    fn into_response(self) -> axum::response::Response {
+        match self {
+            OtlpTraceResponse::FullSuccess(outcome) => {
+                let mut header_map = write_cost_header_map(outcome.write_cost);
+                header_map.insert(header::CONTENT_TYPE, CONTENT_TYPE_PROTOBUF.clone());
+                let body = ExportTraceServiceResponse {
+                    partial_success: None,
+                };
+                (header_map, body.encode_to_vec()).into_response()
+            }
+            OtlpTraceResponse::PartialSuccess(outcome) => {
+                let mut header_map = write_cost_header_map(outcome.write_cost);
+                header_map.insert(header::CONTENT_TYPE, CONTENT_TYPE_PROTOBUF.clone());
+                let body = ExportTraceServiceResponse {
+                    partial_success: outcome.error_message.map(|error_message| {
+                        ExportTracePartialSuccess {
+                            rejected_spans: outcome.rejected_spans as i64,
+                            error_message,
+                        }
+                    }),
+                };
+                (header_map, body.encode_to_vec()).into_response()
+            }
+            OtlpTraceResponse::Failure(outcome) => {
+                let status = GoogleRpcStatus {
+                    code: 0,
+                    message: outcome.error_message.unwrap_or_default(),
+                };
+                (
+                    StatusCode::BAD_REQUEST,
+                    [(header::CONTENT_TYPE, CONTENT_TYPE_PROTOBUF.as_ref())],
+                    status.encode_to_vec(),
+                )
+                    .into_response()
+            }
+        }
+    }
+}
diff --git a/src/servers/src/http/prom_store.rs b/src/servers/src/http/prom_store.rs
index 58c6e0eddd..bfc072e84e 100644
--- a/src/servers/src/http/prom_store.rs
+++ b/src/servers/src/http/prom_store.rs
@@ -35,6 +35,7 @@ use snafu::prelude::*;
 use crate::error::{self, InternalSnafu, PipelineSnafu, Result};
 use crate::http::extractor::PipelineInfo;
 use crate::http::header::{GREPTIME_DB_HEADER_METRICS, write_cost_header_map};
+use crate::pending_rows_batcher::PendingRowsBatcher;
 use crate::prom_remote_write::decode::PromSeriesProcessor;
 use crate::prom_remote_write::decode_remote_write_request;
 use crate::prom_remote_write::validation::PromValidationMode;
@@ -52,6 +53,7 @@ pub struct PromStoreState {
     pub pipeline_handler: Option<PipelineHandlerRef>,
     pub prom_store_with_metric_engine: bool,
     pub prom_validation_mode: PromValidationMode,
+    pub pending_rows_batcher: Option<Arc<PendingRowsBatcher>>,
 }
 
 #[derive(Debug, Serialize, Deserialize)]
@@ -92,6 +94,7 @@ pub async fn remote_write(
         pipeline_handler,
         prom_store_with_metric_engine,
         prom_validation_mode,
+        pending_rows_batcher,
     } = state;
 
     if let Some(_vm_handshake) = params.get_vm_proto_version {
@@ -100,9 +103,11 @@ pub async fn remote_write(
 
     let db = params.db.clone().unwrap_or_default();
     query_ctx.set_channel(Channel::Prometheus);
-    if let Some(physical_table) = params.physical_table {
-        query_ctx.set_extension(PHYSICAL_TABLE_PARAM, physical_table);
-    }
+    let physical_table = params
+        .physical_table
+        .clone()
+        .unwrap_or_else(|| GREPTIME_PHYSICAL_TABLE.to_string());
+    query_ctx.set_extension(PHYSICAL_TABLE_PARAM, physical_table.clone());
     let query_ctx = Arc::new(query_ctx);
     let _timer = crate::metrics::METRIC_HTTP_PROM_STORE_WRITE_ELAPSED
         .with_label_values(&[db.as_str()])
@@ -135,6 +140,19 @@ pub async fn remote_write(
         req.as_insert_requests()
     };
 
+    if prom_store_with_metric_engine && let Some(batcher) = pending_rows_batcher {
+        for (temp_ctx, reqs) in req.as_req_iter(query_ctx) {
+            prom_store_handler
+                .pre_write(&reqs, temp_ctx.clone())
+                .await?;
+            let rows = batcher.submit(reqs, temp_ctx).await?;
+            crate::metrics::PROM_STORE_REMOTE_WRITE_SAMPLES
+                .with_label_values(&[db.as_str()])
+                .inc_by(rows);
+        }
+        return Ok((StatusCode::NO_CONTENT, write_cost_header_map(0)).into_response());
+    }
+
     let mut cost = 0;
     for (temp_ctx, reqs) in req.as_req_iter(query_ctx) {
         let cnt: u64 = reqs
diff --git a/src/servers/src/http/prometheus.rs b/src/servers/src/http/prometheus.rs
index 488a49df99..63149948a5 100644
--- a/src/servers/src/http/prometheus.rs
+++ b/src/servers/src/http/prometheus.rs
@@ -65,9 +65,8 @@ use store_api::metric_engine_consts::{
 
 pub use super::result::prometheus_resp::PrometheusJsonResponse;
 use crate::error::{
-    CatalogSnafu, CollectRecordbatchSnafu, ConvertScalarValueSnafu, DataFusionSnafu, Error,
-    InvalidQuerySnafu, NotSupportedSnafu, ParseTimestampSnafu, Result, TableNotFoundSnafu,
-    UnexpectedResultSnafu,
+    CollectRecordbatchSnafu, ConvertScalarValueSnafu, DataFusionSnafu, Error, InvalidQuerySnafu,
+    NotSupportedSnafu, ParseTimestampSnafu, Result, TableNotFoundSnafu, UnexpectedResultSnafu,
 };
 use crate::http::header::collect_plan_metrics;
 use crate::prom_store::{FIELD_NAME_LABEL, METRIC_NAME_LABEL, is_database_selection_label};
@@ -662,8 +661,7 @@ async fn retrieve_series_from_query_result(
             table_name,
             Some(query_ctx),
         )
-        .await
-        .context(CatalogSnafu)?
+        .await?
         .with_context(|| TableNotFoundSnafu {
             catalog: query_ctx.current_catalog(),
             schema: query_ctx.current_schema(),
@@ -1118,17 +1116,17 @@ fn collect_metric_names(expr: &PromqlExpr, metric_names: &mut HashSet<String>) {
     match expr {
         PromqlExpr::Aggregate(AggregateExpr { modifier, expr, .. }) => {
             match modifier {
-                Some(LabelModifier::Include(labels)) => {
-                    if !labels.labels.contains(&METRIC_NAME.to_string()) {
-                        metric_names.clear();
-                        return;
-                    }
+                Some(LabelModifier::Include(labels))
+                    if !labels.labels.contains(&METRIC_NAME.to_string()) =>
+                {
+                    metric_names.clear();
+                    return;
                 }
-                Some(LabelModifier::Exclude(labels)) => {
-                    if labels.labels.contains(&METRIC_NAME.to_string()) {
-                        metric_names.clear();
-                        return;
-                    }
+                Some(LabelModifier::Exclude(labels))
+                    if labels.labels.contains(&METRIC_NAME.to_string()) =>
+                {
+                    metric_names.clear();
+                    return;
                 }
                 _ => {}
             }
@@ -1440,7 +1438,7 @@ async fn retrieve_table_names(
         });
 
     while let Some(table) = tables_stream.next().await {
-        let table = table.context(CatalogSnafu)?;
+        let table = table?;
         if !table
             .table_info()
             .meta
@@ -1497,7 +1495,7 @@ async fn retrieve_field_names(
             .next()
             .await
         {
-            let table = table.context(CatalogSnafu)?;
+            let table = table?;
             for column in table.field_columns() {
                 field_columns.insert(column.name);
             }
@@ -1508,8 +1506,7 @@ async fn retrieve_field_names(
     for table_name in matches {
         let table = manager
             .table(catalog, &schema, &table_name, Some(query_ctx))
-            .await
-            .context(CatalogSnafu)?
+            .await?
             .with_context(|| TableNotFoundSnafu {
                 catalog: catalog.to_string(),
                 schema: schema.clone(),
@@ -1533,8 +1530,7 @@ async fn retrieve_schema_names(
 
     let candidate_schemas = catalog_manager
         .schema_names(catalog, Some(query_ctx))
-        .await
-        .context(CatalogSnafu)?;
+        .await?;
 
     for schema in candidate_schemas {
         let mut found = true;
@@ -1542,8 +1538,7 @@ async fn retrieve_schema_names(
             if let Some(table_name) = retrieve_metric_name_from_promql(match_item) {
                 let exists = catalog_manager
                     .table_exists(catalog, &schema, &table_name, Some(query_ctx))
-                    .await
-                    .context(CatalogSnafu)?;
+                    .await?;
                 if !exists {
                     found = false;
                     break;
diff --git a/src/servers/src/http/result/arrow_result.rs b/src/servers/src/http/result/arrow_result.rs
index d583c3a590..90f8513827 100644
--- a/src/servers/src/http/result/arrow_result.rs
+++ b/src/servers/src/http/result/arrow_result.rs
@@ -17,7 +17,7 @@ use std::sync::Arc;
 
 use arrow::datatypes::Schema;
 use arrow_ipc::CompressionType;
-use arrow_ipc::writer::{FileWriter, IpcWriteOptions};
+use arrow_ipc::writer::{IpcWriteOptions, StreamWriter};
 use axum::http::{HeaderValue, header};
 use axum::response::{IntoResponse, Response};
 use common_error::status_code::StatusCode;
@@ -48,7 +48,7 @@ async fn write_arrow_bytes(
         let options = IpcWriteOptions::default()
             .try_with_compression(compression)
             .context(error::ArrowSnafu)?;
-        let mut writer = FileWriter::try_new_with_options(&mut bytes, schema, options)
+        let mut writer = StreamWriter::try_new_with_options(&mut bytes, schema, options)
             .context(error::ArrowSnafu)?;
 
         while let Some(rb) = recordbatches.next().await {
@@ -164,7 +164,7 @@ impl IntoResponse for ArrowResponse {
 mod test {
     use std::io::Cursor;
 
-    use arrow_ipc::reader::FileReader;
+    use arrow_ipc::reader::StreamReader;
     use arrow_schema::DataType;
     use common_recordbatch::{RecordBatch, RecordBatches};
     use datatypes::prelude::*;
@@ -200,8 +200,8 @@ mod test {
             match http_resp {
                 HttpResponse::Arrow(resp) => {
                     let output = resp.data;
-                    let mut reader =
-                        FileReader::try_new(Cursor::new(output), None).expect("Arrow reader error");
+                    let mut reader = StreamReader::try_new(Cursor::new(output), None)
+                        .expect("Arrow reader error");
                     let schema = reader.schema();
                     assert_eq!(schema.fields[0].name(), "numbers");
                     assert_eq!(schema.fields[0].data_type(), &DataType::UInt32);
diff --git a/src/servers/src/http/result/error_result.rs b/src/servers/src/http/result/error_result.rs
index 7b70066b68..9bd6e1a7a3 100644
--- a/src/servers/src/http/result/error_result.rs
+++ b/src/servers/src/http/result/error_result.rs
@@ -32,17 +32,24 @@ pub struct ErrorResponse {
 impl ErrorResponse {
     pub fn from_error(error: impl ErrorExt) -> Self {
         let code = error.status_code();
-
         if code.should_log_error() {
             error!(error; "Failed to handle HTTP request");
         } else {
             debug!("Failed to handle HTTP request, err: {:?}", error);
         }
-
-        Self::from_error_message(code, error.output_msg())
+        ErrorResponse {
+            code: code as u32,
+            error: error.output_msg(),
+            execution_time_ms: 0,
+        }
     }
 
     pub fn from_error_message(code: StatusCode, msg: String) -> Self {
+        if code.should_log_error() {
+            error!("Failed to handle HTTP request: {}", msg);
+        } else {
+            debug!("Failed to handle HTTP request: {}", msg);
+        }
         ErrorResponse {
             code: code as u32,
             error: msg,
diff --git a/src/servers/src/http/result/greptime_manage_resp.rs b/src/servers/src/http/result/greptime_manage_resp.rs
index 3f7f3c6eec..2b3a5d455c 100644
--- a/src/servers/src/http/result/greptime_manage_resp.rs
+++ b/src/servers/src/http/result/greptime_manage_resp.rs
@@ -62,6 +62,25 @@ impl GreptimedbManageResponse {
         }
     }
 
+    pub fn from_dashboard(name: String, execution_time_ms: u64) -> Self {
+        GreptimedbManageResponse {
+            manage_result: ManageResult::Dashboards {
+                dashboards: vec![DashboardOutput {
+                    name,
+                    definition: String::new(),
+                }],
+            },
+            execution_time_ms,
+        }
+    }
+
+    pub fn from_dashboards(dashboards: Vec<DashboardOutput>, execution_time_ms: u64) -> Self {
+        GreptimedbManageResponse {
+            manage_result: ManageResult::Dashboards { dashboards },
+            execution_time_ms,
+        }
+    }
+
     pub fn with_execution_time(mut self, execution_time: u64) -> Self {
         self.execution_time_ms = execution_time;
         self
@@ -77,6 +96,7 @@ impl GreptimedbManageResponse {
 pub enum ManageResult {
     Pipelines { pipelines: Vec<PipelineOutput> },
     Sql { sql: SqlOutput },
+    Dashboards { dashboards: Vec<DashboardOutput> },
 }
 
 #[derive(Serialize, Deserialize, Debug)]
@@ -87,6 +107,13 @@ pub struct PipelineOutput {
     pipeline: Option<String>,
 }
 
+#[derive(Serialize, Deserialize, Debug)]
+pub struct DashboardOutput {
+    pub name: String,
+    #[serde(skip_serializing_if = "String::is_empty")]
+    pub definition: String,
+}
+
 #[derive(Serialize, Deserialize, Debug)]
 pub struct SqlOutput {
     pub(crate) sql: String,
diff --git a/src/servers/src/http/result/prometheus_resp.rs b/src/servers/src/http/result/prometheus_resp.rs
index 9ecbe671b4..6f2b115686 100644
--- a/src/servers/src/http/result/prometheus_resp.rs
+++ b/src/servers/src/http/result/prometheus_resp.rs
@@ -204,10 +204,10 @@ impl PrometheusJsonResponse {
 
         for (i, column) in batches.schema().column_schemas().iter().enumerate() {
             match column.data_type {
-                ConcreteDataType::Timestamp(datatypes::types::TimestampType::Millisecond(_)) => {
-                    if timestamp_column_index.is_none() {
-                        timestamp_column_index = Some(i);
-                    }
+                ConcreteDataType::Timestamp(datatypes::types::TimestampType::Millisecond(_))
+                    if timestamp_column_index.is_none() =>
+                {
+                    timestamp_column_index = Some(i);
                 }
                 // Treat all value types as field
                 ConcreteDataType::Float32(_)
@@ -219,10 +219,10 @@ impl PrometheusJsonResponse {
                 | ConcreteDataType::UInt8(_)
                 | ConcreteDataType::UInt16(_)
                 | ConcreteDataType::UInt32(_)
-                | ConcreteDataType::UInt64(_) => {
-                    if first_field_column_index.is_none() {
-                        first_field_column_index = Some(i);
-                    }
+                | ConcreteDataType::UInt64(_)
+                    if first_field_column_index.is_none() =>
+                {
+                    first_field_column_index = Some(i);
                 }
                 ConcreteDataType::String(_) => {
                     tag_column_indices.push(i);
diff --git a/src/servers/src/lib.rs b/src/servers/src/lib.rs
index 9ee7395691..55d3b65fed 100644
--- a/src/servers/src/lib.rs
+++ b/src/servers/src/lib.rs
@@ -12,14 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
 #![feature(try_blocks)]
 #![feature(exclusive_wrapper)]
-#![feature(if_let_guard)]
-#![feature(box_patterns)]
 
 use datafusion_expr::LogicalPlan;
-use datatypes::schema::Schema;
 use sql::statements::statement::Statement;
 // Re-export for use in add_service! macro
 #[doc(hidden)]
@@ -41,9 +37,11 @@ pub mod mysql;
 pub mod opentsdb;
 pub mod otel_arrow;
 pub mod otlp;
+pub mod pending_rows_batcher;
 mod pipeline;
 pub mod postgres;
 pub mod prom_remote_write;
+pub(crate) mod prom_row_builder;
 pub mod prom_store;
 pub mod prometheus;
 pub mod prometheus_handler;
@@ -55,23 +53,28 @@ mod row_writer;
 pub mod server;
 pub mod tls;
 
-/// Cached SQL and logical plan for database interfaces
+/// Cached sql plan or statement for database interfaces
 #[derive(Clone, Debug)]
-pub struct SqlPlan {
-    query: String,
-    // Store the parsed statement to determine if it is a query and whether to track it.
-    statement: Option<Statement>,
-    plan: Option<LogicalPlan>,
-    schema: Option<Schema>,
+pub enum SqlPlan {
+    /// Empty Query
+    Empty,
+    /// Hardcoded SQL shortcuts
+    Shortcut(String),
+    /// Datafusion parsed execution plan with the original statement
+    Plan(LogicalPlan, Statement),
+    /// Parsed statement when execution is not managed by datafusion
+    /// eg. CREATE TABLE
+    /// The String is the original query string to avoid AST round-trip issues
+    Statement(Statement, String),
 }
 
-/// Install the ring crypto provider for rustls process-wide. see:
+/// Install the default crypto provider for rustls process-wide. see:
 ///
 ///  https://docs.rs/rustls/latest/rustls/crypto/struct.CryptoProvider.html#using-the-per-process-default-cryptoprovider
 ///
 /// for more information.
-pub fn install_ring_crypto_provider() -> Result<(), String> {
-    rustls::crypto::CryptoProvider::install_default(rustls::crypto::ring::default_provider())
+pub fn install_default_crypto_provider() -> Result<(), String> {
+    rustls::crypto::CryptoProvider::install_default(rustls::crypto::aws_lc_rs::default_provider())
         .map_err(|ret| {
             format!(
                 "CryptoProvider already installed as: {:?}, but providing {:?}",
diff --git a/src/servers/src/metrics.rs b/src/servers/src/metrics.rs
index 25a900ed3d..6f151db539 100644
--- a/src/servers/src/metrics.rs
+++ b/src/servers/src/metrics.rs
@@ -121,13 +121,69 @@ lazy_static! {
     /// Duration to convert prometheus write request to gRPC request.
     pub static ref METRIC_HTTP_PROM_STORE_CONVERT_ELAPSED: Histogram = METRIC_HTTP_PROM_STORE_CODEC_ELAPSED
         .with_label_values(&["convert"]);
-        /// The samples count of Prometheus remote write.
+    /// The samples count of Prometheus remote write.
     pub static ref PROM_STORE_REMOTE_WRITE_SAMPLES: IntCounterVec = register_int_counter_vec!(
         "greptime_servers_prometheus_remote_write_samples",
         "frontend prometheus remote write samples",
         &[METRIC_DB_LABEL]
     )
     .unwrap();
+    pub static ref PENDING_BATCHES: IntGauge = register_int_gauge!(
+        "greptime_prom_store_pending_batches",
+        "Number of pending batches waiting to be flushed"
+    )
+    .unwrap();
+    pub static ref PENDING_ROWS: IntGauge = register_int_gauge!(
+        "greptime_prom_store_pending_rows",
+        "Number of pending rows waiting to be flushed"
+    )
+    .unwrap();
+    pub static ref PENDING_WORKERS: IntGauge = register_int_gauge!(
+        "greptime_prom_store_pending_workers",
+        "Number of active pending rows batch workers"
+    )
+    .unwrap();
+    pub static ref FLUSH_TOTAL: IntCounter = register_int_counter!(
+        "greptime_prom_store_flush_total",
+        "Total number of batch flushes"
+    )
+    .unwrap();
+    pub static ref FLUSH_ROWS: Histogram = register_histogram!(
+        "greptime_prom_store_flush_rows",
+        "Number of rows per flush",
+        vec![100.0, 1000.0, 10000.0, 50000.0, 100000.0, 500000.0]
+    )
+    .unwrap();
+    pub static ref FLUSH_ELAPSED: Histogram = register_histogram!(
+        "greptime_prom_store_flush_elapsed",
+        "Elapsed time of pending rows batch flush in seconds",
+        vec![0.005, 0.01, 0.05, 0.1, 0.5, 1.0, 5.0, 10.0, 60.0, 300.0]
+    )
+    .unwrap();
+    pub static ref FLUSH_DROPPED_ROWS: IntCounter = register_int_counter!(
+        "greptime_pending_rows_flush_dropped_rows",
+        "Total rows dropped due to pending rows flush failures"
+    )
+    .unwrap();
+    pub static ref FLUSH_FAILURES: IntCounter = register_int_counter!(
+        "greptime_pending_rows_flush_failures",
+        "Total pending rows flush failures"
+    )
+    .unwrap();
+    pub static ref PENDING_ROWS_BATCH_INGEST_STAGE_ELAPSED: HistogramVec = register_histogram_vec!(
+        "greptime_prom_store_pending_rows_batch_ingest_stage_elapsed",
+        "Elapsed time of pending rows batch ingestion stages in seconds",
+        &["stage"],
+        vec![0.0005, 0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1.0, 5.0, 10.0, 60.0]
+    )
+    .unwrap();
+    pub static ref PENDING_ROWS_BATCH_FLUSH_STAGE_ELAPSED: HistogramVec = register_histogram_vec!(
+        "greptime_prom_store_pending_rows_batch_flush_stage_elapsed",
+        "Elapsed time of pending rows batch flush stages in seconds",
+        &["stage"],
+        vec![0.0005, 0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1.0, 5.0, 10.0, 60.0]
+    )
+    .unwrap();
     /// Http prometheus read duration per database.
     pub static ref METRIC_HTTP_PROM_STORE_READ_ELAPSED: HistogramVec = register_histogram_vec!(
         "greptime_servers_http_prometheus_read_elapsed",
@@ -312,7 +368,7 @@ lazy_static! {
         "maximum bytes allowed for all concurrent request bodies and messages"
     ).unwrap();
 
-    /// Total number of rejected requests due to memory exhaustion.
+    /// Total number of requests rejected due to memory exhaustion.
     pub static ref REQUEST_MEMORY_REJECTED: IntCounterVec = register_int_counter_vec!(
         "greptime_servers_request_memory_rejected_total",
         "number of requests rejected due to memory limit",
diff --git a/src/servers/src/mysql/federated.rs b/src/servers/src/mysql/federated.rs
index 92085f1e58..fa3aec144c 100644
--- a/src/servers/src/mysql/federated.rs
+++ b/src/servers/src/mysql/federated.rs
@@ -21,6 +21,7 @@ use std::sync::Arc;
 use common_query::Output;
 use common_recordbatch::RecordBatches;
 use common_time::timezone::system_timezone_name;
+use common_version;
 use datatypes::prelude::ConcreteDataType;
 use datatypes::schema::{ColumnSchema, Schema};
 use datatypes::vectors::StringVector;
@@ -119,7 +120,7 @@ static VAR_VALUES: Lazy<HashMap<&str, &str>> = Lazy::new(|| {
         ("interactive_timeout", "31536000"),
         ("wait_timeout", "31536000"),
         ("net_write_timeout", "31536000"),
-        ("version_comment", "Greptime"),
+        ("version_comment", common_version::product_name()),
     ])
 });
 
@@ -380,7 +381,7 @@ mod test {
 +-------------------+
 | @@version_comment |
 +-------------------+
-| Greptime          |
+| GreptimeDB        |
 +-------------------+";
         test(query, expected);
 
diff --git a/src/servers/src/mysql/handler.rs b/src/servers/src/mysql/handler.rs
index dd67012a52..3a80593c63 100644
--- a/src/servers/src/mysql/handler.rs
+++ b/src/servers/src/mysql/handler.rs
@@ -28,6 +28,7 @@ use common_telemetry::{debug, error, tracing, warn};
 use datafusion_common::ParamValues;
 use datafusion_expr::LogicalPlan;
 use datatypes::prelude::ConcreteDataType;
+use datatypes::schema::Schema;
 use itertools::Itertools;
 use opensrv_mysql::{
     AsyncMysqlShim, Column, ErrorKind, InitWriter, ParamParser, ParamValue, QueryResultWriter,
@@ -138,23 +139,6 @@ impl MysqlInstanceShim {
         }
     }
 
-    /// Execute the logical plan and return the output
-    async fn do_exec_plan(
-        &self,
-        query: &str,
-        stmt: Option<Statement>,
-        plan: LogicalPlan,
-        query_ctx: QueryContextRef,
-    ) -> Result<Output> {
-        if let Some(output) =
-            crate::mysql::federated::check(query, query_ctx.clone(), self.session.clone())
-        {
-            Ok(output)
-        } else {
-            self.query_handler.do_exec_plan(stmt, plan, query_ctx).await
-        }
-    }
-
     /// Describe the statement
     async fn do_describe(
         &self,
@@ -198,6 +182,16 @@ impl MysqlInstanceShim {
         query_ctx: QueryContextRef,
         stmt_key: String,
     ) -> Result<(Vec<Column>, Vec<Column>)> {
+        if crate::mysql::federated::check(raw_query, query_ctx.clone(), self.session.clone())
+            .is_some()
+        {
+            self.save_plan(SqlPlan::Shortcut(raw_query.to_string()), stmt_key)
+                .inspect_err(|e| {
+                    error!(e; "Failed to save prepared statement");
+                })?;
+            return Ok((vec![], vec![]));
+        }
+
         let (query, param_num) = replace_placeholders(raw_query);
 
         let statement = validate_query(raw_query).await?;
@@ -209,15 +203,7 @@ impl MysqlInstanceShim {
         let describe_result = self
             .do_describe(statement.clone(), query_ctx.clone())
             .await?;
-        let (plan, schema) = if let Some(DescribeResult {
-            logical_plan,
-            schema,
-        }) = describe_result
-        {
-            (Some(logical_plan), Some(schema))
-        } else {
-            (None, None)
-        };
+        let plan = describe_result.map(|DescribeResult { logical_plan }| logical_plan);
 
         let params = if let Some(plan) = &plan {
             let param_types = DfLogicalPlanner::get_inferred_parameter_types(plan)
@@ -230,49 +216,41 @@ impl MysqlInstanceShim {
             dummy_params(param_num)?
         };
 
-        let columns = schema
-            .as_ref()
-            .map(|schema| {
-                schema
-                    .column_schemas()
-                    .iter()
-                    .map(|column_schema| {
-                        create_mysql_column(&column_schema.data_type, &column_schema.name)
-                    })
-                    .collect::<Result<Vec<_>>>()
-            })
-            .transpose()?
-            .unwrap_or_default();
+        let columns =
+            plan.as_ref()
+                .map(|plan| {
+                    let schema: Schema = plan.schema().clone().try_into().map_err(
+                        |e: datatypes::error::Error| {
+                            error::InternalSnafu {
+                                err_msg: e.to_string(),
+                            }
+                            .build()
+                        },
+                    )?;
+                    schema
+                        .column_schemas()
+                        .iter()
+                        .map(|column_schema| {
+                            create_mysql_column(&column_schema.data_type, &column_schema.name)
+                        })
+                        .collect::<Result<Vec<_>>>()
+                })
+                .transpose()?
+                .unwrap_or_default();
 
-        // DataFusion may optimize the plan so that some parameters are not used.
-        if params.len() != param_num - 1 {
-            self.save_plan(
-                SqlPlan {
-                    query: query.clone(),
-                    statement: Some(statement),
-                    plan: None,
-                    schema: None,
-                },
-                stmt_key,
-            )
-            .map_err(|e| {
-                error!(e; "Failed to save prepared statement");
-                e
-            })?;
-        } else {
-            self.save_plan(
-                SqlPlan {
-                    query: query.clone(),
-                    statement: Some(statement),
-                    plan,
-                    schema,
-                },
-                stmt_key,
-            )
-            .map_err(|e| {
-                error!(e; "Failed to save prepared statement");
-                e
-            })?;
+        match plan {
+            Some(plan) if params.len() == param_num - 1 => {
+                self.save_plan(SqlPlan::Plan(plan, statement), stmt_key)
+                    .inspect_err(|e| {
+                        error!(e; "Failed to save prepared statement");
+                    })?;
+            }
+            _ => {
+                self.save_plan(SqlPlan::Statement(statement, query), stmt_key)
+                    .inspect_err(|e| {
+                        error!(e; "Failed to save prepared statement");
+                    })?;
+            }
         }
 
         Ok((params, columns))
@@ -291,8 +269,8 @@ impl MysqlInstanceShim {
             Some(sql_plan) => sql_plan,
         };
 
-        let outputs = match sql_plan.plan {
-            Some(plan) => {
+        let outputs = match sql_plan {
+            SqlPlan::Plan(plan, stmt) => {
                 let param_types = DfLogicalPlanner::get_inferred_parameter_types(&plan)
                     .context(InferParameterTypesSnafu)?
                     .into_iter()
@@ -306,7 +284,7 @@ impl MysqlInstanceShim {
                     .fail();
                 }
 
-                let plan = match params {
+                let replaced_plan = match params {
                     Params::ProtocolParams(params) => {
                         replace_params_with_values(&plan, param_types, &params)
                     }
@@ -315,18 +293,26 @@ impl MysqlInstanceShim {
                     }
                 }?;
 
-                debug!("Mysql execute prepared plan: {}", plan.display_indent());
+                debug!(
+                    "Mysql execute prepared plan: {}",
+                    replaced_plan.display_indent()
+                );
                 vec![
-                    self.do_exec_plan(
-                        &sql_plan.query,
-                        sql_plan.statement.clone(),
-                        plan,
-                        query_ctx.clone(),
-                    )
-                    .await,
+                    self.query_handler
+                        .do_exec_plan(replaced_plan, Some(stmt), query_ctx.clone())
+                        .await,
                 ]
             }
-            None => {
+            SqlPlan::Shortcut(query) => {
+                if let Some(output) =
+                    crate::mysql::federated::check(&query, query_ctx.clone(), self.session.clone())
+                {
+                    vec![Ok(output)]
+                } else {
+                    self.do_query(&query, query_ctx.clone()).await
+                }
+            }
+            SqlPlan::Statement(_stmt, query) => {
                 let param_strs = match params {
                     Params::ProtocolParams(params) => {
                         params.iter().map(convert_param_value_to_string).collect()
@@ -335,12 +321,15 @@ impl MysqlInstanceShim {
                 };
                 debug!(
                     "do_execute Replacing with Params: {:?}, Original Query: {}",
-                    param_strs, sql_plan.query
+                    param_strs, query
                 );
-                let query = replace_params(param_strs, sql_plan.query);
+                let query = replace_params(param_strs, query);
                 debug!("Mysql execute replaced query: {}", query);
                 self.do_query(&query, query_ctx.clone()).await
             }
+            _ => {
+                return error::PrepareStatementNotFoundSnafu { name: stmt_key }.fail();
+            }
         };
 
         Ok(outputs)
@@ -455,9 +444,17 @@ impl<W: AsyncWrite + Send + Sync + Unpin> AsyncMysqlShim<W> for MysqlInstanceShi
         let query_ctx = self.session.new_query_context();
         let stmt_id = self.prepared_stmts_counter.fetch_add(1, Ordering::Relaxed);
         let stmt_key = uuid::Uuid::from_u128(stmt_id as u128).to_string();
-        let (params, columns) = self
+        let (params, columns) = match self
             .do_prepare(raw_query, query_ctx.clone(), stmt_key)
-            .await?;
+            .await
+        {
+            Ok(x) => x,
+            Err(e) => {
+                let (kind, msg) = handle_err(e, query_ctx.clone());
+                w.error(kind, msg.as_bytes()).await?;
+                return Ok(());
+            }
+        };
         debug!("on_prepare: Params: {:?}, Columns: {:?}", params, columns);
         w.reply(stmt_id, &params, &columns).await?;
         crate::metrics::METRIC_MYSQL_PREPARED_COUNT
@@ -674,10 +671,8 @@ fn convert_param_value_to_string(param: &ParamValue) -> String {
 
 fn replace_params(params: Vec<String>, query: String) -> String {
     let mut query = query;
-    let mut index = 1;
-    for param in params {
+    for (index, param) in (1..).zip(params) {
         query = query.replace(&format_placeholder(index), &param);
-        index += 1;
     }
     query
 }
@@ -796,3 +791,152 @@ fn prepared_params(param_types: &HashMap<String, Option<ConcreteDataType>>) -> R
 
     Ok(params)
 }
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use async_trait::async_trait;
+    use common_query::Output;
+    use datafusion_expr::LogicalPlan;
+    use query::parser::PromQuery;
+    use query::query_engine::DescribeResult;
+    use session::context::QueryContext;
+    use sql::statements::statement::Statement;
+
+    use super::*;
+    use crate::error::Result;
+    use crate::query_handler::sql::SqlQueryHandler;
+
+    struct DummyQueryHandler;
+
+    #[async_trait]
+    impl SqlQueryHandler for DummyQueryHandler {
+        async fn do_query(&self, _: &str, _: QueryContextRef) -> Vec<Result<Output>> {
+            unimplemented!()
+        }
+
+        async fn do_promql_query(&self, _: &PromQuery, _: QueryContextRef) -> Vec<Result<Output>> {
+            unimplemented!()
+        }
+
+        async fn do_exec_plan(
+            &self,
+            _: LogicalPlan,
+            _: Option<Statement>,
+            _: QueryContextRef,
+        ) -> Result<Output> {
+            unimplemented!()
+        }
+
+        async fn do_describe(
+            &self,
+            _: Statement,
+            _: QueryContextRef,
+        ) -> Result<Option<DescribeResult>> {
+            unimplemented!()
+        }
+
+        async fn is_valid_schema(&self, _: &str, _: &str) -> Result<bool> {
+            Ok(true)
+        }
+    }
+
+    fn create_shim() -> MysqlInstanceShim {
+        MysqlInstanceShim::create(
+            Arc::new(DummyQueryHandler),
+            None,
+            "127.0.0.1:3306".parse().unwrap(),
+            1,
+            1024,
+        )
+    }
+
+    #[tokio::test]
+    async fn test_prepare_federated_query() {
+        let mut shim = create_shim();
+        let query_ctx = QueryContext::arc();
+        let stmt_key = "test_federated".to_string();
+
+        let (params, columns) = shim
+            .do_prepare(
+                "SELECT @@version_comment",
+                query_ctx.clone(),
+                stmt_key.clone(),
+            )
+            .await
+            .unwrap();
+
+        assert!(params.is_empty());
+        assert!(columns.is_empty());
+
+        let plan = shim.plan(&stmt_key).unwrap();
+        assert!(matches!(plan, SqlPlan::Shortcut(q) if q == "SELECT @@version_comment"));
+    }
+
+    #[tokio::test]
+    async fn test_execute_federated_shortcut() {
+        let mut shim = create_shim();
+        let query_ctx = QueryContext::arc();
+        let stmt_key = "test_federated_exec".to_string();
+
+        shim.do_prepare(
+            "SELECT @@version_comment",
+            query_ctx.clone(),
+            stmt_key.clone(),
+        )
+        .await
+        .unwrap();
+
+        let outputs = shim
+            .do_execute(query_ctx.clone(), stmt_key, Params::CliParams(vec![]))
+            .await
+            .unwrap();
+
+        assert_eq!(outputs.len(), 1);
+        let output = outputs.into_iter().next().unwrap().unwrap();
+        let pretty = output.data.pretty_print().await;
+        assert!(pretty.contains("GreptimeDB"));
+    }
+
+    #[tokio::test]
+    async fn test_prepare_non_federated_query_not_shortcut() {
+        let mut shim = create_shim();
+        let query_ctx = QueryContext::arc();
+        let stmt_key = "test_non_federated".to_string();
+
+        let result = shim
+            .do_prepare("SET NAMES utf8", query_ctx.clone(), stmt_key.clone())
+            .await;
+
+        assert!(result.is_ok());
+        let plan = shim.plan(&stmt_key).unwrap();
+        assert!(matches!(plan, SqlPlan::Shortcut(_)));
+    }
+
+    #[tokio::test]
+    async fn test_execute_set_shortcut() {
+        let mut shim = create_shim();
+        let query_ctx = QueryContext::arc();
+        let stmt_key = "test_set_shortcut".to_string();
+
+        shim.do_prepare("SET NAMES utf8", query_ctx.clone(), stmt_key.clone())
+            .await
+            .unwrap();
+
+        let outputs = shim
+            .do_execute(query_ctx.clone(), stmt_key, Params::CliParams(vec![]))
+            .await
+            .unwrap();
+
+        assert_eq!(outputs.len(), 1);
+        let output = outputs.into_iter().next().unwrap().unwrap();
+        match output.data {
+            common_query::OutputData::RecordBatches(batches) => {
+                let total_rows: usize = batches.iter().map(|b| b.num_rows()).sum();
+                assert_eq!(total_rows, 0);
+            }
+            other => panic!("Expected RecordBatches, got {:?}", other),
+        }
+    }
+}
diff --git a/src/servers/src/mysql/writer.rs b/src/servers/src/mysql/writer.rs
index 5c9b9c7ef7..2ff2c4481c 100644
--- a/src/servers/src/mysql/writer.rs
+++ b/src/servers/src/mysql/writer.rs
@@ -47,7 +47,7 @@ use crate::metrics::*;
 
 /// Try to write multiple output to the writer if possible.
 pub async fn write_output<W: AsyncWrite + Send + Sync + Unpin>(
-    w: QueryResultWriter<'_, W>,
+    mut writer: QueryResultWriter<'_, W>,
     query_context: QueryContextRef,
     session: SessionRef,
     outputs: Vec<Result<Output>>,
@@ -56,21 +56,87 @@ pub async fn write_output<W: AsyncWrite + Send + Sync + Unpin>(
         session.add_warning(warning);
     }
 
-    let mut writer = Some(MysqlResultWriter::new(
-        w,
-        query_context.clone(),
-        session.clone(),
-    ));
-    for output in outputs {
-        let result_writer = writer.take().context(error::InternalSnafu {
-            err_msg: "Sending multiple result set is unsupported",
-        })?;
-        writer = result_writer.try_write_one(output).await?;
+    enum Response {
+        ResultSet {
+            columns: Vec<Column>,
+            stream: SendableRecordBatchStream,
+        },
+        AffectedRows(usize),
     }
 
-    if let Some(result_writer) = writer {
-        result_writer.finish().await?;
+    let mut responses = Vec::with_capacity(outputs.len());
+    for output in outputs {
+        match output {
+            Ok(x) => {
+                let output = match x.data {
+                    OutputData::Stream(stream) => either::Left(stream),
+                    OutputData::RecordBatches(record_batches) => {
+                        either::Left(record_batches.as_stream())
+                    }
+                    OutputData::AffectedRows(rows) => either::Right(rows),
+                };
+                responses.push(match output {
+                    either::Left(stream) => {
+                        let schema = stream.schema();
+                        let columns = match create_mysql_column_def(&schema) {
+                            Ok(columns) => columns,
+                            Err(e) => {
+                                MysqlResultWriter::write_query_error(
+                                    e,
+                                    writer,
+                                    query_context.clone(),
+                                )
+                                .await?;
+                                return Ok(());
+                            }
+                        };
+                        Response::ResultSet { columns, stream }
+                    }
+                    either::Right(rows) => Response::AffectedRows(rows),
+                });
+            }
+            Err(e) => {
+                MysqlResultWriter::write_query_error(e, writer, query_context.clone()).await?;
+                return Ok(());
+            }
+        }
     }
+
+    for response in &mut responses {
+        writer = match response {
+            Response::ResultSet { columns, stream } => {
+                let mut row_writer = writer.start(columns).await?;
+                while let Some(record_batch) = stream.next().await {
+                    match record_batch {
+                        Ok(record_batch) => {
+                            if let Err(e) = MysqlResultWriter::write_recordbatch(
+                                &mut row_writer,
+                                record_batch,
+                                query_context.clone(),
+                            )
+                            .await
+                            {
+                                let (kind, err) = handle_err(e, query_context);
+                                row_writer.finish_error(kind, &err.as_bytes()).await?;
+                                return Ok(());
+                            }
+                        }
+                        Err(e) => {
+                            let (kind, err) = handle_err(e, query_context);
+                            row_writer.finish_error(kind, &err.as_bytes()).await?;
+                            return Ok(());
+                        }
+                    }
+                }
+                row_writer.finish_one().await?
+            }
+            Response::AffectedRows(rows) => {
+                MysqlResultWriter::write_affected_rows(writer, *rows, &session).await?
+            }
+        }
+    }
+
+    writer.no_more_results().await?;
     Ok(())
 }
 
@@ -97,75 +163,10 @@ pub fn handle_err(e: impl ErrorExt, query_ctx: QueryContextRef) -> (ErrorKind, S
     (kind, err_msg)
 }
 
-struct QueryResult {
-    schema: SchemaRef,
-    stream: SendableRecordBatchStream,
-}
+struct MysqlResultWriter;
 
-pub struct MysqlResultWriter<'a, W: AsyncWrite + Unpin> {
-    writer: QueryResultWriter<'a, W>,
-    query_context: QueryContextRef,
-    session: SessionRef,
-}
-
-impl<'a, W: AsyncWrite + Unpin> MysqlResultWriter<'a, W> {
-    pub fn new(
-        writer: QueryResultWriter<'a, W>,
-        query_context: QueryContextRef,
-        session: SessionRef,
-    ) -> MysqlResultWriter<'a, W> {
-        MysqlResultWriter::<'a, W> {
-            writer,
-            query_context,
-            session,
-        }
-    }
-
-    /// Try to write one result set. If there are more than one result set, return `Some`.
-    pub async fn try_write_one(
-        self,
-        output: Result<Output>,
-    ) -> io::Result<Option<MysqlResultWriter<'a, W>>> {
-        // We don't support sending multiple query result because the RowWriter's lifetime is bound to
-        // a local variable.
-        match output {
-            Ok(output) => match output.data {
-                OutputData::Stream(stream) => {
-                    let query_result = QueryResult {
-                        schema: stream.schema(),
-                        stream,
-                    };
-                    Self::write_query_result(query_result, self.writer, self.query_context).await?;
-                }
-                OutputData::RecordBatches(recordbatches) => {
-                    let query_result = QueryResult {
-                        schema: recordbatches.schema(),
-                        stream: recordbatches.as_stream(),
-                    };
-                    Self::write_query_result(query_result, self.writer, self.query_context).await?;
-                }
-                OutputData::AffectedRows(rows) => {
-                    let next_writer =
-                        Self::write_affected_rows(self.writer, rows, &self.session).await?;
-                    return Ok(Some(MysqlResultWriter::new(
-                        next_writer,
-                        self.query_context,
-                        self.session,
-                    )));
-                }
-            },
-            Err(error) => Self::write_query_error(error, self.writer, self.query_context).await?,
-        }
-        Ok(None)
-    }
-
-    /// Indicate no more result set to write. No need to call this if there is only one result set.
-    pub async fn finish(self) -> Result<()> {
-        self.writer.no_more_results().await?;
-        Ok(())
-    }
-
-    async fn write_affected_rows(
+impl MysqlResultWriter {
+    async fn write_affected_rows<'a, W: AsyncWrite + Unpin>(
         w: QueryResultWriter<'a, W>,
         rows: usize,
         session: &SessionRef,
@@ -182,54 +183,12 @@ impl<'a, W: AsyncWrite + Unpin> MysqlResultWriter<'a, W> {
         Ok(next_writer)
     }
 
-    async fn write_query_result(
-        mut query_result: QueryResult,
-        writer: QueryResultWriter<'a, W>,
-        query_context: QueryContextRef,
-    ) -> io::Result<()> {
-        match create_mysql_column_def(&query_result.schema) {
-            Ok(column_def) => {
-                // The RowWriter's lifetime is bound to `column_def` thus we can't use finish_one()
-                // to return a new QueryResultWriter.
-                let mut row_writer = writer.start(&column_def).await?;
-                while let Some(record_batch) = query_result.stream.next().await {
-                    match record_batch {
-                        Ok(record_batch) => {
-                            if let Err(e) = Self::write_recordbatch(
-                                &mut row_writer,
-                                record_batch,
-                                query_context.clone(),
-                                &query_result.schema,
-                            )
-                            .await
-                            {
-                                let (kind, err) = handle_err(e, query_context);
-                                row_writer.finish_error(kind, &err.as_bytes()).await?;
-                                return Ok(());
-                            }
-                        }
-                        Err(e) => {
-                            let (kind, err) = handle_err(e, query_context);
-                            debug!("Failed to get result, kind: {:?}, err: {}", kind, err);
-                            row_writer.finish_error(kind, &err.as_bytes()).await?;
-
-                            return Ok(());
-                        }
-                    }
-                }
-                row_writer.finish().await?;
-                Ok(())
-            }
-            Err(error) => Self::write_query_error(error, writer, query_context).await,
-        }
-    }
-
-    async fn write_recordbatch(
-        row_writer: &mut RowWriter<'_, W>,
+    async fn write_recordbatch<W: AsyncWrite + Unpin>(
+        row_writer: &mut RowWriter<'_, '_, W>,
         record_batch: RecordBatch,
         query_context: QueryContextRef,
-        schema: &SchemaRef,
     ) -> Result<()> {
+        let schema = record_batch.schema.clone();
         let record_batch = record_batch.into_df_record_batch();
         for i in 0..record_batch.num_rows() {
             for (j, column) in record_batch.columns().iter().enumerate() {
@@ -358,7 +317,7 @@ impl<'a, W: AsyncWrite + Unpin> MysqlResultWriter<'a, W> {
         Ok(())
     }
 
-    async fn write_query_error(
+    async fn write_query_error<'a, W: AsyncWrite + Unpin>(
         error: impl ErrorExt,
         w: QueryResultWriter<'a, W>,
         query_context: QueryContextRef,
diff --git a/src/servers/src/otlp/trace.rs b/src/servers/src/otlp/trace.rs
index b724bb1d22..98f4441923 100644
--- a/src/servers/src/otlp/trace.rs
+++ b/src/servers/src/otlp/trace.rs
@@ -13,19 +13,22 @@
 // limitations under the License.
 
 pub mod attributes;
+pub mod coerce;
 pub mod span;
 pub mod v0;
 pub mod v1;
 
+use std::collections::HashSet;
+
 use api::v1::RowInsertRequests;
 pub use common_catalog::consts::{
     PARENT_SPAN_ID_COLUMN, SPAN_ID_COLUMN, SPAN_NAME_COLUMN, TRACE_ID_COLUMN,
 };
-use opentelemetry_proto::tonic::collector::trace::v1::ExportTraceServiceRequest;
 use pipeline::{GreptimePipelineParams, PipelineWay};
 use session::context::QueryContextRef;
 
 use crate::error::{NotSupportedSnafu, Result};
+use crate::otlp::trace::span::TraceSpan;
 use crate::query_handler::PipelineHandlerRef;
 
 // column names
@@ -64,27 +67,58 @@ pub const SPAN_STATUS_PREFIX: &str = "STATUS_CODE_";
 pub const SPAN_STATUS_UNSET: &str = "STATUS_CODE_UNSET";
 pub const SPAN_STATUS_ERROR: &str = "STATUS_CODE_ERROR";
 
-/// Convert SpanTraces to GreptimeDB row insert requests.
-/// Returns `InsertRequests` and total number of rows to ingest
-pub fn to_grpc_insert_requests(
-    request: ExportTraceServiceRequest,
-    pipeline: PipelineWay,
-    pipeline_params: GreptimePipelineParams,
-    table_name: String,
+/// Deduplicated auxiliary trace entities derived from successfully ingested
+/// spans.
+///
+/// The main trace table is written first. Once a span is confirmed accepted, we
+/// record the service and operation tuples here so the auxiliary tables can be
+/// updated separately without affecting span acceptance accounting.
+#[derive(Debug, Default)]
+pub struct TraceAuxData {
+    pub services: HashSet<String>,
+    pub operations: HashSet<(String, String, String)>,
+}
+
+impl TraceAuxData {
+    /// Records the auxiliary service and operation rows implied by one accepted
+    /// span.
+    pub fn observe_span(&mut self, span: &TraceSpan) {
+        if let Some(service_name) = &span.service_name {
+            self.services.insert(service_name.clone());
+            self.operations.insert((
+                service_name.clone(),
+                span.span_name.clone(),
+                span.span_kind.clone(),
+            ));
+        }
+    }
+
+    /// Returns true when no auxiliary table updates are needed.
+    pub fn is_empty(&self) -> bool {
+        self.services.is_empty() && self.operations.is_empty()
+    }
+}
+
+/// Convert a subset of trace spans to GreptimeDB row insert requests.
+pub fn to_grpc_insert_requests_from_spans(
+    spans: &[TraceSpan],
+    pipeline: &PipelineWay,
+    pipeline_params: &GreptimePipelineParams,
+    table_name: &str,
     query_ctx: &QueryContextRef,
     pipeline_handler: PipelineHandlerRef,
 ) -> Result<(RowInsertRequests, usize)> {
     match pipeline {
-        PipelineWay::OtlpTraceDirectV0 => v0::v0_to_grpc_insert_requests(
-            request,
+        PipelineWay::OtlpTraceDirectV0 => v0::v0_to_grpc_main_insert_requests(
+            spans,
             pipeline,
             pipeline_params,
             table_name,
             query_ctx,
             pipeline_handler,
         ),
-        PipelineWay::OtlpTraceDirectV1 => v1::v1_to_grpc_insert_requests(
-            request,
+        PipelineWay::OtlpTraceDirectV1 => v1::v1_to_grpc_main_insert_requests(
+            spans,
             pipeline,
             pipeline_params,
             table_name,
@@ -97,3 +131,23 @@ pub fn to_grpc_insert_requests(
         .fail(),
     }
 }
+
+/// Build insert requests for the auxiliary trace tables derived from accepted
+/// spans.
+///
+/// "Aux" here refers to the trace service and trace operation tables, not the
+/// main trace span table itself.
+pub fn to_grpc_insert_requests_for_aux_tables(
+    aux_data: TraceAuxData,
+    pipeline: &PipelineWay,
+    table_name: &str,
+) -> Result<(RowInsertRequests, usize)> {
+    match pipeline {
+        PipelineWay::OtlpTraceDirectV0 => v0::build_aux_table_requests(aux_data, table_name),
+        PipelineWay::OtlpTraceDirectV1 => v1::build_aux_table_requests(aux_data, table_name),
+        _ => NotSupportedSnafu {
+            feat: "Unsupported pipeline for trace",
+        }
+        .fail(),
+    }
+}
diff --git a/src/servers/src/otlp/trace/coerce.rs b/src/servers/src/otlp/trace/coerce.rs
new file mode 100644
index 0000000000..febec0fda9
--- /dev/null
+++ b/src/servers/src/otlp/trace/coerce.rs
@@ -0,0 +1,343 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use api::v1::ColumnDataType;
+use api::v1::value::ValueData;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum TraceCoerceError {
+    Unsupported,
+}
+
+// For now we support the following coercions:
+// - Int64 to Float64
+// - Int64 to String
+// - Float64 to String
+// - Boolean to String
+// The following coercions are supported with parse, which could fail:
+// If fails, we will return TraceCoerceError::Unsupported.
+// - String to Int64
+// - String to Float64
+// - String to Boolean
+pub fn is_supported_trace_coercion(
+    request_type: ColumnDataType,
+    target_type: ColumnDataType,
+) -> bool {
+    matches!(
+        (request_type, target_type),
+        (ColumnDataType::Int64, ColumnDataType::Float64)
+            | (ColumnDataType::Int64, ColumnDataType::String)
+            | (ColumnDataType::Float64, ColumnDataType::String)
+            | (ColumnDataType::Boolean, ColumnDataType::String)
+            | (ColumnDataType::String, ColumnDataType::Int64)
+            | (ColumnDataType::String, ColumnDataType::Float64)
+            | (ColumnDataType::String, ColumnDataType::Boolean)
+    )
+}
+
+pub fn coerce_value_data(
+    value: &Option<ValueData>,
+    target: ColumnDataType,
+    request_type: ColumnDataType,
+) -> Result<Option<ValueData>, TraceCoerceError> {
+    let Some(v) = value else {
+        return Ok(None);
+    };
+
+    let Some(value) = coerce_non_null_value(target, request_type, v) else {
+        return Err(TraceCoerceError::Unsupported);
+    };
+    Ok(Some(value))
+}
+
+pub fn coerce_non_null_value(
+    target: ColumnDataType,
+    request_type: ColumnDataType,
+    value: &ValueData,
+) -> Option<ValueData> {
+    match (request_type, target, value) {
+        (ColumnDataType::Int64, ColumnDataType::Float64, ValueData::I64Value(n)) => {
+            Some(ValueData::F64Value(*n as f64))
+        }
+        (ColumnDataType::Int64, ColumnDataType::String, ValueData::I64Value(n)) => {
+            Some(ValueData::StringValue(n.to_string()))
+        }
+        (ColumnDataType::Float64, ColumnDataType::String, ValueData::F64Value(n)) => {
+            Some(ValueData::StringValue(n.to_string()))
+        }
+        (ColumnDataType::Boolean, ColumnDataType::String, ValueData::BoolValue(b)) => {
+            Some(ValueData::StringValue(b.to_string()))
+        }
+        (ColumnDataType::String, ColumnDataType::Int64, ValueData::StringValue(s)) => {
+            s.parse::<i64>().ok().map(ValueData::I64Value)
+        }
+        (ColumnDataType::String, ColumnDataType::Float64, ValueData::StringValue(s)) => {
+            s.parse::<f64>().ok().map(ValueData::F64Value)
+        }
+        (ColumnDataType::String, ColumnDataType::Boolean, ValueData::StringValue(s)) => {
+            s.parse::<bool>().ok().map(ValueData::BoolValue)
+        }
+        _ => None,
+    }
+}
+
+pub fn trace_value_datatype(value: &ValueData) -> Option<ColumnDataType> {
+    match value {
+        ValueData::StringValue(_) => Some(ColumnDataType::String),
+        ValueData::BoolValue(_) => Some(ColumnDataType::Boolean),
+        ValueData::I64Value(_) => Some(ColumnDataType::Int64),
+        ValueData::F64Value(_) => Some(ColumnDataType::Float64),
+        ValueData::BinaryValue(_) => Some(ColumnDataType::Binary),
+        _ => None,
+    }
+}
+
+/// Resolves the final datatype for a new trace column when there is no existing
+/// table schema to override the request-local observations.
+pub fn resolve_new_trace_column_type(
+    observed_types: impl IntoIterator<Item = ColumnDataType>,
+) -> Result<Option<ColumnDataType>, TraceCoerceError> {
+    let mut observed_types = observed_types.into_iter().collect::<Vec<_>>();
+    observed_types.dedup();
+
+    match observed_types.as_slice() {
+        [] => Ok(None),
+        [datatype] => Ok(Some(*datatype)),
+        [_, _]
+            if observed_types.contains(&ColumnDataType::String)
+                && observed_types.contains(&ColumnDataType::Boolean) =>
+        {
+            Ok(Some(ColumnDataType::Boolean))
+        }
+        [_, _]
+            if observed_types.contains(&ColumnDataType::String)
+                && observed_types.contains(&ColumnDataType::Int64) =>
+        {
+            Ok(Some(ColumnDataType::Int64))
+        }
+        [_, _]
+            if observed_types.contains(&ColumnDataType::String)
+                && observed_types.contains(&ColumnDataType::Float64) =>
+        {
+            Ok(Some(ColumnDataType::Float64))
+        }
+        [_, _]
+            if observed_types.contains(&ColumnDataType::Int64)
+                && observed_types.contains(&ColumnDataType::Float64) =>
+        {
+            Ok(Some(ColumnDataType::Float64))
+        }
+        _ => Err(TraceCoerceError::Unsupported),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_coerce_int64_to_float64() {
+        let result = coerce_value_data(
+            &Some(ValueData::I64Value(42)),
+            ColumnDataType::Float64,
+            ColumnDataType::Int64,
+        );
+        assert_eq!(result, Ok(Some(ValueData::F64Value(42.0))));
+    }
+
+    #[test]
+    fn test_coerce_string_to_int64() {
+        let result = coerce_value_data(
+            &Some(ValueData::StringValue("123".to_string())),
+            ColumnDataType::Int64,
+            ColumnDataType::String,
+        );
+        assert_eq!(result, Ok(Some(ValueData::I64Value(123))));
+    }
+
+    #[test]
+    fn test_coerce_int64_to_string() {
+        let result = coerce_value_data(
+            &Some(ValueData::I64Value(123)),
+            ColumnDataType::String,
+            ColumnDataType::Int64,
+        );
+        assert_eq!(result, Ok(Some(ValueData::StringValue("123".to_string()))));
+    }
+
+    #[test]
+    fn test_coerce_string_to_float64() {
+        let result = coerce_value_data(
+            &Some(ValueData::StringValue("1.5".to_string())),
+            ColumnDataType::Float64,
+            ColumnDataType::String,
+        );
+        assert_eq!(result, Ok(Some(ValueData::F64Value(1.5))));
+    }
+
+    #[test]
+    fn test_coerce_float64_to_string() {
+        let result = coerce_value_data(
+            &Some(ValueData::F64Value(1.5)),
+            ColumnDataType::String,
+            ColumnDataType::Float64,
+        );
+        assert_eq!(result, Ok(Some(ValueData::StringValue("1.5".to_string()))));
+    }
+
+    #[test]
+    fn test_coerce_string_to_boolean() {
+        let result = coerce_value_data(
+            &Some(ValueData::StringValue("true".to_string())),
+            ColumnDataType::Boolean,
+            ColumnDataType::String,
+        );
+        assert_eq!(result, Ok(Some(ValueData::BoolValue(true))));
+
+        let result = coerce_value_data(
+            &Some(ValueData::StringValue("false".to_string())),
+            ColumnDataType::Boolean,
+            ColumnDataType::String,
+        );
+        assert_eq!(result, Ok(Some(ValueData::BoolValue(false))));
+    }
+
+    #[test]
+    fn test_coerce_boolean_to_string() {
+        let result = coerce_value_data(
+            &Some(ValueData::BoolValue(true)),
+            ColumnDataType::String,
+            ColumnDataType::Boolean,
+        );
+        assert_eq!(result, Ok(Some(ValueData::StringValue("true".to_string()))));
+    }
+
+    #[test]
+    fn test_coerce_unparsable_string() {
+        let result = coerce_value_data(
+            &Some(ValueData::StringValue("not_a_number".to_string())),
+            ColumnDataType::Int64,
+            ColumnDataType::String,
+        );
+        assert_eq!(result, Err(TraceCoerceError::Unsupported));
+    }
+
+    #[test]
+    fn test_coerce_float64_to_int64_not_supported() {
+        let result = coerce_value_data(
+            &Some(ValueData::F64Value(1.5)),
+            ColumnDataType::Int64,
+            ColumnDataType::Float64,
+        );
+        assert_eq!(result, Err(TraceCoerceError::Unsupported));
+    }
+
+    #[test]
+    fn test_coerce_none_value() {
+        let result = coerce_value_data(&None, ColumnDataType::Float64, ColumnDataType::Int64);
+        assert_eq!(result, Ok(None));
+    }
+
+    #[test]
+    fn test_is_supported_trace_coercion() {
+        assert!(is_supported_trace_coercion(
+            ColumnDataType::Int64,
+            ColumnDataType::Float64
+        ));
+        assert!(is_supported_trace_coercion(
+            ColumnDataType::Int64,
+            ColumnDataType::String
+        ));
+        assert!(is_supported_trace_coercion(
+            ColumnDataType::Float64,
+            ColumnDataType::String
+        ));
+        assert!(is_supported_trace_coercion(
+            ColumnDataType::Boolean,
+            ColumnDataType::String
+        ));
+        assert!(is_supported_trace_coercion(
+            ColumnDataType::String,
+            ColumnDataType::Int64
+        ));
+        assert!(is_supported_trace_coercion(
+            ColumnDataType::String,
+            ColumnDataType::Float64
+        ));
+        assert!(is_supported_trace_coercion(
+            ColumnDataType::String,
+            ColumnDataType::Boolean
+        ));
+        assert!(!is_supported_trace_coercion(
+            ColumnDataType::Binary,
+            ColumnDataType::Json
+        ));
+    }
+
+    #[test]
+    fn test_trace_value_datatype() {
+        assert_eq!(
+            trace_value_datatype(&ValueData::StringValue("x".to_string())),
+            Some(ColumnDataType::String)
+        );
+        assert_eq!(
+            trace_value_datatype(&ValueData::BoolValue(true)),
+            Some(ColumnDataType::Boolean)
+        );
+        assert_eq!(
+            trace_value_datatype(&ValueData::I64Value(1)),
+            Some(ColumnDataType::Int64)
+        );
+        assert_eq!(
+            trace_value_datatype(&ValueData::F64Value(1.0)),
+            Some(ColumnDataType::Float64)
+        );
+        assert_eq!(
+            trace_value_datatype(&ValueData::BinaryValue(vec![1_u8])),
+            Some(ColumnDataType::Binary)
+        );
+    }
+
+    #[test]
+    fn test_resolve_new_trace_column_type() {
+        assert_eq!(
+            resolve_new_trace_column_type([ColumnDataType::Int64]),
+            Ok(Some(ColumnDataType::Int64))
+        );
+        assert_eq!(
+            resolve_new_trace_column_type([ColumnDataType::String, ColumnDataType::Int64]),
+            Ok(Some(ColumnDataType::Int64))
+        );
+        assert_eq!(
+            resolve_new_trace_column_type([ColumnDataType::String, ColumnDataType::Float64]),
+            Ok(Some(ColumnDataType::Float64))
+        );
+        assert_eq!(
+            resolve_new_trace_column_type([ColumnDataType::String, ColumnDataType::Boolean]),
+            Ok(Some(ColumnDataType::Boolean))
+        );
+        assert_eq!(
+            resolve_new_trace_column_type([ColumnDataType::Int64, ColumnDataType::Float64]),
+            Ok(Some(ColumnDataType::Float64))
+        );
+        assert_eq!(
+            resolve_new_trace_column_type([
+                ColumnDataType::String,
+                ColumnDataType::Int64,
+                ColumnDataType::Float64,
+            ]),
+            Err(TraceCoerceError::Unsupported)
+        );
+    }
+}
diff --git a/src/servers/src/otlp/trace/span.rs b/src/servers/src/otlp/trace/span.rs
index d96bc17277..19103240f6 100644
--- a/src/servers/src/otlp/trace/span.rs
+++ b/src/servers/src/otlp/trace/span.rs
@@ -53,6 +53,18 @@ pub struct TraceSpan {
 
 pub type TraceSpans = Vec<TraceSpan>;
 
+#[derive(Debug, Clone)]
+pub struct TraceSpanGroup {
+    pub service_name: Option<String>,
+    pub resource_attributes: Attributes,
+    pub scope_name: String,
+    pub scope_version: String,
+    pub scope_attributes: Attributes,
+    pub spans: TraceSpans,
+}
+
+pub type TraceSpanGroups = Vec<TraceSpanGroup>;
+
 #[derive(Debug, Clone, Serialize)]
 pub struct SpanLink {
     pub trace_id: String,
@@ -241,14 +253,13 @@ pub fn status_to_string(status: &Option<Status>) -> (String, String) {
 /// See
 /// <https://github.com/open-telemetry/opentelemetry-proto/blob/main/opentelemetry/proto/trace/v1/trace.proto>
 /// for data structure of OTLP traces.
-pub fn parse(request: ExportTraceServiceRequest) -> TraceSpans {
-    let span_size = request
+pub fn parse(request: ExportTraceServiceRequest) -> TraceSpanGroups {
+    let group_size = request
         .resource_spans
         .iter()
         .flat_map(|res| res.scope_spans.iter())
-        .flat_map(|scope| scope.spans.iter())
         .count();
-    let mut spans = Vec::with_capacity(span_size);
+    let mut groups = Vec::with_capacity(group_size);
     for resource_spans in request.resource_spans {
         let resource_attrs = resource_spans
             .resource
@@ -268,6 +279,7 @@ pub fn parse(request: ExportTraceServiceRequest) -> TraceSpans {
 
         for scope_spans in resource_spans.scope_spans {
             let scope = scope_spans.scope.unwrap_or_default();
+            let mut spans = Vec::with_capacity(scope_spans.spans.len());
             for span in scope_spans.spans {
                 spans.push(parse_span(
                     service_name.clone(),
@@ -276,16 +288,47 @@ pub fn parse(request: ExportTraceServiceRequest) -> TraceSpans {
                     span,
                 ));
             }
+            groups.push(TraceSpanGroup {
+                service_name: service_name.clone(),
+                resource_attributes: Attributes::from(&resource_attrs[..]),
+                scope_name: scope.name,
+                scope_version: scope.version,
+                scope_attributes: Attributes::from(scope.attributes),
+                spans,
+            });
         }
     }
-    spans
+    groups
 }
 
 #[cfg(test)]
 mod tests {
-    use opentelemetry_proto::tonic::trace::v1::Status;
+    use opentelemetry_proto::tonic::collector::trace::v1::ExportTraceServiceRequest;
+    use opentelemetry_proto::tonic::common::v1::{
+        AnyValue, InstrumentationScope, KeyValue, any_value,
+    };
+    use opentelemetry_proto::tonic::resource::v1::Resource;
+    use opentelemetry_proto::tonic::trace::v1::{ResourceSpans, ScopeSpans, Span, Status};
 
-    use crate::otlp::trace::span::{bytes_to_hex_string, status_to_string};
+    use crate::otlp::trace::KEY_SERVICE_NAME;
+    use crate::otlp::trace::span::{bytes_to_hex_string, parse, status_to_string};
+
+    fn make_kv(key: &str, value: &str) -> KeyValue {
+        KeyValue {
+            key: key.to_string(),
+            value: Some(AnyValue {
+                value: Some(any_value::Value::StringValue(value.to_string())),
+            }),
+        }
+    }
+
+    fn make_span(trace_id: u8, span_id: u8) -> Span {
+        Span {
+            trace_id: vec![trace_id; 16],
+            span_id: vec![span_id; 8],
+            ..Default::default()
+        }
+    }
 
     #[test]
     fn test_bytes_to_hex_string() {
@@ -315,4 +358,62 @@ mod tests {
             status_to_string(&Some(status)),
         );
     }
+
+    #[test]
+    fn test_parse_preserves_resource_scope_groups() {
+        let request = ExportTraceServiceRequest {
+            resource_spans: vec![
+                ResourceSpans {
+                    resource: Some(Resource {
+                        attributes: vec![make_kv(KEY_SERVICE_NAME, "svc-a")],
+                        ..Default::default()
+                    }),
+                    scope_spans: vec![
+                        ScopeSpans {
+                            scope: Some(InstrumentationScope {
+                                name: "scope-1".to_string(),
+                                ..Default::default()
+                            }),
+                            spans: vec![make_span(0x11, 0x21), make_span(0x12, 0x22)],
+                            ..Default::default()
+                        },
+                        ScopeSpans {
+                            scope: Some(InstrumentationScope {
+                                name: "scope-2".to_string(),
+                                ..Default::default()
+                            }),
+                            spans: vec![make_span(0x13, 0x23)],
+                            ..Default::default()
+                        },
+                    ],
+                    ..Default::default()
+                },
+                ResourceSpans {
+                    resource: Some(Resource {
+                        attributes: vec![make_kv(KEY_SERVICE_NAME, "svc-b")],
+                        ..Default::default()
+                    }),
+                    scope_spans: vec![ScopeSpans {
+                        scope: Some(InstrumentationScope {
+                            name: "scope-3".to_string(),
+                            ..Default::default()
+                        }),
+                        spans: vec![make_span(0x14, 0x24)],
+                        ..Default::default()
+                    }],
+                    ..Default::default()
+                },
+            ],
+        };
+
+        let groups = parse(request);
+        assert_eq!(groups.len(), 3);
+        assert_eq!(groups[0].service_name.as_deref(), Some("svc-a"));
+        assert_eq!(groups[0].scope_name, "scope-1");
+        assert_eq!(groups[0].spans.len(), 2);
+        assert_eq!(groups[1].scope_name, "scope-2");
+        assert_eq!(groups[1].spans.len(), 1);
+        assert_eq!(groups[2].service_name.as_deref(), Some("svc-b"));
+        assert_eq!(groups[2].scope_name, "scope-3");
+    }
 }
diff --git a/src/servers/src/otlp/trace/v0.rs b/src/servers/src/otlp/trace/v0.rs
index b52b406fb2..fa10dcc00f 100644
--- a/src/servers/src/otlp/trace/v0.rs
+++ b/src/servers/src/otlp/trace/v0.rs
@@ -18,16 +18,16 @@ use api::v1::value::ValueData;
 use api::v1::{ColumnDataType, RowInsertRequests};
 use common_catalog::consts::{trace_operations_table_name, trace_services_table_name};
 use common_grpc::precision::Precision;
-use opentelemetry_proto::tonic::collector::trace::v1::ExportTraceServiceRequest;
 use pipeline::{GreptimePipelineParams, PipelineWay};
 use session::context::QueryContextRef;
 
 use crate::error::Result;
-use crate::otlp::trace::span::{TraceSpan, parse};
+use crate::otlp::trace::span::TraceSpan;
 use crate::otlp::trace::{
     DURATION_NANO_COLUMN, PARENT_SPAN_ID_COLUMN, SERVICE_NAME_COLUMN, SPAN_ATTRIBUTES_COLUMN,
     SPAN_EVENTS_COLUMN, SPAN_ID_COLUMN, SPAN_KIND_COLUMN, SPAN_NAME_COLUMN, SPAN_STATUS_CODE,
     SPAN_STATUS_MESSAGE_COLUMN, TIMESTAMP_COLUMN, TRACE_ID_COLUMN, TRACE_STATE_COLUMN,
+    TraceAuxData,
 };
 use crate::otlp::utils::{make_column_data, make_string_column_data};
 use crate::query_handler::PipelineHandlerRef;
@@ -38,56 +38,52 @@ const APPROXIMATE_COLUMN_COUNT: usize = 24;
 // Use a timestamp(2100-01-01 00:00:00) as large as possible.
 const MAX_TIMESTAMP: i64 = 4102444800000000000;
 
-/// Convert SpanTraces to GreptimeDB row insert requests.
-/// Returns `InsertRequests` and total number of rows to ingest
-pub fn v0_to_grpc_insert_requests(
-    request: ExportTraceServiceRequest,
-    _pipeline: PipelineWay,
-    _pipeline_params: GreptimePipelineParams,
-    table_name: String,
+/// Converts trace spans into row insert requests for the main v0 trace table.
+///
+/// Auxiliary service and operation table writes are built separately so the
+/// caller can update them only after the main span write succeeds.
+pub fn v0_to_grpc_main_insert_requests(
+    spans: &[TraceSpan],
+    _pipeline: &PipelineWay,
+    _pipeline_params: &GreptimePipelineParams,
+    table_name: &str,
     _query_ctx: &QueryContextRef,
     _pipeline_handler: PipelineHandlerRef,
 ) -> Result<(RowInsertRequests, usize)> {
-    let spans = parse(request);
     let mut multi_table_writer = MultiTableData::default();
+    let trace_writer = build_trace_table_data(spans)?;
+    multi_table_writer.add_table_data(table_name, trace_writer);
+
+    Ok(multi_table_writer.into_row_insert_requests())
+}
+
+/// Builds the row-oriented payload for the main v0 trace table.
+pub fn build_trace_table_data(spans: &[TraceSpan]) -> Result<TableData> {
     let mut trace_writer = TableData::new(APPROXIMATE_COLUMN_COUNT, spans.len());
+    for span in spans.iter().cloned() {
+        write_span_to_row(&mut trace_writer, span)?;
+    }
+
+    Ok(trace_writer)
+}
+
+/// Builds row insert requests for the v0 trace auxiliary tables.
+pub fn build_aux_table_requests(
+    aux_data: TraceAuxData,
+    table_name: &str,
+) -> Result<(RowInsertRequests, usize)> {
+    let mut multi_table_writer = MultiTableData::default();
     let mut trace_services_writer = TableData::new(APPROXIMATE_COLUMN_COUNT, 1);
     let mut trace_operations_writer = TableData::new(APPROXIMATE_COLUMN_COUNT, 1);
 
-    let mut services = HashSet::new();
-    let mut operations = HashSet::new();
-    for span in spans {
-        if let Some(service_name) = &span.service_name {
-            // Only insert the service name if it's not already in the set.
-            if !services.contains(service_name) {
-                services.insert(service_name.clone());
-            }
-
-            // Collect operations (service_name + span_name + span_kind).
-            let operation = (
-                service_name.clone(),
-                span.span_name.clone(),
-                span.span_kind.clone(),
-            );
-            if !operations.contains(&operation) {
-                operations.insert(operation);
-            }
-        }
-        write_span_to_row(&mut trace_writer, span)?;
-    }
-    write_trace_services_to_row(&mut trace_services_writer, services)?;
-    write_trace_operations_to_row(&mut trace_operations_writer, operations)?;
+    write_trace_services_to_row(&mut trace_services_writer, aux_data.services)?;
+    write_trace_operations_to_row(&mut trace_operations_writer, aux_data.operations)?;
 
+    multi_table_writer.add_table_data(trace_services_table_name(table_name), trace_services_writer);
     multi_table_writer.add_table_data(
-        trace_services_table_name(&table_name),
-        trace_services_writer,
-    );
-    multi_table_writer.add_table_data(
-        trace_operations_table_name(&table_name),
+        trace_operations_table_name(table_name),
         trace_operations_writer,
     );
-    multi_table_writer.add_table_data(table_name, trace_writer);
-
     Ok(multi_table_writer.into_row_insert_requests())
 }
 
@@ -232,3 +228,63 @@ fn write_trace_operations_to_row(
 
     Ok(())
 }
+
+#[cfg(test)]
+mod tests {
+    use super::{build_aux_table_requests, build_trace_table_data};
+    use crate::otlp::trace::TraceAuxData;
+    use crate::otlp::trace::attributes::Attributes;
+    use crate::otlp::trace::span::{SpanEvents, SpanLinks, TraceSpan};
+
+    fn make_span(service_name: &str, trace_id: &str, span_id: &str) -> TraceSpan {
+        TraceSpan {
+            service_name: Some(service_name.to_string()),
+            trace_id: trace_id.to_string(),
+            span_id: span_id.to_string(),
+            parent_span_id: None,
+            resource_attributes: Attributes::from(vec![]),
+            scope_name: "scope".to_string(),
+            scope_version: "v1".to_string(),
+            scope_attributes: Attributes::from(vec![]),
+            trace_state: String::new(),
+            span_name: "op".to_string(),
+            span_kind: "SPAN_KIND_SERVER".to_string(),
+            span_status_code: "STATUS_CODE_UNSET".to_string(),
+            span_status_message: String::new(),
+            span_attributes: Attributes::from(vec![]),
+            span_events: SpanEvents::from(vec![]),
+            span_links: SpanLinks::from(vec![]),
+            start_in_nanosecond: 1,
+            end_in_nanosecond: 2,
+        }
+    }
+
+    #[test]
+    fn test_build_trace_table_data_from_span_subset() {
+        let spans = [
+            make_span("svc-a", "trace-a", "span-a"),
+            make_span("svc-b", "trace-b", "span-b"),
+        ];
+
+        let writer = build_trace_table_data(&spans[..1]).unwrap();
+        let (_, rows) = writer.into_schema_and_rows();
+        assert_eq!(rows.len(), 1);
+    }
+
+    #[test]
+    fn test_build_aux_table_requests_deduplicates_services_and_operations() {
+        let spans = vec![
+            make_span("svc-a", "trace-a", "span-a"),
+            make_span("svc-a", "trace-b", "span-b"),
+        ];
+        let mut aux_data = TraceAuxData::default();
+        for span in &spans {
+            aux_data.observe_span(span);
+        }
+
+        let (requests, total_rows) =
+            build_aux_table_requests(aux_data, "opentelemetry_traces").unwrap();
+        assert_eq!(requests.inserts.len(), 2);
+        assert_eq!(total_rows, 2);
+    }
+}
diff --git a/src/servers/src/otlp/trace/v1.rs b/src/servers/src/otlp/trace/v1.rs
index 86f8229769..cce6891b0d 100644
--- a/src/servers/src/otlp/trace/v1.rs
+++ b/src/servers/src/otlp/trace/v1.rs
@@ -18,19 +18,18 @@ use api::v1::value::ValueData;
 use api::v1::{ColumnDataType, RowInsertRequests, Value};
 use common_catalog::consts::{trace_operations_table_name, trace_services_table_name};
 use common_grpc::precision::Precision;
-use opentelemetry_proto::tonic::collector::trace::v1::ExportTraceServiceRequest;
 use opentelemetry_proto::tonic::common::v1::any_value::Value as OtlpValue;
 use pipeline::{GreptimePipelineParams, PipelineWay};
 use session::context::QueryContextRef;
 
 use crate::error::Result;
 use crate::otlp::trace::attributes::Attributes;
-use crate::otlp::trace::span::{TraceSpan, parse};
+use crate::otlp::trace::span::TraceSpan;
 use crate::otlp::trace::{
     DURATION_NANO_COLUMN, KEY_SERVICE_NAME, PARENT_SPAN_ID_COLUMN, SCOPE_NAME_COLUMN,
     SCOPE_VERSION_COLUMN, SERVICE_NAME_COLUMN, SPAN_EVENTS_COLUMN, SPAN_ID_COLUMN,
     SPAN_KIND_COLUMN, SPAN_NAME_COLUMN, SPAN_STATUS_CODE, SPAN_STATUS_MESSAGE_COLUMN,
-    TIMESTAMP_COLUMN, TRACE_ID_COLUMN, TRACE_STATE_COLUMN,
+    TIMESTAMP_COLUMN, TRACE_ID_COLUMN, TRACE_STATE_COLUMN, TraceAuxData,
 };
 use crate::otlp::utils::{any_value_to_jsonb, make_column_data, make_string_column_data};
 use crate::query_handler::PipelineHandlerRef;
@@ -41,64 +40,52 @@ const APPROXIMATE_COLUMN_COUNT: usize = 30;
 // Use a timestamp(2100-01-01 00:00:00) as large as possible.
 const MAX_TIMESTAMP: i64 = 4102444800000000000;
 
-/// Convert SpanTraces to GreptimeDB row insert requests.
-/// Returns `InsertRequests` and total number of rows to ingest
+/// Converts trace spans into row insert requests for the main v1 trace table.
 ///
-/// Compared with v0, this v1 implementation:
-/// 1. flattens all attribute data into columns.
-/// 2. treat `span_id` and `parent_trace_id` as fields.
-/// 3. removed `service_name` column because it's already in
-///    `resource_attributes.service_name`
-///
-/// For other compound data structures like span_links and span_events here we
-/// are still using `json` data structure.
-pub fn v1_to_grpc_insert_requests(
-    request: ExportTraceServiceRequest,
-    _pipeline: PipelineWay,
-    _pipeline_params: GreptimePipelineParams,
-    table_name: String,
+/// Auxiliary service and operation table writes are built separately so the
+/// caller can update them only after the main span write succeeds.
+pub fn v1_to_grpc_main_insert_requests(
+    spans: &[TraceSpan],
+    _pipeline: &PipelineWay,
+    _pipeline_params: &GreptimePipelineParams,
+    table_name: &str,
     _query_ctx: &QueryContextRef,
     _pipeline_handler: PipelineHandlerRef,
 ) -> Result<(RowInsertRequests, usize)> {
-    let spans = parse(request);
     let mut multi_table_writer = MultiTableData::default();
+    let trace_writer = build_trace_table_data(spans)?;
+    multi_table_writer.add_table_data(table_name, trace_writer);
+
+    Ok(multi_table_writer.into_row_insert_requests())
+}
+
+/// Builds the row-oriented payload for the main v1 trace table.
+pub fn build_trace_table_data(spans: &[TraceSpan]) -> Result<TableData> {
     let mut trace_writer = TableData::new(APPROXIMATE_COLUMN_COUNT, spans.len());
+    for span in spans.iter().cloned() {
+        write_span_to_row(&mut trace_writer, span)?;
+    }
+
+    Ok(trace_writer)
+}
+
+/// Builds row insert requests for the v1 trace auxiliary tables.
+pub fn build_aux_table_requests(
+    aux_data: TraceAuxData,
+    table_name: &str,
+) -> Result<(RowInsertRequests, usize)> {
+    let mut multi_table_writer = MultiTableData::default();
     let mut trace_services_writer = TableData::new(APPROXIMATE_COLUMN_COUNT, 1);
     let mut trace_operations_writer = TableData::new(APPROXIMATE_COLUMN_COUNT, 1);
 
-    let mut services = HashSet::new();
-    let mut operations = HashSet::new();
-    for span in spans {
-        if let Some(service_name) = &span.service_name {
-            // Only insert the service name if it's not already in the set.
-            if !services.contains(service_name) {
-                services.insert(service_name.clone());
-            }
-
-            // Only insert the operation if it's not already in the set.
-            let operation = (
-                service_name.clone(),
-                span.span_name.clone(),
-                span.span_kind.clone(),
-            );
-            if !operations.contains(&operation) {
-                operations.insert(operation);
-            }
-        }
-        write_span_to_row(&mut trace_writer, span)?;
-    }
-    write_trace_services_to_row(&mut trace_services_writer, services)?;
-    write_trace_operations_to_row(&mut trace_operations_writer, operations)?;
+    write_trace_services_to_row(&mut trace_services_writer, aux_data.services)?;
+    write_trace_operations_to_row(&mut trace_operations_writer, aux_data.operations)?;
 
+    multi_table_writer.add_table_data(trace_services_table_name(table_name), trace_services_writer);
     multi_table_writer.add_table_data(
-        trace_services_table_name(&table_name),
-        trace_services_writer,
-    );
-    multi_table_writer.add_table_data(
-        trace_operations_table_name(&table_name),
+        trace_operations_table_name(table_name),
         trace_operations_writer,
     );
-    multi_table_writer.add_table_data(table_name, trace_writer);
 
     Ok(multi_table_writer.into_row_insert_requests())
 }
@@ -230,7 +217,7 @@ fn write_trace_operations_to_row(
     Ok(())
 }
 
-fn write_attributes(
+pub(crate) fn write_attributes(
     writer: &mut TableData,
     prefix: &str,
     attributes: Attributes,
@@ -247,44 +234,40 @@ fn write_attributes(
         let key = format!("{}.{}", prefix, key_suffix);
         match attr.value.and_then(|v| v.value) {
             Some(OtlpValue::StringValue(v)) => {
-                row_writer::write_fields(
-                    writer,
-                    std::iter::once(make_string_column_data(&key, Some(v))),
+                // Keep the raw request value here. Mixed trace types are reconciled later
+                // in the frontend once we can also see the existing table schema.
+                writer.write_field_unchecked(
+                    &key,
+                    ColumnDataType::String,
+                    Some(ValueData::StringValue(v)),
                     row,
-                )?;
+                );
             }
             Some(OtlpValue::BoolValue(v)) => {
-                row_writer::write_fields(
-                    writer,
-                    std::iter::once(make_column_data(
-                        &key,
-                        ColumnDataType::Boolean,
-                        Some(ValueData::BoolValue(v)),
-                    )),
+                // Do not coerce or promote types while building the request-local rows.
+                writer.write_field_unchecked(
+                    &key,
+                    ColumnDataType::Boolean,
+                    Some(ValueData::BoolValue(v)),
                     row,
-                )?;
+                );
             }
             Some(OtlpValue::IntValue(v)) => {
-                row_writer::write_fields(
-                    writer,
-                    std::iter::once(make_column_data(
-                        &key,
-                        ColumnDataType::Int64,
-                        Some(ValueData::I64Value(v)),
-                    )),
+                // Preserving the original value avoids order-dependent behavior inside one batch.
+                writer.write_field_unchecked(
+                    &key,
+                    ColumnDataType::Int64,
+                    Some(ValueData::I64Value(v)),
                     row,
-                )?;
+                );
             }
             Some(OtlpValue::DoubleValue(v)) => {
-                row_writer::write_fields(
-                    writer,
-                    std::iter::once(make_column_data(
-                        &key,
-                        ColumnDataType::Float64,
-                        Some(ValueData::F64Value(v)),
-                    )),
+                writer.write_field_unchecked(
+                    &key,
+                    ColumnDataType::Float64,
+                    Some(ValueData::F64Value(v)),
                     row,
-                )?;
+                );
             }
             Some(OtlpValue::ArrayValue(v)) => row_writer::write_json(
                 writer,
@@ -315,3 +298,256 @@ fn write_attributes(
 
     Ok(())
 }
+
+#[cfg(test)]
+mod tests {
+    use api::v1::value::ValueData;
+    use opentelemetry_proto::tonic::common::v1::any_value::Value as OtlpValue;
+    use opentelemetry_proto::tonic::common::v1::{AnyValue, KeyValue};
+
+    use super::*;
+    use crate::otlp::trace::TraceAuxData;
+    use crate::otlp::trace::attributes::Attributes;
+    use crate::otlp::trace::span::{SpanEvents, SpanLinks};
+    use crate::row_writer::TableData;
+
+    fn make_kv(key: &str, value: OtlpValue) -> KeyValue {
+        KeyValue {
+            key: key.to_string(),
+            value: Some(AnyValue { value: Some(value) }),
+        }
+    }
+
+    fn make_span(service_name: &str, trace_id: &str, span_id: &str) -> TraceSpan {
+        TraceSpan {
+            service_name: Some(service_name.to_string()),
+            trace_id: trace_id.to_string(),
+            span_id: span_id.to_string(),
+            parent_span_id: None,
+            resource_attributes: Attributes::from(vec![]),
+            scope_name: "scope".to_string(),
+            scope_version: "v1".to_string(),
+            scope_attributes: Attributes::from(vec![]),
+            trace_state: String::new(),
+            span_name: "op".to_string(),
+            span_kind: "SPAN_KIND_SERVER".to_string(),
+            span_status_code: "STATUS_CODE_UNSET".to_string(),
+            span_status_message: String::new(),
+            span_attributes: Attributes::from(vec![]),
+            span_events: SpanEvents::from(vec![]),
+            span_links: SpanLinks::from(vec![]),
+            start_in_nanosecond: 1,
+            end_in_nanosecond: 2,
+        }
+    }
+
+    #[test]
+    fn test_keep_mixed_numeric_values_until_frontend_reconciliation() {
+        let mut writer = TableData::new(4, 2);
+
+        let attrs1 = Attributes::from(vec![make_kv("val", OtlpValue::DoubleValue(1.5))]);
+        let mut row1 = writer.alloc_one_row();
+        write_attributes(&mut writer, "attr", attrs1, &mut row1).unwrap();
+        writer.add_row(row1);
+
+        let attrs2 = Attributes::from(vec![make_kv("val", OtlpValue::IntValue(42))]);
+        let mut row2 = writer.alloc_one_row();
+        write_attributes(&mut writer, "attr", attrs2, &mut row2).unwrap();
+        writer.add_row(row2);
+
+        let (schema, rows) = writer.into_schema_and_rows();
+
+        let col_idx = schema
+            .iter()
+            .position(|c| c.column_name == "attr.val")
+            .unwrap();
+        assert_eq!(schema[col_idx].datatype, ColumnDataType::Float64 as i32);
+
+        assert_eq!(
+            rows[0].values[col_idx].value_data,
+            Some(ValueData::F64Value(1.5))
+        );
+        assert_eq!(
+            rows[1].values[col_idx].value_data,
+            Some(ValueData::I64Value(42))
+        );
+    }
+
+    #[test]
+    fn test_keep_mixed_string_and_int_values_until_frontend_reconciliation() {
+        let mut writer = TableData::new(4, 2);
+
+        let attrs1 = Attributes::from(vec![make_kv("val", OtlpValue::IntValue(10))]);
+        let mut row1 = writer.alloc_one_row();
+        write_attributes(&mut writer, "attr", attrs1, &mut row1).unwrap();
+        writer.add_row(row1);
+
+        let attrs2 = Attributes::from(vec![make_kv(
+            "val",
+            OtlpValue::StringValue("20".to_string()),
+        )]);
+        let mut row2 = writer.alloc_one_row();
+        write_attributes(&mut writer, "attr", attrs2, &mut row2).unwrap();
+        writer.add_row(row2);
+
+        let (schema, rows) = writer.into_schema_and_rows();
+        let col_idx = schema
+            .iter()
+            .position(|c| c.column_name == "attr.val")
+            .unwrap();
+        assert_eq!(schema[col_idx].datatype, ColumnDataType::Int64 as i32);
+        assert_eq!(
+            rows[1].values[col_idx].value_data,
+            Some(ValueData::StringValue("20".to_string()))
+        );
+    }
+
+    #[test]
+    fn test_keep_first_seen_schema_until_frontend_reconciliation() {
+        let mut writer = TableData::new(4, 2);
+
+        let attrs1 = Attributes::from(vec![make_kv(
+            "val",
+            OtlpValue::StringValue("10".to_string()),
+        )]);
+        let mut row1 = writer.alloc_one_row();
+        write_attributes(&mut writer, "attr", attrs1, &mut row1).unwrap();
+        writer.add_row(row1);
+
+        let attrs2 = Attributes::from(vec![make_kv("val", OtlpValue::IntValue(20))]);
+        let mut row2 = writer.alloc_one_row();
+        write_attributes(&mut writer, "attr", attrs2, &mut row2).unwrap();
+        writer.add_row(row2);
+
+        let (schema, rows) = writer.into_schema_and_rows();
+        let col_idx = schema
+            .iter()
+            .position(|c| c.column_name == "attr.val")
+            .unwrap();
+        assert_eq!(schema[col_idx].datatype, ColumnDataType::String as i32);
+        assert_eq!(
+            rows[0].values[col_idx].value_data,
+            Some(ValueData::StringValue("10".to_string()))
+        );
+        assert_eq!(
+            rows[1].values[col_idx].value_data,
+            Some(ValueData::I64Value(20))
+        );
+    }
+
+    #[test]
+    fn test_keep_mixed_string_and_float_values_until_frontend_reconciliation() {
+        let mut writer = TableData::new(4, 2);
+
+        let attrs1 = Attributes::from(vec![make_kv("val", OtlpValue::DoubleValue(1.5))]);
+        let mut row1 = writer.alloc_one_row();
+        write_attributes(&mut writer, "attr", attrs1, &mut row1).unwrap();
+        writer.add_row(row1);
+
+        let attrs2 = Attributes::from(vec![make_kv(
+            "val",
+            OtlpValue::StringValue("1.5".to_string()),
+        )]);
+        let mut row2 = writer.alloc_one_row();
+        write_attributes(&mut writer, "attr", attrs2, &mut row2).unwrap();
+        writer.add_row(row2);
+
+        let (schema, rows) = writer.into_schema_and_rows();
+        let col_idx = schema
+            .iter()
+            .position(|c| c.column_name == "attr.val")
+            .unwrap();
+        assert_eq!(schema[col_idx].datatype, ColumnDataType::Float64 as i32);
+        assert_eq!(
+            rows[1].values[col_idx].value_data,
+            Some(ValueData::StringValue("1.5".to_string()))
+        );
+    }
+
+    #[test]
+    fn test_keep_mixed_string_and_bool_values_until_frontend_reconciliation() {
+        let mut writer = TableData::new(4, 2);
+
+        let attrs1 = Attributes::from(vec![make_kv(
+            "val",
+            OtlpValue::StringValue("true".to_string()),
+        )]);
+        let mut row1 = writer.alloc_one_row();
+        write_attributes(&mut writer, "attr", attrs1, &mut row1).unwrap();
+        writer.add_row(row1);
+
+        let attrs2 = Attributes::from(vec![make_kv("val", OtlpValue::BoolValue(false))]);
+        let mut row2 = writer.alloc_one_row();
+        write_attributes(&mut writer, "attr", attrs2, &mut row2).unwrap();
+        writer.add_row(row2);
+
+        let (schema, rows) = writer.into_schema_and_rows();
+        let col_idx = schema
+            .iter()
+            .position(|c| c.column_name == "attr.val")
+            .unwrap();
+        assert_eq!(schema[col_idx].datatype, ColumnDataType::String as i32);
+        assert_eq!(
+            rows[0].values[col_idx].value_data,
+            Some(ValueData::StringValue("true".to_string()))
+        );
+        assert_eq!(
+            rows[1].values[col_idx].value_data,
+            Some(ValueData::BoolValue(false))
+        );
+    }
+
+    #[test]
+    fn test_keep_mixed_binary_and_string_values_until_frontend_reconciliation() {
+        let mut writer = TableData::new(4, 2);
+
+        let attrs1 = Attributes::from(vec![make_kv(
+            "val",
+            OtlpValue::BytesValue(vec![1_u8, 2, 3]),
+        )]);
+        let mut row1 = writer.alloc_one_row();
+        write_attributes(&mut writer, "attr", attrs1, &mut row1).unwrap();
+        writer.add_row(row1);
+
+        let attrs2 = Attributes::from(vec![make_kv(
+            "val",
+            OtlpValue::StringValue("false".to_string()),
+        )]);
+        let mut row2 = writer.alloc_one_row();
+        write_attributes(&mut writer, "attr", attrs2, &mut row2).unwrap();
+        writer.add_row(row2);
+
+        let (schema, rows) = writer.into_schema_and_rows();
+        let col_idx = schema
+            .iter()
+            .position(|c| c.column_name == "attr.val")
+            .unwrap();
+        assert_eq!(schema[col_idx].datatype, ColumnDataType::Binary as i32);
+        assert_eq!(
+            rows[0].values[col_idx].value_data,
+            Some(ValueData::BinaryValue(vec![1_u8, 2, 3]))
+        );
+        assert_eq!(
+            rows[1].values[col_idx].value_data,
+            Some(ValueData::StringValue("false".to_string()))
+        );
+    }
+
+    #[test]
+    fn test_build_aux_table_requests_deduplicates_services_and_operations() {
+        let spans = vec![
+            make_span("svc-a", "trace-a", "span-a"),
+            make_span("svc-a", "trace-b", "span-b"),
+        ];
+        let mut aux_data = TraceAuxData::default();
+        for span in &spans {
+            aux_data.observe_span(span);
+        }
+
+        let (requests, total_rows) =
+            build_aux_table_requests(aux_data, "opentelemetry_traces").unwrap();
+        assert_eq!(requests.inserts.len(), 2);
+        assert_eq!(total_rows, 2);
+    }
+    // Conversion matrix coverage lives in the shared coercion helper tests.
+}
diff --git a/src/servers/src/pending_rows_batcher.rs b/src/servers/src/pending_rows_batcher.rs
new file mode 100644
index 0000000000..307311a4e2
--- /dev/null
+++ b/src/servers/src/pending_rows_batcher.rs
@@ -0,0 +1,2704 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::{HashMap, HashSet};
+use std::sync::Arc;
+use std::time::{Duration, Instant};
+
+use api::v1::meta::Peer;
+use api::v1::region::{
+    BulkInsertRequest, RegionRequest, RegionRequestHeader, bulk_insert_request, region_request,
+};
+use api::v1::{ArrowIpc, ColumnSchema, RowInsertRequests, Rows};
+use arrow::compute::{concat_batches, filter_record_batch};
+use arrow::datatypes::{DataType as ArrowDataType, Schema as ArrowSchema, TimeUnit};
+use arrow::record_batch::RecordBatch;
+use async_trait::async_trait;
+use bytes::Bytes;
+use catalog::CatalogManagerRef;
+use common_grpc::flight::{FlightEncoder, FlightMessage};
+use common_meta::node_manager::NodeManagerRef;
+use common_query::prelude::{GREPTIME_PHYSICAL_TABLE, greptime_timestamp, greptime_value};
+use common_telemetry::tracing_context::TracingContext;
+use common_telemetry::{debug, warn};
+use dashmap::DashMap;
+use dashmap::mapref::entry::Entry;
+use metric_engine::batch_modifier::{TagColumnInfo, modify_batch_sparse};
+use partition::manager::PartitionRuleManagerRef;
+use partition::partition::PartitionRuleRef;
+use session::context::QueryContextRef;
+use smallvec::SmallVec;
+use snafu::{OptionExt, ResultExt, ensure};
+use store_api::storage::{RegionId, TableId};
+use table::metadata::{TableInfo, TableInfoRef};
+use tokio::sync::{OwnedSemaphorePermit, Semaphore, broadcast, mpsc, oneshot};
+
+use crate::error;
+use crate::error::{Error, Result};
+use crate::metrics::{
+    FLUSH_DROPPED_ROWS, FLUSH_ELAPSED, FLUSH_FAILURES, FLUSH_ROWS, FLUSH_TOTAL, PENDING_BATCHES,
+    PENDING_ROWS, PENDING_ROWS_BATCH_FLUSH_STAGE_ELAPSED, PENDING_ROWS_BATCH_INGEST_STAGE_ELAPSED,
+    PENDING_WORKERS,
+};
+use crate::prom_row_builder::{
+    build_prom_create_table_schema_from_proto, identify_missing_columns_from_proto,
+    rows_to_aligned_record_batch,
+};
+
+const PHYSICAL_TABLE_KEY: &str = "physical_table";
+/// Whether wait for ingestion result before reply to client.
+const PENDING_ROWS_BATCH_SYNC_ENV: &str = "PENDING_ROWS_BATCH_SYNC";
+const WORKER_IDLE_TIMEOUT_MULTIPLIER: u32 = 3;
+const PHYSICAL_REGION_ESSENTIAL_COLUMN_COUNT: usize = 3;
+
+#[async_trait]
+pub trait PendingRowsSchemaAlterer: Send + Sync {
+    /// Batch-create multiple logical tables that are missing.
+    /// Each entry is `(table_name, request_schema)`.
+    async fn create_tables_if_missing_batch(
+        &self,
+        catalog: &str,
+        schema: &str,
+        tables: &[(&str, &[ColumnSchema])],
+        with_metric_engine: bool,
+        ctx: QueryContextRef,
+    ) -> Result<()>;
+
+    /// Batch-alter multiple logical tables to add missing tag columns.
+    /// Each entry is `(table_name, missing_column_names)`.
+    async fn add_missing_prom_tag_columns_batch(
+        &self,
+        catalog: &str,
+        schema: &str,
+        tables: &[(&str, &[String])],
+        ctx: QueryContextRef,
+    ) -> Result<()>;
+}
+
+pub type PendingRowsSchemaAltererRef = Arc<dyn PendingRowsSchemaAlterer>;
+
+#[derive(Clone)]
+pub struct PhysicalTableMetadata {
+    pub table_info: TableInfoRef,
+    /// Mapping from column name to column id
+    pub col_name_to_ids: Option<HashMap<String, u32>>,
+}
+
+#[async_trait]
+pub trait PhysicalFlushCatalogProvider: Send + Sync {
+    async fn physical_table(
+        &self,
+        catalog: &str,
+        schema: &str,
+        table_name: &str,
+        query_ctx: &session::context::QueryContext,
+    ) -> catalog::error::Result<Option<PhysicalTableMetadata>>;
+}
+
+#[async_trait]
+pub trait PhysicalFlushPartitionProvider: Send + Sync {
+    async fn find_table_partition_rule(
+        &self,
+        table_info: &TableInfo,
+    ) -> partition::error::Result<PartitionRuleRef>;
+
+    async fn find_region_leader(&self, region_id: RegionId) -> Result<Peer>;
+}
+
+#[async_trait]
+pub trait PhysicalFlushNodeRequester: Send + Sync {
+    async fn handle(
+        &self,
+        peer: &Peer,
+        request: RegionRequest,
+    ) -> Result<api::region::RegionResponse>;
+}
+
+#[derive(Clone)]
+struct CatalogManagerPhysicalFlushAdapter {
+    catalog_manager: CatalogManagerRef,
+}
+
+#[async_trait]
+impl PhysicalFlushCatalogProvider for CatalogManagerPhysicalFlushAdapter {
+    async fn physical_table(
+        &self,
+        catalog: &str,
+        schema: &str,
+        table_name: &str,
+        query_ctx: &session::context::QueryContext,
+    ) -> catalog::error::Result<Option<PhysicalTableMetadata>> {
+        self.catalog_manager
+            .table(catalog, schema, table_name, Some(query_ctx))
+            .await
+            .map(|table| {
+                table.map(|table| {
+                    let table_info = table.table_info();
+                    let name_to_ids = table_info.name_to_ids();
+                    PhysicalTableMetadata {
+                        table_info,
+                        col_name_to_ids: name_to_ids,
+                    }
+                })
+            })
+    }
+}
+
+#[derive(Clone)]
+struct PartitionManagerPhysicalFlushAdapter {
+    partition_manager: PartitionRuleManagerRef,
+}
+
+#[async_trait]
+impl PhysicalFlushPartitionProvider for PartitionManagerPhysicalFlushAdapter {
+    async fn find_table_partition_rule(
+        &self,
+        table_info: &TableInfo,
+    ) -> partition::error::Result<PartitionRuleRef> {
+        self.partition_manager
+            .find_table_partition_rule(table_info)
+            .await
+            .map(|(rule, _)| rule)
+    }
+
+    async fn find_region_leader(&self, region_id: RegionId) -> Result<Peer> {
+        let peer = self.partition_manager.find_region_leader(region_id).await?;
+        Ok(peer)
+    }
+}
+
+#[derive(Clone)]
+struct NodeManagerPhysicalFlushAdapter {
+    node_manager: NodeManagerRef,
+}
+
+#[async_trait]
+impl PhysicalFlushNodeRequester for NodeManagerPhysicalFlushAdapter {
+    async fn handle(
+        &self,
+        peer: &Peer,
+        request: RegionRequest,
+    ) -> error::Result<api::region::RegionResponse> {
+        let datanode = self.node_manager.datanode(peer).await;
+        datanode
+            .handle(request)
+            .await
+            .context(error::CommonMetaSnafu)
+    }
+}
+
+#[derive(Debug, Clone, Hash, Eq, PartialEq)]
+struct BatchKey {
+    catalog: String,
+    schema: String,
+    physical_table: String,
+}
+
+#[derive(Debug)]
+pub struct TableBatch {
+    pub table_name: String,
+    pub table_id: TableId,
+    pub batches: Vec<RecordBatch>,
+    pub row_count: usize,
+}
+
+/// Intermediate planning state for resolving and preparing logical tables
+/// before row-to-batch alignment.
+struct TableResolutionPlan {
+    /// Resolved table schema and table id by logical table name.
+    region_schemas: HashMap<String, (Arc<ArrowSchema>, u32)>,
+    /// Missing tables that need to be created before alignment.
+    tables_to_create: Vec<(String, Vec<ColumnSchema>)>,
+    /// Existing tables that need tag-column schema evolution.
+    tables_to_alter: Vec<(String, Vec<String>)>,
+}
+
+struct PendingBatch {
+    tables: HashMap<String, TableBatch>,
+    created_at: Instant,
+    total_row_count: usize,
+    db_string: String,
+    ctx: QueryContextRef,
+    waiters: Vec<FlushWaiter>,
+}
+
+struct FlushWaiter {
+    response_tx: oneshot::Sender<std::result::Result<(), Arc<Error>>>,
+    _permit: OwnedSemaphorePermit,
+}
+
+struct FlushBatch {
+    table_batches: Vec<TableBatch>,
+    total_row_count: usize,
+    db_string: String,
+    ctx: QueryContextRef,
+    waiters: Vec<FlushWaiter>,
+}
+
+#[derive(Clone)]
+struct PendingWorker {
+    tx: mpsc::Sender<WorkerCommand>,
+}
+
+enum WorkerCommand {
+    Submit {
+        table_batches: Vec<(String, u32, RecordBatch)>,
+        total_rows: usize,
+        ctx: QueryContextRef,
+        response_tx: oneshot::Sender<std::result::Result<(), Arc<Error>>>,
+        _permit: OwnedSemaphorePermit,
+    },
+}
+
+// Batch key is derived from QueryContext; it assumes catalog/schema/physical_table fully
+// define the write target and must remain consistent across the batch.
+fn batch_key_from_ctx(ctx: &QueryContextRef) -> BatchKey {
+    let physical_table = ctx
+        .extension(PHYSICAL_TABLE_KEY)
+        .unwrap_or(GREPTIME_PHYSICAL_TABLE)
+        .to_string();
+    BatchKey {
+        catalog: ctx.current_catalog().to_string(),
+        schema: ctx.current_schema(),
+        physical_table,
+    }
+}
+
+/// Prometheus remote write pending rows batcher.
+pub struct PendingRowsBatcher {
+    workers: Arc<DashMap<BatchKey, PendingWorker>>,
+    flush_interval: Duration,
+    max_batch_rows: usize,
+    partition_manager: PartitionRuleManagerRef,
+    node_manager: NodeManagerRef,
+    catalog_manager: CatalogManagerRef,
+    flush_semaphore: Arc<Semaphore>,
+    inflight_semaphore: Arc<Semaphore>,
+    worker_channel_capacity: usize,
+    prom_store_with_metric_engine: bool,
+    schema_alterer: PendingRowsSchemaAltererRef,
+    pending_rows_batch_sync: bool,
+    shutdown: broadcast::Sender<()>,
+}
+
+impl PendingRowsBatcher {
+    #[allow(clippy::too_many_arguments)]
+    pub fn try_new(
+        partition_manager: PartitionRuleManagerRef,
+        node_manager: NodeManagerRef,
+        catalog_manager: CatalogManagerRef,
+        prom_store_with_metric_engine: bool,
+        schema_alterer: PendingRowsSchemaAltererRef,
+        flush_interval: Duration,
+        max_batch_rows: usize,
+        max_concurrent_flushes: usize,
+        worker_channel_capacity: usize,
+        max_inflight_requests: usize,
+    ) -> Option<Arc<Self>> {
+        // Disable the batcher if flush is disabled or configuration is invalid.
+        // Zero values for these knobs either cause panics (e.g., zero-capacity channels)
+        // or deadlocks (e.g., semaphores with no permits).
+        if flush_interval.is_zero()
+            || max_batch_rows == 0
+            || max_concurrent_flushes == 0
+            || worker_channel_capacity == 0
+            || max_inflight_requests == 0
+        {
+            return None;
+        }
+
+        let (shutdown, _) = broadcast::channel(1);
+        let pending_rows_batch_sync = std::env::var(PENDING_ROWS_BATCH_SYNC_ENV)
+            .ok()
+            .as_deref()
+            .and_then(|v| v.parse::<bool>().ok())
+            .unwrap_or(true);
+        let workers = Arc::new(DashMap::new());
+        PENDING_WORKERS.set(workers.len() as i64);
+
+        Some(Arc::new(Self {
+            workers,
+            flush_interval,
+            max_batch_rows,
+            partition_manager,
+            node_manager,
+            catalog_manager,
+            prom_store_with_metric_engine,
+            schema_alterer,
+            flush_semaphore: Arc::new(Semaphore::new(max_concurrent_flushes)),
+            inflight_semaphore: Arc::new(Semaphore::new(max_inflight_requests)),
+            worker_channel_capacity,
+            pending_rows_batch_sync,
+            shutdown,
+        }))
+    }
+
+    pub async fn submit(&self, requests: RowInsertRequests, ctx: QueryContextRef) -> Result<u64> {
+        let (table_batches, total_rows) = {
+            let _timer = PENDING_ROWS_BATCH_INGEST_STAGE_ELAPSED
+                .with_label_values(&["submit_build_and_align"])
+                .start_timer();
+            self.build_and_align_table_batches(requests, &ctx).await?
+        };
+        if total_rows == 0 {
+            return Ok(0);
+        }
+
+        let permit = {
+            let _timer = PENDING_ROWS_BATCH_INGEST_STAGE_ELAPSED
+                .with_label_values(&["submit_acquire_inflight_permit"])
+                .start_timer();
+            self.inflight_semaphore
+                .clone()
+                .acquire_owned()
+                .await
+                .map_err(|_| error::BatcherChannelClosedSnafu.build())?
+        };
+
+        let (response_tx, response_rx) = oneshot::channel();
+
+        let batch_key = batch_key_from_ctx(&ctx);
+        let mut cmd = Some(WorkerCommand::Submit {
+            table_batches,
+            total_rows,
+            ctx,
+            response_tx,
+            _permit: permit,
+        });
+
+        {
+            let _timer = PENDING_ROWS_BATCH_INGEST_STAGE_ELAPSED
+                .with_label_values(&["submit_send_to_worker"])
+                .start_timer();
+
+            for _ in 0..2 {
+                let worker = self.get_or_spawn_worker(batch_key.clone());
+                let Some(worker_cmd) = cmd.take() else {
+                    break;
+                };
+
+                match worker.tx.send(worker_cmd).await {
+                    Ok(()) => break,
+                    Err(err) => {
+                        cmd = Some(err.0);
+                        remove_worker_if_same_channel(
+                            self.workers.as_ref(),
+                            &batch_key,
+                            &worker.tx,
+                        );
+                    }
+                }
+            }
+
+            if cmd.is_some() {
+                return Err(Error::BatcherChannelClosed);
+            }
+        }
+
+        if self.pending_rows_batch_sync {
+            let result = {
+                let _timer = PENDING_ROWS_BATCH_INGEST_STAGE_ELAPSED
+                    .with_label_values(&["submit_wait_flush_result"])
+                    .start_timer();
+                response_rx
+                    .await
+                    .map_err(|_| error::BatcherChannelClosedSnafu.build())?
+            };
+            result
+                .context(error::SubmitBatchSnafu)
+                .map(|()| total_rows as u64)
+        } else {
+            Ok(total_rows as u64)
+        }
+    }
+
+    /// Converts proto `RowInsertRequests` directly into aligned `RecordBatch`es
+    /// in a single pass, handling table creation, schema alteration, column
+    /// renaming, reordering, and null-filling without building intermediate
+    /// RecordBatches.
+    async fn build_and_align_table_batches(
+        &self,
+        requests: RowInsertRequests,
+        ctx: &QueryContextRef,
+    ) -> Result<(Vec<(String, u32, RecordBatch)>, usize)> {
+        let catalog = ctx.current_catalog().to_string();
+        let schema = ctx.current_schema();
+
+        let (table_rows, total_rows) = Self::collect_non_empty_table_rows(requests);
+        if total_rows == 0 {
+            return Ok((Vec::new(), 0));
+        }
+
+        let unique_tables = Self::collect_unique_table_schemas(&table_rows)?;
+        let mut plan = self
+            .plan_table_resolution(&catalog, &schema, ctx, &unique_tables)
+            .await?;
+
+        self.create_missing_tables_and_refresh_schemas(
+            &catalog,
+            &schema,
+            ctx,
+            &table_rows,
+            &mut plan,
+        )
+        .await?;
+
+        self.alter_tables_and_refresh_schemas(&catalog, &schema, ctx, &mut plan)
+            .await?;
+
+        let aligned_batches = Self::build_aligned_batches(&table_rows, &plan.region_schemas)?;
+
+        Ok((aligned_batches, total_rows))
+    }
+
+    /// Extracts non-empty `(table_name, rows)` pairs and computes total row
+    /// count across the retained entries.
+    fn collect_non_empty_table_rows(requests: RowInsertRequests) -> (Vec<(String, Rows)>, usize) {
+        let mut table_rows: Vec<(String, Rows)> = Vec::with_capacity(requests.inserts.len());
+        let mut total_rows = 0;
+
+        for request in requests.inserts {
+            let Some(rows) = request.rows else {
+                continue;
+            };
+            if rows.rows.is_empty() {
+                continue;
+            }
+
+            total_rows += rows.rows.len();
+            table_rows.push((request.table_name, rows));
+        }
+
+        (table_rows, total_rows)
+    }
+
+    /// Returns unique `(table_name, proto_schema)` pairs while keeping the
+    /// first-seen schema for duplicate table names.
+    fn collect_unique_table_schemas(
+        table_rows: &[(String, Rows)],
+    ) -> Result<Vec<(&str, &[ColumnSchema])>> {
+        let mut unique_tables: Vec<(&str, &[ColumnSchema])> = Vec::with_capacity(table_rows.len());
+        let mut seen = HashSet::new();
+
+        for (table_name, rows) in table_rows {
+            if seen.insert(table_name.as_str()) {
+                unique_tables.push((table_name.as_str(), &rows.schema));
+            } else {
+                // table_rows should group rows by table name.
+                return error::InvalidPromRemoteRequestSnafu {
+                    msg: format!(
+                        "Found duplicated table name in RowInsertRequest: {}",
+                        table_name
+                    ),
+                }
+                .fail();
+            }
+        }
+
+        Ok(unique_tables)
+    }
+
+    /// Resolves table metadata and classifies each table into existing,
+    /// to-create, and to-alter groups used by subsequent DDL steps.
+    async fn plan_table_resolution(
+        &self,
+        catalog: &str,
+        schema: &str,
+        ctx: &QueryContextRef,
+        unique_tables: &[(&str, &[ColumnSchema])],
+    ) -> Result<TableResolutionPlan> {
+        let mut plan = TableResolutionPlan {
+            region_schemas: HashMap::with_capacity(unique_tables.len()),
+            tables_to_create: Vec::new(),
+            tables_to_alter: Vec::new(),
+        };
+
+        let resolved_tables = {
+            let _timer = PENDING_ROWS_BATCH_INGEST_STAGE_ELAPSED
+                .with_label_values(&["align_resolve_table"])
+                .start_timer();
+            futures::future::join_all(unique_tables.iter().map(|(table_name, _)| {
+                self.catalog_manager
+                    .table(catalog, schema, table_name, Some(ctx.as_ref()))
+            }))
+            .await
+        };
+
+        for ((table_name, rows_schema), table_result) in unique_tables.iter().zip(resolved_tables) {
+            let table = table_result?;
+
+            if let Some(table) = table {
+                let table_info = table.table_info();
+                let table_id = table_info.ident.table_id;
+                let region_schema = table_info.meta.schema.arrow_schema().clone();
+
+                let missing_columns = {
+                    let _timer = PENDING_ROWS_BATCH_INGEST_STAGE_ELAPSED
+                        .with_label_values(&["align_identify_missing_columns"])
+                        .start_timer();
+                    identify_missing_columns_from_proto(rows_schema, region_schema.as_ref())?
+                };
+                if !missing_columns.is_empty() {
+                    plan.tables_to_alter
+                        .push(((*table_name).to_string(), missing_columns));
+                }
+                plan.region_schemas
+                    .insert((*table_name).to_string(), (region_schema, table_id));
+            } else {
+                let request_schema = {
+                    let _timer = PENDING_ROWS_BATCH_INGEST_STAGE_ELAPSED
+                        .with_label_values(&["align_build_create_table_schema"])
+                        .start_timer();
+                    build_prom_create_table_schema_from_proto(rows_schema)?
+                };
+                plan.tables_to_create
+                    .push(((*table_name).to_string(), request_schema));
+            }
+        }
+
+        Ok(plan)
+    }
+
+    /// Batch-creates missing tables, refreshes their schema metadata, and
+    /// enqueues follow-up alters for extra tag columns discovered in later rows.
+    async fn create_missing_tables_and_refresh_schemas(
+        &self,
+        catalog: &str,
+        schema: &str,
+        ctx: &QueryContextRef,
+        table_rows: &[(String, Rows)],
+        plan: &mut TableResolutionPlan,
+    ) -> Result<()> {
+        if plan.tables_to_create.is_empty() {
+            return Ok(());
+        }
+
+        let create_refs: Vec<(&str, &[ColumnSchema])> = plan
+            .tables_to_create
+            .iter()
+            .map(|(name, schema)| (name.as_str(), schema.as_slice()))
+            .collect();
+
+        {
+            let _timer = PENDING_ROWS_BATCH_INGEST_STAGE_ELAPSED
+                .with_label_values(&["align_batch_create_tables"])
+                .start_timer();
+            self.schema_alterer
+                .create_tables_if_missing_batch(
+                    catalog,
+                    schema,
+                    &create_refs,
+                    self.prom_store_with_metric_engine,
+                    ctx.clone(),
+                )
+                .await?;
+        }
+
+        let created_table_results = {
+            let _timer = PENDING_ROWS_BATCH_INGEST_STAGE_ELAPSED
+                .with_label_values(&["align_resolve_table_after_create"])
+                .start_timer();
+            futures::future::join_all(plan.tables_to_create.iter().map(|(table_name, _)| {
+                self.catalog_manager
+                    .table(catalog, schema, table_name, Some(ctx.as_ref()))
+            }))
+            .await
+        };
+
+        for ((table_name, _), table_result) in
+            plan.tables_to_create.iter().zip(created_table_results)
+        {
+            let table = table_result?.with_context(|| error::UnexpectedResultSnafu {
+                reason: format!(
+                    "Table not found after pending batch create attempt: {}",
+                    table_name
+                ),
+            })?;
+            let table_info = table.table_info();
+            let table_id = table_info.ident.table_id;
+            let region_schema = table_info.meta.schema.arrow_schema().clone();
+            plan.region_schemas
+                .insert(table_name.clone(), (region_schema, table_id));
+        }
+
+        Self::enqueue_alter_for_new_tables(table_rows, plan)?;
+
+        Ok(())
+    }
+
+    /// For newly created tables, re-checks all row schemas and appends alter
+    /// operations when additional tag columns are still missing.
+    fn enqueue_alter_for_new_tables(
+        table_rows: &[(String, Rows)],
+        plan: &mut TableResolutionPlan,
+    ) -> Result<()> {
+        let created_tables: HashSet<&str> = plan
+            .tables_to_create
+            .iter()
+            .map(|(table_name, _)| table_name.as_str())
+            .collect();
+
+        for (table_name, rows) in table_rows {
+            if !created_tables.contains(table_name.as_str()) {
+                continue;
+            }
+
+            let Some((region_schema, _)) = plan.region_schemas.get(table_name) else {
+                continue;
+            };
+
+            let missing_columns = identify_missing_columns_from_proto(&rows.schema, region_schema)?;
+            if missing_columns.is_empty()
+                || plan
+                    .tables_to_alter
+                    .iter()
+                    .any(|(existing_name, _)| existing_name == table_name)
+            {
+                continue;
+            }
+
+            plan.tables_to_alter
+                .push((table_name.clone(), missing_columns));
+        }
+
+        Ok(())
+    }
+
+    /// Batch-alters tables that have missing tag columns and refreshes the
+    /// in-memory schema map used for row alignment.
+    async fn alter_tables_and_refresh_schemas(
+        &self,
+        catalog: &str,
+        schema: &str,
+        ctx: &QueryContextRef,
+        plan: &mut TableResolutionPlan,
+    ) -> Result<()> {
+        if plan.tables_to_alter.is_empty() {
+            return Ok(());
+        }
+
+        let alter_refs: Vec<(&str, &[String])> = plan
+            .tables_to_alter
+            .iter()
+            .map(|(name, cols)| (name.as_str(), cols.as_slice()))
+            .collect();
+        {
+            let _timer = PENDING_ROWS_BATCH_INGEST_STAGE_ELAPSED
+                .with_label_values(&["align_batch_add_missing_columns"])
+                .start_timer();
+            self.schema_alterer
+                .add_missing_prom_tag_columns_batch(catalog, schema, &alter_refs, ctx.clone())
+                .await?;
+        }
+
+        let altered_table_results = {
+            let _timer = PENDING_ROWS_BATCH_INGEST_STAGE_ELAPSED
+                .with_label_values(&["align_resolve_table_after_schema_alter"])
+                .start_timer();
+            futures::future::join_all(plan.tables_to_alter.iter().map(|(table_name, _)| {
+                self.catalog_manager
+                    .table(catalog, schema, table_name, Some(ctx.as_ref()))
+            }))
+            .await
+        };
+
+        for ((table_name, _), table_result) in
+            plan.tables_to_alter.iter().zip(altered_table_results)
+        {
+            let table = table_result?.with_context(|| error::UnexpectedResultSnafu {
+                reason: format!(
+                    "Table not found after pending batch schema alter: {}",
+                    table_name
+                ),
+            })?;
+            let table_info = table.table_info();
+            let table_id = table_info.ident.table_id;
+            let refreshed_region_schema = table_info.meta.schema.arrow_schema().clone();
+            plan.region_schemas
+                .insert(table_name.clone(), (refreshed_region_schema, table_id));
+        }
+
+        Ok(())
+    }
+
+    /// Converts proto rows to `RecordBatch` values aligned to resolved region
+    /// schemas and returns `(table_name, table_id, batch)` tuples.
+    fn build_aligned_batches(
+        table_rows: &[(String, Rows)],
+        region_schemas: &HashMap<String, (Arc<ArrowSchema>, u32)>,
+    ) -> Result<Vec<(String, u32, RecordBatch)>> {
+        let mut aligned_batches = Vec::with_capacity(table_rows.len());
+        for (table_name, rows) in table_rows {
+            let (region_schema, table_id) =
+                region_schemas.get(table_name).cloned().with_context(|| {
+                    error::UnexpectedResultSnafu {
+                        reason: format!("Region schema not resolved for table: {}", table_name),
+                    }
+                })?;
+
+            let record_batch = {
+                let _timer = PENDING_ROWS_BATCH_INGEST_STAGE_ELAPSED
+                    .with_label_values(&["align_rows_to_record_batch"])
+                    .start_timer();
+                rows_to_aligned_record_batch(rows, region_schema.as_ref())?
+            };
+            aligned_batches.push((table_name.clone(), table_id, record_batch));
+        }
+
+        Ok(aligned_batches)
+    }
+
+    fn get_or_spawn_worker(&self, key: BatchKey) -> PendingWorker {
+        if let Some(worker) = self.workers.get(&key)
+            && !worker.tx.is_closed()
+        {
+            return worker.clone();
+        }
+
+        let entry = self.workers.entry(key.clone());
+        match entry {
+            Entry::Occupied(mut worker) => {
+                if worker.get().tx.is_closed() {
+                    let new_worker = self.spawn_worker(key);
+                    worker.insert(new_worker.clone());
+                    PENDING_WORKERS.set(self.workers.len() as i64);
+                    new_worker
+                } else {
+                    worker.get().clone()
+                }
+            }
+            Entry::Vacant(vacant) => {
+                let worker = self.spawn_worker(key);
+
+                vacant.insert(worker.clone());
+                PENDING_WORKERS.set(self.workers.len() as i64);
+                worker
+            }
+        }
+    }
+
+    fn spawn_worker(&self, key: BatchKey) -> PendingWorker {
+        let (tx, rx) = mpsc::channel(self.worker_channel_capacity);
+        let worker = PendingWorker { tx: tx.clone() };
+        let worker_idle_timeout = self
+            .flush_interval
+            .checked_mul(WORKER_IDLE_TIMEOUT_MULTIPLIER)
+            .unwrap_or(self.flush_interval);
+
+        start_worker(
+            key,
+            worker.tx.clone(),
+            self.workers.clone(),
+            rx,
+            self.shutdown.clone(),
+            self.partition_manager.clone(),
+            self.node_manager.clone(),
+            self.catalog_manager.clone(),
+            self.flush_interval,
+            worker_idle_timeout,
+            self.max_batch_rows,
+            self.flush_semaphore.clone(),
+        );
+
+        worker
+    }
+}
+
+impl Drop for PendingRowsBatcher {
+    fn drop(&mut self) {
+        let _ = self.shutdown.send(());
+    }
+}
+
+impl PendingBatch {
+    fn new(ctx: QueryContextRef) -> Self {
+        let db_string = ctx.get_db_string();
+        Self {
+            tables: HashMap::new(),
+            created_at: Instant::now(),
+            total_row_count: 0,
+            db_string,
+            ctx,
+            waiters: Vec::new(),
+        }
+    }
+}
+
+#[allow(clippy::too_many_arguments)]
+fn start_worker(
+    key: BatchKey,
+    worker_tx: mpsc::Sender<WorkerCommand>,
+    workers: Arc<DashMap<BatchKey, PendingWorker>>,
+    mut rx: mpsc::Receiver<WorkerCommand>,
+    shutdown: broadcast::Sender<()>,
+    partition_manager: PartitionRuleManagerRef,
+    node_manager: NodeManagerRef,
+    catalog_manager: CatalogManagerRef,
+    flush_interval: Duration,
+    worker_idle_timeout: Duration,
+    max_batch_rows: usize,
+    flush_semaphore: Arc<Semaphore>,
+) {
+    tokio::spawn(async move {
+        let mut batch = None;
+        let mut interval = tokio::time::interval(flush_interval);
+        let mut shutdown_rx = shutdown.subscribe();
+        let idle_deadline = tokio::time::Instant::now() + worker_idle_timeout;
+        let idle_timer = tokio::time::sleep_until(idle_deadline);
+        tokio::pin!(idle_timer);
+
+        loop {
+            tokio::select! {
+                cmd = rx.recv() => {
+                    match cmd {
+                        Some(WorkerCommand::Submit { table_batches, total_rows, ctx, response_tx, _permit }) => {
+                            idle_timer.as_mut().reset(tokio::time::Instant::now() + worker_idle_timeout);
+
+                            let pending_batch = batch.get_or_insert_with(||{
+                                PENDING_BATCHES.inc();
+                                PendingBatch::new(ctx)
+                            });
+
+                            pending_batch.waiters.push(FlushWaiter { response_tx, _permit });
+
+                            for (table_name, table_id, record_batch) in table_batches {
+                                let entry = pending_batch.tables.entry(table_name.clone()).or_insert_with(|| TableBatch {
+                                    table_name,
+                                    table_id,
+                                    batches: Vec::new(),
+                                    row_count: 0,
+                                });
+                                entry.row_count += record_batch.num_rows();
+                                entry.batches.push(record_batch);
+                            }
+
+                            pending_batch.total_row_count += total_rows;
+                            PENDING_ROWS.add(total_rows as i64);
+
+                            if pending_batch.total_row_count >= max_batch_rows
+                                && let Some(flush) = drain_batch(&mut batch) {
+                                    spawn_flush(
+                                        flush,
+                                        partition_manager.clone(),
+                                        node_manager.clone(),
+                                        catalog_manager.clone(),
+                                        flush_semaphore.clone(),
+                                    ).await;
+                            }
+                        }
+                        None => {
+                            if let Some(flush) = drain_batch(&mut batch) {
+                                flush_batch(
+                                    flush,
+                                    partition_manager.clone(),
+                                    node_manager.clone(),
+                                    catalog_manager.clone(),
+                                ).await;
+                            }
+                            break;
+                        }
+                    }
+                }
+                _ = &mut idle_timer => {
+                    if !should_close_worker_on_idle_timeout(
+                        batch.as_ref().map_or(0, |batch| batch.total_row_count),
+                        rx.len(),
+                    ) {
+                        idle_timer
+                            .as_mut()
+                            .reset(tokio::time::Instant::now() + worker_idle_timeout);
+                        continue;
+                    }
+
+                    debug!(
+                        "Closing idle pending rows worker due to timeout: catalog={}, schema={}, physical_table={}",
+                        key.catalog,
+                        key.schema,
+                        key.physical_table
+                    );
+                    break;
+                }
+                _ = interval.tick() => {
+                    if batch
+                        .as_ref()
+                        .is_some_and(|batch| batch.created_at.elapsed() >= flush_interval)
+                        && let Some(flush) = drain_batch(&mut batch) {
+                            spawn_flush(
+                                flush,
+                                partition_manager.clone(),
+                                node_manager.clone(),
+                                catalog_manager.clone(),
+                                flush_semaphore.clone(),
+                            ).await;
+                    }
+                }
+                _ = shutdown_rx.recv() => {
+                    if let Some(flush) = drain_batch(&mut batch) {
+                        flush_batch(
+                            flush,
+                            partition_manager.clone(),
+                            node_manager.clone(),
+                            catalog_manager.clone(),
+                        ).await;
+                    }
+                    break;
+                }
+            }
+        }
+
+        remove_worker_if_same_channel(workers.as_ref(), &key, &worker_tx);
+    });
+}
+
+fn remove_worker_if_same_channel(
+    workers: &DashMap<BatchKey, PendingWorker>,
+    key: &BatchKey,
+    worker_tx: &mpsc::Sender<WorkerCommand>,
+) -> bool {
+    if let Some(worker) = workers.get(key)
+        && worker.tx.same_channel(worker_tx)
+    {
+        drop(worker);
+        workers.remove(key);
+        PENDING_WORKERS.set(workers.len() as i64);
+        return true;
+    }
+
+    false
+}
+
+fn should_close_worker_on_idle_timeout(total_row_count: usize, queued_requests: usize) -> bool {
+    total_row_count == 0 && queued_requests == 0
+}
+
+fn drain_batch(batch: &mut Option<PendingBatch>) -> Option<FlushBatch> {
+    let batch = batch.take()?;
+    let total_row_count = batch.total_row_count;
+
+    if total_row_count == 0 {
+        return None;
+    }
+
+    let table_batches = batch.tables.into_values().collect();
+    let waiters = batch.waiters;
+
+    PENDING_ROWS.sub(total_row_count as i64);
+    PENDING_BATCHES.dec();
+
+    Some(FlushBatch {
+        table_batches,
+        total_row_count,
+        db_string: batch.db_string,
+        ctx: batch.ctx,
+        waiters,
+    })
+}
+
+async fn spawn_flush(
+    flush: FlushBatch,
+    partition_manager: PartitionRuleManagerRef,
+    node_manager: NodeManagerRef,
+    catalog_manager: CatalogManagerRef,
+    semaphore: Arc<Semaphore>,
+) {
+    match semaphore.acquire_owned().await {
+        Ok(permit) => {
+            tokio::spawn(async move {
+                let _permit = permit;
+                flush_batch(flush, partition_manager, node_manager, catalog_manager).await;
+            });
+        }
+        Err(err) => {
+            warn!(err; "Flush semaphore closed, flushing inline");
+            flush_batch(flush, partition_manager, node_manager, catalog_manager).await;
+        }
+    }
+}
+
+struct FlushRegionWrite {
+    datanode: Peer,
+    request: RegionRequest,
+}
+
+struct PlannedRegionBatch {
+    region_id: RegionId,
+    batch: RecordBatch,
+}
+
+#[cfg(test)]
+impl PlannedRegionBatch {
+    fn num_rows(&self) -> usize {
+        self.batch.num_rows()
+    }
+}
+
+struct ResolvedRegionBatch {
+    planned: PlannedRegionBatch,
+    datanode: Peer,
+}
+
+fn should_dispatch_concurrently(region_write_count: usize) -> bool {
+    region_write_count > 1
+}
+
+/// Classifies columns in a logical-table batch for sparse primary-key conversion.
+///
+/// Returns:
+/// - `Vec<TagColumnInfo>`: all Utf8 tag columns sorted by tag name, used for
+///   TSID and sparse primary-key encoding.
+/// - `SmallVec<[usize; 3]>`: indices of columns copied into the physical batch
+///   after `__primary_key`, ordered as `[greptime_timestamp, greptime_value,
+///   partition_tag_columns...]`.
+fn columns_taxonomy(
+    batch_schema: &Arc<ArrowSchema>,
+    table_name: &str,
+    name_to_ids: &HashMap<String, u32>,
+    partition_columns: &HashSet<&str>,
+) -> Result<(Vec<TagColumnInfo>, SmallVec<[usize; 3]>)> {
+    let mut tag_columns = Vec::new();
+    let mut essential_column_indices =
+        SmallVec::<[usize; 3]>::with_capacity(2 + partition_columns.len());
+    // Placeholder for greptime_timestamp and greptime_value
+    essential_column_indices.push(0);
+    essential_column_indices.push(0);
+
+    let mut timestamp_index = None;
+    let mut value_index = None;
+
+    for (index, field) in batch_schema.fields().iter().enumerate() {
+        match field.data_type() {
+            ArrowDataType::Utf8 => {
+                let column_id = name_to_ids.get(field.name()).copied().with_context(|| {
+                    error::InvalidPromRemoteRequestSnafu {
+                        msg: format!(
+                            "Column '{}' from logical table '{}' not found in physical table column IDs",
+                            field.name(),
+                            table_name
+                        ),
+                    }
+                })?;
+                tag_columns.push(TagColumnInfo {
+                    name: field.name().clone(),
+                    index,
+                    column_id,
+                });
+
+                if partition_columns.contains(field.name().as_str()) {
+                    essential_column_indices.push(index);
+                }
+            }
+            ArrowDataType::Timestamp(TimeUnit::Millisecond, _) => {
+                ensure!(
+                    timestamp_index.replace(index).is_none(),
+                    error::InvalidPromRemoteRequestSnafu {
+                        msg: format!(
+                            "Duplicated timestamp column in logical table '{}' batch schema",
+                            table_name
+                        ),
+                    }
+                );
+            }
+            ArrowDataType::Float64 => {
+                ensure!(
+                    value_index.replace(index).is_none(),
+                    error::InvalidPromRemoteRequestSnafu {
+                        msg: format!(
+                            "Duplicated value column in logical table '{}' batch schema",
+                            table_name
+                        ),
+                    }
+                );
+            }
+            datatype => {
+                return error::InvalidPromRemoteRequestSnafu {
+                    msg: format!(
+                        "Unexpected data type '{datatype:?}' in logical table '{}' batch schema",
+                        table_name
+                    ),
+                }
+                .fail();
+            }
+        }
+    }
+
+    let timestamp_index =
+        timestamp_index.with_context(|| error::InvalidPromRemoteRequestSnafu {
+            msg: format!(
+                "Missing essential column '{}' in logical table '{}' batch schema",
+                greptime_timestamp(),
+                table_name
+            ),
+        })?;
+    let value_index = value_index.with_context(|| error::InvalidPromRemoteRequestSnafu {
+        msg: format!(
+            "Missing essential column '{}' in logical table '{}' batch schema",
+            greptime_value(),
+            table_name
+        ),
+    })?;
+
+    tag_columns.sort_by(|a, b| a.name.cmp(&b.name));
+
+    essential_column_indices[0] = timestamp_index;
+    essential_column_indices[1] = value_index;
+
+    Ok((tag_columns, essential_column_indices))
+}
+
+fn strip_partition_columns_from_batch(batch: RecordBatch) -> Result<RecordBatch> {
+    ensure!(
+        batch.num_columns() >= PHYSICAL_REGION_ESSENTIAL_COLUMN_COUNT,
+        error::InternalSnafu {
+            err_msg: format!(
+                "Expected at least {} columns in physical batch, got {}",
+                PHYSICAL_REGION_ESSENTIAL_COLUMN_COUNT,
+                batch.num_columns()
+            ),
+        }
+    );
+    let essential_indices: Vec<usize> = (0..PHYSICAL_REGION_ESSENTIAL_COLUMN_COUNT).collect();
+    batch.project(&essential_indices).context(error::ArrowSnafu)
+}
+
+async fn flush_region_writes_concurrently(
+    node_manager: &(impl PhysicalFlushNodeRequester + ?Sized),
+    writes: Vec<FlushRegionWrite>,
+) -> Result<usize> {
+    let mut affected_rows = 0;
+    if !should_dispatch_concurrently(writes.len()) {
+        for write in writes {
+            let _timer = PENDING_ROWS_BATCH_FLUSH_STAGE_ELAPSED
+                .with_label_values(&["flush_write_region"])
+                .start_timer();
+            affected_rows += node_manager
+                .handle(&write.datanode, write.request)
+                .await?
+                .affected_rows;
+        }
+        return Ok(affected_rows);
+    }
+
+    let write_futures = writes.into_iter().map(|write| async move {
+        let _timer = PENDING_ROWS_BATCH_FLUSH_STAGE_ELAPSED
+            .with_label_values(&["flush_write_region"])
+            .start_timer();
+
+        let response = node_manager.handle(&write.datanode, write.request).await?;
+        Ok::<_, Error>(response.affected_rows)
+    });
+
+    // todo(hl): should be bounded.
+    let affected_rows = futures::future::try_join_all(write_futures)
+        .await?
+        .into_iter()
+        .sum();
+    Ok(affected_rows)
+}
+
+async fn flush_batch(
+    flush: FlushBatch,
+    partition_manager: PartitionRuleManagerRef,
+    node_manager: NodeManagerRef,
+    catalog_manager: CatalogManagerRef,
+) {
+    let FlushBatch {
+        table_batches,
+        total_row_count,
+        db_string,
+        ctx,
+        waiters,
+    } = flush;
+    let start = Instant::now();
+
+    // Physical-table-level flush: transform all logical table batches
+    // into physical format and write them together.
+    let physical_table_name = ctx
+        .extension(PHYSICAL_TABLE_KEY)
+        .unwrap_or(GREPTIME_PHYSICAL_TABLE)
+        .to_string();
+    let partition_provider = PartitionManagerPhysicalFlushAdapter { partition_manager };
+    let node_requester = NodeManagerPhysicalFlushAdapter { node_manager };
+    let catalog_provider = CatalogManagerPhysicalFlushAdapter { catalog_manager };
+    let result = flush_batch_physical(
+        &table_batches,
+        &physical_table_name,
+        &ctx,
+        &partition_provider,
+        &node_requester,
+        &catalog_provider,
+    )
+    .await;
+
+    let elapsed = start.elapsed().as_secs_f64();
+    FLUSH_ELAPSED.observe(elapsed);
+
+    if result.is_err() {
+        FLUSH_FAILURES.inc();
+        FLUSH_DROPPED_ROWS.inc_by(total_row_count as u64);
+    } else if let Ok(affected_rows) = &result {
+        FLUSH_TOTAL.inc();
+        FLUSH_ROWS.observe(total_row_count as f64);
+        operator::metrics::DIST_INGEST_ROW_COUNT
+            .with_label_values(&[db_string.as_str()])
+            .inc_by(*affected_rows as u64);
+    }
+
+    debug!(
+        "Pending rows batch flushed, total rows: {}, elapsed time: {}s",
+        total_row_count, elapsed
+    );
+
+    notify_waiters(waiters, result.map(|_| ()));
+}
+
+/// Flushes a batch of logical table rows by transforming them into the physical table format
+/// and writing them to the appropriate datanode regions.
+///
+/// This function performs the end-to-end physical flush pipeline:
+/// 1. Resolves the physical table metadata and column ID mapping.
+/// 2. Fetches the physical table's partition rule.
+/// 3. Transforms each logical table batch into the physical (sparse primary key) format.
+/// 4. Concatenates all transformed batches into a single combined batch.
+/// 5. Splits the combined batch by partition rule and sends region write requests
+///    concurrently to the target datanodes.
+pub async fn flush_batch_physical(
+    table_batches: &[TableBatch],
+    physical_table_name: &str,
+    ctx: &QueryContextRef,
+    partition_manager: &(impl PhysicalFlushPartitionProvider + ?Sized),
+    node_manager: &(impl PhysicalFlushNodeRequester + ?Sized),
+    catalog_manager: &(impl PhysicalFlushCatalogProvider + ?Sized),
+) -> Result<usize> {
+    // 1. Resolve the physical table and get column ID mapping
+    let physical_table = {
+        let _timer = PENDING_ROWS_BATCH_FLUSH_STAGE_ELAPSED
+            .with_label_values(&["flush_physical_resolve_table"])
+            .start_timer();
+        catalog_manager
+            .physical_table(
+                ctx.current_catalog(),
+                &ctx.current_schema(),
+                physical_table_name,
+                ctx.as_ref(),
+            )
+            .await?
+            .with_context(|| error::InternalSnafu {
+                err_msg: format!(
+                    "Physical table '{}' not found during pending flush",
+                    physical_table_name
+                ),
+            })?
+    };
+
+    let physical_table_info = physical_table.table_info;
+    let name_to_ids = physical_table
+        .col_name_to_ids
+        .with_context(|| error::InternalSnafu {
+            err_msg: format!(
+                "Physical table '{}' has no column IDs for pending flush",
+                physical_table_name
+            ),
+        })?;
+
+    // 2. Get the physical table's partition rule (one lookup instead of N)
+    let partition_rule = {
+        let _timer = PENDING_ROWS_BATCH_FLUSH_STAGE_ELAPSED
+            .with_label_values(&["flush_physical_fetch_partition_rule"])
+            .start_timer();
+        partition_manager
+            .find_table_partition_rule(physical_table_info.as_ref())
+            .await?
+    };
+    let partition_columns = partition_rule.partition_columns();
+    let partition_columns_set: HashSet<&str> =
+        partition_columns.iter().map(String::as_str).collect();
+
+    // 3. Transform each logical table batch into physical format
+    let modified_batches =
+        transform_logical_batches_to_physical(table_batches, &name_to_ids, &partition_columns_set)?;
+
+    // 4. Concatenate all modified batches (all share the same physical schema)
+    let combined_batch = concat_modified_batches(&modified_batches)?;
+
+    // 5. Split by physical partition rule and send to regions
+    let physical_table_id = physical_table_info.table_id();
+    let planned_batches = plan_region_batches(
+        combined_batch,
+        physical_table_id,
+        partition_rule.as_ref(),
+        partition_columns,
+    )?;
+
+    let resolved_batches = resolve_region_targets(planned_batches, partition_manager).await?;
+    let region_writes = encode_region_write_requests(resolved_batches)?;
+    flush_region_writes_concurrently(node_manager, region_writes).await
+}
+
+/// Transforms logical table batches into physical format (sparse primary key encoding).
+///
+/// It identifies tag columns and essential columns (timestamp, value) for each logical batch
+/// and applies sparse primary key modification.
+fn transform_logical_batches_to_physical(
+    table_batches: &[TableBatch],
+    name_to_ids: &HashMap<String, u32>,
+    partition_columns_set: &HashSet<&str>,
+) -> Result<Vec<RecordBatch>> {
+    let mut modified_batches: Vec<RecordBatch> =
+        Vec::with_capacity(table_batches.iter().map(|b| b.batches.len()).sum());
+
+    let mut modify_elapsed = Duration::ZERO;
+    let mut columns_taxonomy_elapsed = Duration::ZERO;
+
+    for table_batch in table_batches {
+        let table_id = table_batch.table_id;
+
+        for batch in &table_batch.batches {
+            let batch_schema = batch.schema();
+            let start = Instant::now();
+            let (tag_columns, essential_col_indices) = columns_taxonomy(
+                &batch_schema,
+                &table_batch.table_name,
+                name_to_ids,
+                partition_columns_set,
+            )?;
+
+            columns_taxonomy_elapsed += start.elapsed();
+            if tag_columns.is_empty() && essential_col_indices.is_empty() {
+                continue;
+            }
+
+            let modified = {
+                let start = Instant::now();
+                let batch = modify_batch_sparse(
+                    batch.clone(),
+                    table_id,
+                    &tag_columns,
+                    &essential_col_indices,
+                )?;
+                modify_elapsed += start.elapsed();
+                batch
+            };
+
+            modified_batches.push(modified);
+        }
+    }
+
+    PENDING_ROWS_BATCH_FLUSH_STAGE_ELAPSED
+        .with_label_values(&["flush_physical_modify_batch"])
+        .observe(modify_elapsed.as_secs_f64());
+    PENDING_ROWS_BATCH_FLUSH_STAGE_ELAPSED
+        .with_label_values(&["flush_physical_columns_taxonomy"])
+        .observe(columns_taxonomy_elapsed.as_secs_f64());
+
+    ensure!(
+        !modified_batches.is_empty(),
+        error::InternalSnafu {
+            err_msg: "No batches can be transformed during pending flush",
+        }
+    );
+    Ok(modified_batches)
+}
+
+/// Concatenates all modified batches into a single large batch.
+///
+/// All modified batches share the same physical schema.
+fn concat_modified_batches(modified_batches: &[RecordBatch]) -> Result<RecordBatch> {
+    let combined_schema = modified_batches[0].schema();
+    let _timer = PENDING_ROWS_BATCH_FLUSH_STAGE_ELAPSED
+        .with_label_values(&["flush_physical_concat_all"])
+        .start_timer();
+    concat_batches(&combined_schema, modified_batches).context(error::ArrowSnafu)
+}
+
+fn split_combined_batch_by_region(
+    combined_batch: &RecordBatch,
+    partition_rule: &dyn partition::partition::PartitionRule,
+) -> Result<HashMap<u32, partition::partition::RegionMask>> {
+    let _timer = PENDING_ROWS_BATCH_FLUSH_STAGE_ELAPSED
+        .with_label_values(&["flush_physical_split_record_batch"])
+        .start_timer();
+    let map = partition_rule.split_record_batch(combined_batch)?;
+    Ok(map)
+}
+
+fn prepare_physical_region_routing_batch(
+    combined_batch: RecordBatch,
+    partition_columns: &[String],
+) -> Result<RecordBatch> {
+    if partition_columns.is_empty() {
+        return Ok(combined_batch);
+    }
+    strip_partition_columns_from_batch(combined_batch)
+}
+
+fn plan_region_batch(
+    stripped_batch: &RecordBatch,
+    physical_table_id: TableId,
+    region_number: u32,
+    mask: &partition::partition::RegionMask,
+) -> Result<Option<PlannedRegionBatch>> {
+    if mask.select_none() {
+        return Ok(None);
+    }
+
+    let region_batch = if mask.select_all() {
+        stripped_batch.clone()
+    } else {
+        let _timer = PENDING_ROWS_BATCH_FLUSH_STAGE_ELAPSED
+            .with_label_values(&["flush_physical_filter_record_batch"])
+            .start_timer();
+        filter_record_batch(stripped_batch, mask.array()).context(error::ArrowSnafu)?
+    };
+
+    let row_count = region_batch.num_rows();
+    if row_count == 0 {
+        return Ok(None);
+    }
+
+    Ok(Some(PlannedRegionBatch {
+        region_id: RegionId::new(physical_table_id, region_number),
+        batch: region_batch,
+    }))
+}
+
+fn plan_region_batches(
+    combined_batch: RecordBatch,
+    physical_table_id: TableId,
+    partition_rule: &dyn partition::partition::PartitionRule,
+    partition_columns: &[String],
+) -> Result<Vec<PlannedRegionBatch>> {
+    let region_masks = split_combined_batch_by_region(&combined_batch, partition_rule)?;
+    let stripped_batch = prepare_physical_region_routing_batch(combined_batch, partition_columns)?;
+
+    let mut planned_batches = Vec::new();
+    for (region_number, mask) in region_masks {
+        if let Some(planned_batch) =
+            plan_region_batch(&stripped_batch, physical_table_id, region_number, &mask)?
+        {
+            planned_batches.push(planned_batch);
+        }
+    }
+
+    Ok(planned_batches)
+}
+
+async fn resolve_region_targets(
+    planned_batches: Vec<PlannedRegionBatch>,
+    partition_manager: &(impl PhysicalFlushPartitionProvider + ?Sized),
+) -> Result<Vec<ResolvedRegionBatch>> {
+    let mut resolved_batches = Vec::with_capacity(planned_batches.len());
+    for planned in planned_batches {
+        let datanode = {
+            let _timer = PENDING_ROWS_BATCH_FLUSH_STAGE_ELAPSED
+                .with_label_values(&["flush_physical_resolve_region_leader"])
+                .start_timer();
+            partition_manager
+                .find_region_leader(planned.region_id)
+                .await?
+        };
+
+        resolved_batches.push(ResolvedRegionBatch { planned, datanode });
+    }
+
+    Ok(resolved_batches)
+}
+
+fn encode_region_write_requests(
+    resolved_batches: Vec<ResolvedRegionBatch>,
+) -> Result<Vec<FlushRegionWrite>> {
+    let mut region_writes = Vec::with_capacity(resolved_batches.len());
+    for resolved in resolved_batches {
+        let region_id = resolved.planned.region_id;
+        let (schema_bytes, data_header, payload) = {
+            let _timer = PENDING_ROWS_BATCH_FLUSH_STAGE_ELAPSED
+                .with_label_values(&["flush_physical_encode_ipc"])
+                .start_timer();
+            record_batch_to_ipc(resolved.planned.batch)?
+        };
+
+        let request = RegionRequest {
+            header: Some(RegionRequestHeader {
+                tracing_context: TracingContext::from_current_span().to_w3c(),
+                ..Default::default()
+            }),
+            body: Some(region_request::Body::BulkInsert(BulkInsertRequest {
+                region_id: region_id.as_u64(),
+                partition_expr_version: None,
+                body: Some(bulk_insert_request::Body::ArrowIpc(ArrowIpc {
+                    schema: schema_bytes,
+                    data_header,
+                    payload,
+                })),
+            })),
+        };
+
+        region_writes.push(FlushRegionWrite {
+            datanode: resolved.datanode,
+            request,
+        });
+    }
+
+    Ok(region_writes)
+}
+
+fn notify_waiters(waiters: Vec<FlushWaiter>, result: Result<()>) {
+    let shared_result = result.map_err(Arc::new);
+    for waiter in waiters {
+        let _ = waiter.response_tx.send(match &shared_result {
+            Ok(()) => Ok(()),
+            Err(error) => Err(Arc::clone(error)),
+        });
+        // waiter._permit is dropped here, releasing the inflight semaphore slot
+    }
+}
+
+fn record_batch_to_ipc(record_batch: RecordBatch) -> Result<(Bytes, Bytes, Bytes)> {
+    let mut encoder = FlightEncoder::default();
+    let schema = encoder.encode_schema(record_batch.schema().as_ref());
+    let mut iter = encoder
+        .encode(FlightMessage::RecordBatch(record_batch))
+        .into_iter();
+    let Some(flight_data) = iter.next() else {
+        return Err(Error::Internal {
+            err_msg: "Failed to encode empty flight data".to_string(),
+        });
+    };
+    if iter.next().is_some() {
+        return Err(Error::NotSupported {
+            feat: "bulk insert RecordBatch with dictionary arrays".to_string(),
+        });
+    }
+
+    Ok((
+        schema.data_header,
+        flight_data.data_header,
+        flight_data.data_body,
+    ))
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::{HashMap, HashSet};
+    use std::sync::Arc;
+    use std::sync::atomic::{AtomicUsize, Ordering};
+    use std::time::{Duration, Instant};
+
+    use api::region::RegionResponse;
+    use api::v1::flow::{DirtyWindowRequests, FlowRequest, FlowResponse};
+    use api::v1::meta::Peer;
+    use api::v1::region::{InsertRequests, RegionRequest, region_request};
+    use api::v1::{ColumnSchema, Row, RowInsertRequest, RowInsertRequests, Rows};
+    use arrow::array::{BinaryArray, BooleanArray, StringArray, TimestampMillisecondArray};
+    use arrow::datatypes::{DataType as ArrowDataType, Field, Schema as ArrowSchema};
+    use arrow::record_batch::RecordBatch;
+    use async_trait::async_trait;
+    use catalog::error::Result as CatalogResult;
+    use common_meta::error::Result as MetaResult;
+    use common_meta::node_manager::{
+        Datanode, DatanodeManager, DatanodeRef, Flownode, FlownodeManager, FlownodeRef,
+    };
+    use common_query::request::QueryRequest;
+    use common_recordbatch::SendableRecordBatchStream;
+    use dashmap::DashMap;
+    use datatypes::schema::{ColumnSchema as DtColumnSchema, Schema as DtSchema};
+    use partition::error::Result as PartitionResult;
+    use partition::partition::{PartitionRule, PartitionRuleRef, RegionMask};
+    use smallvec::SmallVec;
+    use snafu::ResultExt;
+    use store_api::storage::RegionId;
+    use table::metadata::TableId;
+    use table::test_util::table_info::test_table_info;
+    use tokio::sync::{Semaphore, mpsc, oneshot};
+    use tokio::time::sleep;
+
+    use super::{
+        BatchKey, Error, FlushRegionWrite, FlushWaiter, PendingBatch, PendingRowsBatcher,
+        PendingWorker, PhysicalFlushCatalogProvider, PhysicalFlushNodeRequester,
+        PhysicalFlushPartitionProvider, PhysicalTableMetadata, PlannedRegionBatch,
+        ResolvedRegionBatch, TableBatch, WorkerCommand, columns_taxonomy, drain_batch,
+        encode_region_write_requests, flush_batch_physical, flush_region_writes_concurrently,
+        plan_region_batches, remove_worker_if_same_channel, should_close_worker_on_idle_timeout,
+        should_dispatch_concurrently, strip_partition_columns_from_batch,
+        transform_logical_batches_to_physical,
+    };
+    use crate::error;
+
+    fn mock_rows(row_count: usize, schema_name: &str) -> Rows {
+        Rows {
+            schema: vec![ColumnSchema {
+                column_name: schema_name.to_string(),
+                ..Default::default()
+            }],
+            rows: (0..row_count).map(|_| Row { values: vec![] }).collect(),
+        }
+    }
+
+    fn mock_tag_batch(tag_name: &str, tag_value: &str, ts: i64, val: f64) -> RecordBatch {
+        let schema = Arc::new(ArrowSchema::new(vec![
+            Field::new(
+                "greptime_timestamp",
+                ArrowDataType::Timestamp(arrow::datatypes::TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new("greptime_value", ArrowDataType::Float64, true),
+            Field::new(tag_name, ArrowDataType::Utf8, true),
+        ]));
+
+        RecordBatch::try_new(
+            schema,
+            vec![
+                Arc::new(TimestampMillisecondArray::from(vec![ts])),
+                Arc::new(arrow::array::Float64Array::from(vec![val])),
+                Arc::new(StringArray::from(vec![tag_value])),
+            ],
+        )
+        .unwrap()
+    }
+
+    fn mock_physical_table_metadata(table_id: TableId) -> PhysicalTableMetadata {
+        let schema = Arc::new(
+            DtSchema::try_new(vec![
+                DtColumnSchema::new(
+                    "__primary_key",
+                    datatypes::prelude::ConcreteDataType::binary_datatype(),
+                    false,
+                ),
+                DtColumnSchema::new(
+                    "greptime_timestamp",
+                    datatypes::prelude::ConcreteDataType::timestamp_millisecond_datatype(),
+                    false,
+                ),
+                DtColumnSchema::new(
+                    "greptime_value",
+                    datatypes::prelude::ConcreteDataType::float64_datatype(),
+                    true,
+                ),
+                DtColumnSchema::new(
+                    "tag1",
+                    datatypes::prelude::ConcreteDataType::string_datatype(),
+                    true,
+                ),
+            ])
+            .unwrap(),
+        );
+        let mut table_info = test_table_info(table_id, "phy", "public", "greptime", schema);
+        table_info.meta.column_ids = vec![0, 1, 2, 3];
+
+        PhysicalTableMetadata {
+            table_info: Arc::new(table_info),
+            col_name_to_ids: Some(HashMap::from([("tag1".to_string(), 3)])),
+        }
+    }
+
+    struct MockFlushCatalogProvider {
+        table: Option<PhysicalTableMetadata>,
+    }
+
+    #[async_trait]
+    impl PhysicalFlushCatalogProvider for MockFlushCatalogProvider {
+        async fn physical_table(
+            &self,
+            _catalog: &str,
+            _schema: &str,
+            _table_name: &str,
+            _query_ctx: &session::context::QueryContext,
+        ) -> CatalogResult<Option<PhysicalTableMetadata>> {
+            Ok(self.table.clone())
+        }
+    }
+
+    struct SingleRegionPartitionRule;
+
+    impl PartitionRule for SingleRegionPartitionRule {
+        fn as_any(&self) -> &dyn std::any::Any {
+            self
+        }
+
+        fn partition_columns(&self) -> &[String] {
+            &[]
+        }
+
+        fn find_region(
+            &self,
+            _values: &[datatypes::prelude::Value],
+        ) -> partition::error::Result<store_api::storage::RegionNumber> {
+            unimplemented!()
+        }
+
+        fn split_record_batch(
+            &self,
+            record_batch: &RecordBatch,
+        ) -> partition::error::Result<HashMap<store_api::storage::RegionNumber, RegionMask>>
+        {
+            Ok(HashMap::from([(
+                1,
+                RegionMask::new(
+                    arrow::array::BooleanArray::from(vec![true; record_batch.num_rows()]),
+                    record_batch.num_rows(),
+                ),
+            )]))
+        }
+    }
+
+    struct TwoRegionPartitionRule {
+        partition_columns: Vec<String>,
+    }
+
+    impl PartitionRule for TwoRegionPartitionRule {
+        fn as_any(&self) -> &dyn std::any::Any {
+            self
+        }
+
+        fn partition_columns(&self) -> &[String] {
+            &self.partition_columns
+        }
+
+        fn find_region(
+            &self,
+            _values: &[datatypes::prelude::Value],
+        ) -> partition::error::Result<store_api::storage::RegionNumber> {
+            unimplemented!()
+        }
+
+        fn split_record_batch(
+            &self,
+            _record_batch: &RecordBatch,
+        ) -> partition::error::Result<HashMap<store_api::storage::RegionNumber, RegionMask>>
+        {
+            Ok(HashMap::from([
+                (1, RegionMask::new(BooleanArray::from(vec![true, false]), 1)),
+                (2, RegionMask::new(BooleanArray::from(vec![false, true]), 1)),
+                (
+                    3,
+                    RegionMask::new(BooleanArray::from(vec![false, false]), 0),
+                ),
+            ]))
+        }
+    }
+
+    struct MockFlushPartitionProvider {
+        partition_rule_calls: Arc<AtomicUsize>,
+        region_leader_calls: Arc<AtomicUsize>,
+    }
+
+    #[async_trait]
+    impl PhysicalFlushPartitionProvider for MockFlushPartitionProvider {
+        async fn find_table_partition_rule(
+            &self,
+            _table_info: &table::metadata::TableInfo,
+        ) -> PartitionResult<PartitionRuleRef> {
+            self.partition_rule_calls.fetch_add(1, Ordering::SeqCst);
+            Ok(Arc::new(SingleRegionPartitionRule))
+        }
+
+        async fn find_region_leader(&self, _region_id: RegionId) -> error::Result<Peer> {
+            self.region_leader_calls.fetch_add(1, Ordering::SeqCst);
+            Ok(Peer {
+                id: 1,
+                addr: "node-1".to_string(),
+            })
+        }
+    }
+
+    #[derive(Default)]
+    struct MockFlushNodeRequester {
+        writes: Arc<AtomicUsize>,
+    }
+
+    #[async_trait]
+    impl PhysicalFlushNodeRequester for MockFlushNodeRequester {
+        async fn handle(
+            &self,
+            _peer: &Peer,
+            _request: RegionRequest,
+        ) -> error::Result<RegionResponse> {
+            self.writes.fetch_add(1, Ordering::SeqCst);
+            Ok(RegionResponse::new(0))
+        }
+    }
+
+    #[test]
+    fn test_collect_non_empty_table_rows_filters_empty_payloads() {
+        let requests = RowInsertRequests {
+            inserts: vec![
+                RowInsertRequest {
+                    table_name: "cpu".to_string(),
+                    rows: Some(mock_rows(2, "host")),
+                },
+                RowInsertRequest {
+                    table_name: "mem".to_string(),
+                    rows: Some(mock_rows(0, "host")),
+                },
+                RowInsertRequest {
+                    table_name: "disk".to_string(),
+                    rows: None,
+                },
+            ],
+        };
+
+        let (table_rows, total_rows) = PendingRowsBatcher::collect_non_empty_table_rows(requests);
+
+        assert_eq!(2, total_rows);
+        assert_eq!(1, table_rows.len());
+        assert_eq!("cpu", table_rows[0].0);
+        assert_eq!(2, table_rows[0].1.rows.len());
+    }
+
+    #[test]
+    fn test_drain_batch_takes_initialized_pending_batch_from_option() {
+        let ctx = session::context::QueryContext::arc();
+        let (response_tx, _response_rx) = oneshot::channel();
+        let permit = Arc::new(Semaphore::new(1)).try_acquire_owned().unwrap();
+        let mut batch = Some(PendingBatch {
+            tables: HashMap::from([(
+                "cpu".to_string(),
+                TableBatch {
+                    table_name: "cpu".to_string(),
+                    table_id: 42,
+                    batches: vec![mock_tag_batch("tag1", "host-1", 1000, 1.0)],
+                    row_count: 1,
+                },
+            )]),
+            created_at: Instant::now(),
+            total_row_count: 1,
+            db_string: ctx.get_db_string(),
+            ctx: ctx.clone(),
+            waiters: vec![FlushWaiter {
+                response_tx,
+                _permit: permit,
+            }],
+        });
+
+        let flush = drain_batch(&mut batch).unwrap();
+
+        assert!(batch.is_none());
+        assert_eq!(1, flush.total_row_count);
+        assert_eq!(1, flush.table_batches.len());
+        assert_eq!(ctx.get_db_string(), flush.db_string);
+        assert_eq!(ctx.current_catalog(), flush.ctx.current_catalog());
+    }
+
+    #[derive(Clone)]
+    struct ConcurrentMockDatanode {
+        delay: Duration,
+        inflight: Arc<AtomicUsize>,
+        max_inflight: Arc<AtomicUsize>,
+    }
+
+    #[async_trait]
+    impl Datanode for ConcurrentMockDatanode {
+        async fn handle(&self, _request: RegionRequest) -> MetaResult<RegionResponse> {
+            let now = self.inflight.fetch_add(1, Ordering::SeqCst) + 1;
+            loop {
+                let max = self.max_inflight.load(Ordering::SeqCst);
+                if now <= max {
+                    break;
+                }
+                if self
+                    .max_inflight
+                    .compare_exchange(max, now, Ordering::SeqCst, Ordering::SeqCst)
+                    .is_ok()
+                {
+                    break;
+                }
+            }
+
+            sleep(self.delay).await;
+            self.inflight.fetch_sub(1, Ordering::SeqCst);
+            Ok(RegionResponse::new(0))
+        }
+
+        async fn handle_query(
+            &self,
+            _request: QueryRequest,
+        ) -> MetaResult<SendableRecordBatchStream> {
+            unimplemented!()
+        }
+    }
+
+    #[derive(Clone)]
+    struct ConcurrentMockNodeManager {
+        datanodes: Arc<HashMap<u64, DatanodeRef>>,
+    }
+
+    #[async_trait]
+    impl DatanodeManager for ConcurrentMockNodeManager {
+        async fn datanode(&self, node: &Peer) -> DatanodeRef {
+            self.datanodes
+                .get(&node.id)
+                .expect("datanode not found")
+                .clone()
+        }
+    }
+
+    struct NoopFlownode;
+
+    #[async_trait]
+    impl Flownode for NoopFlownode {
+        async fn handle(&self, _request: FlowRequest) -> MetaResult<FlowResponse> {
+            unimplemented!()
+        }
+
+        async fn handle_inserts(&self, _request: InsertRequests) -> MetaResult<FlowResponse> {
+            unimplemented!()
+        }
+
+        async fn handle_mark_window_dirty(
+            &self,
+            _req: DirtyWindowRequests,
+        ) -> MetaResult<FlowResponse> {
+            unimplemented!()
+        }
+    }
+
+    #[async_trait]
+    impl FlownodeManager for ConcurrentMockNodeManager {
+        async fn flownode(&self, _node: &Peer) -> FlownodeRef {
+            Arc::new(NoopFlownode)
+        }
+    }
+
+    #[async_trait]
+    impl PhysicalFlushNodeRequester for ConcurrentMockNodeManager {
+        async fn handle(
+            &self,
+            peer: &Peer,
+            request: RegionRequest,
+        ) -> error::Result<RegionResponse> {
+            let datanode = self.datanode(peer).await;
+            datanode
+                .handle(request)
+                .await
+                .context(error::CommonMetaSnafu)
+        }
+    }
+
+    #[test]
+    fn test_remove_worker_if_same_channel_removes_matching_entry() {
+        let workers = DashMap::new();
+        let key = BatchKey {
+            catalog: "greptime".to_string(),
+            schema: "public".to_string(),
+            physical_table: "phy".to_string(),
+        };
+
+        let (tx, _rx) = mpsc::channel::<WorkerCommand>(1);
+        workers.insert(key.clone(), PendingWorker { tx: tx.clone() });
+
+        assert!(remove_worker_if_same_channel(&workers, &key, &tx));
+        assert!(!workers.contains_key(&key));
+    }
+
+    #[test]
+    fn test_remove_worker_if_same_channel_keeps_newer_entry() {
+        let workers = DashMap::new();
+        let key = BatchKey {
+            catalog: "greptime".to_string(),
+            schema: "public".to_string(),
+            physical_table: "phy".to_string(),
+        };
+
+        let (stale_tx, _stale_rx) = mpsc::channel::<WorkerCommand>(1);
+        let (fresh_tx, _fresh_rx) = mpsc::channel::<WorkerCommand>(1);
+        workers.insert(
+            key.clone(),
+            PendingWorker {
+                tx: fresh_tx.clone(),
+            },
+        );
+
+        assert!(!remove_worker_if_same_channel(&workers, &key, &stale_tx));
+        assert!(workers.contains_key(&key));
+        assert!(workers.get(&key).unwrap().tx.same_channel(&fresh_tx));
+    }
+
+    #[test]
+    fn test_worker_idle_timeout_close_decision() {
+        assert!(should_close_worker_on_idle_timeout(0, 0));
+        assert!(!should_close_worker_on_idle_timeout(1, 0));
+        assert!(!should_close_worker_on_idle_timeout(0, 1));
+    }
+
+    #[tokio::test]
+    async fn test_flush_region_writes_concurrently_dispatches_multiple_datanodes() {
+        let inflight = Arc::new(AtomicUsize::new(0));
+        let max_inflight = Arc::new(AtomicUsize::new(0));
+        let datanode1: DatanodeRef = Arc::new(ConcurrentMockDatanode {
+            delay: Duration::from_millis(100),
+            inflight: inflight.clone(),
+            max_inflight: max_inflight.clone(),
+        });
+        let datanode2: DatanodeRef = Arc::new(ConcurrentMockDatanode {
+            delay: Duration::from_millis(100),
+            inflight,
+            max_inflight: max_inflight.clone(),
+        });
+
+        let mut datanodes = HashMap::new();
+        datanodes.insert(1, datanode1);
+        datanodes.insert(2, datanode2);
+        let node_manager = Arc::new(ConcurrentMockNodeManager {
+            datanodes: Arc::new(datanodes),
+        });
+
+        let writes = vec![
+            FlushRegionWrite {
+                datanode: Peer {
+                    id: 1,
+                    addr: "node1".to_string(),
+                },
+                request: RegionRequest::default(),
+            },
+            FlushRegionWrite {
+                datanode: Peer {
+                    id: 2,
+                    addr: "node2".to_string(),
+                },
+                request: RegionRequest::default(),
+            },
+        ];
+
+        flush_region_writes_concurrently(node_manager.as_ref(), writes)
+            .await
+            .unwrap();
+        assert!(max_inflight.load(Ordering::SeqCst) >= 2);
+    }
+
+    #[test]
+    fn test_should_dispatch_concurrently_by_region_count() {
+        assert!(!should_dispatch_concurrently(0));
+        assert!(!should_dispatch_concurrently(1));
+        assert!(should_dispatch_concurrently(2));
+    }
+
+    #[test]
+    fn test_strip_partition_columns_from_batch_removes_partition_tags() {
+        let batch = RecordBatch::try_new(
+            Arc::new(ArrowSchema::new(vec![
+                Field::new("__primary_key", ArrowDataType::Binary, false),
+                Field::new(
+                    "greptime_timestamp",
+                    ArrowDataType::Timestamp(arrow::datatypes::TimeUnit::Millisecond, None),
+                    false,
+                ),
+                Field::new("greptime_value", ArrowDataType::Float64, true),
+                Field::new("host", ArrowDataType::Utf8, true),
+            ])),
+            vec![
+                Arc::new(BinaryArray::from(vec![b"k1".as_slice()])),
+                Arc::new(TimestampMillisecondArray::from(vec![1000_i64])),
+                Arc::new(arrow::array::Float64Array::from(vec![42.0_f64])),
+                Arc::new(StringArray::from(vec!["node-1"])),
+            ],
+        )
+        .unwrap();
+
+        let stripped = strip_partition_columns_from_batch(batch).unwrap();
+
+        assert_eq!(3, stripped.num_columns());
+        assert_eq!("__primary_key", stripped.schema().field(0).name());
+        assert_eq!("greptime_timestamp", stripped.schema().field(1).name());
+        assert_eq!("greptime_value", stripped.schema().field(2).name());
+    }
+
+    #[test]
+    fn test_strip_partition_columns_from_batch_projects_essential_columns_without_lookup() {
+        let batch = RecordBatch::try_new(
+            Arc::new(ArrowSchema::new(vec![
+                Field::new("__primary_key", ArrowDataType::Binary, false),
+                Field::new(
+                    "greptime_timestamp",
+                    ArrowDataType::Timestamp(arrow::datatypes::TimeUnit::Millisecond, None),
+                    false,
+                ),
+                Field::new("greptime_value", ArrowDataType::Float64, true),
+                Field::new("host", ArrowDataType::Utf8, true),
+            ])),
+            vec![
+                Arc::new(BinaryArray::from(vec![b"k1".as_slice()])),
+                Arc::new(TimestampMillisecondArray::from(vec![1000_i64])),
+                Arc::new(arrow::array::Float64Array::from(vec![42.0_f64])),
+                Arc::new(StringArray::from(vec!["node-1"])),
+            ],
+        )
+        .unwrap();
+
+        let stripped = strip_partition_columns_from_batch(batch).unwrap();
+
+        assert_eq!(3, stripped.num_columns());
+        assert_eq!("__primary_key", stripped.schema().field(0).name());
+        assert_eq!("greptime_timestamp", stripped.schema().field(1).name());
+        assert_eq!("greptime_value", stripped.schema().field(2).name());
+    }
+
+    #[test]
+    fn test_collect_tag_columns_and_non_tag_indices_keeps_partition_tag_column() {
+        let schema = Arc::new(ArrowSchema::new(vec![
+            Field::new(
+                "greptime_timestamp",
+                ArrowDataType::Timestamp(arrow::datatypes::TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new("greptime_value", ArrowDataType::Float64, true),
+            Field::new("host", ArrowDataType::Utf8, true),
+            Field::new("region", ArrowDataType::Utf8, true),
+        ]));
+        let name_to_ids =
+            HashMap::from([("host".to_string(), 1_u32), ("region".to_string(), 2_u32)]);
+        let partition_columns = HashSet::from(["host"]);
+
+        let (tag_columns, non_tag_indices) =
+            columns_taxonomy(&schema, "cpu", &name_to_ids, &partition_columns).unwrap();
+
+        assert_eq!(2, tag_columns.len());
+        assert_eq!(&[0, 1, 2], non_tag_indices.as_slice());
+    }
+
+    #[test]
+    fn test_collect_tag_columns_and_non_tag_indices_prioritizes_essential_columns() {
+        let schema = Arc::new(ArrowSchema::new(vec![
+            Field::new("host", ArrowDataType::Utf8, true),
+            Field::new("greptime_value", ArrowDataType::Float64, true),
+            Field::new(
+                "greptime_timestamp",
+                ArrowDataType::Timestamp(arrow::datatypes::TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new("region", ArrowDataType::Utf8, true),
+        ]));
+        let name_to_ids =
+            HashMap::from([("host".to_string(), 1_u32), ("region".to_string(), 2_u32)]);
+        let partition_columns = HashSet::from(["host", "region"]);
+
+        let (_tag_columns, non_tag_indices): (_, SmallVec<[usize; 3]>) =
+            columns_taxonomy(&schema, "cpu", &name_to_ids, &partition_columns).unwrap();
+
+        assert_eq!(&[2, 1, 0, 3], non_tag_indices.as_slice());
+    }
+
+    #[test]
+    fn test_collect_tag_columns_and_non_tag_indices_rejects_unexpected_data_type() {
+        let schema = Arc::new(ArrowSchema::new(vec![
+            Field::new(
+                "greptime_timestamp",
+                ArrowDataType::Timestamp(arrow::datatypes::TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new("greptime_value", ArrowDataType::Float64, true),
+            Field::new("host", ArrowDataType::Utf8, true),
+            Field::new("invalid", ArrowDataType::Boolean, true),
+        ]));
+        let name_to_ids = HashMap::from([("host".to_string(), 1_u32)]);
+        let partition_columns = HashSet::from(["host"]);
+
+        let result = columns_taxonomy(&schema, "cpu", &name_to_ids, &partition_columns);
+
+        assert!(matches!(
+            result,
+            Err(Error::InvalidPromRemoteRequest { .. })
+        ));
+    }
+
+    #[test]
+    fn test_collect_tag_columns_and_non_tag_indices_rejects_int64_timestamp_column() {
+        let schema = Arc::new(ArrowSchema::new(vec![
+            Field::new("greptime_timestamp", ArrowDataType::Int64, false),
+            Field::new("greptime_value", ArrowDataType::Float64, true),
+            Field::new("host", ArrowDataType::Utf8, true),
+        ]));
+        let name_to_ids = HashMap::from([("host".to_string(), 1_u32)]);
+        let partition_columns = HashSet::from(["host"]);
+
+        let result = columns_taxonomy(&schema, "cpu", &name_to_ids, &partition_columns);
+
+        assert!(matches!(
+            result,
+            Err(Error::InvalidPromRemoteRequest { .. })
+        ));
+    }
+
+    #[test]
+    fn test_collect_tag_columns_and_non_tag_indices_rejects_duplicated_timestamp_column() {
+        let schema = Arc::new(ArrowSchema::new(vec![
+            Field::new(
+                "ts1",
+                ArrowDataType::Timestamp(arrow::datatypes::TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new(
+                "ts2",
+                ArrowDataType::Timestamp(arrow::datatypes::TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new("greptime_value", ArrowDataType::Float64, true),
+            Field::new("host", ArrowDataType::Utf8, true),
+        ]));
+        let name_to_ids = HashMap::from([("host".to_string(), 1_u32)]);
+        let partition_columns = HashSet::from(["host"]);
+
+        let result = columns_taxonomy(&schema, "cpu", &name_to_ids, &partition_columns);
+
+        assert!(matches!(
+            result,
+            Err(Error::InvalidPromRemoteRequest { .. })
+        ));
+    }
+
+    #[test]
+    fn test_collect_tag_columns_and_non_tag_indices_rejects_duplicated_value_column() {
+        let schema = Arc::new(ArrowSchema::new(vec![
+            Field::new(
+                "greptime_timestamp",
+                ArrowDataType::Timestamp(arrow::datatypes::TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new("value1", ArrowDataType::Float64, true),
+            Field::new("value2", ArrowDataType::Float64, true),
+            Field::new("host", ArrowDataType::Utf8, true),
+        ]));
+        let name_to_ids = HashMap::from([("host".to_string(), 1_u32)]);
+        let partition_columns = HashSet::from(["host"]);
+
+        let result = columns_taxonomy(&schema, "cpu", &name_to_ids, &partition_columns);
+
+        assert!(matches!(
+            result,
+            Err(Error::InvalidPromRemoteRequest { .. })
+        ));
+    }
+
+    #[test]
+    fn test_modify_batch_sparse_with_taxonomy_per_batch() {
+        use arrow::array::BinaryArray;
+        use metric_engine::batch_modifier::modify_batch_sparse;
+
+        let schema1 = Arc::new(ArrowSchema::new(vec![
+            Field::new(
+                "greptime_timestamp",
+                ArrowDataType::Timestamp(arrow::datatypes::TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new("greptime_value", ArrowDataType::Float64, true),
+            Field::new("tag1", ArrowDataType::Utf8, true),
+        ]));
+
+        let schema2 = Arc::new(ArrowSchema::new(vec![
+            Field::new(
+                "greptime_timestamp",
+                ArrowDataType::Timestamp(arrow::datatypes::TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new("greptime_value", ArrowDataType::Float64, true),
+            Field::new("tag1", ArrowDataType::Utf8, true),
+            Field::new("tag2", ArrowDataType::Utf8, true),
+        ]));
+        let batch2 = RecordBatch::try_new(
+            schema2.clone(),
+            vec![
+                Arc::new(TimestampMillisecondArray::from(vec![2000])),
+                Arc::new(arrow::array::Float64Array::from(vec![2.0])),
+                Arc::new(StringArray::from(vec!["v1"])),
+                Arc::new(StringArray::from(vec!["v2"])),
+            ],
+        )
+        .unwrap();
+
+        let name_to_ids = HashMap::from([("tag1".to_string(), 1), ("tag2".to_string(), 2)]);
+        let partition_columns = HashSet::new();
+
+        // A batch that only has tag1, same values as batch2 for ts and val.
+        let batch3 = RecordBatch::try_new(
+            schema1.clone(),
+            vec![
+                Arc::new(TimestampMillisecondArray::from(vec![2000])),
+                Arc::new(arrow::array::Float64Array::from(vec![2.0])),
+                Arc::new(StringArray::from(vec!["v1"])),
+            ],
+        )
+        .unwrap();
+
+        // Simulate the new loop logic in flush_batch_physical:
+        // Resolve taxonomy FOR EACH BATCH.
+        let (tag_columns2, indices2) =
+            columns_taxonomy(&batch2.schema(), "table", &name_to_ids, &partition_columns).unwrap();
+        let modified2 = modify_batch_sparse(batch2, 123, &tag_columns2, &indices2).unwrap();
+
+        let (tag_columns3, indices3) =
+            columns_taxonomy(&batch3.schema(), "table", &name_to_ids, &partition_columns).unwrap();
+        let modified3 = modify_batch_sparse(batch3, 123, &tag_columns3, &indices3).unwrap();
+
+        let pk2 = modified2
+            .column(0)
+            .as_any()
+            .downcast_ref::<BinaryArray>()
+            .unwrap();
+        let pk3 = modified3
+            .column(0)
+            .as_any()
+            .downcast_ref::<BinaryArray>()
+            .unwrap();
+
+        // Now they SHOULD be different because tag2 is included in pk2 but not in pk3.
+        assert_ne!(
+            pk2.value(0),
+            pk3.value(0),
+            "PK should be different because batch2 has tag2!"
+        );
+    }
+
+    #[test]
+    fn test_transform_logical_batches_to_physical_success() {
+        let batch = mock_tag_batch("tag1", "v1", 1000, 1.0);
+
+        let table_batches = vec![TableBatch {
+            table_name: "t1".to_string(),
+            table_id: 1,
+            batches: vec![batch],
+            row_count: 1,
+        }];
+
+        let name_to_ids = HashMap::from([("tag1".to_string(), 1)]);
+        let partition_columns = HashSet::new();
+        let modified =
+            transform_logical_batches_to_physical(&table_batches, &name_to_ids, &partition_columns)
+                .unwrap();
+
+        assert_eq!(1, modified.len());
+        assert_eq!(3, modified[0].num_columns());
+        assert_eq!("__primary_key", modified[0].schema().field(0).name());
+        assert_eq!("greptime_timestamp", modified[0].schema().field(1).name());
+        assert_eq!("greptime_value", modified[0].schema().field(2).name());
+    }
+
+    #[test]
+    fn test_transform_logical_batches_to_physical_taxonomy_failure() {
+        let batch = mock_tag_batch("tag1", "v1", 1000, 1.0);
+
+        let table_batches = vec![TableBatch {
+            table_name: "t1".to_string(),
+            table_id: 1,
+            batches: vec![batch],
+            row_count: 1,
+        }];
+
+        // tag1 is missing from name_to_ids, causing columns_taxonomy to fail.
+        let name_to_ids = HashMap::new();
+        let partition_columns = HashSet::new();
+        let err =
+            transform_logical_batches_to_physical(&table_batches, &name_to_ids, &partition_columns)
+                .unwrap_err();
+
+        assert!(
+            err.to_string()
+                .contains("not found in physical table column IDs")
+        );
+    }
+
+    #[test]
+    fn test_transform_logical_batches_to_physical_multiple_batches() {
+        let batch1 = mock_tag_batch("tag1", "v1", 1000, 1.0);
+        let batch2 = mock_tag_batch("tag2", "v2", 2000, 2.0);
+
+        let table_batches = vec![
+            TableBatch {
+                table_name: "t1".to_string(),
+                table_id: 1,
+                batches: vec![batch1],
+                row_count: 1,
+            },
+            TableBatch {
+                table_name: "t2".to_string(),
+                table_id: 2,
+                batches: vec![batch2],
+                row_count: 1,
+            },
+        ];
+
+        let name_to_ids = HashMap::from([("tag1".to_string(), 1), ("tag2".to_string(), 2)]);
+        let partition_columns = HashSet::new();
+        let modified =
+            transform_logical_batches_to_physical(&table_batches, &name_to_ids, &partition_columns)
+                .unwrap();
+
+        assert_eq!(2, modified.len());
+    }
+
+    #[test]
+    fn test_transform_logical_batches_to_physical_mixed_success_failure() {
+        let batch1 = mock_tag_batch("tag1", "v1", 1000, 1.0);
+        let batch2 = mock_tag_batch("tag2", "v2", 2000, 2.0);
+
+        let table_batches = vec![
+            TableBatch {
+                table_name: "t1".to_string(),
+                table_id: 1,
+                batches: vec![batch1],
+                row_count: 1,
+            },
+            TableBatch {
+                table_name: "t2".to_string(),
+                table_id: 2,
+                batches: vec![batch2],
+                row_count: 1,
+            },
+        ];
+
+        // tag1 is missing from name_to_ids, causing batch1 to fail.
+        let name_to_ids = HashMap::from([("tag2".to_string(), 2)]);
+        let partition_columns = HashSet::new();
+        let err =
+            transform_logical_batches_to_physical(&table_batches, &name_to_ids, &partition_columns)
+                .unwrap_err();
+
+        assert!(err.to_string().contains("tag1"));
+    }
+
+    #[tokio::test]
+    async fn test_flush_batch_physical_uses_mockable_trait_dependencies() {
+        let table_batches = vec![TableBatch {
+            table_name: "t1".to_string(),
+            table_id: 11,
+            batches: vec![mock_tag_batch("tag1", "host-1", 1000, 1.0)],
+            row_count: 1,
+        }];
+        let partition_calls = Arc::new(AtomicUsize::new(0));
+        let leader_calls = Arc::new(AtomicUsize::new(0));
+        let node = MockFlushNodeRequester::default();
+        let ctx = session::context::QueryContext::arc();
+
+        flush_batch_physical(
+            &table_batches,
+            "phy",
+            &ctx,
+            &MockFlushPartitionProvider {
+                partition_rule_calls: partition_calls.clone(),
+                region_leader_calls: leader_calls.clone(),
+            },
+            &node,
+            &MockFlushCatalogProvider {
+                table: Some(mock_physical_table_metadata(1024)),
+            },
+        )
+        .await
+        .unwrap();
+
+        assert_eq!(1, partition_calls.load(Ordering::SeqCst));
+        assert_eq!(1, leader_calls.load(Ordering::SeqCst));
+        assert_eq!(1, node.writes.load(Ordering::SeqCst));
+    }
+
+    #[derive(Default)]
+    struct AffectedRowsFlushNodeRequester {
+        affected_rows: usize,
+    }
+
+    #[async_trait]
+    impl PhysicalFlushNodeRequester for AffectedRowsFlushNodeRequester {
+        async fn handle(
+            &self,
+            _peer: &Peer,
+            _request: RegionRequest,
+        ) -> error::Result<RegionResponse> {
+            Ok(RegionResponse::new(self.affected_rows))
+        }
+    }
+
+    #[tokio::test]
+    async fn test_flush_batch_physical_returns_actual_affected_rows() {
+        let table_batches = vec![TableBatch {
+            table_name: "t1".to_string(),
+            table_id: 11,
+            batches: vec![mock_tag_batch("tag1", "host-1", 1000, 1.0)],
+            row_count: 1,
+        }];
+        let ctx = session::context::QueryContext::arc();
+
+        let affected_rows = flush_batch_physical(
+            &table_batches,
+            "phy",
+            &ctx,
+            &MockFlushPartitionProvider {
+                partition_rule_calls: Arc::new(AtomicUsize::new(0)),
+                region_leader_calls: Arc::new(AtomicUsize::new(0)),
+            },
+            &AffectedRowsFlushNodeRequester { affected_rows: 7 },
+            &MockFlushCatalogProvider {
+                table: Some(mock_physical_table_metadata(1024)),
+            },
+        )
+        .await
+        .unwrap();
+
+        assert_eq!(7, affected_rows);
+    }
+
+    #[tokio::test]
+    async fn test_flush_batch_physical_stops_before_partition_and_node_when_table_missing() {
+        let table_batches = vec![TableBatch {
+            table_name: "t1".to_string(),
+            table_id: 11,
+            batches: vec![mock_tag_batch("tag1", "host-1", 1000, 1.0)],
+            row_count: 1,
+        }];
+        let partition_calls = Arc::new(AtomicUsize::new(0));
+        let leader_calls = Arc::new(AtomicUsize::new(0));
+        let node = MockFlushNodeRequester::default();
+        let ctx = session::context::QueryContext::arc();
+
+        let err = flush_batch_physical(
+            &table_batches,
+            "missing_phy",
+            &ctx,
+            &MockFlushPartitionProvider {
+                partition_rule_calls: partition_calls.clone(),
+                region_leader_calls: leader_calls.clone(),
+            },
+            &node,
+            &MockFlushCatalogProvider { table: None },
+        )
+        .await
+        .unwrap_err();
+
+        assert!(
+            err.to_string()
+                .contains("Physical table 'missing_phy' not found")
+        );
+        assert_eq!(0, partition_calls.load(Ordering::SeqCst));
+        assert_eq!(0, leader_calls.load(Ordering::SeqCst));
+        assert_eq!(0, node.writes.load(Ordering::SeqCst));
+    }
+
+    #[tokio::test]
+    async fn test_flush_batch_physical_aborts_immediately_on_transform_error() {
+        let table_batches = vec![
+            TableBatch {
+                table_name: "broken".to_string(),
+                table_id: 11,
+                batches: vec![mock_tag_batch("unknown_tag", "host-1", 1000, 1.0)],
+                row_count: 1,
+            },
+            TableBatch {
+                table_name: "healthy".to_string(),
+                table_id: 12,
+                batches: vec![mock_tag_batch("tag1", "host-2", 2000, 2.0)],
+                row_count: 1,
+            },
+        ];
+        let partition_calls = Arc::new(AtomicUsize::new(0));
+        let leader_calls = Arc::new(AtomicUsize::new(0));
+        let node = MockFlushNodeRequester::default();
+        let ctx = session::context::QueryContext::arc();
+
+        let err = flush_batch_physical(
+            &table_batches,
+            "phy",
+            &ctx,
+            &MockFlushPartitionProvider {
+                partition_rule_calls: partition_calls.clone(),
+                region_leader_calls: leader_calls.clone(),
+            },
+            &node,
+            &MockFlushCatalogProvider {
+                table: Some(mock_physical_table_metadata(1024)),
+            },
+        )
+        .await
+        .unwrap_err();
+
+        assert!(err.to_string().contains("unknown_tag"));
+        assert_eq!(1, partition_calls.load(Ordering::SeqCst));
+        assert_eq!(0, leader_calls.load(Ordering::SeqCst));
+        assert_eq!(0, node.writes.load(Ordering::SeqCst));
+    }
+
+    #[test]
+    fn test_plan_region_batches_splits_and_strips_partition_columns() {
+        let combined_batch = RecordBatch::try_new(
+            Arc::new(ArrowSchema::new(vec![
+                Field::new("__primary_key", ArrowDataType::Binary, false),
+                Field::new(
+                    "greptime_timestamp",
+                    ArrowDataType::Timestamp(arrow::datatypes::TimeUnit::Millisecond, None),
+                    false,
+                ),
+                Field::new("greptime_value", ArrowDataType::Float64, true),
+                Field::new("host", ArrowDataType::Utf8, true),
+            ])),
+            vec![
+                Arc::new(BinaryArray::from(vec![b"k1".as_slice(), b"k2".as_slice()])),
+                Arc::new(TimestampMillisecondArray::from(vec![1000_i64, 2000_i64])),
+                Arc::new(arrow::array::Float64Array::from(vec![1.0_f64, 2.0_f64])),
+                Arc::new(StringArray::from(vec!["node-1", "node-2"])),
+            ],
+        )
+        .unwrap();
+        let mut planned_batches = plan_region_batches(
+            combined_batch,
+            1024,
+            &TwoRegionPartitionRule {
+                partition_columns: vec!["host".to_string()],
+            },
+            &["host".to_string()],
+        )
+        .unwrap();
+        planned_batches.sort_by_key(|planned| planned.region_id.region_number());
+
+        assert_eq!(2, planned_batches.len());
+        assert_eq!(RegionId::new(1024, 1), planned_batches[0].region_id);
+        assert_eq!(1, planned_batches[0].num_rows());
+        assert_eq!(3, planned_batches[0].batch.num_columns());
+        assert_eq!(RegionId::new(1024, 2), planned_batches[1].region_id);
+        assert_eq!(1, planned_batches[1].num_rows());
+        assert_eq!(3, planned_batches[1].batch.num_columns());
+    }
+
+    #[test]
+    fn test_encode_region_write_requests_builds_bulk_insert_requests() {
+        let planned_batch = PlannedRegionBatch {
+            region_id: RegionId::new(1024, 1),
+            batch: RecordBatch::try_new(
+                Arc::new(ArrowSchema::new(vec![
+                    Field::new("__primary_key", ArrowDataType::Binary, false),
+                    Field::new(
+                        "greptime_timestamp",
+                        ArrowDataType::Timestamp(arrow::datatypes::TimeUnit::Millisecond, None),
+                        false,
+                    ),
+                    Field::new("greptime_value", ArrowDataType::Float64, true),
+                ])),
+                vec![
+                    Arc::new(BinaryArray::from(vec![b"k1".as_slice()])),
+                    Arc::new(TimestampMillisecondArray::from(vec![1000_i64])),
+                    Arc::new(arrow::array::Float64Array::from(vec![1.0_f64])),
+                ],
+            )
+            .unwrap(),
+        };
+        let resolved_batch = ResolvedRegionBatch {
+            planned: planned_batch,
+            datanode: Peer {
+                id: 1,
+                addr: "node-1".to_string(),
+            },
+        };
+        let writes = encode_region_write_requests(vec![resolved_batch]).unwrap();
+
+        assert_eq!(1, writes.len());
+        assert_eq!(1, writes[0].datanode.id);
+        let Some(region_request::Body::BulkInsert(request)) = &writes[0].request.body else {
+            panic!("expected bulk insert request");
+        };
+        assert_eq!(RegionId::new(1024, 1).as_u64(), request.region_id);
+    }
+}
diff --git a/src/servers/src/pipeline.rs b/src/servers/src/pipeline.rs
index fe7a2f48f3..9de2d63b97 100644
--- a/src/servers/src/pipeline.rs
+++ b/src/servers/src/pipeline.rs
@@ -28,7 +28,7 @@ use session::context::{Channel, QueryContextRef};
 use snafu::ResultExt;
 use vrl::value::Value as VrlValue;
 
-use crate::error::{CatalogSnafu, PipelineSnafu, Result};
+use crate::error::{PipelineSnafu, Result};
 use crate::http::event::PipelineIngestRequest;
 use crate::metrics::{
     METRIC_FAILURE_VALUE, METRIC_HTTP_LOGS_TRANSFORM_ELAPSED, METRIC_SUCCESS_VALUE,
@@ -89,10 +89,7 @@ async fn run_identity_pipeline(
     let table = if pipeline_ctx.channel == Channel::Prometheus {
         None
     } else {
-        handler
-            .get_table(&table_name, query_ctx)
-            .await
-            .context(CatalogSnafu)?
+        handler.get_table(&table_name, query_ctx).await?
     };
     identity_pipeline(data_array, table, pipeline_ctx)
         .map(|opt_map| ContextReq::from_opt_map(opt_map, table_name))
@@ -141,10 +138,7 @@ async fn run_custom_pipeline(
         }
     };
 
-    let table = handler
-        .get_table(&table_name, query_ctx)
-        .await
-        .context(CatalogSnafu)?;
+    let table = handler.get_table(&table_name, query_ctx).await?;
     schema_info.set_table(table);
 
     for pipeline_map in pipeline_maps {
diff --git a/src/servers/src/postgres.rs b/src/servers/src/postgres.rs
index 7533fe084c..58ef4fdd7b 100644
--- a/src/servers/src/postgres.rs
+++ b/src/servers/src/postgres.rs
@@ -50,7 +50,11 @@ pub(crate) struct GreptimeDBStartupParameters {
 impl GreptimeDBStartupParameters {
     fn new() -> GreptimeDBStartupParameters {
         GreptimeDBStartupParameters {
-            version: format!("16.3-greptimedb-{}", common_version::version()),
+            version: format!(
+                "16.3-{}-{}",
+                common_version::product_name(),
+                common_version::version()
+            ),
         }
     }
 }
diff --git a/src/servers/src/postgres/handler.rs b/src/servers/src/postgres/handler.rs
index daf4bfc646..e372a54237 100644
--- a/src/servers/src/postgres/handler.rs
+++ b/src/servers/src/postgres/handler.rs
@@ -20,12 +20,13 @@ use async_trait::async_trait;
 use common_query::{Output, OutputData};
 use common_recordbatch::RecordBatch;
 use common_recordbatch::error::Result as RecordBatchResult;
-use common_telemetry::{debug, tracing};
+use common_telemetry::{debug, info, tracing};
 use datafusion::sql::sqlparser::ast::{CopyOption, CopyTarget, Statement as SqlParserStatement};
 use datafusion_common::ParamValues;
+use datafusion_expr::LogicalPlan;
 use datafusion_pg_catalog::sql::PostgresCompatibilityParser;
 use datatypes::prelude::ConcreteDataType;
-use datatypes::schema::SchemaRef;
+use datatypes::schema::{Schema, SchemaRef};
 use futures::{Sink, SinkExt, Stream, StreamExt, future, stream};
 use pgwire::api::portal::{Format, Portal};
 use pgwire::api::query::{ExtendedQueryHandler, SimpleQueryHandler};
@@ -300,8 +301,8 @@ impl DefaultQueryParser {
 /// A container type of parse result types
 #[derive(Clone, Debug)]
 pub struct PgSqlPlan {
-    plan: SqlPlan,
-    copy_to_stdout_format: Option<String>,
+    pub(crate) plan: SqlPlan,
+    pub(crate) copy_to_stdout_format: Option<String>,
 }
 
 #[async_trait]
@@ -318,14 +319,16 @@ impl QueryParser for DefaultQueryParser {
         let query_ctx = self.session.new_query_context();
 
         // do not parse if query is empty or matches rules
-        if sql.is_empty() || fixtures::matches(sql) {
+        if sql.is_empty() {
             return Ok(PgSqlPlan {
-                plan: SqlPlan {
-                    query: sql.to_owned(),
-                    statement: None,
-                    plan: None,
-                    schema: None,
-                },
+                plan: SqlPlan::Empty,
+                copy_to_stdout_format: None,
+            });
+        }
+
+        if fixtures::matches(sql) {
+            return Ok(PgSqlPlan {
+                plan: SqlPlan::Shortcut(sql.to_string()),
                 copy_to_stdout_format: None,
             });
         }
@@ -354,31 +357,23 @@ impl QueryParser for DefaultQueryParser {
         } else {
             let stmt = stmts.remove(0);
 
-            let describe_result = self
+            if let Some(logical_plan) = self
                 .query_handler
                 .do_describe(stmt.clone(), query_ctx)
                 .await
-                .map_err(convert_err)?;
-
-            let (plan, schema) = if let Some(DescribeResult {
-                logical_plan,
-                schema,
-            }) = describe_result
+                .map_err(convert_err)?
+                .map(|DescribeResult { logical_plan }| logical_plan)
             {
-                (Some(logical_plan), Some(schema))
+                Ok(PgSqlPlan {
+                    plan: SqlPlan::Plan(logical_plan, stmt),
+                    copy_to_stdout_format,
+                })
             } else {
-                (None, None)
-            };
-
-            Ok(PgSqlPlan {
-                plan: SqlPlan {
-                    query: sql.clone(),
-                    statement: Some(stmt),
-                    plan,
-                    schema,
-                },
-                copy_to_stdout_format,
-            })
+                Ok(PgSqlPlan {
+                    plan: SqlPlan::Statement(stmt, sql),
+                    copy_to_stdout_format,
+                })
+            }
         }
     }
 
@@ -432,42 +427,45 @@ impl ExtendedQueryHandler for PostgresServerHandlerInner {
         let pg_sql_plan = &portal.statement.statement;
         let sql_plan = &pg_sql_plan.plan;
 
-        if sql_plan.query.is_empty() {
-            // early return if query is empty
-            return Ok(Response::EmptyQuery);
-        }
-
-        if let Some(mut resps) = fixtures::process(&sql_plan.query, query_ctx.clone()) {
-            send_warning_opt(client, query_ctx).await?;
-            // if the statement matches our predefined rules, return it early
-            return Ok(resps.remove(0));
-        }
-
-        let output = if let Some(plan) = &sql_plan.plan {
-            let values = parameters_to_scalar_values(plan, portal)?;
-            let plan = plan
-                .clone()
-                .replace_params_with_values(&ParamValues::List(
-                    values.into_iter().map(Into::into).collect(),
-                ))
-                .context(DataFusionSnafu)
-                .map_err(convert_err)?;
-            self.query_handler
-                .do_exec_plan(sql_plan.statement.clone(), plan, query_ctx.clone())
-                .await
-        } else {
-            // manually replace variables in prepared statement when no
-            // logical_plan is generated. This happens when logical plan is not
-            // supported for certain statements.
-            let mut sql = sql_plan.query.clone();
-            for i in 0..portal.parameter_len() {
-                sql = sql.replace(&format!("${}", i + 1), &parameter_to_string(portal, i)?);
+        let output = match sql_plan {
+            SqlPlan::Empty => {
+                // early return if query is empty
+                return Ok(Response::EmptyQuery);
+            }
+            SqlPlan::Shortcut(query) => {
+                if let Some(mut resps) = fixtures::process(query, query_ctx.clone()) {
+                    send_warning_opt(client, query_ctx).await?;
+                    // if the statement matches our predefined rules, return it early
+                    return Ok(resps.remove(0));
+                } else {
+                    // unreachable logic
+                    return Ok(Response::EmptyQuery);
+                }
+            }
+            SqlPlan::Plan(plan, stmt) => {
+                let values = parameters_to_scalar_values(plan, portal)?;
+                let plan = plan
+                    .clone()
+                    .replace_params_with_values(&ParamValues::List(
+                        values.into_iter().map(Into::into).collect(),
+                    ))
+                    .context(DataFusionSnafu)
+                    .map_err(convert_err)?;
+                self.query_handler
+                    .do_exec_plan(plan, Some(stmt.clone()), query_ctx.clone())
+                    .await
+            }
+            SqlPlan::Statement(_stmt, query) => {
+                // We won't replace params from statement manually any more.
+                // Newer version of datafusion can generate plan for SELECT/INSERT/UPDATE/DELETE.
+                // Only CREATE TABLE and others minor statements cannot generate sql plan,
+                // in this case, we assume these statements will not carry parameters
+                // and execute them directly.
+                self.query_handler
+                    .do_query(query, query_ctx.clone())
+                    .await
+                    .remove(0)
             }
-
-            self.query_handler
-                .do_query(&sql, query_ctx.clone())
-                .await
-                .remove(0)
         };
 
         send_warning_opt(client, query_ctx.clone()).await?;
@@ -490,7 +488,7 @@ impl ExtendedQueryHandler for PostgresServerHandlerInner {
         let sql_plan = &stmt.statement.plan;
         // client provided parameter types, can be empty if client doesn't try to parse statement
         let provided_param_types = &stmt.parameter_types;
-        let server_inferenced_types = if let Some(plan) = &sql_plan.plan {
+        let server_inferenced_types = if let SqlPlan::Plan(plan, _) = &sql_plan {
             let param_types = DfLogicalPlanner::get_inferred_parameter_types(plan)
                 .context(InferParameterTypesSnafu)
                 .map_err(convert_err)?
@@ -528,23 +526,9 @@ impl ExtendedQueryHandler for PostgresServerHandlerInner {
             })
             .collect::<Vec<_>>();
 
-        if let Some(schema) = &sql_plan.schema {
-            schema_to_pg(schema, &Format::UnifiedBinary, None)
-                .map(|fields| DescribeStatementResponse::new(param_types, fields))
-                .map_err(convert_err)
-        } else {
-            if let Some(mut resp) =
-                fixtures::process(&sql_plan.query, self.session.new_query_context())
-                && let Response::Query(query_response) = resp.remove(0)
-            {
-                return Ok(DescribeStatementResponse::new(
-                    param_types,
-                    (*query_response.row_schema()).clone(),
-                ));
-            }
+        let fields = describe_fields(sql_plan, &Format::UnifiedText, &self.session)?;
 
-            Ok(DescribeStatementResponse::new(param_types, vec![]))
-        }
+        Ok(DescribeStatementResponse::new(param_types, fields))
     }
 
     async fn do_describe_portal<C>(
@@ -558,68 +542,100 @@ impl ExtendedQueryHandler for PostgresServerHandlerInner {
         let sql_plan = &portal.statement.statement.plan;
         let format = &portal.result_column_format;
 
-        match sql_plan.statement.as_ref() {
-            Some(Statement::Query(_)) => {
-                // if the query has a schema, it is managed by datafusion, use the schema
-                if let Some(schema) = &sql_plan.schema {
-                    schema_to_pg(schema, format, None)
-                        .map(DescribePortalResponse::new)
-                        .map_err(convert_err)
-                } else {
-                    // fallback to NoData
-                    Ok(DescribePortalResponse::new(vec![]))
-                }
-            }
-            // We can cover only part of show statements
-            // these show create statements will return 2 columns
-            Some(Statement::ShowCreateDatabase(_))
-            | Some(Statement::ShowCreateTable(_))
-            | Some(Statement::ShowCreateFlow(_))
-            | Some(Statement::ShowCreateView(_)) => Ok(DescribePortalResponse::new(vec![
-                FieldInfo::new(
-                    "name".to_string(),
-                    None,
-                    None,
-                    Type::TEXT,
-                    format.format_for(0),
-                ),
-                FieldInfo::new(
-                    "create_statement".to_string(),
-                    None,
-                    None,
-                    Type::TEXT,
-                    format.format_for(1),
-                ),
-            ])),
-            // single column show statements
-            Some(Statement::ShowTables(_))
-            | Some(Statement::ShowFlows(_))
-            | Some(Statement::ShowViews(_)) => {
-                Ok(DescribePortalResponse::new(vec![FieldInfo::new(
-                    "name".to_string(),
-                    None,
-                    None,
-                    Type::TEXT,
-                    format.format_for(0),
-                )]))
-            }
-            // we will not support other show statements for extended query protocol at least for now.
-            // because the return columns is not predictable at this stage
-            _ => {
-                // test if query caught by fixture
-                if let Some(mut resp) =
-                    fixtures::process(&sql_plan.query, self.session.new_query_context())
-                    && let Response::Query(query_response) = resp.remove(0)
-                {
-                    Ok(DescribePortalResponse::new(
-                        (*query_response.row_schema()).clone(),
-                    ))
-                } else {
-                    // fallback to NoData
-                    Ok(DescribePortalResponse::new(vec![]))
-                }
+        let fields = describe_fields(sql_plan, format, &self.session)?;
+
+        Ok(DescribePortalResponse::new(fields))
+    }
+}
+
+fn describe_fields(
+    sql_plan: &SqlPlan,
+    format: &Format,
+    session: &Arc<Session>,
+) -> PgWireResult<Vec<FieldInfo>> {
+    match sql_plan {
+        // query
+        SqlPlan::Plan(plan, _) if !matches!(plan, LogicalPlan::Dml(_) | LogicalPlan::Ddl(_)) => {
+            let schema: Schema = plan.schema().clone().try_into().map_err(convert_err)?;
+            schema_to_pg(&schema, format, None).map_err(convert_err)
+        }
+        // We can cover only part of show statements
+        // these show create statements will return 2 columns
+        SqlPlan::Statement(
+            Statement::ShowCreateDatabase(_)
+            | Statement::ShowCreateTable(_)
+            | Statement::ShowCreateFlow(_)
+            | Statement::ShowCreateView(_),
+            _,
+        ) => Ok(vec![
+            FieldInfo::new(
+                "name".to_string(),
+                None,
+                None,
+                Type::TEXT,
+                format.format_for(0),
+            ),
+            FieldInfo::new(
+                "create_statement".to_string(),
+                None,
+                None,
+                Type::TEXT,
+                format.format_for(1),
+            ),
+        ]),
+        #[cfg(feature = "enterprise")]
+        SqlPlan::Statement(Statement::ShowCreateTrigger(_), _) => Ok(vec![
+            FieldInfo::new(
+                "name".to_string(),
+                None,
+                None,
+                Type::TEXT,
+                format.format_for(0),
+            ),
+            FieldInfo::new(
+                "create_statement".to_string(),
+                None,
+                None,
+                Type::TEXT,
+                format.format_for(1),
+            ),
+        ]),
+        // single column show statements
+        SqlPlan::Statement(
+            Statement::ShowTables(_) | Statement::ShowFlows(_) | Statement::ShowViews(_),
+            _,
+        ) => Ok(vec![FieldInfo::new(
+            "name".to_string(),
+            None,
+            None,
+            Type::TEXT,
+            format.format_for(0),
+        )]),
+        #[cfg(feature = "enterprise")]
+        SqlPlan::Statement(Statement::ShowTriggers(_), _) => Ok(vec![FieldInfo::new(
+            "name".to_string(),
+            None,
+            None,
+            Type::TEXT,
+            format.format_for(0),
+        )]),
+        // we will not support other show statements for extended query protocol at least for now.
+        // because the return columns is not predictable at this stage
+        SqlPlan::Shortcut(query) => {
+            // test if query caught by fixture
+            if let Some(mut resp) = fixtures::process(query, session.new_query_context())
+                && let Response::Query(query_response) = resp.remove(0)
+            {
+                Ok((*query_response.row_schema()).clone())
+            } else {
+                // fallback to NoData
+                Ok(vec![])
             }
         }
+        _ => {
+            // NoData
+            Ok(vec![])
+        }
     }
 }
 
@@ -628,7 +644,10 @@ impl ErrorHandler for PostgresServerHandlerInner {
     where
         C: ClientInfo,
     {
-        debug!("Postgres interface error {}", error)
+        match error {
+            PgWireError::IoError(e) => debug!("Postgres client disconnected: {}", e),
+            _ => info!("Postgres interface error: {}", error),
+        }
     }
 }
 
diff --git a/src/servers/src/postgres/types.rs b/src/servers/src/postgres/types.rs
index a95890e78c..9fc5dccf5f 100644
--- a/src/servers/src/postgres/types.rs
+++ b/src/servers/src/postgres/types.rs
@@ -33,7 +33,7 @@ use datatypes::arrow::datatypes::DataType as ArrowDataType;
 use datatypes::json::JsonStructureSettings;
 use datatypes::prelude::{ConcreteDataType, Value};
 use datatypes::schema::{Schema, SchemaRef};
-use datatypes::types::{IntervalType, TimestampType, jsonb_to_string};
+use datatypes::types::{Decimal128Type, IntervalType, TimestampType, jsonb_to_string};
 use datatypes::value::StructValue;
 use futures::Stream;
 use pg_interval::Interval as PgInterval;
@@ -43,6 +43,8 @@ use pgwire::api::results::FieldInfo;
 use pgwire::error::{PgWireError, PgWireResult};
 use pgwire::types::format::FormatOptions as PgFormatOptions;
 use query::planner::DfLogicalPlanner;
+use rust_decimal::Decimal;
+use rust_decimal::prelude::ToPrimitive;
 use session::context::QueryContextRef;
 use snafu::ResultExt;
 
@@ -235,7 +237,7 @@ pub(super) fn type_gt_to_pg(origin: &ConcreteDataType) -> Result<Type> {
     match origin {
         &ConcreteDataType::Null(_) => Ok(Type::UNKNOWN),
         &ConcreteDataType::Boolean(_) => Ok(Type::BOOL),
-        &ConcreteDataType::Int8(_) => Ok(Type::CHAR),
+        &ConcreteDataType::Int8(_) => Ok(Type::INT2),
         &ConcreteDataType::Int16(_) | &ConcreteDataType::UInt8(_) => Ok(Type::INT2),
         &ConcreteDataType::Int32(_) | &ConcreteDataType::UInt16(_) => Ok(Type::INT4),
         &ConcreteDataType::Int64(_) | &ConcreteDataType::UInt32(_) => Ok(Type::INT8),
@@ -253,7 +255,7 @@ pub(super) fn type_gt_to_pg(origin: &ConcreteDataType) -> Result<Type> {
         ConcreteDataType::List(list) => match list.item_type() {
             &ConcreteDataType::Null(_) => Ok(Type::UNKNOWN),
             &ConcreteDataType::Boolean(_) => Ok(Type::BOOL_ARRAY),
-            &ConcreteDataType::Int8(_) => Ok(Type::CHAR_ARRAY),
+            &ConcreteDataType::Int8(_) => Ok(Type::INT2_ARRAY),
             &ConcreteDataType::Int16(_) | &ConcreteDataType::UInt8(_) => Ok(Type::INT2_ARRAY),
             &ConcreteDataType::Int32(_) | &ConcreteDataType::UInt16(_) => Ok(Type::INT4_ARRAY),
             &ConcreteDataType::Int64(_) | &ConcreteDataType::UInt32(_) => Ok(Type::INT8_ARRAY),
@@ -293,11 +295,11 @@ pub(super) fn type_pg_to_gt(origin: &Type) -> Result<ConcreteDataType> {
     // Note that we only support a small amount of pg data types
     match origin {
         &Type::BOOL => Ok(ConcreteDataType::boolean_datatype()),
-        &Type::CHAR => Ok(ConcreteDataType::int8_datatype()),
         &Type::INT2 => Ok(ConcreteDataType::int16_datatype()),
         &Type::INT4 => Ok(ConcreteDataType::int32_datatype()),
         &Type::INT8 => Ok(ConcreteDataType::int64_datatype()),
-        &Type::VARCHAR | &Type::TEXT => Ok(ConcreteDataType::string_datatype()),
+        &Type::NUMERIC => Ok(ConcreteDataType::uint64_datatype()),
+        &Type::VARCHAR | &Type::CHAR | &Type::TEXT => Ok(ConcreteDataType::string_datatype()),
         &Type::TIMESTAMP | &Type::TIMESTAMPTZ => Ok(ConcreteDataType::timestamp_datatype(
             common_time::timestamp::TimeUnit::Millisecond,
         )),
@@ -305,9 +307,6 @@ pub(super) fn type_pg_to_gt(origin: &Type) -> Result<ConcreteDataType> {
         &Type::TIME => Ok(ConcreteDataType::timestamp_datatype(
             common_time::timestamp::TimeUnit::Microsecond,
         )),
-        &Type::CHAR_ARRAY => Ok(ConcreteDataType::list_datatype(Arc::new(
-            ConcreteDataType::int8_datatype(),
-        ))),
         &Type::INT2_ARRAY => Ok(ConcreteDataType::list_datatype(Arc::new(
             ConcreteDataType::int16_datatype(),
         ))),
@@ -317,9 +316,12 @@ pub(super) fn type_pg_to_gt(origin: &Type) -> Result<ConcreteDataType> {
         &Type::INT8_ARRAY => Ok(ConcreteDataType::list_datatype(Arc::new(
             ConcreteDataType::int64_datatype(),
         ))),
-        &Type::VARCHAR_ARRAY => Ok(ConcreteDataType::list_datatype(Arc::new(
-            ConcreteDataType::string_datatype(),
+        &Type::NUMERIC_ARRAY => Ok(ConcreteDataType::list_datatype(Arc::new(
+            ConcreteDataType::uint64_datatype(),
         ))),
+        &Type::VARCHAR_ARRAY | &Type::CHAR_ARRAY | &Type::TEXT_ARRAY => Ok(
+            ConcreteDataType::list_datatype(Arc::new(ConcreteDataType::string_datatype())),
+        ),
         _ => server_error::InternalSnafu {
             err_msg: format!("unimplemented datatype {origin:?}"),
         }
@@ -327,63 +329,6 @@ pub(super) fn type_pg_to_gt(origin: &Type) -> Result<ConcreteDataType> {
     }
 }
 
-pub(super) fn parameter_to_string(portal: &Portal<PgSqlPlan>, idx: usize) -> PgWireResult<String> {
-    // the index is managed from portal's parameters count so it's safe to
-    // unwrap here.
-    let param_type = portal
-        .statement
-        .parameter_types
-        .get(idx)
-        .unwrap()
-        .as_ref()
-        .unwrap_or(&Type::UNKNOWN);
-    match param_type {
-        &Type::VARCHAR | &Type::TEXT => Ok(format!(
-            "'{}'",
-            portal
-                .parameter::<String>(idx, param_type)?
-                .as_deref()
-                .unwrap_or("")
-        )),
-        &Type::BOOL => Ok(portal
-            .parameter::<bool>(idx, param_type)?
-            .map(|v| v.to_string())
-            .unwrap_or_else(|| "".to_owned())),
-        &Type::INT4 => Ok(portal
-            .parameter::<i32>(idx, param_type)?
-            .map(|v| v.to_string())
-            .unwrap_or_else(|| "".to_owned())),
-        &Type::INT8 => Ok(portal
-            .parameter::<i64>(idx, param_type)?
-            .map(|v| v.to_string())
-            .unwrap_or_else(|| "".to_owned())),
-        &Type::FLOAT4 => Ok(portal
-            .parameter::<f32>(idx, param_type)?
-            .map(|v| v.to_string())
-            .unwrap_or_else(|| "".to_owned())),
-        &Type::FLOAT8 => Ok(portal
-            .parameter::<f64>(idx, param_type)?
-            .map(|v| v.to_string())
-            .unwrap_or_else(|| "".to_owned())),
-        &Type::DATE => Ok(portal
-            .parameter::<NaiveDate>(idx, param_type)?
-            .map(|v| v.format("%Y-%m-%d").to_string())
-            .unwrap_or_else(|| "".to_owned())),
-        &Type::TIMESTAMP => Ok(portal
-            .parameter::<NaiveDateTime>(idx, param_type)?
-            .map(|v| v.format("%Y-%m-%d %H:%M:%S%.6f").to_string())
-            .unwrap_or_else(|| "".to_owned())),
-        &Type::INTERVAL => Ok(portal
-            .parameter::<PgInterval>(idx, param_type)?
-            .map(|v| v.to_sql())
-            .unwrap_or_else(|| "".to_owned())),
-        _ => Err(invalid_parameter_error(
-            "unsupported_parameter_type",
-            Some(param_type.to_string()),
-        )),
-    }
-}
-
 pub(super) fn invalid_parameter_error(msg: &str, detail: Option<String>) -> PgWireError {
     let mut error_info = PgErrorCode::Ec22023.to_err_info(msg.to_string());
     error_info.detail = detail;
@@ -407,6 +352,24 @@ where
     }
 }
 
+fn to_decimal_scalar_value(data: Option<Decimal>, ctype: &Decimal128Type) -> ScalarValue {
+    if let Some(data) = data {
+        let mut value = data;
+        value.rescale(ctype.scale() as u32);
+
+        ScalarValue::Decimal128(Some(value.mantissa()), ctype.precision(), ctype.scale())
+    } else {
+        ScalarValue::Decimal128(None, ctype.precision(), ctype.scale())
+    }
+}
+
+fn numeric_out_of_range_error(value: impl std::fmt::Display) -> PgWireError {
+    invalid_parameter_error(
+        "numeric_value_out_of_range",
+        Some(format!("value {} is out of range for target type", value)),
+    )
+}
+
 pub(super) fn parameters_to_scalar_values(
     plan: &LogicalPlan,
     portal: &Portal<PgSqlPlan>,
@@ -435,14 +398,14 @@ pub(super) fn parameters_to_scalar_values(
             return Err(invalid_parameter_error(
                 "unknown_parameter_type",
                 Some(format!(
-                    "Cannot get parameter type information for parameter {}",
-                    idx
+                    "Cannot get type for parameter {}, try to provide a type using ${}::<type>",
+                    idx, idx
                 )),
             ));
         };
 
         let value = match &client_type {
-            &Type::VARCHAR | &Type::TEXT => {
+            &Type::VARCHAR | &Type::TEXT | &Type::CHAR => {
                 let data = portal.parameter::<String>(idx, &client_type)?;
                 if let Some(server_type) = &server_type {
                     match server_type {
@@ -484,14 +447,29 @@ pub(super) fn parameters_to_scalar_values(
                 let data = portal.parameter::<i16>(idx, &client_type)?;
                 if let Some(server_type) = &server_type {
                     match server_type {
-                        ConcreteDataType::Int8(_) => ScalarValue::Int8(data.map(|n| n as i8)),
+                        ConcreteDataType::Int8(_) => ScalarValue::Int8(
+                            data.map(|n| n.to_i8().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
                         ConcreteDataType::Int16(_) => ScalarValue::Int16(data),
                         ConcreteDataType::Int32(_) => ScalarValue::Int32(data.map(|n| n as i32)),
                         ConcreteDataType::Int64(_) => ScalarValue::Int64(data.map(|n| n as i64)),
-                        ConcreteDataType::UInt8(_) => ScalarValue::UInt8(data.map(|n| n as u8)),
-                        ConcreteDataType::UInt16(_) => ScalarValue::UInt16(data.map(|n| n as u16)),
-                        ConcreteDataType::UInt32(_) => ScalarValue::UInt32(data.map(|n| n as u32)),
-                        ConcreteDataType::UInt64(_) => ScalarValue::UInt64(data.map(|n| n as u64)),
+                        ConcreteDataType::UInt8(_) => ScalarValue::UInt8(
+                            data.map(|n| n.to_u8().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt16(_) => ScalarValue::UInt16(
+                            data.map(|n| n.to_u16().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt32(_) => ScalarValue::UInt32(
+                            data.map(|n| n.to_u32().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt64(_) => ScalarValue::UInt64(
+                            data.map(|n| n.to_u64().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
                         ConcreteDataType::Timestamp(unit) => {
                             to_timestamp_scalar_value(data, unit, server_type)?
                         }
@@ -510,14 +488,32 @@ pub(super) fn parameters_to_scalar_values(
                 let data = portal.parameter::<i32>(idx, &client_type)?;
                 if let Some(server_type) = &server_type {
                     match server_type {
-                        ConcreteDataType::Int8(_) => ScalarValue::Int8(data.map(|n| n as i8)),
-                        ConcreteDataType::Int16(_) => ScalarValue::Int16(data.map(|n| n as i16)),
+                        ConcreteDataType::Int8(_) => ScalarValue::Int8(
+                            data.map(|n| n.to_i8().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::Int16(_) => ScalarValue::Int16(
+                            data.map(|n| n.to_i16().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
                         ConcreteDataType::Int32(_) => ScalarValue::Int32(data),
                         ConcreteDataType::Int64(_) => ScalarValue::Int64(data.map(|n| n as i64)),
-                        ConcreteDataType::UInt8(_) => ScalarValue::UInt8(data.map(|n| n as u8)),
-                        ConcreteDataType::UInt16(_) => ScalarValue::UInt16(data.map(|n| n as u16)),
-                        ConcreteDataType::UInt32(_) => ScalarValue::UInt32(data.map(|n| n as u32)),
-                        ConcreteDataType::UInt64(_) => ScalarValue::UInt64(data.map(|n| n as u64)),
+                        ConcreteDataType::UInt8(_) => ScalarValue::UInt8(
+                            data.map(|n| n.to_u8().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt16(_) => ScalarValue::UInt16(
+                            data.map(|n| n.to_u16().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt32(_) => ScalarValue::UInt32(
+                            data.map(|n| n.to_u32().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt64(_) => ScalarValue::UInt64(
+                            data.map(|n| n.to_u64().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
                         ConcreteDataType::Timestamp(unit) => {
                             to_timestamp_scalar_value(data, unit, server_type)?
                         }
@@ -536,14 +532,35 @@ pub(super) fn parameters_to_scalar_values(
                 let data = portal.parameter::<i64>(idx, &client_type)?;
                 if let Some(server_type) = &server_type {
                     match server_type {
-                        ConcreteDataType::Int8(_) => ScalarValue::Int8(data.map(|n| n as i8)),
-                        ConcreteDataType::Int16(_) => ScalarValue::Int16(data.map(|n| n as i16)),
-                        ConcreteDataType::Int32(_) => ScalarValue::Int32(data.map(|n| n as i32)),
+                        ConcreteDataType::Int8(_) => ScalarValue::Int8(
+                            data.map(|n| n.to_i8().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::Int16(_) => ScalarValue::Int16(
+                            data.map(|n| n.to_i16().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::Int32(_) => ScalarValue::Int32(
+                            data.map(|n| n.to_i32().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
                         ConcreteDataType::Int64(_) => ScalarValue::Int64(data),
-                        ConcreteDataType::UInt8(_) => ScalarValue::UInt8(data.map(|n| n as u8)),
-                        ConcreteDataType::UInt16(_) => ScalarValue::UInt16(data.map(|n| n as u16)),
-                        ConcreteDataType::UInt32(_) => ScalarValue::UInt32(data.map(|n| n as u32)),
-                        ConcreteDataType::UInt64(_) => ScalarValue::UInt64(data.map(|n| n as u64)),
+                        ConcreteDataType::UInt8(_) => ScalarValue::UInt8(
+                            data.map(|n| n.to_u8().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt16(_) => ScalarValue::UInt16(
+                            data.map(|n| n.to_u16().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt32(_) => ScalarValue::UInt32(
+                            data.map(|n| n.to_u32().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt64(_) => ScalarValue::UInt64(
+                            data.map(|n| n.to_u64().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
                         ConcreteDataType::Timestamp(unit) => {
                             to_timestamp_scalar_value(data, unit, server_type)?
                         }
@@ -558,18 +575,60 @@ pub(super) fn parameters_to_scalar_values(
                     ScalarValue::Int64(data)
                 }
             }
+            &Type::NUMERIC => {
+                let data = portal.parameter::<Decimal>(idx, &client_type)?;
+                match &server_type {
+                    Some(ConcreteDataType::Decimal128(dt)) => to_decimal_scalar_value(data, dt),
+                    Some(st @ ConcreteDataType::Timestamp(unit)) => {
+                        to_timestamp_scalar_value(data.and_then(|n| n.to_i64()), unit, st)?
+                    }
+                    Some(ConcreteDataType::UInt64(_)) | None => {
+                        ScalarValue::UInt64(data.and_then(|n| n.to_u64()))
+                    }
+                    Some(st) => {
+                        return Err(invalid_parameter_error(
+                            "invalid_parameter_type",
+                            Some(format!("Expected: {}, found: {}", st, client_type)),
+                        ));
+                    }
+                }
+            }
             &Type::FLOAT4 => {
                 let data = portal.parameter::<f32>(idx, &client_type)?;
                 if let Some(server_type) = &server_type {
                     match server_type {
-                        ConcreteDataType::Int8(_) => ScalarValue::Int8(data.map(|n| n as i8)),
-                        ConcreteDataType::Int16(_) => ScalarValue::Int16(data.map(|n| n as i16)),
-                        ConcreteDataType::Int32(_) => ScalarValue::Int32(data.map(|n| n as i32)),
-                        ConcreteDataType::Int64(_) => ScalarValue::Int64(data.map(|n| n as i64)),
-                        ConcreteDataType::UInt8(_) => ScalarValue::UInt8(data.map(|n| n as u8)),
-                        ConcreteDataType::UInt16(_) => ScalarValue::UInt16(data.map(|n| n as u16)),
-                        ConcreteDataType::UInt32(_) => ScalarValue::UInt32(data.map(|n| n as u32)),
-                        ConcreteDataType::UInt64(_) => ScalarValue::UInt64(data.map(|n| n as u64)),
+                        ConcreteDataType::Int8(_) => ScalarValue::Int8(
+                            data.map(|n| n.to_i8().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::Int16(_) => ScalarValue::Int16(
+                            data.map(|n| n.to_i16().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::Int32(_) => ScalarValue::Int32(
+                            data.map(|n| n.to_i32().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::Int64(_) => ScalarValue::Int64(
+                            data.map(|n| n.to_i64().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt8(_) => ScalarValue::UInt8(
+                            data.map(|n| n.to_u8().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt16(_) => ScalarValue::UInt16(
+                            data.map(|n| n.to_u16().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt32(_) => ScalarValue::UInt32(
+                            data.map(|n| n.to_u32().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt64(_) => ScalarValue::UInt64(
+                            data.map(|n| n.to_u64().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
                         ConcreteDataType::Float32(_) => ScalarValue::Float32(data),
                         ConcreteDataType::Float64(_) => {
                             ScalarValue::Float64(data.map(|n| n as f64))
@@ -589,17 +648,42 @@ pub(super) fn parameters_to_scalar_values(
                 let data = portal.parameter::<f64>(idx, &client_type)?;
                 if let Some(server_type) = &server_type {
                     match server_type {
-                        ConcreteDataType::Int8(_) => ScalarValue::Int8(data.map(|n| n as i8)),
-                        ConcreteDataType::Int16(_) => ScalarValue::Int16(data.map(|n| n as i16)),
-                        ConcreteDataType::Int32(_) => ScalarValue::Int32(data.map(|n| n as i32)),
-                        ConcreteDataType::Int64(_) => ScalarValue::Int64(data.map(|n| n as i64)),
-                        ConcreteDataType::UInt8(_) => ScalarValue::UInt8(data.map(|n| n as u8)),
-                        ConcreteDataType::UInt16(_) => ScalarValue::UInt16(data.map(|n| n as u16)),
-                        ConcreteDataType::UInt32(_) => ScalarValue::UInt32(data.map(|n| n as u32)),
-                        ConcreteDataType::UInt64(_) => ScalarValue::UInt64(data.map(|n| n as u64)),
-                        ConcreteDataType::Float32(_) => {
-                            ScalarValue::Float32(data.map(|n| n as f32))
-                        }
+                        ConcreteDataType::Int8(_) => ScalarValue::Int8(
+                            data.map(|n| n.to_i8().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::Int16(_) => ScalarValue::Int16(
+                            data.map(|n| n.to_i16().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::Int32(_) => ScalarValue::Int32(
+                            data.map(|n| n.to_i32().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::Int64(_) => ScalarValue::Int64(
+                            data.map(|n| n.to_i64().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt8(_) => ScalarValue::UInt8(
+                            data.map(|n| n.to_u8().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt16(_) => ScalarValue::UInt16(
+                            data.map(|n| n.to_u16().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt32(_) => ScalarValue::UInt32(
+                            data.map(|n| n.to_u32().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::UInt64(_) => ScalarValue::UInt64(
+                            data.map(|n| n.to_u64().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
+                        ConcreteDataType::Float32(_) => ScalarValue::Float32(
+                            data.map(|n| n.to_f32().ok_or_else(|| numeric_out_of_range_error(n)))
+                                .transpose()?,
+                        ),
                         ConcreteDataType::Float64(_) => ScalarValue::Float64(data),
                         _ => {
                             return Err(invalid_parameter_error(
@@ -837,7 +921,67 @@ pub(super) fn parameters_to_scalar_values(
                     ScalarValue::Null
                 }
             }
-            &Type::VARCHAR_ARRAY => {
+            &Type::NUMERIC_ARRAY => {
+                let data = portal.parameter::<Vec<Option<Decimal>>>(idx, &client_type)?;
+                if let Some(data) = data {
+                    let build_u64_list = |data: Vec<Option<Decimal>>| {
+                        let values = data
+                            .into_iter()
+                            .map(|n| ScalarValue::UInt64(n.and_then(|n| n.to_u64())))
+                            .collect::<Vec<_>>();
+                        ScalarValue::List(ScalarValue::new_list(
+                            &values,
+                            &ArrowDataType::UInt64,
+                            true,
+                        ))
+                    };
+                    if let Some(server_type) = &server_type {
+                        match server_type {
+                            ConcreteDataType::List(list_type) => match list_type.item_type() {
+                                ConcreteDataType::UInt64(_) => build_u64_list(data),
+                                ConcreteDataType::Decimal128(dt) => {
+                                    let values = data
+                                        .into_iter()
+                                        .map(|n| to_decimal_scalar_value(n, dt))
+                                        .collect::<Vec<_>>();
+                                    ScalarValue::List(ScalarValue::new_list(
+                                        &values,
+                                        &ArrowDataType::Decimal128(dt.precision(), dt.scale()),
+                                        true,
+                                    ))
+                                }
+                                _ => {
+                                    // the server type is not a list of decimal or uint64
+                                    return Err(invalid_parameter_error(
+                                        "invalid_parameter_type",
+                                        Some(format!(
+                                            "Expected: {}, found: {}",
+                                            list_type.item_type(),
+                                            client_type
+                                        )),
+                                    ));
+                                }
+                            },
+                            _ => {
+                                // the server type is not a list
+                                return Err(invalid_parameter_error(
+                                    "invalid_parameter_type",
+                                    Some(format!(
+                                        "Expected: {}, found: {}",
+                                        server_type, client_type
+                                    )),
+                                ));
+                            }
+                        }
+                    } else {
+                        // server type not provided
+                        build_u64_list(data)
+                    }
+                } else {
+                    ScalarValue::Null
+                }
+            }
+            &Type::VARCHAR_ARRAY | &Type::TEXT_ARRAY | &Type::CHAR_ARRAY => {
                 let data = portal.parameter::<Vec<Option<String>>>(idx, &client_type)?;
                 if let Some(data) = data {
                     let values = data.into_iter().map(|i| i.into()).collect::<Vec<_>>();
@@ -1098,12 +1242,16 @@ pub fn format_options_from_query_ctx(query_ctx: &QueryContextRef) -> Arc<PgForma
 
 #[cfg(test)]
 mod test {
+    use std::str::FromStr;
     use std::sync::Arc;
 
     use arrow::array::{
         Float64Builder, Int64Builder, ListBuilder, StringBuilder, TimestampSecondBuilder,
     };
     use arrow_schema::{Field, IntervalUnit};
+    use bytes::Bytes;
+    use datafusion_expr::expr::Placeholder;
+    use datafusion_expr::{Expr, LogicalPlanBuilder};
     use datatypes::schema::{ColumnSchema, Schema};
     use datatypes::vectors::{
         BinaryVector, BooleanVector, DateVector, Float32Vector, Float64Vector, Int8Vector,
@@ -1113,10 +1261,15 @@ mod test {
     };
     use futures::{StreamExt as FuturesStreamExt, stream};
     use pgwire::api::Type;
+    use pgwire::api::portal::{Format, Portal};
     use pgwire::api::results::{DataRowEncoder, FieldFormat, FieldInfo};
+    use pgwire::api::stmt::StoredStatement;
+    use pgwire::messages::extendedquery::Bind;
     use session::context::QueryContextBuilder;
 
     use super::*;
+    use crate::SqlPlan;
+    use crate::postgres::handler::PgSqlPlan;
 
     #[test]
     fn test_schema_convert() {
@@ -1151,7 +1304,7 @@ mod test {
         let pg_field_info = vec![
             FieldInfo::new("nulls".into(), None, None, Type::UNKNOWN, FieldFormat::Text),
             FieldInfo::new("bools".into(), None, None, Type::BOOL, FieldFormat::Text),
-            FieldInfo::new("int8s".into(), None, None, Type::CHAR, FieldFormat::Text),
+            FieldInfo::new("int8s".into(), None, None, Type::INT2, FieldFormat::Text),
             FieldInfo::new("int16s".into(), None, None, Type::INT2, FieldFormat::Text),
             FieldInfo::new("int32s".into(), None, None, Type::INT4, FieldFormat::Text),
             FieldInfo::new("int64s".into(), None, None, Type::INT8, FieldFormat::Text),
@@ -1230,7 +1383,7 @@ mod test {
                 Type::NUMERIC,
                 FieldFormat::Text,
             ),
-            FieldInfo::new("int8s".into(), None, None, Type::CHAR, FieldFormat::Text),
+            FieldInfo::new("int8s".into(), None, None, Type::INT2, FieldFormat::Text),
             FieldInfo::new("int16s".into(), None, None, Type::INT2, FieldFormat::Text),
             FieldInfo::new("int32s".into(), None, None, Type::INT4, FieldFormat::Text),
             FieldInfo::new("int64s".into(), None, None, Type::INT8, FieldFormat::Text),
@@ -1516,4 +1669,360 @@ mod test {
             panic!("test_invalid_parameter failed");
         }
     }
+
+    #[test]
+    fn test_to_decimal_scalar_value() {
+        let dt = Decimal128Type::new(18, 4);
+
+        let d = Decimal::from_str("12345.6789").unwrap();
+        assert_eq!(d.mantissa(), 123456789i128);
+        let scalar = to_decimal_scalar_value(Some(d), &dt);
+        assert_eq!(scalar, ScalarValue::Decimal128(Some(123456789), 18, 4));
+
+        let d = Decimal::from_str("100.5").unwrap();
+        assert_eq!(d.mantissa(), 1005);
+        let scalar = to_decimal_scalar_value(Some(d), &dt);
+        assert_eq!(scalar, ScalarValue::Decimal128(Some(1005000), 18, 4));
+
+        let d = Decimal::from_str("-9876.5432").unwrap();
+        let scalar = to_decimal_scalar_value(Some(d), &dt);
+        assert_eq!(scalar, ScalarValue::Decimal128(Some(-98765432), 18, 4));
+
+        let scalar = to_decimal_scalar_value(None, &dt);
+        assert_eq!(scalar, ScalarValue::Decimal128(None, 18, 4));
+    }
+
+    fn s(v: &str) -> Option<String> {
+        Some(v.to_string())
+    }
+
+    fn typed_param(id: &str, dt: DataType) -> Expr {
+        Expr::Placeholder(Placeholder::new_with_field(
+            id.to_string(),
+            Some(Arc::new(arrow_schema::Field::new(id, dt, true))),
+        ))
+    }
+
+    fn build_plan_with_params(params: Vec<(&str, DataType)>) -> LogicalPlan {
+        let exprs: Vec<Expr> = params
+            .into_iter()
+            .map(|(id, dt)| typed_param(id, dt))
+            .collect();
+        LogicalPlanBuilder::empty(true)
+            .project(exprs)
+            .unwrap()
+            .build()
+            .unwrap()
+    }
+
+    fn make_portal(
+        client_param_types: Vec<Option<Type>>,
+        param_data: Vec<Option<String>>,
+    ) -> Portal<PgSqlPlan> {
+        let bind = Bind::new(
+            None,
+            None,
+            vec![],
+            param_data
+                .into_iter()
+                .map(|opt| opt.map(Bytes::from))
+                .collect(),
+            vec![],
+        );
+        let statement = Arc::new(StoredStatement::new(
+            String::new(),
+            PgSqlPlan {
+                plan: SqlPlan::Empty,
+                copy_to_stdout_format: None,
+            },
+            client_param_types,
+        ));
+        Portal::try_new(&bind, statement).unwrap()
+    }
+
+    #[test]
+    fn test_int2_coerce_in_range() {
+        let plan = build_plan_with_params(vec![
+            ("$1", DataType::Int8),
+            ("$2", DataType::Int16),
+            ("$3", DataType::Int32),
+            ("$4", DataType::Int64),
+            ("$5", DataType::UInt8),
+            ("$6", DataType::UInt16),
+            ("$7", DataType::UInt32),
+            ("$8", DataType::UInt64),
+        ]);
+        let portal = make_portal(
+            vec![
+                Some(Type::INT2),
+                Some(Type::INT2),
+                Some(Type::INT2),
+                Some(Type::INT2),
+                Some(Type::INT2),
+                Some(Type::INT2),
+                Some(Type::INT2),
+                Some(Type::INT2),
+            ],
+            vec![
+                s("100"),
+                s("100"),
+                s("100"),
+                s("100"),
+                s("100"),
+                s("100"),
+                s("100"),
+                s("100"),
+            ],
+        );
+
+        let values = parameters_to_scalar_values(&plan, &portal).unwrap();
+        assert_eq!(values[0], ScalarValue::Int8(Some(100)));
+        assert_eq!(values[1], ScalarValue::Int16(Some(100)));
+        assert_eq!(values[2], ScalarValue::Int32(Some(100)));
+        assert_eq!(values[3], ScalarValue::Int64(Some(100)));
+        assert_eq!(values[4], ScalarValue::UInt8(Some(100)));
+        assert_eq!(values[5], ScalarValue::UInt16(Some(100)));
+        assert_eq!(values[6], ScalarValue::UInt32(Some(100)));
+        assert_eq!(values[7], ScalarValue::UInt64(Some(100)));
+    }
+
+    #[test]
+    fn test_int2_coerce_out_of_range() {
+        let plan = build_plan_with_params(vec![("$1", DataType::Int8)]);
+        let portal = make_portal(vec![Some(Type::INT2)], vec![s("200")]);
+        let result = parameters_to_scalar_values(&plan, &portal);
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_int2_coerce_negative_to_unsigned_out_of_range() {
+        let plan = build_plan_with_params(vec![("$1", DataType::UInt64)]);
+        let portal = make_portal(vec![Some(Type::INT2)], vec![s("-1")]);
+        let result = parameters_to_scalar_values(&plan, &portal);
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_int4_coerce_in_range() {
+        let plan = build_plan_with_params(vec![
+            ("$1", DataType::Int8),
+            ("$2", DataType::Int16),
+            ("$3", DataType::Int32),
+            ("$4", DataType::Int64),
+            ("$5", DataType::UInt8),
+            ("$6", DataType::UInt16),
+            ("$7", DataType::UInt32),
+            ("$8", DataType::UInt64),
+        ]);
+        let portal = make_portal(
+            vec![
+                Some(Type::INT4),
+                Some(Type::INT4),
+                Some(Type::INT4),
+                Some(Type::INT4),
+                Some(Type::INT4),
+                Some(Type::INT4),
+                Some(Type::INT4),
+                Some(Type::INT4),
+            ],
+            vec![
+                s("100"),
+                s("1000"),
+                s("100000"),
+                s("100000"),
+                s("200"),
+                s("1000"),
+                s("100000"),
+                s("100000"),
+            ],
+        );
+
+        let values = parameters_to_scalar_values(&plan, &portal).unwrap();
+        assert_eq!(values[0], ScalarValue::Int8(Some(100)));
+        assert_eq!(values[1], ScalarValue::Int16(Some(1000)));
+        assert_eq!(values[2], ScalarValue::Int32(Some(100000)));
+        assert_eq!(values[3], ScalarValue::Int64(Some(100000)));
+        assert_eq!(values[4], ScalarValue::UInt8(Some(200)));
+        assert_eq!(values[5], ScalarValue::UInt16(Some(1000)));
+        assert_eq!(values[6], ScalarValue::UInt32(Some(100000)));
+        assert_eq!(values[7], ScalarValue::UInt64(Some(100000)));
+    }
+
+    #[test]
+    fn test_int4_coerce_out_of_range() {
+        let plan = build_plan_with_params(vec![("$1", DataType::Int8)]);
+        let portal = make_portal(vec![Some(Type::INT4)], vec![s("200")]);
+        let result = parameters_to_scalar_values(&plan, &portal);
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_int4_coerce_i32_max_to_i16_out_of_range() {
+        let plan = build_plan_with_params(vec![("$1", DataType::Int16)]);
+        let portal = make_portal(vec![Some(Type::INT4)], vec![Some(i32::MAX.to_string())]);
+        let result = parameters_to_scalar_values(&plan, &portal);
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_int8_coerce_in_range() {
+        let plan = build_plan_with_params(vec![
+            ("$1", DataType::Int8),
+            ("$2", DataType::Int16),
+            ("$3", DataType::Int32),
+            ("$4", DataType::Int64),
+            ("$5", DataType::UInt8),
+            ("$6", DataType::UInt16),
+            ("$7", DataType::UInt32),
+            ("$8", DataType::UInt64),
+        ]);
+        let portal = make_portal(
+            vec![
+                Some(Type::INT8),
+                Some(Type::INT8),
+                Some(Type::INT8),
+                Some(Type::INT8),
+                Some(Type::INT8),
+                Some(Type::INT8),
+                Some(Type::INT8),
+                Some(Type::INT8),
+            ],
+            vec![
+                s("100"),
+                s("1000"),
+                s("100000"),
+                s("100000"),
+                s("200"),
+                s("1000"),
+                s("3000000000"),
+                s("3000000000"),
+            ],
+        );
+
+        let values = parameters_to_scalar_values(&plan, &portal).unwrap();
+        assert_eq!(values[0], ScalarValue::Int8(Some(100)));
+        assert_eq!(values[1], ScalarValue::Int16(Some(1000)));
+        assert_eq!(values[2], ScalarValue::Int32(Some(100000)));
+        assert_eq!(values[3], ScalarValue::Int64(Some(100000)));
+        assert_eq!(values[4], ScalarValue::UInt8(Some(200)));
+        assert_eq!(values[5], ScalarValue::UInt16(Some(1000)));
+        assert_eq!(values[6], ScalarValue::UInt32(Some(3000000000)));
+        assert_eq!(values[7], ScalarValue::UInt64(Some(3000000000)));
+    }
+
+    #[test]
+    fn test_int8_coerce_out_of_range() {
+        let plan = build_plan_with_params(vec![("$1", DataType::Int32)]);
+        let portal = make_portal(
+            vec![Some(Type::INT8)],
+            vec![Some((i32::MAX as i64 + 1).to_string())],
+        );
+        let result = parameters_to_scalar_values(&plan, &portal);
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_int8_coerce_negative_to_unsigned_out_of_range() {
+        let plan = build_plan_with_params(vec![("$1", DataType::UInt64)]);
+        let portal = make_portal(vec![Some(Type::INT8)], vec![s("-1")]);
+        let result = parameters_to_scalar_values(&plan, &portal);
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_float4_coerce_in_range() {
+        let plan =
+            build_plan_with_params(vec![("$1", DataType::Float32), ("$2", DataType::Float64)]);
+        let portal = make_portal(
+            vec![Some(Type::FLOAT4), Some(Type::FLOAT4)],
+            vec![s("1.5"), s("2.5")],
+        );
+
+        let values = parameters_to_scalar_values(&plan, &portal).unwrap();
+        assert_eq!(values[0], ScalarValue::Float32(Some(1.5)));
+        assert_eq!(values[1], ScalarValue::Float64(Some(2.5)));
+    }
+
+    #[test]
+    fn test_float4_coerce_to_int_in_range() {
+        let plan = build_plan_with_params(vec![
+            ("$1", DataType::Int8),
+            ("$2", DataType::Int32),
+            ("$3", DataType::UInt64),
+        ]);
+        let portal = make_portal(
+            vec![Some(Type::FLOAT4), Some(Type::FLOAT4), Some(Type::FLOAT4)],
+            vec![s("100"), s("1000"), s("200")],
+        );
+
+        let values = parameters_to_scalar_values(&plan, &portal).unwrap();
+        assert_eq!(values[0], ScalarValue::Int8(Some(100)));
+        assert_eq!(values[1], ScalarValue::Int32(Some(1000)));
+        assert_eq!(values[2], ScalarValue::UInt64(Some(200)));
+    }
+
+    #[test]
+    fn test_float4_coerce_to_int_out_of_range() {
+        let plan = build_plan_with_params(vec![("$1", DataType::Int8)]);
+        let portal = make_portal(vec![Some(Type::FLOAT4)], vec![s("200")]);
+        let result = parameters_to_scalar_values(&plan, &portal);
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_float8_coerce_in_range() {
+        let plan =
+            build_plan_with_params(vec![("$1", DataType::Float32), ("$2", DataType::Float64)]);
+        let portal = make_portal(
+            vec![Some(Type::FLOAT8), Some(Type::FLOAT8)],
+            vec![s("1.5"), s("2.5")],
+        );
+
+        let values = parameters_to_scalar_values(&plan, &portal).unwrap();
+        assert_eq!(values[0], ScalarValue::Float32(Some(1.5)));
+        assert_eq!(values[1], ScalarValue::Float64(Some(2.5)));
+    }
+
+    #[test]
+    fn test_float8_coerce_to_int_in_range() {
+        let plan = build_plan_with_params(vec![
+            ("$1", DataType::Int8),
+            ("$2", DataType::Int64),
+            ("$3", DataType::UInt64),
+        ]);
+        let portal = make_portal(
+            vec![Some(Type::FLOAT8), Some(Type::FLOAT8), Some(Type::FLOAT8)],
+            vec![s("100"), s("1000000"), s("200")],
+        );
+
+        let values = parameters_to_scalar_values(&plan, &portal).unwrap();
+        assert_eq!(values[0], ScalarValue::Int8(Some(100)));
+        assert_eq!(values[1], ScalarValue::Int64(Some(1000000)));
+        assert_eq!(values[2], ScalarValue::UInt64(Some(200)));
+    }
+
+    #[test]
+    fn test_float8_coerce_to_int_out_of_range() {
+        let plan = build_plan_with_params(vec![("$1", DataType::Int8)]);
+        let portal = make_portal(vec![Some(Type::FLOAT8)], vec![s("200")]);
+        let result = parameters_to_scalar_values(&plan, &portal);
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_float8_coerce_negative_to_unsigned_out_of_range() {
+        let plan = build_plan_with_params(vec![("$1", DataType::UInt64)]);
+        let portal = make_portal(vec![Some(Type::FLOAT8)], vec![s("-1")]);
+        let result = parameters_to_scalar_values(&plan, &portal);
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_null_parameter() {
+        let plan = build_plan_with_params(vec![("$1", DataType::Int8)]);
+        let portal = make_portal(vec![Some(Type::INT2)], vec![None]);
+
+        let values = parameters_to_scalar_values(&plan, &portal).unwrap();
+        assert_eq!(values[0], ScalarValue::Int8(None));
+    }
 }
diff --git a/src/servers/src/prom_row_builder.rs b/src/servers/src/prom_row_builder.rs
new file mode 100644
index 0000000000..0fddc0938a
--- /dev/null
+++ b/src/servers/src/prom_row_builder.rs
@@ -0,0 +1,557 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Prometheus row-level helpers for converting proto `Rows` into Arrow
+//! `RecordBatch`es and aligning / normalizing their schemas against
+//! existing table schemas in the catalog.
+
+use std::collections::{HashMap, HashSet};
+use std::sync::Arc;
+
+use api::helper::ColumnDataTypeWrapper;
+use api::v1::value::ValueData;
+use api::v1::{ColumnSchema, Rows, SemanticType};
+use arrow::array::{
+    ArrayRef, Float64Builder, StringBuilder, TimestampMicrosecondBuilder,
+    TimestampMillisecondBuilder, TimestampNanosecondBuilder, TimestampSecondBuilder,
+    new_null_array,
+};
+use arrow::datatypes::{DataType as ArrowDataType, Schema as ArrowSchema};
+use arrow::record_batch::RecordBatch;
+use arrow_schema::TimeUnit;
+use common_query::prelude::{greptime_timestamp, greptime_value};
+use datatypes::data_type::DataType;
+use datatypes::prelude::ConcreteDataType;
+use snafu::{OptionExt, ResultExt, ensure};
+
+use crate::error;
+use crate::error::Result;
+
+/// Extract timestamp, field, and tag column names from a logical region schema.
+fn unzip_logical_region_schema(
+    target_schema: &ArrowSchema,
+) -> Result<(String, String, HashSet<String>)> {
+    let mut timestamp_column = None;
+    let mut field_column = None;
+    let mut tag_columns = HashSet::with_capacity(target_schema.fields.len().saturating_sub(2));
+    for field in target_schema.fields() {
+        if field.name() == greptime_timestamp() {
+            timestamp_column = Some(field.name().clone());
+            continue;
+        }
+
+        if field.name() == greptime_value() {
+            field_column = Some(field.name().clone());
+            continue;
+        }
+
+        if timestamp_column.is_none() && matches!(field.data_type(), ArrowDataType::Timestamp(_, _))
+        {
+            timestamp_column = Some(field.name().clone());
+            continue;
+        }
+
+        if field_column.is_none() && matches!(field.data_type(), ArrowDataType::Float64) {
+            field_column = Some(field.name().clone());
+            continue;
+        }
+        tag_columns.insert(field.name().clone());
+    }
+
+    let timestamp_column = timestamp_column.with_context(|| error::UnexpectedResultSnafu {
+        reason: "Failed to locate timestamp column in target schema".to_string(),
+    })?;
+    let field_column = field_column.with_context(|| error::UnexpectedResultSnafu {
+        reason: "Failed to locate field column in target schema".to_string(),
+    })?;
+
+    Ok((timestamp_column, field_column, tag_columns))
+}
+
+/// Directly converts proto `Rows` into a `RecordBatch` aligned to the given
+/// `target_schema`, handling Prometheus column renaming (timestamp/value),
+/// reordering, type casting, and null-filling in a single pass.
+pub(crate) fn rows_to_aligned_record_batch(
+    rows: &Rows,
+    target_schema: &ArrowSchema,
+) -> Result<RecordBatch> {
+    let row_count = rows.rows.len();
+    let column_count = rows.schema.len();
+
+    for (idx, row) in rows.rows.iter().enumerate() {
+        ensure!(
+            row.values.len() == column_count,
+            error::InternalSnafu {
+                err_msg: format!(
+                    "Column count mismatch in row {}, expected {}, got {}",
+                    idx,
+                    column_count,
+                    row.values.len()
+                )
+            }
+        );
+    }
+
+    let (target_ts_name, target_field_name, _target_tags) =
+        unzip_logical_region_schema(target_schema)?;
+
+    // Map effective target column name → (source column index, source arrow type).
+    // Handles prom renames: Timestamp → target ts name, Float64 → target field name.
+    let mut source_map: HashMap<&str, (usize, ArrowDataType)> =
+        HashMap::with_capacity(rows.schema.len());
+
+    for (src_idx, col) in rows.schema.iter().enumerate() {
+        let wrapper = ColumnDataTypeWrapper::try_new(col.datatype, col.datatype_extension.clone())?;
+        let src_arrow_type = ConcreteDataType::from(wrapper).as_arrow_type();
+
+        match &src_arrow_type {
+            ArrowDataType::Float64 => {
+                source_map.insert(&target_field_name, (src_idx, src_arrow_type));
+            }
+            ArrowDataType::Timestamp(unit, _) => {
+                ensure!(
+                    unit == &TimeUnit::Millisecond,
+                    error::InvalidPromRemoteRequestSnafu {
+                        msg: format!(
+                            "Unexpected remote write batch timestamp unit, expect millisecond, got: {}",
+                            unit
+                        )
+                    }
+                );
+                source_map.insert(&target_ts_name, (src_idx, src_arrow_type));
+            }
+            ArrowDataType::Utf8 => {
+                source_map.insert(&col.column_name, (src_idx, src_arrow_type));
+            }
+            other => {
+                return error::InvalidPromRemoteRequestSnafu {
+                    msg: format!(
+                        "Unexpected remote write batch field type {}, field name: {}",
+                        other, col.column_name
+                    ),
+                }
+                .fail();
+            }
+        }
+    }
+
+    // Build columns in target schema order
+    let mut columns = Vec::with_capacity(target_schema.fields().len());
+    for target_field in target_schema.fields() {
+        if let Some((src_idx, src_arrow_type)) = source_map.get(target_field.name().as_str()) {
+            let array = build_arrow_array(
+                rows,
+                *src_idx,
+                &rows.schema[*src_idx].column_name,
+                src_arrow_type.clone(),
+                row_count,
+            )?;
+            columns.push(array);
+        } else {
+            columns.push(new_null_array(target_field.data_type(), row_count));
+        }
+    }
+
+    let batch = RecordBatch::try_new(Arc::new(target_schema.clone()), columns)
+        .context(error::ArrowSnafu)?;
+    Ok(batch)
+}
+
+/// Identify tag columns in the proto `rows_schema` that are absent from the
+/// target region schema, without building an intermediate `RecordBatch`.
+pub(crate) fn identify_missing_columns_from_proto(
+    rows_schema: &[ColumnSchema],
+    target_schema: &ArrowSchema,
+) -> Result<Vec<String>> {
+    let (_, _, target_tags) = unzip_logical_region_schema(target_schema)?;
+    let mut missing = Vec::new();
+    for col in rows_schema {
+        let wrapper = ColumnDataTypeWrapper::try_new(col.datatype, col.datatype_extension.clone())?;
+        let arrow_type = ConcreteDataType::from(wrapper).as_arrow_type();
+        if matches!(arrow_type, ArrowDataType::Utf8)
+            && !target_tags.contains(&col.column_name)
+            && target_schema.column_with_name(&col.column_name).is_none()
+        {
+            missing.push(col.column_name.clone());
+        }
+    }
+    Ok(missing)
+}
+
+/// Build a `Vec<ColumnSchema>` suitable for creating a new Prometheus logical table
+/// directly from the proto `rows.schema`, avoiding the round-trip through Arrow schema.
+pub fn build_prom_create_table_schema_from_proto(
+    rows_schema: &[ColumnSchema],
+) -> Result<Vec<ColumnSchema>> {
+    rows_schema
+        .iter()
+        .map(|col| {
+            let semantic_type = if col.datatype == api::v1::ColumnDataType::TimestampMillisecond as i32 {
+                SemanticType::Timestamp
+            } else if col.datatype == api::v1::ColumnDataType::Float64 as i32 {
+                SemanticType::Field
+            } else {
+                // tag columns must be String type
+                ensure!(col.datatype == api::v1::ColumnDataType::String as i32, error::InvalidPromRemoteRequestSnafu{
+                                        msg: format!(
+                        "Failed to build create table schema, tag column '{}' must be String but got datatype {}",
+                        col.column_name, col.datatype
+                    )
+                });
+                SemanticType::Tag
+            };
+
+            Ok(ColumnSchema {
+                column_name: col.column_name.clone(),
+                datatype: col.datatype,
+                semantic_type: semantic_type as i32,
+                datatype_extension: col.datatype_extension.clone(),
+                options: None,
+            })
+        })
+        .collect()
+}
+
+/// Build a single Arrow array for the given column index from proto `Rows`.
+fn build_arrow_array(
+    rows: &Rows,
+    col_idx: usize,
+    column_name: &String,
+    column_data_type: arrow::datatypes::DataType,
+    row_count: usize,
+) -> Result<ArrayRef> {
+    macro_rules! build_array {
+        ($builder:expr, $( $pattern:pat => $value:expr ),+ $(,)?) => {{
+            let mut builder = $builder;
+            for row in &rows.rows {
+                match row.values[col_idx].value_data.as_ref() {
+                    $(Some($pattern) => builder.append_value($value),)+
+                    Some(v) => {
+                        return error::InvalidPromRemoteRequestSnafu {
+                            msg: format!("Unexpected value: {:?}", v),
+                        }
+                        .fail();
+                    }
+                    None => builder.append_null(),
+                }
+            }
+            Arc::new(builder.finish()) as ArrayRef
+        }};
+    }
+
+    let array: ArrayRef = match column_data_type {
+        arrow::datatypes::DataType::Float64 => {
+            build_array!(Float64Builder::with_capacity(row_count), ValueData::F64Value(v) => *v)
+        }
+        arrow::datatypes::DataType::Utf8 => build_array!(
+            StringBuilder::with_capacity(row_count, 0),
+            ValueData::StringValue(v) => v
+        ),
+        arrow::datatypes::DataType::Timestamp(u, _) => match u {
+            TimeUnit::Second => build_array!(
+                TimestampSecondBuilder::with_capacity(row_count),
+                ValueData::TimestampSecondValue(v) => *v
+            ),
+            TimeUnit::Millisecond => build_array!(
+                TimestampMillisecondBuilder::with_capacity(row_count),
+                ValueData::TimestampMillisecondValue(v) => *v
+            ),
+            TimeUnit::Microsecond => build_array!(
+                TimestampMicrosecondBuilder::with_capacity(row_count),
+                ValueData::DatetimeValue(v) => *v,
+                ValueData::TimestampMicrosecondValue(v) => *v
+            ),
+            TimeUnit::Nanosecond => build_array!(
+                TimestampNanosecondBuilder::with_capacity(row_count),
+                ValueData::TimestampNanosecondValue(v) => *v
+            ),
+        },
+        ty => {
+            return error::InvalidPromRemoteRequestSnafu {
+                msg: format!(
+                    "Unexpected column type {:?}, column name: {}",
+                    ty, column_name
+                ),
+            }
+            .fail();
+        }
+    };
+
+    Ok(array)
+}
+
+#[cfg(test)]
+mod tests {
+    use api::v1::value::ValueData;
+    use api::v1::{ColumnDataType, ColumnSchema, Row, Rows, SemanticType, Value};
+    use arrow::array::{Array, Float64Array, StringArray, TimestampMillisecondArray};
+    use arrow::datatypes::{DataType, Field, Schema as ArrowSchema, TimeUnit};
+
+    use super::{
+        build_prom_create_table_schema_from_proto, identify_missing_columns_from_proto,
+        rows_to_aligned_record_batch,
+    };
+
+    #[test]
+    fn test_rows_to_aligned_record_batch_renames_and_reorders() {
+        let rows = Rows {
+            schema: vec![
+                ColumnSchema {
+                    column_name: "greptime_timestamp".to_string(),
+                    datatype: ColumnDataType::TimestampMillisecond as i32,
+                    semantic_type: SemanticType::Timestamp as i32,
+                    ..Default::default()
+                },
+                ColumnSchema {
+                    column_name: "host".to_string(),
+                    datatype: ColumnDataType::String as i32,
+                    semantic_type: SemanticType::Tag as i32,
+                    ..Default::default()
+                },
+                ColumnSchema {
+                    column_name: "greptime_value".to_string(),
+                    datatype: ColumnDataType::Float64 as i32,
+                    semantic_type: SemanticType::Field as i32,
+                    ..Default::default()
+                },
+            ],
+            rows: vec![
+                Row {
+                    values: vec![
+                        Value {
+                            value_data: Some(ValueData::TimestampMillisecondValue(1000)),
+                        },
+                        Value {
+                            value_data: Some(ValueData::StringValue("h1".to_string())),
+                        },
+                        Value {
+                            value_data: Some(ValueData::F64Value(42.0)),
+                        },
+                    ],
+                },
+                Row {
+                    values: vec![
+                        Value {
+                            value_data: Some(ValueData::TimestampMillisecondValue(2000)),
+                        },
+                        Value {
+                            value_data: Some(ValueData::StringValue("h2".to_string())),
+                        },
+                        Value {
+                            value_data: Some(ValueData::F64Value(99.0)),
+                        },
+                    ],
+                },
+            ],
+        };
+
+        // Target schema has renamed columns and different ordering.
+        let target = ArrowSchema::new(vec![
+            Field::new(
+                "my_ts",
+                DataType::Timestamp(TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new("host", DataType::Utf8, true),
+            Field::new("my_value", DataType::Float64, true),
+        ]);
+
+        let batch = rows_to_aligned_record_batch(&rows, &target).unwrap();
+        assert_eq!(batch.schema().as_ref(), &target);
+        assert_eq!(2, batch.num_rows());
+        assert_eq!(3, batch.num_columns());
+
+        let ts = batch
+            .column(0)
+            .as_any()
+            .downcast_ref::<TimestampMillisecondArray>()
+            .unwrap();
+        assert_eq!(ts.value(0), 1000);
+        assert_eq!(ts.value(1), 2000);
+
+        let hosts = batch
+            .column(1)
+            .as_any()
+            .downcast_ref::<StringArray>()
+            .unwrap();
+        assert_eq!(hosts.value(0), "h1");
+        assert_eq!(hosts.value(1), "h2");
+
+        let values = batch
+            .column(2)
+            .as_any()
+            .downcast_ref::<Float64Array>()
+            .unwrap();
+        assert_eq!(values.value(0), 42.0);
+        assert_eq!(values.value(1), 99.0);
+    }
+
+    #[test]
+    fn test_rows_to_aligned_record_batch_fills_nulls() {
+        let rows = Rows {
+            schema: vec![
+                ColumnSchema {
+                    column_name: "greptime_timestamp".to_string(),
+                    datatype: ColumnDataType::TimestampMillisecond as i32,
+                    semantic_type: SemanticType::Timestamp as i32,
+                    ..Default::default()
+                },
+                ColumnSchema {
+                    column_name: "host".to_string(),
+                    datatype: ColumnDataType::String as i32,
+                    semantic_type: SemanticType::Tag as i32,
+                    ..Default::default()
+                },
+                ColumnSchema {
+                    column_name: "instance".to_string(),
+                    datatype: ColumnDataType::String as i32,
+                    semantic_type: SemanticType::Tag as i32,
+                    ..Default::default()
+                },
+                ColumnSchema {
+                    column_name: "greptime_value".to_string(),
+                    datatype: ColumnDataType::Float64 as i32,
+                    semantic_type: SemanticType::Field as i32,
+                    ..Default::default()
+                },
+            ],
+            rows: vec![Row {
+                values: vec![
+                    Value {
+                        value_data: Some(ValueData::TimestampMillisecondValue(1000)),
+                    },
+                    Value {
+                        value_data: Some(ValueData::StringValue("h1".to_string())),
+                    },
+                    Value {
+                        value_data: Some(ValueData::StringValue("i1".to_string())),
+                    },
+                    Value {
+                        value_data: Some(ValueData::F64Value(1.0)),
+                    },
+                ],
+            }],
+        };
+
+        // Target schema has "host" but not "instance"; also has "region" which is missing from source.
+        let target = ArrowSchema::new(vec![
+            Field::new(
+                "my_ts",
+                DataType::Timestamp(TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new("host", DataType::Utf8, true),
+            Field::new("region", DataType::Utf8, true),
+            Field::new("my_value", DataType::Float64, true),
+        ]);
+
+        let batch = rows_to_aligned_record_batch(&rows, &target).unwrap();
+        assert_eq!(batch.schema().as_ref(), &target);
+        assert_eq!(1, batch.num_rows());
+        assert_eq!(4, batch.num_columns());
+
+        // "region" column should be null-filled.
+        let region = batch
+            .column(2)
+            .as_any()
+            .downcast_ref::<StringArray>()
+            .unwrap();
+        assert!(region.is_null(0));
+    }
+
+    #[test]
+    fn test_identify_missing_columns_from_proto() {
+        let rows_schema = vec![
+            ColumnSchema {
+                column_name: "greptime_timestamp".to_string(),
+                datatype: ColumnDataType::TimestampMillisecond as i32,
+                semantic_type: SemanticType::Timestamp as i32,
+                ..Default::default()
+            },
+            ColumnSchema {
+                column_name: "host".to_string(),
+                datatype: ColumnDataType::String as i32,
+                semantic_type: SemanticType::Tag as i32,
+                ..Default::default()
+            },
+            ColumnSchema {
+                column_name: "instance".to_string(),
+                datatype: ColumnDataType::String as i32,
+                semantic_type: SemanticType::Tag as i32,
+                ..Default::default()
+            },
+            ColumnSchema {
+                column_name: "greptime_value".to_string(),
+                datatype: ColumnDataType::Float64 as i32,
+                semantic_type: SemanticType::Field as i32,
+                ..Default::default()
+            },
+        ];
+
+        let target = ArrowSchema::new(vec![
+            Field::new(
+                "my_ts",
+                DataType::Timestamp(TimeUnit::Millisecond, None),
+                false,
+            ),
+            Field::new("host", DataType::Utf8, true),
+            Field::new("my_value", DataType::Float64, true),
+        ]);
+
+        let missing = identify_missing_columns_from_proto(&rows_schema, &target).unwrap();
+        assert_eq!(missing, vec!["instance".to_string()]);
+    }
+
+    #[test]
+    fn test_build_prom_create_table_schema_from_proto() {
+        let rows_schema = vec![
+            ColumnSchema {
+                column_name: "greptime_timestamp".to_string(),
+                datatype: ColumnDataType::TimestampMillisecond as i32,
+                semantic_type: SemanticType::Timestamp as i32,
+                ..Default::default()
+            },
+            ColumnSchema {
+                column_name: "job".to_string(),
+                datatype: ColumnDataType::String as i32,
+                semantic_type: SemanticType::Tag as i32,
+                ..Default::default()
+            },
+            ColumnSchema {
+                column_name: "greptime_value".to_string(),
+                datatype: ColumnDataType::Float64 as i32,
+                semantic_type: SemanticType::Field as i32,
+                ..Default::default()
+            },
+        ];
+
+        let schema = build_prom_create_table_schema_from_proto(&rows_schema).unwrap();
+        assert_eq!(3, schema.len());
+
+        assert_eq!("greptime_timestamp", schema[0].column_name);
+        assert_eq!(SemanticType::Timestamp as i32, schema[0].semantic_type);
+        assert_eq!(
+            ColumnDataType::TimestampMillisecond as i32,
+            schema[0].datatype
+        );
+
+        assert_eq!("job", schema[1].column_name);
+        assert_eq!(SemanticType::Tag as i32, schema[1].semantic_type);
+        assert_eq!(ColumnDataType::String as i32, schema[1].datatype);
+
+        assert_eq!("greptime_value", schema[2].column_name);
+        assert_eq!(SemanticType::Field as i32, schema[2].semantic_type);
+        assert_eq!(ColumnDataType::Float64 as i32, schema[2].datatype);
+    }
+}
diff --git a/src/servers/src/query_handler.rs b/src/servers/src/query_handler.rs
index 60efe69faa..d4b272de12 100644
--- a/src/servers/src/query_handler.rs
+++ b/src/servers/src/query_handler.rs
@@ -44,6 +44,12 @@ use pipeline::{GreptimePipelineParams, Pipeline, PipelineInfo, PipelineVersion,
 use serde_json::Value;
 use session::context::{QueryContext, QueryContextRef};
 
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+pub struct DashboardDefinition {
+    pub name: String,
+    pub definition: String,
+}
+
 use crate::error::Result;
 use crate::http::jaeger::QueryTraceParams;
 use crate::influxdb::InfluxdbRequest;
@@ -57,6 +63,14 @@ pub type PipelineHandlerRef = Arc<dyn PipelineHandler + Send + Sync>;
 pub type LogQueryHandlerRef = Arc<dyn LogQueryHandler + Send + Sync>;
 pub type JaegerQueryHandlerRef = Arc<dyn JaegerQueryHandler + Send + Sync>;
 
+#[derive(Debug, Default, Clone)]
+pub struct TraceIngestOutcome {
+    pub write_cost: usize,
+    pub accepted_spans: usize,
+    pub rejected_spans: usize,
+    pub error_message: Option<String>,
+}
+
 #[async_trait]
 pub trait InfluxdbLineProtocolHandler {
     /// A successful request will not return a response.
@@ -80,6 +94,11 @@ pub struct PromStoreResponse {
 
 #[async_trait]
 pub trait PromStoreProtocolHandler {
+    /// Runs pre-write checks/hooks for prometheus remote write requests.
+    async fn pre_write(&self, _request: &RowInsertRequests, _ctx: QueryContextRef) -> Result<()> {
+        Ok(())
+    }
+
     /// Handling prometheus remote write requests
     async fn write(
         &self,
@@ -112,7 +131,7 @@ pub trait OpenTelemetryProtocolHandler: PipelineHandler {
         pipeline_params: GreptimePipelineParams,
         table_name: String,
         ctx: QueryContextRef,
-    ) -> Result<Output>;
+    ) -> Result<TraceIngestOutcome>;
 
     async fn logs(
         &self,
@@ -176,6 +195,18 @@ pub trait PipelineHandler {
     ) -> Result<(String, TimestampNanosecond)>;
 }
 
+/// Handling dashboard as code CRUD
+pub type DashboardHandlerRef = Arc<dyn DashboardHandler + Send + Sync>;
+
+#[async_trait]
+pub trait DashboardHandler {
+    async fn save(&self, name: &str, definition: &str, ctx: QueryContextRef) -> Result<()>;
+
+    async fn list(&self, ctx: QueryContextRef) -> Result<Vec<DashboardDefinition>>;
+
+    async fn delete(&self, name: &str, ctx: QueryContextRef) -> Result<()>;
+}
+
 /// Handle log query requests.
 #[async_trait]
 pub trait LogQueryHandler {
diff --git a/src/servers/src/query_handler/grpc.rs b/src/servers/src/query_handler/grpc.rs
index 67d8b3890e..d66a76464e 100644
--- a/src/servers/src/query_handler/grpc.rs
+++ b/src/servers/src/query_handler/grpc.rs
@@ -17,15 +17,13 @@ use std::sync::Arc;
 
 use api::v1::greptime_request::Request;
 use async_trait::async_trait;
-use common_base::AffectedRows;
 use common_grpc::flight::do_put::DoPutResponse;
 use common_query::Output;
 use futures::Stream;
 use session::context::QueryContextRef;
-use table::TableRef;
 
 use crate::error::Result;
-use crate::grpc::flight::{PutRecordBatchRequest, PutRecordBatchRequestStream};
+use crate::grpc::flight::PutRecordBatchRequestStream;
 
 pub type ServerGrpcQueryHandlerRef = Arc<dyn GrpcQueryHandler + Send + Sync>;
 
@@ -35,13 +33,6 @@ pub type RawRecordBatch = bytes::Bytes;
 pub trait GrpcQueryHandler {
     async fn do_query(&self, query: Request, ctx: QueryContextRef) -> Result<Output>;
 
-    async fn put_record_batch(
-        &self,
-        request: PutRecordBatchRequest,
-        table_ref: &mut Option<TableRef>,
-        ctx: QueryContextRef,
-    ) -> Result<AffectedRows>;
-
     fn handle_put_record_batch_stream(
         &self,
         stream: PutRecordBatchRequestStream,
diff --git a/src/servers/src/query_handler/sql.rs b/src/servers/src/query_handler/sql.rs
index 544f5bceb6..eb72190309 100644
--- a/src/servers/src/query_handler/sql.rs
+++ b/src/servers/src/query_handler/sql.rs
@@ -32,8 +32,8 @@ pub trait SqlQueryHandler {
 
     async fn do_exec_plan(
         &self,
-        stmt: Option<Statement>,
         plan: LogicalPlan,
+        stmt: Option<Statement>,
         query_ctx: QueryContextRef,
     ) -> Result<Output>;
 
diff --git a/src/servers/src/request_memory_metrics.rs b/src/servers/src/request_memory_metrics.rs
index 4298830f18..68f52816f4 100644
--- a/src/servers/src/request_memory_metrics.rs
+++ b/src/servers/src/request_memory_metrics.rs
@@ -34,7 +34,7 @@ impl MemoryMetrics for RequestMemoryMetrics {
         REQUEST_MEMORY_IN_USE.set(bytes);
     }
 
-    fn inc_rejected(&self, reason: &str) {
+    fn inc_exhausted(&self, reason: &str) {
         REQUEST_MEMORY_REJECTED.with_label_values(&[reason]).inc();
     }
 }
diff --git a/src/servers/src/row_writer.rs b/src/servers/src/row_writer.rs
index ec439a8659..fca2c21b41 100644
--- a/src/servers/src/row_writer.rs
+++ b/src/servers/src/row_writer.rs
@@ -77,6 +77,35 @@ impl TableData {
     pub fn into_schema_and_rows(self) -> (Vec<ColumnSchema>, Vec<Row>) {
         (self.schema, self.rows)
     }
+
+    /// Writes a field value without enforcing that later writes use the same datatype
+    /// as the first-seen schema entry.
+    ///
+    /// The OTLP trace v1 path uses this to preserve raw mixed values inside one request
+    /// so the frontend can reconcile them later against both the full batch and the
+    /// existing table schema.
+    pub fn write_field_unchecked(
+        &mut self,
+        name: impl ToString,
+        datatype: ColumnDataType,
+        value: Option<ValueData>,
+        one_row: &mut Vec<Value>,
+    ) {
+        let name = name.to_string();
+        if let Some(index) = self.column_indexes.get(&name).copied() {
+            one_row[index].value_data = value;
+        } else {
+            let index = self.schema.len();
+            self.schema.push(ColumnSchema {
+                column_name: name.clone(),
+                datatype: datatype as i32,
+                semantic_type: SemanticType::Field as i32,
+                ..Default::default()
+            });
+            self.column_indexes.insert(name, index);
+            one_row.push(Value { value_data: value });
+        }
+    }
 }
 
 pub struct MultiTableData {
diff --git a/src/servers/src/tls.rs b/src/servers/src/tls.rs
index 5f461731ef..51f5818b49 100644
--- a/src/servers/src/tls.rs
+++ b/src/servers/src/tls.rs
@@ -239,7 +239,7 @@ pub fn maybe_watch_server_tls_config(
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::install_ring_crypto_provider;
+    use crate::install_default_crypto_provider;
     use crate::tls::TlsMode::Disable;
 
     #[test]
@@ -510,7 +510,7 @@ mod tests {
     #[test]
     fn test_tls_file_change_watch() {
         common_telemetry::init_default_ut_logging();
-        let _ = install_ring_crypto_provider();
+        let _ = install_default_crypto_provider();
 
         let dir = tempfile::tempdir().unwrap();
         let cert_path = dir.path().join("server.crt");
diff --git a/src/servers/tests/http/authorize.rs b/src/servers/tests/http/authorize.rs
index 214ab7d9f9..be54d47e78 100644
--- a/src/servers/tests/http/authorize.rs
+++ b/src/servers/tests/http/authorize.rs
@@ -17,14 +17,15 @@ use std::sync::Arc;
 use auth::UserProvider;
 use auth::tests::MockUserProvider;
 use axum::http;
+use base64::prelude::{BASE64_STANDARD, Engine as _};
 use hyper::{Request, StatusCode};
 use servers::http::AUTHORIZATION_HEADER;
 use servers::http::authorize::inner_auth;
 use session::context::QueryContext;
 
 async fn check_http_auth(header_key: &str) {
-    // base64encode("username:password") == "dXNlcm5hbWU6cGFzc3dvcmQ="
-    let req = mock_http_request(header_key, Some("Basic dXNlcm5hbWU6cGFzc3dvcmQ="), None).unwrap();
+    let req =
+        mock_http_request(header_key, Some(&basic_auth("username", "password")), None).unwrap();
     let req = inner_auth(None, req).await.unwrap();
     let ctx: &QueryContext = req.extensions().get().unwrap();
     let user_info = ctx.current_user();
@@ -34,8 +35,8 @@ async fn check_http_auth(header_key: &str) {
     // In mock user provider, right username:password == "greptime:greptime"
     let mock_user_provider = Some(Arc::new(MockUserProvider::default()) as Arc<dyn UserProvider>);
 
-    // base64encode("greptime:greptime") == "Z3JlcHRpbWU6Z3JlcHRpbWU="
-    let req = mock_http_request(header_key, Some("Basic Z3JlcHRpbWU6Z3JlcHRpbWU="), None).unwrap();
+    let req =
+        mock_http_request(header_key, Some(&basic_auth("greptime", "greptime")), None).unwrap();
     let req = inner_auth(mock_user_provider.clone(), req).await.unwrap();
     let ctx: &QueryContext = req.extensions().get().unwrap();
     let user_info = ctx.current_user();
@@ -52,9 +53,8 @@ async fn check_http_auth(header_key: &str) {
         axum::body::to_bytes(resp.into_body(), usize::MAX).await.unwrap().as_ref()
     );
 
-    // base64encode("username:password") == "dXNlcm5hbWU6cGFzc3dvcmQ="
     let wrong_req =
-        mock_http_request(header_key, Some("Basic dXNlcm5hbWU6cGFzc3dvcmQ="), None).unwrap();
+        mock_http_request(header_key, Some(&basic_auth("username", "password")), None).unwrap();
     let auth_res = inner_auth(mock_user_provider, wrong_req).await;
     assert!(auth_res.is_err());
     let resp = auth_res.unwrap_err();
@@ -78,12 +78,11 @@ async fn check_schema_validating(header: &str) {
     // In mock user provider, right username:password == "greptime:greptime"
     let mock_user_provider = Some(Arc::new(MockUserProvider::default()) as Arc<dyn UserProvider>);
 
-    // base64encode("greptime:greptime") == "Z3JlcHRpbWU6Z3JlcHRpbWU="
     // http://localhost/{http_api_version}/sql?db=greptime
     let version = servers::http::HTTP_API_VERSION;
     let req = mock_http_request(
         header,
-        Some("Basic Z3JlcHRpbWU6Z3JlcHRpbWU="),
+        Some(&basic_auth("greptime", "greptime")),
         Some(format!("http://localhost/{version}/sql?db=public").as_str()),
     )
     .unwrap();
@@ -96,7 +95,7 @@ async fn check_schema_validating(header: &str) {
     // wrong database
     let req = mock_http_request(
         header,
-        Some("Basic Z3JlcHRpbWU6Z3JlcHRpbWU="),
+        Some(&basic_auth("greptime", "greptime")),
         Some(format!("http://localhost/{version}/sql?db=wrong").as_str()),
     )
     .unwrap();
@@ -120,7 +119,6 @@ async fn check_auth_header(header_key: &str) {
     // In mock user provider, right username:password == "greptime:greptime"
     let mock_user_provider = Some(Arc::new(MockUserProvider::default()) as Arc<dyn UserProvider>);
 
-    // base64encode("greptime:greptime") == "Z3JlcHRpbWU6Z3JlcHRpbWU="
     // try auth path first
     let req = mock_http_request(header_key, None, None).unwrap();
     let auth_res = inner_auth(mock_user_provider.clone(), req).await;
@@ -144,6 +142,13 @@ async fn test_whitelist_no_auth() {
     check_auth_header(AUTHORIZATION_HEADER).await;
 }
 
+fn basic_auth(username: &str, password: &str) -> String {
+    format!(
+        "Basic {}",
+        BASE64_STANDARD.encode(format!("{username}:{password}"))
+    )
+}
+
 // copy from http::authorize
 fn mock_http_request(
     auth_header_key: &str,
diff --git a/src/servers/tests/http/influxdb_test.rs b/src/servers/tests/http/influxdb_test.rs
index 2939d2a11c..59ec9c138c 100644
--- a/src/servers/tests/http/influxdb_test.rs
+++ b/src/servers/tests/http/influxdb_test.rs
@@ -18,6 +18,7 @@ use api::v1::RowInsertRequests;
 use async_trait::async_trait;
 use auth::tests::{DatabaseAuthInfo, MockUserProvider};
 use axum::{Router, http};
+use base64::prelude::{BASE64_STANDARD, Engine as _};
 use common_query::Output;
 use common_test_util::ports;
 use datafusion_expr::LogicalPlan;
@@ -34,6 +35,13 @@ use session::context::QueryContextRef;
 use sql::statements::statement::Statement;
 use tokio::sync::mpsc;
 
+fn basic_auth(username: &str, password: &str) -> String {
+    format!(
+        "basic {}",
+        BASE64_STANDARD.encode(format!("{username}:{password}"))
+    )
+}
+
 struct DummyInstance {
     tx: Arc<mpsc::Sender<(String, String)>>,
 }
@@ -58,8 +66,8 @@ impl SqlQueryHandler for DummyInstance {
 
     async fn do_exec_plan(
         &self,
-        _stmt: Option<Statement>,
         _plan: LogicalPlan,
+        _stmt: Option<Statement>,
         _query_ctx: QueryContextRef,
     ) -> Result<Output> {
         unimplemented!()
@@ -146,7 +154,7 @@ async fn test_influxdb_write() {
         .body("monitor,host=host1 cpu=1.2 1664370459457010101")
         .header(
             http::header::AUTHORIZATION,
-            "basic Z3JlcHRpbWU6Z3JlcHRpbWU=",
+            basic_auth("greptime", "greptime"),
         )
         .send()
         .await;
diff --git a/src/servers/tests/http/opentsdb_test.rs b/src/servers/tests/http/opentsdb_test.rs
index a7a044d4db..49acd8625b 100644
--- a/src/servers/tests/http/opentsdb_test.rs
+++ b/src/servers/tests/http/opentsdb_test.rs
@@ -58,8 +58,8 @@ impl SqlQueryHandler for DummyInstance {
 
     async fn do_exec_plan(
         &self,
-        _stmt: Option<Statement>,
         _plan: LogicalPlan,
+        _stmt: Option<Statement>,
         _query_ctx: QueryContextRef,
     ) -> Result<Output> {
         unimplemented!()
diff --git a/src/servers/tests/http/prom_store_test.rs b/src/servers/tests/http/prom_store_test.rs
index b1e974d3d3..0da9f70cb7 100644
--- a/src/servers/tests/http/prom_store_test.rs
+++ b/src/servers/tests/http/prom_store_test.rs
@@ -87,8 +87,8 @@ impl SqlQueryHandler for DummyInstance {
 
     async fn do_exec_plan(
         &self,
-        _stmt: Option<Statement>,
         _plan: LogicalPlan,
+        _stmt: Option<Statement>,
         _query_ctx: QueryContextRef,
     ) -> Result<Output> {
         unimplemented!()
@@ -120,7 +120,7 @@ fn make_test_app(tx: mpsc::Sender<(String, Vec<u8>)>) -> Router {
     let instance = Arc::new(DummyInstance { tx });
     let server = HttpServerBuilder::new(http_opts)
         .with_sql_handler(instance.clone())
-        .with_prom_handler(instance, None, true, PromValidationMode::Unchecked)
+        .with_prom_handler(instance, None, true, PromValidationMode::Unchecked, None)
         .build();
     server.build(server.make_app()).unwrap()
 }
diff --git a/src/servers/tests/mod.rs b/src/servers/tests/mod.rs
index e3f8f8fc79..767a5aa26c 100644
--- a/src/servers/tests/mod.rs
+++ b/src/servers/tests/mod.rs
@@ -18,20 +18,22 @@ use api::v1::greptime_request::Request;
 use api::v1::query_request::Query;
 use async_trait::async_trait;
 use catalog::memory::MemoryCatalogManager;
-use common_base::AffectedRows;
 use common_catalog::consts::{DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME};
+use common_error::ext::BoxedError;
 use common_grpc::flight::do_put::DoPutResponse;
 use common_query::Output;
 use datafusion_expr::LogicalPlan;
+use futures_util::TryFutureExt;
 use query::options::QueryOptions;
-use query::parser::{PromQuery, QueryLanguageParser, QueryStatement};
+use query::parser::{PromQuery, QueryStatement};
 use query::query_engine::DescribeResult;
 use query::{QueryEngineFactory, QueryEngineRef};
-use servers::error::{NotSupportedSnafu, Result};
+use servers::error::{ExecuteQuerySnafu, NotSupportedSnafu, Result};
 use servers::query_handler::grpc::GrpcQueryHandler;
 use servers::query_handler::sql::{ServerSqlQueryHandlerRef, SqlQueryHandler};
 use session::context::QueryContextRef;
-use snafu::ensure;
+use snafu::{ResultExt, ensure};
+use sql::parser::{ParseOptions, ParserContext};
 use sql::statements::statement::Statement;
 use table::TableRef;
 
@@ -53,21 +55,35 @@ impl DummyInstance {
 #[async_trait]
 impl SqlQueryHandler for DummyInstance {
     async fn do_query(&self, query: &str, query_ctx: QueryContextRef) -> Vec<Result<Output>> {
-        let stmt = QueryLanguageParser::parse_sql(query, &query_ctx).unwrap();
-        let plan = self
-            .query_engine
-            .planner()
-            .plan(&stmt, query_ctx.clone())
-            .await
-            .unwrap();
-        let output = self.query_engine.execute(plan, query_ctx).await.unwrap();
-        vec![Ok(output)]
+        let mut results = vec![];
+
+        let statements = ParserContext::create_with_dialect(
+            query,
+            query_ctx.sql_dialect(),
+            ParseOptions::default(),
+        )
+        .map(|x| x.into_iter().map(QueryStatement::Sql).collect::<Vec<_>>())
+        .unwrap();
+
+        for statement in &statements {
+            let result = self
+                .query_engine
+                .planner()
+                .plan(statement, query_ctx.clone())
+                .and_then(|plan| self.query_engine.execute(plan, query_ctx.clone()))
+                .await
+                .map_err(BoxedError::new)
+                .context(ExecuteQuerySnafu);
+            results.push(result);
+        }
+
+        results
     }
 
     async fn do_exec_plan(
         &self,
-        _stmt: Option<Statement>,
         plan: LogicalPlan,
+        _stmt: Option<Statement>,
         query_ctx: QueryContextRef,
     ) -> Result<Output> {
         Ok(self.query_engine.execute(plan, query_ctx).await.unwrap())
@@ -149,15 +165,6 @@ impl GrpcQueryHandler for DummyInstance {
         Ok(output)
     }
 
-    async fn put_record_batch(
-        &self,
-        _request: servers::grpc::flight::PutRecordBatchRequest,
-        _table_ref: &mut Option<TableRef>,
-        _ctx: QueryContextRef,
-    ) -> Result<AffectedRows> {
-        unimplemented!()
-    }
-
     fn handle_put_record_batch_stream(
         &self,
         _stream: servers::grpc::flight::PutRecordBatchRequestStream,
diff --git a/src/servers/tests/mysql/mysql_server_test.rs b/src/servers/tests/mysql/mysql_server_test.rs
index 3aa7b98f39..e0cb086dda 100644
--- a/src/servers/tests/mysql/mysql_server_test.rs
+++ b/src/servers/tests/mysql/mysql_server_test.rs
@@ -27,7 +27,7 @@ use datatypes::value::Value;
 use mysql_async::prelude::*;
 use mysql_async::{Conn, Row, SslOpts};
 use servers::error::Result;
-use servers::install_ring_crypto_provider;
+use servers::install_default_crypto_provider;
 use servers::mysql::server::{MysqlServer, MysqlSpawnConfig, MysqlSpawnRef};
 use servers::server::Server;
 use servers::tls::{ReloadableTlsServerConfig, TlsOption};
@@ -45,7 +45,7 @@ struct MysqlOpts<'a> {
 }
 
 fn create_mysql_server(table: TableRef, opts: MysqlOpts<'_>) -> Result<Box<dyn Server>> {
-    let _ = install_ring_crypto_provider();
+    let _ = install_default_crypto_provider();
     let query_handler = create_testing_sql_query_handler(table);
     let io_runtime = RuntimeBuilder::default()
         .worker_threads(4)
@@ -505,6 +505,17 @@ async fn test_query_prepared() -> Result<()> {
 
     test_prepare_all_type(column_schemas, columns, &mut connection).await;
 
+    match connection
+        .prep("SELECT `timestamp` FROM t WHERE `timestamp` > NOW() - INTERVAL '1 hour'")
+        .await
+    {
+        Err(mysql_async::Error::Server(e)) => assert_eq!(
+            "ERROR HY000 (1210): (InvalidArguments): Invalid prepare statement: Invalid SQL syntax: sql parser error: INTERVAL requires a unit after the literal value",
+            e.to_string()
+        ),
+        _ => unreachable!(),
+    }
+
     Ok(())
 }
 
@@ -514,8 +525,7 @@ async fn test_prepare_all_type(
     connection: &mut Conn,
 ) {
     let mut column_index = 0;
-    let mut stmt_id = 1;
-    for schema in column_schemas {
+    for (stmt_id, schema) in (1..).zip(column_schemas) {
         let query = format!(
             "SELECT {} FROM all_datatypes WHERE {} = ?",
             schema.name, schema.name
@@ -523,7 +533,6 @@ async fn test_prepare_all_type(
         let statement = connection.prep(query).await;
         let statement = statement.unwrap();
         assert_eq!(stmt_id, statement.id());
-        stmt_id += 1;
 
         let vector_ref = columns.get(column_index).unwrap();
         for vector_index in 0..vector_ref.len() {
diff --git a/src/servers/tests/postgres/mod.rs b/src/servers/tests/postgres/mod.rs
index 311e630e0e..5a3e97e3e7 100644
--- a/src/servers/tests/postgres/mod.rs
+++ b/src/servers/tests/postgres/mod.rs
@@ -27,7 +27,7 @@ use rustls::client::danger::{ServerCertVerified, ServerCertVerifier};
 use rustls::{Error, SignatureScheme};
 use rustls_pki_types::{CertificateDer, ServerName};
 use servers::error::Result;
-use servers::install_ring_crypto_provider;
+use servers::install_default_crypto_provider;
 use servers::postgres::PostgresServer;
 use servers::server::Server;
 use servers::tls::{ReloadableTlsServerConfig, TlsOption};
@@ -365,7 +365,7 @@ async fn test_extended_query() -> Result<()> {
 
 async fn start_test_server(server_tls: TlsOption) -> Result<u16> {
     common_telemetry::init_default_ut_logging();
-    let _ = install_ring_crypto_provider();
+    let _ = install_default_crypto_provider();
 
     let table = MemTable::default_numbers_table();
     let mut pg_server = create_postgres_server(table, false, server_tls, None)?;
diff --git a/src/session/src/context.rs b/src/session/src/context.rs
index 2b9483aca8..5f16ea8b5a 100644
--- a/src/session/src/context.rs
+++ b/src/session/src/context.rs
@@ -433,10 +433,21 @@ impl QueryContext {
         self.snapshot_seqs.read().unwrap().clone()
     }
 
+    pub fn sst_min_sequences(&self) -> HashMap<u64, u64> {
+        self.sst_min_sequences.read().unwrap().clone()
+    }
+
     pub fn get_snapshot(&self, region_id: u64) -> Option<u64> {
         self.snapshot_seqs.read().unwrap().get(&region_id).cloned()
     }
 
+    pub fn set_snapshot(&self, region_id: u64, sequence: u64) {
+        self.snapshot_seqs
+            .write()
+            .unwrap()
+            .insert(region_id, sequence);
+    }
+
     /// Returns `true` if the session can cast strings to numbers in MySQL style.
     pub fn auto_string_to_numeric(&self) -> bool {
         matches!(self.channel, Channel::Mysql)
@@ -669,6 +680,8 @@ impl ConfigurationVariables {
 
 #[cfg(test)]
 mod test {
+    use std::collections::HashMap;
+
     use common_catalog::consts::DEFAULT_CATALOG_NAME;
 
     use super::*;
@@ -704,4 +717,30 @@ mod test {
         let context = QueryContext::with(DEFAULT_CATALOG_NAME, "test");
         assert_eq!("test", context.get_db_string());
     }
+
+    #[test]
+    fn test_api_query_context_roundtrip_with_sequences() {
+        let api_ctx = api::v1::QueryContext {
+            current_catalog: "c1".to_string(),
+            current_schema: "s1".to_string(),
+            timezone: "UTC".to_string(),
+            extensions: HashMap::from([("flow.return_region_seq".to_string(), "true".to_string())]),
+            channel: Channel::Grpc as u32,
+            snapshot_seqs: Some(api::v1::SnapshotSequences {
+                snapshot_seqs: HashMap::from([(1, 100)]),
+                sst_min_sequences: HashMap::from([(1, 90)]),
+            }),
+            explain: None,
+        };
+
+        let session_ctx: QueryContext = api_ctx.clone().into();
+        let roundtrip_api: api::v1::QueryContext = session_ctx.into();
+
+        assert_eq!(roundtrip_api.current_catalog, api_ctx.current_catalog);
+        assert_eq!(roundtrip_api.current_schema, api_ctx.current_schema);
+        assert_eq!(roundtrip_api.timezone, api_ctx.timezone);
+        assert_eq!(roundtrip_api.extensions, api_ctx.extensions);
+        assert_eq!(roundtrip_api.channel, api_ctx.channel);
+        assert_eq!(roundtrip_api.snapshot_seqs, api_ctx.snapshot_seqs);
+    }
 }
diff --git a/src/sql/src/lib.rs b/src/sql/src/lib.rs
index b2b151d439..e8c6bdf8ef 100644
--- a/src/sql/src/lib.rs
+++ b/src/sql/src/lib.rs
@@ -13,8 +13,6 @@
 // limitations under the License.
 
 #![feature(box_patterns)]
-#![feature(assert_matches)]
-#![feature(if_let_guard)]
 
 pub mod ast;
 pub mod dialect;
diff --git a/src/sql/src/parsers/alter_parser.rs b/src/sql/src/parsers/alter_parser.rs
index df89f90922..e5e1575a20 100644
--- a/src/sql/src/parsers/alter_parser.rs
+++ b/src/sql/src/parsers/alter_parser.rs
@@ -677,7 +677,7 @@ fn parse_string_option_names(parser: &mut Parser) -> std::result::Result<String,
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_error::ext::ErrorExt;
     use datatypes::schema::{FulltextAnalyzer, FulltextBackend, FulltextOptions};
diff --git a/src/sql/src/parsers/comment_parser.rs b/src/sql/src/parsers/comment_parser.rs
index bd2cb53c5d..4e35e414ac 100644
--- a/src/sql/src/parsers/comment_parser.rs
+++ b/src/sql/src/parsers/comment_parser.rs
@@ -130,7 +130,7 @@ impl ParserContext<'_> {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use crate::dialect::GreptimeDbDialect;
     use crate::parser::{ParseOptions, ParserContext};
diff --git a/src/sql/src/parsers/copy_parser.rs b/src/sql/src/parsers/copy_parser.rs
index d975d884f6..9a2eddcc78 100644
--- a/src/sql/src/parsers/copy_parser.rs
+++ b/src/sql/src/parsers/copy_parser.rs
@@ -267,7 +267,7 @@ impl ParserContext<'_> {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
 
     use sqlparser::ast::{Ident, ObjectName};
diff --git a/src/sql/src/parsers/create_parser.rs b/src/sql/src/parsers/create_parser.rs
index d69a1af61d..a82590c603 100644
--- a/src/sql/src/parsers/create_parser.rs
+++ b/src/sql/src/parsers/create_parser.rs
@@ -68,6 +68,17 @@ pub const VECTOR: &str = "VECTOR";
 
 pub type RawIntervalExpr = String;
 
+// Preserve raw CREATE FLOW option entries until operator-side validation.
+// Do not use `OptionMap::new()` here: it can drop non-string values for
+// redacted keys before the flow option allowlist rejects them.
+fn flow_option_map(options: HashMap<String, OptionValue>) -> OptionMap {
+    let mut flow_options = OptionMap::default();
+    for (key, value) in options {
+        flow_options.insert_options(&key, value);
+    }
+    flow_options
+}
+
 /// Parses create [table] statement
 impl<'a> ParserContext<'a> {
     pub(crate) fn parse_create(&mut self) -> Result<Statement> {
@@ -339,6 +350,14 @@ impl<'a> ParserContext<'a> {
             None
         };
 
+        let flow_options = self
+            .parser
+            .parse_options(Keyword::WITH)
+            .context(SyntaxSnafu)?
+            .into_iter()
+            .map(parse_option_string)
+            .collect::<Result<HashMap<String, OptionValue>>>()?;
+
         self.parser
             .expect_keyword(Keyword::AS)
             .context(SyntaxSnafu)?;
@@ -353,6 +372,7 @@ impl<'a> ParserContext<'a> {
             expire_after,
             eval_interval,
             comment,
+            flow_options: flow_option_map(flow_options),
             query,
         }))
     }
@@ -1242,7 +1262,7 @@ fn ensure_partition_columns_defined<'a>(
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
     use std::collections::HashMap;
 
     use common_catalog::consts::FILE_ENGINE;
@@ -1256,6 +1276,20 @@ mod tests {
     use crate::dialect::GreptimeDbDialect;
     use crate::parser::ParseOptions;
 
+    fn string_option_map(
+        entries: impl IntoIterator<Item = (&'static str, &'static str)>,
+    ) -> OptionMap {
+        OptionMap::new(entries.into_iter().map(|(key, value)| {
+            (
+                key.to_string(),
+                OptionValue::try_new(Expr::Value(
+                    Value::SingleQuotedString(value.to_string()).into(),
+                ))
+                .unwrap(),
+            )
+        }))
+    }
+
     #[test]
     fn test_parse_create_table_like() {
         let sql = "CREATE TABLE t1 LIKE t2";
@@ -1498,6 +1532,8 @@ mod tests {
             pub expire_after: Option<i64>,
             /// Comment string
             pub comment: Option<String>,
+            /// Flow creation options
+            pub flow_options: OptionMap,
         }
         let testcases = vec![
             (
@@ -1518,6 +1554,7 @@ SELECT max(c1), min(c2) FROM schema_2.table_2;",
                     if_not_exists: true,
                     expire_after: Some(300),
                     comment: Some("test comment".to_string()),
+                    flow_options: OptionMap::default(),
                 },
             ),
             (
@@ -1538,6 +1575,7 @@ SELECT max(c1), min(c2) FROM schema_2.table_2;",
                     if_not_exists: true,
                     expire_after: Some(300),
                     comment: Some("test comment".to_string()),
+                    flow_options: OptionMap::default(),
                 },
             ),
             (
@@ -1558,6 +1596,7 @@ SELECT max(c1), min(c2) FROM schema_2.table_2;",
                     if_not_exists: true,
                     expire_after: Some(300),
                     comment: Some("test comment".to_string()),
+                    flow_options: OptionMap::default(),
                 },
             ),
             (
@@ -1578,6 +1617,7 @@ SELECT max(c1), min(c2) FROM schema_2.table_2;",
                     if_not_exists: true,
                     expire_after: Some(300),
                     comment: Some("test comment".to_string()),
+                    flow_options: OptionMap::default(),
                 },
             ),
             (
@@ -1597,6 +1637,7 @@ SELECT max(c1), min(c2) FROM schema_2.table_2;",
                     if_not_exists: false,
                     expire_after: Some(2 * 86400 + 3600 + 2 * 60),
                     comment: None,
+                    flow_options: OptionMap::default(),
                 },
             ),
             (
@@ -1616,6 +1657,7 @@ select max(c1), min(c2) from schema_2.table_2;",
                     if_not_exists: false,
                     expire_after: Some(600), // 10 minutes in seconds
                     comment: None,
+                    flow_options: OptionMap::default(),
                 },
             ),
             (
@@ -1636,6 +1678,27 @@ select max(c1), min(c2) from schema_2.table_2;",
                     if_not_exists: true,
                     expire_after: Some(7200), // 2 hours in seconds
                     comment: Some("lowercase test".to_string()),
+                    flow_options: OptionMap::default(),
+                },
+            ),
+            (
+                r"
+CREATE FLOW task_5
+SINK TO schema_1.table_1
+WITH (defer_on_missing_source = 'true')
+AS
+SELECT max(c1), min(c2) FROM schema_2.table_2;",
+                CreateFlowWoutQuery {
+                    flow_name: ObjectName::from(vec![Ident::new("task_5")]),
+                    sink_table_name: ObjectName::from(vec![
+                        Ident::new("schema_1"),
+                        Ident::new("table_1"),
+                    ]),
+                    or_replace: false,
+                    if_not_exists: false,
+                    expire_after: None,
+                    comment: None,
+                    flow_options: string_option_map([("defer_on_missing_source", "true")]),
                 },
             ),
         ];
@@ -1651,6 +1714,7 @@ select max(c1), min(c2) from schema_2.table_2;",
                 expire_after: expected.expire_after,
                 eval_interval: None,
                 comment: expected.comment,
+                flow_options: expected.flow_options,
                 // ignore query parse result
                 query: create_task.query.clone(),
             };
@@ -1696,6 +1760,8 @@ select max(c1), min(c2) from schema_2.table_2;",
             pub eval_interval: Option<i64>,
             /// Comment string
             pub comment: Option<String>,
+            /// Flow creation options
+            pub flow_options: OptionMap,
         }
 
         // create flow without `OR REPLACE`, `IF NOT EXISTS`, `EXPIRE AFTER` and `COMMENT`
@@ -1719,6 +1785,7 @@ SELECT max(c1), min(c2) FROM schema_2.table_2;",
                     expire_after: Some(300),
                     eval_interval: None,
                     comment: Some("test comment".to_string()),
+                    flow_options: OptionMap::default(),
                 },
             ),
             (
@@ -1740,6 +1807,7 @@ SELECT max(c1), min(c2) FROM schema_2.table_2;",
                     expire_after: Some(300),
                     eval_interval: None,
                     comment: Some("test comment".to_string()),
+                    flow_options: OptionMap::default(),
                 },
             ),
             (
@@ -1762,6 +1830,7 @@ SELECT max(c1), min(c2) FROM schema_2.table_2;",
                     expire_after: Some(300),
                     eval_interval: Some(10),
                     comment: Some("test comment".to_string()),
+                    flow_options: OptionMap::default(),
                 },
             ),
             (
@@ -1784,6 +1853,7 @@ SELECT max(c1), min(c2) FROM schema_2.table_2;",
                     expire_after: Some(300),
                     eval_interval: Some(10),
                     comment: Some("test comment".to_string()),
+                    flow_options: OptionMap::default(),
                 },
             ),
             (
@@ -1806,6 +1876,32 @@ SELECT max(c1), min(c2) FROM schema_2.table_2;",
                     expire_after: Some(2 * 86400 + 3600 + 2 * 60),
                     eval_interval: None,
                     comment: None,
+                    flow_options: OptionMap::default(),
+                },
+            ),
+            (
+                r"
+CREATE FLOW task_3
+SINK TO schema_1.table_1
+EVAL INTERVAL '10 seconds'
+WITH (defer_on_missing_source = 'true', foo = 'bar')
+AS
+SELECT max(c1), min(c2) FROM schema_2.table_2;",
+                CreateFlowWoutQuery {
+                    flow_name: ObjectName(vec![ObjectNamePart::Identifier(Ident::new("task_3"))]),
+                    sink_table_name: ObjectName(vec![
+                        ObjectNamePart::Identifier(Ident::new("schema_1")),
+                        ObjectNamePart::Identifier(Ident::new("table_1")),
+                    ]),
+                    or_replace: false,
+                    if_not_exists: false,
+                    expire_after: None,
+                    eval_interval: Some(10),
+                    comment: None,
+                    flow_options: string_option_map([
+                        ("defer_on_missing_source", "true"),
+                        ("foo", "bar"),
+                    ]),
                 },
             ),
         ];
@@ -1821,6 +1917,7 @@ SELECT max(c1), min(c2) FROM schema_2.table_2;",
                 expire_after: expected.expire_after,
                 eval_interval: expected.eval_interval,
                 comment: expected.comment,
+                flow_options: expected.flow_options,
                 // ignore query parse result
                 query: create_task.query.clone(),
             };
diff --git a/src/sql/src/parsers/delete_parser.rs b/src/sql/src/parsers/delete_parser.rs
index 3f13d18f97..be1a93b6c6 100644
--- a/src/sql/src/parsers/delete_parser.rs
+++ b/src/sql/src/parsers/delete_parser.rs
@@ -43,7 +43,7 @@ impl ParserContext<'_> {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use super::*;
     use crate::dialect::GreptimeDbDialect;
diff --git a/src/sql/src/parsers/insert_parser.rs b/src/sql/src/parsers/insert_parser.rs
index a7d5f311f6..d121181774 100644
--- a/src/sql/src/parsers/insert_parser.rs
+++ b/src/sql/src/parsers/insert_parser.rs
@@ -64,7 +64,7 @@ impl ParserContext<'_> {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use super::*;
     use crate::dialect::GreptimeDbDialect;
diff --git a/src/sql/src/parsers/show_parser.rs b/src/sql/src/parsers/show_parser.rs
index e2e5fc50ac..d6fc35c675 100644
--- a/src/sql/src/parsers/show_parser.rs
+++ b/src/sql/src/parsers/show_parser.rs
@@ -599,7 +599,7 @@ impl ParserContext<'_> {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use sqlparser::ast::{Ident, ObjectName};
 
diff --git a/src/sql/src/partition.rs b/src/sql/src/partition.rs
index 0f1a7526dc..f562fabee2 100644
--- a/src/sql/src/partition.rs
+++ b/src/sql/src/partition.rs
@@ -46,10 +46,13 @@ macro_rules! between_string {
     };
 }
 
-pub fn partition_rule_for_hexstring(ident: &str) -> Result<Partitions> {
+pub fn partition_rule_for_hexstring(ident: &str, partition_num: Option<u32>) -> Result<Partitions> {
     Ok(Partitions {
         column_list: vec![Ident::new(ident)],
-        exprs: partition_rules_for_uuid(DEFAULT_PARTITION_NUM_FOR_TRACES, ident)?,
+        exprs: partition_rules_for_uuid(
+            partition_num.unwrap_or(DEFAULT_PARTITION_NUM_FOR_TRACES),
+            ident,
+        )?,
     })
 }
 
@@ -173,7 +176,30 @@ mod tests {
 
         assert_eq!(
             results,
-            partition_rule_for_hexstring("trace_id").unwrap().exprs
+            partition_rule_for_hexstring("trace_id", None)
+                .unwrap()
+                .exprs
+        );
+
+        // custom partition number
+        let expr = vec![
+            "trace_id < '4'",
+            "trace_id >= '4' AND trace_id < '8'",
+            "trace_id >= '8' AND trace_id < 'c'",
+            "trace_id >= 'c'",
+        ];
+        let results = expr
+            .into_iter()
+            .map(|s| {
+                let mut parser = Parser::new(&dialect).try_with_sql(s).unwrap();
+                parser.parse_expr().unwrap()
+            })
+            .collect::<Vec<Expr>>();
+        assert_eq!(
+            results,
+            partition_rule_for_hexstring("trace_id", Some(4))
+                .unwrap()
+                .exprs
         );
     }
 
diff --git a/src/sql/src/statements.rs b/src/sql/src/statements.rs
index 211fc5598e..f51735b769 100644
--- a/src/sql/src/statements.rs
+++ b/src/sql/src/statements.rs
@@ -63,6 +63,7 @@ pub use crate::statements::option_map::OptionMap;
 pub(crate) use crate::statements::transform::transform_statements;
 
 const VECTOR_TYPE_NAME: &str = "VECTOR";
+const JSON2_TYPE_NAME: &str = "JSON2";
 
 pub fn value_to_sql_value(val: &Value) -> Result<SqlValue> {
     Ok(match val {
@@ -153,7 +154,16 @@ pub fn column_to_schema(
 
     column_schema.set_inverted_index(column.extensions.inverted_index_options.is_some());
 
-    if matches!(column.data_type(), SqlDataType::JSON) {
+    let is_json2_column = if let SqlDataType::Custom(object_name, _) = column.data_type() {
+        object_name
+            .0
+            .first()
+            .map(|x| x.to_string_unquoted().eq_ignore_ascii_case(JSON2_TYPE_NAME))
+            .unwrap_or_default()
+    } else {
+        false
+    };
+    if is_json2_column {
         let settings = column
             .extensions
             .build_json_structure_settings()?
@@ -273,39 +283,42 @@ pub fn sql_data_type_to_concrete_data_type(
                 Ok(ConcreteDataType::decimal128_datatype(*p as u8, *s as i8))
             }
         },
-        SqlDataType::JSON => {
-            let format = if let Some(x) = column_extensions.build_json_structure_settings()? {
-                if let Some(fields) = match x {
-                    JsonStructureSettings::Structured(fields) => fields,
-                    JsonStructureSettings::UnstructuredRaw => None,
-                    JsonStructureSettings::PartialUnstructuredByKey { fields, .. } => fields,
-                } {
-                    let datatype = &ConcreteDataType::Struct(fields);
-                    JsonFormat::Native(Box::new(datatype.into()))
-                } else {
-                    JsonFormat::Native(Box::new(JsonNativeType::Null))
+        SqlDataType::JSON => Ok(ConcreteDataType::Json(JsonType::new(JsonFormat::Jsonb))),
+        // Vector type and JSON2 type
+        SqlDataType::Custom(name, args) if name.0.len() == 1 => {
+            let name = name.0[0].to_string_unquoted().to_ascii_uppercase();
+            match name.as_str() {
+                VECTOR_TYPE_NAME if args.len() == 1 => {
+                    let dim = &args[0];
+                    let dim = dim.parse().map_err(|e| {
+                        error::ParseSqlValueSnafu {
+                            msg: format!("Failed to parse vector dimension '{}': {}", dim, e),
+                        }
+                        .build()
+                    })?;
+                    Ok(ConcreteDataType::vector_datatype(dim))
                 }
-            } else {
-                JsonFormat::Jsonb
-            };
-            Ok(ConcreteDataType::Json(JsonType::new(format)))
-        }
-        // Vector type
-        SqlDataType::Custom(name, d)
-            if name.0.as_slice().len() == 1
-                && name.0.as_slice()[0]
-                    .to_string_unquoted()
-                    .to_ascii_uppercase()
-                    == VECTOR_TYPE_NAME
-                && d.len() == 1 =>
-        {
-            let dim = d[0].parse().map_err(|e| {
-                error::ParseSqlValueSnafu {
-                    msg: format!("Failed to parse vector dimension: {}", e),
+                JSON2_TYPE_NAME if args.is_empty() => {
+                    let native_type = column_extensions
+                        .build_json_structure_settings()?
+                        .and_then(|x| match x {
+                            JsonStructureSettings::Structured(Some(fields))
+                            | JsonStructureSettings::PartialUnstructuredByKey {
+                                fields: Some(fields),
+                                ..
+                            } => Some(JsonNativeType::from(&ConcreteDataType::Struct(fields))),
+                            JsonStructureSettings::UnstructuredRaw => Some(JsonNativeType::Variant),
+                            _ => None,
+                        })
+                        .unwrap_or(JsonNativeType::Object(Default::default()));
+                    let format = JsonFormat::Json2(Box::new(native_type));
+                    Ok(ConcreteDataType::Json(JsonType::new(format)))
                 }
-                .build()
-            })?;
-            Ok(ConcreteDataType::vector_datatype(dim))
+                _ => error::SqlTypeNotSupportedSnafu {
+                    t: data_type.clone(),
+                }
+                .fail(),
+            }
         }
         _ => error::SqlTypeNotSupportedSnafu {
             t: data_type.clone(),
diff --git a/src/sql/src/statements/alter.rs b/src/sql/src/statements/alter.rs
index 5b6a5ab5e6..ab35e5bd34 100644
--- a/src/sql/src/statements/alter.rs
+++ b/src/sql/src/statements/alter.rs
@@ -390,7 +390,7 @@ impl Display for AlterDatabaseOperation {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use crate::dialect::GreptimeDbDialect;
     use crate::parser::{ParseOptions, ParserContext};
diff --git a/src/sql/src/statements/copy.rs b/src/sql/src/statements/copy.rs
index 7aa099c53c..7a59175538 100644
--- a/src/sql/src/statements/copy.rs
+++ b/src/sql/src/statements/copy.rs
@@ -164,7 +164,7 @@ impl CopyTableArgument {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use crate::dialect::GreptimeDbDialect;
     use crate::parser::{ParseOptions, ParserContext};
diff --git a/src/sql/src/statements/create.rs b/src/sql/src/statements/create.rs
index 817b31518d..74ab8aee18 100644
--- a/src/sql/src/statements/create.rs
+++ b/src/sql/src/statements/create.rs
@@ -377,32 +377,35 @@ impl ColumnExtensions {
             None
         };
 
-        options
+        let format = options
             .get(JSON_OPT_FORMAT)
-            .map(|format| match format {
-                JSON_FORMAT_FULL_STRUCTURED => Ok(JsonStructureSettings::Structured(fields)),
-                JSON_FORMAT_PARTIAL => {
-                    let fields = fields.map(|fields| {
-                        let mut fields = Arc::unwrap_or_clone(fields.fields());
-                        fields.push(datatypes::types::StructField::new(
-                            JsonStructureSettings::RAW_FIELD.to_string(),
-                            ConcreteDataType::string_datatype(),
-                            true,
-                        ));
-                        StructType::new(Arc::new(fields))
-                    });
-                    Ok(JsonStructureSettings::PartialUnstructuredByKey {
-                        fields,
-                        unstructured_keys,
-                    })
+            .unwrap_or(JSON_FORMAT_FULL_STRUCTURED);
+        let settings = match format {
+            JSON_FORMAT_FULL_STRUCTURED => JsonStructureSettings::Structured(fields),
+            JSON_FORMAT_PARTIAL => {
+                let fields = fields.map(|fields| {
+                    let mut fields = Arc::unwrap_or_clone(fields.fields());
+                    fields.push(datatypes::types::StructField::new(
+                        JsonStructureSettings::RAW_FIELD.to_string(),
+                        ConcreteDataType::string_datatype(),
+                        true,
+                    ));
+                    StructType::new(Arc::new(fields))
+                });
+                JsonStructureSettings::PartialUnstructuredByKey {
+                    fields,
+                    unstructured_keys,
                 }
-                JSON_FORMAT_RAW => Ok(JsonStructureSettings::UnstructuredRaw),
-                _ => InvalidSqlSnafu {
+            }
+            JSON_FORMAT_RAW => JsonStructureSettings::UnstructuredRaw,
+            _ => {
+                return InvalidSqlSnafu {
                     msg: format!("unknown JSON datatype 'format': {format}"),
                 }
-                .fail(),
-            })
-            .transpose()
+                .fail();
+            }
+        };
+        Ok(Some(settings))
     }
 
     pub fn set_json_structure_settings(&mut self, settings: JsonStructureSettings) {
@@ -612,6 +615,8 @@ pub struct CreateFlow {
     pub eval_interval: Option<i64>,
     /// Comment string
     pub comment: Option<String>,
+    /// Flow creation options from `WITH (...)`
+    pub flow_options: OptionMap,
     /// SQL statement
     pub query: Box<SqlOrTql>,
 }
@@ -669,6 +674,10 @@ impl Display for CreateFlow {
         if let Some(comment) = &self.comment {
             writeln!(f, "COMMENT '{}'", comment)?;
         }
+        if !self.flow_options.is_empty() {
+            let options = self.flow_options.kv_pairs();
+            writeln!(f, "WITH ({})", format_list_comma!(options))?;
+        }
         write!(f, "AS {}", &self.query)
     }
 }
@@ -709,7 +718,7 @@ impl Display for CreateView {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use crate::dialect::GreptimeDbDialect;
     use crate::error::Error;
diff --git a/src/sql/src/statements/describe.rs b/src/sql/src/statements/describe.rs
index 2a87725dfc..ec48f4be81 100644
--- a/src/sql/src/statements/describe.rs
+++ b/src/sql/src/statements/describe.rs
@@ -44,7 +44,7 @@ impl Display for DescribeTable {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use crate::dialect::GreptimeDbDialect;
     use crate::parser::{ParseOptions, ParserContext};
diff --git a/src/sql/src/statements/drop.rs b/src/sql/src/statements/drop.rs
index 1e97bee25c..26190b48a0 100644
--- a/src/sql/src/statements/drop.rs
+++ b/src/sql/src/statements/drop.rs
@@ -167,7 +167,7 @@ impl Display for DropView {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use crate::dialect::GreptimeDbDialect;
     use crate::parser::{ParseOptions, ParserContext};
diff --git a/src/sql/src/statements/option_map.rs b/src/sql/src/statements/option_map.rs
index 61f8badd0a..d562a61ce0 100644
--- a/src/sql/src/statements/option_map.rs
+++ b/src/sql/src/statements/option_map.rs
@@ -49,6 +49,18 @@ impl OptionMap {
         }
     }
 
+    pub fn from_filtered_string_map(
+        options: &HashMap<String, String>,
+        hidden_keys: &[&str],
+    ) -> Self {
+        Self::from(
+            options
+                .iter()
+                .filter(|(key, _)| !hidden_keys.contains(&key.as_str()))
+                .map(|(key, value)| (key.clone(), value.clone())),
+        )
+    }
+
     pub fn insert(&mut self, k: String, v: String) {
         if REDACTED_OPTIONS.contains(&k.as_str()) {
             self.secrets.insert(k, SecretString::new(Box::new(v)));
@@ -221,6 +233,8 @@ impl VisitMut for OptionMap {
 
 #[cfg(test)]
 mod tests {
+    use std::collections::HashMap;
+
     use crate::statements::OptionMap;
 
     #[test]
@@ -237,4 +251,18 @@ mod tests {
         map.insert("a.b".to_string(), "中文comment\n".to_string());
         assert_eq!("'a.b' = '中文comment\\n'", map.kv_pairs()[0]);
     }
+
+    #[test]
+    fn test_from_filtered_string_map() {
+        let map = OptionMap::from_filtered_string_map(
+            &HashMap::from([
+                ("visible".to_string(), "1".to_string()),
+                ("hidden".to_string(), "2".to_string()),
+            ]),
+            &["hidden"],
+        );
+
+        assert_eq!(map.get("visible"), Some("1"));
+        assert_eq!(map.get("hidden"), None);
+    }
 }
diff --git a/src/sql/src/statements/set_variables.rs b/src/sql/src/statements/set_variables.rs
index 748d077d84..4c6eb16692 100644
--- a/src/sql/src/statements/set_variables.rs
+++ b/src/sql/src/statements/set_variables.rs
@@ -41,7 +41,7 @@ impl Display for SetVariables {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use crate::dialect::GreptimeDbDialect;
     use crate::parser::{ParseOptions, ParserContext};
diff --git a/src/sql/src/statements/show.rs b/src/sql/src/statements/show.rs
index 0dfdd1de7d..77880e4a50 100644
--- a/src/sql/src/statements/show.rs
+++ b/src/sql/src/statements/show.rs
@@ -340,7 +340,7 @@ impl Display for ShowProcessList {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use sqlparser::ast::UnaryOperator;
 
diff --git a/src/sql/src/statements/truncate.rs b/src/sql/src/statements/truncate.rs
index b9c299601c..307326d9c7 100644
--- a/src/sql/src/statements/truncate.rs
+++ b/src/sql/src/statements/truncate.rs
@@ -98,7 +98,7 @@ impl Display for TruncateTable {
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use crate::dialect::GreptimeDbDialect;
     use crate::parser::{ParseOptions, ParserContext};
diff --git a/src/standalone/src/information_extension.rs b/src/standalone/src/information_extension.rs
index fb3bc3e5fb..a7de19e48a 100644
--- a/src/standalone/src/information_extension.rs
+++ b/src/standalone/src/information_extension.rs
@@ -88,6 +88,7 @@ impl InformationExtension for StandaloneInformationExtension {
                 .unwrap_or_default()
                 .to_string_lossy()
                 .to_string(),
+            env_vars: Default::default(),
         };
         Ok(vec![node_info])
     }
diff --git a/src/standalone/src/options.rs b/src/standalone/src/options.rs
index 35915742a7..dece6389f0 100644
--- a/src/standalone/src/options.rs
+++ b/src/standalone/src/options.rs
@@ -67,6 +67,8 @@ pub struct StandaloneOptions {
     pub slow_query: SlowQueryOptions,
     pub query: QueryOptions,
     pub memory: MemoryOptions,
+    /// Environment variable keys to read and report in heartbeat messages.
+    pub heartbeat_env_vars: Vec<String>,
 }
 
 impl Default for StandaloneOptions {
@@ -102,13 +104,14 @@ impl Default for StandaloneOptions {
             slow_query: SlowQueryOptions::default(),
             query: QueryOptions::default(),
             memory: MemoryOptions::default(),
+            heartbeat_env_vars: vec![],
         }
     }
 }
 
 impl Configurable for StandaloneOptions {
     fn env_list_keys() -> Option<&'static [&'static str]> {
-        Some(&["wal.broker_endpoints"])
+        Some(&["heartbeat_env_vars", "wal.broker_endpoints"])
     }
 }
 
@@ -141,6 +144,7 @@ impl StandaloneOptions {
             logging: cloned_opts.logging,
             user_provider: cloned_opts.user_provider,
             slow_query: cloned_opts.slow_query,
+            heartbeat_env_vars: cloned_opts.heartbeat_env_vars.clone(),
             ..Default::default()
         }
     }
@@ -157,6 +161,7 @@ impl StandaloneOptions {
             init_regions_in_background: cloned_opts.init_regions_in_background,
             init_regions_parallelism: cloned_opts.init_regions_parallelism,
             query: cloned_opts.query,
+            heartbeat_env_vars: cloned_opts.heartbeat_env_vars,
             ..Default::default()
         }
     }
diff --git a/src/store-api/src/lib.rs b/src/store-api/src/lib.rs
index 4df594fc67..cb39875d74 100644
--- a/src/store-api/src/lib.rs
+++ b/src/store-api/src/lib.rs
@@ -14,8 +14,6 @@
 
 //! Storage related APIs
 
-#![feature(iterator_try_collect)]
-
 pub mod codec;
 pub mod data_source;
 pub mod logstore;
diff --git a/src/store-api/src/metadata.rs b/src/store-api/src/metadata.rs
index d571a5392f..0c663bccc0 100644
--- a/src/store-api/src/metadata.rs
+++ b/src/store-api/src/metadata.rs
@@ -18,8 +18,8 @@
 
 use std::any::Any;
 use std::collections::{HashMap, HashSet};
-use std::fmt;
 use std::sync::Arc;
+use std::{fmt, mem};
 
 use api::v1::SemanticType;
 use api::v1::column_def::try_as_column_schema;
@@ -99,6 +99,12 @@ impl ColumnMetadata {
     pub fn is_same_datatype(&self, other: &Self) -> bool {
         self.column_schema.data_type == other.column_schema.data_type
     }
+
+    /// Returns the estimated memory footprint of this metadata.
+    pub fn estimated_size(&self) -> usize {
+        mem::size_of_val(self) - mem::size_of_val(&self.column_schema)
+            + self.column_schema.estimated_size()
+    }
 }
 
 #[cfg_attr(doc, aquamarine::aquamarine)]
@@ -226,6 +232,25 @@ impl RegionMetadata {
         serde_json::from_str(s).context(SerdeJsonSnafu)
     }
 
+    /// Returns the estimated memory footprint of this metadata.
+    pub fn estimated_size(&self) -> usize {
+        mem::size_of_val(self)
+            + mem::size_of::<ColumnMetadata>() * self.column_metadatas.capacity()
+            + self
+                .column_metadatas
+                .iter()
+                .map(|column| column.estimated_size() - mem::size_of::<ColumnMetadata>())
+                .sum::<usize>()
+            + mem::size_of::<ColumnId>() * self.primary_key.capacity()
+            + mem::size_of::<(ColumnId, usize)>() * self.id_to_index.capacity()
+            + self.schema.estimated_size()
+            + self
+                .partition_expr
+                .as_ref()
+                .map(|expr| expr.capacity())
+                .unwrap_or_default()
+    }
+
     /// Encode the metadata to a JSON string.
     pub fn to_json(&self) -> Result<String> {
         serde_json::to_string(&self).context(SerdeJsonSnafu)
diff --git a/src/store-api/src/region_engine.rs b/src/store-api/src/region_engine.rs
index b3f460d01d..b235fcffc7 100644
--- a/src/store-api/src/region_engine.rs
+++ b/src/store-api/src/region_engine.rs
@@ -23,7 +23,7 @@ use api::greptime_proto::v1::meta::{GrantedRegion as PbGrantedRegion, RegionRole
 use api::region::RegionResponse;
 use async_trait::async_trait;
 use common_error::ext::BoxedError;
-use common_recordbatch::{EmptyRecordBatchStream, MemoryPermit, SendableRecordBatchStream};
+use common_recordbatch::{EmptyRecordBatchStream, QueryMemoryTracker, SendableRecordBatchStream};
 use common_time::Timestamp;
 use datafusion_physical_plan::metrics::ExecutionPlanMetricsSet;
 use datafusion_physical_plan::{DisplayAs, DisplayFormatType, PhysicalExpr};
@@ -67,7 +67,7 @@ impl From<SettableRegionRoleState> for RegionRole {
             SettableRegionRoleState::Follower => RegionRole::Follower,
             SettableRegionRoleState::DowngradingLeader => RegionRole::DowngradingLeader,
             SettableRegionRoleState::Leader => RegionRole::Leader,
-            SettableRegionRoleState::StagingLeader => RegionRole::Leader, // Still a leader role
+            SettableRegionRoleState::StagingLeader => RegionRole::StagingLeader,
         }
     }
 }
@@ -210,6 +210,11 @@ pub enum RegionRole {
     Follower,
     // Writable region(mito2), Readonly region(file).
     Leader,
+    // Leader is in staging mode.
+    //
+    // This is leader-like and writable, but it follows the staging workflow
+    // semantics instead of a normal leader's steady state.
+    StagingLeader,
     // Leader is downgrading to follower.
     //
     // This state is used to prevent new write requests.
@@ -221,6 +226,7 @@ impl Display for RegionRole {
         match self {
             RegionRole::Follower => write!(f, "Follower"),
             RegionRole::Leader => write!(f, "Leader"),
+            RegionRole::StagingLeader => write!(f, "Leader(Staging)"),
             RegionRole::DowngradingLeader => write!(f, "Leader(Downgrading)"),
         }
     }
@@ -228,7 +234,7 @@ impl Display for RegionRole {
 
 impl RegionRole {
     pub fn writable(&self) -> bool {
-        matches!(self, RegionRole::Leader)
+        matches!(self, RegionRole::Leader | RegionRole::StagingLeader)
     }
 }
 
@@ -237,6 +243,7 @@ impl From<RegionRole> for PbRegionRole {
         match value {
             RegionRole::Follower => PbRegionRole::Follower,
             RegionRole::Leader => PbRegionRole::Leader,
+            RegionRole::StagingLeader => PbRegionRole::StagingLeader,
             RegionRole::DowngradingLeader => PbRegionRole::DowngradingLeader,
         }
     }
@@ -246,6 +253,7 @@ impl From<PbRegionRole> for RegionRole {
     fn from(value: PbRegionRole) -> Self {
         match value {
             PbRegionRole::Leader => RegionRole::Leader,
+            PbRegionRole::StagingLeader => RegionRole::StagingLeader,
             PbRegionRole::Follower => RegionRole::Follower,
             PbRegionRole::DowngradingLeader => RegionRole::DowngradingLeader,
         }
@@ -462,6 +470,10 @@ pub trait RegionScanner: Debug + DisplayAs + Send {
 
     /// Sets whether the scanner is reading a logical region.
     fn set_logical_region(&mut self, logical_region: bool);
+
+    fn snapshot_sequence(&self) -> Option<SequenceNumber> {
+        None
+    }
 }
 
 pub type RegionScannerRef = Box<dyn RegionScanner>;
@@ -886,8 +898,8 @@ pub trait RegionEngine: Send + Sync {
         request: ScanRequest,
     ) -> Result<RegionScannerRef, BoxedError>;
 
-    /// Registers and returns a query memory permit.
-    fn register_query_memory_permit(&self) -> Option<Arc<MemoryPermit>> {
+    /// Returns the query memory tracker for scan execution.
+    fn query_memory_tracker(&self) -> Option<QueryMemoryTracker> {
         None
     }
 
@@ -945,6 +957,7 @@ pub struct SinglePartitionScanner {
     schema: SchemaRef,
     properties: ScannerProperties,
     metadata: RegionMetadataRef,
+    snapshot_sequence: Option<SequenceNumber>,
 }
 
 impl SinglePartitionScanner {
@@ -953,6 +966,7 @@ impl SinglePartitionScanner {
         stream: SendableRecordBatchStream,
         append_mode: bool,
         metadata: RegionMetadataRef,
+        snapshot_sequence: Option<SequenceNumber>,
     ) -> Self {
         let schema = stream.schema();
         Self {
@@ -960,6 +974,7 @@ impl SinglePartitionScanner {
             schema,
             properties: ScannerProperties::default().with_append_mode(append_mode),
             metadata,
+            snapshot_sequence,
         }
     }
 }
@@ -1019,6 +1034,10 @@ impl RegionScanner for SinglePartitionScanner {
     fn set_logical_region(&mut self, logical_region: bool) {
         self.properties.set_logical_region(logical_region);
     }
+
+    fn snapshot_sequence(&self) -> Option<SequenceNumber> {
+        self.snapshot_sequence
+    }
 }
 
 impl DisplayAs for SinglePartitionScanner {
diff --git a/src/store-api/src/region_request.rs b/src/store-api/src/region_request.rs
index 99d3a87dd3..0759fccf67 100644
--- a/src/store-api/src/region_request.rs
+++ b/src/store-api/src/region_request.rs
@@ -346,9 +346,7 @@ fn make_region_flush(flush: FlushRequest) -> Result<Vec<(RegionId, RegionRequest
     let region_id = flush.region_id.into();
     Ok(vec![(
         region_id,
-        RegionRequest::Flush(RegionFlushRequest {
-            row_group_size: None,
-        }),
+        RegionRequest::Flush(RegionFlushRequest::default()),
     )])
 }
 
@@ -1404,9 +1402,20 @@ impl Display for UnsetRegionOption {
     }
 }
 
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+pub enum RegionFlushReason {
+    /// Flush triggered before region migration.
+    RegionMigration,
+    /// Flush triggered by repartition procedure.
+    Repartition,
+    /// Flush triggered by remote WAL pruning.
+    RemoteWalPrune,
+}
+
 #[derive(Debug, Clone, Default)]
 pub struct RegionFlushRequest {
     pub row_group_size: Option<usize>,
+    pub reason: Option<RegionFlushReason>,
 }
 
 #[derive(Debug)]
diff --git a/src/store-api/src/sst_entry.rs b/src/store-api/src/sst_entry.rs
index 832bfc1155..840fef2268 100644
--- a/src/store-api/src/sst_entry.rs
+++ b/src/store-api/src/sst_entry.rs
@@ -384,11 +384,8 @@ fn build_plan_helper(
 ) -> Result<LogicalPlan, DataFusionError> {
     let table_source = LogicalTableSource::new(schema.arrow_schema().clone());
 
-    let mut builder = LogicalPlanBuilder::scan(
-        table_name,
-        Arc::new(table_source),
-        scan_request.projection.clone(),
-    )?;
+    let projection = scan_request.projection_input.map(|input| input.projection);
+    let mut builder = LogicalPlanBuilder::scan(table_name, Arc::new(table_source), projection)?;
 
     for filter in scan_request.filters {
         builder = builder.filter(filter)?;
@@ -910,8 +907,9 @@ mod tests {
     #[test]
     fn test_manifest_build_plan() {
         // Note: filter must reference a column in the projected schema
+        let projection_input = Some(vec![0, 1, 2].into());
         let request = ScanRequest {
-            projection: Some(vec![0, 1, 2]),
+            projection_input,
             filters: vec![binary_expr(col("table_id"), Operator::Gt, lit(0))],
             limit: Some(5),
             ..Default::default()
@@ -941,8 +939,9 @@ mod tests {
 
     #[test]
     fn test_storage_build_plan() {
+        let projection_input = Some(vec![0, 2].into());
         let request = ScanRequest {
-            projection: Some(vec![0, 2]),
+            projection_input,
             filters: vec![binary_expr(col("file_path"), Operator::Eq, lit("/a"))],
             limit: Some(1),
             ..Default::default()
diff --git a/src/store-api/src/storage.rs b/src/store-api/src/storage.rs
index b97fe0b3ad..33aa3a7b41 100644
--- a/src/store-api/src/storage.rs
+++ b/src/store-api/src/storage.rs
@@ -17,6 +17,7 @@
 pub mod consts;
 mod descriptors;
 mod file;
+mod projection;
 mod requests;
 mod types;
 
@@ -27,6 +28,7 @@ pub use datatypes::schema::{
 
 pub use self::descriptors::*;
 pub use self::file::{FileId, FileRef, FileRefsManifest, GcReport, IndexVersion, ParseIdError};
+pub use self::projection::{NestedPath, ProjectionInput};
 pub use self::requests::{
     ScanRequest, TimeSeriesDistribution, TimeSeriesRowSelector, VectorDistanceMetric,
     VectorIndexEngine, VectorIndexEngineType, VectorSearchMatches, VectorSearchRequest,
diff --git a/src/store-api/src/storage/projection.rs b/src/store-api/src/storage/projection.rs
new file mode 100644
index 0000000000..f8f27fe5a4
--- /dev/null
+++ b/src/store-api/src/storage/projection.rs
@@ -0,0 +1,78 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::fmt::{Display, Formatter};
+
+/// A nested field access path.
+///
+/// Each path represents a field access on a nested column.
+///
+/// Example:
+/// - `j.a.b` -> `["j", "a", "b"]`
+pub type NestedPath = Vec<String>;
+
+/// Projection information for a table scan.
+#[derive(Default, Debug, Clone, PartialEq)]
+pub struct ProjectionInput {
+    /// Top-level column projection.
+    ///
+    /// The indices are based on the schema exposed by the table scan input,
+    /// such as the schema passed to `TableProvider::scan`.
+    ///
+    /// Only the root columns with the specified schema indices are needed.
+    pub projection: Vec<usize>,
+    /// Nested field access paths used for sub-field projection.
+    ///
+    /// It extends and refines the top-level projection by specifying nested
+    /// field accesses inside complex columns such as JSON or struct columns.
+    ///
+    /// In other words:
+    /// - `projection` determines **which root columns are needed**
+    /// - `nested_paths` further determines **which sub-fields inside those
+    ///   columns are required**
+    ///
+    /// Each path starts with the root column name and continues with
+    /// nested field names.
+    pub nested_paths: Vec<NestedPath>,
+}
+
+impl ProjectionInput {
+    pub fn new(projection: Vec<usize>) -> Self {
+        Self {
+            projection,
+            nested_paths: Vec::new(),
+        }
+    }
+
+    pub fn with_nested_paths(mut self, nested_paths: Vec<NestedPath>) -> Self {
+        self.nested_paths = nested_paths;
+        self
+    }
+}
+
+impl From<Vec<usize>> for ProjectionInput {
+    fn from(projection: Vec<usize>) -> Self {
+        Self::new(projection)
+    }
+}
+
+impl Display for ProjectionInput {
+    fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
+        write!(
+            f,
+            "ProjectionInput {{ projection: {:?}, nested_paths: {:?} }}",
+            self.projection, self.nested_paths
+        )
+    }
+}
diff --git a/src/store-api/src/storage/requests.rs b/src/store-api/src/storage/requests.rs
index 6725de92e3..57b2ca8e88 100644
--- a/src/store-api/src/storage/requests.rs
+++ b/src/store-api/src/storage/requests.rs
@@ -12,6 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+use std::collections::HashMap;
 use std::fmt::{Display, Formatter};
 
 use common_error::ext::BoxedError;
@@ -19,9 +20,11 @@ use common_recordbatch::OrderOption;
 use datafusion_expr::expr::Expr;
 // Re-export vector types from datatypes to avoid duplication
 pub use datatypes::schema::{VectorDistanceMetric, VectorIndexEngineType};
+use datatypes::types::json_type::JsonNativeType;
+use itertools::Itertools;
 use strum::Display;
 
-use crate::storage::{ColumnId, SequenceNumber};
+use crate::storage::{ColumnId, ProjectionInput, SequenceNumber};
 
 /// A hint for KNN vector search.
 #[derive(Debug, Clone, PartialEq)]
@@ -95,9 +98,9 @@ pub enum TimeSeriesDistribution {
 
 #[derive(Default, Clone, Debug, PartialEq)]
 pub struct ScanRequest {
-    /// Indices of columns to read, `None` to read all columns. This indices is
-    /// based on table schema.
-    pub projection: Option<Vec<usize>>,
+    /// Optional projection information for the scan. `None` reads all root
+    /// columns.
+    pub projection_input: Option<ProjectionInput>,
     /// Filters pushed down
     pub filters: Vec<Expr>,
     /// Expected output ordering. This is only a hint and isn't guaranteed.
@@ -112,6 +115,8 @@ pub struct ScanRequest {
     /// Optional constraint on the sequence number of the rows to read.
     /// If set, only rows with a sequence number **lesser or equal** to this value
     /// will be returned.
+    /// This is the effective memtable upper bound used by the scan, whether provided
+    /// explicitly or bound on scan open.
     pub memtable_max_sequence: Option<SequenceNumber>,
     /// Optional constraint on the minimal sequence number in the memtable.
     /// If set, only the memtables that contain sequences **greater than** this value will be scanned
@@ -119,13 +124,24 @@ pub struct ScanRequest {
     /// Optional constraint on the minimal sequence number in the SST files.
     /// If set, only the SST files that contain sequences greater than this value will be scanned.
     pub sst_min_sequence: Option<SequenceNumber>,
+    /// Whether to bind the effective snapshot upper bound when opening the scan.
+    pub snapshot_on_scan: bool,
     /// Optional hint for the distribution of time-series data.
     pub distribution: Option<TimeSeriesDistribution>,
     /// Optional hint for KNN vector search. When set, the scan should use
     /// vector index to find the k nearest neighbors.
     pub vector_search: Option<VectorSearchRequest>,
-    /// Whether to force reading region data in flat format.
-    pub force_flat_format: bool,
+    /// Optional hint from query-driven JSON type concretization.
+    pub json_type_hint: HashMap<String, JsonNativeType>,
+}
+
+impl ScanRequest {
+    /// Returns the top-level projected column indices.
+    pub fn projection_indices(&self) -> Option<&[usize]> {
+        self.projection_input
+            .as_ref()
+            .map(|projection_input| projection_input.projection.as_slice())
+    }
 }
 
 impl Display for ScanRequest {
@@ -150,7 +166,7 @@ impl Display for ScanRequest {
         let mut delimiter = Delimiter::None;
 
         write!(f, "ScanRequest {{ ")?;
-        if let Some(projection) = &self.projection {
+        if let Some(projection) = &self.projection_input {
             write!(f, "{}projection: {:?}", delimiter.as_str(), projection)?;
         }
         if !self.filters.is_empty() {
@@ -195,6 +211,14 @@ impl Display for ScanRequest {
                 sst_min_sequence
             )?;
         }
+        if self.snapshot_on_scan {
+            write!(
+                f,
+                "{}snapshot_on_scan: {}",
+                delimiter.as_str(),
+                self.snapshot_on_scan
+            )?;
+        }
         if let Some(distribution) = &self.distribution {
             write!(f, "{}distribution: {}", delimiter.as_str(), distribution)?;
         }
@@ -208,12 +232,15 @@ impl Display for ScanRequest {
                 vector_search.metric
             )?;
         }
-        if self.force_flat_format {
+        if !self.json_type_hint.is_empty() {
             write!(
                 f,
-                "{}force_flat_format: {}",
+                "{}json_type_hint: {}",
                 delimiter.as_str(),
-                self.force_flat_format
+                self.json_type_hint
+                    .iter()
+                    .map(|(column, json_type)| format!("({column}: {json_type})"))
+                    .join(", ")
             )?;
         }
         write!(f, " }}")
@@ -233,8 +260,9 @@ mod tests {
         };
         assert_eq!(request.to_string(), "ScanRequest {  }");
 
+        let projection_input = Some(vec![1, 2].into());
         let request = ScanRequest {
-            projection: Some(vec![1, 2]),
+            projection_input,
             filters: vec![
                 binary_expr(col("i"), Operator::Gt, lit(1)),
                 binary_expr(col("s"), Operator::Eq, lit("x")),
@@ -244,7 +272,7 @@ mod tests {
         };
         assert_eq!(
             request.to_string(),
-            r#"ScanRequest { projection: [1, 2], filters: [i > Int32(1), s = Utf8("x")], limit: 10 }"#
+            r#"ScanRequest { projection: ProjectionInput { projection: [1, 2], nested_paths: [] }, filters: [i > Int32(1), s = Utf8("x")], limit: 10 }"#
         );
 
         let request = ScanRequest {
@@ -260,23 +288,38 @@ mod tests {
             r#"ScanRequest { filters: [i > Int32(1), s = Utf8("x")], limit: 10 }"#
         );
 
+        let projection_input = Some(vec![1, 2].into());
         let request = ScanRequest {
-            projection: Some(vec![1, 2]),
+            projection_input,
             limit: Some(10),
             ..Default::default()
         };
         assert_eq!(
             request.to_string(),
-            "ScanRequest { projection: [1, 2], limit: 10 }"
+            "ScanRequest { projection: ProjectionInput { projection: [1, 2], nested_paths: [] }, limit: 10 }"
         );
 
+        let projection_input = Some(ProjectionInput::new(vec![1, 2]).with_nested_paths(vec![
+            vec!["j".to_string(), "a".to_string(), "b".to_string()],
+            vec!["s".to_string(), "x".to_string()],
+        ]));
         let request = ScanRequest {
-            force_flat_format: true,
+            projection_input,
+            limit: Some(10),
             ..Default::default()
         };
         assert_eq!(
             request.to_string(),
-            "ScanRequest { force_flat_format: true }"
+            r#"ScanRequest { projection: ProjectionInput { projection: [1, 2], nested_paths: [["j", "a", "b"], ["s", "x"]] }, limit: 10 }"#
+        );
+
+        let request = ScanRequest {
+            snapshot_on_scan: true,
+            ..Default::default()
+        };
+        assert_eq!(
+            request.to_string(),
+            "ScanRequest { snapshot_on_scan: true }"
         );
     }
 }
diff --git a/src/table/src/lib.rs b/src/table/src/lib.rs
index 64e72029b8..abc7dc8300 100644
--- a/src/table/src/lib.rs
+++ b/src/table/src/lib.rs
@@ -12,7 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
 #![feature(try_blocks)]
 
 pub mod dist_table;
diff --git a/src/table/src/metadata.rs b/src/table/src/metadata.rs
index ca8bc30aa1..42f66044d9 100644
--- a/src/table/src/metadata.rs
+++ b/src/table/src/metadata.rs
@@ -1388,7 +1388,7 @@ fn unset_column_skipping_index_options(
 
 #[cfg(test)]
 mod tests {
-    use std::assert_matches::assert_matches;
+    use std::assert_matches;
 
     use common_error::ext::ErrorExt;
     use common_error::status_code::StatusCode;
diff --git a/src/table/src/predicate.rs b/src/table/src/predicate.rs
index f9be7be16e..daf404d5d6 100644
--- a/src/table/src/predicate.rs
+++ b/src/table/src/predicate.rs
@@ -120,11 +120,11 @@ impl Predicate {
             .context(error::DatafusionSnafu)
     }
 
-    /// Builds physical exprs according to provided schema.
-    pub fn to_physical_exprs(
-        &self,
+    /// Builds a single physical expr according to provided schema.
+    pub fn to_physical_expr(
+        expr: &Expr,
         schema: &arrow::datatypes::SchemaRef,
-    ) -> error::Result<Vec<Arc<dyn PhysicalExpr>>> {
+    ) -> error::Result<Arc<dyn PhysicalExpr>> {
         let df_schema = schema
             .clone()
             .to_dfschema_ref()
@@ -135,12 +135,21 @@ impl Predicate {
         // registering variables.
         let execution_props = &ExecutionProps::new();
 
+        create_physical_expr(expr, df_schema.as_ref(), execution_props)
+            .context(error::DatafusionSnafu)
+    }
+
+    /// Builds physical exprs according to provided schema.
+    pub fn to_physical_exprs(
+        &self,
+        schema: &arrow::datatypes::SchemaRef,
+    ) -> error::Result<Vec<Arc<dyn PhysicalExpr>>> {
         let dyn_filters = self.dyn_filter_phy_exprs()?;
 
         Ok(self
             .exprs
             .iter()
-            .filter_map(|expr| create_physical_expr(expr, df_schema.as_ref(), execution_props).ok())
+            .filter_map(|expr| Self::to_physical_expr(expr, schema).ok())
             .chain(dyn_filters)
             .collect::<Vec<_>>())
     }
@@ -203,7 +212,7 @@ pub fn build_time_range_predicate(
 
 /// Extract time range filter from `WHERE`/`IN (...)`/`BETWEEN` clauses.
 /// Return None if no time range can be found in expr.
-fn extract_time_range_from_expr(
+pub fn extract_time_range_from_expr(
     ts_col_name: &str,
     ts_col_unit: TimeUnit,
     expr: &Expr,
@@ -730,5 +739,8 @@ mod tests {
 
         let predicates = predicate.to_physical_exprs(&schema).unwrap();
         assert!(!predicates.is_empty());
+
+        let physical_expr = Predicate::to_physical_expr(&col("host").eq(lit("host_a")), &schema);
+        assert!(physical_expr.is_ok());
     }
 }
diff --git a/src/table/src/requests.rs b/src/table/src/requests.rs
index 43fc36644b..7e56026d93 100644
--- a/src/table/src/requests.rs
+++ b/src/table/src/requests.rs
@@ -36,8 +36,9 @@ use store_api::metric_engine_consts::{
     LOGICAL_TABLE_METADATA_KEY, PHYSICAL_TABLE_METADATA_KEY, is_metric_engine_option_key,
 };
 use store_api::mito_engine_options::{
-    APPEND_MODE_KEY, COMPACTION_TYPE, MEMTABLE_TYPE, MERGE_MODE_KEY, TWCS_FALLBACK_TO_LOCAL,
-    TWCS_MAX_OUTPUT_FILE_SIZE, TWCS_TIME_WINDOW, TWCS_TRIGGER_FILE_NUM, is_mito_engine_option_key,
+    APPEND_MODE_KEY, COMPACTION_TYPE, MEMTABLE_TYPE, MERGE_MODE_KEY, SST_FORMAT_KEY,
+    TWCS_FALLBACK_TO_LOCAL, TWCS_MAX_OUTPUT_FILE_SIZE, TWCS_TIME_WINDOW, TWCS_TRIGGER_FILE_NUM,
+    is_mito_engine_option_key,
 };
 use store_api::region_request::{SetRegionOption, UnsetRegionOption};
 
@@ -56,13 +57,14 @@ pub const TABLE_DATA_MODEL_TRACE_V1: &str = "greptime_trace_v1";
 pub const OTLP_METRIC_COMPAT_KEY: &str = "otlp_metric_compat";
 pub const OTLP_METRIC_COMPAT_PROM: &str = "prom";
 
-pub const VALID_TABLE_OPTION_KEYS: [&str; 12] = [
+pub const VALID_TABLE_OPTION_KEYS: [&str; 13] = [
     // common keys:
     WRITE_BUFFER_SIZE_KEY,
     TTL_KEY,
     STORAGE_KEY,
     COMMENT_KEY,
     SKIP_WAL_KEY,
+    SST_FORMAT_KEY,
     // file engine keys:
     FILE_TABLE_LOCATION_KEY,
     FILE_TABLE_FORMAT_KEY,
@@ -94,6 +96,7 @@ static VALID_DB_OPT_KEYS: Lazy<HashSet<&str>> = Lazy::new(|| {
     set.insert(TWCS_TIME_WINDOW);
     set.insert(TWCS_TRIGGER_FILE_NUM);
     set.insert(TWCS_MAX_OUTPUT_FILE_SIZE);
+    set.insert(SST_FORMAT_KEY);
     set
 });
 
@@ -142,6 +145,7 @@ pub const STORAGE_KEY: &str = "storage";
 pub const COMMENT_KEY: &str = "comment";
 pub const AUTO_CREATE_TABLE_KEY: &str = "auto_create_table";
 pub const SKIP_WAL_KEY: &str = store_api::mito_engine_options::SKIP_WAL_KEY;
+pub const TRACE_TABLE_PARTITIONS_HINT_KEY: &str = "trace_table_partitions";
 
 impl TableOptions {
     pub fn try_from_iter<T: ToString, U: IntoIterator<Item = (T, T)>>(
diff --git a/src/table/src/table/adapter.rs b/src/table/src/table/adapter.rs
index e33bf53c84..55cd5ba017 100644
--- a/src/table/src/table/adapter.rs
+++ b/src/table/src/table/adapter.rs
@@ -104,10 +104,11 @@ impl TableProvider for DfTableProviderAdapter {
         limit: Option<usize>,
     ) -> DfResult<Arc<dyn ExecutionPlan>> {
         let filters: Vec<Expr> = filters.iter().map(Clone::clone).collect();
+        let projection_input = projection.map(|p| p.clone().into());
         let request = {
             let mut request = self.scan_req.lock().unwrap();
             request.filters = filters;
-            request.projection = projection.cloned();
+            request.projection_input = projection_input;
             request.limit = limit;
             request.clone()
         };
diff --git a/src/table/src/table/numbers.rs b/src/table/src/table/numbers.rs
index 6dd40fe747..cb4b8ce317 100644
--- a/src/table/src/table/numbers.rs
+++ b/src/table/src/table/numbers.rs
@@ -110,7 +110,8 @@ impl NumbersDataSource {
 
 impl DataSource for NumbersDataSource {
     fn get_stream(&self, request: ScanRequest) -> Result<SendableRecordBatchStream, BoxedError> {
-        let projected_schema = match &request.projection {
+        let projection = request.projection_input.map(|input| input.projection);
+        let projected_schema = match &projection {
             Some(projection) => Arc::new(self.schema.try_project(projection).unwrap()),
             None => self.schema.clone(),
         };
@@ -118,7 +119,7 @@ impl DataSource for NumbersDataSource {
             limit: request.limit.unwrap_or(100) as u32,
             schema: self.schema.clone(),
             already_run: false,
-            projection: request.projection,
+            projection,
             projected_schema,
         }))
     }
diff --git a/src/table/src/table/scan.rs b/src/table/src/table/scan.rs
index e2d8f794da..02511456ae 100644
--- a/src/table/src/table/scan.rs
+++ b/src/table/src/table/scan.rs
@@ -20,7 +20,7 @@ use std::time::Instant;
 
 use common_error::ext::BoxedError;
 use common_recordbatch::{
-    DfRecordBatch, DfSendableRecordBatchStream, MemoryPermit, MemoryTrackedStream,
+    DfRecordBatch, DfSendableRecordBatchStream, MemoryTrackedStream, QueryMemoryTracker,
     SendableRecordBatchStream,
 };
 use common_telemetry::tracing::Span;
@@ -67,7 +67,7 @@ pub struct RegionScanExec {
     // TODO(ruihang): handle TimeWindowed dist via this parameter
     distribution: Option<TimeSeriesDistribution>,
     explain_verbose: bool,
-    query_memory_permit: Option<Arc<MemoryPermit>>,
+    query_memory_tracker: Option<QueryMemoryTracker>,
 }
 
 impl std::fmt::Debug for RegionScanExec {
@@ -91,7 +91,7 @@ impl RegionScanExec {
     pub fn new(
         scanner: RegionScannerRef,
         request: ScanRequest,
-        query_memory_permit: Option<Arc<MemoryPermit>>,
+        query_memory_tracker: Option<QueryMemoryTracker>,
     ) -> DfResult<Self> {
         let arrow_schema = scanner.schema().arrow_schema().clone();
         let scanner_props = scanner.properties();
@@ -226,7 +226,7 @@ impl RegionScanExec {
             is_partition_set: false,
             distribution: request.distribution,
             explain_verbose: false,
-            query_memory_permit,
+            query_memory_tracker,
         })
     }
 
@@ -299,7 +299,7 @@ impl RegionScanExec {
             is_partition_set: true,
             distribution: self.distribution,
             explain_verbose: self.explain_verbose,
-            query_memory_permit: self.query_memory_permit.clone(),
+            query_memory_tracker: self.query_memory_tracker.clone(),
         })
     }
 
@@ -387,8 +387,8 @@ impl ExecutionPlan for RegionScanExec {
             .scan_partition(&ctx, &self.metric, partition)
             .map_err(|e| DataFusionError::External(Box::new(e)))?;
 
-        let stream = if let Some(permit) = &self.query_memory_permit {
-            Box::pin(MemoryTrackedStream::new(stream, permit.clone()))
+        let stream = if let Some(tracker) = &self.query_memory_tracker {
+            Box::pin(MemoryTrackedStream::new(stream, tracker.clone()))
         } else {
             stream
         };
@@ -599,7 +599,12 @@ mod test {
             .primary_key(vec![1]);
         let region_metadata = Arc::new(builder.build().unwrap());
 
-        let scanner = Box::new(SinglePartitionScanner::new(stream, false, region_metadata));
+        let scanner = Box::new(SinglePartitionScanner::new(
+            stream,
+            false,
+            region_metadata,
+            None,
+        ));
         let plan = RegionScanExec::new(scanner, ScanRequest::default(), None).unwrap();
         let actual: SchemaRef = Arc::new(
             plan.properties
diff --git a/src/table/src/test_util/memtable.rs b/src/table/src/test_util/memtable.rs
index f1108bc533..ee73123aca 100644
--- a/src/table/src/test_util/memtable.rs
+++ b/src/table/src/test_util/memtable.rs
@@ -121,10 +121,10 @@ impl DataSource for MemtableDataSource {
         &self,
         request: ScanRequest,
     ) -> std::result::Result<SendableRecordBatchStream, BoxedError> {
-        let df_recordbatch = if let Some(indices) = request.projection {
+        let df_recordbatch = if let Some(indices) = request.projection_indices() {
             self.recordbatch
                 .df_record_batch()
-                .project(&indices)
+                .project(indices)
                 .context(TableProjectionSnafu)
                 .map_err(BoxedError::new)?
         } else {
@@ -198,8 +198,9 @@ mod test {
     async fn test_scan_with_projection() {
         let table = build_testing_table();
 
+        let projection_input = Some(vec![1].into());
         let scan_req = ScanRequest {
-            projection: Some(vec![1]),
+            projection_input,
             ..Default::default()
         };
         let stream = table.scan_to_stream(scan_req).await.unwrap();
diff --git a/tests-fuzz/Cargo.toml b/tests-fuzz/Cargo.toml
index a537ca0687..a38a4a13f8 100644
--- a/tests-fuzz/Cargo.toml
+++ b/tests-fuzz/Cargo.toml
@@ -46,6 +46,7 @@ paste.workspace = true
 rand = { workspace = true }
 rand_chacha = "0.9"
 reqwest = { workspace = true }
+rustls = { workspace = true, default-features = false, features = ["aws_lc_rs", "std", "tls12"] }
 schemars = "0.8"
 serde = { workspace = true }
 serde_json = { workspace = true }
@@ -100,6 +101,20 @@ test = false
 bench = false
 doc = false
 
+[[bin]]
+name = "fuzz_repartition_table_chaos"
+path = "targets/ddl/fuzz_repartition_table_chaos.rs"
+test = false
+bench = false
+doc = false
+
+[[bin]]
+name = "fuzz_repartition_metric_table"
+path = "targets/ddl/fuzz_repartition_metric_table.rs"
+test = false
+bench = false
+doc = false
+
 [[bin]]
 name = "fuzz_alter_table"
 path = "targets/ddl/fuzz_alter_table.rs"
diff --git a/tests-fuzz/README.md b/tests-fuzz/README.md
index 6807e19a1c..cc9d7eb84e 100644
--- a/tests-fuzz/README.md
+++ b/tests-fuzz/README.md
@@ -66,3 +66,23 @@ GT_FUZZ_OVERRIDE_SEED=6666 GT_FUZZ_OVERRIDE_ACTIONS=175 cargo fuzz run fuzz_targ
 ```
 
 For more details, visit [cargo fuzz](https://rust-fuzz.github.io/book/cargo-fuzz/tutorial.html) or run the command `cargo fuzz --help`.
+
+## Repartition Metric Dump Artifacts
+
+For `fuzz_repartition_metric_table`, dump artifacts are written under one run directory.
+
+- Table data snapshots: `<logical_table>.table-data.csv`
+- SQL traces per logical table: `<logical_table>.trace.sql`
+- Seed metadata: `seed.meta`
+
+SQL trace behavior:
+
+- Insert SQL is appended after successful execution with comment fields including
+  `started_at_ms` and `elapsed_ms`.
+- Repartition events are broadcast to all logical table trace files with comment fields including
+  `action_idx`, `started_at_ms`, `elapsed_ms`, and SQL text.
+
+Run directory lifecycle:
+
+- On success, the run directory is cleaned up.
+- On failure, the run directory is retained for CI/local diffing.
diff --git a/tests-fuzz/src/context.rs b/tests-fuzz/src/context.rs
index 1180e8ac7f..2f65ad56aa 100644
--- a/tests-fuzz/src/context.rs
+++ b/tests-fuzz/src/context.rs
@@ -376,6 +376,7 @@ mod tests {
                 table_name: expr.table_name.clone(),
                 target: partitions.last().unwrap().clone(),
                 into: vec![expected_exprs[2].clone(), expected_exprs[3].clone()],
+                wait: true,
             }))
             .unwrap();
         let partition_def = table_ctx.partition.as_ref().unwrap();
@@ -417,6 +418,7 @@ mod tests {
             .repartition(RepartitionExpr::Merge(MergePartitionExpr {
                 table_name: expr.table_name.clone(),
                 targets: vec![partitions[1].clone(), partitions[2].clone()],
+                wait: true,
             }))
             .unwrap();
         let partition_def = table_ctx.partition.as_ref().unwrap();
diff --git a/tests-fuzz/src/fake.rs b/tests-fuzz/src/fake.rs
index aa92e0293a..8910a39206 100644
--- a/tests-fuzz/src/fake.rs
+++ b/tests-fuzz/src/fake.rs
@@ -65,6 +65,26 @@ where
     _v: PhantomData<V>,
 }
 
+pub struct ConstGenerator<V> {
+    value: V,
+}
+
+impl<V> ConstGenerator<V> {
+    pub fn new(value: V) -> Self {
+        Self { value }
+    }
+}
+
+impl<R, V> Random<V, R> for ConstGenerator<V>
+where
+    R: Rng,
+    V: Clone,
+{
+    fn choose(&self, _rng: &mut R, amount: usize) -> Vec<V> {
+        vec![self.value.clone(); amount]
+    }
+}
+
 pub fn random_capitalize_map<R: Rng + 'static>(rng: &mut R, s: Ident) -> Ident {
     let mut v = s.value.chars().collect::<Vec<_>>();
 
diff --git a/tests-fuzz/src/generator/create_expr.rs b/tests-fuzz/src/generator/create_expr.rs
index fae6a95eda..261a310db2 100644
--- a/tests-fuzz/src/generator/create_expr.rs
+++ b/tests-fuzz/src/generator/create_expr.rs
@@ -193,6 +193,26 @@ fn generate_partition_def(
     }
 }
 
+fn generate_metric_partition(partitions: usize) -> Option<(Column, PartitionDef)> {
+    if partitions <= 1 {
+        return None;
+    }
+
+    let partition_column = Column {
+        name: Ident::new("host"),
+        column_type: ConcreteDataType::string_datatype(),
+        options: vec![ColumnOption::PrimaryKey],
+    };
+    let bounds = generate_partition_bounds(&partition_column.column_type, partitions - 1);
+    let partitions = SimplePartitions::new(partition_column.name.clone(), bounds);
+    let partition_def = PartitionDef {
+        columns: vec![partitions.column_name.clone()],
+        exprs: partitions.generate().unwrap(),
+    };
+
+    Some((partition_column, partition_def))
+}
+
 /// Generate a physical table with 2 columns: ts of TimestampType::Millisecond as time index and val of Float64Type.
 #[derive(Builder)]
 #[builder(pattern = "owned")]
@@ -201,6 +221,8 @@ pub struct CreatePhysicalTableExprGenerator<R: Rng + 'static> {
     name_generator: Box<dyn Random<Ident, R>>,
     #[builder(default = "false")]
     if_not_exists: bool,
+    #[builder(default = "0")]
+    partition: usize,
     #[builder(default, setter(into))]
     with_clause: HashMap<String, String>,
 }
@@ -215,25 +237,35 @@ impl<R: Rng + 'static> Generator<CreateTableExpr, R> for CreatePhysicalTableExpr
             options.insert(key.clone(), Value::from(value.clone()));
         }
 
+        let mut columns = vec![
+            Column {
+                name: Ident::new("ts"),
+                column_type: ConcreteDataType::timestamp_millisecond_datatype(),
+                options: vec![ColumnOption::TimeIndex],
+            },
+            Column {
+                name: Ident::new("val"),
+                column_type: ConcreteDataType::float64_datatype(),
+                options: vec![],
+            },
+        ];
+
+        let mut partition = None;
+        let mut primary_keys = vec![];
+        if let Some((partition_column, partition_def)) = generate_metric_partition(self.partition) {
+            columns.push(partition_column);
+            partition = Some(partition_def);
+            primary_keys.push(columns.len() - 1);
+        }
+
         Ok(CreateTableExpr {
             table_name: self.name_generator.generate(rng),
-            columns: vec![
-                Column {
-                    name: Ident::new("ts"),
-                    column_type: ConcreteDataType::timestamp_millisecond_datatype(),
-                    options: vec![ColumnOption::TimeIndex],
-                },
-                Column {
-                    name: Ident::new("val"),
-                    column_type: ConcreteDataType::float64_datatype(),
-                    options: vec![],
-                },
-            ],
+            columns,
             if_not_exists: self.if_not_exists,
-            partition: None,
+            partition,
             engine: "metric".to_string(),
             options,
-            primary_keys: vec![],
+            primary_keys,
         })
     }
 }
@@ -245,6 +277,8 @@ pub struct CreateLogicalTableExprGenerator<R: Rng + 'static> {
     physical_table_ctx: TableContextRef,
     labels: usize,
     if_not_exists: bool,
+    #[builder(default = "true")]
+    include_partition_column: bool,
     #[builder(default = "Box::new(WordGenerator)")]
     name_generator: Box<dyn Random<Ident, R>>,
 }
@@ -253,11 +287,11 @@ impl<R: Rng + 'static> Generator<CreateTableExpr, R> for CreateLogicalTableExprG
     type Error = Error;
 
     fn generate(&self, rng: &mut R) -> Result<CreateTableExpr> {
-        // Currently we mock the usage of GreptimeDB as Prometheus' backend, the physical table must have two columns.
+        // Currently we mock the usage of GreptimeDB as Prometheus' backend, the physical table must have ts and val.
         ensure!(
-            self.physical_table_ctx.columns.len() == 2,
+            self.physical_table_ctx.columns.len() >= 2,
             error::UnexpectedSnafu {
-                violated: "The physical table must have two columns"
+                violated: "The physical table must have at least two columns"
             }
         );
 
@@ -265,9 +299,16 @@ impl<R: Rng + 'static> Generator<CreateTableExpr, R> for CreateLogicalTableExprG
         let logical_table_name = self
             .physical_table_ctx
             .generate_unique_table_name(rng, self.name_generator.as_ref());
+        let mut physical_columns = self.physical_table_ctx.columns.clone();
+        if !self.include_partition_column
+            && let Some(partition_def) = &self.physical_table_ctx.partition
+        {
+            physical_columns.retain(|column| !partition_def.columns.contains(&column.name));
+        }
+
         let mut logical_table = CreateTableExpr {
             table_name: logical_table_name,
-            columns: self.physical_table_ctx.columns.clone(),
+            columns: physical_columns,
             if_not_exists: self.if_not_exists,
             partition: None,
             engine: "metric".to_string(),
@@ -459,6 +500,58 @@ mod tests {
         }));
     }
 
+    #[test]
+    fn test_create_physical_table_expr_generator_with_partition() {
+        let mut rng = rand::rng();
+        let physical_table_expr = CreatePhysicalTableExprGeneratorBuilder::default()
+            .partition(3)
+            .if_not_exists(false)
+            .build()
+            .unwrap()
+            .generate(&mut rng)
+            .unwrap();
+
+        assert_eq!(physical_table_expr.engine, "metric");
+        assert!(physical_table_expr.partition.is_some());
+        assert_eq!(physical_table_expr.partition.unwrap().exprs.len(), 3);
+    }
+
+    #[test]
+    fn test_create_logical_table_expr_generator_without_partition_column() {
+        let mut rng = rand::rng();
+        let physical_table_expr = CreatePhysicalTableExprGeneratorBuilder::default()
+            .partition(3)
+            .if_not_exists(false)
+            .build()
+            .unwrap()
+            .generate(&mut rng)
+            .unwrap();
+        let partition_columns = physical_table_expr
+            .partition
+            .as_ref()
+            .unwrap()
+            .columns
+            .clone();
+        let physical_table_ctx = Arc::new(TableContext::from(&physical_table_expr));
+
+        let logical_table_expr = CreateLogicalTableExprGeneratorBuilder::default()
+            .physical_table_ctx(physical_table_ctx)
+            .labels(3)
+            .include_partition_column(false)
+            .if_not_exists(false)
+            .build()
+            .unwrap()
+            .generate(&mut rng)
+            .unwrap();
+
+        assert!(
+            logical_table_expr
+                .columns
+                .iter()
+                .all(|column| !partition_columns.contains(&column.name))
+        );
+    }
+
     #[test]
     fn test_create_logical_table_expr_generator_deterministic() {
         let mut rng = rand_chacha::ChaCha8Rng::seed_from_u64(0);
diff --git a/tests-fuzz/src/generator/repartition_expr.rs b/tests-fuzz/src/generator/repartition_expr.rs
index bf2c98567c..a6e748a0c0 100644
--- a/tests-fuzz/src/generator/repartition_expr.rs
+++ b/tests-fuzz/src/generator/repartition_expr.rs
@@ -25,6 +25,8 @@ use crate::ir::repartition_expr::{MergePartitionExpr, SplitPartitionExpr};
 #[builder(pattern = "owned")]
 pub struct SplitPartitionExprGenerator {
     table_ctx: TableContextRef,
+    #[builder(default = "true")]
+    wait: bool,
 }
 
 impl<R: Rng + 'static> Generator<SplitPartitionExpr, R> for SplitPartitionExprGenerator {
@@ -66,6 +68,7 @@ impl<R: Rng + 'static> Generator<SplitPartitionExpr, R> for SplitPartitionExprGe
             table_name,
             target: from_expr,
             into: vec![left, right],
+            wait: self.wait,
         })
     }
 }
@@ -74,6 +77,8 @@ impl<R: Rng + 'static> Generator<SplitPartitionExpr, R> for SplitPartitionExprGe
 #[builder(pattern = "owned")]
 pub struct MergePartitionExprGenerator {
     table_ctx: TableContextRef,
+    #[builder(default = "true")]
+    wait: bool,
 }
 
 impl<R: Rng + 'static> Generator<MergePartitionExpr, R> for MergePartitionExprGenerator {
@@ -96,6 +101,7 @@ impl<R: Rng + 'static> Generator<MergePartitionExpr, R> for MergePartitionExprGe
         Ok(MergePartitionExpr {
             table_name,
             targets: vec![left, right],
+            wait: self.wait,
         })
     }
 }
diff --git a/tests-fuzz/src/ir.rs b/tests-fuzz/src/ir.rs
index e8c15dcf95..ce1628cd61 100644
--- a/tests-fuzz/src/ir.rs
+++ b/tests-fuzz/src/ir.rs
@@ -20,6 +20,7 @@ pub(crate) mod insert_expr;
 pub(crate) mod partition_expr;
 pub(crate) mod repartition_expr;
 pub(crate) mod select_expr;
+pub(crate) mod string_value;
 
 use core::fmt;
 use std::collections::HashMap;
@@ -126,20 +127,7 @@ pub fn generate_partition_bounds(datatype: &ConcreteDataType, bounds: usize) ->
         ConcreteDataType::Int64(_) => generate_values!(i64, bounds),
         ConcreteDataType::Float32(_) => generate_values!(f32, bounds),
         ConcreteDataType::Float64(_) => generate_values!(f64, bounds),
-        ConcreteDataType::String(_) => {
-            let base = b'A';
-            let range = b'z' - b'A';
-            let step = range / (bounds as u8 + 1);
-            (1..=bounds)
-                .map(|i| {
-                    Value::from(
-                        char::from(base + step * i as u8)
-                            .escape_default()
-                            .to_string(),
-                    )
-                })
-                .collect()
-        }
+        ConcreteDataType::String(_) => string_value::generate_partition_bounds(bounds),
         _ => unimplemented!("unsupported type: {datatype}"),
     }
 }
@@ -157,10 +145,7 @@ pub fn generate_random_value<R: Rng>(
         ConcreteDataType::Int64(_) => Value::from(rng.random::<i64>()),
         ConcreteDataType::Float32(_) => Value::from(rng.random::<f32>()),
         ConcreteDataType::Float64(_) => Value::from(rng.random::<f64>()),
-        ConcreteDataType::String(_) => match random_str {
-            Some(random) => Value::from(random.generate(rng).value),
-            None => Value::from(rng.random::<char>().to_string()),
-        },
+        ConcreteDataType::String(_) => string_value::generate_data_string_value(rng, random_str),
         ConcreteDataType::Date(_) => generate_random_date(rng),
 
         _ => unimplemented!("unsupported type: {datatype}"),
@@ -341,21 +326,7 @@ pub fn generate_partition_value<R: Rng + 'static>(
             }
         }
         datatypes::data_type::ConcreteDataType::String(_) => {
-            let upper = match first {
-                datatypes::value::Value::String(v) => v.as_utf8(),
-                _ => "",
-            };
-            if bound_idx == 0 {
-                if upper <= "A" {
-                    datatypes::value::Value::from("")
-                } else {
-                    datatypes::value::Value::from("A")
-                }
-            } else if bound_idx < bounds.len() {
-                bounds[bound_idx - 1].clone()
-            } else {
-                last.clone()
-            }
+            string_value::generate_partition_value(bounds, bound_idx)
         }
         _ => unimplemented!("unsupported partition column type: {column_type}"),
     }
diff --git a/tests-fuzz/src/ir/partition_expr.rs b/tests-fuzz/src/ir/partition_expr.rs
index c91dd487ae..908223366c 100644
--- a/tests-fuzz/src/ir/partition_expr.rs
+++ b/tests-fuzz/src/ir/partition_expr.rs
@@ -20,7 +20,7 @@ use snafu::ensure;
 
 use crate::context::TableContext;
 use crate::error::{self, Result};
-use crate::ir::{Ident, generate_random_value};
+use crate::ir::{Ident, generate_random_value, string_value};
 
 /// A partitioning scheme that divides a single column into multiple ranges based on provided bounds.
 ///
@@ -245,6 +245,10 @@ pub fn generate_unique_bound<R: Rng + 'static>(
     datatype: &ConcreteDataType,
     bounds: &[Value],
 ) -> Result<Value> {
+    if matches!(datatype, ConcreteDataType::String(_)) {
+        return string_value::generate_unique_partition_bound(rng, bounds);
+    }
+
     for _ in 0..16 {
         let candidate = generate_random_value(rng, datatype, None);
         if !bounds.contains(&candidate) {
diff --git a/tests-fuzz/src/ir/repartition_expr.rs b/tests-fuzz/src/ir/repartition_expr.rs
index f048308bce..5c8b401c8d 100644
--- a/tests-fuzz/src/ir/repartition_expr.rs
+++ b/tests-fuzz/src/ir/repartition_expr.rs
@@ -22,12 +22,16 @@ pub struct SplitPartitionExpr {
     pub table_name: Ident,
     pub target: PartitionExpr,
     pub into: Vec<PartitionExpr>,
+    #[serde(default = "default_wait")]
+    pub wait: bool,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct MergePartitionExpr {
     pub table_name: Ident,
     pub targets: Vec<PartitionExpr>,
+    #[serde(default = "default_wait")]
+    pub wait: bool,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -35,3 +39,7 @@ pub enum RepartitionExpr {
     Split(SplitPartitionExpr),
     Merge(MergePartitionExpr),
 }
+
+const fn default_wait() -> bool {
+    true
+}
diff --git a/tests-fuzz/src/ir/string_value.rs b/tests-fuzz/src/ir/string_value.rs
new file mode 100644
index 0000000000..6a53aa69de
--- /dev/null
+++ b/tests-fuzz/src/ir/string_value.rs
@@ -0,0 +1,162 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use datatypes::value::Value;
+use rand::Rng;
+
+use crate::error::{self, Result};
+use crate::generator::Random;
+use crate::ir::Ident;
+
+const READABLE_CHARSET: &[u8] = b"0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
+
+fn readable_token(index: usize) -> String {
+    let base = READABLE_CHARSET.len();
+    let mut n = index + 1;
+    let mut buf = Vec::new();
+
+    while n > 0 {
+        let rem = (n - 1) % base;
+        buf.push(READABLE_CHARSET[rem] as char);
+        n = (n - 1) / base;
+    }
+
+    buf.iter().rev().collect()
+}
+
+pub fn generate_data_string_value<R: Rng>(
+    rng: &mut R,
+    random_str: Option<&dyn Random<Ident, R>>,
+) -> Value {
+    match random_str {
+        Some(random) => Value::from(random.generate(rng).value),
+        None => {
+            let idx = rng.random_range(0..(READABLE_CHARSET.len() * READABLE_CHARSET.len() * 4));
+            Value::from(readable_token(idx))
+        }
+    }
+}
+
+/// Generates ordered readable string bounds for partition expressions.
+pub fn generate_partition_bounds(bounds: usize) -> Vec<Value> {
+    let token_space = READABLE_CHARSET.len() * READABLE_CHARSET.len() * 1024;
+    (1..=bounds)
+        .map(|i| {
+            let idx = i * token_space / (bounds + 1);
+            Value::from(readable_token(idx))
+        })
+        .collect()
+}
+
+/// Picks a representative string value for the target partition range.
+pub fn generate_partition_value(bounds: &[Value], bound_idx: usize) -> Value {
+    let first = bounds.first().unwrap();
+    let last = bounds.last().unwrap();
+    let upper = match first {
+        Value::String(v) => v.as_utf8(),
+        _ => "",
+    };
+
+    if bound_idx == 0 {
+        if upper <= "0" {
+            Value::from("")
+        } else {
+            Value::from("0")
+        }
+    } else if bound_idx < bounds.len() {
+        bounds[bound_idx - 1].clone()
+    } else {
+        last.clone()
+    }
+}
+
+/// Generates a unique readable bound not present in existing bounds.
+pub fn generate_unique_partition_bound<R: Rng>(rng: &mut R, bounds: &[Value]) -> Result<Value> {
+    let search_space = READABLE_CHARSET.len() * READABLE_CHARSET.len() * 1024;
+    let start = rng.random_range(0..search_space);
+    for offset in 0..search_space {
+        let idx = start + offset;
+        let candidate = Value::from(readable_token(idx));
+        if !bounds.contains(&candidate) {
+            return Ok(candidate);
+        }
+    }
+
+    error::UnexpectedSnafu {
+        violated: "unable to generate unique string partition bound".to_string(),
+    }
+    .fail()
+}
+
+#[cfg(test)]
+mod tests {
+    use rand::SeedableRng;
+    use rand_chacha::ChaCha8Rng;
+
+    use super::*;
+
+    #[test]
+    fn test_readable_token_grows_length() {
+        assert_eq!("0", readable_token(0));
+        assert_eq!("9", readable_token(9));
+        assert_eq!("A", readable_token(10));
+        assert_eq!("z", readable_token(61));
+        assert_eq!("00", readable_token(62));
+    }
+
+    #[test]
+    fn test_generate_partition_bounds_are_readable_and_unique() {
+        let bounds = generate_partition_bounds(8);
+        assert_eq!(8, bounds.len());
+
+        let mut values = bounds
+            .iter()
+            .map(|v| match v {
+                Value::String(s) => s.as_utf8().to_string(),
+                _ => panic!("expected string value"),
+            })
+            .collect::<Vec<_>>();
+        let mut dedup = values.clone();
+        dedup.sort();
+        dedup.dedup();
+        assert_eq!(values.len(), dedup.len());
+
+        for s in values.drain(..) {
+            assert!(s.chars().all(|c| c.is_ascii_alphanumeric()));
+        }
+    }
+
+    #[test]
+    fn test_generate_partition_value_for_string_bounds() {
+        let bounds = vec![Value::from("A"), Value::from("M")];
+        assert_eq!(Value::from("0"), generate_partition_value(&bounds, 0));
+        assert_eq!(Value::from("A"), generate_partition_value(&bounds, 1));
+        assert_eq!(Value::from("M"), generate_partition_value(&bounds, 2));
+    }
+
+    #[test]
+    fn test_generate_unique_partition_bound_not_in_existing() {
+        let mut rng = ChaCha8Rng::seed_from_u64(42);
+        let bounds = vec![Value::from("0"), Value::from("1"), Value::from("2")];
+        let candidate = generate_unique_partition_bound(&mut rng, &bounds).unwrap();
+        assert!(!bounds.contains(&candidate));
+        match candidate {
+            Value::String(s) => {
+                assert!(!s.as_utf8().is_empty());
+                assert!(s.as_utf8().chars().all(|c| c.is_ascii_alphanumeric()));
+            }
+            _ => panic!("expected string value"),
+        }
+    }
+}
diff --git a/tests-fuzz/src/lib.rs b/tests-fuzz/src/lib.rs
index 8900350224..1ecf4909f3 100644
--- a/tests-fuzz/src/lib.rs
+++ b/tests-fuzz/src/lib.rs
@@ -21,5 +21,19 @@ pub mod translator;
 pub mod utils;
 pub mod validator;
 
+use std::sync::OnceLock;
+
 #[cfg(test)]
 pub mod test_utils;
+
+static RUSTLS_CRYPTO_PROVIDER_INIT: OnceLock<()> = OnceLock::new();
+
+pub fn install_rustls_crypto_provider() {
+    RUSTLS_CRYPTO_PROVIDER_INIT.get_or_init(|| {
+        if rustls::crypto::CryptoProvider::get_default().is_none() {
+            let _ = rustls::crypto::CryptoProvider::install_default(
+                rustls::crypto::aws_lc_rs::default_provider(),
+            );
+        }
+    });
+}
diff --git a/tests-fuzz/src/translator.rs b/tests-fuzz/src/translator.rs
index 673b543f2c..4c5e0bb6a4 100644
--- a/tests-fuzz/src/translator.rs
+++ b/tests-fuzz/src/translator.rs
@@ -13,6 +13,8 @@
 // limitations under the License.
 
 mod common;
+/// Translator that converts insert expressions into CSV records.
+pub mod csv;
 pub mod mysql;
 pub mod postgres;
 
diff --git a/tests-fuzz/src/translator/csv.rs b/tests-fuzz/src/translator/csv.rs
new file mode 100644
index 0000000000..e95956862c
--- /dev/null
+++ b/tests-fuzz/src/translator/csv.rs
@@ -0,0 +1,121 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use crate::error::Error;
+use crate::ir::insert_expr::{InsertIntoExpr, RowValue};
+use crate::translator::DslTranslator;
+
+/// One CSV record converted from an insert row.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct CsvRecord {
+    /// Cell values in column order.
+    pub values: Vec<String>,
+}
+
+/// CSV records converted from an insert expression.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct CsvRecords {
+    /// Target table name from insert expression.
+    pub table_name: String,
+    /// Header values from insert columns.
+    pub headers: Vec<String>,
+    /// Converted row records.
+    pub records: Vec<CsvRecord>,
+}
+
+/// Translates `InsertIntoExpr` into CSV-writer-ready records.
+pub struct InsertExprToCsvRecordsTranslator;
+
+impl DslTranslator<InsertIntoExpr, CsvRecords> for InsertExprToCsvRecordsTranslator {
+    type Error = Error;
+
+    fn translate(&self, input: &InsertIntoExpr) -> Result<CsvRecords, Self::Error> {
+        let headers = input
+            .columns
+            .iter()
+            .map(|column| column.name.to_string())
+            .collect::<Vec<_>>();
+        let records = input
+            .values_list
+            .iter()
+            .map(|row| CsvRecord {
+                values: row.iter().map(Self::format_row_value).collect(),
+            })
+            .collect::<Vec<_>>();
+
+        Ok(CsvRecords {
+            table_name: input.table_name.to_string(),
+            headers,
+            records,
+        })
+    }
+}
+
+impl InsertExprToCsvRecordsTranslator {
+    fn format_row_value(value: &RowValue) -> String {
+        match value {
+            RowValue::Value(datatypes::value::Value::Null) => String::new(),
+            RowValue::Value(v) => v.to_string(),
+            RowValue::Default => "DEFAULT".to_string(),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use datatypes::data_type::ConcreteDataType;
+
+    use super::InsertExprToCsvRecordsTranslator;
+    use crate::ir::create_expr::ColumnOption;
+    use crate::ir::insert_expr::{InsertIntoExpr, RowValue};
+    use crate::ir::{Column, Ident};
+    use crate::translator::DslTranslator;
+
+    #[test]
+    fn test_translate_insert_expr_to_csv_records() {
+        let input = InsertIntoExpr {
+            table_name: Ident::new("metric_a"),
+            omit_column_list: false,
+            columns: vec![
+                Column {
+                    name: "host".into(),
+                    column_type: ConcreteDataType::string_datatype(),
+                    options: vec![ColumnOption::PrimaryKey],
+                },
+                Column {
+                    name: "value".into(),
+                    column_type: ConcreteDataType::float64_datatype(),
+                    options: vec![],
+                },
+            ],
+            values_list: vec![
+                vec![
+                    RowValue::Value(datatypes::value::Value::String("web-1".into())),
+                    RowValue::Value(datatypes::value::Value::Int32(15)),
+                ],
+                vec![
+                    RowValue::Value(datatypes::value::Value::Null),
+                    RowValue::Default,
+                ],
+            ],
+        };
+
+        let output = InsertExprToCsvRecordsTranslator.translate(&input).unwrap();
+        assert_eq!(output.table_name, "metric_a");
+        assert_eq!(output.headers, vec!["host", "value"]);
+        assert_eq!(output.records.len(), 2);
+        assert_eq!(output.records[0].values, vec!["web-1", "15"]);
+        assert_eq!(output.records[1].values, vec!["", "DEFAULT"]);
+    }
+}
diff --git a/tests-fuzz/src/translator/mysql/repartition_expr.rs b/tests-fuzz/src/translator/mysql/repartition_expr.rs
index b620d38f8f..56c27594b3 100644
--- a/tests-fuzz/src/translator/mysql/repartition_expr.rs
+++ b/tests-fuzz/src/translator/mysql/repartition_expr.rs
@@ -29,6 +29,7 @@ impl DslTranslator<RepartitionExpr, String> for RepartitionExprTranslator {
                 table_name,
                 target,
                 into,
+                wait,
             }) => {
                 let target_expr = format_partition_expr_sql(target);
                 let into_exprs = into
@@ -36,23 +37,26 @@ impl DslTranslator<RepartitionExpr, String> for RepartitionExprTranslator {
                     .map(format_partition_expr_sql)
                     .collect::<Vec<_>>()
                     .join(",\n  ");
+                let wait_clause = format_wait_clause(*wait);
                 Ok(format!(
-                    "ALTER TABLE {} SPLIT PARTITION (\n  {}\n) INTO (\n  {}\n);",
-                    table_name, target_expr, into_exprs
+                    "ALTER TABLE {} SPLIT PARTITION (\n  {}\n) INTO (\n  {}\n){};",
+                    table_name, target_expr, into_exprs, wait_clause
                 ))
             }
             RepartitionExpr::Merge(MergePartitionExpr {
                 table_name,
                 targets,
+                wait,
             }) => {
                 let merge_exprs = targets
                     .iter()
                     .map(format_partition_expr_sql)
                     .collect::<Vec<_>>()
                     .join(",\n  ");
+                let wait_clause = format_wait_clause(*wait);
                 Ok(format!(
-                    "ALTER TABLE {} MERGE PARTITION (\n  {}\n);",
-                    table_name, merge_exprs
+                    "ALTER TABLE {} MERGE PARTITION (\n  {}\n){};",
+                    table_name, merge_exprs, wait_clause
                 ))
             }
         }
@@ -63,6 +67,14 @@ fn format_partition_expr_sql(expr: &PartitionExpr) -> String {
     expr.to_parser_expr().to_string()
 }
 
+fn format_wait_clause(wait: bool) -> String {
+    if wait {
+        String::new()
+    } else {
+        " WITH (\n  WAIT = false\n)".to_string()
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use datatypes::value::Value;
@@ -83,6 +95,7 @@ mod tests {
                     .gt_eq(Value::Int32(5))
                     .and(col("id").lt(Value::Int32(10))),
             ],
+            wait: true,
         });
         let sql = RepartitionExprTranslator.translate(&expr).unwrap();
         let expected = r#"ALTER TABLE demo SPLIT PARTITION (
@@ -102,11 +115,37 @@ mod tests {
                 col("id").gt_eq(Value::Int32(10)),
                 col("id").gt_eq(Value::Int32(20)),
             ],
+            wait: true,
         });
         let sql = RepartitionExprTranslator.translate(&expr).unwrap();
         let expected = r#"ALTER TABLE demo MERGE PARTITION (
   id >= 10,
   id >= 20
+);"#;
+        assert_eq!(sql, expected);
+    }
+
+    #[test]
+    fn test_translate_split_expr_wait_false() {
+        let expr = RepartitionExpr::Split(SplitPartitionExpr {
+            table_name: "demo".into(),
+            target: col("id").lt(Value::Int32(10)),
+            into: vec![
+                col("id").lt(Value::Int32(5)),
+                col("id")
+                    .gt_eq(Value::Int32(5))
+                    .and(col("id").lt(Value::Int32(10))),
+            ],
+            wait: false,
+        });
+        let sql = RepartitionExprTranslator.translate(&expr).unwrap();
+        let expected = r#"ALTER TABLE demo SPLIT PARTITION (
+  id < 10
+) INTO (
+  id < 5,
+  id >= 5 AND id < 10
+) WITH (
+  WAIT = false
 );"#;
         assert_eq!(sql, expected);
     }
diff --git a/tests-fuzz/src/utils.rs b/tests-fuzz/src/utils.rs
index 0780f6c93d..2c5717bb71 100644
--- a/tests-fuzz/src/utils.rs
+++ b/tests-fuzz/src/utils.rs
@@ -15,17 +15,25 @@
 pub mod cluster_info;
 pub mod config;
 pub mod crd;
+/// CSV dump writer utilities for fuzz tests.
+pub mod csv_dump_writer;
 pub mod health;
 pub mod migration;
+pub mod network_chaos;
 pub mod partition;
 pub mod pod_failure;
 pub mod procedure;
 #[cfg(feature = "unstable")]
 pub mod process;
+pub mod retry;
+/// SQL dump writer utilities for fuzz tests.
+pub mod sql_dump_writer;
 pub mod wait;
 
 use std::env;
+use std::str::FromStr;
 
+use common_base::readable_size::ReadableSize;
 use common_telemetry::info;
 use common_telemetry::tracing::log::LevelFilter;
 use paste::paste;
@@ -45,6 +53,8 @@ const GT_MYSQL_ADDR: &str = "GT_MYSQL_ADDR";
 
 /// Connects to GreptimeDB via env variables.
 pub async fn init_greptime_connections_via_env() -> Connections {
+    crate::install_rustls_crypto_provider();
+
     let _ = dotenv::dotenv();
     let mysql = if let Ok(addr) = env::var(GT_MYSQL_ADDR) {
         Some(addr)
@@ -126,6 +136,14 @@ pub const GT_FUZZ_INPUT_MAX_COLUMNS: &str = "GT_FUZZ_INPUT_MAX_COLUMNS";
 pub const GT_FUZZ_INPUT_MAX_ALTER_ACTIONS: &str = "GT_FUZZ_INPUT_MAX_ALTER_ACTIONS";
 pub const GT_FUZZ_INPUT_MAX_INSERT_ACTIONS: &str = "GT_FUZZ_INPUT_MAX_INSERT_ACTIONS";
 pub const FUZZ_OVERRIDE_PREFIX: &str = "GT_FUZZ_OVERRIDE_";
+/// Enables CSV dump generation for fuzz runs.
+pub const GT_FUZZ_DUMP_TABLE_CSV: &str = "GT_FUZZ_DUMP_TABLE_CSV";
+/// Base directory for CSV dump sessions.
+pub const GT_FUZZ_DUMP_DIR: &str = "GT_FUZZ_DUMP_DIR";
+/// Directory suffix used by one CSV dump session.
+pub const GT_FUZZ_DUMP_SUFFIX: &str = "GT_FUZZ_DUMP_SUFFIX";
+/// Max in-memory CSV buffer size before auto flush.
+pub const GT_FUZZ_DUMP_BUFFER_MAX_BYTES: &str = "GT_FUZZ_DUMP_BUFFER_MAX_BYTES";
 
 /// Reads an override value for a fuzz parameter from env `GT_FUZZ_OVERRIDE_<NAME>`.
 pub fn get_fuzz_override<T>(name: &str) -> Option<T>
@@ -137,6 +155,33 @@ where
     env::var(&key).ok().and_then(|v| v.parse().ok())
 }
 
+/// Returns CSV dump base directory.
+pub fn get_gt_fuzz_dump_dir() -> String {
+    let _ = dotenv::dotenv();
+    env::var(GT_FUZZ_DUMP_DIR).unwrap_or_else(|_| "/tmp/greptime-fuzz-dumps".to_string())
+}
+
+/// Returns CSV dump directory suffix.
+pub fn get_gt_fuzz_dump_suffix() -> String {
+    let _ = dotenv::dotenv();
+    env::var(GT_FUZZ_DUMP_SUFFIX).unwrap_or_else(|_| ".repartition-metric-csv".to_string())
+}
+
+/// Returns max CSV in-memory buffer size.
+pub fn get_gt_fuzz_dump_buffer_max_bytes() -> usize {
+    let _ = dotenv::dotenv();
+    env::var(GT_FUZZ_DUMP_BUFFER_MAX_BYTES)
+        .ok()
+        .and_then(|value| {
+            value.parse::<usize>().ok().or_else(|| {
+                ReadableSize::from_str(&value)
+                    .ok()
+                    .map(|size| size.as_bytes() as usize)
+            })
+        })
+        .unwrap_or(8 * 1024 * 1024)
+}
+
 macro_rules! make_get_from_env_helper {
     ($key:expr, $default: expr) => {
         paste! {
diff --git a/tests-fuzz/src/utils/cluster_info.rs b/tests-fuzz/src/utils/cluster_info.rs
index 9f43711ca4..035f7908ac 100644
--- a/tests-fuzz/src/utils/cluster_info.rs
+++ b/tests-fuzz/src/utils/cluster_info.rs
@@ -77,3 +77,42 @@ pub async fn wait_for_all_datanode_online(greptime: MySqlPool, timeout: Duration
     )
     .await
 }
+
+pub async fn wait_for_all_datanode_offline(greptime: MySqlPool, timeout: Duration) {
+    wait_condition_fn(
+        timeout,
+        || {
+            let greptime = greptime.clone();
+            Box::pin(async move {
+                let nodes = fetch_nodes(&greptime)
+                    .await
+                    .unwrap()
+                    .into_iter()
+                    .flat_map(|node| {
+                        if node.peer_type == PEER_TYPE_DATANODE {
+                            Some(node)
+                        } else {
+                            None
+                        }
+                    })
+                    .collect::<Vec<_>>();
+                info!("Waits for datanode offline: {nodes:?}");
+                nodes
+            })
+        },
+        |nodes| {
+            nodes
+                .into_iter()
+                .map(|node| {
+                    info!(
+                        "Waits for datanode {} offline, active_time: {:?}",
+                        node.peer_id, node.active_time
+                    );
+                    parse_duration(&node.active_time.unwrap()).unwrap()
+                })
+                .all(|duration| duration >= Duration::from_secs(3))
+        },
+        Duration::from_secs(2),
+    )
+    .await
+}
diff --git a/tests-fuzz/src/utils/crd.rs b/tests-fuzz/src/utils/crd.rs
index c2fdcad9e0..df3a2cb86e 100644
--- a/tests-fuzz/src/utils/crd.rs
+++ b/tests-fuzz/src/utils/crd.rs
@@ -13,4 +13,5 @@
 // limitations under the License.
 
 pub mod common;
+pub mod network;
 pub mod pod;
diff --git a/tests-fuzz/src/utils/crd/network.rs b/tests-fuzz/src/utils/crd/network.rs
new file mode 100644
index 0000000000..cbae991a49
--- /dev/null
+++ b/tests-fuzz/src/utils/crd/network.rs
@@ -0,0 +1,291 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use derive_builder::Builder;
+use kube::CustomResource;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use super::common::{Mode, Selector};
+
+#[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize, Builder, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct Target {
+    mode: Mode,
+    selector: Selector,
+}
+
+#[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "kebab-case")]
+pub enum Action {
+    Netem,
+    Delay,
+    Loss,
+    Corrupt,
+    Partition,
+    Bandwidth,
+}
+
+#[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "kebab-case")]
+pub enum Direction {
+    From,
+    To,
+    Both,
+}
+
+#[derive(
+    CustomResource, Debug, PartialEq, Eq, Clone, Serialize, Deserialize, Builder, JsonSchema,
+)]
+#[kube(
+    group = "chaos-mesh.org",
+    version = "v1alpha1",
+    namespaced,
+    kind = "NetworkChaos",
+    plural = "networkchaos",
+    singular = "networkchaos",
+    derive = "PartialEq"
+)]
+#[serde(rename_all = "camelCase")]
+pub struct NetworkChaosSpec {
+    // Indicates the specific fault type. Available types include:
+    // netem, delay (network delay), loss (packet loss), duplicate (packet duplicating),
+    // corrupt (packet corrupt), partition (network partition), and bandwidth (network bandwidth limit).
+    // After you specify action field,
+    // refer to Description for action-related fields for other necessary field configuration.
+    action: Action,
+    // Used in combination with direction, making Chaos only effective for some packets.
+    #[builder(setter(into), default = "None")]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    target: Option<Target>,
+    // Indicates the direction of target packets. Available value include `from` (the packets from target),
+    // `to` (the packets to target), and `both` ( the packets from or to target).
+    // This parameter makes Chaos only take effect for a specific direction of packets.
+    #[builder(setter(into), default = "None")]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    direction: Option<Direction>,
+    // Specifies the mode of the experiment. The mode options include one (selecting a random Pod),
+    // all (selecting all eligible Pods), fixed (selecting a specified number of eligible Pods),
+    // fixed-percent (selecting a specified percentage of Pods from the eligible Pods),
+    // and random-max-percent (selecting the maximum percentage of Pods from the eligible Pods).
+    mode: Mode,
+    // Provides a parameter for the mode configuration, depending on mode.
+    // For example, when mode is set to fixed-percent, value specifies the percentage of Pods.
+    #[builder(setter(into), default = "None")]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    value: Option<String>,
+    // Specifies the target Pod.
+    selector: Selector,
+    // Indicates the network targets except for Kubernetes, which can be IPv4 addresses or domains.
+    // This parameter only works with `direction: to`.
+    #[builder(setter(into), default = "Vec::new()")]
+    #[serde(skip_serializing_if = "Vec::is_empty", default)]
+    external_targets: Vec<String>,
+    // Specifies the affected network interface
+    #[builder(setter(into), default = "None")]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    device: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into), default = "None")]
+    delay: Option<Delay>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into), default = "None")]
+    loss: Option<Loss>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into), default = "None")]
+    duplicate: Option<Duplicate>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into), default = "None")]
+    corrupt: Option<Corrupt>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into), default = "None")]
+    bandwidth: Option<Bandwidth>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into), default = "None")]
+    duration: Option<String>,
+}
+
+#[derive(Debug, Default, PartialEq, Eq, Clone, Serialize, Deserialize, Builder, JsonSchema)]
+#[serde(default)]
+#[builder(default)]
+pub struct Delay {
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into, strip_option))]
+    // Indicates the network latency.
+    latency: Option<String>,
+    // Indicates the correlation between the current latency and the previous one.
+    // Range of value: [0, 100]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into, strip_option))]
+    correlation: Option<String>,
+    // Indicates the range of the network latency.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into, strip_option))]
+    jitter: Option<String>,
+    // Indicates the status of network packet reordering
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into, strip_option))]
+    reorder: Option<Reorder>,
+}
+
+#[derive(Debug, Default, PartialEq, Eq, Clone, Serialize, Deserialize, JsonSchema, Builder)]
+#[serde(default)]
+#[builder(default)]
+pub struct Reorder {
+    // Indicates the probability to reorder
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into, strip_option))]
+    reorder: Option<String>,
+    // Indicates the correlation between this time's length of delay time
+    // and the previous time's length of delay time. Range of value: [0, 100]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into, strip_option))]
+    correlation: Option<String>,
+    // Indicates the gap before and after packet reordering.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into, strip_option))]
+    gap: Option<i32>,
+}
+
+#[derive(Debug, Default, PartialEq, Eq, Clone, Serialize, Deserialize, JsonSchema, Builder)]
+#[serde(default)]
+#[builder(default, setter(into, strip_option))]
+pub struct Loss {
+    // Indicates the probability of packet loss.
+    // Range of value: [0, 100].
+    #[serde(skip_serializing_if = "Option::is_none")]
+    loss: Option<String>,
+    // Indicates the correlation between the probability of current packet loss
+    // and the previous time's packet loss. Range of value: [0, 100].
+    #[serde(skip_serializing_if = "Option::is_none")]
+    correlation: Option<String>,
+}
+
+#[derive(Debug, Default, PartialEq, Eq, Clone, Serialize, Deserialize, JsonSchema, Builder)]
+#[serde(default)]
+#[builder(default, setter(into, strip_option))]
+pub struct Duplicate {
+    // Indicates the probability of packet duplicating.
+    // Range of value: [0, 100]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    duplicate: Option<String>,
+    // Indicates the correlation between the probability of current packet duplicating
+    // and the previous time's packet duplicating. Range of value: [0, 100]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    correlation: Option<String>,
+}
+
+#[derive(Debug, Default, PartialEq, Eq, Clone, Serialize, Deserialize, JsonSchema, Builder)]
+#[serde(default)]
+#[builder(default, setter(into, strip_option))]
+pub struct Corrupt {
+    // Indicates the probability of packet corruption. Range of value: [0, 100]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    corrupt: Option<String>,
+    // Indicates the correlation between the probability of current packet corruption
+    // and the previous time's packet corruption. Range of value: [0, 100]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    correlation: Option<String>,
+}
+
+#[derive(Debug, Default, PartialEq, Eq, Clone, Serialize, Deserialize, JsonSchema, Builder)]
+#[serde(default)]
+#[builder(default)]
+pub struct Bandwidth {
+    // Indicates the rate of bandwidth limit. e.g., 1mbps.
+    rate: String,
+    // Indicates the number of bytes waiting in queue.
+    limit: u32,
+    // Indicates the maximum number of bytes that can be sent instantaneously.
+    buffer: u32,
+    // Indicates the maximum consumption of `bucket` (usually not set)
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into, strip_option))]
+    peakrate: Option<u64>,
+    // Indicates the size of peakrate bucket (usually not set).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[builder(setter(into, strip_option))]
+    minburst: Option<u32>,
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::BTreeMap;
+
+    use super::{Action, Direction, NetworkChaosSpecBuilder, TargetBuilder};
+    use crate::utils::crd::common::{Mode, SelectorBuilder};
+    use crate::utils::crd::network::{DelayBuilder, NetworkChaos};
+
+    #[test]
+    fn test_serde() {
+        let mut fs = BTreeMap::new();
+        fs.insert("app".into(), "datanode".into());
+        let selector = SelectorBuilder::default()
+            .namespaces(vec!["default".into()])
+            .label_selectors(fs)
+            .build()
+            .unwrap();
+        let target = TargetBuilder::default()
+            .mode(Mode::All)
+            .selector(selector.clone())
+            .build()
+            .unwrap();
+        let delay = DelayBuilder::default()
+            .latency("100ms")
+            .correlation("100")
+            .jitter("0ms")
+            .build()
+            .unwrap();
+        let spec = NetworkChaosSpecBuilder::default()
+            .action(Action::Delay)
+            .target(target)
+            .direction(Direction::Both)
+            .mode(Mode::One)
+            .selector(selector)
+            .delay(delay)
+            .duration(Some("10s".to_string()))
+            .build()
+            .unwrap();
+        let crd = NetworkChaos::new("my-delay", spec);
+        let ser = serde_yaml::to_string(&crd).unwrap();
+
+        let expected = r#"apiVersion: chaos-mesh.org/v1alpha1
+kind: NetworkChaos
+metadata:
+  name: my-delay
+spec:
+  action: delay
+  target:
+    mode: all
+    selector:
+      namespaces:
+      - default
+      labelSelectors:
+        app: datanode
+  direction: both
+  mode: one
+  selector:
+    namespaces:
+    - default
+    labelSelectors:
+      app: datanode
+  delay:
+    latency: 100ms
+    correlation: '100'
+    jitter: 0ms
+  duration: 10s
+"#;
+        assert_eq!(expected, ser);
+    }
+}
diff --git a/tests-fuzz/src/utils/csv_dump_writer.rs b/tests-fuzz/src/utils/csv_dump_writer.rs
new file mode 100644
index 0000000000..de16a23c24
--- /dev/null
+++ b/tests-fuzz/src/utils/csv_dump_writer.rs
@@ -0,0 +1,383 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::{HashMap, HashSet};
+use std::fs::{File, OpenOptions, create_dir_all, remove_dir_all};
+use std::io::Write;
+use std::path::{Path, PathBuf};
+
+use common_telemetry::{info, warn};
+use common_time::util::current_time_millis;
+use snafu::ResultExt;
+
+use crate::error::{self, Result};
+use crate::translator::csv::CsvRecords;
+use crate::utils::{
+    get_gt_fuzz_dump_buffer_max_bytes, get_gt_fuzz_dump_dir, get_gt_fuzz_dump_suffix,
+};
+
+/// Metadata for one CSV dump session.
+#[derive(Debug, Clone)]
+pub struct CsvDumpMetadata {
+    /// Fuzz target name.
+    pub target: String,
+    /// Seed used by current fuzz input.
+    pub seed: u64,
+    /// Repartition action count.
+    pub actions: usize,
+    /// Initial partition count.
+    pub partitions: usize,
+    /// Logical table count.
+    pub tables: usize,
+    /// Session start time in unix milliseconds.
+    pub started_at_unix_ms: i64,
+}
+
+impl CsvDumpMetadata {
+    /// Builds dump metadata with current timestamp.
+    pub fn new(
+        target: impl Into<String>,
+        seed: u64,
+        actions: usize,
+        partitions: usize,
+        tables: usize,
+    ) -> Self {
+        Self {
+            target: target.into(),
+            seed,
+            actions,
+            partitions,
+            tables,
+            started_at_unix_ms: current_time_millis(),
+        }
+    }
+}
+
+/// Session writer for staged CSV dump records.
+#[derive(Debug)]
+pub struct CsvDumpSession {
+    /// Session metadata.
+    pub metadata: CsvDumpMetadata,
+    /// Session directory path.
+    pub run_dir: PathBuf,
+    /// Max in-memory buffer size before auto flush.
+    pub max_buffer_bytes: usize,
+    records: Vec<CsvRecords>,
+    buffered_bytes: usize,
+    written_tables: HashSet<String>,
+    full_headers_by_table: HashMap<String, Vec<String>>,
+}
+
+impl CsvDumpSession {
+    /// Creates session directory and writes seed metadata file.
+    pub fn new(metadata: CsvDumpMetadata) -> Result<Self> {
+        Self::new_with_buffer_limit(metadata, get_gt_fuzz_dump_buffer_max_bytes())
+    }
+
+    /// Creates session with a custom in-memory buffer limit.
+    pub fn new_with_buffer_limit(
+        metadata: CsvDumpMetadata,
+        max_buffer_bytes: usize,
+    ) -> Result<Self> {
+        let run_dir = build_run_dir(&metadata);
+        create_dir_all(&run_dir).context(error::CreateFileSnafu {
+            path: run_dir.to_string_lossy().to_string(),
+        })?;
+        write_seed_meta(&run_dir, &metadata)?;
+        info!(
+            "Create csv dump session, target: {}, run_dir: {}, max_buffer_bytes: {}",
+            metadata.target,
+            run_dir.display(),
+            max_buffer_bytes
+        );
+
+        Ok(Self {
+            metadata,
+            run_dir,
+            max_buffer_bytes,
+            records: Vec::new(),
+            buffered_bytes: 0,
+            written_tables: HashSet::new(),
+            full_headers_by_table: HashMap::new(),
+        })
+    }
+
+    /// Appends one table CSV records batch with full table headers.
+    pub fn append(&mut self, records: CsvRecords, full_headers: Vec<String>) -> Result<()> {
+        self.full_headers_by_table
+            .entry(records.table_name.clone())
+            .or_insert(full_headers);
+        self.buffered_bytes += estimate_csv_records_size(&records);
+        self.records.push(records);
+        if self.buffered_bytes >= self.max_buffer_bytes {
+            self.flush_buffered_records()?;
+        }
+        Ok(())
+    }
+
+    /// Flushes all appended batches to CSV files.
+    pub fn flush_all(&mut self) -> Result<()> {
+        self.flush_buffered_records()
+    }
+
+    /// Removes session directory after successful validation.
+    pub fn cleanup_on_success(&self) -> std::io::Result<()> {
+        match remove_dir_all(&self.run_dir) {
+            Ok(_) => {
+                info!(
+                    "Cleanup csv dump directory on success: {}",
+                    self.run_dir.display()
+                );
+                Ok(())
+            }
+            Err(err) => {
+                warn!(
+                    "Cleanup csv dump directory failed: {}, error: {:?}",
+                    self.run_dir.display(),
+                    err
+                );
+                Err(err)
+            }
+        }
+    }
+
+    fn flush_buffered_records(&mut self) -> Result<()> {
+        if self.records.is_empty() {
+            return Ok(());
+        }
+        for batch in &self.records {
+            write_batch_csv(
+                &self.run_dir,
+                batch,
+                &mut self.written_tables,
+                &self.full_headers_by_table,
+            )?;
+        }
+        self.records.clear();
+        self.buffered_bytes = 0;
+        Ok(())
+    }
+}
+
+fn write_seed_meta(run_dir: &Path, metadata: &CsvDumpMetadata) -> Result<()> {
+    let path = run_dir.join("seed.meta");
+    let mut file = File::create(&path).context(error::CreateFileSnafu {
+        path: path.to_string_lossy().to_string(),
+    })?;
+
+    let content = format!(
+        "target={}\nseed={}\nactions={}\npartitions={}\ntables={}\nstarted_at_unix_ms={}\n",
+        metadata.target,
+        metadata.seed,
+        metadata.actions,
+        metadata.partitions,
+        metadata.tables,
+        metadata.started_at_unix_ms,
+    );
+    file.write_all(content.as_bytes())
+        .context(error::WriteFileSnafu {
+            path: path.to_string_lossy().to_string(),
+        })
+}
+
+fn write_batch_csv(
+    run_dir: &Path,
+    batch: &CsvRecords,
+    written_tables: &mut HashSet<String>,
+    full_headers_by_table: &HashMap<String, Vec<String>>,
+) -> Result<()> {
+    let output_headers = full_headers_by_table
+        .get(&batch.table_name)
+        .cloned()
+        .unwrap_or_else(|| batch.headers.clone());
+    let file_name = format!("{}.table-data.csv", sanitize_file_name(&batch.table_name));
+    let path = run_dir.join(file_name);
+    let mut file = OpenOptions::new()
+        .create(true)
+        .append(true)
+        .open(&path)
+        .context(error::CreateFileSnafu {
+            path: path.to_string_lossy().to_string(),
+        })?;
+
+    if written_tables.insert(batch.table_name.clone()) {
+        file.write_all(join_line(&output_headers).as_bytes())
+            .context(error::WriteFileSnafu {
+                path: path.to_string_lossy().to_string(),
+            })?;
+        file.write_all(b"\n").context(error::WriteFileSnafu {
+            path: path.to_string_lossy().to_string(),
+        })?;
+    }
+
+    let header_index = batch
+        .headers
+        .iter()
+        .enumerate()
+        .map(|(idx, header)| (header.as_str(), idx))
+        .collect::<HashMap<_, _>>();
+
+    for record in &batch.records {
+        let aligned_values = output_headers
+            .iter()
+            .map(|header| {
+                header_index
+                    .get(header.as_str())
+                    .and_then(|idx| record.values.get(*idx))
+                    .cloned()
+                    .unwrap_or_default()
+            })
+            .collect::<Vec<_>>();
+        file.write_all(join_line(&aligned_values).as_bytes())
+            .context(error::WriteFileSnafu {
+                path: path.to_string_lossy().to_string(),
+            })?;
+        file.write_all(b"\n").context(error::WriteFileSnafu {
+            path: path.to_string_lossy().to_string(),
+        })?;
+    }
+
+    Ok(())
+}
+
+fn estimate_csv_records_size(records: &CsvRecords) -> usize {
+    let headers = records.headers.iter().map(String::len).sum::<usize>();
+    let rows = records
+        .records
+        .iter()
+        .flat_map(|record| record.values.iter())
+        .map(String::len)
+        .sum::<usize>();
+    headers + rows
+}
+
+fn join_line(cells: &[String]) -> String {
+    cells
+        .iter()
+        .map(|cell| escape_csv_cell(cell))
+        .collect::<Vec<_>>()
+        .join(",")
+}
+
+fn escape_csv_cell(value: &str) -> String {
+    if value.contains([',', '"', '\n', '\r']) {
+        format!("\"{}\"", value.replace('"', "\"\""))
+    } else {
+        value.to_string()
+    }
+}
+
+fn sanitize_file_name(raw: &str) -> String {
+    raw.chars()
+        .map(|ch| {
+            if ch.is_ascii_alphanumeric() || ch == '_' || ch == '-' {
+                ch
+            } else {
+                '_'
+            }
+        })
+        .collect()
+}
+
+fn build_run_dir(metadata: &CsvDumpMetadata) -> PathBuf {
+    let base = PathBuf::from(get_gt_fuzz_dump_dir());
+    let suffix = get_gt_fuzz_dump_suffix();
+    let name = format!(
+        "{}_seed_{}_actions_{}_ts_{}{}",
+        metadata.target, metadata.seed, metadata.actions, metadata.started_at_unix_ms, suffix
+    );
+    base.join(name)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{CsvDumpMetadata, CsvDumpSession};
+    use crate::translator::csv::{CsvRecord, CsvRecords};
+
+    #[test]
+    fn test_create_session_and_flush() {
+        let mut session = CsvDumpSession::new_with_buffer_limit(
+            CsvDumpMetadata::new("fuzz_case", 1, 2, 3, 4),
+            1024,
+        )
+        .unwrap();
+        session
+            .append(
+                CsvRecords {
+                    table_name: "metric-a".to_string(),
+                    headers: vec!["host".to_string(), "value".to_string()],
+                    records: vec![CsvRecord {
+                        values: vec!["web-1".to_string(), "10".to_string()],
+                    }],
+                },
+                vec!["host".to_string(), "value".to_string()],
+            )
+            .unwrap();
+        session.flush_all().unwrap();
+
+        assert!(session.run_dir.exists());
+        assert!(session.run_dir.join("seed.meta").exists());
+        assert!(session.run_dir.join("metric-a.table-data.csv").exists());
+    }
+
+    #[test]
+    fn test_auto_flush_on_buffer_limit() {
+        let mut session =
+            CsvDumpSession::new_with_buffer_limit(CsvDumpMetadata::new("fuzz_case", 5, 2, 3, 4), 1)
+                .unwrap();
+        session
+            .append(
+                CsvRecords {
+                    table_name: "metric-b".to_string(),
+                    headers: vec!["host".to_string()],
+                    records: vec![CsvRecord {
+                        values: vec!["web-2".to_string()],
+                    }],
+                },
+                vec!["host".to_string()],
+            )
+            .unwrap();
+
+        assert!(session.run_dir.join("metric-b.table-data.csv").exists());
+        assert_eq!(session.buffered_bytes, 0);
+    }
+
+    #[test]
+    fn test_flush_with_partial_headers_uses_full_headers() {
+        let mut session = CsvDumpSession::new_with_buffer_limit(
+            CsvDumpMetadata::new("fuzz_case", 7, 2, 3, 4),
+            1024,
+        )
+        .unwrap();
+        session
+            .append(
+                CsvRecords {
+                    table_name: "metric-c".to_string(),
+                    headers: vec!["host".to_string(), "value".to_string()],
+                    records: vec![CsvRecord {
+                        values: vec!["web-3".to_string(), "12".to_string()],
+                    }],
+                },
+                vec!["host".to_string(), "idc".to_string(), "value".to_string()],
+            )
+            .unwrap();
+        session.flush_all().unwrap();
+
+        let file =
+            std::fs::read_to_string(session.run_dir.join("metric-c.table-data.csv")).unwrap();
+        let mut lines = file.lines();
+        assert_eq!(lines.next().unwrap(), "host,idc,value");
+        assert_eq!(lines.next().unwrap(), "web-3,,12");
+    }
+}
diff --git a/tests-fuzz/src/utils/network_chaos.rs b/tests-fuzz/src/utils/network_chaos.rs
new file mode 100644
index 0000000000..26d4becfe8
--- /dev/null
+++ b/tests-fuzz/src/utils/network_chaos.rs
@@ -0,0 +1,94 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::BTreeMap;
+
+use kube::Api;
+
+use crate::error::Result;
+use crate::utils::crd::common::{Mode, Selector, SelectorBuilder};
+use crate::utils::crd::network::{
+    Action, Direction, NetworkChaos, NetworkChaosSpecBuilder, TargetBuilder,
+};
+
+fn build_datanode_selector(namespace: &str, cluster_name: &str) -> Selector {
+    let mut selector = BTreeMap::new();
+    selector.insert(
+        "app.greptime.io/component".into(),
+        format!("{cluster_name}-datanode"),
+    );
+
+    SelectorBuilder::default()
+        .namespaces(vec![namespace.to_string()])
+        .label_selectors(selector)
+        .build()
+        .unwrap()
+}
+
+fn build_metasrv_selector(namespace: &str, cluster_name: &str) -> Selector {
+    let mut selector = BTreeMap::new();
+    selector.insert(
+        "app.greptime.io/component".into(),
+        format!("{cluster_name}-meta"),
+    );
+
+    SelectorBuilder::default()
+        .namespaces(vec![namespace.to_string()])
+        .label_selectors(selector)
+        .build()
+        .unwrap()
+}
+
+/// Injects a network partition between a datanode pod and metasrv.
+pub async fn inject_datanode_metasrv_network_partition(
+    client: kube::Client,
+    namespace: &str,
+    cluster_name: &str,
+    duration_secs: usize,
+) -> Result<String> {
+    let selector = build_datanode_selector(namespace, cluster_name);
+    let target = TargetBuilder::default()
+        .mode(Mode::All)
+        .selector(build_metasrv_selector(namespace, cluster_name))
+        .build()
+        .unwrap();
+
+    let spec = NetworkChaosSpecBuilder::default()
+        .action(Action::Partition)
+        .direction(Direction::Both)
+        .mode(Mode::All)
+        .selector(selector)
+        .target(target)
+        .duration(format!("{duration_secs}s"))
+        .build()
+        .unwrap();
+
+    let chaos_name = "datanode-metasrv-network-partition".to_string();
+    let cr = NetworkChaos::new(&chaos_name, spec);
+    let api: Api<NetworkChaos> = Api::namespaced(client, namespace);
+    api.create(&Default::default(), &cr).await.unwrap();
+
+    Ok(chaos_name)
+}
+
+/// Recovers network chaos by deleting the associated NetworkChaos resource.
+pub async fn recover_network_chaos(
+    client: kube::Client,
+    namespace: &str,
+    name: &str,
+) -> Result<()> {
+    let api: Api<NetworkChaos> = Api::namespaced(client, namespace);
+    api.delete(name, &Default::default()).await.unwrap();
+    Ok(())
+}
diff --git a/tests-fuzz/src/utils/partition.rs b/tests-fuzz/src/utils/partition.rs
index d3dc30061d..89a684326b 100644
--- a/tests-fuzz/src/utils/partition.rs
+++ b/tests-fuzz/src/utils/partition.rs
@@ -36,7 +36,7 @@ pub struct PartitionCount {
 }
 
 pub async fn count_partitions(db: &MySqlPool, datanode_id: u64) -> Result<PartitionCount> {
-    let sql = "select count(1) as count from information_schema.region_peers where peer_id == ?";
+    let sql = "select count(1) as count from information_schema.region_peers where peer_id = ?";
     sqlx::query_as::<_, PartitionCount>(sql)
         .bind(datanode_id)
         .fetch_one(db)
diff --git a/tests-fuzz/src/utils/retry.rs b/tests-fuzz/src/utils/retry.rs
new file mode 100644
index 0000000000..06d1ede54f
--- /dev/null
+++ b/tests-fuzz/src/utils/retry.rs
@@ -0,0 +1,49 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::future::Future;
+use std::time::Duration;
+
+use common_telemetry::warn;
+
+pub async fn retry_with_backoff<T, E, Fut, F>(
+    mut operation: F,
+    max_attempts: usize,
+    init_backoff: Duration,
+    max_backoff: Duration,
+) -> Result<T, E>
+where
+    F: FnMut() -> Fut,
+    Fut: Future<Output = Result<T, E>>,
+    E: std::fmt::Debug,
+{
+    let mut backoff = init_backoff;
+    for attempt in 0..max_attempts {
+        match operation().await {
+            Ok(result) => return Ok(result),
+            Err(err) if attempt + 1 == max_attempts => return Err(err),
+            Err(err) => {
+                let current_attempt = attempt + 1;
+                warn!(
+                    "Retryable operation failed, attempt: {}, max_attempts: {}, backoff: {:?}, error: {:?}",
+                    current_attempt, max_attempts, backoff, err
+                );
+                tokio::time::sleep(backoff).await;
+                backoff = std::cmp::min(backoff * 2, max_backoff);
+            }
+        }
+    }
+
+    panic!("retry loop should always return")
+}
diff --git a/tests-fuzz/src/utils/sql_dump_writer.rs b/tests-fuzz/src/utils/sql_dump_writer.rs
new file mode 100644
index 0000000000..6f098d9584
--- /dev/null
+++ b/tests-fuzz/src/utils/sql_dump_writer.rs
@@ -0,0 +1,267 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::HashMap;
+use std::fs::{OpenOptions, create_dir_all};
+use std::io::Write;
+use std::path::PathBuf;
+
+use snafu::ResultExt;
+
+use crate::error::{self, Result};
+use crate::utils::get_gt_fuzz_dump_buffer_max_bytes;
+
+/// Session writer for table-scoped SQL trace files.
+#[derive(Debug)]
+pub struct SqlDumpSession {
+    /// Session directory path.
+    pub run_dir: PathBuf,
+    /// Max in-memory buffer size before auto flush.
+    pub max_buffer_bytes: usize,
+    buffered_bytes: usize,
+    entries_by_table: HashMap<String, Vec<String>>,
+}
+
+impl SqlDumpSession {
+    /// Creates SQL dump session with default buffer limit.
+    pub fn new(run_dir: PathBuf) -> Result<Self> {
+        Self::new_with_buffer_limit(run_dir, get_gt_fuzz_dump_buffer_max_bytes())
+    }
+
+    /// Creates SQL dump session with custom buffer limit.
+    pub fn new_with_buffer_limit(run_dir: PathBuf, max_buffer_bytes: usize) -> Result<Self> {
+        create_dir_all(&run_dir).context(error::CreateFileSnafu {
+            path: run_dir.to_string_lossy().to_string(),
+        })?;
+
+        Ok(Self {
+            run_dir,
+            max_buffer_bytes,
+            buffered_bytes: 0,
+            entries_by_table: HashMap::new(),
+        })
+    }
+
+    /// Appends one SQL statement for a logical table.
+    pub fn append_sql(&mut self, table: &str, sql: &str, comment: Option<&str>) -> Result<()> {
+        let entry = format_sql_entry(sql, comment);
+        self.push_entry(table, entry)?;
+        Ok(())
+    }
+
+    /// Broadcasts one comment event to all table trace files.
+    pub fn broadcast_event<I, T>(&mut self, tables: I, event: &str, sql: &str) -> Result<()>
+    where
+        I: IntoIterator<Item = T>,
+        T: AsRef<str>,
+    {
+        let entry = format_sql_entry(sql, Some(event));
+        for table in tables {
+            self.push_entry(table.as_ref(), entry.clone())?;
+        }
+        Ok(())
+    }
+
+    /// Flushes all staged SQL traces to table-scoped files.
+    pub fn flush_all(&mut self) -> Result<()> {
+        self.flush_buffered_entries()
+    }
+
+    fn push_entry(&mut self, table: &str, entry: String) -> Result<()> {
+        self.buffered_bytes += entry.len();
+        self.entries_by_table
+            .entry(table.to_string())
+            .or_default()
+            .push(entry);
+
+        if self.buffered_bytes >= self.max_buffer_bytes {
+            self.flush_buffered_entries()?;
+        }
+        Ok(())
+    }
+
+    fn flush_buffered_entries(&mut self) -> Result<()> {
+        if self.entries_by_table.is_empty() {
+            return Ok(());
+        }
+
+        for (table, entries) in &self.entries_by_table {
+            let path = self
+                .run_dir
+                .join(format!("{}.trace.sql", sanitize_file_name(table)));
+            let mut file = OpenOptions::new()
+                .create(true)
+                .append(true)
+                .open(&path)
+                .context(error::CreateFileSnafu {
+                    path: path.to_string_lossy().to_string(),
+                })?;
+
+            for entry in entries {
+                file.write_all(entry.as_bytes())
+                    .context(error::WriteFileSnafu {
+                        path: path.to_string_lossy().to_string(),
+                    })?;
+                file.write_all(b"\n").context(error::WriteFileSnafu {
+                    path: path.to_string_lossy().to_string(),
+                })?;
+            }
+        }
+
+        self.entries_by_table.clear();
+        self.buffered_bytes = 0;
+        Ok(())
+    }
+}
+
+fn format_sql_entry(sql: &str, comment: Option<&str>) -> String {
+    let normalized_sql = normalize_sql(sql);
+    if let Some(comment) = comment {
+        format!("{}\n{normalized_sql}", format_comment(comment))
+    } else {
+        normalized_sql
+    }
+}
+
+fn format_comment(comment: &str) -> String {
+    comment
+        .lines()
+        .map(|line| format!("-- {line}"))
+        .collect::<Vec<_>>()
+        .join("\n")
+}
+
+fn normalize_sql(sql: &str) -> String {
+    let trimmed = sql.trim_end();
+    if trimmed.ends_with(';') {
+        trimmed.to_string()
+    } else {
+        format!("{trimmed};")
+    }
+}
+
+fn sanitize_file_name(raw: &str) -> String {
+    raw.chars()
+        .map(|ch| {
+            if ch.is_ascii_alphanumeric() || ch == '_' || ch == '-' {
+                ch
+            } else {
+                '_'
+            }
+        })
+        .collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use std::time::{SystemTime, UNIX_EPOCH};
+
+    use super::SqlDumpSession;
+
+    #[test]
+    fn test_append_sql_writes_table_trace_file() {
+        let run_dir = std::env::temp_dir().join(format!(
+            "tests-fuzz-sql-dump-{}",
+            SystemTime::now()
+                .duration_since(UNIX_EPOCH)
+                .unwrap()
+                .as_millis()
+        ));
+
+        let mut session = SqlDumpSession::new_with_buffer_limit(run_dir.clone(), 1024).unwrap();
+        session
+            .append_sql(
+                "metric-a",
+                "INSERT INTO t VALUES(1)",
+                Some("kind=insert elapsed_ms=10"),
+            )
+            .unwrap();
+        session.flush_all().unwrap();
+
+        let content = std::fs::read_to_string(run_dir.join("metric-a.trace.sql")).unwrap();
+        assert!(content.contains("-- kind=insert elapsed_ms=10"));
+        assert!(content.contains("INSERT INTO t VALUES(1);"));
+    }
+
+    #[test]
+    fn test_broadcast_event_writes_to_all_tables() {
+        let run_dir = std::env::temp_dir().join(format!(
+            "tests-fuzz-sql-broadcast-{}",
+            SystemTime::now()
+                .duration_since(UNIX_EPOCH)
+                .unwrap()
+                .as_millis()
+        ));
+
+        let mut session = SqlDumpSession::new_with_buffer_limit(run_dir.clone(), 1024).unwrap();
+        session
+            .broadcast_event(
+                ["metric-a", "metric-b"],
+                "repartition action_idx=3",
+                "ALTER TABLE t REPARTITION",
+            )
+            .unwrap();
+        session.flush_all().unwrap();
+
+        let content_a = std::fs::read_to_string(run_dir.join("metric-a.trace.sql")).unwrap();
+        let content_b = std::fs::read_to_string(run_dir.join("metric-b.trace.sql")).unwrap();
+        assert!(content_a.contains("-- repartition action_idx=3"));
+        assert!(content_a.contains("ALTER TABLE t REPARTITION;"));
+        assert!(content_b.contains("-- repartition action_idx=3"));
+        assert!(content_b.contains("ALTER TABLE t REPARTITION;"));
+    }
+
+    #[test]
+    fn test_multiline_comment_is_prefixed_per_line() {
+        let run_dir = std::env::temp_dir().join(format!(
+            "tests-fuzz-sql-dump-comment-{}",
+            SystemTime::now()
+                .duration_since(UNIX_EPOCH)
+                .unwrap()
+                .as_millis()
+        ));
+
+        let mut session = SqlDumpSession::new_with_buffer_limit(run_dir.clone(), 1024).unwrap();
+        session
+            .append_sql(
+                "metric-a",
+                "INSERT INTO t VALUES(1)",
+                Some("kind=insert\nstarted_at_ms=1 elapsed_ms=2"),
+            )
+            .unwrap();
+        session.flush_all().unwrap();
+
+        let content = std::fs::read_to_string(run_dir.join("metric-a.trace.sql")).unwrap();
+        assert!(content.contains("-- kind=insert\n-- started_at_ms=1 elapsed_ms=2"));
+    }
+
+    #[test]
+    fn test_auto_flush_on_buffer_limit() {
+        let run_dir = std::env::temp_dir().join(format!(
+            "tests-fuzz-sql-dump-limit-{}",
+            SystemTime::now()
+                .duration_since(UNIX_EPOCH)
+                .unwrap()
+                .as_millis()
+        ));
+
+        let mut session = SqlDumpSession::new_with_buffer_limit(run_dir.clone(), 1).unwrap();
+        session
+            .append_sql("metric-a", "INSERT INTO t VALUES(1)", None)
+            .unwrap();
+
+        assert!(run_dir.join("metric-a.trace.sql").exists());
+        assert_eq!(session.buffered_bytes, 0);
+    }
+}
diff --git a/tests-fuzz/targets/ddl/fuzz_repartition_metric_table.rs b/tests-fuzz/targets/ddl/fuzz_repartition_metric_table.rs
new file mode 100644
index 0000000000..7932bc7759
--- /dev/null
+++ b/tests-fuzz/targets/ddl/fuzz_repartition_metric_table.rs
@@ -0,0 +1,684 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#![no_main]
+
+use std::collections::{BTreeMap, HashMap};
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, Instant};
+
+use arbitrary::{Arbitrary, Unstructured};
+use common_telemetry::{info, warn};
+use common_time::Timestamp;
+use common_time::util::current_time_millis;
+use libfuzzer_sys::fuzz_target;
+use rand::{Rng, SeedableRng};
+use rand_chacha::ChaChaRng;
+use snafu::{ResultExt, ensure};
+use sqlx::{MySql, Pool};
+use tests_fuzz::context::{TableContext, TableContextRef};
+use tests_fuzz::error::{self, Result};
+use tests_fuzz::fake::{
+    ConstGenerator, MappedGenerator, WordGenerator, merge_two_word_map_fn, random_capitalize_map,
+    uppercase_and_keyword_backtick_map,
+};
+use tests_fuzz::generator::Generator;
+use tests_fuzz::generator::create_expr::{
+    CreateLogicalTableExprGeneratorBuilder, CreatePhysicalTableExprGeneratorBuilder,
+};
+use tests_fuzz::generator::insert_expr::InsertExprGeneratorBuilder;
+use tests_fuzz::generator::repartition_expr::{
+    MergePartitionExprGeneratorBuilder, SplitPartitionExprGeneratorBuilder,
+};
+use tests_fuzz::ir::{
+    CreateTableExpr, Ident, InsertIntoExpr, RepartitionExpr, generate_random_value,
+    generate_unique_timestamp_for_mysql_with_clock,
+};
+use tests_fuzz::translator::DslTranslator;
+use tests_fuzz::translator::csv::InsertExprToCsvRecordsTranslator;
+use tests_fuzz::translator::mysql::create_expr::CreateTableExprTranslator;
+use tests_fuzz::translator::mysql::insert_expr::InsertIntoExprTranslator;
+use tests_fuzz::translator::mysql::repartition_expr::RepartitionExprTranslator;
+use tests_fuzz::utils::csv_dump_writer::{CsvDumpMetadata, CsvDumpSession};
+use tests_fuzz::utils::retry::retry_with_backoff;
+use tests_fuzz::utils::sql_dump_writer::SqlDumpSession;
+use tests_fuzz::utils::{
+    Connections, get_fuzz_override, get_gt_fuzz_input_max_alter_actions,
+    get_gt_fuzz_input_max_tables, init_greptime_connections_via_env,
+};
+use tests_fuzz::validator::row::count_values;
+use tokio::sync::{mpsc, oneshot};
+
+const BARRIER_ACK_TIMEOUT_SECS: u64 = 10;
+const VALIDATE_QUERY_MAX_ATTEMPTS: usize = 6;
+const VALIDATE_QUERY_INIT_BACKOFF: Duration = Duration::from_millis(50);
+const VALIDATE_QUERY_MAX_BACKOFF: Duration = Duration::from_millis(800);
+
+#[derive(Clone)]
+struct FuzzContext {
+    greptime: Pool<MySql>,
+}
+
+impl FuzzContext {
+    async fn close(self) {
+        self.greptime.close().await;
+    }
+}
+
+#[derive(Clone, Debug)]
+struct FuzzInput {
+    seed: u64,
+    actions: usize,
+    partitions: usize,
+    tables: usize,
+}
+
+fn generate_create_physical_table_expr<R: Rng + 'static>(
+    partitions: usize,
+    rng: &mut R,
+) -> Result<CreateTableExpr> {
+    CreatePhysicalTableExprGeneratorBuilder::default()
+        .name_generator(Box::new(ConstGenerator::new(Ident::new(
+            "fuzz_repartition_metric_physical",
+        ))))
+        .if_not_exists(rng.random_bool(0.5))
+        .partition(partitions)
+        .build()
+        .unwrap()
+        .generate(rng)
+}
+
+fn generate_create_logical_table_expr<R: Rng + 'static>(
+    physical_table_ctx: TableContextRef,
+    include_partition_column: bool,
+    rng: &mut R,
+) -> Result<CreateTableExpr> {
+    CreateLogicalTableExprGeneratorBuilder::default()
+        .name_generator(Box::new(MappedGenerator::new(
+            WordGenerator,
+            merge_two_word_map_fn(random_capitalize_map, uppercase_and_keyword_backtick_map),
+        )))
+        .physical_table_ctx(physical_table_ctx)
+        .labels(rng.random_range(1..=5))
+        .if_not_exists(rng.random_bool(0.5))
+        .include_partition_column(include_partition_column)
+        .build()
+        .unwrap()
+        .generate(rng)
+}
+
+fn generate_insert_expr<R: Rng + 'static>(
+    rows: usize,
+    rng: &mut R,
+    table_ctx: TableContextRef,
+    clock: Arc<Mutex<Timestamp>>,
+) -> Result<InsertIntoExpr> {
+    let ts_value_generator = generate_unique_timestamp_for_mysql_with_clock(clock);
+    InsertExprGeneratorBuilder::default()
+        .omit_column_list(false)
+        .table_ctx(table_ctx)
+        .rows(rows)
+        .value_generator(Box::new(generate_random_value))
+        .ts_value_generator(ts_value_generator)
+        .build()
+        .unwrap()
+        .generate(rng)
+}
+
+async fn create_metric_tables<R: Rng + 'static>(
+    ctx: &FuzzContext,
+    rng: &mut R,
+    partitions: usize,
+    table_count: usize,
+) -> Result<(
+    TableContextRef,
+    BTreeMap<Ident, TableContextRef>,
+    HashMap<String, String>,
+    String,
+)> {
+    let create_physical_expr = generate_create_physical_table_expr(partitions, rng)?;
+    let translator = CreateTableExprTranslator;
+    let create_physical_sql = translator.translate(&create_physical_expr)?;
+    let result = sqlx::query(&create_physical_sql)
+        .execute(&ctx.greptime)
+        .await
+        .context(error::ExecuteQuerySnafu {
+            sql: &create_physical_sql,
+        })?;
+    info!("Create physical table: {create_physical_sql}, result: {result:?}");
+    let physical_table_ctx = Arc::new(TableContext::from(&create_physical_expr));
+    ensure!(
+        physical_table_ctx.partition.is_some(),
+        error::AssertSnafu {
+            reason: "Physical metric table must have partition".to_string()
+        }
+    );
+
+    let mut logical_tables = BTreeMap::new();
+    let mut create_logical_sqls = HashMap::new();
+    let max_attempts = table_count * 3;
+    for _ in 0..max_attempts {
+        if logical_tables.len() >= table_count {
+            break;
+        }
+
+        let include_partition_column = rng.random_bool(0.5);
+        let create_logical_expr = generate_create_logical_table_expr(
+            physical_table_ctx.clone(),
+            include_partition_column,
+            rng,
+        )?;
+        if logical_tables.contains_key(&create_logical_expr.table_name) {
+            continue;
+        }
+
+        let create_logical_sql = translator.translate(&create_logical_expr)?;
+        let result = sqlx::query(&create_logical_sql)
+            .execute(&ctx.greptime)
+            .await
+            .context(error::ExecuteQuerySnafu {
+                sql: &create_logical_sql,
+            })?;
+        info!("Create logical table: {create_logical_sql}, result: {result:?}");
+        let logical_ctx = Arc::new(TableContext::from(&create_logical_expr));
+        create_logical_sqls.insert(logical_ctx.name.to_string(), create_logical_sql);
+        logical_tables.insert(logical_ctx.name.clone(), logical_ctx);
+    }
+
+    ensure!(
+        !logical_tables.is_empty(),
+        error::AssertSnafu {
+            reason: "No logical table created".to_string()
+        }
+    );
+
+    Ok((
+        physical_table_ctx,
+        logical_tables,
+        create_logical_sqls,
+        create_physical_sql,
+    ))
+}
+
+async fn execute_insert_with_retry(ctx: &FuzzContext, sql: &str) -> Result<()> {
+    let mut delay = Duration::from_millis(100);
+    let mut attempt = 0;
+    let max_attempts = 10;
+    loop {
+        match sqlx::query(sql)
+            .persistent(false)
+            .execute(&ctx.greptime)
+            .await
+        {
+            Ok(_) => return Ok(()),
+            Err(err) => {
+                tokio::time::sleep(delay).await;
+                delay = std::cmp::min(delay * 2, Duration::from_secs(1));
+                attempt += 1;
+                warn!("Execute insert with retry: {sql}, attempt: {attempt}, error: {err:?}");
+                if attempt >= max_attempts {
+                    return Err(err).context(error::ExecuteQuerySnafu { sql });
+                }
+            }
+        }
+    }
+}
+
+struct SharedState {
+    clock: Arc<Mutex<Timestamp>>,
+    inserted_rows: HashMap<String, u64>,
+    csv_dump_session: Option<CsvDumpSession>,
+    sql_dump_session: Option<SqlDumpSession>,
+    running: bool,
+}
+
+enum WriterControl {
+    Barrier {
+        epoch: usize,
+        ack: oneshot::Sender<()>,
+    },
+    Resume {
+        epoch: usize,
+    },
+    Stop,
+}
+
+fn handle_writer_control(control: WriterControl, paused: &mut bool) -> bool {
+    match control {
+        WriterControl::Barrier { epoch, ack } => {
+            info!("Writer received barrier control, epoch: {epoch}");
+            *paused = true;
+            let _ = ack.send(());
+            false
+        }
+        WriterControl::Resume { epoch } => {
+            info!("Writer received resume control, epoch: {epoch}");
+            *paused = false;
+            false
+        }
+        WriterControl::Stop => {
+            info!("Writer received stop control");
+            true
+        }
+    }
+}
+
+async fn write_loop<R: Rng + 'static>(
+    mut rng: R,
+    ctx: FuzzContext,
+    logical_tables: BTreeMap<Ident, TableContextRef>,
+    shared_state: Arc<Mutex<SharedState>>,
+    mut control_rx: mpsc::UnboundedReceiver<WriterControl>,
+) -> Result<()> {
+    info!("Start write loop");
+    let mut paused = false;
+    loop {
+        while let Ok(control) = control_rx.try_recv() {
+            if handle_writer_control(control, &mut paused) {
+                return Ok(());
+            }
+        }
+
+        if paused {
+            match control_rx.recv().await {
+                Some(control) => {
+                    if handle_writer_control(control, &mut paused) {
+                        return Ok(());
+                    }
+                }
+                None => return Ok(()),
+            }
+            continue;
+        }
+
+        let (running, clock) = {
+            let state = shared_state.lock().unwrap();
+            (state.running, state.clock.clone())
+        };
+        if !running {
+            break;
+        }
+
+        for table_ctx in logical_tables.values() {
+            let rows = rng.random_range(1..=3);
+            let insert_expr =
+                generate_insert_expr(rows, &mut rng, table_ctx.clone(), clock.clone())?;
+            let translator = InsertIntoExprTranslator;
+            let sql = translator.translate(&insert_expr)?;
+            let inserted = insert_expr.values_list.len() as u64;
+            let csv_records = InsertExprToCsvRecordsTranslator.translate(&insert_expr)?;
+            let table_name = table_ctx.name.to_string();
+            let full_headers = table_ctx
+                .columns
+                .iter()
+                .map(|column| column.name.value.clone())
+                .collect::<Vec<_>>();
+
+            let started_at_ms = current_time_millis();
+            let now = Instant::now();
+            execute_insert_with_retry(&ctx, &sql).await?;
+            let elapsed = now.elapsed();
+            info!("Execute insert sql: {sql}, elapsed: {elapsed:?}");
+
+            let mut state = shared_state.lock().unwrap();
+            if let Some(csv_dump_session) = state.csv_dump_session.as_mut() {
+                csv_dump_session.append(csv_records, full_headers)?;
+            }
+            if let Some(sql_dump_session) = state.sql_dump_session.as_mut() {
+                let comment = format!(
+                    "kind=insert table={} started_at_ms={} elapsed_ms={}",
+                    table_name,
+                    started_at_ms,
+                    elapsed.as_millis()
+                );
+                sql_dump_session.append_sql(&table_name, &sql, Some(&comment))?;
+            }
+            *state.inserted_rows.entry(table_name).or_insert(0) += inserted;
+        }
+
+        tokio::time::sleep(Duration::from_millis(100)).await;
+    }
+    info!("Write loop ended");
+
+    Ok(())
+}
+
+async fn validate_rows(
+    ctx: &FuzzContext,
+    logical_tables: &BTreeMap<Ident, TableContextRef>,
+    inserted_rows: &HashMap<String, u64>,
+) -> Result<()> {
+    for table_ctx in logical_tables.values() {
+        let expected = *inserted_rows.get(&table_ctx.name.to_string()).unwrap_or(&0) as usize;
+        let count_sql = format!("SELECT COUNT(1) AS count FROM {}", table_ctx.name);
+        let count = retry_with_backoff(
+            || count_values(&ctx.greptime, &count_sql),
+            VALIDATE_QUERY_MAX_ATTEMPTS,
+            VALIDATE_QUERY_INIT_BACKOFF,
+            VALIDATE_QUERY_MAX_BACKOFF,
+        )
+        .await?;
+        let distinct_count_sql = format!(
+            "SELECT COUNT(DISTINCT {}) AS count FROM {}",
+            table_ctx.timestamp_column().unwrap().name,
+            table_ctx.name
+        );
+        let distinct_count = retry_with_backoff(
+            || count_values(&ctx.greptime, &distinct_count_sql),
+            VALIDATE_QUERY_MAX_ATTEMPTS,
+            VALIDATE_QUERY_INIT_BACKOFF,
+            VALIDATE_QUERY_MAX_BACKOFF,
+        )
+        .await?;
+        info!(
+            "Validate rows for table: {}, expected: {}, count: {}, distinct_count: {}",
+            table_ctx.name, expected, count.count as usize, distinct_count.count as usize
+        );
+        assert_eq!(count.count as usize, expected);
+
+        assert_eq!(distinct_count.count as usize, expected);
+    }
+    Ok(())
+}
+
+fn flush_dump_sessions_and_snapshot(
+    shared_state: &Arc<Mutex<SharedState>>,
+) -> Result<HashMap<String, u64>> {
+    let mut state = shared_state.lock().unwrap();
+    if let Some(csv_dump_session) = state.csv_dump_session.as_mut() {
+        csv_dump_session.flush_all()?;
+    }
+    if let Some(sql_dump_session) = state.sql_dump_session.as_mut() {
+        sql_dump_session.flush_all()?;
+    }
+    Ok(state.inserted_rows.clone())
+}
+
+async fn cleanup_tables(
+    ctx: &FuzzContext,
+    physical_table_ctx: &TableContextRef,
+    logical_tables: &BTreeMap<Ident, TableContextRef>,
+) -> Result<()> {
+    for table_ctx in logical_tables.values() {
+        let drop_logical_sql = format!("DROP TABLE {}", table_ctx.name);
+        let result = sqlx::query(&drop_logical_sql)
+            .execute(&ctx.greptime)
+            .await
+            .context(error::ExecuteQuerySnafu {
+                sql: &drop_logical_sql,
+            })?;
+        info!("Drop logical table: {drop_logical_sql}, result: {result:?}");
+    }
+
+    let drop_physical_sql = format!("DROP TABLE {}", physical_table_ctx.name);
+    let result = sqlx::query(&drop_physical_sql)
+        .execute(&ctx.greptime)
+        .await
+        .context(error::ExecuteQuerySnafu {
+            sql: &drop_physical_sql,
+        })?;
+    info!("Drop physical table: {drop_physical_sql}, result: {result:?}");
+    Ok(())
+}
+
+fn repartition_operation<R: Rng + 'static>(
+    table_ctx: &TableContextRef,
+    rng: &mut R,
+) -> Result<RepartitionExpr> {
+    let split = rng.random_bool(0.5);
+    if table_ctx.partition.as_ref().unwrap().exprs.len() <= 2 || split {
+        let expr = SplitPartitionExprGeneratorBuilder::default()
+            .table_ctx(table_ctx.clone())
+            .build()
+            .unwrap()
+            .generate(rng)?;
+        Ok(RepartitionExpr::Split(expr))
+    } else {
+        let expr = MergePartitionExprGeneratorBuilder::default()
+            .table_ctx(table_ctx.clone())
+            .build()
+            .unwrap()
+            .generate(rng)?;
+        Ok(RepartitionExpr::Merge(expr))
+    }
+}
+
+impl Arbitrary<'_> for FuzzInput {
+    fn arbitrary(u: &mut Unstructured<'_>) -> arbitrary::Result<Self> {
+        let seed = get_fuzz_override::<u64>("SEED").unwrap_or(u.int_in_range(u64::MIN..=u64::MAX)?);
+        let mut rng = ChaChaRng::seed_from_u64(seed);
+        let partitions =
+            get_fuzz_override::<usize>("PARTITIONS").unwrap_or_else(|| rng.random_range(2..8));
+        let max_tables = get_gt_fuzz_input_max_tables();
+        let tables = get_fuzz_override::<usize>("TABLES")
+            .unwrap_or_else(|| rng.random_range(1..=std::cmp::max(1, max_tables)));
+        let max_actions = get_gt_fuzz_input_max_alter_actions();
+        let actions = get_fuzz_override::<usize>("ACTIONS")
+            .unwrap_or_else(|| rng.random_range(1..max_actions));
+
+        Ok(FuzzInput {
+            seed,
+            actions,
+            partitions,
+            tables,
+        })
+    }
+}
+
+async fn execute_repartition_metric_table(ctx: FuzzContext, input: FuzzInput) -> Result<()> {
+    info!("input: {input:?}");
+    let mut rng = ChaChaRng::seed_from_u64(input.seed);
+    let clock = Arc::new(Mutex::new(Timestamp::current_millis()));
+
+    let (mut physical_table_ctx, logical_tables, create_logical_sqls, create_physical_sql) =
+        create_metric_tables(&ctx, &mut rng, input.partitions, input.tables).await?;
+
+    let mut inserted_rows = HashMap::with_capacity(logical_tables.len());
+    for table_ctx in logical_tables.values() {
+        inserted_rows.insert(table_ctx.name.to_string(), 0);
+    }
+    let csv_dump_session = CsvDumpSession::new(CsvDumpMetadata::new(
+        "fuzz_repartition_metric_table",
+        input.seed,
+        input.actions,
+        input.partitions,
+        input.tables,
+    ))?;
+    let sql_dump_session = SqlDumpSession::new(csv_dump_session.run_dir.clone())?;
+    let logical_table_names = logical_tables
+        .values()
+        .map(|table_ctx| table_ctx.name.to_string())
+        .collect::<Vec<_>>();
+
+    let mut sql_dump_session = sql_dump_session;
+    sql_dump_session.append_sql(
+        &physical_table_ctx.name.to_string(),
+        &create_physical_sql,
+        Some("kind=create_physical_table"),
+    )?;
+    for table_name in &logical_table_names {
+        if let Some(create_sql) = create_logical_sqls.get(table_name) {
+            sql_dump_session.append_sql(
+                table_name,
+                create_sql,
+                Some("kind=create_logical_table"),
+            )?;
+        }
+    }
+
+    let shared_state = Arc::new(Mutex::new(SharedState {
+        clock,
+        inserted_rows,
+        csv_dump_session: Some(csv_dump_session),
+        sql_dump_session: Some(sql_dump_session),
+        running: true,
+    }));
+    let writer_rng = ChaChaRng::seed_from_u64(input.seed ^ 0xA5A5_A5A5_A5A5_A5A5);
+    let (control_tx, control_rx) = mpsc::unbounded_channel::<WriterControl>();
+    let writer_task = tokio::spawn(write_loop(
+        writer_rng,
+        ctx.clone(),
+        logical_tables.clone(),
+        shared_state.clone(),
+        control_rx,
+    ));
+    tokio::time::sleep(Duration::from_millis(100)).await;
+
+    for i in 0..input.actions {
+        let partition_num = physical_table_ctx.partition.as_ref().unwrap().exprs.len();
+        info!(
+            "partition_num: {partition_num}, action: {}/{}, table: {}, logical table num: {}",
+            i + 1,
+            input.actions,
+            physical_table_ctx.name,
+            logical_tables.len()
+        );
+
+        let repartition_expr = repartition_operation(&physical_table_ctx, &mut rng)?;
+        let translator = RepartitionExprTranslator;
+        let sql = translator.translate(&repartition_expr)?;
+        info!("Repartition sql: {sql}");
+        let started_at_ms = current_time_millis();
+        let now = Instant::now();
+        let result = sqlx::query(&sql)
+            .execute(&ctx.greptime)
+            .await
+            .context(error::ExecuteQuerySnafu { sql: &sql })?;
+        let elapsed = now.elapsed();
+        info!("Repartition result: {result:?}, elapsed: {elapsed:?}");
+
+        physical_table_ctx = Arc::new(
+            Arc::unwrap_or_clone(physical_table_ctx)
+                .repartition(repartition_expr)
+                .unwrap(),
+        );
+
+        let partition_entries = tests_fuzz::validator::partition::fetch_partitions_info_schema(
+            &ctx.greptime,
+            "public".into(),
+            &physical_table_ctx.name,
+        )
+        .await?;
+        tests_fuzz::validator::partition::assert_partitions(
+            physical_table_ctx.partition.as_ref().unwrap(),
+            &partition_entries,
+        )?;
+
+        {
+            let mut state = shared_state.lock().unwrap();
+            if let Some(sql_dump_session) = state.sql_dump_session.as_mut() {
+                let repartition_comment = format!(
+                    "kind=repartition table={} action_idx={} started_at_ms={} elapsed_ms={}",
+                    physical_table_ctx.name,
+                    i + 1,
+                    started_at_ms,
+                    elapsed.as_millis()
+                );
+                sql_dump_session.append_sql(
+                    &physical_table_ctx.name.to_string(),
+                    &sql,
+                    Some(&repartition_comment),
+                )?;
+                let event = format!(
+                    "repartition action_idx={} started_at_ms={} elapsed_ms={} sql={}",
+                    i + 1,
+                    started_at_ms,
+                    elapsed.as_millis(),
+                    sql
+                );
+                sql_dump_session.broadcast_event(logical_table_names.iter(), &event, &sql)?;
+            }
+        }
+
+        let (ack_tx, ack_rx) = oneshot::channel();
+        control_tx
+            .send(WriterControl::Barrier {
+                epoch: i + 1,
+                ack: ack_tx,
+            })
+            .expect("barrier control send must succeed");
+        tokio::time::timeout(Duration::from_secs(BARRIER_ACK_TIMEOUT_SECS), ack_rx)
+            .await
+            .expect("barrier ack timeout")
+            .expect("barrier ack dropped");
+
+        let inserted_rows_snapshot = flush_dump_sessions_and_snapshot(&shared_state)?;
+        info!("validate rows, epoch: {}", i + 1);
+        validate_rows(&ctx, &logical_tables, &inserted_rows_snapshot).await?;
+
+        control_tx
+            .send(WriterControl::Resume { epoch: i + 1 })
+            .expect("resume control send must succeed");
+    }
+
+    let _ = control_tx.send(WriterControl::Stop);
+    shared_state.lock().unwrap().running = false;
+    writer_task.await.unwrap().unwrap();
+    let inserted_rows = flush_dump_sessions_and_snapshot(&shared_state)?;
+    let (mut csv_dump_session, mut sql_dump_session) = {
+        let mut state = shared_state.lock().unwrap();
+        (state.csv_dump_session.take(), state.sql_dump_session.take())
+    };
+
+    let run_result = async {
+        validate_rows(&ctx, &logical_tables, &inserted_rows).await?;
+        cleanup_tables(&ctx, &physical_table_ctx, &logical_tables).await?;
+        Ok(())
+    }
+    .await;
+
+    if let Some(csv_dump_session) = csv_dump_session.take() {
+        match &run_result {
+            Ok(_) => {
+                if let Err(err) = csv_dump_session.cleanup_on_success() {
+                    warn!(
+                        "Cleanup csv dump directory failed, path: {}, error: {:?}",
+                        csv_dump_session.run_dir.display(),
+                        err
+                    );
+                }
+            }
+            Err(_) => {
+                warn!(
+                    "Keep csv dump directory for failure analysis, path: {}",
+                    csv_dump_session.run_dir.display()
+                );
+            }
+        }
+    }
+    if let Some(sql_dump_session) = sql_dump_session.take()
+        && run_result.is_err()
+    {
+        warn!(
+            "Keep sql dump directory for failure analysis, path: {}",
+            sql_dump_session.run_dir.display()
+        );
+    }
+
+    ctx.close().await;
+    run_result
+}
+
+fuzz_target!(|input: FuzzInput| {
+    common_telemetry::init_default_ut_logging();
+    common_runtime::block_on_global(async {
+        let Connections { mysql } = init_greptime_connections_via_env().await;
+        let ctx = FuzzContext {
+            greptime: mysql.expect("mysql connection init must be succeed"),
+        };
+        execute_repartition_metric_table(ctx, input)
+            .await
+            .unwrap_or_else(|err| panic!("fuzz test must be succeed: {err:?}"));
+    })
+});
diff --git a/tests-fuzz/targets/ddl/fuzz_repartition_table_chaos.rs b/tests-fuzz/targets/ddl/fuzz_repartition_table_chaos.rs
new file mode 100644
index 0000000000..d3789b696c
--- /dev/null
+++ b/tests-fuzz/targets/ddl/fuzz_repartition_table_chaos.rs
@@ -0,0 +1,447 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#![no_main]
+
+use std::env;
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::sync::{Arc, Mutex};
+use std::time::Duration;
+
+use arbitrary::{Arbitrary, Unstructured};
+use common_telemetry::info;
+use common_time::Timestamp;
+use libfuzzer_sys::fuzz_target;
+use rand::{Rng, SeedableRng};
+use rand_chacha::ChaChaRng;
+use snafu::ResultExt;
+use sqlx::{Executor, MySql, Pool};
+use tests_fuzz::context::{TableContext, TableContextRef};
+use tests_fuzz::error::{self, Result};
+use tests_fuzz::fake::{
+    MappedGenerator, WordGenerator, merge_two_word_map_fn, random_capitalize_map,
+    uppercase_and_keyword_backtick_map,
+};
+use tests_fuzz::generator::Generator;
+use tests_fuzz::generator::create_expr::CreateTableExprGeneratorBuilder;
+use tests_fuzz::generator::insert_expr::InsertExprGeneratorBuilder;
+use tests_fuzz::generator::repartition_expr::{
+    MergePartitionExprGeneratorBuilder, SplitPartitionExprGeneratorBuilder,
+};
+use tests_fuzz::ir::{
+    CreateTableExpr, InsertIntoExpr, MySQLTsColumnTypeGenerator, RepartitionExpr, RowValue,
+    SimplePartitions, generate_partition_value, generate_unique_timestamp_for_mysql_with_clock,
+};
+use tests_fuzz::translator::DslTranslator;
+use tests_fuzz::translator::mysql::create_expr::CreateTableExprTranslator;
+use tests_fuzz::translator::mysql::insert_expr::InsertIntoExprTranslator;
+use tests_fuzz::translator::mysql::repartition_expr::RepartitionExprTranslator;
+use tests_fuzz::utils::cluster_info::{
+    wait_for_all_datanode_offline, wait_for_all_datanode_online,
+};
+use tests_fuzz::utils::network_chaos::{
+    inject_datanode_metasrv_network_partition, recover_network_chaos,
+};
+use tests_fuzz::utils::procedure::procedure_state as fetch_procedure_state_json;
+use tests_fuzz::utils::{
+    Connections, GT_FUZZ_CLUSTER_NAME, GT_FUZZ_CLUSTER_NAMESPACE, get_fuzz_override,
+    get_gt_fuzz_input_max_rows, init_greptime_connections_via_env,
+};
+use tests_fuzz::validator;
+use tests_fuzz::validator::row::count_values;
+
+#[derive(Clone)]
+struct FuzzContext {
+    greptime: Pool<MySql>,
+    kube: kube::client::Client,
+    namespace: String,
+    cluster_name: String,
+}
+
+const PROCEDURE_TIMEOUT: Duration = Duration::from_secs(300);
+const NETWORK_CHAOS_DURATION_SECS: usize = 360;
+
+impl FuzzContext {
+    async fn close(self) {
+        self.greptime.close().await;
+    }
+}
+
+#[derive(Clone, Debug)]
+struct FuzzInput {
+    seed: u64,
+    rows: usize,
+    partitions: usize,
+    chaos_delay_ms: u64,
+    chaos_hold_secs: u64,
+}
+
+impl Arbitrary<'_> for FuzzInput {
+    fn arbitrary(u: &mut Unstructured<'_>) -> arbitrary::Result<Self> {
+        let seed = get_fuzz_override::<u64>("SEED").unwrap_or(u.int_in_range(u64::MIN..=u64::MAX)?);
+        let mut rng = ChaChaRng::seed_from_u64(seed);
+        let rows = get_fuzz_override::<usize>("ROWS")
+            .unwrap_or_else(|| rng.random_range(2..get_gt_fuzz_input_max_rows()));
+        let partitions =
+            get_fuzz_override::<usize>("PARTITIONS").unwrap_or_else(|| rng.random_range(2..8));
+        let chaos_delay_ms =
+            get_fuzz_override::<u64>("CHAOS_DELAY_MS").unwrap_or_else(|| rng.random_range(0..5000));
+        let chaos_hold_secs =
+            get_fuzz_override::<u64>("CHAOS_HOLD_SECS").unwrap_or_else(|| rng.random_range(10..20));
+
+        Ok(FuzzInput {
+            seed,
+            rows,
+            partitions,
+            chaos_delay_ms,
+            chaos_hold_secs,
+        })
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum ProcedureTerminalState {
+    Done,
+    Failed,
+}
+
+fn generate_create_expr<R: Rng + 'static>(
+    input: &FuzzInput,
+    rng: &mut R,
+) -> Result<CreateTableExpr> {
+    let create_table_generator = CreateTableExprGeneratorBuilder::default()
+        .name_generator(Box::new(MappedGenerator::new(
+            WordGenerator,
+            merge_two_word_map_fn(random_capitalize_map, uppercase_and_keyword_backtick_map),
+        )))
+        .columns(5)
+        .partition(input.partitions)
+        .engine("mito")
+        .ts_column_type_generator(Box::new(MySQLTsColumnTypeGenerator))
+        .build()
+        .unwrap();
+    create_table_generator.generate(rng)
+}
+
+fn build_insert_expr<R: Rng + 'static>(
+    table_ctx: &TableContextRef,
+    rng: &mut R,
+    partitions: &SimplePartitions,
+    clock: &Arc<Mutex<Timestamp>>,
+    rows: usize,
+) -> InsertIntoExpr {
+    let ts_value_generator = generate_unique_timestamp_for_mysql_with_clock(clock.clone());
+    let counter = Arc::new(AtomicUsize::new(0));
+    let counter_clone = counter.clone();
+    let partition_len = table_ctx.partition.as_ref().unwrap().exprs.len();
+    let moved_partitions = partitions.clone();
+    let insert_generator = InsertExprGeneratorBuilder::default()
+        .table_ctx(table_ctx.clone())
+        .omit_column_list(false)
+        .rows(rows.max(partition_len))
+        .required_columns(vec![partitions.column_name.clone()])
+        .value_overrides(Some(Box::new(move |column, rng| {
+            if column.name.value == moved_partitions.column_name.value {
+                let bound_idx = counter_clone.fetch_add(1, Ordering::Relaxed) % partition_len;
+                return Some(RowValue::Value(generate_partition_value(
+                    rng,
+                    &column.column_type,
+                    &moved_partitions.bounds,
+                    bound_idx,
+                )));
+            }
+            None
+        })))
+        .ts_value_generator(ts_value_generator)
+        .build()
+        .unwrap();
+    insert_generator.generate(rng).unwrap()
+}
+
+async fn execute_insert_with_retry(ctx: &FuzzContext, sql: &str) -> Result<()> {
+    let mut delay = Duration::from_millis(100);
+    let mut attempt = 0;
+    let max_attempts = 10;
+    loop {
+        match ctx.greptime.execute(sql).await {
+            Ok(_) => return Ok(()),
+            Err(err) => {
+                tokio::time::sleep(delay).await;
+                delay = std::cmp::min(delay * 2, Duration::from_secs(1));
+                attempt += 1;
+                if attempt >= max_attempts {
+                    return Err(err).context(error::ExecuteQuerySnafu { sql });
+                }
+            }
+        }
+    }
+}
+
+async fn create_table(ctx: &FuzzContext, expr: &CreateTableExpr) -> Result<TableContextRef> {
+    let translator = CreateTableExprTranslator;
+    let sql = translator.translate(expr)?;
+    let result = sqlx::query(&sql)
+        .execute(&ctx.greptime)
+        .await
+        .context(error::ExecuteQuerySnafu { sql: &sql })?;
+    info!("Create table: {sql}, result: {result:?}");
+    Ok(Arc::new(TableContext::from(expr)))
+}
+
+async fn insert_initial_rows<R: Rng + 'static>(
+    ctx: &FuzzContext,
+    table_ctx: &TableContextRef,
+    rng: &mut R,
+    rows: usize,
+) -> Result<u64> {
+    let partitions = SimplePartitions::from_table_ctx(table_ctx).unwrap();
+    let clock = Arc::new(Mutex::new(Timestamp::current_millis()));
+    let insert_expr = build_insert_expr(table_ctx, rng, &partitions, &clock, rows);
+    let inserted_rows = insert_expr.values_list.len() as u64;
+    let translator = InsertIntoExprTranslator;
+    let sql = translator.translate(&insert_expr)?;
+    execute_insert_with_retry(ctx, &sql).await?;
+    info!("Insert initial rows: {sql}");
+    Ok(inserted_rows)
+}
+
+async fn validate_table_rows(
+    ctx: &FuzzContext,
+    table_ctx: &TableContextRef,
+    inserted_rows: u64,
+) -> Result<()> {
+    let count_sql = format!("SELECT COUNT(1) AS count FROM {}", table_ctx.name);
+    let count = count_values(&ctx.greptime, &count_sql).await?;
+    assert_eq!(count.count as u64, inserted_rows);
+
+    let timestamp_column_name = table_ctx.timestamp_column().unwrap().name.clone();
+    let distinct_count_sql = format!(
+        "SELECT COUNT(DISTINCT {}) AS count FROM {}",
+        timestamp_column_name, table_ctx.name
+    );
+    let distinct_count = count_values(&ctx.greptime, &distinct_count_sql).await?;
+    assert_eq!(distinct_count.count as u64, inserted_rows);
+    Ok(())
+}
+
+fn repartition_operation<R: Rng + 'static>(
+    table_ctx: &TableContextRef,
+    rng: &mut R,
+    wait: bool,
+) -> Result<RepartitionExpr> {
+    let split = rng.random_bool(0.5);
+    if table_ctx.partition.as_ref().unwrap().exprs.len() <= 2 || split {
+        let expr = SplitPartitionExprGeneratorBuilder::default()
+            .table_ctx(table_ctx.clone())
+            .wait(wait)
+            .build()
+            .unwrap()
+            .generate(rng)?;
+        Ok(RepartitionExpr::Split(expr))
+    } else {
+        let expr = MergePartitionExprGeneratorBuilder::default()
+            .table_ctx(table_ctx.clone())
+            .wait(wait)
+            .build()
+            .unwrap()
+            .generate(rng)?;
+        Ok(RepartitionExpr::Merge(expr))
+    }
+}
+
+async fn submit_repartition_procedure(ctx: &FuzzContext, expr: &RepartitionExpr) -> Result<String> {
+    let translator = RepartitionExprTranslator;
+    let sql = translator.translate(expr)?;
+    let (procedure_id,) = sqlx::query_as::<_, (String,)>(&sql)
+        .fetch_one(&ctx.greptime)
+        .await
+        .context(error::ExecuteQuerySnafu { sql: &sql })?;
+    info!("Submit repartition procedure: {sql}, procedure_id: {procedure_id}");
+    Ok(procedure_id)
+}
+
+async fn fetch_procedure_state(
+    ctx: &FuzzContext,
+    procedure_id: &str,
+) -> Result<Option<ProcedureTerminalState>> {
+    let output = fetch_procedure_state_json(&ctx.greptime, procedure_id).await;
+
+    Ok(if output.contains("Done") {
+        Some(ProcedureTerminalState::Done)
+    } else if output.contains("Failed") {
+        Some(ProcedureTerminalState::Failed)
+    } else {
+        None
+    })
+}
+
+async fn wait_for_procedure_terminal_state(
+    ctx: &FuzzContext,
+    procedure_id: &str,
+    timeout: Duration,
+) -> Result<ProcedureTerminalState> {
+    let deadline = tokio::time::Instant::now() + timeout;
+    loop {
+        if let Some(state) = fetch_procedure_state(ctx, procedure_id).await? {
+            info!("Procedure terminal state: {state:?}, procedure_id: {procedure_id}");
+            return Ok(state);
+        }
+
+        if tokio::time::Instant::now() >= deadline {
+            return error::AssertSnafu {
+                reason: format!(
+                    "procedure {procedure_id} did not reach terminal state before timeout"
+                ),
+            }
+            .fail();
+        }
+
+        tokio::time::sleep(Duration::from_secs(1)).await;
+    }
+}
+
+async fn validate_terminal_metadata(
+    ctx: &FuzzContext,
+    table_name: &tests_fuzz::ir::Ident,
+    terminal: ProcedureTerminalState,
+    before_table_ctx: &TableContextRef,
+    after_table_ctx: &TableContextRef,
+) -> Result<()> {
+    let partition_entries = validator::partition::fetch_partitions_info_schema(
+        &ctx.greptime,
+        "public".into(),
+        table_name,
+    )
+    .await?;
+
+    info!(
+        "Validate terminal metadata: procedure_terminal_state={terminal:?}, partition_entries={partition_entries:?}"
+    );
+
+    match terminal {
+        ProcedureTerminalState::Done => validator::partition::assert_partitions(
+            after_table_ctx.partition.as_ref().unwrap(),
+            &partition_entries,
+        )?,
+        ProcedureTerminalState::Failed => validator::partition::assert_partitions(
+            before_table_ctx.partition.as_ref().unwrap(),
+            &partition_entries,
+        )?,
+    }
+
+    Ok(())
+}
+
+async fn drop_table(ctx: &FuzzContext, table_name: &tests_fuzz::ir::Ident) -> Result<()> {
+    let sql = format!("DROP TABLE {}", table_name);
+    let result = sqlx::query(&sql)
+        .execute(&ctx.greptime)
+        .await
+        .context(error::ExecuteQuerySnafu { sql: &sql })?;
+    info!("Drop table: {table_name}, result: {result:?}");
+    Ok(())
+}
+
+async fn execute_repartition_chaos(ctx: FuzzContext, input: FuzzInput) -> Result<()> {
+    info!("input: {input:?}");
+    let mut rng = ChaChaRng::seed_from_u64(input.seed);
+
+    let create_expr = generate_create_expr(&input, &mut rng)?;
+    let before_table_ctx = create_table(&ctx, &create_expr).await?;
+    let inserted_rows = insert_initial_rows(&ctx, &before_table_ctx, &mut rng, input.rows).await?;
+    validate_table_rows(&ctx, &before_table_ctx, inserted_rows).await?;
+
+    let before_entries = validator::partition::fetch_partitions_info_schema(
+        &ctx.greptime,
+        "public".into(),
+        &before_table_ctx.name,
+    )
+    .await?;
+    info!("Before repartition partition entries: {before_entries:?}");
+
+    let repartition_expr = repartition_operation(&before_table_ctx, &mut rng, false)?;
+    let after_table_ctx = Arc::new(
+        Arc::unwrap_or_clone(before_table_ctx.clone())
+            .repartition(repartition_expr.clone())
+            .unwrap(),
+    );
+
+    let namespace = ctx.namespace.clone();
+    let cluster_name = ctx.cluster_name.clone();
+    let kube_client = ctx.kube.clone();
+    let handle = tokio::spawn(async move {
+        let chaos_name = inject_datanode_metasrv_network_partition(
+            kube_client.clone(),
+            &namespace,
+            &cluster_name,
+            NETWORK_CHAOS_DURATION_SECS,
+        )
+        .await
+        .unwrap();
+        info!(
+            "Injected network chaos: {chaos_name}, holding for {} seconds",
+            input.chaos_hold_secs
+        );
+
+        tokio::time::sleep(Duration::from_secs(input.chaos_hold_secs)).await;
+        recover_network_chaos(kube_client.clone(), &namespace, &chaos_name)
+            .await
+            .unwrap();
+    });
+
+    wait_for_all_datanode_offline(ctx.greptime.clone(), Duration::from_secs(30)).await;
+    info!(
+        "Waiting for {} ms before submitting procedure",
+        input.chaos_delay_ms
+    );
+    tokio::time::sleep(Duration::from_millis(input.chaos_delay_ms)).await;
+
+    let procedure_id = submit_repartition_procedure(&ctx, &repartition_expr).await?;
+    let terminal =
+        wait_for_procedure_terminal_state(&ctx, &procedure_id, PROCEDURE_TIMEOUT).await?;
+    handle.await.unwrap();
+
+    wait_for_all_datanode_online(ctx.greptime.clone(), Duration::from_secs(30)).await;
+    validate_terminal_metadata(
+        &ctx,
+        &before_table_ctx.name,
+        terminal,
+        &before_table_ctx,
+        &after_table_ctx,
+    )
+    .await?;
+    validate_table_rows(&ctx, &before_table_ctx, inserted_rows).await?;
+
+    drop_table(&ctx, &before_table_ctx.name).await?;
+    ctx.close().await;
+    Ok(())
+}
+
+fuzz_target!(|input: FuzzInput| {
+    common_telemetry::init_default_ut_logging();
+    common_runtime::block_on_global(async {
+        let Connections { mysql } = init_greptime_connections_via_env().await;
+        let ctx = FuzzContext {
+            greptime: mysql.expect("mysql connection init must be succeed"),
+            kube: kube::client::Client::try_default()
+                .await
+                .expect("init kube client"),
+            namespace: env::var(GT_FUZZ_CLUSTER_NAMESPACE).unwrap_or("my-greptimedb".to_string()),
+            cluster_name: env::var(GT_FUZZ_CLUSTER_NAME).unwrap_or("my-greptimedb".to_string()),
+        };
+        execute_repartition_chaos(ctx, input)
+            .await
+            .unwrap_or_else(|err| panic!("fuzz test must be succeed: {err:?}"));
+    })
+});
diff --git a/tests-fuzz/targets/migration/fuzz_migrate_mito_regions.rs b/tests-fuzz/targets/migration/fuzz_migrate_mito_regions.rs
index c8ebbb54af..17cbfb9251 100644
--- a/tests-fuzz/targets/migration/fuzz_migrate_mito_regions.rs
+++ b/tests-fuzz/targets/migration/fuzz_migrate_mito_regions.rs
@@ -261,13 +261,18 @@ async fn migrate_regions(ctx: &FuzzContext, migrations: &[Migration]) -> Result<
                     {
                         let output = procedure_state(&greptime, &procedure_id).await;
                         info!("Checking procedure: {procedure_id}, output: {output}");
-                        (fetch_partition(&greptime, region_id).await.unwrap(), output)
+                        (fetch_partition(&greptime, region_id).await.ok(), output)
                     }
                 })
             },
             |(partition, output)| {
-                info!("Region: {region_id},  datanode: {}", partition.datanode_id);
-                partition.datanode_id == migration.to_peer && output.contains("Done")
+                if let Some(partition) = partition {
+                    info!("Region: {region_id},  datanode: {}", partition.datanode_id);
+                    partition.datanode_id == migration.to_peer && output.contains("Done")
+                } else {
+                    info!("Region: {region_id}, partition not found yet");
+                    false
+                }
             },
             Duration::from_secs(5),
         )
diff --git a/tests-integration/Cargo.toml b/tests-integration/Cargo.toml
index 0c6b965fd3..43850e4ed3 100644
--- a/tests-integration/Cargo.toml
+++ b/tests-integration/Cargo.toml
@@ -3,9 +3,14 @@ name = "tests-integration"
 version.workspace = true
 edition.workspace = true
 license.workspace = true
+autotests = false
+
+[[test]]
+name = "main"
+path = "tests/main.rs"
 
 [features]
-dashboard = []
+dashboard = ["servers/dashboard"]
 vector_index = []
 
 [lints]
@@ -18,6 +23,7 @@ async-stream.workspace = true
 async-trait.workspace = true
 auth.workspace = true
 axum.workspace = true
+base64.workspace = true
 cache.workspace = true
 catalog.workspace = true
 chrono.workspace = true
@@ -33,6 +39,7 @@ common-frontend.workspace = true
 common-grpc.workspace = true
 common-memory-manager.workspace = true
 common-meta = { workspace = true, features = ["testing"] }
+common-options.workspace = true
 common-procedure.workspace = true
 common-query.workspace = true
 common-recordbatch.workspace = true
@@ -57,23 +64,15 @@ loki-proto.workspace = true
 meta-client.workspace = true
 meta-srv = { workspace = true, features = ["mock"] }
 mito2.workspace = true
-moka.workspace = true
-mysql_async = { version = "0.35", default-features = false, features = [
-    "time",
-    "default-rustls-ring",
-] }
 object-store.workspace = true
 operator = { workspace = true, features = ["testing"] }
+plugins.workspace = true
 prost.workspace = true
 query.workspace = true
 rand.workspace = true
-rstest.workspace = true
-rstest_reuse.workspace = true
-sea-query.workspace = true
 serde_json.workspace = true
 servers = { workspace = true, features = ["testing"] }
 session.workspace = true
-similar-asserts.workspace = true
 snafu.workspace = true
 sql.workspace = true
 sqlx = { workspace = true, features = [
@@ -99,10 +98,15 @@ common-procedure-test.workspace = true
 common-wal = { workspace = true, features = ["testing"] }
 datafusion.workspace = true
 datafusion-expr.workspace = true
+either.workspace = true
 hex.workspace = true
 http.workspace = true
 itertools.workspace = true
 jsonb.workspace = true
+mysql_async = { version = "0.35", default-features = false, features = [
+    "time",
+    "default-rustls-ring",
+] }
 opentelemetry-proto.workspace = true
 otel-arrow-rust.workspace = true
 partition.workspace = true
@@ -110,7 +114,11 @@ paste.workspace = true
 pipeline.workspace = true
 prost.workspace = true
 rand.workspace = true
+rstest.workspace = true
+rstest_reuse.workspace = true
+sea-query.workspace = true
 session = { workspace = true, features = ["testing"] }
+similar-asserts.workspace = true
 store-api.workspace = true
 tokio-postgres = { workspace = true }
 url = "2.3"
diff --git a/tests-integration/src/cluster.rs b/tests-integration/src/cluster.rs
index daa2f64919..0719fcd9f1 100644
--- a/tests-integration/src/cluster.rs
+++ b/tests-integration/src/cluster.rs
@@ -26,7 +26,9 @@ use cache::{
     with_default_composite_cache_registry,
 };
 use catalog::information_extension::DistributedInformationExtension;
-use catalog::kvbackend::{CachedKvBackendBuilder, KvBackendCatalogManagerBuilder, MetaKvBackend};
+use catalog::kvbackend::{
+    CachedKvBackendBuilder, KvBackendCatalogManagerBuilder, new_read_only_meta_kv_backend,
+};
 use catalog::process_manager::ProcessManager;
 use client::Client;
 use client::client_manager::NodeClients;
@@ -127,8 +129,8 @@ impl GreptimeDbCluster {
                 .await
                 .into_iter()
                 .flat_map(|e| {
-                    if e.index_file_path.is_some() {
-                        vec![e.file_path, e.index_file_path.unwrap()]
+                    if let Some(index_file_path) = e.index_file_path {
+                        vec![e.file_path, index_file_path]
                     } else {
                         vec![e.file_path]
                     }
@@ -285,6 +287,9 @@ impl GreptimeDbClusterBuilder {
             ..Default::default()
         };
 
+        test_util::prepare_another_catalog_and_schema_with_kv_backend(self.kv_backend.clone())
+            .await;
+
         let metasrv = meta_srv::mocks::mock(
             opt,
             self.kv_backend.clone(),
@@ -307,8 +312,6 @@ impl GreptimeDbClusterBuilder {
             .build_frontend(metasrv.clone(), datanode_clients, start_frontend_servers)
             .await;
 
-        test_util::prepare_another_catalog_and_schema(&frontend.instance).await;
-
         frontend.start().await.unwrap();
 
         GreptimeDbCluster {
@@ -422,9 +425,7 @@ impl GreptimeDbClusterBuilder {
         meta_client.start(&[&metasrv.server_addr]).await.unwrap();
         let meta_client = Arc::new(meta_client);
 
-        let meta_backend = Arc::new(MetaKvBackend {
-            client: meta_client.clone(),
-        });
+        let meta_backend = new_read_only_meta_kv_backend(meta_client.clone());
 
         let layered_cache_registry = Arc::new(
             LayeredCacheRegistryBuilder::default()
@@ -456,9 +457,9 @@ impl GreptimeDbClusterBuilder {
         meta_client.start(&[&metasrv.server_addr]).await.unwrap();
         let meta_client = Arc::new(meta_client);
 
-        let cached_meta_backend = Arc::new(
-            CachedKvBackendBuilder::new(Arc::new(MetaKvBackend::new(meta_client.clone()))).build(),
-        );
+        let readonly_meta_backend = new_read_only_meta_kv_backend(meta_client.clone());
+        let cached_meta_backend =
+            Arc::new(CachedKvBackendBuilder::new(readonly_meta_backend.clone()).build());
 
         let layered_cache_builder = LayeredCacheRegistryBuilder::default().add_cache_registry(
             CacheRegistryBuilder::default()
@@ -466,7 +467,7 @@ impl GreptimeDbClusterBuilder {
                 .build(),
         );
         let fundamental_cache_registry =
-            build_fundamental_cache_registry(Arc::new(MetaKvBackend::new(meta_client.clone())));
+            build_fundamental_cache_registry(readonly_meta_backend.clone());
         let cache_registry = Arc::new(
             with_default_composite_cache_registry(
                 layered_cache_builder.add_cache_registry(fundamental_cache_registry),
diff --git a/tests-integration/src/grpc/flight.rs b/tests-integration/src/grpc/flight.rs
index 9ed6b8176f..d638926fff 100644
--- a/tests-integration/src/grpc/flight.rs
+++ b/tests-integration/src/grpc/flight.rs
@@ -27,6 +27,7 @@ mod test {
     use common_grpc::flight::{FlightEncoder, FlightMessage};
     use common_query::OutputData;
     use common_recordbatch::RecordBatch;
+    use common_recordbatch::adapter::RegionWatermarkEntry;
     use datatypes::prelude::{ConcreteDataType, ScalarVector, VectorRef};
     use datatypes::schema::{ColumnSchema, Schema};
     use datatypes::vectors::{Int32Vector, StringVector, TimestampMillisecondVector};
@@ -129,6 +130,104 @@ mod test {
 | 1970-01-01T00:00:00.009 | -9 | s9 |
 +-------------------------+----+----+";
         query_and_expect(db.fe_instance().as_ref(), sql, expected).await;
+
+        let output = client.sql(sql).await.unwrap();
+        let OutputData::Stream(mut stream) = output.data else {
+            panic!("expected stream output");
+        };
+        while let Some(batch) = stream.next().await {
+            batch.unwrap();
+        }
+        let metrics = stream.metrics().expect("expected terminal metrics");
+        assert!(metrics.region_watermarks.is_empty());
+
+        let result = client
+            .sql_with_terminal_metrics(sql, &[("flow.return_region_seq", "true")])
+            .await
+            .unwrap();
+        let terminal_metrics = result.metrics.clone();
+        let OutputData::Stream(mut stream) = result.output.data else {
+            panic!("expected stream output");
+        };
+        while let Some(batch) = stream.next().await {
+            batch.unwrap();
+        }
+        assert!(terminal_metrics.is_ready());
+        let regions = db.list_all_regions().await;
+        assert_eq!(regions.len(), 1);
+        let (region_id, region) = regions.into_iter().next().unwrap();
+        let expected_watermark = (region_id.as_u64(), region.find_committed_sequence());
+        assert_eq!(
+            terminal_metrics.region_watermark_map(),
+            Some(std::collections::HashMap::from([expected_watermark]))
+        );
+
+        let output = client
+            .sql_with_hint(sql, &[("flow.return_region_seq", "true")])
+            .await
+            .unwrap();
+        let OutputData::Stream(mut stream) = output.data else {
+            panic!("expected stream output");
+        };
+
+        let mut row_count = 0;
+        while let Some(batch) = stream.next().await {
+            let batch = batch.unwrap();
+            row_count += batch.num_rows();
+        }
+        assert_eq!(row_count, 9);
+
+        let metrics = stream.metrics().expect("expected terminal metrics");
+        let region_watermarks = metrics.region_watermarks;
+        assert_eq!(
+            region_watermarks,
+            vec![RegionWatermarkEntry {
+                region_id: expected_watermark.0,
+                watermark: Some(expected_watermark.1),
+            }]
+        );
+
+        let previous_watermark = expected_watermark;
+
+        create_table_named(&client, "bar").await;
+        let result = client
+            .sql_with_terminal_metrics("insert into bar select ts, a, `B` from foo", &[])
+            .await
+            .unwrap();
+        let OutputData::AffectedRows(affected_rows) = result.output.data else {
+            panic!("expected affected rows output");
+        };
+        assert_eq!(affected_rows, 9);
+        assert!(result.metrics.is_ready());
+        assert!(result.region_watermark_map().is_none());
+
+        let err = client
+            .sql_with_terminal_metrics(
+                "insert into bar select ts, a, `B` from foo",
+                &[("flow.return_region_seq", "not-a-bool")],
+            )
+            .await
+            .unwrap_err();
+        let err_msg = format!("{err:?}");
+        assert!(err_msg.contains("Invalid value for flow.return_region_seq"));
+
+        client.sql("truncate table bar").await.unwrap();
+
+        let result = client
+            .sql_with_terminal_metrics(
+                "insert into bar select ts, a, `B` from foo",
+                &[("flow.return_region_seq", "true")],
+            )
+            .await
+            .unwrap();
+        let OutputData::AffectedRows(affected_rows) = result.output.data else {
+            panic!("expected affected rows output");
+        };
+        assert_eq!(affected_rows, 9);
+        assert_eq!(
+            result.region_watermark_map(),
+            Some(std::collections::HashMap::from([previous_watermark]))
+        );
     }
 
     async fn test_put_record_batches(client: &Database, record_batches: Vec<RecordBatch>) {
@@ -224,6 +323,10 @@ mod test {
     }
 
     async fn create_table(client: &Database) {
+        create_table_named(client, "foo").await;
+    }
+
+    async fn create_table_named(client: &Database, table_name: &str) {
         // create table foo (
         //   ts timestamp time index,
         //   a int primary key,
@@ -232,7 +335,7 @@ mod test {
         let output = client
             .create(CreateTableExpr {
                 schema_name: "public".to_string(),
-                table_name: "foo".to_string(),
+                table_name: table_name.to_string(),
                 column_defs: vec![
                     ColumnDef {
                         name: "ts".to_string(),
diff --git a/tests-integration/src/instance.rs b/tests-integration/src/instance.rs
index 6d39473e97..b11d5c32c6 100644
--- a/tests-integration/src/instance.rs
+++ b/tests-integration/src/instance.rs
@@ -33,7 +33,7 @@ mod tests {
     use datafusion_expr::LogicalPlan;
     use frontend::error::{Error, Result};
     use frontend::instance::Instance;
-    use query::parser::QueryLanguageParser;
+    use query::parser::{QueryLanguageParser, QueryStatement};
     use query::query_engine::DefaultSerializer;
     use servers::interceptor::{SqlQueryInterceptor, SqlQueryInterceptorRef};
     use servers::query_handler::sql::SqlQueryHandler;
@@ -381,6 +381,82 @@ mod tests {
         }
     }
 
+    #[tokio::test(flavor = "multi_thread")]
+    async fn test_exec_plan_interceptor_plugin() {
+        use std::sync::atomic::{AtomicBool, Ordering};
+
+        #[derive(Default)]
+        struct ExecPlanHook {
+            pub(crate) pre_execute_called: AtomicBool,
+            pub(crate) post_execute_called: AtomicBool,
+        }
+
+        impl SqlQueryInterceptor for ExecPlanHook {
+            type Error = Error;
+
+            fn pre_execute(
+                &self,
+                _statement: &Statement,
+                _plan: Option<&LogicalPlan>,
+                _query_ctx: QueryContextRef,
+            ) -> Result<()> {
+                self.pre_execute_called.store(true, Ordering::Relaxed);
+                Ok(())
+            }
+
+            fn post_execute(&self, output: Output, _query_ctx: QueryContextRef) -> Result<Output> {
+                self.post_execute_called.store(true, Ordering::Relaxed);
+                Ok(output)
+            }
+        }
+
+        let plugins = Plugins::new();
+        let hook = Arc::new(ExecPlanHook::default());
+        plugins.insert::<SqlQueryInterceptorRef<Error>>(hook.clone());
+
+        let standalone = GreptimeDbStandaloneBuilder::new("test_exec_plan_interceptor_plugin")
+            .with_plugin(plugins)
+            .build()
+            .await;
+        let instance = standalone.fe_instance().clone();
+
+        let sql = r#"CREATE TABLE demo(
+                            host STRING,
+                            ts TIMESTAMP,
+                            cpu DOUBLE NULL,
+                            TIME INDEX (ts),
+                            PRIMARY KEY(host)
+                        ) engine=mito;"#;
+        SqlQueryHandler::do_query(&*instance, sql, QueryContext::arc())
+            .await
+            .remove(0)
+            .unwrap();
+
+        let query_ctx = QueryContext::arc();
+        let stmt = QueryLanguageParser::parse_sql("SELECT * FROM demo", &query_ctx).unwrap();
+        let plan = instance
+            .statement_executor()
+            .plan(&stmt, query_ctx.clone())
+            .await
+            .unwrap();
+        let QueryStatement::Sql(sql_stmt) = stmt else {
+            unreachable!()
+        };
+
+        SqlQueryHandler::do_exec_plan(&*instance, plan, Some(sql_stmt), query_ctx.clone())
+            .await
+            .unwrap();
+
+        assert!(
+            hook.pre_execute_called.load(Ordering::Relaxed),
+            "pre_execute should be called for do_exec_plan"
+        );
+        assert!(
+            hook.post_execute_called.load(Ordering::Relaxed),
+            "post_execute should be called for do_exec_plan"
+        );
+    }
+
     #[tokio::test(flavor = "multi_thread")]
     async fn test_disable_db_operation_plugin() {
         #[derive(Default)]
diff --git a/tests-integration/src/lib.rs b/tests-integration/src/lib.rs
index 5def9351d0..5a4dca146f 100644
--- a/tests-integration/src/lib.rs
+++ b/tests-integration/src/lib.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#![feature(assert_matches)]
+#![recursion_limit = "256"]
 
 pub mod cluster;
 mod grpc;
diff --git a/tests-integration/src/test_util.rs b/tests-integration/src/test_util.rs
index fd0d1ef3c4..15d65c34ea 100644
--- a/tests-integration/src/test_util.rs
+++ b/tests-integration/src/test_util.rs
@@ -22,8 +22,10 @@ use axum::Router;
 use catalog::kvbackend::KvBackendCatalogManager;
 use common_base::Plugins;
 use common_config::Configurable;
+use common_meta::key::TableMetadataManager;
 use common_meta::key::catalog_name::CatalogNameKey;
 use common_meta::key::schema_name::SchemaNameKey;
+use common_meta::kv_backend::KvBackendRef;
 use common_query::Output;
 use common_runtime::runtime::BuilderBuild;
 use common_runtime::{Builder as RuntimeBuilder, Runtime};
@@ -534,6 +536,7 @@ pub async fn setup_test_http_app_with_frontend_and_custom_options(
         .with_influxdb_handler(instance.fe_instance().clone())
         .with_otlp_handler(instance.fe_instance().clone(), true)
         .with_jaeger_handler(instance.fe_instance().clone())
+        .with_dashboard_handler(instance.fe_instance().clone())
         .with_greptime_config_options(instance.opts.to_toml().unwrap());
 
     if let Some(user_provider) = user_provider {
@@ -622,6 +625,7 @@ pub async fn setup_test_prom_app_with_frontend(
             Some(frontend_ref.clone()),
             true,
             PromValidationMode::Strict,
+            None,
         )
         .with_prometheus_handler(frontend_ref)
         .with_greptime_config_options(instance.opts.datanode_options().to_toml().unwrap())
@@ -905,6 +909,17 @@ pub(crate) async fn prepare_another_catalog_and_schema(instance: &Instance) {
         .unwrap();
 
     let table_metadata_manager = catalog_manager.table_metadata_manager_ref();
+    prepare_another_catalog_and_schema_with_manager(table_metadata_manager).await;
+}
+
+pub(crate) async fn prepare_another_catalog_and_schema_with_kv_backend(kv_backend: KvBackendRef) {
+    let table_metadata_manager = TableMetadataManager::new(kv_backend);
+    prepare_another_catalog_and_schema_with_manager(&table_metadata_manager).await;
+}
+
+async fn prepare_another_catalog_and_schema_with_manager(
+    table_metadata_manager: &TableMetadataManager,
+) {
     table_metadata_manager
         .catalog_manager()
         .create(CatalogNameKey::new("another_catalog"), true)
diff --git a/tests-integration/src/tests/instance_kafka_wal_test.rs b/tests-integration/src/tests/instance_kafka_wal_test.rs
index ed74525aef..57be7e6cd6 100644
--- a/tests-integration/src/tests/instance_kafka_wal_test.rs
+++ b/tests-integration/src/tests/instance_kafka_wal_test.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::assert_matches::assert_matches;
+use std::assert_matches;
 use std::sync::Arc;
 use std::sync::atomic::{AtomicU64, Ordering};
 
diff --git a/tests-integration/src/tests/instance_test.rs b/tests-integration/src/tests/instance_test.rs
index 5879ae9dfa..3be45d8970 100644
--- a/tests-integration/src/tests/instance_test.rs
+++ b/tests-integration/src/tests/instance_test.rs
@@ -2834,36 +2834,9 @@ async fn test_copy_parquet_map_to_binary(instance: Arc<dyn MockInstance>) {
 async fn test_create_table_with_json_datatype(instance: Arc<dyn MockInstance>) {
     let instance = instance.frontend();
 
-    let sql = r#"
-CREATE TABLE a (
-    j JSON(format = "partial", unstructured_keys = ["foo", "foo.bar"]),
-    ts TIMESTAMP TIME INDEX,
-)"#;
-    let output = execute_sql(&instance, sql).await.data;
-    assert!(matches!(output, OutputData::AffectedRows(0)));
-
-    // "show create table" finds the information from table metadata.
-    // So if the output is expected, we know the options are really set.
-    let output = execute_sql(&instance, "SHOW CREATE TABLE a").await.data;
-    let expected = r#"
-+-------+------------------------------------------------------------------------------+
-| Table | Create Table                                                                 |
-+-------+------------------------------------------------------------------------------+
-| a     | CREATE TABLE IF NOT EXISTS "a" (                                             |
-|       |   "j" JSON(format = 'partial', unstructured_keys = ['foo', 'foo.bar']) NULL, |
-|       |   "ts" TIMESTAMP(3) NOT NULL,                                                |
-|       |   TIME INDEX ("ts")                                                          |
-|       | )                                                                            |
-|       |                                                                              |
-|       | ENGINE=mito                                                                  |
-|       |                                                                              |
-+-------+------------------------------------------------------------------------------+"#;
-    check_output_stream(output, expected).await;
-
-    // test the default options
     let sql = r#"
 CREATE TABLE b (
-    j JSON,
+    j JSON2,
     ts TIMESTAMP TIME INDEX,
 )"#;
     let output = execute_sql(&instance, sql).await.data;
diff --git a/tests-integration/src/tests/promql_test.rs b/tests-integration/src/tests/promql_test.rs
index 7fbce91ea6..ede4663118 100644
--- a/tests-integration/src/tests/promql_test.rs
+++ b/tests-integration/src/tests/promql_test.rs
@@ -15,7 +15,9 @@
 use std::sync::Arc;
 use std::time::{Duration, SystemTime, UNIX_EPOCH};
 
-use common_query::Output;
+use common_query::{Output, OutputData};
+use common_recordbatch::util::collect_batches;
+use datatypes::arrow::array::{Float64Array, Int64Array};
 use frontend::instance::Instance;
 use query::parser::{PromQuery, QueryLanguageParser, QueryStatement};
 use rstest::rstest;
@@ -151,6 +153,103 @@ async fn create_insert_tql_assert(
     check_unordered_output_stream(query_output, expected).await;
 }
 
+async fn execute_all(instance: &Arc<Instance>, sql: &str, query_ctx: Arc<QueryContext>) {
+    instance
+        .do_query(sql, query_ctx)
+        .await
+        .into_iter()
+        .for_each(|v| {
+            let _ = v.unwrap();
+        });
+}
+
+#[allow(clippy::too_many_arguments)]
+async fn promql_query_as_batches(
+    ins: Arc<Instance>,
+    promql: &str,
+    alias: Option<String>,
+    query_ctx: Arc<QueryContext>,
+    start: SystemTime,
+    end: SystemTime,
+    interval: Duration,
+    lookback: Duration,
+) -> common_recordbatch::RecordBatches {
+    let output = promql_query(
+        ins, promql, alias, query_ctx, start, end, interval, lookback,
+    )
+    .await
+    .unwrap();
+    match output.data {
+        OutputData::Stream(stream) => collect_batches(stream).await.unwrap(),
+        OutputData::RecordBatches(recordbatches) => recordbatches,
+        _ => unreachable!(),
+    }
+}
+
+const ANON_PROMQL_RATIO_REPRO_DB: &str = "repro_db";
+
+const ANON_PROMQL_RATIO_REPRO_CREATE: &str = r#"
+CREATE TABLE phy (
+    t TIMESTAMP TIME INDEX,
+    v DOUBLE
+) ENGINE=metric WITH ("physical_metric_table" = "");
+
+CREATE TABLE metric_a (
+    l1 STRING NULL,
+    l2 STRING NULL,
+    l3 STRING NULL,
+    l4 STRING NULL,
+    l5 STRING NULL,
+    t TIMESTAMP NOT NULL,
+    v DOUBLE NULL,
+    TIME INDEX (t),
+    PRIMARY KEY (l1, l2, l3, l4, l5)
+) ENGINE=metric WITH (on_physical_table = 'phy');
+
+CREATE TABLE metric_b (
+    l6 STRING NULL,
+    l1 STRING NULL,
+    l2 STRING NULL,
+    l3 STRING NULL,
+    l4 STRING NULL,
+    t TIMESTAMP NOT NULL,
+    v DOUBLE NULL,
+    TIME INDEX (t),
+    PRIMARY KEY (l6, l1, l2, l3, l4)
+) ENGINE=metric WITH (on_physical_table = 'phy');
+"#;
+
+const ANON_PROMQL_RATIO_REPRO_INSERT: &str = r#"
+INSERT INTO metric_a (l1, l2, l3, l4, l5, t, v) VALUES
+    ('v1', 'v2', 'v3', 'v4a', 'v5a', 1, 0),
+    ('v1', 'v2', 'v3', 'v4a', 'v5a', 180000, 120),
+    ('v1', 'v2', 'v3', 'v4a', 'v5a', 360000, 240),
+    ('v1', 'v2', 'v3', 'v4a', 'v5b', 1, 0),
+    ('v1', 'v2', 'v3', 'v4a', 'v5b', 180000, 30),
+    ('v1', 'v2', 'v3', 'v4a', 'v5b', 360000, 60),
+    ('v1', 'v2', 'v3-b', 'v4b', 'v5c', 1, 0),
+    ('v1', 'v2', 'v3-b', 'v4b', 'v5c', 180000, 60),
+    ('v1', 'v2', 'v3-b', 'v4b', 'v5c', 360000, 120);
+
+INSERT INTO metric_b (l6, l1, l2, l3, l4, t, v) VALUES
+    ('v6', 'v1', 'v2', 'v3', 'v4a', 1, 1),
+    ('v6', 'v1', 'v2', 'v3', 'v4a', 180000, 1),
+    ('v6', 'v1', 'v2', 'v3', 'v4a', 360000, 1),
+    ('v6', 'v1', 'v2', 'v3-b', 'v4b', 1, 2),
+    ('v6', 'v1', 'v2', 'v3-b', 'v4b', 180000, 2),
+    ('v6', 'v1', 'v2', 'v3-b', 'v4b', 360000, 2);
+"#;
+
+const ANON_PROMQL_RATIO_REPRO_NUMERATOR: &str = r#"count(((rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)",__schema__="repro_db"}[3m]) / on(l3,l4) group_left metric_b{l6="v6",l1="v1",l2="v2",l3=~"v3(|-a|-b)",__schema__="repro_db"}) > 0.50))"#;
+
+const ANON_PROMQL_RATIO_REPRO_DENOMINATOR: &str =
+    r#"count(rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)",__schema__="repro_db"}[3m]))"#;
+
+const ANON_PROMQL_RATIO_REPRO_WHOLE: &str = r#"(count(((rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)",__schema__="repro_db"}[3m]) / on(l3,l4) group_left metric_b{l6="v6",l1="v1",l2="v2",l3=~"v3(|-a|-b)",__schema__="repro_db"}) > 0.50)) / count(rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)",__schema__="repro_db"}[3m]))) * 100"#;
+
+const ANON_PROMQL_RATIO_REPRO_SCALAR_DIV: &str =
+    r#"count(rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)",__schema__="repro_db"}[3m])) / 2"#;
+
 #[apply(both_instances_cases)]
 async fn sql_insert_tql_query_ceil(instance: Arc<dyn MockInstance>) {
     let instance = instance.frontend();
@@ -709,3 +808,140 @@ async fn cross_schema_query(instance: Arc<dyn MockInstance>) {
 
     check_unordered_output_stream(query_output, expected).await;
 }
+
+#[apply(both_instances_cases)]
+async fn anon_promql_ratio_repro(instance: Arc<dyn MockInstance>) {
+    let ins = instance.frontend();
+
+    execute_all(
+        &ins,
+        &format!("CREATE DATABASE {ANON_PROMQL_RATIO_REPRO_DB}"),
+        QueryContext::arc(),
+    )
+    .await;
+
+    let repro_ctx: Arc<QueryContext> =
+        QueryContext::with_db_name(Some(ANON_PROMQL_RATIO_REPRO_DB)).into();
+    execute_all(&ins, ANON_PROMQL_RATIO_REPRO_CREATE, repro_ctx.clone()).await;
+    execute_all(&ins, ANON_PROMQL_RATIO_REPRO_INSERT, repro_ctx).await;
+
+    let start = UNIX_EPOCH.checked_add(Duration::from_secs(180)).unwrap();
+    let end = UNIX_EPOCH.checked_add(Duration::from_secs(360)).unwrap();
+    let interval = Duration::from_secs(180);
+    let lookback = Duration::from_secs(1);
+
+    let numerator = promql_query_as_batches(
+        ins.clone(),
+        ANON_PROMQL_RATIO_REPRO_NUMERATOR,
+        Some("num".to_string()),
+        QueryContext::arc(),
+        start,
+        end,
+        interval,
+        lookback,
+    )
+    .await;
+    let denominator = promql_query_as_batches(
+        ins.clone(),
+        ANON_PROMQL_RATIO_REPRO_DENOMINATOR,
+        Some("den".to_string()),
+        QueryContext::arc(),
+        start,
+        end,
+        interval,
+        lookback,
+    )
+    .await;
+    let whole = promql_query_as_batches(
+        ins.clone(),
+        ANON_PROMQL_RATIO_REPRO_WHOLE,
+        Some("pct".to_string()),
+        QueryContext::arc(),
+        start,
+        end,
+        interval,
+        lookback,
+    )
+    .await;
+    let scalar_div = promql_query_as_batches(
+        ins,
+        ANON_PROMQL_RATIO_REPRO_SCALAR_DIV,
+        Some("half_den".to_string()),
+        QueryContext::arc(),
+        start,
+        end,
+        interval,
+        lookback,
+    )
+    .await;
+
+    let numerator = numerator.iter().collect::<Vec<_>>();
+    let denominator = denominator.iter().collect::<Vec<_>>();
+    let whole = whole.iter().collect::<Vec<_>>();
+    let scalar_div = scalar_div.iter().collect::<Vec<_>>();
+
+    let numerator_values = numerator[0]
+        .column_by_name("num")
+        .unwrap()
+        .as_any()
+        .downcast_ref::<Int64Array>()
+        .unwrap();
+    let denominator_values = denominator[0]
+        .column_by_name("den")
+        .unwrap()
+        .as_any()
+        .downcast_ref::<Int64Array>()
+        .unwrap();
+    let percentage_values = whole[0]
+        .column_by_name("pct")
+        .unwrap()
+        .as_any()
+        .downcast_ref::<Float64Array>()
+        .unwrap();
+    let scalar_div_values = scalar_div[0]
+        .column_by_name("half_den")
+        .unwrap()
+        .as_any()
+        .downcast_ref::<Float64Array>()
+        .unwrap();
+
+    assert_eq!(numerator_values.len(), 1, "{}", numerator[0].pretty_print());
+    assert_eq!(
+        denominator_values.len(),
+        1,
+        "{}",
+        denominator[0].pretty_print()
+    );
+    assert_eq!(percentage_values.len(), 1, "{}", whole[0].pretty_print());
+    assert_eq!(
+        scalar_div_values.len(),
+        1,
+        "{}",
+        scalar_div[0].pretty_print()
+    );
+
+    assert_eq!(
+        numerator_values.value(0),
+        1,
+        "{}",
+        numerator[0].pretty_print()
+    );
+    assert_eq!(
+        denominator_values.value(0),
+        3,
+        "{}",
+        denominator[0].pretty_print()
+    );
+    assert!(
+        (scalar_div_values.value(0) - 1.5).abs() < 1e-9,
+        "{}",
+        scalar_div[0].pretty_print()
+    );
+
+    let expected = 100.0 / 3.0;
+    assert!(
+        (percentage_values.value(0) - expected).abs() < 1e-9,
+        "{}",
+        whole[0].pretty_print()
+    );
+}
diff --git a/tests-integration/tests/grpc.rs b/tests-integration/tests/grpc.rs
index e31444b45c..0f1112ff4a 100644
--- a/tests-integration/tests/grpc.rs
+++ b/tests-integration/tests/grpc.rs
@@ -22,6 +22,7 @@ use api::v1::{
     column,
 };
 use auth::user_provider_from_option;
+use base64::prelude::{BASE64_STANDARD, Engine as _};
 use client::{Client, DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME, Database, OutputData};
 use common_catalog::consts::MITO_ENGINE;
 use common_grpc::channel_manager::ClientTlsOption;
@@ -336,7 +337,7 @@ pub async fn test_otel_arrow_auth(store_type: StorageType) {
         let mut request = Request::new(stream);
         request.metadata_mut().insert(
             "authorization",
-            MetadataValue::from_static("Basic Z3JlcHRpbWVfdXNlcjpncmVwdGltZV9wd2Q="), // greptime_user:greptime_pwd base64 encoded
+            MetadataValue::try_from(basic_auth("greptime_user", "greptime_pwd")).unwrap(),
         );
         let response = client.arrow_metrics(request).await;
         assert!(response.is_ok());
@@ -356,7 +357,8 @@ pub async fn test_otel_arrow_auth(store_type: StorageType) {
         let mut request = Request::new(stream);
         request.metadata_mut().insert(
             "authorization",
-            MetadataValue::from_static("Z3JlcHRpbWVfdXNlcjpncmVwdGltZV9wd2Q="), // greptime_user:greptime_pwd base64 encoded
+            MetadataValue::try_from(basic_auth_credentials("greptime_user", "greptime_pwd"))
+                .unwrap(),
         );
         let response = client.arrow_metrics(request).await;
         assert!(response.is_ok());
@@ -374,6 +376,14 @@ pub async fn test_otel_arrow_auth(store_type: StorageType) {
     let _ = fe_grpc_server.shutdown().await;
 }
 
+fn basic_auth(username: &str, password: &str) -> String {
+    format!("Basic {}", basic_auth_credentials(username, password))
+}
+
+fn basic_auth_credentials(username: &str, password: &str) -> String {
+    BASE64_STANDARD.encode(format!("{username}:{password}"))
+}
+
 pub async fn test_auto_create_table(store_type: StorageType) {
     let (_db, fe_grpc_server) = setup_grpc_server(store_type, "test_auto_create_table").await;
     let addr = fe_grpc_server.bind_addr().unwrap().to_string();
diff --git a/tests-integration/tests/http.rs b/tests-integration/tests/http.rs
index 68fa2a228d..7fe12cc8b9 100644
--- a/tests-integration/tests/http.rs
+++ b/tests-integration/tests/http.rs
@@ -21,18 +21,23 @@ use api::prom_store::remote::label_matcher::Type as MatcherType;
 use api::prom_store::remote::{
     Label, LabelMatcher, Query, ReadRequest, ReadResponse, Sample, TimeSeries, WriteRequest,
 };
-use auth::user_provider_from_option;
+use auth::{UserProviderRef, user_provider_from_option};
 use axum::http::{HeaderName, HeaderValue, StatusCode};
+use base64::prelude::{BASE64_STANDARD, Engine as _};
 use chrono::Utc;
+use cmd::options::GreptimeOptions;
+use common_base::Plugins;
 use common_catalog::consts::{
     DEFAULT_PRIVATE_SCHEMA_NAME, TRACE_TABLE_NAME, trace_operations_table_name,
     trace_services_table_name,
 };
+use common_config::Configurable;
 use common_error::status_code::StatusCode as ErrorCode;
 use common_frontend::slow_query_event::{
     SLOW_QUERY_TABLE_NAME, SLOW_QUERY_TABLE_QUERY_COLUMN_NAME,
 };
 use common_memory_manager::OnExhaustedPolicy;
+use common_options::plugin_options::StandaloneFlag;
 use flate2::Compression;
 use flate2::write::GzEncoder;
 use log_query::{Context, Limit, LogQuery, TimeFilter};
@@ -40,7 +45,9 @@ use loki_proto::logproto::{EntryAdapter, LabelPairAdapter, PushRequest, StreamAd
 use loki_proto::prost_types::Timestamp;
 use opentelemetry_proto::tonic::collector::logs::v1::ExportLogsServiceRequest;
 use opentelemetry_proto::tonic::collector::metrics::v1::ExportMetricsServiceRequest;
-use opentelemetry_proto::tonic::collector::trace::v1::ExportTraceServiceRequest;
+use opentelemetry_proto::tonic::collector::trace::v1::{
+    ExportTraceServiceRequest, ExportTraceServiceResponse,
+};
 use pipeline::GREPTIME_INTERNAL_TRACE_PIPELINE_V1_NAME;
 use prost::Message;
 use serde_json::{Value, json};
@@ -50,6 +57,7 @@ use servers::http::header::constants::{
     GREPTIME_LOG_TABLE_NAME_HEADER_NAME, GREPTIME_PIPELINE_NAME_HEADER_NAME,
 };
 use servers::http::header::{GREPTIME_DB_HEADER_NAME, GREPTIME_TIMEZONE_HEADER_NAME};
+use servers::http::otlp::GoogleRpcStatus;
 use servers::http::prometheus::{Column, PrometheusJsonResponse, PrometheusResponse};
 use servers::http::result::error_result::ErrorResponse;
 use servers::http::result::greptime_result_v1::GreptimedbV1Response;
@@ -57,6 +65,7 @@ use servers::http::result::influxdb_result_v1::{InfluxdbOutput, InfluxdbV1Respon
 use servers::http::test_helpers::{TestClient, TestResponse};
 use servers::prom_store::{self, mock_timeseries_new_label};
 use servers::request_memory_limiter::ServerMemoryLimiter;
+use standalone::options::StandaloneOptions;
 use table::table_name::TableName;
 use tests_integration::test_util::{
     StorageType, setup_test_http_app, setup_test_http_app_with_frontend,
@@ -106,6 +115,7 @@ macro_rules! http_tests {
                 test_config_api,
                 test_dynamic_tracer_toggle,
                 test_dashboard_path,
+                test_dashboard_api,
                 test_prometheus_remote_write,
                 test_prometheus_remote_special_labels,
                 test_prometheus_remote_schema_labels,
@@ -131,6 +141,7 @@ macro_rules! http_tests {
                 test_pipeline_skip_error,
                 test_pipeline_filter,
                 test_pipeline_create_table,
+                test_pipeline_index_options,
 
                 test_otlp_metrics_new,
                 test_otlp_traces_v0,
@@ -147,6 +158,8 @@ macro_rules! http_tests {
                 test_jaeger_query_api_for_trace_v1,
 
                 test_influxdb_write,
+                test_influxdb_write_with_hints,
+                test_influxdb_write_with_append_mode_hint,
                 test_http_memory_limit,
             );
         )*
@@ -179,7 +192,7 @@ pub async fn test_http_auth(store_type: StorageType) {
     // 2. wrong auth
     let res = client
         .get("/v1/sql?db=public&sql=show tables;")
-        .header("Authorization", "basic Z3JlcHRpbWVfdXNlcjp3cm9uZ19wd2Q=")
+        .header("Authorization", basic_auth("greptime_user", "wrong_pwd"))
         .send()
         .await;
     assert_eq!(res.status(), StatusCode::UNAUTHORIZED);
@@ -187,10 +200,7 @@ pub async fn test_http_auth(store_type: StorageType) {
     // 3. right auth
     let res = client
         .get("/v1/sql?db=public&sql=show tables;")
-        .header(
-            "Authorization",
-            "basic Z3JlcHRpbWVfdXNlcjpncmVwdGltZV9wd2Q=",
-        )
+        .header("Authorization", basic_auth("greptime_user", "greptime_pwd"))
         .send()
         .await;
     assert_eq!(res.status(), StatusCode::OK);
@@ -198,19 +208,13 @@ pub async fn test_http_auth(store_type: StorageType) {
     // 4. readonly user cannot write
     let res = client
         .get("/v1/sql?db=public&sql=show tables;")
-        .header(
-            "Authorization",
-            "basic cmVhZG9ubHlfdXNlcjpyZWFkb25seV9wd2Q=",
-        )
+        .header("Authorization", basic_auth("readonly_user", "readonly_pwd"))
         .send()
         .await;
     assert_eq!(res.status(), StatusCode::OK);
     let res = client
         .get("/v1/sql?db=public&sql=create table auth_test(ts timestamp time index);")
-        .header(
-            "Authorization",
-            "basic cmVhZG9ubHlfdXNlcjpyZWFkb25seV9wd2Q=",
-        )
+        .header("Authorization", basic_auth("readonly_user", "readonly_pwd"))
         .send()
         .await;
     assert_eq!(res.status(), StatusCode::FORBIDDEN);
@@ -220,7 +224,7 @@ pub async fn test_http_auth(store_type: StorageType) {
         .get("/v1/sql?db=public&sql=show tables;")
         .header(
             "Authorization",
-            "basic d3JpdGVvbmx5X3VzZXI6d3JpdGVvbmx5X3B3ZA==",
+            basic_auth("writeonly_user", "writeonly_pwd"),
         )
         .send()
         .await;
@@ -229,7 +233,7 @@ pub async fn test_http_auth(store_type: StorageType) {
         .get("/v1/sql?db=public&sql=create table auth_test(ts timestamp time index);")
         .header(
             "Authorization",
-            "basic d3JpdGVvbmx5X3VzZXI6d3JpdGVvbmx5X3B3ZA==",
+            basic_auth("writeonly_user", "writeonly_pwd"),
         )
         .send()
         .await;
@@ -238,7 +242,7 @@ pub async fn test_http_auth(store_type: StorageType) {
         .get("/v1/sql?db=public&sql=insert into auth_test values(1);")
         .header(
             "Authorization",
-            "basic d3JpdGVvbmx5X3VzZXI6d3JpdGVvbmx5X3B3ZA==",
+            basic_auth("writeonly_user", "writeonly_pwd"),
         )
         .send()
         .await;
@@ -247,7 +251,7 @@ pub async fn test_http_auth(store_type: StorageType) {
         .get("/v1/sql?db=public&sql=select * from auth_test;")
         .header(
             "Authorization",
-            "basic d3JpdGVvbmx5X3VzZXI6d3JpdGVvbmx5X3B3ZA==",
+            basic_auth("writeonly_user", "writeonly_pwd"),
         )
         .send()
         .await;
@@ -256,6 +260,66 @@ pub async fn test_http_auth(store_type: StorageType) {
     guard.remove_all().await;
 }
 
+fn basic_auth(username: &str, password: &str) -> String {
+    format!(
+        "basic {}",
+        BASE64_STANDARD.encode(format!("{username}:{password}"))
+    )
+}
+
+pub async fn test_http_auth_from_standalone_user_provider_config() {
+    common_telemetry::init_default_ut_logging();
+
+    let config = tempfile::NamedTempFile::new().unwrap();
+    let user_provider = "static_user_provider:cmd:greptime_user=greptime_pwd";
+    std::fs::write(
+        config.path(),
+        format!("user_provider = \"{user_provider}\"\n"),
+    )
+    .unwrap();
+
+    let options =
+        GreptimeOptions::<StandaloneOptions>::load_layered_options(config.path().to_str(), "")
+            .unwrap();
+    let fe_opts = options.component.frontend_options();
+
+    let mut plugins = Plugins::new();
+    plugins.insert(StandaloneFlag);
+    plugins::setup_frontend_plugins(&mut plugins, &[], &fe_opts)
+        .await
+        .unwrap();
+    let user_provider = plugins.get::<UserProviderRef>();
+
+    let (app, mut guard) = setup_test_http_app_with_frontend_and_user_provider(
+        StorageType::File,
+        "sql_api_user_provider_config",
+        user_provider,
+    )
+    .await;
+    let client = TestClient::new(app).await;
+
+    let res = client
+        .get("/v1/sql?db=public&sql=show tables;")
+        .send()
+        .await;
+    let status = res.status();
+    let body = res.text().await;
+    assert_eq!(
+        status,
+        StatusCode::UNAUTHORIZED,
+        "unexpected response body: {body}"
+    );
+
+    let res = client
+        .get("/v1/sql?db=public&sql=show tables;")
+        .header("Authorization", basic_auth("greptime_user", "greptime_pwd"))
+        .send()
+        .await;
+    assert_eq!(res.status(), StatusCode::OK);
+
+    guard.remove_all().await;
+}
+
 #[tokio::test]
 pub async fn test_cors() {
     let (app, mut guard) = setup_test_http_app_with_frontend(StorageType::File, "test_cors").await;
@@ -1172,6 +1236,20 @@ pub async fn test_prom_http_api(store_type: StorageType) {
         .await;
     assert_eq!(res.status(), StatusCode::OK);
 
+    // query non-exist label in metric table
+    let res = client
+        .get("/v1/prometheus/api/v1/label/not_exist_label/values?match[]=demo&start=0&end=600")
+        .send()
+        .await;
+    assert_eq!(res.status(), StatusCode::OK);
+    let prom_resp = res.json::<PrometheusJsonResponse>().await;
+    assert_eq!(prom_resp.status, "success");
+    assert!(prom_resp.error.is_none() && prom_resp.error_type.is_none());
+    assert_eq!(
+        prom_resp.data,
+        serde_json::from_value::<PrometheusResponse>(json!([])).unwrap()
+    );
+
     // query `__name__` without match[]
     // create a physical table and a logical table
     let res = client
@@ -1416,6 +1494,7 @@ max_in_flight_write_bytes = "0KiB"
 write_bytes_exhausted_policy = "wait"
 init_regions_in_background = false
 init_regions_parallelism = 16
+heartbeat_env_vars = []
 
 [http]
 addr = "127.0.0.1:4000"
@@ -1481,6 +1560,11 @@ enable = true
 [prom_store]
 enable = true
 with_metric_engine = true
+pending_rows_flush_interval = "0s"
+max_batch_rows = 100000
+max_concurrent_flushes = 256
+worker_channel_capacity = 65526
+max_inflight_requests = 3000
 
 [wal]
 provider = "raft_engine"
@@ -1513,7 +1597,6 @@ experimental_min_refresh_duration = "5s"
 grpc_conn_timeout = "5s"
 experimental_grpc_max_retries = 3
 experimental_frontend_scan_timeout = "30s"
-experimental_frontend_activity_timeout = "1m"
 experimental_max_filter_num_per_query = 20
 experimental_time_window_merge_threshold = 3
 read_preference = "Leader"
@@ -1543,11 +1626,11 @@ index_cache_percent = 20
 enable_refill_cache_on_read = true
 manifest_cache_size = "256MiB"
 sst_write_buffer_size = "8MiB"
-parallel_scan_channel_size = 32
 max_concurrent_scan_files = 384
 allow_stale_entries = false
+scan_memory_on_exhausted = "fail"
 min_compaction_interval = "0s"
-default_experimental_flat_format = false
+default_flat_format = true
 
 [region_engine.mito.index]
 aux_path = ""
@@ -1640,6 +1723,7 @@ fn drop_lines_with_inconsistent_results(input: String) -> String {
         "metadata_cache_size =",
         "content_cache_size =",
         "result_cache_size =",
+        "range_result_cache_size =",
         "name =",
         "recovery_parallelism =",
         "max_background_index_builds =",
@@ -1720,6 +1804,121 @@ pub async fn test_dashboard_path(store_type: StorageType) {
 #[cfg(not(feature = "dashboard"))]
 pub async fn test_dashboard_path(_: StorageType) {}
 
+#[cfg(feature = "dashboard")]
+pub async fn test_dashboard_api(store_type: StorageType) {
+    common_telemetry::init_default_ut_logging();
+    let (app, mut guard) = setup_test_http_app_with_frontend(store_type, "dashboard_api").await;
+    let client = TestClient::new(app).await;
+
+    // 1. List dashboards - should be empty initially
+    let res = client.get("/v1/dashboards").send().await;
+    assert_eq!(res.status(), StatusCode::OK);
+    let body: Value = res.json().await;
+    let dashboards = body.get("dashboards").unwrap().as_array().unwrap();
+    assert!(dashboards.is_empty());
+
+    // 2. Save a dashboard
+    let dashboard_definition = r#"{"title": "My Dashboard", "panels": []}"#;
+    let res = client
+        .post("/v1/dashboards/test_dashboard")
+        .body(dashboard_definition)
+        .send()
+        .await;
+    assert_eq!(res.status(), StatusCode::OK);
+    let body: Value = res.json().await;
+    let dashboards = body.get("dashboards").unwrap().as_array().unwrap();
+    assert_eq!(dashboards.len(), 1);
+    assert_eq!(dashboards[0].get("name").unwrap(), "test_dashboard");
+
+    // 3. Save another dashboard
+    let res = client
+        .post("/v1/dashboards/another_dashboard")
+        .body(r#"{"title": "Another Dashboard"}"#)
+        .send()
+        .await;
+    assert_eq!(res.status(), StatusCode::OK);
+
+    // 4. List dashboards - should have 2
+    let res = client.get("/v1/dashboards").send().await;
+    assert_eq!(res.status(), StatusCode::OK);
+    let body: Value = res.json().await;
+    let dashboards = body.get("dashboards").unwrap().as_array().unwrap();
+    assert_eq!(dashboards.len(), 2);
+
+    let names: Vec<&str> = dashboards
+        .iter()
+        .map(|d| d.get("name").unwrap().as_str().unwrap())
+        .collect();
+    assert!(names.contains(&"test_dashboard"));
+    assert!(names.contains(&"another_dashboard"));
+
+    // 5. Update a dashboard by posting again with new definition
+    let updated_definition = r#"{"title": "Updated Dashboard", "panels": [{"id": 1}]}"#;
+    let res = client
+        .post("/v1/dashboards/test_dashboard")
+        .body(updated_definition)
+        .send()
+        .await;
+    assert_eq!(res.status(), StatusCode::OK);
+
+    let body: Value = res.json().await;
+    let dashboards = body.get("dashboards").unwrap().as_array().unwrap();
+    assert_eq!(dashboards.len(), 1);
+    assert_eq!(dashboards[0].get("name").unwrap(), "test_dashboard");
+
+    // Verify the definition was updated by listing again
+    let res = client.get("/v1/dashboards").send().await;
+    assert_eq!(res.status(), StatusCode::OK);
+    let body: Value = res.json().await;
+    let dashboards = body.get("dashboards").unwrap().as_array().unwrap();
+    assert_eq!(dashboards.len(), 2);
+
+    // Find test_dashboard and verify it has updated definition
+    let test_db = dashboards
+        .iter()
+        .find(|d| d.get("name").unwrap() == "test_dashboard")
+        .unwrap();
+    assert_eq!(
+        test_db.get("definition").unwrap(),
+        r#"{"title": "Updated Dashboard", "panels": [{"id": 1}]}"#
+    );
+
+    // 6. Delete one dashboard
+    let res = client.delete("/v1/dashboards/test_dashboard").send().await;
+    assert_eq!(res.status(), StatusCode::OK);
+    let body: Value = res.json().await;
+    let dashboards = body.get("dashboards").unwrap().as_array().unwrap();
+    assert_eq!(dashboards.len(), 1);
+    assert_eq!(dashboards[0].get("name").unwrap(), "test_dashboard");
+
+    // 7. List dashboards - should have 1
+    let res = client.get("/v1/dashboards").send().await;
+    assert_eq!(res.status(), StatusCode::OK);
+    let body: Value = res.json().await;
+    let dashboards = body.get("dashboards").unwrap().as_array().unwrap();
+    assert_eq!(dashboards.len(), 1);
+    assert_eq!(dashboards[0].get("name").unwrap(), "another_dashboard");
+
+    // 8. Delete the remaining dashboard
+    let res = client
+        .delete("/v1/dashboards/another_dashboard")
+        .send()
+        .await;
+    assert_eq!(res.status(), StatusCode::OK);
+
+    // 9. List dashboards - should be empty
+    let res = client.get("/v1/dashboards").send().await;
+    assert_eq!(res.status(), StatusCode::OK);
+    let body: Value = res.json().await;
+    let dashboards = body.get("dashboards").unwrap().as_array().unwrap();
+    assert!(dashboards.is_empty());
+
+    guard.remove_all().await;
+}
+
+#[cfg(not(feature = "dashboard"))]
+pub async fn test_dashboard_api(_: StorageType) {}
+
 pub async fn test_prometheus_remote_write(store_type: StorageType) {
     common_telemetry::init_default_ut_logging();
     let (app, mut guard) =
@@ -2936,6 +3135,82 @@ transform:
     guard.remove_all().await;
 }
 
+pub async fn test_pipeline_index_options(store_type: StorageType) {
+    common_telemetry::init_default_ut_logging();
+    let (app, mut guard) =
+        setup_test_http_app_with_frontend(store_type, "test_pipeline_index_options").await;
+
+    let client = TestClient::new(app).await;
+
+    let pipeline_body = r#"
+processors:
+  - date:
+      field: ts
+      formats:
+        - "%Y-%m-%d %H:%M:%S%.3f"
+      ignore_missing: true
+
+transform:
+  - field: message
+    type: string
+    index:
+      type: fulltext
+      options:
+        analyzer: Chinese
+        case_sensitive: true
+        backend: bloom
+        granularity: 2048
+        false_positive_rate: 0.02
+  - field: trace_id
+    type: int64
+    index:
+      type: skipping
+      options:
+        granularity: 1024
+        false_positive_rate: 0.05
+        type: BLOOM
+  - field: ts
+    type: time
+    index: timestamp
+"#;
+
+    let res = client
+        .post("/v1/pipelines/index_options")
+        .header("Content-Type", "application/x-yaml")
+        .body(pipeline_body)
+        .send()
+        .await;
+    assert_eq!(res.status(), StatusCode::OK);
+
+    let data_body = r#"
+[
+  {
+    "message": "hello greptime",
+    "trace_id": 42,
+    "ts": "2024-05-25 20:16:37.217"
+  }
+]
+"#;
+    let res = client
+        .post("/v1/ingest?db=public&table=pipeline_index_options&pipeline_name=index_options")
+        .header("Content-Type", "application/json")
+        .body(data_body)
+        .send()
+        .await;
+    assert_eq!(res.status(), StatusCode::OK);
+
+    let expected_schema = "[[\"pipeline_index_options\",\"CREATE TABLE IF NOT EXISTS \\\"pipeline_index_options\\\" (\\n  \\\"message\\\" STRING NULL FULLTEXT INDEX WITH(analyzer = 'Chinese', backend = 'bloom', case_sensitive = 'true', false_positive_rate = '0.02', granularity = '2048'),\\n  \\\"trace_id\\\" BIGINT NULL SKIPPING INDEX WITH(false_positive_rate = '0.05', granularity = '1024', type = 'BLOOM'),\\n  \\\"ts\\\" TIMESTAMP(9) NOT NULL,\\n  TIME INDEX (\\\"ts\\\")\\n)\\n\\nENGINE=mito\\nWITH(\\n  'comment' = 'Created on insertion',\\n  append_mode = 'true'\\n)\"]]";
+    validate_data(
+        "pipeline_index_options_schema",
+        &client,
+        "show create table pipeline_index_options",
+        expected_schema,
+    )
+    .await;
+
+    guard.remove_all().await;
+}
+
 pub async fn test_pipeline_dispatcher(storage_type: StorageType) {
     common_telemetry::init_default_ut_logging();
     let (app, mut guard) =
@@ -3522,6 +3797,79 @@ transform:
     guard.remove_all().await;
 }
 
+pub async fn test_influxdb_write_with_hints(storage_type: StorageType) {
+    common_telemetry::init_default_ut_logging();
+    let (app, mut guard) =
+        setup_test_http_app_with_frontend(storage_type, "test_influxdb_write_with_hints").await;
+
+    let client = TestClient::new(app).await;
+
+    let result = client
+        .post("/v1/influxdb/write?db=public")
+        .header("x-greptime-hints", "sst_format=flat,ttl=30d,skip_wal=true")
+        .body("sst_fmt_table,host=host1 cpu=1.2 1664370459457010101")
+        .send()
+        .await;
+    assert_eq!(result.status(), 204);
+
+    let res = client
+        .get("/v1/sql?sql=show create table sst_fmt_table")
+        .send()
+        .await;
+    assert_eq!(res.status(), StatusCode::OK);
+    let resp = res.text().await;
+    assert!(
+        resp.contains("sst_format = 'flat'"),
+        "expected sst_format = 'flat' in SHOW CREATE TABLE output, got: {resp}"
+    );
+    assert!(
+        resp.contains("ttl = '30days'"),
+        "expected ttl = '30days' in SHOW CREATE TABLE output, got: {resp}"
+    );
+    assert!(
+        resp.contains("skip_wal = 'true'"),
+        "expected skip_wal = 'true' in SHOW CREATE TABLE output, got: {resp}"
+    );
+
+    guard.remove_all().await;
+}
+
+pub async fn test_influxdb_write_with_append_mode_hint(storage_type: StorageType) {
+    common_telemetry::init_default_ut_logging();
+    let (app, mut guard) = setup_test_http_app_with_frontend(
+        storage_type,
+        "test_influxdb_write_with_append_mode_hint",
+    )
+    .await;
+
+    let client = TestClient::new(app).await;
+
+    let result = client
+        .post("/v1/influxdb/write?db=public")
+        .header("x-greptime-hint-append_mode", "true")
+        .body("append_mode_table,host=host1 cpu=1.2 1664370459457010101")
+        .send()
+        .await;
+    assert_eq!(result.status(), 204);
+
+    let res = client
+        .get("/v1/sql?sql=show create table append_mode_table")
+        .send()
+        .await;
+    assert_eq!(res.status(), StatusCode::OK);
+    let resp = res.text().await;
+    assert!(
+        resp.contains("append_mode = 'true'"),
+        "expected append_mode = 'true' in SHOW CREATE TABLE output, got: {resp}"
+    );
+    assert!(
+        resp.contains("merge_mode = 'last_row'"),
+        "expected merge_mode = 'last_row' in SHOW CREATE TABLE output, got: {resp}"
+    );
+
+    guard.remove_all().await;
+}
+
 /// Test one-to-many VRL pipeline expansion.
 /// This test verifies that a VRL processor can return an array, which results in
 /// multiple output rows from a single input row.
@@ -5089,6 +5437,146 @@ pub async fn test_otlp_traces_v1(store_type: StorageType) {
     let res = client.get("/v1/sql?sql=drop table mytable;").send().await;
     assert_eq!(res.status(), StatusCode::OK);
 
+    // test trace table with custom partitions: 1
+    let trace_table_part1 = "trace_table_part1";
+    let res = send_req(
+        &client,
+        vec![
+            (
+                HeaderName::from_static("content-type"),
+                HeaderValue::from_static("application/x-protobuf"),
+            ),
+            (
+                HeaderName::from_static("x-greptime-pipeline-name"),
+                HeaderValue::from_static(GREPTIME_INTERNAL_TRACE_PIPELINE_V1_NAME),
+            ),
+            (
+                HeaderName::from_static("x-greptime-trace-table-name"),
+                HeaderValue::from_static(trace_table_part1),
+            ),
+            (
+                HeaderName::from_static("x-greptime-hints"),
+                HeaderValue::from_static("trace_table_partitions=1"),
+            ),
+        ],
+        "/v1/otlp/v1/traces",
+        body.clone(),
+        false,
+    )
+    .await;
+    assert_eq!(StatusCode::OK, res.status());
+    let expected_ddl = r#"[["trace_table_part1","CREATE TABLE IF NOT EXISTS \"trace_table_part1\" (\n  \"timestamp\" TIMESTAMP(9) NOT NULL,\n  \"timestamp_end\" TIMESTAMP(9) NULL,\n  \"duration_nano\" BIGINT UNSIGNED NULL,\n  \"parent_span_id\" STRING NULL SKIPPING INDEX WITH(false_positive_rate = '0.01', granularity = '10240', type = 'BLOOM'),\n  \"trace_id\" STRING NULL SKIPPING INDEX WITH(false_positive_rate = '0.01', granularity = '10240', type = 'BLOOM'),\n  \"span_id\" STRING NULL,\n  \"span_kind\" STRING NULL,\n  \"span_name\" STRING NULL,\n  \"span_status_code\" STRING NULL,\n  \"span_status_message\" STRING NULL,\n  \"trace_state\" STRING NULL,\n  \"scope_name\" STRING NULL,\n  \"scope_version\" STRING NULL,\n  \"service_name\" STRING NULL SKIPPING INDEX WITH(false_positive_rate = '0.01', granularity = '10240', type = 'BLOOM'),\n  \"span_attributes.net.peer.ip\" STRING NULL,\n  \"span_attributes.peer.service\" STRING NULL,\n  \"span_events\" JSON NULL,\n  \"span_links\" JSON NULL,\n  TIME INDEX (\"timestamp\"),\n  PRIMARY KEY (\"service_name\")\n)\n\nENGINE=mito\nWITH(\n  'comment' = 'Created on insertion',\n  append_mode = 'true',\n  table_data_model = 'greptime_trace_v1'\n)"]]"#;
+    validate_data(
+        "otlp_traces",
+        &client,
+        "show create table trace_table_part1;",
+        expected_ddl,
+    )
+    .await;
+
+    // test trace table with custom partitions: 4
+    let trace_table_part4 = "trace_table_part4";
+    let res = send_req(
+        &client,
+        vec![
+            (
+                HeaderName::from_static("content-type"),
+                HeaderValue::from_static("application/x-protobuf"),
+            ),
+            (
+                HeaderName::from_static("x-greptime-pipeline-name"),
+                HeaderValue::from_static(GREPTIME_INTERNAL_TRACE_PIPELINE_V1_NAME),
+            ),
+            (
+                HeaderName::from_static("x-greptime-trace-table-name"),
+                HeaderValue::from_static(trace_table_part4),
+            ),
+            (
+                HeaderName::from_static("x-greptime-hints"),
+                HeaderValue::from_static("trace_table_partitions=4"),
+            ),
+        ],
+        "/v1/otlp/v1/traces",
+        body.clone(),
+        false,
+    )
+    .await;
+    assert_eq!(StatusCode::OK, res.status());
+    let expected_ddl = r#"[["trace_table_part4","CREATE TABLE IF NOT EXISTS \"trace_table_part4\" (\n  \"timestamp\" TIMESTAMP(9) NOT NULL,\n  \"timestamp_end\" TIMESTAMP(9) NULL,\n  \"duration_nano\" BIGINT UNSIGNED NULL,\n  \"parent_span_id\" STRING NULL SKIPPING INDEX WITH(false_positive_rate = '0.01', granularity = '10240', type = 'BLOOM'),\n  \"trace_id\" STRING NULL SKIPPING INDEX WITH(false_positive_rate = '0.01', granularity = '10240', type = 'BLOOM'),\n  \"span_id\" STRING NULL,\n  \"span_kind\" STRING NULL,\n  \"span_name\" STRING NULL,\n  \"span_status_code\" STRING NULL,\n  \"span_status_message\" STRING NULL,\n  \"trace_state\" STRING NULL,\n  \"scope_name\" STRING NULL,\n  \"scope_version\" STRING NULL,\n  \"service_name\" STRING NULL SKIPPING INDEX WITH(false_positive_rate = '0.01', granularity = '10240', type = 'BLOOM'),\n  \"span_attributes.net.peer.ip\" STRING NULL,\n  \"span_attributes.peer.service\" STRING NULL,\n  \"span_events\" JSON NULL,\n  \"span_links\" JSON NULL,\n  TIME INDEX (\"timestamp\"),\n  PRIMARY KEY (\"service_name\")\n)\nPARTITION ON COLUMNS (\"trace_id\") (\n  trace_id < '4',\n  trace_id >= '4' AND trace_id < '8',\n  trace_id >= '8' AND trace_id < 'c',\n  trace_id >= 'c'\n)\nENGINE=mito\nWITH(\n  'comment' = 'Created on insertion',\n  append_mode = 'true',\n  table_data_model = 'greptime_trace_v1'\n)"]]"#;
+    validate_data(
+        "otlp_traces",
+        &client,
+        "show create table trace_table_part4;",
+        expected_ddl,
+    )
+    .await;
+
+    // test trace table with custom partitions: 32
+    let trace_table_part32 = "trace_table_part32";
+    let res = send_req(
+        &client,
+        vec![
+            (
+                HeaderName::from_static("content-type"),
+                HeaderValue::from_static("application/x-protobuf"),
+            ),
+            (
+                HeaderName::from_static("x-greptime-pipeline-name"),
+                HeaderValue::from_static(GREPTIME_INTERNAL_TRACE_PIPELINE_V1_NAME),
+            ),
+            (
+                HeaderName::from_static("x-greptime-trace-table-name"),
+                HeaderValue::from_static(trace_table_part32),
+            ),
+            (
+                HeaderName::from_static("x-greptime-hints"),
+                HeaderValue::from_static("trace_table_partitions=32"),
+            ),
+        ],
+        "/v1/otlp/v1/traces",
+        body.clone(),
+        false,
+    )
+    .await;
+    assert_eq!(StatusCode::OK, res.status());
+    let expected_ddl = r#"[["trace_table_part32","CREATE TABLE IF NOT EXISTS \"trace_table_part32\" (\n  \"timestamp\" TIMESTAMP(9) NOT NULL,\n  \"timestamp_end\" TIMESTAMP(9) NULL,\n  \"duration_nano\" BIGINT UNSIGNED NULL,\n  \"parent_span_id\" STRING NULL SKIPPING INDEX WITH(false_positive_rate = '0.01', granularity = '10240', type = 'BLOOM'),\n  \"trace_id\" STRING NULL SKIPPING INDEX WITH(false_positive_rate = '0.01', granularity = '10240', type = 'BLOOM'),\n  \"span_id\" STRING NULL,\n  \"span_kind\" STRING NULL,\n  \"span_name\" STRING NULL,\n  \"span_status_code\" STRING NULL,\n  \"span_status_message\" STRING NULL,\n  \"trace_state\" STRING NULL,\n  \"scope_name\" STRING NULL,\n  \"scope_version\" STRING NULL,\n  \"service_name\" STRING NULL SKIPPING INDEX WITH(false_positive_rate = '0.01', granularity = '10240', type = 'BLOOM'),\n  \"span_attributes.net.peer.ip\" STRING NULL,\n  \"span_attributes.peer.service\" STRING NULL,\n  \"span_events\" JSON NULL,\n  \"span_links\" JSON NULL,\n  TIME INDEX (\"timestamp\"),\n  PRIMARY KEY (\"service_name\")\n)\nPARTITION ON COLUMNS (\"trace_id\") (\n  trace_id < '08',\n  trace_id >= '08' AND trace_id < '10',\n  trace_id >= '10' AND trace_id < '18',\n  trace_id >= '18' AND trace_id < '20',\n  trace_id >= '20' AND trace_id < '28',\n  trace_id >= '28' AND trace_id < '30',\n  trace_id >= '30' AND trace_id < '38',\n  trace_id >= '38' AND trace_id < '40',\n  trace_id >= '40' AND trace_id < '48',\n  trace_id >= '48' AND trace_id < '50',\n  trace_id >= '50' AND trace_id < '58',\n  trace_id >= '58' AND trace_id < '60',\n  trace_id >= '60' AND trace_id < '68',\n  trace_id >= '68' AND trace_id < '70',\n  trace_id >= '70' AND trace_id < '78',\n  trace_id >= '78' AND trace_id < '80',\n  trace_id >= '80' AND trace_id < '88',\n  trace_id >= '88' AND trace_id < '90',\n  trace_id >= '90' AND trace_id < '98',\n  trace_id >= '98' AND trace_id < 'a0',\n  trace_id >= 'a0' AND trace_id < 'a8',\n  trace_id >= 'a8' AND trace_id < 'b0',\n  trace_id >= 'b0' AND trace_id < 'b8',\n  trace_id >= 'b8' AND trace_id < 'c0',\n  trace_id >= 'c0' AND trace_id < 'c8',\n  trace_id >= 'c8' AND trace_id < 'd0',\n  trace_id >= 'd0' AND trace_id < 'd8',\n  trace_id >= 'd8' AND trace_id < 'e0',\n  trace_id >= 'e0' AND trace_id < 'e8',\n  trace_id >= 'e8' AND trace_id < 'f0',\n  trace_id >= 'f0' AND trace_id < 'f8',\n  trace_id >= 'f8'\n)\nENGINE=mito\nWITH(\n  'comment' = 'Created on insertion',\n  append_mode = 'true',\n  table_data_model = 'greptime_trace_v1'\n)"]]"#;
+    validate_data(
+        "otlp_traces",
+        &client,
+        "show create table trace_table_part32;",
+        expected_ddl,
+    )
+    .await;
+
+    // invalid partition count
+    let trace_table_part_abc = "trace_table_part_abc";
+    let res = send_req(
+        &client,
+        vec![
+            (
+                HeaderName::from_static("content-type"),
+                HeaderValue::from_static("application/x-protobuf"),
+            ),
+            (
+                HeaderName::from_static("x-greptime-pipeline-name"),
+                HeaderValue::from_static(GREPTIME_INTERNAL_TRACE_PIPELINE_V1_NAME),
+            ),
+            (
+                HeaderName::from_static("x-greptime-trace-table-name"),
+                HeaderValue::from_static(trace_table_part_abc),
+            ),
+            (
+                HeaderName::from_static("x-greptime-hints"),
+                HeaderValue::from_static("trace_table_partitions=abc"),
+            ),
+        ],
+        "/v1/otlp/v1/traces",
+        body.clone(),
+        false,
+    )
+    .await;
+    assert_eq!(StatusCode::BAD_REQUEST, res.status());
+
     // write traces data with gzip
     let res = send_req(
         &client,
@@ -5122,6 +5610,561 @@ pub async fn test_otlp_traces_v1(store_type: StorageType) {
     )
     .await;
 
+    let coercion_table_name = "trace_type_coercion";
+    let coercion_req = make_trace_v1_request(
+        "type-coercion",
+        vec![make_trace_v1_span(
+            "00000000000000000000000000000001",
+            "0000000000000001",
+            "coercion-seed",
+            1_736_480_942_444_376_000,
+            1_736_480_942_444_499_000,
+            vec![
+                make_double_attr("attr_float", 1.5),
+                make_int_attr("attr_int", 10),
+                make_bool_attr("attr_bool", true),
+            ],
+        )],
+    );
+    let res = send_trace_v1_req(&client, coercion_table_name, coercion_req, false).await;
+    assert_eq!(StatusCode::OK, res.status());
+
+    let coercion_req = make_trace_v1_request(
+        "type-coercion",
+        vec![make_trace_v1_span(
+            "00000000000000000000000000000002",
+            "0000000000000002",
+            "coercion-apply",
+            1_736_480_942_444_589_000,
+            1_736_480_942_444_712_000,
+            vec![
+                make_int_attr("attr_float", 2),
+                make_string_attr("attr_int", "20"),
+                make_string_attr("attr_bool", "false"),
+            ],
+        )],
+    );
+    let res = send_trace_v1_req(&client, coercion_table_name, coercion_req, false).await;
+    assert_eq!(StatusCode::OK, res.status());
+
+    let string_target_table_name = "trace_type_coercion_to_string";
+    let string_target_seed_req = make_trace_v1_request(
+        "type-coercion-string",
+        vec![make_trace_v1_span(
+            "00000000000000000000000000000021",
+            "0000000000000021",
+            "string-target-seed",
+            1_736_480_942_444_720_000,
+            1_736_480_942_444_820_000,
+            vec![
+                make_string_attr("attr_int", "seed"),
+                make_string_attr("attr_float", "seed"),
+                make_string_attr("attr_bool", "seed"),
+            ],
+        )],
+    );
+    let res = send_trace_v1_req(
+        &client,
+        string_target_table_name,
+        string_target_seed_req,
+        false,
+    )
+    .await;
+    assert_eq!(StatusCode::OK, res.status());
+
+    let string_target_req = make_trace_v1_request(
+        "type-coercion-string",
+        vec![make_trace_v1_span(
+            "00000000000000000000000000000022",
+            "0000000000000022",
+            "string-target-apply",
+            1_736_480_942_444_830_000,
+            1_736_480_942_444_930_000,
+            vec![
+                make_int_attr("attr_int", 20),
+                make_double_attr("attr_float", 2.5),
+                make_bool_attr("attr_bool", false),
+            ],
+        )],
+    );
+    let res = send_trace_v1_req(&client, string_target_table_name, string_target_req, false).await;
+    assert_eq!(StatusCode::OK, res.status());
+
+    validate_data(
+        "otlp_traces_v1_type_coercion_to_string_rows",
+        &client,
+        &format!(
+            "select trace_id, \"span_attributes.attr_bool\", \"span_attributes.attr_float\", \"span_attributes.attr_int\" from {} order by trace_id;",
+            string_target_table_name
+        ),
+        r#"[["00000000000000000000000000000021","seed","seed","seed"],["00000000000000000000000000000022","false","2.5","20"]]"#,
+    )
+    .await;
+    validate_data(
+        "otlp_traces_v1_type_coercion_to_string_schema",
+        &client,
+        "select column_name, lower(data_type), semantic_type from information_schema.columns where table_name = 'trace_type_coercion_to_string' and column_name in ('span_attributes.attr_bool', 'span_attributes.attr_float', 'span_attributes.attr_int') order by column_name;",
+        r#"[["span_attributes.attr_bool","string","FIELD"],["span_attributes.attr_float","string","FIELD"],["span_attributes.attr_int","string","FIELD"]]"#,
+    )
+    .await;
+
+    let intra_batch_prefer_non_string_table_name = "trace_type_prefer_non_string";
+    let intra_batch_prefer_non_string_req = make_trace_v1_request(
+        "type-prefer-non-string",
+        vec![
+            make_trace_v1_span(
+                "00000000000000000000000000000031",
+                "0000000000000031",
+                "prefer-non-string-seed",
+                1_736_480_942_444_940_000,
+                1_736_480_942_445_040_000,
+                vec![
+                    make_string_attr("attr_int", "10"),
+                    make_string_attr("attr_float", "1.5"),
+                    make_string_attr("attr_bool", "true"),
+                ],
+            ),
+            make_trace_v1_span(
+                "00000000000000000000000000000032",
+                "0000000000000032",
+                "prefer-non-string-apply",
+                1_736_480_942_445_050_000,
+                1_736_480_942_445_150_000,
+                vec![
+                    make_int_attr("attr_int", 20),
+                    make_double_attr("attr_float", 2.5),
+                    make_bool_attr("attr_bool", false),
+                ],
+            ),
+        ],
+    );
+    let res = send_trace_v1_req(
+        &client,
+        intra_batch_prefer_non_string_table_name,
+        intra_batch_prefer_non_string_req,
+        false,
+    )
+    .await;
+    assert_eq!(StatusCode::OK, res.status());
+
+    validate_data(
+        "otlp_traces_v1_prefer_non_string_rows",
+        &client,
+        &format!(
+            "select trace_id, \"span_attributes.attr_bool\", \"span_attributes.attr_float\", \"span_attributes.attr_int\" from {} order by trace_id;",
+            intra_batch_prefer_non_string_table_name
+        ),
+        r#"[["00000000000000000000000000000031",true,1.5,10],["00000000000000000000000000000032",false,2.5,20]]"#,
+    )
+    .await;
+    validate_data(
+        "otlp_traces_v1_prefer_non_string_schema",
+        &client,
+        "select column_name, lower(data_type), semantic_type from information_schema.columns where table_name = 'trace_type_prefer_non_string' and column_name in ('span_attributes.attr_bool', 'span_attributes.attr_float', 'span_attributes.attr_int') order by column_name;",
+        r#"[["span_attributes.attr_bool","boolean","FIELD"],["span_attributes.attr_float","double","FIELD"],["span_attributes.attr_int","bigint","FIELD"]]"#,
+    )
+    .await;
+
+    let existing_float_table_name = "trace_type_existing_float_prefers_schema";
+    let existing_float_seed_req = make_trace_v1_request(
+        "type-existing-float",
+        vec![make_trace_v1_span(
+            "00000000000000000000000000000041",
+            "0000000000000041",
+            "existing-float-seed",
+            1_736_480_942_445_160_000,
+            1_736_480_942_445_260_000,
+            vec![make_double_attr("attr_num", 1.25)],
+        )],
+    );
+    let res = send_trace_v1_req(
+        &client,
+        existing_float_table_name,
+        existing_float_seed_req,
+        false,
+    )
+    .await;
+    assert_eq!(StatusCode::OK, res.status());
+
+    let existing_float_req = make_trace_v1_request(
+        "type-existing-float",
+        vec![
+            make_trace_v1_span(
+                "00000000000000000000000000000042",
+                "0000000000000042",
+                "existing-float-int-first",
+                1_736_480_942_445_270_000,
+                1_736_480_942_445_370_000,
+                vec![make_int_attr("attr_num", 2)],
+            ),
+            make_trace_v1_span(
+                "00000000000000000000000000000043",
+                "0000000000000043",
+                "existing-float-float-later",
+                1_736_480_942_445_380_000,
+                1_736_480_942_445_480_000,
+                vec![make_double_attr("attr_num", 3.5)],
+            ),
+        ],
+    );
+    let res = send_trace_v1_req(
+        &client,
+        existing_float_table_name,
+        existing_float_req,
+        false,
+    )
+    .await;
+    assert_eq!(StatusCode::OK, res.status());
+
+    validate_data(
+        "otlp_traces_v1_existing_float_prefers_schema_rows",
+        &client,
+        &format!(
+            "select trace_id, \"span_attributes.attr_num\" from {} order by trace_id;",
+            existing_float_table_name
+        ),
+        r#"[["00000000000000000000000000000041",1.25],["00000000000000000000000000000042",2.0],["00000000000000000000000000000043",3.5]]"#,
+    )
+    .await;
+    validate_data(
+        "otlp_traces_v1_existing_float_prefers_schema_type",
+        &client,
+        "select column_name, lower(data_type), semantic_type from information_schema.columns where table_name = 'trace_type_existing_float_prefers_schema' and column_name = 'span_attributes.attr_num';",
+        r#"[["span_attributes.attr_num","double","FIELD"]]"#,
+    )
+    .await;
+
+    let existing_int_table_name = "trace_type_existing_int_widens_to_float";
+    let existing_int_seed_req = make_trace_v1_request(
+        "type-existing-int",
+        vec![make_trace_v1_span(
+            "00000000000000000000000000000051",
+            "0000000000000051",
+            "existing-int-seed",
+            1_736_480_942_445_490_000,
+            1_736_480_942_445_590_000,
+            vec![make_int_attr("attr_num", 1)],
+        )],
+    );
+    let res = send_trace_v1_req(
+        &client,
+        existing_int_table_name,
+        existing_int_seed_req,
+        false,
+    )
+    .await;
+    assert_eq!(StatusCode::OK, res.status());
+
+    let existing_int_req = make_trace_v1_request(
+        "type-existing-int",
+        vec![
+            make_trace_v1_span(
+                "00000000000000000000000000000052",
+                "0000000000000052",
+                "existing-int-upcast-int",
+                1_736_480_942_445_600_000,
+                1_736_480_942_445_700_000,
+                vec![make_int_attr("attr_num", 2)],
+            ),
+            make_trace_v1_span(
+                "00000000000000000000000000000053",
+                "0000000000000053",
+                "existing-int-upcast-float",
+                1_736_480_942_445_710_000,
+                1_736_480_942_445_810_000,
+                vec![make_double_attr("attr_num", 3.5)],
+            ),
+        ],
+    );
+    let res = send_trace_v1_req(&client, existing_int_table_name, existing_int_req, false).await;
+    assert_eq!(StatusCode::OK, res.status());
+
+    validate_data(
+        "otlp_traces_v1_existing_int_widens_rows",
+        &client,
+        &format!(
+            "select trace_id, \"span_attributes.attr_num\" from {} order by trace_id;",
+            existing_int_table_name
+        ),
+        r#"[["00000000000000000000000000000051",1.0],["00000000000000000000000000000052",2.0],["00000000000000000000000000000053",3.5]]"#,
+    )
+    .await;
+    validate_data(
+        "otlp_traces_v1_existing_int_widens_type",
+        &client,
+        "select column_name, lower(data_type), semantic_type from information_schema.columns where table_name = 'trace_type_existing_int_widens_to_float' and column_name = 'span_attributes.attr_num';",
+        r#"[["span_attributes.attr_num","double","FIELD"]]"#,
+    )
+    .await;
+
+    let existing_int_atomic_table_name = "trace_type_existing_int_widen_atomic";
+    let existing_int_atomic_seed_req = make_trace_v1_request(
+        "type-existing-int-atomic",
+        vec![make_trace_v1_span(
+            "00000000000000000000000000000054",
+            "0000000000000054",
+            "existing-int-atomic-seed",
+            1_736_480_942_445_720_000,
+            1_736_480_942_445_820_000,
+            vec![
+                make_int_attr("attr_num", 1),
+                make_int_attr("attr_parse", 10),
+            ],
+        )],
+    );
+    let res = send_trace_v1_req(
+        &client,
+        existing_int_atomic_table_name,
+        existing_int_atomic_seed_req,
+        false,
+    )
+    .await;
+    assert_eq!(StatusCode::OK, res.status());
+
+    let existing_int_atomic_req = make_trace_v1_request(
+        "type-existing-int-atomic",
+        vec![make_trace_v1_span(
+            "00000000000000000000000000000055",
+            "0000000000000055",
+            "existing-int-atomic-invalid",
+            1_736_480_942_445_830_000,
+            1_736_480_942_445_930_000,
+            vec![
+                make_double_attr("attr_num", 3.5),
+                make_string_attr("attr_parse", "not_a_number"),
+            ],
+        )],
+    );
+    let res = send_trace_v1_req(
+        &client,
+        existing_int_atomic_table_name,
+        existing_int_atomic_req,
+        false,
+    )
+    .await;
+    assert_eq!(StatusCode::BAD_REQUEST, res.status());
+    let body = res.bytes().await;
+    let status = GoogleRpcStatus::decode(body.as_ref()).unwrap();
+    assert!(
+        status
+            .message
+            .contains("Accepted 0 spans, rejected 1 spans"),
+        "unexpected error body: {status:?}"
+    );
+
+    validate_data(
+        "otlp_traces_v1_existing_int_widen_atomic_rows",
+        &client,
+        &format!(
+            "select trace_id, \"span_attributes.attr_num\", \"span_attributes.attr_parse\" from {} order by trace_id;",
+            existing_int_atomic_table_name
+        ),
+        r#"[["00000000000000000000000000000054",1,10]]"#,
+    )
+    .await;
+    validate_data(
+        "otlp_traces_v1_existing_int_widen_atomic_types",
+        &client,
+        "select column_name, lower(data_type), semantic_type from information_schema.columns where table_name = 'trace_type_existing_int_widen_atomic' and column_name in ('span_attributes.attr_num', 'span_attributes.attr_parse') order by column_name;",
+        r#"[["span_attributes.attr_num","bigint","FIELD"],["span_attributes.attr_parse","bigint","FIELD"]]"#,
+    )
+    .await;
+
+    let existing_int_float_only_table_name = "trace_type_existing_int_float_only";
+    let existing_int_float_only_seed_req = make_trace_v1_request(
+        "type-existing-int-float-only",
+        vec![make_trace_v1_span(
+            "00000000000000000000000000000061",
+            "0000000000000061",
+            "existing-int-float-only-seed",
+            1_736_480_942_445_820_000,
+            1_736_480_942_445_920_000,
+            vec![make_int_attr("attr_num", 1)],
+        )],
+    );
+    let res = send_trace_v1_req(
+        &client,
+        existing_int_float_only_table_name,
+        existing_int_float_only_seed_req,
+        false,
+    )
+    .await;
+    assert_eq!(StatusCode::OK, res.status());
+
+    let existing_int_float_only_req = make_trace_v1_request(
+        "type-existing-int-float-only",
+        vec![make_trace_v1_span(
+            "00000000000000000000000000000062",
+            "0000000000000062",
+            "existing-int-float-only-apply",
+            1_736_480_942_445_930_000,
+            1_736_480_942_446_030_000,
+            vec![make_double_attr("attr_num", 2.5)],
+        )],
+    );
+    let res = send_trace_v1_req(
+        &client,
+        existing_int_float_only_table_name,
+        existing_int_float_only_req,
+        false,
+    )
+    .await;
+    assert_eq!(StatusCode::OK, res.status());
+
+    validate_data(
+        "otlp_traces_v1_existing_int_float_only_rows",
+        &client,
+        &format!(
+            "select trace_id, \"span_attributes.attr_num\" from {} order by trace_id;",
+            existing_int_float_only_table_name
+        ),
+        r#"[["00000000000000000000000000000061",1.0],["00000000000000000000000000000062",2.5]]"#,
+    )
+    .await;
+    validate_data(
+        "otlp_traces_v1_existing_int_float_only_type",
+        &client,
+        "select column_name, lower(data_type), semantic_type from information_schema.columns where table_name = 'trace_type_existing_int_float_only' and column_name = 'span_attributes.attr_num';",
+        r#"[["span_attributes.attr_num","double","FIELD"]]"#,
+    )
+    .await;
+
+    validate_data(
+        "otlp_traces_v1_type_coercion_rows",
+        &client,
+        &format!(
+            "select trace_id, \"span_attributes.attr_int\", \"span_attributes.attr_bool\" from {} order by trace_id;",
+            coercion_table_name
+        ),
+        r#"[["00000000000000000000000000000001",10,true],["00000000000000000000000000000002",20,false]]"#,
+    )
+    .await;
+    validate_data(
+        "otlp_traces_v1_type_coercion_float_sum",
+        &client,
+        &format!(
+            "select sum(\"span_attributes.attr_float\") from {};",
+            coercion_table_name
+        ),
+        r#"[[3.5]]"#,
+    )
+    .await;
+    validate_data(
+        "otlp_traces_v1_type_coercion_schema",
+        &client,
+        "select column_name, lower(data_type), semantic_type from information_schema.columns where table_name = 'trace_type_coercion' and column_name in ('span_attributes.attr_bool', 'span_attributes.attr_float', 'span_attributes.attr_int') order by column_name;",
+        r#"[["span_attributes.attr_bool","boolean","FIELD"],["span_attributes.attr_float","double","FIELD"],["span_attributes.attr_int","bigint","FIELD"]]"#,
+    )
+    .await;
+
+    let abort_table_name = "trace_type_abort";
+    let abort_seed_req = make_trace_v1_request(
+        "type-abort",
+        vec![make_trace_v1_span(
+            "00000000000000000000000000000011",
+            "0000000000000011",
+            "abort-seed",
+            1_736_480_942_444_800_000,
+            1_736_480_942_444_900_000,
+            vec![make_int_attr("attr_int", 10)],
+        )],
+    );
+    let res = send_trace_v1_req(&client, abort_table_name, abort_seed_req, false).await;
+    assert_eq!(StatusCode::OK, res.status());
+
+    let abort_req = make_trace_v1_request(
+        "type-abort",
+        vec![
+            make_trace_v1_span(
+                "00000000000000000000000000000012",
+                "0000000000000012",
+                "abort-parseable",
+                1_736_480_942_445_000_000,
+                1_736_480_942_445_100_000,
+                vec![make_string_attr("attr_int", "20")],
+            ),
+            make_trace_v1_span(
+                "00000000000000000000000000000013",
+                "0000000000000013",
+                "abort-unparsable",
+                1_736_480_942_445_200_000,
+                1_736_480_942_445_300_000,
+                vec![make_string_attr("attr_int", "not_a_number")],
+            ),
+        ],
+    );
+    let res = send_trace_v1_req(&client, abort_table_name, abort_req, false).await;
+    assert_eq!(StatusCode::OK, res.status());
+    let body = ExportTraceServiceResponse::decode(res.bytes().await).unwrap();
+    let partial_success = body.partial_success.as_ref().unwrap();
+    assert_eq!(partial_success.rejected_spans, 1);
+    assert!(
+        partial_success
+            .error_message
+            .contains("Accepted 1 spans, rejected 1 spans"),
+        "unexpected partial success body: {body:?}"
+    );
+    assert!(
+        partial_success.error_message.contains(
+            "Rejected span 00000000000000000000000000000013:0000000000000013 (InvalidArguments)"
+        ),
+        "unexpected partial success body: {body:?}"
+    );
+
+    validate_data(
+        "otlp_traces_v1_type_abort_rows",
+        &client,
+        &format!(
+            "select trace_id, \"span_attributes.attr_int\" from {} order by trace_id",
+            abort_table_name
+        ),
+        r#"[["00000000000000000000000000000011",10],["00000000000000000000000000000012",20]]"#,
+    )
+    .await;
+
+    let chunk_failure_req = make_trace_v1_request(
+        "type-discard",
+        vec![
+            make_trace_v1_span(
+                "00000000000000000000000000000021",
+                "0000000000000021",
+                "discard-one",
+                1_736_480_942_445_400_000,
+                1_736_480_942_445_500_000,
+                vec![make_string_attr("attr_text", "alpha")],
+            ),
+            make_trace_v1_span(
+                "00000000000000000000000000000022",
+                "0000000000000022",
+                "discard-two",
+                1_736_480_942_445_600_000,
+                1_736_480_942_445_700_000,
+                vec![make_string_attr("attr_text", "beta")],
+            ),
+        ],
+    );
+    let res = send_trace_v1_req_with_db(
+        &client,
+        "nonexistent",
+        "trace_chunk_discard",
+        chunk_failure_req,
+        false,
+    )
+    .await;
+    assert_eq!(StatusCode::BAD_REQUEST, res.status());
+    let body = GoogleRpcStatus::decode(res.bytes().await.as_ref()).unwrap();
+    assert!(
+        body.message.contains("Accepted 0 spans, rejected 2 spans"),
+        "unexpected error body: {body:?}"
+    );
+    assert!(
+        body.message.contains("Chunk fallback triggered by")
+            || body
+                .message
+                .contains("Discarded 2 spans after ambiguous chunk failure"),
+        "unexpected error body: {body:?}"
+    );
+
     guard.remove_all().await;
 }
 
@@ -7354,6 +8397,129 @@ async fn send_req(
     req.header("content-length", len).send().await
 }
 
+async fn send_trace_v1_req(
+    client: &TestClient,
+    table_name: &str,
+    req: ExportTraceServiceRequest,
+    with_gzip: bool,
+) -> TestResponse {
+    send_trace_v1_req_with_db(client, "public", table_name, req, with_gzip).await
+}
+
+async fn send_trace_v1_req_with_db(
+    client: &TestClient,
+    db_name: &str,
+    table_name: &str,
+    req: ExportTraceServiceRequest,
+    with_gzip: bool,
+) -> TestResponse {
+    send_req(
+        client,
+        vec![
+            (
+                HeaderName::from_static("content-type"),
+                HeaderValue::from_static("application/x-protobuf"),
+            ),
+            (
+                HeaderName::from_static("x-greptime-pipeline-name"),
+                HeaderValue::from_static(GREPTIME_INTERNAL_TRACE_PIPELINE_V1_NAME),
+            ),
+            (
+                HeaderName::from_static("x-greptime-trace-table-name"),
+                HeaderValue::from_str(table_name).unwrap(),
+            ),
+            (
+                GREPTIME_DB_HEADER_NAME.clone(),
+                HeaderValue::from_str(db_name).unwrap(),
+            ),
+        ],
+        "/v1/otlp/v1/traces",
+        req.encode_to_vec(),
+        with_gzip,
+    )
+    .await
+}
+
+fn make_trace_v1_request(service_name: &str, spans: Vec<Value>) -> ExportTraceServiceRequest {
+    serde_json::from_value(json!({
+        "resourceSpans": [{
+            "resource": {
+                "attributes": [{
+                    "key": "service.name",
+                    "value": { "stringValue": service_name }
+                }]
+            },
+            "scopeSpans": [{
+                "scope": {
+                    "name": "trace-v1-type-tests"
+                },
+                "spans": spans
+            }],
+            "schemaUrl": "https://opentelemetry.io/schemas/1.4.0"
+        }]
+    }))
+    .unwrap()
+}
+
+fn make_trace_v1_span(
+    trace_id: &str,
+    span_id: &str,
+    name: &str,
+    start_time_unix_nano: i64,
+    end_time_unix_nano: i64,
+    attributes: Vec<Value>,
+) -> Value {
+    json!({
+        "traceId": trace_id,
+        "spanId": span_id,
+        "name": name,
+        "kind": 2,
+        "startTimeUnixNano": start_time_unix_nano.to_string(),
+        "endTimeUnixNano": end_time_unix_nano.to_string(),
+        "attributes": attributes,
+        "status": {
+            "message": "",
+            "code": 0
+        }
+    })
+}
+
+fn make_string_attr(key: &str, value: &str) -> Value {
+    json!({
+        "key": key,
+        "value": {
+            "stringValue": value
+        }
+    })
+}
+
+fn make_int_attr(key: &str, value: i64) -> Value {
+    json!({
+        "key": key,
+        "value": {
+            "intValue": value.to_string()
+        }
+    })
+}
+
+fn make_double_attr(key: &str, value: f64) -> Value {
+    json!({
+        "key": key,
+        "value": {
+            "doubleValue": value
+        }
+    })
+}
+
+fn make_bool_attr(key: &str, value: bool) -> Value {
+    json!({
+        "key": key,
+        "value": {
+            "boolValue": value
+        }
+    })
+}
+
 fn get_rows_from_output(output: &str) -> String {
     let resp: Value = serde_json::from_str(output).unwrap();
     resp.get("output")
diff --git a/tests-integration/tests/jsonbench.rs b/tests-integration/tests/jsonbench.rs
index 60f699c4ce..55cfcd53f0 100644
--- a/tests-integration/tests/jsonbench.rs
+++ b/tests-integration/tests/jsonbench.rs
@@ -25,6 +25,8 @@ use servers::server::ServerHandlers;
 use tests_integration::standalone::GreptimeDbStandaloneBuilder;
 use tests_integration::test_util::execute_sql_and_expect;
 
+// TODO(LFC): Unignore the test when JSON2 is ready.
+#[ignore]
 #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
 async fn test_load_jsonbench_data_by_pipeline() -> io::Result<()> {
     common_telemetry::init_default_ut_logging();
@@ -121,6 +123,8 @@ transform:
     assert!(response.starts_with(pattern));
 }
 
+// TODO(LFC): Unignore the test when JSON2 is ready.
+#[ignore]
 #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
 async fn test_load_jsonbench_data_by_sql() -> io::Result<()> {
     common_telemetry::init_default_ut_logging();
diff --git a/tests-integration/tests/main.rs b/tests-integration/tests/main.rs
index 14cf734291..e8b8232ad8 100644
--- a/tests-integration/tests/main.rs
+++ b/tests-integration/tests/main.rs
@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#![recursion_limit = "256"]
+
 #[macro_use]
 mod grpc;
 #[macro_use]
@@ -25,11 +27,17 @@ mod region_migration;
 mod repartition;
 #[macro_use]
 mod repartition_expr_version;
+mod mysql;
 
 grpc_tests!(File, S3, S3WithCache, Oss, Azblob, Gcs);
 
 http_tests!(File, S3, S3WithCache, Oss, Azblob, Gcs);
 
+#[tokio::test(flavor = "multi_thread")]
+async fn test_http_auth_from_standalone_user_provider_config() {
+    http::test_http_auth_from_standalone_user_provider_config().await;
+}
+
 sql_tests!(File);
 
 region_migration_tests!(File);
diff --git a/tests-integration/tests/mysql.rs b/tests-integration/tests/mysql.rs
new file mode 100644
index 0000000000..5de680b3ba
--- /dev/null
+++ b/tests-integration/tests/mysql.rs
@@ -0,0 +1,124 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use itertools::Itertools;
+use sqlx::mysql::{MySqlQueryResult, MySqlRow};
+use sqlx::{Connection, MySqlConnection, Row};
+use tests_integration::test_util::{StorageType, setup_mysql_server};
+use tokio_stream::StreamExt;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn test_mysql_multiple_statement_execution_support() -> sqlx::Result<()> {
+    let (mut guard, server) = setup_mysql_server(
+        StorageType::File,
+        "test_mysql_multiple_statement_execution_support",
+    )
+    .await;
+    let addr = server.bind_addr().unwrap();
+    let mut conn = MySqlConnection::connect(&format!("mysql://{addr}/public")).await?;
+
+    let query = "create table foo (ts timestamp time index, i int)";
+    let result = sqlx::raw_sql(query).execute(&mut conn).await?;
+    assert_eq!(result.rows_affected(), 0);
+
+    fn to_string(result: either::Either<MySqlQueryResult, MySqlRow>) -> String {
+        match result {
+            either::Left(result) => {
+                format!("OK packet (rows affected: {})", result.rows_affected())
+            }
+            either::Right(result) => {
+                format!(
+                    "Row: [{}]",
+                    (0..result.columns().len())
+                        .map(|i| {
+                            let i: i64 = result.get(i);
+                            i.to_string()
+                        })
+                        .join(", ")
+                )
+            }
+        }
+    }
+
+    let query = "insert into foo values (1, 1); select i from foo";
+    let results = sqlx::raw_sql(query)
+        .fetch_many(&mut conn)
+        .collect::<sqlx::Result<Vec<_>>>()
+        .await?
+        .into_iter()
+        .map(to_string)
+        .join("\n");
+    let expected = r#"
+OK packet (rows affected: 1)
+Row: [1]
+OK packet (rows affected: 0)
+"#;
+    assert_eq!(results, expected.trim());
+
+    let query = "insert into foo values (2, 2); insert into foo values (3, 3)";
+    let results = sqlx::raw_sql(query)
+        .fetch_many(&mut conn)
+        .collect::<sqlx::Result<Vec<_>>>()
+        .await?
+        .into_iter()
+        .map(to_string)
+        .join("\n");
+    let expected = r#"
+OK packet (rows affected: 1)
+OK packet (rows affected: 1)
+"#;
+    assert_eq!(results, expected.trim());
+
+    let query = "select i from foo order by i; select sum(i) from foo";
+    let results = sqlx::raw_sql(query)
+        .fetch_many(&mut conn)
+        .collect::<sqlx::Result<Vec<_>>>()
+        .await?
+        .into_iter()
+        .map(to_string)
+        .join("\n");
+    let expected = r#"
+Row: [1]
+Row: [2]
+Row: [3]
+OK packet (rows affected: 0)
+Row: [6]
+OK packet (rows affected: 0)
+"#;
+    assert_eq!(results, expected.trim());
+
+    let query = "select i from foo; select i from bar";
+    let result = sqlx::raw_sql(query)
+        .fetch_many(&mut conn)
+        .collect::<sqlx::Result<Vec<_>>>()
+        .await
+        .unwrap_err()
+        .to_string();
+    let expected = r#"error returned from database: 1146 (42S02): (TableNotFound): Failed to plan SQL: Table not found: greptime.public.bar"#;
+    assert_eq!(result, expected);
+
+    let query = "select i from bar; select i from foo";
+    let result = sqlx::raw_sql(query)
+        .fetch_many(&mut conn)
+        .collect::<sqlx::Result<Vec<_>>>()
+        .await
+        .unwrap_err()
+        .to_string();
+    let expected = r#"error returned from database: 1146 (42S02): (TableNotFound): Failed to plan SQL: Table not found: greptime.public.bar"#;
+    assert_eq!(result, expected);
+
+    let _ = server.shutdown().await;
+    guard.remove_all().await;
+    Ok(())
+}
diff --git a/tests-integration/tests/sql.rs b/tests-integration/tests/sql.rs
index 28f8535ec2..14972d900b 100644
--- a/tests-integration/tests/sql.rs
+++ b/tests-integration/tests/sql.rs
@@ -82,8 +82,11 @@ macro_rules! sql_tests {
                 test_postgres_datestyle,
                 test_postgres_intervalstyle,
                 test_postgres_parameter_inference,
+                test_postgres_uint64_parameter,
+                test_postgres_explain_bind_parameter,
                 test_postgres_array_types,
                 test_mysql_prepare_stmt_insert_timestamp,
+                test_mysql_federated_prepare_stmt,
                 test_declare_fetch_close_cursor,
                 test_alter_update_on,
             );
@@ -1300,6 +1303,116 @@ pub async fn test_postgres_parameter_inference(store_type: StorageType) {
     guard.remove_all().await;
 }
 
+pub async fn test_postgres_uint64_parameter(store_type: StorageType) {
+    let (mut guard, fe_pg_server) =
+        setup_pg_server(store_type, "test_postgres_uint64_parameter").await;
+    let addr = fe_pg_server.bind_addr().unwrap().to_string();
+
+    let (client, connection) = tokio_postgres::connect(&format!("postgres://{addr}/public"), NoTls)
+        .await
+        .unwrap();
+
+    let (tx, rx) = tokio::sync::oneshot::channel();
+    tokio::spawn(async move {
+        connection.await.unwrap();
+        tx.send(()).unwrap();
+    });
+
+    let _ = client
+        .simple_query("create table demo_u64(v bigint unsigned, ts timestamp time index)")
+        .await
+        .unwrap();
+
+    let dt = NaiveDate::from_yo_opt(2015, 100)
+        .unwrap()
+        .and_hms_opt(0, 0, 0)
+        .unwrap();
+    let _ = client
+        .execute(
+            "INSERT INTO demo_u64 VALUES($1, $2)",
+            &[&Decimal::from(123456u64), &dt],
+        )
+        .await
+        .unwrap();
+
+    let rows = client
+        .query(
+            "SELECT count(*) FROM demo_u64 WHERE v = $1",
+            &[&Decimal::from(123456u64)],
+        )
+        .await
+        .unwrap();
+
+    assert_eq!(1, rows.len());
+    let count: i64 = rows[0].get(0);
+    assert_eq!(count, 1);
+
+    drop(client);
+    rx.await.unwrap();
+
+    let _ = fe_pg_server.shutdown().await;
+    guard.remove_all().await;
+}
+
+pub async fn test_postgres_explain_bind_parameter(store_type: StorageType) {
+    // Regression test for #8029: EXPLAIN / EXPLAIN ANALYZE must accept bind
+    // parameters over the Postgres extended query protocol.
+    let (mut guard, fe_pg_server) =
+        setup_pg_server(store_type, "test_postgres_explain_bind_parameter").await;
+    let addr = fe_pg_server.bind_addr().unwrap().to_string();
+
+    let (client, connection) = tokio_postgres::connect(&format!("postgres://{addr}/public"), NoTls)
+        .await
+        .unwrap();
+
+    let (tx, rx) = tokio::sync::oneshot::channel();
+    tokio::spawn(async move {
+        connection.await.unwrap();
+        tx.send(()).unwrap();
+    });
+
+    let _ = client
+        .simple_query(
+            "create table t (k varchar(36) not null, ts timestamp(3) not null, time index(ts))",
+        )
+        .await
+        .unwrap();
+    let _ = client
+        .simple_query("insert into t (k, ts) values ('a', 1), ('b', 2), ('c', 3)")
+        .await
+        .unwrap();
+
+    // Sanity check: the underlying SELECT with a bind parameter works.
+    let rows = client
+        .query("SELECT k FROM t WHERE k = $1", &[&"a"])
+        .await
+        .unwrap();
+    assert_eq!(1, rows.len());
+
+    // EXPLAIN with a bind parameter must succeed.
+    let rows = client
+        .query("EXPLAIN SELECT k FROM t WHERE k = $1", &[&"a"])
+        .await
+        .unwrap();
+    assert!(!rows.is_empty(), "EXPLAIN should produce at least one row");
+
+    // EXPLAIN ANALYZE with a bind parameter must also succeed.
+    let rows = client
+        .query("EXPLAIN ANALYZE SELECT k FROM t WHERE k = $1", &[&"a"])
+        .await
+        .unwrap();
+    assert!(
+        !rows.is_empty(),
+        "EXPLAIN ANALYZE should produce at least one row"
+    );
+
+    drop(client);
+    rx.await.unwrap();
+
+    let _ = fe_pg_server.shutdown().await;
+    guard.remove_all().await;
+}
+
 pub async fn test_mysql_async_timestamp(store_type: StorageType) {
     use mysql_async::prelude::*;
     use time::PrimitiveDateTime;
@@ -1534,6 +1647,45 @@ pub async fn test_mysql_prepare_stmt_insert_timestamp(store_type: StorageType) {
     guard.remove_all().await;
 }
 
+pub async fn test_mysql_federated_prepare_stmt(store_type: StorageType) {
+    common_telemetry::init_default_ut_logging();
+
+    let (mut guard, fe_mysql_server) =
+        setup_mysql_server(store_type, "test_mysql_federated_prepare_stmt").await;
+    let addr = fe_mysql_server.bind_addr().unwrap().to_string();
+
+    let pool = MySqlPoolOptions::new()
+        .max_connections(2)
+        .connect(&format!("mysql://{addr}/public"))
+        .await
+        .unwrap();
+
+    // sqlx::query uses binary prepared statement protocol (COM_STMT_PREPARE + COM_STMT_EXECUTE)
+    // "SELECT @@version_comment" is a federated query matched by federated::check
+    let rows = sqlx::query("SELECT @@version_comment")
+        .fetch_all(&pool)
+        .await
+        .unwrap();
+    assert_eq!(rows.len(), 1);
+    let val: String = rows[0].get(0);
+    assert!(val.contains("GreptimeDB"));
+
+    // "SET NAMES utf8" is another federated pattern
+    sqlx::query("SET NAMES utf8").execute(&pool).await.unwrap();
+
+    // "SELECT @@tx_isolation" is a federated variable query
+    let rows = sqlx::query("SELECT @@tx_isolation")
+        .fetch_all(&pool)
+        .await
+        .unwrap();
+    assert_eq!(rows.len(), 1);
+    let val: String = rows[0].get(0);
+    assert_eq!(val, "REPEATABLE-READ");
+
+    let _ = fe_mysql_server.shutdown().await;
+    guard.remove_all().await;
+}
+
 pub async fn test_postgres_array_types(store_type: StorageType) {
     let (mut guard, fe_pg_server) = setup_pg_server(store_type, "test_postgres_array_types").await;
     let addr = fe_pg_server.bind_addr().unwrap().to_string();
diff --git a/tests/cases/distributed/explain/order_by.result b/tests/cases/distributed/explain/order_by.result
index 362849afea..6ce8b4e170 100644
--- a/tests/cases/distributed/explain/order_by.result
+++ b/tests/cases/distributed/explain/order_by.result
@@ -126,8 +126,7 @@ EXPLAIN ANALYZE SELECT i, t AS alias_ts FROM test_pk ORDER BY t DESC LIMIT 5;
 |_|_|_|
 | 1_| 0_|_ProjectionExec: expr=[i@0 as i, t@1 as alias_ts] REDACTED
 |_|_|_SortPreservingMergeExec: [test_pk.t__temp__0@2 DESC], fetch=5 REDACTED
-|_|_|_WindowedSortExec: expr=t@1 DESC num_ranges=REDACTED fetch=5 REDACTED
-|_|_|_PartSortExec: expr=t@1 DESC num_ranges=REDACTED limit=5 REDACTED
+|_|_|_SortExec: TopK(fetch=5), expr=[t@1 DESC], preserve_partitioning=[true], filter=[t@1 IS NULL OR t@1 > 2] REDACTED
 |_|_|_ProjectionExec: expr=[i@0 as i, t@1 as t, t@1 as test_pk.t__temp__0] REDACTED
 |_|_|_SeqScan: region=REDACTED, "partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0} REDACTED
 |_|_|_|
@@ -150,8 +149,7 @@ EXPLAIN ANALYZE SELECT i, t AS alias_ts FROM test_pk ORDER BY alias_ts DESC LIMI
 |_|_|_|
 | 1_| 0_|_ProjectionExec: expr=[i@0 as i, t@1 as alias_ts] REDACTED
 |_|_|_SortPreservingMergeExec: [t@1 DESC], fetch=5 REDACTED
-|_|_|_WindowedSortExec: expr=t@1 DESC num_ranges=REDACTED fetch=5 REDACTED
-|_|_|_PartSortExec: expr=t@1 DESC num_ranges=REDACTED limit=5 REDACTED
+|_|_|_SortExec: TopK(fetch=5), expr=[t@1 DESC], preserve_partitioning=[true], filter=[t@1 IS NULL OR t@1 > 2] REDACTED
 |_|_|_SeqScan: region=REDACTED, "partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0} REDACTED
 |_|_|_|
 |_|_| Total rows: 5_|
diff --git a/tests/cases/distributed/explain/step_aggr_advance.result b/tests/cases/distributed/explain/step_aggr_advance.result
index 4bd83b7afa..5938fa202d 100644
--- a/tests/cases/distributed/explain/step_aggr_advance.result
+++ b/tests/cases/distributed/explain/step_aggr_advance.result
@@ -442,54 +442,54 @@ Affected Rows: 0
 -- SQLNESS REPLACE (Hash.*) REDACTED
 tql explain (1752591864, 1752592164, '30s') sum by (a, b, c) (rate(aggr_optimize_not [2m])) / sum by (a, b, c) (rate(aggr_optimize_not_count [2m]));
 
-+---------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
-| plan_type     | plan                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    |
-+---------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
-| logical_plan  | Projection: aggr_optimize_not_count.a, aggr_optimize_not_count.b, aggr_optimize_not_count.c, aggr_optimize_not_count.greptime_timestamp, aggr_optimize_not.sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000))) / aggr_optimize_not_count.sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000))) AS aggr_optimize_not.sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000))) / aggr_optimize_not_count.sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000))) |
-|               |   Inner Join: aggr_optimize_not.a = aggr_optimize_not_count.a, aggr_optimize_not.b = aggr_optimize_not_count.b, aggr_optimize_not.c = aggr_optimize_not_count.c, aggr_optimize_not.greptime_timestamp = aggr_optimize_not_count.greptime_timestamp                                                                                                                                                                                                                                                                                                                                                      |
-|               |     MergeSort: aggr_optimize_not.a ASC NULLS LAST, aggr_optimize_not.b ASC NULLS LAST, aggr_optimize_not.c ASC NULLS LAST, aggr_optimize_not.greptime_timestamp ASC NULLS LAST                                                                                                                                                                                                                                                                                                                                                                                                                          |
-|               |       MergeScan [is_placeholder=false, remote_input=[                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |
-|               | SubqueryAlias: aggr_optimize_not                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
-|               |   Sort: aggr_optimize_not.a ASC NULLS LAST, aggr_optimize_not.b ASC NULLS LAST, aggr_optimize_not.c ASC NULLS LAST, aggr_optimize_not.greptime_timestamp ASC NULLS LAST                                                                                                                                                                                                                                                                                                                                                                                                                                 |
-|               |     Aggregate: groupBy=[[aggr_optimize_not.a, aggr_optimize_not.b, aggr_optimize_not.c, aggr_optimize_not.greptime_timestamp]], aggr=[[sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)))]]                                                                                                                                                                                                                                                                                                                                                                       |
-|               |       Filter: prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)) IS NOT NULL                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           |
-|               |         Projection: aggr_optimize_not.greptime_timestamp, prom_rate(greptime_timestamp_range, greptime_value, aggr_optimize_not.greptime_timestamp, Int64(120000)) AS prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)), aggr_optimize_not.a, aggr_optimize_not.b, aggr_optimize_not.c, aggr_optimize_not.d                                                                                                                                                                                                                                                           |
-|               |           PromRangeManipulate: req range=[1752591864000..1752592164000], interval=[30000], eval range=[120000], time index=[greptime_timestamp], values=["greptime_value"]                                                                                                                                                                                                                                                                                                                                                                                                                              |
-|               |             PromSeriesNormalize: offset=[0], time index=[greptime_timestamp], filter NaN: [true]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
-|               |               PromSeriesDivide: tags=["a", "b", "c", "d"]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               |
-|               |                 Sort: aggr_optimize_not.a ASC NULLS FIRST, aggr_optimize_not.b ASC NULLS FIRST, aggr_optimize_not.c ASC NULLS FIRST, aggr_optimize_not.d ASC NULLS FIRST, aggr_optimize_not.greptime_timestamp ASC NULLS FIRST                                                                                                                                                                                                                                                                                                                                                                          |
-|               |                   Filter: aggr_optimize_not.greptime_timestamp >= TimestampMillisecond(1752591744001, None) AND aggr_optimize_not.greptime_timestamp <= TimestampMillisecond(1752592164000, None)                                                                                                                                                                                                                                                                                                                                                                                                       |
-|               |                     TableScan: aggr_optimize_not                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
-|               | ]]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      |
-|               |     SubqueryAlias: aggr_optimize_not_count                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              |
-|               |       Sort: aggr_optimize_not_count.a ASC NULLS LAST, aggr_optimize_not_count.b ASC NULLS LAST, aggr_optimize_not_count.c ASC NULLS LAST, aggr_optimize_not_count.greptime_timestamp ASC NULLS LAST                                                                                                                                                                                                                                                                                                                                                                                                     |
-|               |         Aggregate: groupBy=[[aggr_optimize_not_count.a, aggr_optimize_not_count.b, aggr_optimize_not_count.c, aggr_optimize_not_count.greptime_timestamp]], aggr=[[sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)))]]                                                                                                                                                                                                                                                                                                                                           |
-|               |           Filter: prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)) IS NOT NULL                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       |
-|               |             Projection: aggr_optimize_not_count.greptime_timestamp, prom_rate(greptime_timestamp_range, greptime_value, aggr_optimize_not_count.greptime_timestamp, Int64(120000)) AS prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)), aggr_optimize_not_count.a, aggr_optimize_not_count.b, aggr_optimize_not_count.c                                                                                                                                                                                                                                              |
-|               |               PromRangeManipulate: req range=[1752591864000..1752592164000], interval=[30000], eval range=[120000], time index=[greptime_timestamp], values=["greptime_value"]                                                                                                                                                                                                                                                                                                                                                                                                                          |
-|               |                 PromSeriesNormalize: offset=[0], time index=[greptime_timestamp], filter NaN: [true]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    |
-|               |                   PromSeriesDivide: tags=["a", "b", "c", "d"]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           |
-|               |                     Sort: aggr_optimize_not_count.a ASC NULLS FIRST, aggr_optimize_not_count.b ASC NULLS FIRST, aggr_optimize_not_count.c ASC NULLS FIRST, aggr_optimize_not_count.d ASC NULLS FIRST, aggr_optimize_not_count.greptime_timestamp ASC NULLS FIRST                                                                                                                                                                                                                                                                                                                                        |
-|               |                       MergeScan [is_placeholder=false, remote_input=[                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |
-|               | Filter: aggr_optimize_not_count.greptime_timestamp >= TimestampMillisecond(1752591744001, None) AND aggr_optimize_not_count.greptime_timestamp <= TimestampMillisecond(1752592164000, None)                                                                                                                                                                                                                                                                                                                                                                                                             |
-|               |   TableScan: aggr_optimize_not_count                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    |
-|               | ]]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      |
-| physical_plan | ProjectionExec: expr=[a@0 as a, b@1 as b, c@2 as c, greptime_timestamp@3 as greptime_timestamp, sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)))@5 / sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)))@4 as aggr_optimize_not.sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000))) / aggr_optimize_not_count.sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)))]                                                                               |
++---------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| plan_type     | plan                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      |
++---------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| logical_plan  | Projection: aggr_optimize_not_count.a, aggr_optimize_not_count.b, aggr_optimize_not_count.c, aggr_optimize_not_count.greptime_timestamp, CAST(aggr_optimize_not.sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000))) AS Float64) / CAST(aggr_optimize_not_count.sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000))) AS Float64) AS aggr_optimize_not.sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000))) / aggr_optimize_not_count.sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000))) |
+|               |   Inner Join: aggr_optimize_not.a = aggr_optimize_not_count.a, aggr_optimize_not.b = aggr_optimize_not_count.b, aggr_optimize_not.c = aggr_optimize_not_count.c, aggr_optimize_not.greptime_timestamp = aggr_optimize_not_count.greptime_timestamp                                                                                                                                                                                                                                                                                                                                                                                        |
+|               |     MergeSort: aggr_optimize_not.a ASC NULLS LAST, aggr_optimize_not.b ASC NULLS LAST, aggr_optimize_not.c ASC NULLS LAST, aggr_optimize_not.greptime_timestamp ASC NULLS LAST                                                                                                                                                                                                                                                                                                                                                                                                                                                            |
+|               |       MergeScan [is_placeholder=false, remote_input=[                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     |
+|               | SubqueryAlias: aggr_optimize_not                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          |
+|               |   Sort: aggr_optimize_not.a ASC NULLS LAST, aggr_optimize_not.b ASC NULLS LAST, aggr_optimize_not.c ASC NULLS LAST, aggr_optimize_not.greptime_timestamp ASC NULLS LAST                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |
+|               |     Aggregate: groupBy=[[aggr_optimize_not.a, aggr_optimize_not.b, aggr_optimize_not.c, aggr_optimize_not.greptime_timestamp]], aggr=[[sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)))]]                                                                                                                                                                                                                                                                                                                                                                                                         |
+|               |       Filter: prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)) IS NOT NULL                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             |
+|               |         Projection: aggr_optimize_not.greptime_timestamp, prom_rate(greptime_timestamp_range, greptime_value, aggr_optimize_not.greptime_timestamp, Int64(120000)) AS prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)), aggr_optimize_not.a, aggr_optimize_not.b, aggr_optimize_not.c, aggr_optimize_not.d                                                                                                                                                                                                                                                                                             |
+|               |           PromRangeManipulate: req range=[1752591864000..1752592164000], interval=[30000], eval range=[120000], time index=[greptime_timestamp], values=["greptime_value"]                                                                                                                                                                                                                                                                                                                                                                                                                                                                |
+|               |             PromSeriesNormalize: offset=[0], time index=[greptime_timestamp], filter NaN: [true]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          |
+|               |               PromSeriesDivide: tags=["a", "b", "c", "d"]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 |
+|               |                 Sort: aggr_optimize_not.a ASC NULLS FIRST, aggr_optimize_not.b ASC NULLS FIRST, aggr_optimize_not.c ASC NULLS FIRST, aggr_optimize_not.d ASC NULLS FIRST, aggr_optimize_not.greptime_timestamp ASC NULLS FIRST                                                                                                                                                                                                                                                                                                                                                                                                            |
+|               |                   Filter: aggr_optimize_not.greptime_timestamp >= TimestampMillisecond(1752591744001, None) AND aggr_optimize_not.greptime_timestamp <= TimestampMillisecond(1752592164000, None)                                                                                                                                                                                                                                                                                                                                                                                                                                         |
+|               |                     TableScan: aggr_optimize_not                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          |
+|               | ]]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
+|               |     SubqueryAlias: aggr_optimize_not_count                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                |
+|               |       Sort: aggr_optimize_not_count.a ASC NULLS LAST, aggr_optimize_not_count.b ASC NULLS LAST, aggr_optimize_not_count.c ASC NULLS LAST, aggr_optimize_not_count.greptime_timestamp ASC NULLS LAST                                                                                                                                                                                                                                                                                                                                                                                                                                       |
+|               |         Aggregate: groupBy=[[aggr_optimize_not_count.a, aggr_optimize_not_count.b, aggr_optimize_not_count.c, aggr_optimize_not_count.greptime_timestamp]], aggr=[[sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)))]]                                                                                                                                                                                                                                                                                                                                                                             |
+|               |           Filter: prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)) IS NOT NULL                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         |
+|               |             Projection: aggr_optimize_not_count.greptime_timestamp, prom_rate(greptime_timestamp_range, greptime_value, aggr_optimize_not_count.greptime_timestamp, Int64(120000)) AS prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)), aggr_optimize_not_count.a, aggr_optimize_not_count.b, aggr_optimize_not_count.c                                                                                                                                                                                                                                                                                |
+|               |               PromRangeManipulate: req range=[1752591864000..1752592164000], interval=[30000], eval range=[120000], time index=[greptime_timestamp], values=["greptime_value"]                                                                                                                                                                                                                                                                                                                                                                                                                                                            |
+|               |                 PromSeriesNormalize: offset=[0], time index=[greptime_timestamp], filter NaN: [true]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      |
+|               |                   PromSeriesDivide: tags=["a", "b", "c", "d"]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             |
+|               |                     Sort: aggr_optimize_not_count.a ASC NULLS FIRST, aggr_optimize_not_count.b ASC NULLS FIRST, aggr_optimize_not_count.c ASC NULLS FIRST, aggr_optimize_not_count.d ASC NULLS FIRST, aggr_optimize_not_count.greptime_timestamp ASC NULLS FIRST                                                                                                                                                                                                                                                                                                                                                                          |
+|               |                       MergeScan [is_placeholder=false, remote_input=[                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     |
+|               | Filter: aggr_optimize_not_count.greptime_timestamp >= TimestampMillisecond(1752591744001, None) AND aggr_optimize_not_count.greptime_timestamp <= TimestampMillisecond(1752592164000, None)                                                                                                                                                                                                                                                                                                                                                                                                                                               |
+|               |   TableScan: aggr_optimize_not_count                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      |
+|               | ]]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
+| physical_plan | ProjectionExec: expr=[a@0 as a, b@1 as b, c@2 as c, greptime_timestamp@3 as greptime_timestamp, sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)))@5 / sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)))@4 as aggr_optimize_not.sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000))) / aggr_optimize_not_count.sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)))]                                                                                                                 |
 |               |   REDACTED
-|               |     CoalescePartitionsExec                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              |
-|               |       AggregateExec: mode=SinglePartitioned, gby=[a@2 as a, b@3 as b, c@4 as c, greptime_timestamp@0 as greptime_timestamp], aggr=[sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)))]                                                                                                                                                                                                                                                                                                                                                                            |
-|               |         FilterExec: prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000))@1 IS NOT NULL                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |
-|               |           ProjectionExec: expr=[greptime_timestamp@4 as greptime_timestamp, prom_rate(greptime_timestamp_range@6, greptime_value@5, greptime_timestamp@4, 120000) as prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)), a@0 as a, b@1 as b, c@2 as c]                                                                                                                                                                                                                                                                                                                 |
-|               |             PromRangeManipulateExec: req range=[1752591864000..1752592164000], interval=[30000], eval range=[120000], time index=[greptime_timestamp]                                                                                                                                                                                                                                                                                                                                                                                                                                                   |
-|               |               PromSeriesNormalizeExec: offset=[0], time index=[greptime_timestamp], filter NaN: [true]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  |
-|               |                 PromSeriesDivideExec: tags=["a", "b", "c", "d"]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         |
-|               |                   SortExec: expr=[a@0 ASC, b@1 ASC, c@2 ASC, d@3 ASC, greptime_timestamp@4 ASC], preserve_partitioning=[true]                                                                                                                                                                                                                                                                                                                                                                                                                                                                           |
+|               |     CoalescePartitionsExec                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                |
+|               |       AggregateExec: mode=SinglePartitioned, gby=[a@2 as a, b@3 as b, c@4 as c, greptime_timestamp@0 as greptime_timestamp], aggr=[sum(prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)))]                                                                                                                                                                                                                                                                                                                                                                                                              |
+|               |         FilterExec: prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000))@1 IS NOT NULL                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     |
+|               |           ProjectionExec: expr=[greptime_timestamp@4 as greptime_timestamp, prom_rate(greptime_timestamp_range@6, greptime_value@5, greptime_timestamp@4, 120000) as prom_rate(greptime_timestamp_range,greptime_value,greptime_timestamp,Int64(120000)), a@0 as a, b@1 as b, c@2 as c]                                                                                                                                                                                                                                                                                                                                                   |
+|               |             PromRangeManipulateExec: req range=[1752591864000..1752592164000], interval=[30000], eval range=[120000], time index=[greptime_timestamp]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     |
+|               |               PromSeriesNormalizeExec: offset=[0], time index=[greptime_timestamp], filter NaN: [true]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    |
+|               |                 PromSeriesDivideExec: tags=["a", "b", "c", "d"]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           |
+|               |                   SortExec: expr=[a@0 ASC, b@1 ASC, c@2 ASC, d@3 ASC, greptime_timestamp@4 ASC], preserve_partitioning=[true]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             |
 |               |                     MergeScanExec: REDACTED
-|               |     SortExec: expr=[a@0 ASC NULLS LAST, b@1 ASC NULLS LAST, c@2 ASC NULLS LAST, greptime_timestamp@3 ASC NULLS LAST], preserve_partitioning=[true]                                                                                                                                                                                                                                                                                                                                                                                                                                                      |
-|               |       CooperativeExec                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |
+|               |     SortExec: expr=[a@0 ASC NULLS LAST, b@1 ASC NULLS LAST, c@2 ASC NULLS LAST, greptime_timestamp@3 ASC NULLS LAST], preserve_partitioning=[true]                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
+|               |       CooperativeExec                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     |
 |               |         MergeScanExec: REDACTED
-|               |                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         |
-+---------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+|               |                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           |
++---------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
 
 -- SQLNESS REPLACE (metrics.*) REDACTED
 -- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
diff --git a/tests/cases/distributed/flow-tql/flow_tql_flownode_addrs.result b/tests/cases/distributed/flow-tql/flow_tql_flownode_addrs.result
new file mode 100644
index 0000000000..5494c3f011
--- /dev/null
+++ b/tests/cases/distributed/flow-tql/flow_tql_flownode_addrs.result
@@ -0,0 +1,37 @@
+CREATE TABLE http_requests (
+  ts timestamp(3) time index,
+  host STRING,
+  idc STRING,
+  val DOUBLE,
+  PRIMARY KEY(host, idc),
+);
+
+Affected Rows: 0
+
+CREATE FLOW calc_reqs SINK TO cnt_reqs EVAL INTERVAL '1m' AS
+TQL EVAL (now() - '1m'::interval, now(), '5s') count_values("status_code", http_requests);
+
+Affected Rows: 0
+
+SELECT flownode_addrs IS NOT NULL AS flownode_addrs_is_not_null
+FROM information_schema.flows
+WHERE flow_name = 'calc_reqs';
+
++----------------------------+
+| flownode_addrs_is_not_null |
++----------------------------+
+| true                       |
++----------------------------+
+
+DROP FLOW calc_reqs;
+
+Affected Rows: 0
+
+DROP TABLE http_requests;
+
+Affected Rows: 0
+
+DROP TABLE cnt_reqs;
+
+Affected Rows: 0
+
diff --git a/tests/cases/distributed/flow-tql/flow_tql_flownode_addrs.sql b/tests/cases/distributed/flow-tql/flow_tql_flownode_addrs.sql
new file mode 100644
index 0000000000..eeb0f48ed0
--- /dev/null
+++ b/tests/cases/distributed/flow-tql/flow_tql_flownode_addrs.sql
@@ -0,0 +1,18 @@
+CREATE TABLE http_requests (
+  ts timestamp(3) time index,
+  host STRING,
+  idc STRING,
+  val DOUBLE,
+  PRIMARY KEY(host, idc),
+);
+
+CREATE FLOW calc_reqs SINK TO cnt_reqs EVAL INTERVAL '1m' AS
+TQL EVAL (now() - '1m'::interval, now(), '5s') count_values("status_code", http_requests);
+
+SELECT flownode_addrs IS NOT NULL AS flownode_addrs_is_not_null
+FROM information_schema.flows
+WHERE flow_name = 'calc_reqs';
+
+DROP FLOW calc_reqs;
+DROP TABLE http_requests;
+DROP TABLE cnt_reqs;
diff --git a/tests/cases/distributed/information_schema/cluster_info.result b/tests/cases/distributed/information_schema/cluster_info.result
index 83512e6ffb..662c73a1ea 100644
--- a/tests/cases/distributed/information_schema/cluster_info.result
+++ b/tests/cases/distributed/information_schema/cluster_info.result
@@ -24,55 +24,55 @@ DESC TABLE CLUSTER_INFO;
 +----------------------+----------------------+-----+------+---------+---------------+
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
--- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
--- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE (\s127\.0\.0\.1:\d+\s) Address
+-- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
+-- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO ORDER BY peer_type;
 
 +++++++++|peer_id|peer_type|peer_addr|node_version|git_commit|start_time|uptime|active_time|+++++++++|0|DATANODE|Address|Version|Hash|Start_time|Duration|Duration||1|DATANODE|Address|Version|Hash|Start_time|Duration|Duration||2|DATANODE|Address|Version|Hash|Start_time|Duration|Duration||0|FLOWNODE|Address|Version|Hash|Start_time|Duration|Duration||1|FRONTEND|Address|Version|Hash|Start_time|Duration|Duration||1|METASRV|Address|Version|Hash|Start_time|Duration||+++++++++
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
--- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
--- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE (\s127\.0\.0\.1:\d+\s) Address
+-- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
+-- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO WHERE PEER_TYPE = 'METASRV' ORDER BY peer_type;
 
 +++++++++|peer_id|peer_type|peer_addr|node_version|git_commit|start_time|uptime|active_time|+++++++++|1|METASRV|Address|Version|Hash|Start_time|Duration||+++++++++
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
--- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
--- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE (\s127\.0\.0\.1:\d+\s) Address
+-- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
+-- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO WHERE PEER_TYPE = 'FRONTEND' ORDER BY peer_type;
 
 +++++++++|peer_id|peer_type|peer_addr|node_version|git_commit|start_time|uptime|active_time|+++++++++|1|FRONTEND|Address|Version|Hash|Start_time|Duration|Duration|+++++++++
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
--- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
--- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE (\s127\.0\.0\.1:\d+\s) Address
+-- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
+-- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO WHERE PEER_TYPE != 'FRONTEND' ORDER BY peer_type;
 
 +++++++++|peer_id|peer_type|peer_addr|node_version|git_commit|start_time|uptime|active_time|+++++++++|0|DATANODE|Address|Version|Hash|Start_time|Duration|Duration||1|DATANODE|Address|Version|Hash|Start_time|Duration|Duration||2|DATANODE|Address|Version|Hash|Start_time|Duration|Duration||0|FLOWNODE|Address|Version|Hash|Start_time|Duration|Duration||1|METASRV|Address|Version|Hash|Start_time|Duration||+++++++++
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
--- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
--- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE (\s127\.0\.0\.1:\d+\s) Address
+-- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
+-- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO WHERE PEER_ID > 1 ORDER BY peer_type;
 
diff --git a/tests/cases/distributed/information_schema/cluster_info.sql b/tests/cases/distributed/information_schema/cluster_info.sql
index 3c2dcccaa0..9e85245859 100644
--- a/tests/cases/distributed/information_schema/cluster_info.sql
+++ b/tests/cases/distributed/information_schema/cluster_info.sql
@@ -3,47 +3,47 @@ USE INFORMATION_SCHEMA;
 DESC TABLE CLUSTER_INFO;
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
--- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
--- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE (\s127\.0\.0\.1:\d+\s) Address
+-- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
+-- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO ORDER BY peer_type;
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
--- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
--- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE (\s127\.0\.0\.1:\d+\s) Address
+-- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
+-- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO WHERE PEER_TYPE = 'METASRV' ORDER BY peer_type;
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
--- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
--- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE (\s127\.0\.0\.1:\d+\s) Address
+-- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
+-- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO WHERE PEER_TYPE = 'FRONTEND' ORDER BY peer_type;
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
--- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
--- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE (\s127\.0\.0\.1:\d+\s) Address
+-- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
+-- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO WHERE PEER_TYPE != 'FRONTEND' ORDER BY peer_type;
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
--- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
--- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE (\s127\.0\.0\.1:\d+\s) Address
+-- SQLNESS REPLACE (\s[\-0-9T:\.]{19,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
+-- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO WHERE PEER_ID > 1 ORDER BY peer_type;
 
diff --git a/tests/cases/distributed/optimizer/windowed_sort_advance.result b/tests/cases/distributed/optimizer/windowed_sort_advance.result
new file mode 100644
index 0000000000..e3c92a0a08
--- /dev/null
+++ b/tests/cases/distributed/optimizer/windowed_sort_advance.result
@@ -0,0 +1,118 @@
+create table `a` (`value` double, `status` bigint, ts timestamp(9) time index);
+
+Affected Rows: 0
+
+INSERT INTO `a` (`value`, `status`, `ts`) VALUES
+(46.82, 200, '2026-03-12T08:00:05.000000000+08:00'),
+(46.84, 200, '2026-03-12T08:00:15.000000000+08:00'),
+(46.85, 200, '2026-03-12T08:00:25.000000000+08:00'),
+(46.86, 200, '2026-03-12T08:00:35.000000000+08:00'),
+(46.88, 200, '2026-03-12T08:00:45.000000000+08:00'),
+(46.89, 200, '2026-03-12T08:00:55.000000000+08:00'),
+(46.91, 200, '2026-03-12T08:01:05.000000000+08:00'),
+(46.90, 200, '2026-03-12T08:01:15.000000000+08:00'),
+(46.87, 200, '2026-03-12T08:01:25.000000000+08:00'),
+(46.85, 200, '2026-03-12T08:01:35.000000000+08:00');
+
+Affected Rows: 10
+
+select ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
++---------------------+--------+-------+
+| ts                  | status | value |
++---------------------+--------+-------+
+| 2026-03-12T00:00:05 | 200    | 46.82 |
+| 2026-03-12T00:00:15 | 200    | 46.84 |
+| 2026-03-12T00:00:25 | 200    | 46.85 |
+| 2026-03-12T00:00:35 | 200    | 46.86 |
+| 2026-03-12T00:00:45 | 200    | 46.88 |
+| 2026-03-12T00:00:55 | 200    | 46.89 |
+| 2026-03-12T00:01:05 | 200    | 46.91 |
+| 2026-03-12T00:01:15 | 200    | 46.9  |
+| 2026-03-12T00:01:25 | 200    | 46.87 |
+| 2026-03-12T00:01:35 | 200    | 46.85 |
++---------------------+--------+-------+
+
+select ts as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
++---------------------+--------+-------+
+| ts                  | status | value |
++---------------------+--------+-------+
+| 2026-03-12T00:00:05 | 200    | 46.82 |
+| 2026-03-12T00:00:15 | 200    | 46.84 |
+| 2026-03-12T00:00:25 | 200    | 46.85 |
+| 2026-03-12T00:00:35 | 200    | 46.86 |
+| 2026-03-12T00:00:45 | 200    | 46.88 |
+| 2026-03-12T00:00:55 | 200    | 46.89 |
+| 2026-03-12T00:01:05 | 200    | 46.91 |
+| 2026-03-12T00:01:15 | 200    | 46.9  |
+| 2026-03-12T00:01:25 | 200    | 46.87 |
+| 2026-03-12T00:01:35 | 200    | 46.85 |
++---------------------+--------+-------+
+
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE num_ranges=\d+ num_ranges=REDACTED
+EXPLAIN ANALYZE select ts as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
++-+-+-+
+| stage | node | plan_|
++-+-+-+
+| 0_| 0_|_CooperativeExec REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_|
+| 1_| 0_|_SortPreservingMergeExec: [ts@0 ASC NULLS LAST] REDACTED
+|_|_|_WindowedSortExec: expr=ts@0 ASC NULLS LAST num_ranges=REDACTED REDACTED
+|_|_|_ProjectionExec: expr=[ts@2 as ts, status@1 as status, value@0 as value] REDACTED
+|_|_|_SeqScan: region=REDACTED, "partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0} REDACTED
+|_|_|_|
+|_|_| Total rows: 10_|
++-+-+-+
+
+select to_timestamp_millis(ts) as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
++---------------------+--------+-------+
+| ts                  | status | value |
++---------------------+--------+-------+
+| 2026-03-12T00:00:05 | 200    | 46.82 |
+| 2026-03-12T00:00:15 | 200    | 46.84 |
+| 2026-03-12T00:00:25 | 200    | 46.85 |
+| 2026-03-12T00:00:35 | 200    | 46.86 |
+| 2026-03-12T00:00:45 | 200    | 46.88 |
+| 2026-03-12T00:00:55 | 200    | 46.89 |
+| 2026-03-12T00:01:05 | 200    | 46.91 |
+| 2026-03-12T00:01:15 | 200    | 46.9  |
+| 2026-03-12T00:01:25 | 200    | 46.87 |
+| 2026-03-12T00:01:35 | 200    | 46.85 |
++---------------------+--------+-------+
+
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE num_ranges=\d+ num_ranges=REDACTED
+EXPLAIN ANALYZE select to_timestamp_millis(ts) as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
++-+-+-+
+| stage | node | plan_|
++-+-+-+
+| 0_| 0_|_CooperativeExec REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_|
+| 1_| 0_|_ProjectionExec: expr=[to_timestamp_millis(a.ts)@0 as ts, status@1 as status, value@2 as value] REDACTED
+|_|_|_SortPreservingMergeExec: [to_timestamp_millis(a.ts)@0 ASC NULLS LAST] REDACTED
+|_|_|_WindowedSortExec: expr=to_timestamp_millis(a.ts)@0 ASC NULLS LAST num_ranges=REDACTED REDACTED
+|_|_|_ProjectionExec: expr=[to_timestamp_millis(ts@2) as to_timestamp_millis(a.ts), status@1 as status, value@0 as value] REDACTED
+|_|_|_SeqScan: region=REDACTED, "partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0} REDACTED
+|_|_|_|
+|_|_| Total rows: 10_|
++-+-+-+
+
+DROP TABLE `a`;
+
+Affected Rows: 0
+
diff --git a/tests/cases/distributed/optimizer/windowed_sort_advance.sql b/tests/cases/distributed/optimizer/windowed_sort_advance.sql
new file mode 100644
index 0000000000..fa73b97933
--- /dev/null
+++ b/tests/cases/distributed/optimizer/windowed_sort_advance.sql
@@ -0,0 +1,37 @@
+create table `a` (`value` double, `status` bigint, ts timestamp(9) time index);
+
+INSERT INTO `a` (`value`, `status`, `ts`) VALUES
+(46.82, 200, '2026-03-12T08:00:05.000000000+08:00'),
+(46.84, 200, '2026-03-12T08:00:15.000000000+08:00'),
+(46.85, 200, '2026-03-12T08:00:25.000000000+08:00'),
+(46.86, 200, '2026-03-12T08:00:35.000000000+08:00'),
+(46.88, 200, '2026-03-12T08:00:45.000000000+08:00'),
+(46.89, 200, '2026-03-12T08:00:55.000000000+08:00'),
+(46.91, 200, '2026-03-12T08:01:05.000000000+08:00'),
+(46.90, 200, '2026-03-12T08:01:15.000000000+08:00'),
+(46.87, 200, '2026-03-12T08:01:25.000000000+08:00'),
+(46.85, 200, '2026-03-12T08:01:35.000000000+08:00');
+
+select ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
+select ts as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE num_ranges=\d+ num_ranges=REDACTED
+EXPLAIN ANALYZE select ts as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
+select to_timestamp_millis(ts) as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE num_ranges=\d+ num_ranges=REDACTED
+EXPLAIN ANALYZE select to_timestamp_millis(ts) as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
+DROP TABLE `a`;
diff --git a/tests/cases/distributed/optimizer/windowed_sort_nanos.result b/tests/cases/distributed/optimizer/windowed_sort_nanos.result
new file mode 100644
index 0000000000..2310ec8f28
--- /dev/null
+++ b/tests/cases/distributed/optimizer/windowed_sort_nanos.result
@@ -0,0 +1,79 @@
+create table `a_ms` (`value` double, `status` bigint, ts timestamp(3) time index);
+
+Affected Rows: 0
+
+INSERT INTO `a_ms` (`value`, `status`, `ts`) VALUES
+(46.82, 200, '2026-03-12T08:00:05.123+08:00'),
+(46.84, 200, '2026-03-12T08:00:15.234+08:00'),
+(46.85, 200, '2026-03-12T08:00:25.345+08:00'),
+(46.86, 200, '2026-03-12T08:00:35.456+08:00'),
+(46.88, 200, '2026-03-12T08:00:45.567+08:00'),
+(46.89, 200, '2026-03-12T08:00:55.678+08:00'),
+(46.91, 200, '2026-03-12T08:01:05.789+08:00'),
+(46.90, 200, '2026-03-12T08:01:15.890+08:00'),
+(46.87, 200, '2026-03-12T08:01:25.901+08:00'),
+(46.85, 200, '2026-03-12T08:01:35.999+08:00');
+
+Affected Rows: 10
+
+select ts, status, value from `a_ms` where ts >= '2026-03-12T08:00:00.000+08:00' and ts < '2026-03-12T08:02:01.000+08:00' order by ts asc;
+
++-------------------------+--------+-------+
+| ts                      | status | value |
++-------------------------+--------+-------+
+| 2026-03-12T00:00:05.123 | 200    | 46.82 |
+| 2026-03-12T00:00:15.234 | 200    | 46.84 |
+| 2026-03-12T00:00:25.345 | 200    | 46.85 |
+| 2026-03-12T00:00:35.456 | 200    | 46.86 |
+| 2026-03-12T00:00:45.567 | 200    | 46.88 |
+| 2026-03-12T00:00:55.678 | 200    | 46.89 |
+| 2026-03-12T00:01:05.789 | 200    | 46.91 |
+| 2026-03-12T00:01:15.890 | 200    | 46.9  |
+| 2026-03-12T00:01:25.901 | 200    | 46.87 |
+| 2026-03-12T00:01:35.999 | 200    | 46.85 |
++-------------------------+--------+-------+
+
+select to_timestamp_nanos(ts) as ts, status, value from `a_ms` where ts >= '2026-03-12T08:00:00.000+08:00' and ts < '2026-03-12T08:02:01.000+08:00' order by ts asc;
+
++-------------------------+--------+-------+
+| ts                      | status | value |
++-------------------------+--------+-------+
+| 2026-03-12T00:00:05.123 | 200    | 46.82 |
+| 2026-03-12T00:00:15.234 | 200    | 46.84 |
+| 2026-03-12T00:00:25.345 | 200    | 46.85 |
+| 2026-03-12T00:00:35.456 | 200    | 46.86 |
+| 2026-03-12T00:00:45.567 | 200    | 46.88 |
+| 2026-03-12T00:00:55.678 | 200    | 46.89 |
+| 2026-03-12T00:01:05.789 | 200    | 46.91 |
+| 2026-03-12T00:01:15.890 | 200    | 46.9  |
+| 2026-03-12T00:01:25.901 | 200    | 46.87 |
+| 2026-03-12T00:01:35.999 | 200    | 46.85 |
++-------------------------+--------+-------+
+
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE num_ranges=\d+ num_ranges=REDACTED
+EXPLAIN ANALYZE select to_timestamp_nanos(ts) as ts, status, value from `a_ms` where ts >= '2026-03-12T08:00:00.000+08:00' and ts < '2026-03-12T08:02:01.000+08:00' order by ts asc;
+
++-+-+-+
+| stage | node | plan_|
++-+-+-+
+| 0_| 0_|_CooperativeExec REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_|
+| 1_| 0_|_ProjectionExec: expr=[to_timestamp_nanos(a_ms.ts)@0 as ts, status@1 as status, value@2 as value] REDACTED
+|_|_|_SortPreservingMergeExec: [to_timestamp_nanos(a_ms.ts)@0 ASC NULLS LAST] REDACTED
+|_|_|_WindowedSortExec: expr=to_timestamp_nanos(a_ms.ts)@0 ASC NULLS LAST num_ranges=REDACTED REDACTED
+|_|_|_ProjectionExec: expr=[to_timestamp_nanos(ts@2) as to_timestamp_nanos(a_ms.ts), status@1 as status, value@0 as value] REDACTED
+|_|_|_SeqScan: region=REDACTED, "partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0} REDACTED
+|_|_|_|
+|_|_| Total rows: 10_|
++-+-+-+
+
+DROP TABLE `a_ms`;
+
+Affected Rows: 0
+
diff --git a/tests/cases/distributed/optimizer/windowed_sort_nanos.sql b/tests/cases/distributed/optimizer/windowed_sort_nanos.sql
new file mode 100644
index 0000000000..6a6cf51c4b
--- /dev/null
+++ b/tests/cases/distributed/optimizer/windowed_sort_nanos.sql
@@ -0,0 +1,27 @@
+create table `a_ms` (`value` double, `status` bigint, ts timestamp(3) time index);
+
+INSERT INTO `a_ms` (`value`, `status`, `ts`) VALUES
+(46.82, 200, '2026-03-12T08:00:05.123+08:00'),
+(46.84, 200, '2026-03-12T08:00:15.234+08:00'),
+(46.85, 200, '2026-03-12T08:00:25.345+08:00'),
+(46.86, 200, '2026-03-12T08:00:35.456+08:00'),
+(46.88, 200, '2026-03-12T08:00:45.567+08:00'),
+(46.89, 200, '2026-03-12T08:00:55.678+08:00'),
+(46.91, 200, '2026-03-12T08:01:05.789+08:00'),
+(46.90, 200, '2026-03-12T08:01:15.890+08:00'),
+(46.87, 200, '2026-03-12T08:01:25.901+08:00'),
+(46.85, 200, '2026-03-12T08:01:35.999+08:00');
+
+select ts, status, value from `a_ms` where ts >= '2026-03-12T08:00:00.000+08:00' and ts < '2026-03-12T08:02:01.000+08:00' order by ts asc;
+
+select to_timestamp_nanos(ts) as ts, status, value from `a_ms` where ts >= '2026-03-12T08:00:00.000+08:00' and ts < '2026-03-12T08:02:01.000+08:00' order by ts asc;
+
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE num_ranges=\d+ num_ranges=REDACTED
+EXPLAIN ANALYZE select to_timestamp_nanos(ts) as ts, status, value from `a_ms` where ts >= '2026-03-12T08:00:00.000+08:00' and ts < '2026-03-12T08:02:01.000+08:00' order by ts asc;
+
+DROP TABLE `a_ms`;
diff --git a/tests/cases/standalone/common/alter/alter_database.result b/tests/cases/standalone/common/alter/alter_database.result
index 911ef5ddfc..2fccce10de 100644
--- a/tests/cases/standalone/common/alter/alter_database.result
+++ b/tests/cases/standalone/common/alter/alter_database.result
@@ -314,6 +314,85 @@ SHOW CREATE DATABASE alter_database;
 |                | )                                            |
 +----------------+----------------------------------------------+
 
+-- Test sst_format option
+ALTER DATABASE alter_database SET 'sst_format'='flat';
+
+Affected Rows: 0
+
+SHOW CREATE DATABASE alter_database;
+
++----------------+----------------------------------------------+
+| Database       | Create Database                              |
++----------------+----------------------------------------------+
+| alter_database | CREATE DATABASE IF NOT EXISTS alter_database |
+|                | WITH(                                        |
+|                |   'compaction.twcs.time_window' = '30m',     |
+|                |   'compaction.type' = 'twcs',                |
+|                |   sst_format = 'flat'                        |
+|                | )                                            |
++----------------+----------------------------------------------+
+
+USE alter_database;
+
+Affected Rows: 0
+
+CREATE TABLE monitor(ts TIMESTAMP TIME INDEX);
+
+Affected Rows: 0
+
+SHOW CREATE TABLE monitor;
+
++---------+----------------------------------------+
+| Table   | Create Table                           |
++---------+----------------------------------------+
+| monitor | CREATE TABLE IF NOT EXISTS "monitor" ( |
+|         |   "ts" TIMESTAMP(3) NOT NULL,          |
+|         |   TIME INDEX ("ts")                    |
+|         | )                                      |
+|         |                                        |
+|         | ENGINE=mito                            |
+|         | WITH(                                  |
+|         |   sst_format = 'flat'                  |
+|         | )                                      |
++---------+----------------------------------------+
+
+USE public;
+
+Affected Rows: 0
+
+ALTER DATABASE alter_database SET 'sst_format'='primary_key';
+
+Affected Rows: 0
+
+SHOW CREATE DATABASE alter_database;
+
++----------------+----------------------------------------------+
+| Database       | Create Database                              |
++----------------+----------------------------------------------+
+| alter_database | CREATE DATABASE IF NOT EXISTS alter_database |
+|                | WITH(                                        |
+|                |   'compaction.twcs.time_window' = '30m',     |
+|                |   'compaction.type' = 'twcs',                |
+|                |   sst_format = 'primary_key'                 |
+|                | )                                            |
++----------------+----------------------------------------------+
+
+ALTER DATABASE alter_database UNSET 'sst_format';
+
+Affected Rows: 0
+
+SHOW CREATE DATABASE alter_database;
+
++----------------+----------------------------------------------+
+| Database       | Create Database                              |
++----------------+----------------------------------------------+
+| alter_database | CREATE DATABASE IF NOT EXISTS alter_database |
+|                | WITH(                                        |
+|                |   'compaction.twcs.time_window' = '30m',     |
+|                |   'compaction.type' = 'twcs'                 |
+|                | )                                            |
++----------------+----------------------------------------------+
+
 DROP DATABASE alter_database;
 
 Affected Rows: 0
diff --git a/tests/cases/standalone/common/alter/alter_database.sql b/tests/cases/standalone/common/alter/alter_database.sql
index 1b2f75637a..33b309153e 100644
--- a/tests/cases/standalone/common/alter/alter_database.sql
+++ b/tests/cases/standalone/common/alter/alter_database.sql
@@ -90,5 +90,25 @@ ALTER DATABASE alter_database UNSET 'ttl';
 
 SHOW CREATE DATABASE alter_database;
 
-DROP DATABASE alter_database;
+-- Test sst_format option
+ALTER DATABASE alter_database SET 'sst_format'='flat';
 
+SHOW CREATE DATABASE alter_database;
+
+USE alter_database;
+
+CREATE TABLE monitor(ts TIMESTAMP TIME INDEX);
+
+SHOW CREATE TABLE monitor;
+
+USE public;
+
+ALTER DATABASE alter_database SET 'sst_format'='primary_key';
+
+SHOW CREATE DATABASE alter_database;
+
+ALTER DATABASE alter_database UNSET 'sst_format';
+
+SHOW CREATE DATABASE alter_database;
+
+DROP DATABASE alter_database;
diff --git a/tests/cases/standalone/common/alter/alter_format.result b/tests/cases/standalone/common/alter/alter_format.result
index d38c63997d..a1019a8c93 100644
--- a/tests/cases/standalone/common/alter/alter_format.result
+++ b/tests/cases/standalone/common/alter/alter_format.result
@@ -42,6 +42,26 @@ ALTER TABLE test_alt_format SET 'sst_format' = 'flat';
 
 Affected Rows: 0
 
+SHOW CREATE TABLE test_alt_format;
+
++-----------------+------------------------------------------------+
+| Table           | Create Table                                   |
++-----------------+------------------------------------------------+
+| test_alt_format | CREATE TABLE IF NOT EXISTS "test_alt_format" ( |
+|                 |   "h" INT NULL,                                |
+|                 |   "i" INT NULL DEFAULT 0,                      |
+|                 |   "j" TIMESTAMP(3) NOT NULL,                   |
+|                 |   "k" INT NULL,                                |
+|                 |   TIME INDEX ("j"),                            |
+|                 |   PRIMARY KEY ("h")                            |
+|                 | )                                              |
+|                 |                                                |
+|                 | ENGINE=mito                                    |
+|                 | WITH(                                          |
+|                 |   sst_format = 'flat'                          |
+|                 | )                                              |
++-----------------+------------------------------------------------+
+
 -- SQLNESS SORT_RESULT 3 1
 SELECT * FROM test_alt_format;
 
@@ -116,11 +136,68 @@ SELECT i, h FROM test_alt_format;
 | 23 | 13 |
 +----+----+
 
--- not allow to change from flat to primary_key
--- SQLNESS REPLACE \d+\(\d+,\s+\d+\) REDACTED
+-- allow to change from flat to primary_key
 ALTER TABLE test_alt_format SET 'sst_format' = 'primary_key';
 
-Error: 1004(InvalidArguments), Invalid region request, region_id: REDACTED, err: Only allow changing format type to flat
+Affected Rows: 0
+
+SHOW CREATE TABLE test_alt_format;
+
++-----------------+------------------------------------------------+
+| Table           | Create Table                                   |
++-----------------+------------------------------------------------+
+| test_alt_format | CREATE TABLE IF NOT EXISTS "test_alt_format" ( |
+|                 |   "h" INT NULL,                                |
+|                 |   "i" INT NULL DEFAULT 0,                      |
+|                 |   "j" TIMESTAMP(3) NOT NULL,                   |
+|                 |   "k" INT NULL,                                |
+|                 |   TIME INDEX ("j"),                            |
+|                 |   PRIMARY KEY ("h")                            |
+|                 | )                                              |
+|                 |                                                |
+|                 | ENGINE=mito                                    |
+|                 | WITH(                                          |
+|                 |   sst_format = 'primary_key'                   |
+|                 | )                                              |
++-----------------+------------------------------------------------+
+
+INSERT INTO test_alt_format (h, j, i) VALUES (14, 4, 34);
+
+Affected Rows: 1
+
+-- SQLNESS SORT_RESULT 3 1
+SELECT * FROM test_alt_format;
+
++----+----+-------------------------+----+
+| h  | i  | j                       | k  |
++----+----+-------------------------+----+
+| 10 | 0  | 1970-01-01T00:00:00     |    |
+| 11 | 0  | 1970-01-01T00:00:00.001 |    |
+| 12 | 0  | 1970-01-01T00:00:00.002 |    |
+| 13 | 23 | 1970-01-01T00:00:00.003 | 33 |
+| 14 | 34 | 1970-01-01T00:00:00.004 |    |
++----+----+-------------------------+----+
+
+ADMIN flush_table('test_alt_format');
+
++--------------------------------------+
+| ADMIN flush_table('test_alt_format') |
++--------------------------------------+
+| 0                                    |
++--------------------------------------+
+
+-- SQLNESS SORT_RESULT 3 1
+SELECT * FROM test_alt_format;
+
++----+----+-------------------------+----+
+| h  | i  | j                       | k  |
++----+----+-------------------------+----+
+| 10 | 0  | 1970-01-01T00:00:00     |    |
+| 11 | 0  | 1970-01-01T00:00:00.001 |    |
+| 12 | 0  | 1970-01-01T00:00:00.002 |    |
+| 13 | 23 | 1970-01-01T00:00:00.003 | 33 |
+| 14 | 34 | 1970-01-01T00:00:00.004 |    |
++----+----+-------------------------+----+
 
 DROP TABLE test_alt_format;
 
@@ -167,6 +244,27 @@ ALTER TABLE alt_format_phy SET 'sst_format' = 'flat';
 
 Affected Rows: 0
 
+SHOW CREATE TABLE alt_format_phy;
+
++----------------+-----------------------------------------------+
+| Table          | Create Table                                  |
++----------------+-----------------------------------------------+
+| alt_format_phy | CREATE TABLE IF NOT EXISTS "alt_format_phy" ( |
+|                |   "ts" TIMESTAMP(3) NOT NULL,                 |
+|                |   "val" DOUBLE NULL,                          |
+|                |   "host" STRING NULL,                         |
+|                |   "k" STRING NULL,                            |
+|                |   TIME INDEX ("ts"),                          |
+|                |   PRIMARY KEY ("host", "k")                   |
+|                | )                                             |
+|                |                                               |
+|                | ENGINE=metric                                 |
+|                | WITH(                                         |
+|                |   physical_metric_table = '',                 |
+|                |   sst_format = 'flat'                         |
+|                | )                                             |
++----------------+-----------------------------------------------+
+
 SELECT * FROM t1 ORDER BY ts ASC;
 
 +-------------+---+---------------------+------+
@@ -202,11 +300,47 @@ SELECT host, ts, val FROM t1 where host = 'example.com' ORDER BY ts ASC;
 | example.com | 2022-01-02T00:00:00 | 4.56 |
 +-------------+---------------------+------+
 
--- not allow to change from flat to primary_key
--- SQLNESS REPLACE \d+\(\d+,\s+\d+\) REDACTED
+-- allow to change from flat to primary_key
 ALTER TABLE alt_format_phy SET 'sst_format' = 'primary_key';
 
-Error: 1004(InvalidArguments), Invalid region request, region_id: REDACTED, err: Only allow changing format type to flat
+Affected Rows: 0
+
+SHOW CREATE TABLE alt_format_phy;
+
++----------------+-----------------------------------------------+
+| Table          | Create Table                                  |
++----------------+-----------------------------------------------+
+| alt_format_phy | CREATE TABLE IF NOT EXISTS "alt_format_phy" ( |
+|                |   "ts" TIMESTAMP(3) NOT NULL,                 |
+|                |   "val" DOUBLE NULL,                          |
+|                |   "host" STRING NULL,                         |
+|                |   "k" STRING NULL,                            |
+|                |   TIME INDEX ("ts"),                          |
+|                |   PRIMARY KEY ("host", "k")                   |
+|                | )                                             |
+|                |                                               |
+|                | ENGINE=metric                                 |
+|                | WITH(                                         |
+|                |   physical_metric_table = '',                 |
+|                |   sst_format = 'primary_key'                  |
+|                | )                                             |
++----------------+-----------------------------------------------+
+
+INSERT INTO t1 (ts, val, host) VALUES
+  ('2022-01-01 00:00:02', 5.0, 'example.com');
+
+Affected Rows: 1
+
+SELECT host, ts, val FROM t1 where host = 'example.com' ORDER BY ts ASC;
+
++-------------+---------------------+------+
+| host        | ts                  | val  |
++-------------+---------------------+------+
+| example.com | 2022-01-01T00:00:00 | 1.23 |
+| example.com | 2022-01-01T00:00:01 | 3.0  |
+| example.com | 2022-01-01T00:00:02 | 5.0  |
+| example.com | 2022-01-02T00:00:00 | 4.56 |
++-------------+---------------------+------+
 
 DROP TABLE t1;
 
diff --git a/tests/cases/standalone/common/alter/alter_format.sql b/tests/cases/standalone/common/alter/alter_format.sql
index e1472d28e1..c3b292875c 100644
--- a/tests/cases/standalone/common/alter/alter_format.sql
+++ b/tests/cases/standalone/common/alter/alter_format.sql
@@ -16,6 +16,8 @@ SELECT i, h FROM test_alt_format;
 
 ALTER TABLE test_alt_format SET 'sst_format' = 'flat';
 
+SHOW CREATE TABLE test_alt_format;
+
 -- SQLNESS SORT_RESULT 3 1
 SELECT * FROM test_alt_format;
 
@@ -37,10 +39,21 @@ SELECT * FROM test_alt_format;
 -- SQLNESS SORT_RESULT 3 1
 SELECT i, h FROM test_alt_format;
 
--- not allow to change from flat to primary_key
--- SQLNESS REPLACE \d+\(\d+,\s+\d+\) REDACTED
+-- allow to change from flat to primary_key
 ALTER TABLE test_alt_format SET 'sst_format' = 'primary_key';
 
+SHOW CREATE TABLE test_alt_format;
+
+INSERT INTO test_alt_format (h, j, i) VALUES (14, 4, 34);
+
+-- SQLNESS SORT_RESULT 3 1
+SELECT * FROM test_alt_format;
+
+ADMIN flush_table('test_alt_format');
+
+-- SQLNESS SORT_RESULT 3 1
+SELECT * FROM test_alt_format;
+
 DROP TABLE test_alt_format;
 
 CREATE TABLE alt_format_phy (ts timestamp time index, val double) engine=metric with ("physical_metric_table" = "", "sst_format" = "primary_key");
@@ -62,6 +75,8 @@ SELECT * FROM t1 ORDER BY ts ASC;
 
 ALTER TABLE alt_format_phy SET 'sst_format' = 'flat';
 
+SHOW CREATE TABLE alt_format_phy;
+
 SELECT * FROM t1 ORDER BY ts ASC;
 
 SELECT host, ts, val FROM t1 where host = 'example.com' ORDER BY ts ASC;
@@ -72,10 +87,16 @@ INSERT INTO t1 (ts, val, host) VALUES
 
 SELECT host, ts, val FROM t1 where host = 'example.com' ORDER BY ts ASC;
 
--- not allow to change from flat to primary_key
--- SQLNESS REPLACE \d+\(\d+,\s+\d+\) REDACTED
+-- allow to change from flat to primary_key
 ALTER TABLE alt_format_phy SET 'sst_format' = 'primary_key';
 
+SHOW CREATE TABLE alt_format_phy;
+
+INSERT INTO t1 (ts, val, host) VALUES
+  ('2022-01-01 00:00:02', 5.0, 'example.com');
+
+SELECT host, ts, val FROM t1 where host = 'example.com' ORDER BY ts ASC;
+
 DROP TABLE t1;
 
 DROP TABLE alt_format_phy;
diff --git a/tests/cases/standalone/common/alter/alter_table.result b/tests/cases/standalone/common/alter/alter_table.result
index b1969bbaf7..5ba9fc0ed0 100644
--- a/tests/cases/standalone/common/alter/alter_table.result
+++ b/tests/cases/standalone/common/alter/alter_table.result
@@ -508,42 +508,42 @@ Affected Rows: 0
 ALTER TABLE
     t1
 ADD
-    COLUMN `at` STRING;
+    COLUMN `at` STRING PRIMARY KEY;
 
 Affected Rows: 0
 
 ALTER TABLE
     t2
 ADD
-    COLUMN at3 STRING;
+    COLUMN at3 STRING PRIMARY KEY;
 
 Affected Rows: 0
 
 ALTER TABLE
     t2
 ADD
-    COLUMN `at` STRING;
+    COLUMN `at` STRING PRIMARY KEY;
 
 Affected Rows: 0
 
 ALTER TABLE
     t2
 ADD
-    COLUMN at2 STRING;
+    COLUMN at2 STRING PRIMARY KEY;
 
 Affected Rows: 0
 
 ALTER TABLE
     t2
 ADD
-    COLUMN at4 UINT16;
+    COLUMN at4 STRING PRIMARY KEY;
 
 Affected Rows: 0
 
 INSERT INTO
     t2
 VALUES
-    ("loc_1", "loc_2", "loc_3", 2, 'job1', 0, 1);
+    ("loc_1", "loc_2", "loc_3", "loc_4", 'job1', 0, 1);
 
 Affected Rows: 1
 
@@ -552,11 +552,11 @@ SELECT
 FROM
     t2;
 
-+-------+-------+-------+-----+------+---------------------+-----+
-| at    | at2   | at3   | at4 | job  | ts                  | val |
-+-------+-------+-------+-----+------+---------------------+-----+
-| loc_1 | loc_2 | loc_3 | 2   | job1 | 1970-01-01T00:00:00 | 1.0 |
-+-------+-------+-------+-----+------+---------------------+-----+
++-------+-------+-------+-------+------+---------------------+-----+
+| at    | at2   | at3   | at4   | job  | ts                  | val |
++-------+-------+-------+-------+------+---------------------+-----+
+| loc_1 | loc_2 | loc_3 | loc_4 | job1 | 1970-01-01T00:00:00 | 1.0 |
++-------+-------+-------+-------+------+---------------------+-----+
 
 DROP TABLE t1;
 
diff --git a/tests/cases/standalone/common/alter/alter_table.sql b/tests/cases/standalone/common/alter/alter_table.sql
index 6088ea3bd2..767aca0510 100644
--- a/tests/cases/standalone/common/alter/alter_table.sql
+++ b/tests/cases/standalone/common/alter/alter_table.sql
@@ -184,32 +184,32 @@ CREATE TABLE t2 (
 ALTER TABLE
     t1
 ADD
-    COLUMN `at` STRING;
+    COLUMN `at` STRING PRIMARY KEY;
 
 ALTER TABLE
     t2
 ADD
-    COLUMN at3 STRING;
+    COLUMN at3 STRING PRIMARY KEY;
 
 ALTER TABLE
     t2
 ADD
-    COLUMN `at` STRING;
+    COLUMN `at` STRING PRIMARY KEY;
 
 ALTER TABLE
     t2
 ADD
-    COLUMN at2 STRING;
+    COLUMN at2 STRING PRIMARY KEY;
 
 ALTER TABLE
     t2
 ADD
-    COLUMN at4 UINT16;
+    COLUMN at4 STRING PRIMARY KEY;
 
 INSERT INTO
     t2
 VALUES
-    ("loc_1", "loc_2", "loc_3", 2, 'job1', 0, 1);
+    ("loc_1", "loc_2", "loc_3", "loc_4", 'job1', 0, 1);
 
 SELECT
     *
diff --git a/tests/cases/standalone/common/alter/change_col_type_inverted_index.result b/tests/cases/standalone/common/alter/change_col_type_inverted_index.result
new file mode 100644
index 0000000000..14d3d754a0
--- /dev/null
+++ b/tests/cases/standalone/common/alter/change_col_type_inverted_index.result
@@ -0,0 +1,52 @@
+-- Regression test for issue #8074
+-- https://github.com/GreptimeTeam/greptimedb/issues/8074
+CREATE TABLE monitoring_data (
+    host STRING INVERTED INDEX,
+    `region` STRING,
+    cpu_usage DOUBLE INVERTED INDEX,
+    `timestamp` TIMESTAMP TIME INDEX
+) WITH ('append_mode'='true');
+
+Affected Rows: 0
+
+INSERT INTO monitoring_data (host, region, cpu_usage, `timestamp`) VALUES
+('web-01', 'us-east', 12.5, '2026-05-06 10:00:00'),
+('web-02', 'us-east', 18.3, '2026-05-06 10:00:00'),
+('web-03', 'us-east', 91.2, '2026-05-06 10:00:00'),
+('web-04', 'us-west', 73.8, '2026-05-06 10:00:00');
+
+Affected Rows: 4
+
+INSERT INTO monitoring_data (host, region, cpu_usage, `timestamp`) VALUES
+('web-01', 'us-east', 15.2, '2026-05-06 10:01:00'),
+('web-02', 'us-east', 23.7, '2026-05-06 10:01:00'),
+('web-03', 'us-east', 94.5, '2026-05-06 10:01:00'),
+('web-04', 'us-west', 78.1, '2026-05-06 10:01:00');
+
+Affected Rows: 4
+
+ADMIN FLUSH_TABLE('monitoring_data');
+
++--------------------------------------+
+| ADMIN FLUSH_TABLE('monitoring_data') |
++--------------------------------------+
+| 0                                    |
++--------------------------------------+
+
+ALTER TABLE monitoring_data
+MODIFY COLUMN cpu_usage STRING;
+
+Affected Rows: 0
+
+SELECT host FROM monitoring_data WHERE cpu_usage = '23.7' ORDER BY host;
+
++--------+
+| host   |
++--------+
+| web-02 |
++--------+
+
+DROP TABLE monitoring_data;
+
+Affected Rows: 0
+
diff --git a/tests/cases/standalone/common/alter/change_col_type_inverted_index.sql b/tests/cases/standalone/common/alter/change_col_type_inverted_index.sql
new file mode 100644
index 0000000000..7020157758
--- /dev/null
+++ b/tests/cases/standalone/common/alter/change_col_type_inverted_index.sql
@@ -0,0 +1,29 @@
+-- Regression test for issue #8074
+-- https://github.com/GreptimeTeam/greptimedb/issues/8074
+CREATE TABLE monitoring_data (
+    host STRING INVERTED INDEX,
+    `region` STRING,
+    cpu_usage DOUBLE INVERTED INDEX,
+    `timestamp` TIMESTAMP TIME INDEX
+) WITH ('append_mode'='true');
+
+INSERT INTO monitoring_data (host, region, cpu_usage, `timestamp`) VALUES
+('web-01', 'us-east', 12.5, '2026-05-06 10:00:00'),
+('web-02', 'us-east', 18.3, '2026-05-06 10:00:00'),
+('web-03', 'us-east', 91.2, '2026-05-06 10:00:00'),
+('web-04', 'us-west', 73.8, '2026-05-06 10:00:00');
+
+INSERT INTO monitoring_data (host, region, cpu_usage, `timestamp`) VALUES
+('web-01', 'us-east', 15.2, '2026-05-06 10:01:00'),
+('web-02', 'us-east', 23.7, '2026-05-06 10:01:00'),
+('web-03', 'us-east', 94.5, '2026-05-06 10:01:00'),
+('web-04', 'us-west', 78.1, '2026-05-06 10:01:00');
+
+ADMIN FLUSH_TABLE('monitoring_data');
+
+ALTER TABLE monitoring_data
+MODIFY COLUMN cpu_usage STRING;
+
+SELECT host FROM monitoring_data WHERE cpu_usage = '23.7' ORDER BY host;
+
+DROP TABLE monitoring_data;
diff --git a/tests/cases/standalone/common/alter/prefilter_last_row_null.result b/tests/cases/standalone/common/alter/prefilter_last_row_null.result
new file mode 100644
index 0000000000..00aac5d9db
--- /dev/null
+++ b/tests/cases/standalone/common/alter/prefilter_last_row_null.result
@@ -0,0 +1,69 @@
+CREATE TABLE prefilter_last_row_null(
+    host STRING,
+    ts TIMESTAMP TIME INDEX,
+    cpu DOUBLE,
+    PRIMARY KEY(host)
+) ENGINE=mito;
+
+Affected Rows: 0
+
+INSERT INTO prefilter_last_row_null(host, ts, cpu) VALUES ('host1', 0, NULL);
+
+Affected Rows: 1
+
+ADMIN FLUSH_TABLE('prefilter_last_row_null');
+
++----------------------------------------------+
+| ADMIN FLUSH_TABLE('prefilter_last_row_null') |
++----------------------------------------------+
+| 0                                            |
++----------------------------------------------+
+
+INSERT INTO prefilter_last_row_null(host, ts, cpu) VALUES ('host1', 0, 10.0);
+
+Affected Rows: 1
+
+ADMIN FLUSH_TABLE('prefilter_last_row_null');
+
++----------------------------------------------+
+| ADMIN FLUSH_TABLE('prefilter_last_row_null') |
++----------------------------------------------+
+| 0                                            |
++----------------------------------------------+
+
+SELECT COUNT(*) FROM prefilter_last_row_null WHERE cpu IS NULL;
+
++----------+
+| count(*) |
++----------+
+| 0        |
++----------+
+
+ALTER TABLE prefilter_last_row_null ADD COLUMN memory DOUBLE NULL;
+
+Affected Rows: 0
+
+INSERT INTO prefilter_last_row_null(host, ts, cpu, memory) VALUES ('host1', 0, 10.0, 20.0);
+
+Affected Rows: 1
+
+ADMIN FLUSH_TABLE('prefilter_last_row_null');
+
++----------------------------------------------+
+| ADMIN FLUSH_TABLE('prefilter_last_row_null') |
++----------------------------------------------+
+| 0                                            |
++----------------------------------------------+
+
+SELECT COUNT(*) FROM prefilter_last_row_null WHERE memory IS NULL;
+
++----------+
+| count(*) |
++----------+
+| 0        |
++----------+
+
+DROP TABLE prefilter_last_row_null;
+
+Affected Rows: 0
+
diff --git a/tests/cases/standalone/common/alter/prefilter_last_row_null.sql b/tests/cases/standalone/common/alter/prefilter_last_row_null.sql
new file mode 100644
index 0000000000..d92dbf7829
--- /dev/null
+++ b/tests/cases/standalone/common/alter/prefilter_last_row_null.sql
@@ -0,0 +1,26 @@
+CREATE TABLE prefilter_last_row_null(
+    host STRING,
+    ts TIMESTAMP TIME INDEX,
+    cpu DOUBLE,
+    PRIMARY KEY(host)
+) ENGINE=mito;
+
+INSERT INTO prefilter_last_row_null(host, ts, cpu) VALUES ('host1', 0, NULL);
+
+ADMIN FLUSH_TABLE('prefilter_last_row_null');
+
+INSERT INTO prefilter_last_row_null(host, ts, cpu) VALUES ('host1', 0, 10.0);
+
+ADMIN FLUSH_TABLE('prefilter_last_row_null');
+
+SELECT COUNT(*) FROM prefilter_last_row_null WHERE cpu IS NULL;
+
+ALTER TABLE prefilter_last_row_null ADD COLUMN memory DOUBLE NULL;
+
+INSERT INTO prefilter_last_row_null(host, ts, cpu, memory) VALUES ('host1', 0, 10.0, 20.0);
+
+ADMIN FLUSH_TABLE('prefilter_last_row_null');
+
+SELECT COUNT(*) FROM prefilter_last_row_null WHERE memory IS NULL;
+
+DROP TABLE prefilter_last_row_null;
diff --git a/tests/cases/standalone/common/create/create_with_options.result b/tests/cases/standalone/common/create/create_with_options.result
index bb350a78aa..7c76ac6c63 100644
--- a/tests/cases/standalone/common/create/create_with_options.result
+++ b/tests/cases/standalone/common/create/create_with_options.result
@@ -81,6 +81,22 @@ drop table test_mito_options;
 
 Affected Rows: 0
 
+create table if not exists test_compaction_override_without_type(
+    host string,
+    ts timestamp,
+    memory double,
+    TIME INDEX (ts),
+    PRIMARY KEY(host)
+)
+engine=mito
+with('compaction.override'='true');
+
+Affected Rows: 0
+
+drop table test_compaction_override_without_type;
+
+Affected Rows: 0
+
 create table if not exists invalid_compaction(
     host string,
     ts timestamp,
diff --git a/tests/cases/standalone/common/create/create_with_options.sql b/tests/cases/standalone/common/create/create_with_options.sql
index 13824f76aa..2ab8cee1b8 100644
--- a/tests/cases/standalone/common/create/create_with_options.sql
+++ b/tests/cases/standalone/common/create/create_with_options.sql
@@ -67,6 +67,18 @@ with(
 
 drop table test_mito_options;
 
+create table if not exists test_compaction_override_without_type(
+    host string,
+    ts timestamp,
+    memory double,
+    TIME INDEX (ts),
+    PRIMARY KEY(host)
+)
+engine=mito
+with('compaction.override'='true');
+
+drop table test_compaction_override_without_type;
+
 create table if not exists invalid_compaction(
     host string,
     ts timestamp,
diff --git a/tests/cases/standalone/common/flow/flow_last_non_null.result b/tests/cases/standalone/common/flow/flow_last_non_null.result
index 64cc8d7eed..50cb46faa3 100644
--- a/tests/cases/standalone/common/flow/flow_last_non_null.result
+++ b/tests/cases/standalone/common/flow/flow_last_non_null.result
@@ -185,7 +185,7 @@ Affected Rows: 2
 -- should return error
 ADMIN FLUSH_FLOW('find_approx_rate');
 
-Error: 1002(Unexpected), Failed to execute admin function: Failed to execute admin function flush_flow, error: Execution error: Function execution error: Internal error: 1003: Execution error: Function execution error: Internal error: 1003
+Error: 1002(Unexpected), Failed to execute admin function flush_flow: Execution error: Internal error: 1003
 
 DROP FLOW find_approx_rate;
 
diff --git a/tests/cases/standalone/common/flow/show_create_flow.result b/tests/cases/standalone/common/flow/show_create_flow.result
index d18c95afde..113822cd18 100644
--- a/tests/cases/standalone/common/flow/show_create_flow.result
+++ b/tests/cases/standalone/common/flow/show_create_flow.result
@@ -442,6 +442,46 @@ DROP FLOW filter_numbers_show;
 
 Affected Rows: 0
 
+CREATE FLOW filter_numbers_show
+SINK TO out_num_cnt_show
+WITH (defer_on_missing_source = true)
+AS SELECT number AS n1 FROM numbers_input_show where number > 10;
+
+Affected Rows: 0
+
+SHOW CREATE FLOW filter_numbers_show;
+
++---------------------+------------------------------------------------------------------+
+| Flow                | Create Flow                                                      |
++---------------------+------------------------------------------------------------------+
+| filter_numbers_show | CREATE FLOW IF NOT EXISTS filter_numbers_show                    |
+|                     | SINK TO public.out_num_cnt_show                                  |
+|                     | WITH (defer_on_missing_source = 'true')                          |
+|                     | AS SELECT number AS n1 FROM numbers_input_show WHERE number > 10 |
++---------------------+------------------------------------------------------------------+
+
+SELECT flow_definition FROM INFORMATION_SCHEMA.FLOWS WHERE flow_name='filter_numbers_show';
+
++------------------------------------------------------------------+
+| flow_definition                                                  |
++------------------------------------------------------------------+
+| CREATE FLOW IF NOT EXISTS filter_numbers_show                    |
+| SINK TO public.out_num_cnt_show                                  |
+| WITH (defer_on_missing_source = 'true')                          |
+| AS SELECT number AS n1 FROM numbers_input_show WHERE number > 10 |
++------------------------------------------------------------------+
+
+CREATE FLOW invalid_flow_option_show
+SINK TO out_num_cnt_show
+WITH (access_key_id = [true])
+AS SELECT number AS n1 FROM numbers_input_show where number > 10;
+
+Error: 1004(InvalidArguments), Invalid SQL, error: unknown flow option 'access_key_id', supported options: defer_on_missing_source
+
+DROP FLOW filter_numbers_show;
+
+Affected Rows: 0
+
 drop table out_num_cnt_show;
 
 Affected Rows: 0
diff --git a/tests/cases/standalone/common/flow/show_create_flow.sql b/tests/cases/standalone/common/flow/show_create_flow.sql
index fdbc7f9cf1..9856a6041f 100644
--- a/tests/cases/standalone/common/flow/show_create_flow.sql
+++ b/tests/cases/standalone/common/flow/show_create_flow.sql
@@ -168,6 +168,22 @@ SELECT number FROM out_num_cnt_show;
 
 DROP FLOW filter_numbers_show;
 
+CREATE FLOW filter_numbers_show
+SINK TO out_num_cnt_show
+WITH (defer_on_missing_source = true)
+AS SELECT number AS n1 FROM numbers_input_show where number > 10;
+
+SHOW CREATE FLOW filter_numbers_show;
+
+SELECT flow_definition FROM INFORMATION_SCHEMA.FLOWS WHERE flow_name='filter_numbers_show';
+
+CREATE FLOW invalid_flow_option_show
+SINK TO out_num_cnt_show
+WITH (access_key_id = [true])
+AS SELECT number AS n1 FROM numbers_input_show where number > 10;
+
+DROP FLOW filter_numbers_show;
+
 drop table out_num_cnt_show;
 
 drop table numbers_input_show;
diff --git a/tests/cases/standalone/common/function/ip.result b/tests/cases/standalone/common/function/ip.result
index 21e6774f22..c4759c26e3 100644
--- a/tests/cases/standalone/common/function/ip.result
+++ b/tests/cases/standalone/common/function/ip.result
@@ -122,6 +122,15 @@ FROM ip_v4_data;
 | 8  | 0.0.0.0         | 0          | 0                | 0.0.0.0         |
 +----+-----------------+------------+------------------+-----------------+
 
+-- Test IPv4 number to string conversion function without explicit cast
+SELECT ipv4_num_to_string(3232235521);
+
++---------------------------------------+
+| ipv4_num_to_string(Int64(3232235521)) |
++---------------------------------------+
+| 192.168.0.1                           |
++---------------------------------------+
+
 -- Test IPv4 CIDR functions
 -- SQLNESS SORT_RESULT 3 1
 SELECT
diff --git a/tests/cases/standalone/common/function/ip.sql b/tests/cases/standalone/common/function/ip.sql
index 344d43f46e..ddac31f2a5 100644
--- a/tests/cases/standalone/common/function/ip.sql
+++ b/tests/cases/standalone/common/function/ip.sql
@@ -84,6 +84,9 @@ SELECT
     inet_ntoa(ip_numeric) AS computed_addr
 FROM ip_v4_data;
 
+-- Test IPv4 number to string conversion function without explicit cast
+SELECT ipv4_num_to_string(3232235521);
+
 -- Test IPv4 CIDR functions
 -- SQLNESS SORT_RESULT 3 1
 SELECT
diff --git a/tests/cases/standalone/common/function/matches_term.result b/tests/cases/standalone/common/function/matches_term.result
index 38b9f30723..37ecf5a55f 100644
--- a/tests/cases/standalone/common/function/matches_term.result
+++ b/tests/cases/standalone/common/function/matches_term.result
@@ -157,6 +157,111 @@ SELECT matches_term('русский!', 'русский') as result;
 | true   |
 +--------+
 
+-- Phase 1 mixed Chinese and numeric behavior
+SELECT matches_term('登录手机号18888888888的动态key', '手机号') as result;
+
++--------+
+| result |
++--------+
+| true   |
++--------+
+
+SELECT matches_term('登录手机号18888888888的动态key', '18888888888') as result;
+
++--------+
+| result |
++--------+
+| true   |
++--------+
+
+SELECT matches_term('登录手机号18888888888的动态key', '手机') as result;
+
++--------+
+| result |
++--------+
+| true   |
++--------+
+
+SELECT matches_term('登录手机号18888888888的动态key', '机号') as result;
+
++--------+
+| result |
++--------+
+| true   |
++--------+
+
+SELECT matches_term('登录手机号18888888888的动态key', '机号1888') as result;
+
++--------+
+| result |
++--------+
+| true   |
++--------+
+
+SELECT matches_term('中国农业银行', '农业') as result;
+
++--------+
+| result |
++--------+
+| true   |
++--------+
+
+SELECT matches_term('中国农业银行账号', '行账号') as result;
+
++--------+
+| result |
++--------+
+| true   |
++--------+
+
+SELECT matches_term('错误error日志', 'error') as result;
+
++--------+
+| result |
++--------+
+| true   |
++--------+
+
+SELECT matches_term('trace_id=abc', 'trace_id') as result;
+
++--------+
+| result |
++--------+
+| true   |
++--------+
+
+SELECT matches_term('ship__ship', 'ship__ship') as result;
+
++--------+
+| result |
++--------+
+| true   |
++--------+
+
+SELECT matches_term('__IDENTIFIER__', '__IDENTIFIER__') as result;
+
++--------+
+| result |
++--------+
+| true   |
++--------+
+
+SELECT matches_term('_ship', '_ship') as result;
+
++--------+
+| result |
++--------+
+| true   |
++--------+
+
+SELECT matches_term('ship_', 'ship_') as result;
+
++--------+
+| result |
++--------+
+| true   |
++--------+
+
 -- Test complete word matching
 CREATE TABLE logs (
     `id` TIMESTAMP TIME INDEX,
@@ -312,3 +417,94 @@ DROP TABLE logs;
 
 Affected Rows: 0
 
+CREATE TABLE zh_logs (
+    `id` TIMESTAMP TIME INDEX,
+    `log_message` STRING FULLTEXT INDEX WITH(analyzer = 'Chinese', backend = 'bloom', case_sensitive = 'true', false_positive_rate = '0.01', granularity = '10240')
+);
+
+Affected Rows: 0
+
+INSERT INTO zh_logs VALUES
+    (1, '[2026/04/09/ 13:56:11.031]2026-04-09 13:56:11.031 - [ trace_id=340a6a44b0bd8e37bb7697ss7da61ff0 span_id=085ff5ttf1e0a23b trace_flags=01] - [http-nio-8081-exec-16] INFO c.h.p.xx.web.service.impl.CCCXForwardKKKServiceImpl.pushout(188) - 登录手机号18888888888的动态key：829889AC8 ship_ship ship__ship __IDENTIFIER__ _ship ship_'),
+    (2, '哈基米曼波');
+
+Affected Rows: 2
+
+ADMIN flush_table('zh_logs');
+
++------------------------------+
+| ADMIN flush_table('zh_logs') |
++------------------------------+
+| 0                            |
++------------------------------+
+
+SELECT * FROM zh_logs where `log_message` @@ 'trace_id';
+
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| id                      | log_message                                                                                                                                                                                                                                                                                                                          |
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| 1970-01-01T00:00:00.001 | [2026/04/09/ 13:56:11.031]2026-04-09 13:56:11.031 - [ trace_id=340a6a44b0bd8e37bb7697ss7da61ff0 span_id=085ff5ttf1e0a23b trace_flags=01] - [http-nio-8081-exec-16] INFO c.h.p.xx.web.service.impl.CCCXForwardKKKServiceImpl.pushout(188) - 登录手机号18888888888的动态key：829889AC8 ship_ship ship__ship __IDENTIFIER__ _ship ship_ |
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+
+SELECT * FROM zh_logs where `log_message` @@ 'ship_ship';
+
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| id                      | log_message                                                                                                                                                                                                                                                                                                                          |
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| 1970-01-01T00:00:00.001 | [2026/04/09/ 13:56:11.031]2026-04-09 13:56:11.031 - [ trace_id=340a6a44b0bd8e37bb7697ss7da61ff0 span_id=085ff5ttf1e0a23b trace_flags=01] - [http-nio-8081-exec-16] INFO c.h.p.xx.web.service.impl.CCCXForwardKKKServiceImpl.pushout(188) - 登录手机号18888888888的动态key：829889AC8 ship_ship ship__ship __IDENTIFIER__ _ship ship_ |
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+
+SELECT * FROM zh_logs where `log_message` @@ 'ship__ship';
+
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| id                      | log_message                                                                                                                                                                                                                                                                                                                          |
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| 1970-01-01T00:00:00.001 | [2026/04/09/ 13:56:11.031]2026-04-09 13:56:11.031 - [ trace_id=340a6a44b0bd8e37bb7697ss7da61ff0 span_id=085ff5ttf1e0a23b trace_flags=01] - [http-nio-8081-exec-16] INFO c.h.p.xx.web.service.impl.CCCXForwardKKKServiceImpl.pushout(188) - 登录手机号18888888888的动态key：829889AC8 ship_ship ship__ship __IDENTIFIER__ _ship ship_ |
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+
+SELECT * FROM zh_logs where `log_message` @@ '__IDENTIFIER__';
+
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| id                      | log_message                                                                                                                                                                                                                                                                                                                          |
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| 1970-01-01T00:00:00.001 | [2026/04/09/ 13:56:11.031]2026-04-09 13:56:11.031 - [ trace_id=340a6a44b0bd8e37bb7697ss7da61ff0 span_id=085ff5ttf1e0a23b trace_flags=01] - [http-nio-8081-exec-16] INFO c.h.p.xx.web.service.impl.CCCXForwardKKKServiceImpl.pushout(188) - 登录手机号18888888888的动态key：829889AC8 ship_ship ship__ship __IDENTIFIER__ _ship ship_ |
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+
+SELECT * FROM zh_logs where `log_message` @@ '_ship';
+
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| id                      | log_message                                                                                                                                                                                                                                                                                                                          |
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| 1970-01-01T00:00:00.001 | [2026/04/09/ 13:56:11.031]2026-04-09 13:56:11.031 - [ trace_id=340a6a44b0bd8e37bb7697ss7da61ff0 span_id=085ff5ttf1e0a23b trace_flags=01] - [http-nio-8081-exec-16] INFO c.h.p.xx.web.service.impl.CCCXForwardKKKServiceImpl.pushout(188) - 登录手机号18888888888的动态key：829889AC8 ship_ship ship__ship __IDENTIFIER__ _ship ship_ |
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+
+SELECT * FROM zh_logs where `log_message` @@ 'ship_';
+
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| id                      | log_message                                                                                                                                                                                                                                                                                                                          |
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| 1970-01-01T00:00:00.001 | [2026/04/09/ 13:56:11.031]2026-04-09 13:56:11.031 - [ trace_id=340a6a44b0bd8e37bb7697ss7da61ff0 span_id=085ff5ttf1e0a23b trace_flags=01] - [http-nio-8081-exec-16] INFO c.h.p.xx.web.service.impl.CCCXForwardKKKServiceImpl.pushout(188) - 登录手机号18888888888的动态key：829889AC8 ship_ship ship__ship __IDENTIFIER__ _ship ship_ |
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+
+SELECT * FROM zh_logs where `log_message` @@ '登录_id';
+
+++
+++
+
+SELECT * FROM zh_logs where `log_message` @@ '手机号_trace';
+
+++
+++
+
+SELECT * FROM zh_logs where `log_message` @@ '手机';
+
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| id                      | log_message                                                                                                                                                                                                                                                                                                                          |
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| 1970-01-01T00:00:00.001 | [2026/04/09/ 13:56:11.031]2026-04-09 13:56:11.031 - [ trace_id=340a6a44b0bd8e37bb7697ss7da61ff0 span_id=085ff5ttf1e0a23b trace_flags=01] - [http-nio-8081-exec-16] INFO c.h.p.xx.web.service.impl.CCCXForwardKKKServiceImpl.pushout(188) - 登录手机号18888888888的动态key：829889AC8 ship_ship ship__ship __IDENTIFIER__ _ship ship_ |
++-------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+
+DROP TABLE zh_logs;
+
+Affected Rows: 0
+
diff --git a/tests/cases/standalone/common/function/matches_term.sql b/tests/cases/standalone/common/function/matches_term.sql
index e91b9fdf5f..bbffd9cbe9 100644
--- a/tests/cases/standalone/common/function/matches_term.sql
+++ b/tests/cases/standalone/common/function/matches_term.sql
@@ -47,6 +47,21 @@ SELECT matches_term('café>', 'café') as result;
 -- Expect: true
 SELECT matches_term('русский!', 'русский') as result;
 
+-- Phase 1 mixed Chinese and numeric behavior
+SELECT matches_term('登录手机号18888888888的动态key', '手机号') as result;
+SELECT matches_term('登录手机号18888888888的动态key', '18888888888') as result;
+SELECT matches_term('登录手机号18888888888的动态key', '手机') as result;
+SELECT matches_term('登录手机号18888888888的动态key', '机号') as result;
+SELECT matches_term('登录手机号18888888888的动态key', '机号1888') as result;
+SELECT matches_term('中国农业银行', '农业') as result;
+SELECT matches_term('中国农业银行账号', '行账号') as result;
+SELECT matches_term('错误error日志', 'error') as result;
+SELECT matches_term('trace_id=abc', 'trace_id') as result;
+SELECT matches_term('ship__ship', 'ship__ship') as result;
+SELECT matches_term('__IDENTIFIER__', '__IDENTIFIER__') as result;
+SELECT matches_term('_ship', '_ship') as result;
+SELECT matches_term('ship_', 'ship_') as result;
+
 -- Test complete word matching
 CREATE TABLE logs (
     `id` TIMESTAMP TIME INDEX,
@@ -140,3 +155,34 @@ FROM logs WHERE `id` BETWEEN 20 AND 23 ORDER BY `id`;
 SELECT `id`, `log_message`, matches_term(lower(`log_message`), 'warning') as `matches_warning` FROM logs WHERE `id` >= 24 ORDER BY `id`;
 
 DROP TABLE logs;
+
+CREATE TABLE zh_logs (
+    `id` TIMESTAMP TIME INDEX,
+    `log_message` STRING FULLTEXT INDEX WITH(analyzer = 'Chinese', backend = 'bloom', case_sensitive = 'true', false_positive_rate = '0.01', granularity = '10240')
+);
+
+INSERT INTO zh_logs VALUES
+    (1, '[2026/04/09/ 13:56:11.031]2026-04-09 13:56:11.031 - [ trace_id=340a6a44b0bd8e37bb7697ss7da61ff0 span_id=085ff5ttf1e0a23b trace_flags=01] - [http-nio-8081-exec-16] INFO c.h.p.xx.web.service.impl.CCCXForwardKKKServiceImpl.pushout(188) - 登录手机号18888888888的动态key：829889AC8 ship_ship ship__ship __IDENTIFIER__ _ship ship_'),
+    (2, '哈基米曼波');
+
+ADMIN flush_table('zh_logs');
+
+SELECT * FROM zh_logs where `log_message` @@ 'trace_id';
+
+SELECT * FROM zh_logs where `log_message` @@ 'ship_ship';
+
+SELECT * FROM zh_logs where `log_message` @@ 'ship__ship';
+
+SELECT * FROM zh_logs where `log_message` @@ '__IDENTIFIER__';
+
+SELECT * FROM zh_logs where `log_message` @@ '_ship';
+
+SELECT * FROM zh_logs where `log_message` @@ 'ship_';
+
+SELECT * FROM zh_logs where `log_message` @@ '登录_id';
+
+SELECT * FROM zh_logs where `log_message` @@ '手机号_trace';
+
+SELECT * FROM zh_logs where `log_message` @@ '手机';
+
+DROP TABLE zh_logs;
diff --git a/tests/cases/standalone/common/function/system.result b/tests/cases/standalone/common/function/system.result
index 0cb6839292..d2007b5240 100644
--- a/tests/cases/standalone/common/function/system.result
+++ b/tests/cases/standalone/common/function/system.result
@@ -8,7 +8,7 @@ SELECT build();
 
 ++|build()|++|branch:BRANCH|commit:COMMIT|commit_short:COMMIT_SHORT|clean:CLEAN|version:VERSION++
 
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) VERSION
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) VERSION
 -- SQLNESS REPLACE [\s\-]+
 SELECT version();
 
diff --git a/tests/cases/standalone/common/function/system.sql b/tests/cases/standalone/common/function/system.sql
index 8ae1475311..6504b48679 100644
--- a/tests/cases/standalone/common/function/system.sql
+++ b/tests/cases/standalone/common/function/system.sql
@@ -6,7 +6,7 @@
 -- SQLNESS REPLACE [\s\-]+
 SELECT build();
 
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) VERSION
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) VERSION
 -- SQLNESS REPLACE [\s\-]+
 SELECT version();
 
diff --git a/tests/cases/standalone/common/function/vector/vector_index_explain.result b/tests/cases/standalone/common/function/vector/vector_index_explain.result
index 246a49f405..10351cce11 100644
--- a/tests/cases/standalone/common/function/vector/vector_index_explain.result
+++ b/tests/cases/standalone/common/function/vector/vector_index_explain.result
@@ -41,6 +41,9 @@ ADMIN FLUSH_TABLE('vectors_explain');
 -- SQLNESS REPLACE Total\s+rows:\s+\d+ Total rows: REDACTED
 -- SQLNESS REPLACE (peers.*) REDACTED
 -- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE ,\s"dyn_filters":\s\[.* REDACTED
+-- SQLNESS REPLACE ,\s"vector_index_k":\s\d+
+-- SQLNESS REPLACE "index_size":\d+ "index_size":REDACTED
 EXPLAIN ANALYZE VERBOSE
 SELECT vec_id
 FROM vectors_explain
@@ -56,7 +59,7 @@ LIMIT 2;
 | 1_| 0_|_ProjectionExec: expr=[vec_id@0 as vec_id] metrics=REDACTED_|
 |_|_|_SortPreservingMergeExec: [vec_l2sq_distance(embedding@1, [1.0, 0.0]) ASC NULLS LAST, vec_id@0 ASC NULLS LAST], fetch=2 metrics=REDACTED_|
 |_|_|_SortExec: TopK(fetch=2), expr=[vec_l2sq_distance(embedding@1, [1.0, 0.0]) ASC NULLS LAST, vec_id@0 ASC NULLS LAST], preserve_partitioning=[true] metrics=REDACTED_|
-|_|_|_SeqScan: region=REDACTED, {"partition_count":{"count":1, "mem_ranges":0, "files":1, "file_ranges":1}, "projection": ["vec_id", "embedding"], "dyn_filters": ["DynamicFilter [ vec_l2sq_distance(embedding@1, [1.0, 0.0]) < 0.010000004 OR vec_l2sq_distance(embedding@1, [1.0, 0.0]) = 0.010000004 AND vec_id@0 < 2 ]"], "vector_index_k": 2, "files": [{"file_id":"REDACTED","time_range_start":"REDACTED","time_range_end":"REDACTED","rows":4,"size":REDACTED,"index_size":893}], "flat_format":REDACTED, "metrics_per_partition": REDACTED metrics=REDACTED |
+|_|_|_SeqScan: region=REDACTED, {"partition_count":{"count":1, "mem_ranges":0, "files":1, "file_ranges":1}, "projection": ["vec_id", "embedding"]REDACTED
 |_|_|_|
 |_|_| Total rows: REDACTED_|
 +-+-+-+
@@ -122,6 +125,7 @@ ADMIN FLUSH_TABLE('vectors_explain_left');
 -- SQLNESS REPLACE Total\s+rows:\s+\d+ Total rows: REDACTED
 -- SQLNESS REPLACE (peers.*) REDACTED
 -- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE SeqScan:.* SeqScan: REDACTED 
 EXPLAIN ANALYZE VERBOSE
 SELECT l.vec_id
 FROM vectors_explain_left l
@@ -144,10 +148,10 @@ LIMIT 1;
 |_|_|_MergeScanExec: REDACTED
 |_|_|_|
 | 1_| 0_|_CooperativeExec metrics=REDACTED_|
-|_|_|_SeqScan: region=REDACTED, {"partition_count":{"count":1, "mem_ranges":0, "files":1, "file_ranges":1}, "projection": ["vec_id", "ts", "embedding"], "files": [{"file_id":"REDACTED","time_range_start":"REDACTED","time_range_end":"REDACTED","rows":4,"size":REDACTED,"index_size":893}], "flat_format":REDACTED, "metrics_per_partition": REDACTED metrics=REDACTED |
+|_|_|_SeqScan: REDACTED
 |_|_|_|
 | 1_| 0_|_CooperativeExec metrics=REDACTED_|
-|_|_|_SeqScan: region=REDACTED, {"partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0}, "projection": ["vec_id", "note", "ts"], "flat_format":REDACTED, "metrics_per_partition": REDACTED metrics=REDACTED_|
+|_|_|_SeqScan: REDACTED
 |_|_|_|
 |_|_| Total rows: REDACTED_|
 +-+-+-+
@@ -202,6 +206,9 @@ ADMIN FLUSH_TABLE('vectors_explain_metric');
 -- SQLNESS REPLACE Total\s+rows:\s+\d+ Total rows: REDACTED
 -- SQLNESS REPLACE (peers.*) REDACTED
 -- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE SeqScan:.* SeqScan: REDACTED 
+-- SQLNESS REPLACE ,\s"vector_index_k":\s\d+
+-- SQLNESS REPLACE "index_size":\d+ "index_size":REDACTED
 EXPLAIN ANALYZE VERBOSE
 SELECT vec_id
 FROM vectors_explain_metric
@@ -217,7 +224,7 @@ LIMIT 2;
 | 1_| 0_|_ProjectionExec: expr=[vec_id@0 as vec_id] metrics=REDACTED_|
 |_|_|_SortPreservingMergeExec: [vec_cos_distance(embedding@1, [1.0, 0.0]) ASC NULLS LAST, vec_id@0 ASC NULLS LAST], fetch=2 metrics=REDACTED_|
 |_|_|_SortExec: TopK(fetch=2), expr=[vec_cos_distance(embedding@1, [1.0, 0.0]) ASC NULLS LAST, vec_id@0 ASC NULLS LAST], preserve_partitioning=[true] metrics=REDACTED_|
-|_|_|_SeqScan: region=REDACTED, {"partition_count":{"count":1, "mem_ranges":0, "files":1, "file_ranges":1}, "projection": ["vec_id", "embedding"], "dyn_filters": ["DynamicFilter [ vec_cos_distance(embedding@1, [1.0, 0.0]) < 1 OR vec_cos_distance(embedding@1, [1.0, 0.0]) = 1 AND vec_id@0 < 4 ]"], "vector_index_k": 2, "files": [{"file_id":"REDACTED","time_range_start":"REDACTED","time_range_end":"REDACTED","rows":4,"size":REDACTED,"index_size":895}], "flat_format":REDACTED, "metrics_per_partition": REDACTED metrics=REDACTED |
+|_|_|_SeqScan: REDACTED
 |_|_|_|
 |_|_| Total rows: REDACTED_|
 +-+-+-+
@@ -236,6 +243,7 @@ LIMIT 2;
 -- SQLNESS REPLACE Total\s+rows:\s+\d+ Total rows: REDACTED
 -- SQLNESS REPLACE (peers.*) REDACTED
 -- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE SeqScan:.* SeqScan: REDACTED 
 EXPLAIN ANALYZE VERBOSE
 SELECT vec_id
 FROM vectors_explain_metric
@@ -251,7 +259,7 @@ LIMIT 2;
 | 1_| 0_|_ProjectionExec: expr=[vec_id@0 as vec_id] metrics=REDACTED_|
 |_|_|_SortPreservingMergeExec: [vec_dot_product(embedding@1, [1.0, 0.0]) DESC, vec_id@0 ASC NULLS LAST], fetch=2 metrics=REDACTED_|
 |_|_|_SortExec: TopK(fetch=2), expr=[vec_dot_product(embedding@1, [1.0, 0.0]) DESC, vec_id@0 ASC NULLS LAST], preserve_partitioning=[true] metrics=REDACTED_|
-|_|_|_SeqScan: region=REDACTED, {"partition_count":{"count":1, "mem_ranges":0, "files":1, "file_ranges":1}, "projection": ["vec_id", "embedding"], "dyn_filters": ["DynamicFilter [ vec_dot_product(embedding@1, [1.0, 0.0]) IS NULL OR vec_dot_product(embedding@1, [1.0, 0.0]) > 0 OR vec_dot_product(embedding@1, [1.0, 0.0]) = 0 AND vec_id@0 < 2 ]"], "vector_index_k": 2, "files": [{"file_id":"REDACTED","time_range_start":"REDACTED","time_range_end":"REDACTED","rows":4,"size":REDACTED,"index_size":895}], "flat_format":REDACTED, "metrics_per_partition": REDACTED metrics=REDACTED |
+|_|_|_SeqScan: REDACTED
 |_|_|_|
 |_|_| Total rows: REDACTED_|
 +-+-+-+
diff --git a/tests/cases/standalone/common/function/vector/vector_index_explain.sql b/tests/cases/standalone/common/function/vector/vector_index_explain.sql
index 50bf3cdbaa..7a0330b0ea 100644
--- a/tests/cases/standalone/common/function/vector/vector_index_explain.sql
+++ b/tests/cases/standalone/common/function/vector/vector_index_explain.sql
@@ -33,6 +33,9 @@ ADMIN FLUSH_TABLE('vectors_explain');
 -- SQLNESS REPLACE Total\s+rows:\s+\d+ Total rows: REDACTED
 -- SQLNESS REPLACE (peers.*) REDACTED
 -- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE ,\s"dyn_filters":\s\[.* REDACTED
+-- SQLNESS REPLACE ,\s"vector_index_k":\s\d+
+-- SQLNESS REPLACE "index_size":\d+ "index_size":REDACTED
 EXPLAIN ANALYZE VERBOSE
 SELECT vec_id
 FROM vectors_explain
@@ -84,6 +87,7 @@ ADMIN FLUSH_TABLE('vectors_explain_left');
 -- SQLNESS REPLACE Total\s+rows:\s+\d+ Total rows: REDACTED
 -- SQLNESS REPLACE (peers.*) REDACTED
 -- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE SeqScan:.* SeqScan: REDACTED 
 EXPLAIN ANALYZE VERBOSE
 SELECT l.vec_id
 FROM vectors_explain_left l
@@ -126,6 +130,9 @@ ADMIN FLUSH_TABLE('vectors_explain_metric');
 -- SQLNESS REPLACE Total\s+rows:\s+\d+ Total rows: REDACTED
 -- SQLNESS REPLACE (peers.*) REDACTED
 -- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE SeqScan:.* SeqScan: REDACTED 
+-- SQLNESS REPLACE ,\s"vector_index_k":\s\d+
+-- SQLNESS REPLACE "index_size":\d+ "index_size":REDACTED
 EXPLAIN ANALYZE VERBOSE
 SELECT vec_id
 FROM vectors_explain_metric
@@ -146,6 +153,7 @@ LIMIT 2;
 -- SQLNESS REPLACE Total\s+rows:\s+\d+ Total rows: REDACTED
 -- SQLNESS REPLACE (peers.*) REDACTED
 -- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE SeqScan:.* SeqScan: REDACTED 
 EXPLAIN ANALYZE VERBOSE
 SELECT vec_id
 FROM vectors_explain_metric
diff --git a/tests/cases/standalone/common/information_schema/ssts.result b/tests/cases/standalone/common/information_schema/ssts.result
index bf0642f667..880a4878cb 100644
--- a/tests/cases/standalone/common/information_schema/ssts.result
+++ b/tests/cases/standalone/common/information_schema/ssts.result
@@ -92,7 +92,7 @@ ADMIN FLUSH_TABLE('sst_case');
 
 -- SQLNESS REPLACE (\s+\d+\s+) <NUM>
 -- SQLNESS REPLACE ([0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}) <UUID>
--- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3,9}) <DATETIME>
+-- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(\.\d{3,9})?) <DATETIME>
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>
 -- SQLNESS REPLACE (/public/\d+) /public/<TABLE_ID>
 SELECT * FROM information_schema.ssts_manifest order by file_path;
@@ -107,7 +107,7 @@ SELECT * FROM information_schema.ssts_manifest order by file_path;
 
 -- SQLNESS REPLACE (\s+\d+\s+) <NUM>
 -- SQLNESS REPLACE ([0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}) <UUID>
--- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3,9}) <DATETIME>
+-- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(\.\d{3,9})?) <DATETIME>
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+/index/<UUID>\.puffin) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>
 -- SQLNESS REPLACE (/public/\d+) /public/<TABLE_ID>
@@ -129,19 +129,20 @@ SELECT * FROM information_schema.ssts_index_meta ORDER BY meta_json;
 
 -- SQLNESS REPLACE (\s+\d+\s+) <NUM>
 -- SQLNESS REPLACE ([0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}) <UUID>
+-- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(\.\d{3,9})?) <DATETIME>
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>
 SELECT * FROM information_schema.ssts_storage order by file_path;
 
-+---------------------------------------------------------------------------------------------+-----------+------------------+---------+
-| file_path                                                                                   | file_size | last_modified_ms | node_id |
-+---------------------------------------------------------------------------------------------+-----------+------------------+---------+
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |           |                  |<NUM>|
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |           |                  |<NUM>|
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |           |                  |<NUM>|
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |           |                  |<NUM>|
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |           |                  |<NUM>|
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |           |                  |<NUM>|
-+---------------------------------------------------------------------------------------------+-----------+------------------+---------+
++---------------------------------------------------------------------------------------------+-----------+-------------------------+---------+
+| file_path                                                                                   | file_size | last_modified_ms        | node_id |
++---------------------------------------------------------------------------------------------+-----------+-------------------------+---------+
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |<NUM>| <DATETIME> |<NUM>|
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |<NUM>| <DATETIME> |<NUM>|
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |<NUM>| <DATETIME> |<NUM>|
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |<NUM>| <DATETIME> |<NUM>|
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |<NUM>| <DATETIME> |<NUM>|
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |<NUM>| <DATETIME> |<NUM>|
++---------------------------------------------------------------------------------------------+-----------+-------------------------+---------+
 
 INSERT INTO sst_case VALUES
   (24, 'foo', 'foo', 100),
@@ -160,7 +161,7 @@ ADMIN FLUSH_TABLE('sst_case');
 
 -- SQLNESS REPLACE (\s+\d+\s+) <NUM>
 -- SQLNESS REPLACE ([0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}) <UUID>
--- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3,9}) <DATETIME>
+-- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(\.\d{3,9})?) <DATETIME>
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>
 -- SQLNESS REPLACE (/public/\d+) /public/<TABLE_ID>
 SELECT * FROM information_schema.ssts_manifest order by file_path;
@@ -177,7 +178,7 @@ SELECT * FROM information_schema.ssts_manifest order by file_path;
 
 -- SQLNESS REPLACE (\s+\d+\s+) <NUM>
 -- SQLNESS REPLACE ([0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}) <UUID>
--- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3,9}) <DATETIME>
+-- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(\.\d{3,9})?) <DATETIME>
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+/index/<UUID>\.puffin) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>
 -- SQLNESS REPLACE (/public/\d+) /public/<TABLE_ID>
@@ -205,23 +206,24 @@ SELECT * FROM information_schema.ssts_index_meta ORDER BY meta_json;
 
 -- SQLNESS REPLACE (\s+\d+\s+) <NUM>
 -- SQLNESS REPLACE ([0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}) <UUID>
+-- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(\.\d{3,9})?) <DATETIME>
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>
 SELECT * FROM information_schema.ssts_storage order by file_path;
 
-+---------------------------------------------------------------------------------------------+-----------+------------------+---------+
-| file_path                                                                                   | file_size | last_modified_ms | node_id |
-+---------------------------------------------------------------------------------------------+-----------+------------------+---------+
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |           |                  |<NUM>|
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |           |                  |<NUM>|
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |           |                  |<NUM>|
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |           |                  |<NUM>|
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |           |                  |<NUM>|
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |           |                  |<NUM>|
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |           |                  |<NUM>|
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |           |                  |<NUM>|
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |           |                  |<NUM>|
-| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |           |                  |<NUM>|
-+---------------------------------------------------------------------------------------------+-----------+------------------+---------+
++---------------------------------------------------------------------------------------------+-----------+-------------------------+---------+
+| file_path                                                                                   | file_size | last_modified_ms        | node_id |
++---------------------------------------------------------------------------------------------+-----------+-------------------------+---------+
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |<NUM>| <DATETIME> |<NUM>|
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |<NUM>| <DATETIME> |<NUM>|
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |<NUM>| <DATETIME> |<NUM>|
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |<NUM>| <DATETIME> |<NUM>|
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |<NUM>| <DATETIME> |<NUM>|
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |<NUM>| <DATETIME> |<NUM>|
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |<NUM>| <DATETIME> |<NUM>|
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |<NUM>| <DATETIME> |<NUM>|
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/<UUID>.parquet      |<NUM>| <DATETIME> |<NUM>|
+| data/greptime/public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin |<NUM>| <DATETIME> |<NUM>|
++---------------------------------------------------------------------------------------------+-----------+-------------------------+---------+
 
 DROP TABLE sst_case;
 
diff --git a/tests/cases/standalone/common/information_schema/ssts.sql b/tests/cases/standalone/common/information_schema/ssts.sql
index f411c8f67a..408cec353c 100644
--- a/tests/cases/standalone/common/information_schema/ssts.sql
+++ b/tests/cases/standalone/common/information_schema/ssts.sql
@@ -25,14 +25,14 @@ ADMIN FLUSH_TABLE('sst_case');
 
 -- SQLNESS REPLACE (\s+\d+\s+) <NUM>
 -- SQLNESS REPLACE ([0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}) <UUID>
--- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3,9}) <DATETIME>
+-- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(\.\d{3,9})?) <DATETIME>
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>
 -- SQLNESS REPLACE (/public/\d+) /public/<TABLE_ID>
 SELECT * FROM information_schema.ssts_manifest order by file_path;
 
 -- SQLNESS REPLACE (\s+\d+\s+) <NUM>
 -- SQLNESS REPLACE ([0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}) <UUID>
--- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3,9}) <DATETIME>
+-- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(\.\d{3,9})?) <DATETIME>
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+/index/<UUID>\.puffin) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>
 -- SQLNESS REPLACE (/public/\d+) /public/<TABLE_ID>
@@ -40,6 +40,7 @@ SELECT * FROM information_schema.ssts_index_meta ORDER BY meta_json;
 
 -- SQLNESS REPLACE (\s+\d+\s+) <NUM>
 -- SQLNESS REPLACE ([0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}) <UUID>
+-- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(\.\d{3,9})?) <DATETIME>
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>
 SELECT * FROM information_schema.ssts_storage order by file_path;
 
@@ -52,14 +53,14 @@ ADMIN FLUSH_TABLE('sst_case');
 
 -- SQLNESS REPLACE (\s+\d+\s+) <NUM>
 -- SQLNESS REPLACE ([0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}) <UUID>
--- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3,9}) <DATETIME>
+-- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(\.\d{3,9})?) <DATETIME>
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>
 -- SQLNESS REPLACE (/public/\d+) /public/<TABLE_ID>
 SELECT * FROM information_schema.ssts_manifest order by file_path;
 
 -- SQLNESS REPLACE (\s+\d+\s+) <NUM>
 -- SQLNESS REPLACE ([0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}) <UUID>
--- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3,9}) <DATETIME>
+-- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(\.\d{3,9})?) <DATETIME>
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+/index/<UUID>\.puffin) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>/index/<UUID>.puffin
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>
 -- SQLNESS REPLACE (/public/\d+) /public/<TABLE_ID>
@@ -67,6 +68,7 @@ SELECT * FROM information_schema.ssts_index_meta ORDER BY meta_json;
 
 -- SQLNESS REPLACE (\s+\d+\s+) <NUM>
 -- SQLNESS REPLACE ([0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}) <UUID>
+-- SQLNESS REPLACE (\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(\.\d{3,9})?) <DATETIME>
 -- SQLNESS REPLACE (/public/\d+/\d+_\d+) /public/<TABLE_ID>/<REGION_ID>_<REGION_NUMBER>
 SELECT * FROM information_schema.ssts_storage order by file_path;
 
diff --git a/tests/cases/standalone/common/insert/logical_metric_table.result b/tests/cases/standalone/common/insert/logical_metric_table.result
index 80e765ccd6..32085175ea 100644
--- a/tests/cases/standalone/common/insert/logical_metric_table.result
+++ b/tests/cases/standalone/common/insert/logical_metric_table.result
@@ -77,6 +77,40 @@ DROP TABLE phy;
 
 Affected Rows: 0
 
+CREATE TABLE phy_default (ts timestamp time index, val double default 42) engine=metric with ("physical_metric_table" = "");
+
+Affected Rows: 0
+
+CREATE TABLE t_default (ts timestamp time index, val double default 42, host string primary key) engine = metric with ("on_physical_table" = "phy_default");
+
+Affected Rows: 0
+
+INSERT INTO t_default (host, ts) VALUES ('host1', 0), ('host2', 1);
+
+Affected Rows: 2
+
+SELECT host, ts, val FROM t_default ORDER BY host;
+
++-------+-------------------------+------+
+| host  | ts                      | val  |
++-------+-------------------------+------+
+| host1 | 1970-01-01T00:00:00     | 42.0 |
+| host2 | 1970-01-01T00:00:00.001 | 42.0 |
++-------+-------------------------+------+
+
+-- SQLNESS REPLACE (region\s\d+\(\d+\,\s\d+\)) region
+INSERT INTO t_default (host, val) VALUES ('host3', 3);
+
+Error: 1004(InvalidArguments), Invalid request for region, reason: missing required time index column ts
+
+DROP TABLE t_default;
+
+Affected Rows: 0
+
+DROP TABLE phy_default;
+
+Affected Rows: 0
+
 CREATE TABLE phy (
     ts timestamp time index,
     val double
diff --git a/tests/cases/standalone/common/insert/logical_metric_table.sql b/tests/cases/standalone/common/insert/logical_metric_table.sql
index 9899699c66..6a3bed401d 100644
--- a/tests/cases/standalone/common/insert/logical_metric_table.sql
+++ b/tests/cases/standalone/common/insert/logical_metric_table.sql
@@ -24,6 +24,21 @@ SELECT ts, val, __tsid, host, job FROM phy;
 
 DROP TABLE phy;
 
+CREATE TABLE phy_default (ts timestamp time index, val double default 42) engine=metric with ("physical_metric_table" = "");
+
+CREATE TABLE t_default (ts timestamp time index, val double default 42, host string primary key) engine = metric with ("on_physical_table" = "phy_default");
+
+INSERT INTO t_default (host, ts) VALUES ('host1', 0), ('host2', 1);
+
+SELECT host, ts, val FROM t_default ORDER BY host;
+
+-- SQLNESS REPLACE (region\s\d+\(\d+\,\s\d+\)) region
+INSERT INTO t_default (host, val) VALUES ('host3', 3);
+
+DROP TABLE t_default;
+
+DROP TABLE phy_default;
+
 CREATE TABLE phy (
     ts timestamp time index,
     val double
diff --git a/tests/cases/standalone/common/insert/merge_mode.result b/tests/cases/standalone/common/insert/merge_mode.result
index a98f6b6e38..a5e03ae8b9 100644
--- a/tests/cases/standalone/common/insert/merge_mode.result
+++ b/tests/cases/standalone/common/insert/merge_mode.result
@@ -181,5 +181,5 @@ create table if not exists invalid_merge_mode(
 engine=mito
 with('merge_mode'='last_non_null', 'append_mode'='true');
 
-Error: 1004(InvalidArguments), Invalid region options, merge_mode is not allowed when append_mode is enabled
+Error: 1004(InvalidArguments), Invalid region options, only last_row merge_mode is allowed when append_mode is enabled
 
diff --git a/tests/cases/standalone/common/mysql.result b/tests/cases/standalone/common/mysql.result
index 232cd3aed2..7414d2290d 100644
--- a/tests/cases/standalone/common/mysql.result
+++ b/tests/cases/standalone/common/mysql.result
@@ -13,7 +13,7 @@ SELECT @@version_comment;
 +-------------------+
 | @@version_comment |
 +-------------------+
-| Greptime          |
+| GreptimeDB        |
 +-------------------+
 
 -- SQLNESS PROTOCOL MYSQL
diff --git a/tests/cases/standalone/common/order/order_by.result b/tests/cases/standalone/common/order/order_by.result
index 13ac8caebe..6a2807e4b6 100644
--- a/tests/cases/standalone/common/order/order_by.result
+++ b/tests/cases/standalone/common/order/order_by.result
@@ -288,6 +288,7 @@ select tag from t where num > 6 order by ts;
 -- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
 -- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
 -- SQLNESS REPLACE num_ranges=\d+ num_ranges=REDACTED
+-- SQLNESS REPLACE ,\sfilter=\[[^]]+\] 
 explain analyze select tag from t where num > 6 order by ts desc limit 2;
 
 +-+-+-+
@@ -295,18 +296,16 @@ explain analyze select tag from t where num > 6 order by ts desc limit 2;
 +-+-+-+
 | 0_| 0_|_ProjectionExec: expr=[tag@0 as tag] REDACTED
 |_|_|_SortPreservingMergeExec: [ts@1 DESC], fetch=2 REDACTED
-|_|_|_SortExec: TopK(fetch=2), expr=[ts@1 DESC], preserve_partitioning=[true], filter=[ts@1 IS NULL OR ts@1 > 6000] REDACTED
+|_|_|_SortExec: TopK(fetch=2), expr=[ts@1 DESC], preserve_partitioning=[true] REDACTED
 |_|_|_MergeScanExec: REDACTED
 |_|_|_|
 | 1_| 0_|_SortPreservingMergeExec: [ts@1 DESC], fetch=2 REDACTED
-|_|_|_WindowedSortExec: expr=ts@1 DESC num_ranges=REDACTED fetch=2 REDACTED
-|_|_|_PartSortExec: expr=ts@1 DESC num_ranges=REDACTED limit=2 REDACTED
+|_|_|_SortExec: TopK(fetch=2), expr=[ts@1 DESC], preserve_partitioning=[true] REDACTED
 |_|_|_FilterExec: num@2 > 6, projection=[tag@0, ts@1] REDACTED
 |_|_|_SeqScan: region=REDACTED, "partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0} REDACTED
 |_|_|_|
 | 1_| 1_|_SortPreservingMergeExec: [ts@1 DESC], fetch=2 REDACTED
-|_|_|_WindowedSortExec: expr=ts@1 DESC num_ranges=REDACTED fetch=2 REDACTED
-|_|_|_PartSortExec: expr=ts@1 DESC num_ranges=REDACTED limit=2 REDACTED
+|_|_|_SortExec: TopK(fetch=2), expr=[ts@1 DESC], preserve_partitioning=[true] REDACTED
 |_|_|_FilterExec: num@2 > 6, projection=[tag@0, ts@1] REDACTED
 |_|_|_SeqScan: region=REDACTED, "partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0} REDACTED
 |_|_|_|
diff --git a/tests/cases/standalone/common/order/order_by.sql b/tests/cases/standalone/common/order/order_by.sql
index dd641613d9..95fd2c6f18 100644
--- a/tests/cases/standalone/common/order/order_by.sql
+++ b/tests/cases/standalone/common/order/order_by.sql
@@ -95,6 +95,7 @@ select tag from t where num > 6 order by ts;
 -- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
 -- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
 -- SQLNESS REPLACE num_ranges=\d+ num_ranges=REDACTED
+-- SQLNESS REPLACE ,\sfilter=\[[^]]+\] 
 explain analyze select tag from t where num > 6 order by ts desc limit 2;
 
 drop table t;
diff --git a/tests/cases/standalone/common/order/windowed_sort.result b/tests/cases/standalone/common/order/windowed_sort.result
index 4e550bf311..f85aa8c04e 100644
--- a/tests/cases/standalone/common/order/windowed_sort.result
+++ b/tests/cases/standalone/common/order/windowed_sort.result
@@ -70,7 +70,7 @@ EXPLAIN ANALYZE SELECT * FROM test ORDER BY t LIMIT 5;
 |_|_|_MergeScanExec: REDACTED
 |_|_|_|
 | 1_| 0_|_SortPreservingMergeExec: [t@1 ASC NULLS LAST], fetch=5 REDACTED
-|_|_|_WindowedSortExec: expr=t@1 ASC NULLS LAST num_ranges=REDACTED fetch=5 REDACTED
+|_|_|_SortExec: TopK(fetch=5), expr=[t@1 ASC NULLS LAST], preserve_partitioning=[true] REDACTED
 |_|_|_SeqScan: region=REDACTED, "partition_count":{"count":4, "mem_ranges":1, "files":3, "file_ranges":3} REDACTED
 |_|_|_|
 |_|_| Total rows: 5_|
@@ -103,8 +103,7 @@ EXPLAIN ANALYZE SELECT * FROM test ORDER BY t DESC LIMIT 5;
 |_|_|_MergeScanExec: REDACTED
 |_|_|_|
 | 1_| 0_|_SortPreservingMergeExec: [t@1 DESC], fetch=5 REDACTED
-|_|_|_WindowedSortExec: expr=t@1 DESC num_ranges=REDACTED fetch=5 REDACTED
-|_|_|_PartSortExec: expr=t@1 DESC num_ranges=REDACTED limit=5 REDACTED
+|_|_|_SortExec: TopK(fetch=5), expr=[t@1 DESC], preserve_partitioning=[true] REDACTED
 |_|_|_SeqScan: region=REDACTED, "partition_count":{"count":4, "mem_ranges":1, "files":3, "file_ranges":3} REDACTED
 |_|_|_|
 |_|_| Total rows: 5_|
@@ -137,7 +136,7 @@ EXPLAIN ANALYZE SELECT * FROM test where i > 2 ORDER BY t LIMIT 4;
 |_|_|_MergeScanExec: REDACTED
 |_|_|_|
 | 1_| 0_|_SortPreservingMergeExec: [t@1 ASC NULLS LAST], fetch=4 REDACTED
-|_|_|_WindowedSortExec: expr=t@1 ASC NULLS LAST num_ranges=REDACTED fetch=4 REDACTED
+|_|_|_SortExec: TopK(fetch=4), expr=[t@1 ASC NULLS LAST], preserve_partitioning=[true] REDACTED
 |_|_|_FilterExec: i@0 > 2 REDACTED
 |_|_|_SeqScan: region=REDACTED, "partition_count":{"count":4, "mem_ranges":1, "files":3, "file_ranges":3} REDACTED
 |_|_|_|
@@ -171,8 +170,7 @@ EXPLAIN ANALYZE SELECT * FROM test where i > 2 ORDER BY t DESC LIMIT 4;
 |_|_|_MergeScanExec: REDACTED
 |_|_|_|
 | 1_| 0_|_SortPreservingMergeExec: [t@1 DESC], fetch=4 REDACTED
-|_|_|_WindowedSortExec: expr=t@1 DESC num_ranges=REDACTED fetch=4 REDACTED
-|_|_|_PartSortExec: expr=t@1 DESC num_ranges=REDACTED limit=4 REDACTED
+|_|_|_SortExec: TopK(fetch=4), expr=[t@1 DESC], preserve_partitioning=[true] REDACTED
 |_|_|_FilterExec: i@0 > 2 REDACTED
 |_|_|_SeqScan: region=REDACTED, "partition_count":{"count":4, "mem_ranges":1, "files":3, "file_ranges":3} REDACTED
 |_|_|_|
@@ -206,8 +204,7 @@ EXPLAIN ANALYZE SELECT * FROM test where t > 8 ORDER BY t DESC LIMIT 4;
 |_|_|_MergeScanExec: REDACTED
 |_|_|_|
 | 1_| 0_|_SortPreservingMergeExec: [t@1 DESC], fetch=4 REDACTED
-|_|_|_WindowedSortExec: expr=t@1 DESC num_ranges=REDACTED fetch=4 REDACTED
-|_|_|_PartSortExec: expr=t@1 DESC num_ranges=REDACTED limit=4 REDACTED
+|_|_|_SortExec: TopK(fetch=4), expr=[t@1 DESC], preserve_partitioning=[true] REDACTED
 |_|_|_SeqScan: region=REDACTED, "partition_count":{"count":2, "mem_ranges":1, "files":1, "file_ranges":1} REDACTED
 |_|_|_|
 |_|_| Total rows: 4_|
@@ -289,8 +286,7 @@ EXPLAIN ANALYZE SELECT * FROM test_pk ORDER BY t LIMIT 5;
 |_|_|_MergeScanExec: REDACTED
 |_|_|_|
 | 1_| 0_|_SortPreservingMergeExec: [t@2 ASC NULLS LAST], fetch=5 REDACTED
-|_|_|_WindowedSortExec: expr=t@2 ASC NULLS LAST num_ranges=REDACTED fetch=5 REDACTED
-|_|_|_PartSortExec: expr=t@2 ASC NULLS LAST num_ranges=REDACTED limit=5 REDACTED
+|_|_|_SortExec: TopK(fetch=5), expr=[t@2 ASC NULLS LAST], preserve_partitioning=[true] REDACTED
 |_|_|_SeqScan: region=REDACTED, "partition_count":{"count":4, "mem_ranges":1, "files":3, "file_ranges":3} REDACTED
 |_|_|_|
 |_|_| Total rows: 5_|
@@ -312,8 +308,7 @@ EXPLAIN ANALYZE VERBOSE SELECT * FROM test_pk ORDER BY t LIMIT 5;
 |_|_|_MergeScanExec: REDACTED
 |_|_|_|
 | 1_| 0_|_SortPreservingMergeExec: [t@2 ASC NULLS LAST], fetch=5 REDACTED
-|_|_|_WindowedSortExec: expr=t@2 ASC NULLS LAST num_ranges=REDACTED fetch=5 REDACTED
-|_|_|_PartSortExec: expr=t@2 ASC NULLS LAST num_ranges=REDACTED limit=5 REDACTED
+|_|_|_SortExec: TopK(fetch=5), expr=[t@2 ASC NULLS LAST], preserve_partitioning=[true] REDACTED
 |_|_|_SeqScan: region=REDACTED, {"partition_count":{"count":4, "mem_ranges":1, "REDACTED
 |_|_|_|
 |_|_| Total rows: 5_|
@@ -346,8 +341,7 @@ EXPLAIN ANALYZE SELECT * FROM test_pk ORDER BY t DESC LIMIT 5;
 |_|_|_MergeScanExec: REDACTED
 |_|_|_|
 | 1_| 0_|_SortPreservingMergeExec: [t@2 DESC], fetch=5 REDACTED
-|_|_|_WindowedSortExec: expr=t@2 DESC num_ranges=REDACTED fetch=5 REDACTED
-|_|_|_PartSortExec: expr=t@2 DESC num_ranges=REDACTED limit=5 REDACTED
+|_|_|_SortExec: TopK(fetch=5), expr=[t@2 DESC], preserve_partitioning=[true] REDACTED
 |_|_|_SeqScan: region=REDACTED, "partition_count":{"count":4, "mem_ranges":1, "files":3, "file_ranges":3} REDACTED
 |_|_|_|
 |_|_| Total rows: 5_|
@@ -381,8 +375,7 @@ EXPLAIN ANALYZE SELECT * FROM test_pk where pk > 7 ORDER BY t LIMIT 5;
 |_|_|_MergeScanExec: REDACTED
 |_|_|_|
 | 1_| 0_|_SortPreservingMergeExec: [t@2 ASC NULLS LAST], fetch=5 REDACTED
-|_|_|_WindowedSortExec: expr=t@2 ASC NULLS LAST num_ranges=REDACTED fetch=5 REDACTED
-|_|_|_PartSortExec: expr=t@2 ASC NULLS LAST num_ranges=REDACTED limit=5 REDACTED
+|_|_|_SortExec: TopK(fetch=5), expr=[t@2 ASC NULLS LAST], preserve_partitioning=[true] REDACTED
 |_|_|_SeqScan: region=REDACTED, "partition_count":{"count":4, "mem_ranges":1, "files":3, "file_ranges":3} REDACTED
 |_|_|_|
 |_|_| Total rows: 5_|
@@ -404,8 +397,7 @@ EXPLAIN ANALYZE VERBOSE SELECT * FROM test_pk where pk > 7 ORDER BY t LIMIT 5;
 |_|_|_MergeScanExec: REDACTED
 |_|_|_|
 | 1_| 0_|_SortPreservingMergeExec: [t@2 ASC NULLS LAST], fetch=5 REDACTED
-|_|_|_WindowedSortExec: expr=t@2 ASC NULLS LAST num_ranges=REDACTED fetch=5 REDACTED
-|_|_|_PartSortExec: expr=t@2 ASC NULLS LAST num_ranges=REDACTED limit=5 REDACTED
+|_|_|_SortExec: TopK(fetch=5), expr=[t@2 ASC NULLS LAST], preserve_partitioning=[true] REDACTED
 |_|_|_SeqScan: region=REDACTED, {"partition_count":{"count":4, "mem_ranges":1, "REDACTED
 |_|_|_|
 |_|_| Total rows: 5_|
diff --git a/tests/cases/standalone/common/prepare/mysql_prepare.result b/tests/cases/standalone/common/prepare/mysql_prepare.result
index abc267b50e..5ef242a891 100644
--- a/tests/cases/standalone/common/prepare/mysql_prepare.result
+++ b/tests/cases/standalone/common/prepare/mysql_prepare.result
@@ -42,7 +42,7 @@ affected_rows: 0
 -- SQLNESS PROTOCOL MYSQL
 EXECUTE stmt USING 'a';
 
-Failed to execute query, err: MySqlError { ERROR 1815 (HY000): (EngineExecuteQuery): Cast error: Cannot cast string 'a' to value of Int32 type }
+Failed to execute query, err: MySqlError { ERROR 1210 (HY000): (InvalidArguments): Invalid request parameter: Unable to convert a to datatype Int32(Int32Type) }
 
 -- SQLNESS PROTOCOL MYSQL
 DEALLOCATE stmt;
@@ -124,6 +124,25 @@ DEALLOCATE stmt;
 
 affected_rows: 0
 
+-- SQLNESS PROTOCOL MYSQL
+PREPARE stmt FROM 'SELECT table_name, table_schema FROM information_schema.tables WHERE table_name = ?';
+
+affected_rows: 0
+
+-- SQLNESS PROTOCOL MYSQL
+EXECUTE stmt USING 'cake';
+
++------------+--------------+
+| table_name | table_schema |
++------------+--------------+
+| cake       | public       |
++------------+--------------+
+
+-- SQLNESS PROTOCOL MYSQL
+DEALLOCATE stmt;
+
+affected_rows: 0
+
 -- SQLNESS PROTOCOL MYSQL
 DROP TABLE cake;
 
diff --git a/tests/cases/standalone/common/prepare/mysql_prepare.sql b/tests/cases/standalone/common/prepare/mysql_prepare.sql
index 8e80a0a867..e96e945f88 100644
--- a/tests/cases/standalone/common/prepare/mysql_prepare.sql
+++ b/tests/cases/standalone/common/prepare/mysql_prepare.sql
@@ -72,5 +72,14 @@ EXECUTE stmt USING 'happy', 42, 0;
 -- SQLNESS PROTOCOL MYSQL
 DEALLOCATE stmt;
 
+-- SQLNESS PROTOCOL MYSQL
+PREPARE stmt FROM 'SELECT table_name, table_schema FROM information_schema.tables WHERE table_name = ?';
+
+-- SQLNESS PROTOCOL MYSQL
+EXECUTE stmt USING 'cake';
+
+-- SQLNESS PROTOCOL MYSQL
+DEALLOCATE stmt;
+
 -- SQLNESS PROTOCOL MYSQL
 DROP TABLE cake;
diff --git a/tests/cases/standalone/common/promql/anon_promql_ratio_repro.result b/tests/cases/standalone/common/promql/anon_promql_ratio_repro.result
new file mode 100644
index 0000000000..ab3c4db715
--- /dev/null
+++ b/tests/cases/standalone/common/promql/anon_promql_ratio_repro.result
@@ -0,0 +1,106 @@
+CREATE TABLE phy (
+    t TIMESTAMP TIME INDEX,
+    v DOUBLE
+) ENGINE=metric WITH ("physical_metric_table" = "");
+
+Affected Rows: 0
+
+CREATE TABLE metric_a (
+    l1 STRING NULL,
+    l2 STRING NULL,
+    l3 STRING NULL,
+    l4 STRING NULL,
+    l5 STRING NULL,
+    t TIMESTAMP NOT NULL,
+    v DOUBLE NULL,
+    TIME INDEX (t),
+    PRIMARY KEY (l1, l2, l3, l4, l5)
+) ENGINE=metric WITH (on_physical_table = 'phy');
+
+Affected Rows: 0
+
+CREATE TABLE metric_b (
+    l6 STRING NULL,
+    l1 STRING NULL,
+    l2 STRING NULL,
+    l3 STRING NULL,
+    l4 STRING NULL,
+    t TIMESTAMP NOT NULL,
+    v DOUBLE NULL,
+    TIME INDEX (t),
+    PRIMARY KEY (l6, l1, l2, l3, l4)
+) ENGINE=metric WITH (on_physical_table = 'phy');
+
+Affected Rows: 0
+
+INSERT INTO metric_a (l1, l2, l3, l4, l5, t, v) VALUES
+    ('v1', 'v2', 'v3', 'v4a', 'v5a', 1, 0),
+    ('v1', 'v2', 'v3', 'v4a', 'v5a', 180000, 120),
+    ('v1', 'v2', 'v3', 'v4a', 'v5a', 360000, 240),
+    ('v1', 'v2', 'v3', 'v4a', 'v5b', 1, 0),
+    ('v1', 'v2', 'v3', 'v4a', 'v5b', 180000, 30),
+    ('v1', 'v2', 'v3', 'v4a', 'v5b', 360000, 60),
+    ('v1', 'v2', 'v3-b', 'v4b', 'v5c', 1, 0),
+    ('v1', 'v2', 'v3-b', 'v4b', 'v5c', 180000, 60),
+    ('v1', 'v2', 'v3-b', 'v4b', 'v5c', 360000, 120);
+
+Affected Rows: 9
+
+INSERT INTO metric_b (l6, l1, l2, l3, l4, t, v) VALUES
+    ('v6', 'v1', 'v2', 'v3', 'v4a', 1, 1),
+    ('v6', 'v1', 'v2', 'v3', 'v4a', 180000, 1),
+    ('v6', 'v1', 'v2', 'v3', 'v4a', 360000, 1),
+    ('v6', 'v1', 'v2', 'v3-b', 'v4b', 1, 2),
+    ('v6', 'v1', 'v2', 'v3-b', 'v4b', 180000, 2),
+    ('v6', 'v1', 'v2', 'v3-b', 'v4b', 360000, 2);
+
+Affected Rows: 6
+
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (180, 360, '180s') count(((rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)"}[3m]) / on(l3,l4) group_left metric_b{l6="v6",l1="v1",l2="v2",l3=~"v3(|-a|-b)"}) > 0.50));
+
++---------------------+-------------------------------------------------------------------+
+| t                   | count(metric_a.prom_rate(t_range,v,t,Int64(180000)) / metric_b.v) |
++---------------------+-------------------------------------------------------------------+
+| 1970-01-01T00:03:00 | 1                                                                 |
++---------------------+-------------------------------------------------------------------+
+
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (180, 360, '180s') count(rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)"}[3m]));
+
++---------------------+---------------------------------------------+
+| t                   | count(prom_rate(t_range,v,t,Int64(180000))) |
++---------------------+---------------------------------------------+
+| 1970-01-01T00:03:00 | 3                                           |
++---------------------+---------------------------------------------+
+
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (180, 360, '180s') count(rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)"}[3m])) / 2;
+
++---------------------+----------------------------------------------------------+
+| t                   | count(prom_rate(t_range,v,t,Int64(180000))) / Float64(2) |
++---------------------+----------------------------------------------------------+
+| 1970-01-01T00:03:00 | 1.5                                                      |
++---------------------+----------------------------------------------------------+
+
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (180, 360, '180s') (count(((rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)"}[3m]) / on(l3,l4) group_left metric_b{l6="v6",l1="v1",l2="v2",l3=~"v3(|-a|-b)"}) > 0.50)) / count(rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)"}[3m]))) * 100;
+
++---------------------+--------------------------------------------------------------------------------------------------------------------------------------------------+
+| t                   | metric_b.count(metric_a.prom_rate(t_range,v,t,Int64(180000)) / metric_b.v) / metric_a.count(prom_rate(t_range,v,t,Int64(180000))) * Float64(100) |
++---------------------+--------------------------------------------------------------------------------------------------------------------------------------------------+
+| 1970-01-01T00:03:00 | 33.33333333333333                                                                                                                                |
++---------------------+--------------------------------------------------------------------------------------------------------------------------------------------------+
+
+DROP TABLE metric_a;
+
+Affected Rows: 0
+
+DROP TABLE metric_b;
+
+Affected Rows: 0
+
+DROP TABLE phy;
+
+Affected Rows: 0
+
diff --git a/tests/cases/standalone/common/promql/anon_promql_ratio_repro.sql b/tests/cases/standalone/common/promql/anon_promql_ratio_repro.sql
new file mode 100644
index 0000000000..946d4f93a1
--- /dev/null
+++ b/tests/cases/standalone/common/promql/anon_promql_ratio_repro.sql
@@ -0,0 +1,63 @@
+CREATE TABLE phy (
+    t TIMESTAMP TIME INDEX,
+    v DOUBLE
+) ENGINE=metric WITH ("physical_metric_table" = "");
+
+CREATE TABLE metric_a (
+    l1 STRING NULL,
+    l2 STRING NULL,
+    l3 STRING NULL,
+    l4 STRING NULL,
+    l5 STRING NULL,
+    t TIMESTAMP NOT NULL,
+    v DOUBLE NULL,
+    TIME INDEX (t),
+    PRIMARY KEY (l1, l2, l3, l4, l5)
+) ENGINE=metric WITH (on_physical_table = 'phy');
+
+CREATE TABLE metric_b (
+    l6 STRING NULL,
+    l1 STRING NULL,
+    l2 STRING NULL,
+    l3 STRING NULL,
+    l4 STRING NULL,
+    t TIMESTAMP NOT NULL,
+    v DOUBLE NULL,
+    TIME INDEX (t),
+    PRIMARY KEY (l6, l1, l2, l3, l4)
+) ENGINE=metric WITH (on_physical_table = 'phy');
+
+INSERT INTO metric_a (l1, l2, l3, l4, l5, t, v) VALUES
+    ('v1', 'v2', 'v3', 'v4a', 'v5a', 1, 0),
+    ('v1', 'v2', 'v3', 'v4a', 'v5a', 180000, 120),
+    ('v1', 'v2', 'v3', 'v4a', 'v5a', 360000, 240),
+    ('v1', 'v2', 'v3', 'v4a', 'v5b', 1, 0),
+    ('v1', 'v2', 'v3', 'v4a', 'v5b', 180000, 30),
+    ('v1', 'v2', 'v3', 'v4a', 'v5b', 360000, 60),
+    ('v1', 'v2', 'v3-b', 'v4b', 'v5c', 1, 0),
+    ('v1', 'v2', 'v3-b', 'v4b', 'v5c', 180000, 60),
+    ('v1', 'v2', 'v3-b', 'v4b', 'v5c', 360000, 120);
+
+INSERT INTO metric_b (l6, l1, l2, l3, l4, t, v) VALUES
+    ('v6', 'v1', 'v2', 'v3', 'v4a', 1, 1),
+    ('v6', 'v1', 'v2', 'v3', 'v4a', 180000, 1),
+    ('v6', 'v1', 'v2', 'v3', 'v4a', 360000, 1),
+    ('v6', 'v1', 'v2', 'v3-b', 'v4b', 1, 2),
+    ('v6', 'v1', 'v2', 'v3-b', 'v4b', 180000, 2),
+    ('v6', 'v1', 'v2', 'v3-b', 'v4b', 360000, 2);
+
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (180, 360, '180s') count(((rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)"}[3m]) / on(l3,l4) group_left metric_b{l6="v6",l1="v1",l2="v2",l3=~"v3(|-a|-b)"}) > 0.50));
+
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (180, 360, '180s') count(rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)"}[3m]));
+
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (180, 360, '180s') count(rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)"}[3m])) / 2;
+
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (180, 360, '180s') (count(((rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)"}[3m]) / on(l3,l4) group_left metric_b{l6="v6",l1="v1",l2="v2",l3=~"v3(|-a|-b)"}) > 0.50)) / count(rate(metric_a{l1="v1",l2="v2",l3=~"v3(|-a|-b)"}[3m]))) * 100;
+
+DROP TABLE metric_a;
+DROP TABLE metric_b;
+DROP TABLE phy;
diff --git a/tests/cases/standalone/common/promql/scalar.result b/tests/cases/standalone/common/promql/scalar.result
index c5c3e5ebd1..c3292b4f5c 100644
--- a/tests/cases/standalone/common/promql/scalar.result
+++ b/tests/cases/standalone/common/promql/scalar.result
@@ -136,6 +136,42 @@ TQL EVAL (0, 15, '5s') scalar(count(count(host) by (host)));
 | 1970-01-01T00:00:15 | 2.0                            |
 +---------------------+--------------------------------+
 
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 15, '5s') scalar(count(sum(host) by (host)));
+
++---------------------+------------------------------+
+| ts                  | scalar(count(sum(host.val))) |
++---------------------+------------------------------+
+| 1970-01-01T00:00:00 | 2.0                          |
+| 1970-01-01T00:00:05 | 2.0                          |
+| 1970-01-01T00:00:10 | 2.0                          |
+| 1970-01-01T00:00:15 | 2.0                          |
++---------------------+------------------------------+
+
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 15, '5s') scalar(count(avg(host) by (host)));
+
++---------------------+------------------------------+
+| ts                  | scalar(count(avg(host.val))) |
++---------------------+------------------------------+
+| 1970-01-01T00:00:00 | 2.0                          |
+| 1970-01-01T00:00:05 | 2.0                          |
+| 1970-01-01T00:00:10 | 2.0                          |
+| 1970-01-01T00:00:15 | 2.0                          |
++---------------------+------------------------------+
+
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 15, '5s') scalar(count(stddev(host) by (host)));
+
++---------------------+-------------------------------------+
+| ts                  | scalar(count(stddev_pop(host.val))) |
++---------------------+-------------------------------------+
+| 1970-01-01T00:00:00 | 2.0                                 |
+| 1970-01-01T00:00:05 | 2.0                                 |
+| 1970-01-01T00:00:10 | 2.0                                 |
+| 1970-01-01T00:00:15 | 2.0                                 |
++---------------------+-------------------------------------+
+
 -- SQLNESS SORT_RESULT 3 1
 TQL EVAL (0, 15, '5s') scalar(host{host="host1"} + scalar(host{host="host2"}));
 
@@ -516,7 +552,99 @@ TQL EVAL (0, 15, '5s') clamp_max(clamp(host{host="host1"}, 0, 15), 6);
 | 1970-01-01T00:00:15 | 6.0                                                     | host1 |
 +---------------------+---------------------------------------------------------+-------+
 
-Drop table host;
+DROP TABLE host;
+
+Affected Rows: 0
+
+CREATE TABLE presence_metric (
+  ts timestamp(3) time index,
+  instance STRING,
+  cpu STRING,
+  shard STRING,
+  val DOUBLE,
+  PRIMARY KEY (instance, cpu, shard),
+);
+
+Affected Rows: 0
+
+INSERT INTO TABLE presence_metric VALUES
+    (0,      'i1', 'cpu0', 'a', 1.0),
+    (0,      'i1', 'cpu0', 'b', 2.0),
+    (0,      'i1', 'cpu1', 'a', 10.0),
+    (0,      'i1', 'cpu2', 'a', 20.0),
+    (0,      'i2', 'cpu9', 'a', 100.0),
+    (200000, 'i1', 'cpu0', 'a', 'NAN'::DOUBLE),
+    (200000, 'i1', 'cpu0', 'b', 'NAN'::DOUBLE),
+    (200000, 'i1', 'cpu1', 'a', 11.0),
+    (200000, 'i1', 'cpu2', 'a', NULL),
+    (200000, 'i2', 'cpu9', 'a', 101.0),
+    (400000, 'i1', 'cpu1', 'a', 12.0),
+    (400000, 'i2', 'cpu9', 'a', 102.0),
+    (600000, 'i1', 'cpu0', 'a', 7.0),
+    (600000, 'i1', 'cpu0', 'b', 8.0),
+    (600000, 'i2', 'cpu9', 'a', 103.0);
+
+Affected Rows: 15
+
+-- NaN drops `cpu0` from the grouped count, while the NULL sample on `cpu2`
+-- still leaves a zero-valued row in `count(...) by (cpu)`.
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 600, '200s') count(presence_metric{instance="i1"}) by (cpu);
+
++------+---------------------+----------------------------+
+| cpu  | ts                  | count(presence_metric.val) |
++------+---------------------+----------------------------+
+| cpu0 | 1970-01-01T00:00:00 | 2                          |
+| cpu0 | 1970-01-01T00:10:00 | 2                          |
+| cpu1 | 1970-01-01T00:00:00 | 1                          |
+| cpu1 | 1970-01-01T00:03:20 | 1                          |
+| cpu1 | 1970-01-01T00:06:40 | 1                          |
+| cpu1 | 1970-01-01T00:10:00 | 1                          |
+| cpu2 | 1970-01-01T00:00:00 | 1                          |
+| cpu2 | 1970-01-01T00:03:20 | 0                          |
+| cpu2 | 1970-01-01T00:06:40 | 0                          |
++------+---------------------+----------------------------+
+
+-- Nested-count rewrite should preserve grouped presence after stale-NaN filtering and null-value pruning.
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 600, '200s') scalar(count(count(presence_metric{instance="i1"}) by (cpu)));
+
++---------------------+-------------------------------------------+
+| ts                  | scalar(count(count(presence_metric.val))) |
++---------------------+-------------------------------------------+
+| 1970-01-01T00:00:00 | 3.0                                       |
+| 1970-01-01T00:03:20 | 2.0                                       |
+| 1970-01-01T00:06:40 | 2.0                                       |
+| 1970-01-01T00:10:00 | 2.0                                       |
++---------------------+-------------------------------------------+
+
+-- Non-count inner aggregates must drop NULL-only groups before the outer count.
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 600, '200s') scalar(count(sum(presence_metric{instance="i1"}) by (cpu)));
+
++---------------------+-----------------------------------------+
+| ts                  | scalar(count(sum(presence_metric.val))) |
++---------------------+-----------------------------------------+
+| 1970-01-01T00:00:00 | 3.0                                     |
+| 1970-01-01T00:03:20 | 1.0                                     |
+| 1970-01-01T00:06:40 | 1.0                                     |
+| 1970-01-01T00:10:00 | 2.0                                     |
++---------------------+-----------------------------------------+
+
+-- False case: outer `by (instance)` keeps multiple series at the scalar input, so scalar should still yield NaN.
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 600, '200s') scalar(count(count(presence_metric) by (instance, cpu)) by (instance));
+
++---------------------+-------------------------------------------+
+| ts                  | scalar(count(count(presence_metric.val))) |
++---------------------+-------------------------------------------+
+| 1970-01-01T00:00:00 | NaN                                       |
+| 1970-01-01T00:03:20 | NaN                                       |
+| 1970-01-01T00:06:40 | NaN                                       |
+| 1970-01-01T00:10:00 | NaN                                       |
++---------------------+-------------------------------------------+
+
+DROP TABLE presence_metric;
 
 Affected Rows: 0
 
diff --git a/tests/cases/standalone/common/promql/scalar.sql b/tests/cases/standalone/common/promql/scalar.sql
index b4007bbf15..662f9665fe 100644
--- a/tests/cases/standalone/common/promql/scalar.sql
+++ b/tests/cases/standalone/common/promql/scalar.sql
@@ -43,6 +43,15 @@ TQL EVAL (0, 15, '5s') scalar(host{host="host1"}) + host;
 -- SQLNESS SORT_RESULT 3 1
 TQL EVAL (0, 15, '5s') scalar(count(count(host) by (host)));
 
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 15, '5s') scalar(count(sum(host) by (host)));
+
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 15, '5s') scalar(count(avg(host) by (host)));
+
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 15, '5s') scalar(count(stddev(host) by (host)));
+
 -- SQLNESS SORT_RESULT 3 1
 TQL EVAL (0, 15, '5s') scalar(host{host="host1"} + scalar(host{host="host2"}));
 
@@ -149,4 +158,49 @@ TQL EVAL (0, 15, '5s') clamp(clamp_min(host{host="host1"}, 1), 0, 12);
 -- SQLNESS SORT_RESULT 3 1
 TQL EVAL (0, 15, '5s') clamp_max(clamp(host{host="host1"}, 0, 15), 6);
 
-Drop table host;
+DROP TABLE host;
+
+CREATE TABLE presence_metric (
+  ts timestamp(3) time index,
+  instance STRING,
+  cpu STRING,
+  shard STRING,
+  val DOUBLE,
+  PRIMARY KEY (instance, cpu, shard),
+);
+
+INSERT INTO TABLE presence_metric VALUES
+    (0,      'i1', 'cpu0', 'a', 1.0),
+    (0,      'i1', 'cpu0', 'b', 2.0),
+    (0,      'i1', 'cpu1', 'a', 10.0),
+    (0,      'i1', 'cpu2', 'a', 20.0),
+    (0,      'i2', 'cpu9', 'a', 100.0),
+    (200000, 'i1', 'cpu0', 'a', 'NAN'::DOUBLE),
+    (200000, 'i1', 'cpu0', 'b', 'NAN'::DOUBLE),
+    (200000, 'i1', 'cpu1', 'a', 11.0),
+    (200000, 'i1', 'cpu2', 'a', NULL),
+    (200000, 'i2', 'cpu9', 'a', 101.0),
+    (400000, 'i1', 'cpu1', 'a', 12.0),
+    (400000, 'i2', 'cpu9', 'a', 102.0),
+    (600000, 'i1', 'cpu0', 'a', 7.0),
+    (600000, 'i1', 'cpu0', 'b', 8.0),
+    (600000, 'i2', 'cpu9', 'a', 103.0);
+
+-- NaN drops `cpu0` from the grouped count, while the NULL sample on `cpu2`
+-- still leaves a zero-valued row in `count(...) by (cpu)`.
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 600, '200s') count(presence_metric{instance="i1"}) by (cpu);
+
+-- Nested-count rewrite should preserve grouped presence after stale-NaN filtering and null-value pruning.
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 600, '200s') scalar(count(count(presence_metric{instance="i1"}) by (cpu)));
+
+-- Non-count inner aggregates must drop NULL-only groups before the outer count.
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 600, '200s') scalar(count(sum(presence_metric{instance="i1"}) by (cpu)));
+
+-- False case: outer `by (instance)` keeps multiple series at the scalar input, so scalar should still yield NaN.
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 600, '200s') scalar(count(count(presence_metric) by (instance, cpu)) by (instance));
+
+DROP TABLE presence_metric;
diff --git a/tests/cases/standalone/common/promql/stats_schema_mismatch_regression.result b/tests/cases/standalone/common/promql/stats_schema_mismatch_regression.result
index 0bd2acfa4f..f376dfe59c 100644
--- a/tests/cases/standalone/common/promql/stats_schema_mismatch_regression.result
+++ b/tests/cases/standalone/common/promql/stats_schema_mismatch_regression.result
@@ -125,29 +125,26 @@ Affected Rows: 0
 CREATE TABLE promql_instant_mismatch_nested (
   ts TIMESTAMP(3) TIME INDEX,
   k STRING PRIMARY KEY,
-  v JSON(format = "structured"),
+  v JSON2,
 );
 
 Affected Rows: 0
 
-INSERT INTO promql_instant_mismatch_nested VALUES
-  (0, 'a', '{"x": 1}'),
-  (1000, 'a', '{"x": 2}');
-
-Affected Rows: 2
-
+-- TODO(LFC): Uncomment the following SQLs and results when JSON2 is ready.
+-- INSERT INTO promql_instant_mismatch_nested VALUES
+--   (0, 'a', '{"x": 1}'),
+--   (1000, 'a', '{"x": 2}');
+-- Affected Rows: 0
 -- This used to error due to InstantManipulateExec returning `column_statistics` sized by
 -- `schema.flattened_fields().len()` when the schema contains nested fields (Arrow Struct).
 -- SQLNESS SORT_RESULT 3 1
-TQL EVAL (0, 1, '1s') promql_instant_mismatch_nested == promql_instant_mismatch_nested;
-
-+---------------------+--------+---+
-| ts                  | v      | k |
-+---------------------+--------+---+
-| 1970-01-01T00:00:00 | {x: 1} | a |
-| 1970-01-01T00:00:01 | {x: 2} | a |
-+---------------------+--------+---+
-
+-- TQL EVAL (0, 1, '1s') promql_instant_mismatch_nested == promql_instant_mismatch_nested;
+-- +---------------------+--------+---+
+-- | ts                  | v      | k |
+-- +---------------------+--------+---+
+-- | 1970-01-01T00:00:00 | {x: 1} | a |
+-- | 1970-01-01T00:00:01 | {x: 2} | a |
+-- +---------------------+--------+---+
 DROP TABLE promql_instant_mismatch_nested;
 
 Affected Rows: 0
diff --git a/tests/cases/standalone/common/promql/stats_schema_mismatch_regression.sql b/tests/cases/standalone/common/promql/stats_schema_mismatch_regression.sql
index 372d0387e0..db69796e74 100644
--- a/tests/cases/standalone/common/promql/stats_schema_mismatch_regression.sql
+++ b/tests/cases/standalone/common/promql/stats_schema_mismatch_regression.sql
@@ -107,16 +107,27 @@ DROP TABLE promql_stats_mismatch_physical;
 CREATE TABLE promql_instant_mismatch_nested (
   ts TIMESTAMP(3) TIME INDEX,
   k STRING PRIMARY KEY,
-  v JSON(format = "structured"),
+  v JSON2,
 );
 
-INSERT INTO promql_instant_mismatch_nested VALUES
-  (0, 'a', '{"x": 1}'),
-  (1000, 'a', '{"x": 2}');
+-- TODO(LFC): Uncomment the following SQLs and results when JSON2 is ready.
+
+-- INSERT INTO promql_instant_mismatch_nested VALUES
+--   (0, 'a', '{"x": 1}'),
+--   (1000, 'a', '{"x": 2}');
+
+-- Affected Rows: 0
 
 -- This used to error due to InstantManipulateExec returning `column_statistics` sized by
 -- `schema.flattened_fields().len()` when the schema contains nested fields (Arrow Struct).
 -- SQLNESS SORT_RESULT 3 1
-TQL EVAL (0, 1, '1s') promql_instant_mismatch_nested == promql_instant_mismatch_nested;
+-- TQL EVAL (0, 1, '1s') promql_instant_mismatch_nested == promql_instant_mismatch_nested;
+
+-- +---------------------+--------+---+
+-- | ts                  | v      | k |
+-- +---------------------+--------+---+
+-- | 1970-01-01T00:00:00 | {x: 1} | a |
+-- | 1970-01-01T00:00:01 | {x: 2} | a |
+-- +---------------------+--------+---+
 
 DROP TABLE promql_instant_mismatch_nested;
diff --git a/tests/cases/standalone/common/promql/tsid_binary_join_regression.result b/tests/cases/standalone/common/promql/tsid_binary_join_regression.result
new file mode 100644
index 0000000000..3640291dc3
--- /dev/null
+++ b/tests/cases/standalone/common/promql/tsid_binary_join_regression.result
@@ -0,0 +1,237 @@
+-- Regression test for TSID-backed PromQL binary joins on metric-engine tables.
+-- Default arithmetic and comparison joins should use `__tsid` when matching is the
+-- default one-to-one case. Label modifiers still have to stay label-based.
+CREATE TABLE tsid_binary_join_physical (
+  ts TIMESTAMP(3) TIME INDEX,
+  greptime_value DOUBLE,
+) ENGINE = metric WITH ("physical_metric_table" = "");
+
+Affected Rows: 0
+
+CREATE TABLE tsid_binary_join_left (
+  host STRING NULL,
+  job STRING NULL,
+  ts TIMESTAMP(3) NOT NULL,
+  greptime_value DOUBLE NULL,
+  TIME INDEX (ts),
+  PRIMARY KEY(host, job),
+)
+ENGINE = metric
+WITH(
+  on_physical_table = 'tsid_binary_join_physical'
+);
+
+Affected Rows: 0
+
+CREATE TABLE tsid_binary_join_right (
+  host STRING NULL,
+  job STRING NULL,
+  ts TIMESTAMP(3) NOT NULL,
+  greptime_value DOUBLE NULL,
+  TIME INDEX (ts),
+  PRIMARY KEY(host, job),
+)
+ENGINE = metric
+WITH(
+  on_physical_table = 'tsid_binary_join_physical'
+);
+
+Affected Rows: 0
+
+INSERT INTO tsid_binary_join_left (host, job, ts, greptime_value) VALUES
+  ('host1', 'job1', 0, 12),
+  ('host2', 'job2', 0, 18),
+  ('host1', 'job1', 5000, 15),
+  ('host2', 'job2', 5000, 21);
+
+Affected Rows: 4
+
+INSERT INTO tsid_binary_join_right (host, job, ts, greptime_value) VALUES
+  ('host1', 'job1', 0, 3),
+  ('host2', 'job2', 0, 6),
+  ('host1', 'job1', 5000, 5),
+  ('host2', 'job2', 5000, 7);
+
+Affected Rows: 4
+
+-- Default vector-vector arithmetic should join on `__tsid` and time index.
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE Hash\(\[__tsid@1,\sts@2\],.* Hash([__tsid@1, ts@2],REDACTED
+-- SQLNESS REPLACE Hash\(\[__tsid@3,\sts@4\],.* Hash([__tsid@3, ts@4],REDACTED
+-- SQLNESS REPLACE input_partitions=\d+ input_partitions=REDACTED
+-- SQLNESS REPLACE "partition_count":\{(.*?)\} "partition_count":REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+TQL ANALYZE (0, 5, '5s') tsid_binary_join_left / tsid_binary_join_right;
+
++-+-+-+
+| stage | node | plan_|
++-+-+-+
+| 0_| 0_|_ProjectionExec: expr=[host@2 as host, job@3 as job, ts@5 as ts, __tsid@4 as __tsid, greptime_value@0 / greptime_value@1 as tsid_binary_join_left.greptime_value / tsid_binary_join_right.greptime_value] REDACTED
+|_|_|_HashJoinExec: mode=Partitioned, join_type=Inner, on=[(__tsid@1, __tsid@3), (ts@2, ts@4)], projection=[greptime_value@0, greptime_value@3, host@4, job@5, __tsid@6, ts@7], NullsEqual: true REDACTED
+|_|_|_RepartitionExec: partitioning=Hash([__tsid@1, ts@2],REDACTED
+|_|_|_ProjectionExec: expr=[greptime_value@0 as greptime_value, __tsid@3 as __tsid, ts@4 as ts] REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_RepartitionExec: partitioning=Hash([__tsid@3, ts@4],REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_|
+| 1_| 0_|_PromInstantManipulateExec: range=[0..5000], lookback=[300000], interval=[5000], time index=[ts] REDACTED
+|_|_|_PromSeriesDivideExec: tags=["__tsid"] REDACTED
+|_|_|_ProjectionExec: expr=[greptime_value@1 as greptime_value, host@3 as host, job@4 as job, __tsid@2 as __tsid, ts@0 as ts] REDACTED
+|_|_|_CooperativeExec REDACTED
+|_|_|_SeriesScan: region=REDACTED, "partition_count":REDACTED, "distribution":"PerSeries" REDACTED
+|_|_|_|
+| 1_| 0_|_PromInstantManipulateExec: range=[0..5000], lookback=[300000], interval=[5000], time index=[ts] REDACTED
+|_|_|_PromSeriesDivideExec: tags=["__tsid"] REDACTED
+|_|_|_ProjectionExec: expr=[greptime_value@1 as greptime_value, host@3 as host, job@4 as job, __tsid@2 as __tsid, ts@0 as ts] REDACTED
+|_|_|_CooperativeExec REDACTED
+|_|_|_SeriesScan: region=REDACTED, "partition_count":REDACTED, "distribution":"PerSeries" REDACTED
+|_|_|_|
+|_|_| Total rows: 4_|
++-+-+-+
+
+-- Label modifiers must disable the TSID shortcut and keep matching on the remaining labels.
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE Hash\(\[job@1,\sts@2\],.* Hash([job@1, ts@2],REDACTED
+-- SQLNESS REPLACE Hash\(\[job@2,\sts@4\],.* Hash([job@2, ts@4],REDACTED
+-- SQLNESS REPLACE input_partitions=\d+ input_partitions=REDACTED
+-- SQLNESS REPLACE "partition_count":\{(.*?)\} "partition_count":REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+TQL ANALYZE (0, 5, '5s') tsid_binary_join_left / ignoring(host) tsid_binary_join_right;
+
++-+-+-+
+| stage | node | plan_|
++-+-+-+
+| 0_| 0_|_ProjectionExec: expr=[host@2 as host, job@3 as job, ts@5 as ts, __tsid@4 as __tsid, greptime_value@0 / greptime_value@1 as tsid_binary_join_left.greptime_value / tsid_binary_join_right.greptime_value] REDACTED
+|_|_|_HashJoinExec: mode=Partitioned, join_type=Inner, on=[(job@1, job@2), (ts@2, ts@4)], projection=[greptime_value@0, greptime_value@3, host@4, job@5, __tsid@6, ts@7], NullsEqual: true REDACTED
+|_|_|_RepartitionExec: partitioning=Hash([job@1, ts@2],REDACTED
+|_|_|_ProjectionExec: expr=[greptime_value@0 as greptime_value, job@2 as job, ts@4 as ts] REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_RepartitionExec: partitioning=Hash([job@2, ts@4],REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_|
+| 1_| 0_|_PromInstantManipulateExec: range=[0..5000], lookback=[300000], interval=[5000], time index=[ts] REDACTED
+|_|_|_PromSeriesDivideExec: tags=["__tsid"] REDACTED
+|_|_|_ProjectionExec: expr=[greptime_value@1 as greptime_value, host@3 as host, job@4 as job, __tsid@2 as __tsid, ts@0 as ts] REDACTED
+|_|_|_CooperativeExec REDACTED
+|_|_|_SeriesScan: region=REDACTED, "partition_count":REDACTED, "distribution":"PerSeries" REDACTED
+|_|_|_|
+| 1_| 0_|_PromInstantManipulateExec: range=[0..5000], lookback=[300000], interval=[5000], time index=[ts] REDACTED
+|_|_|_PromSeriesDivideExec: tags=["__tsid"] REDACTED
+|_|_|_ProjectionExec: expr=[greptime_value@1 as greptime_value, host@3 as host, job@4 as job, __tsid@2 as __tsid, ts@0 as ts] REDACTED
+|_|_|_CooperativeExec REDACTED
+|_|_|_SeriesScan: region=REDACTED, "partition_count":REDACTED, "distribution":"PerSeries" REDACTED
+|_|_|_|
+|_|_| Total rows: 4_|
++-+-+-+
+
+-- Comparison filters can join on `__tsid`, but the filtered result must still behave like
+-- a regular derived vector downstream.
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE Hash\(\[__tsid@1,\sts@2\],.* Hash([__tsid@1, ts@2],REDACTED
+-- SQLNESS REPLACE Hash\(\[__tsid@3,\sts@4\],.* Hash([__tsid@3, ts@4],REDACTED
+-- SQLNESS REPLACE input_partitions=\d+ input_partitions=REDACTED
+-- SQLNESS REPLACE "partition_count":\{(.*?)\} "partition_count":REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+TQL ANALYZE (0, 5, '5s') tsid_binary_join_left > tsid_binary_join_right;
+
++-+-+-+
+| stage | node | plan_|
++-+-+-+
+| 0_| 0_|_ProjectionExec: expr=[ts@4 as ts, greptime_value@0 as greptime_value, host@1 as host, job@2 as job, __tsid@3 as __tsid] REDACTED
+|_|_|_HashJoinExec: mode=Partitioned, join_type=Inner, on=[(__tsid@3, __tsid@1), (ts@4, ts@2)], filter=greptime_value@1 < greptime_value@0, projection=[greptime_value@0, host@1, job@2, __tsid@3, ts@4], NullsEqual: true REDACTED
+|_|_|_RepartitionExec: partitioning=Hash([__tsid@3, ts@4],REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_RepartitionExec: partitioning=Hash([__tsid@1, ts@2],REDACTED
+|_|_|_ProjectionExec: expr=[greptime_value@0 as greptime_value, __tsid@3 as __tsid, ts@4 as ts] REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_|
+| 1_| 0_|_PromInstantManipulateExec: range=[0..5000], lookback=[300000], interval=[5000], time index=[ts] REDACTED
+|_|_|_PromSeriesDivideExec: tags=["__tsid"] REDACTED
+|_|_|_ProjectionExec: expr=[greptime_value@1 as greptime_value, host@3 as host, job@4 as job, __tsid@2 as __tsid, ts@0 as ts] REDACTED
+|_|_|_CooperativeExec REDACTED
+|_|_|_SeriesScan: region=REDACTED, "partition_count":REDACTED, "distribution":"PerSeries" REDACTED
+|_|_|_|
+| 1_| 0_|_PromInstantManipulateExec: range=[0..5000], lookback=[300000], interval=[5000], time index=[ts] REDACTED
+|_|_|_PromSeriesDivideExec: tags=["__tsid"] REDACTED
+|_|_|_ProjectionExec: expr=[greptime_value@1 as greptime_value, host@3 as host, job@4 as job, __tsid@2 as __tsid, ts@0 as ts] REDACTED
+|_|_|_CooperativeExec REDACTED
+|_|_|_SeriesScan: region=REDACTED, "partition_count":REDACTED, "distribution":"PerSeries" REDACTED
+|_|_|_|
+|_|_| Total rows: 4_|
++-+-+-+
+
+-- `bool` comparison should follow the same TSID-backed matching path.
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE Hash\(\[__tsid@1,\sts@2\],.* Hash([__tsid@1, ts@2],REDACTED
+-- SQLNESS REPLACE Hash\(\[__tsid@3,\sts@4\],.* Hash([__tsid@3, ts@4],REDACTED
+-- SQLNESS REPLACE input_partitions=\d+ input_partitions=REDACTED
+-- SQLNESS REPLACE "partition_count":\{(.*?)\} "partition_count":REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+TQL ANALYZE (0, 5, '5s') tsid_binary_join_left > bool tsid_binary_join_right;
+
++-+-+-+
+| stage | node | plan_|
++-+-+-+
+| 0_| 0_|_ProjectionExec: expr=[host@2 as host, job@3 as job, ts@5 as ts, __tsid@4 as __tsid, CAST(greptime_value@1 < greptime_value@0 AS Float64) as tsid_binary_join_left.greptime_value > tsid_binary_join_right.greptime_value] REDACTED
+|_|_|_HashJoinExec: mode=Partitioned, join_type=Inner, on=[(__tsid@1, __tsid@3), (ts@2, ts@4)], projection=[greptime_value@0, greptime_value@3, host@4, job@5, __tsid@6, ts@7], NullsEqual: true REDACTED
+|_|_|_RepartitionExec: partitioning=Hash([__tsid@1, ts@2],REDACTED
+|_|_|_ProjectionExec: expr=[greptime_value@0 as greptime_value, __tsid@3 as __tsid, ts@4 as ts] REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_RepartitionExec: partitioning=Hash([__tsid@3, ts@4],REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_|
+| 1_| 0_|_PromInstantManipulateExec: range=[0..5000], lookback=[300000], interval=[5000], time index=[ts] REDACTED
+|_|_|_PromSeriesDivideExec: tags=["__tsid"] REDACTED
+|_|_|_ProjectionExec: expr=[greptime_value@1 as greptime_value, host@3 as host, job@4 as job, __tsid@2 as __tsid, ts@0 as ts] REDACTED
+|_|_|_CooperativeExec REDACTED
+|_|_|_SeriesScan: region=REDACTED, "partition_count":REDACTED, "distribution":"PerSeries" REDACTED
+|_|_|_|
+| 1_| 0_|_PromInstantManipulateExec: range=[0..5000], lookback=[300000], interval=[5000], time index=[ts] REDACTED
+|_|_|_PromSeriesDivideExec: tags=["__tsid"] REDACTED
+|_|_|_ProjectionExec: expr=[greptime_value@1 as greptime_value, host@3 as host, job@4 as job, __tsid@2 as __tsid, ts@0 as ts] REDACTED
+|_|_|_CooperativeExec REDACTED
+|_|_|_SeriesScan: region=REDACTED, "partition_count":REDACTED, "distribution":"PerSeries" REDACTED
+|_|_|_|
+|_|_| Total rows: 4_|
++-+-+-+
+
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 5, '5s') tsid_binary_join_left / tsid_binary_join_right;
+
++-------+------+---------------------+------------------------------------------------------------------------------+
+| host  | job  | ts                  | tsid_binary_join_left.greptime_value / tsid_binary_join_right.greptime_value |
++-------+------+---------------------+------------------------------------------------------------------------------+
+| host1 | job1 | 1970-01-01T00:00:00 | 4.0                                                                          |
+| host1 | job1 | 1970-01-01T00:00:05 | 3.0                                                                          |
+| host2 | job2 | 1970-01-01T00:00:00 | 3.0                                                                          |
+| host2 | job2 | 1970-01-01T00:00:05 | 3.0                                                                          |
++-------+------+---------------------+------------------------------------------------------------------------------+
+
+DROP TABLE tsid_binary_join_right;
+
+Affected Rows: 0
+
+DROP TABLE tsid_binary_join_left;
+
+Affected Rows: 0
+
+DROP TABLE tsid_binary_join_physical;
+
+Affected Rows: 0
+
diff --git a/tests/cases/standalone/common/promql/tsid_binary_join_regression.sql b/tests/cases/standalone/common/promql/tsid_binary_join_regression.sql
new file mode 100644
index 0000000000..61dfd23765
--- /dev/null
+++ b/tests/cases/standalone/common/promql/tsid_binary_join_regression.sql
@@ -0,0 +1,106 @@
+-- Regression test for TSID-backed PromQL binary joins on metric-engine tables.
+-- Default arithmetic and comparison joins should use `__tsid` when matching is the
+-- default one-to-one case. Label modifiers still have to stay label-based.
+
+CREATE TABLE tsid_binary_join_physical (
+  ts TIMESTAMP(3) TIME INDEX,
+  greptime_value DOUBLE,
+) ENGINE = metric WITH ("physical_metric_table" = "");
+
+CREATE TABLE tsid_binary_join_left (
+  host STRING NULL,
+  job STRING NULL,
+  ts TIMESTAMP(3) NOT NULL,
+  greptime_value DOUBLE NULL,
+  TIME INDEX (ts),
+  PRIMARY KEY(host, job),
+)
+ENGINE = metric
+WITH(
+  on_physical_table = 'tsid_binary_join_physical'
+);
+
+CREATE TABLE tsid_binary_join_right (
+  host STRING NULL,
+  job STRING NULL,
+  ts TIMESTAMP(3) NOT NULL,
+  greptime_value DOUBLE NULL,
+  TIME INDEX (ts),
+  PRIMARY KEY(host, job),
+)
+ENGINE = metric
+WITH(
+  on_physical_table = 'tsid_binary_join_physical'
+);
+
+INSERT INTO tsid_binary_join_left (host, job, ts, greptime_value) VALUES
+  ('host1', 'job1', 0, 12),
+  ('host2', 'job2', 0, 18),
+  ('host1', 'job1', 5000, 15),
+  ('host2', 'job2', 5000, 21);
+
+INSERT INTO tsid_binary_join_right (host, job, ts, greptime_value) VALUES
+  ('host1', 'job1', 0, 3),
+  ('host2', 'job2', 0, 6),
+  ('host1', 'job1', 5000, 5),
+  ('host2', 'job2', 5000, 7);
+
+-- Default vector-vector arithmetic should join on `__tsid` and time index.
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE Hash\(\[__tsid@1,\sts@2\],.* Hash([__tsid@1, ts@2],REDACTED
+-- SQLNESS REPLACE Hash\(\[__tsid@3,\sts@4\],.* Hash([__tsid@3, ts@4],REDACTED
+-- SQLNESS REPLACE input_partitions=\d+ input_partitions=REDACTED
+-- SQLNESS REPLACE "partition_count":\{(.*?)\} "partition_count":REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+TQL ANALYZE (0, 5, '5s') tsid_binary_join_left / tsid_binary_join_right;
+
+-- Label modifiers must disable the TSID shortcut and keep matching on the remaining labels.
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE Hash\(\[job@1,\sts@2\],.* Hash([job@1, ts@2],REDACTED
+-- SQLNESS REPLACE Hash\(\[job@2,\sts@4\],.* Hash([job@2, ts@4],REDACTED
+-- SQLNESS REPLACE input_partitions=\d+ input_partitions=REDACTED
+-- SQLNESS REPLACE "partition_count":\{(.*?)\} "partition_count":REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+TQL ANALYZE (0, 5, '5s') tsid_binary_join_left / ignoring(host) tsid_binary_join_right;
+
+-- Comparison filters can join on `__tsid`, but the filtered result must still behave like
+-- a regular derived vector downstream.
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE Hash\(\[__tsid@1,\sts@2\],.* Hash([__tsid@1, ts@2],REDACTED
+-- SQLNESS REPLACE Hash\(\[__tsid@3,\sts@4\],.* Hash([__tsid@3, ts@4],REDACTED
+-- SQLNESS REPLACE input_partitions=\d+ input_partitions=REDACTED
+-- SQLNESS REPLACE "partition_count":\{(.*?)\} "partition_count":REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+TQL ANALYZE (0, 5, '5s') tsid_binary_join_left > tsid_binary_join_right;
+
+-- `bool` comparison should follow the same TSID-backed matching path.
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE Hash\(\[__tsid@1,\sts@2\],.* Hash([__tsid@1, ts@2],REDACTED
+-- SQLNESS REPLACE Hash\(\[__tsid@3,\sts@4\],.* Hash([__tsid@3, ts@4],REDACTED
+-- SQLNESS REPLACE input_partitions=\d+ input_partitions=REDACTED
+-- SQLNESS REPLACE "partition_count":\{(.*?)\} "partition_count":REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+TQL ANALYZE (0, 5, '5s') tsid_binary_join_left > bool tsid_binary_join_right;
+
+-- SQLNESS SORT_RESULT 3 1
+TQL EVAL (0, 5, '5s') tsid_binary_join_left / tsid_binary_join_right;
+
+DROP TABLE tsid_binary_join_right;
+DROP TABLE tsid_binary_join_left;
+DROP TABLE tsid_binary_join_physical;
diff --git a/tests/cases/standalone/common/range/error.result b/tests/cases/standalone/common/range/error.result
index e3f12646e7..ff0795f450 100644
--- a/tests/cases/standalone/common/range/error.result
+++ b/tests/cases/standalone/common/range/error.result
@@ -60,6 +60,12 @@ SELECT min(val) RANGE '5s' FILL PREV FROM host;
 
 Error: 2000(InvalidSyntax), Invalid SQL syntax: sql parser error: ALIGN argument cannot be omitted in the range select query
 
+SELECT tmp.ts, tmp.host, min(tmp.val) RANGE '5s'
+FROM (SELECT ts, host, val FROM host) AS tmp
+ALIGN '5s';
+
+Error: 2000(InvalidSyntax), Range Query: Cannot infer default BY columns from derived range query input
+
 -- 2.3 type mismatch
 SELECT covar(ceil(val), floor(val)) RANGE '20s' FROM host ALIGN '10s';
 
diff --git a/tests/cases/standalone/common/range/error.sql b/tests/cases/standalone/common/range/error.sql
index 3659be1c79..d43b67ea4d 100644
--- a/tests/cases/standalone/common/range/error.sql
+++ b/tests/cases/standalone/common/range/error.sql
@@ -42,6 +42,10 @@ SELECT min(val) RANGE '5s' FROM host;
 
 SELECT min(val) RANGE '5s' FILL PREV FROM host;
 
+SELECT tmp.ts, tmp.host, min(tmp.val) RANGE '5s'
+FROM (SELECT ts, host, val FROM host) AS tmp
+ALIGN '5s';
+
 -- 2.3 type mismatch
 
 SELECT covar(ceil(val), floor(val)) RANGE '20s' FROM host ALIGN '10s';
diff --git a/tests/cases/standalone/common/range/nest.result b/tests/cases/standalone/common/range/nest.result
index 0730a0b997..7bc3e39a7b 100644
--- a/tests/cases/standalone/common/range/nest.result
+++ b/tests/cases/standalone/common/range/nest.result
@@ -45,6 +45,82 @@ SELECT ts, b, min(c) RANGE '5s' FROM (SELECT ts, host AS b, val AS c FROM host W
 | 1970-01-01T00:00:20 | host1 | 2               |
 +---------------------+-------+-----------------+
 
+CREATE TABLE host_union_0 (
+  ts timestamp(3) time index,
+  host STRING PRIMARY KEY,
+  val BIGINT,
+  ts2 timestamp(3),
+);
+
+Affected Rows: 0
+
+CREATE TABLE host_union_1 (
+  ts timestamp(3) time index,
+  host STRING PRIMARY KEY,
+  val BIGINT,
+  ts2 timestamp(3),
+);
+
+Affected Rows: 0
+
+INSERT INTO TABLE host_union_0 VALUES
+    (0,     'host1', 3, 0),
+    (5000,  'host1', 2, 5000),
+    (10000, 'host1', 1, 10000);
+
+Affected Rows: 3
+
+INSERT INTO TABLE host_union_1 VALUES
+    (0,     'host1', 6, 0),
+    (5000,  'host1', 5, 5000),
+    (10000, 'host1', 4, 10000);
+
+Affected Rows: 3
+
+SELECT ts, host, min(val ORDER BY ts ASC) RANGE '5s'
+FROM (
+  SELECT ts, host, val, ts2 FROM host_union_0
+  UNION ALL
+  SELECT ts, host, val, ts2 FROM host_union_1
+)
+WHERE ts >= '1970-01-01 00:00:00'
+ALIGN '5s' BY (host)
+ORDER BY host, ts;
+
++---------------------+-------+------------------------------------------------+
+| ts                  | host  | min(val) ORDER BY [ts ASC NULLS LAST] RANGE 5s |
++---------------------+-------+------------------------------------------------+
+| 1970-01-01T00:00:00 | host1 | 3                                              |
+| 1970-01-01T00:00:05 | host1 | 2                                              |
+| 1970-01-01T00:00:10 | host1 | 1                                              |
++---------------------+-------+------------------------------------------------+
+
+SELECT tmp.ts, tmp.host, min(tmp.val ORDER BY tmp.ts ASC) RANGE '5s'
+FROM (
+  SELECT ts, host, val, ts2 FROM host_union_0
+  UNION ALL
+  SELECT ts, host, val, ts2 FROM host_union_1
+) AS tmp
+WHERE tmp.ts >= '1970-01-01 00:00:00'
+ALIGN '5s' BY (tmp.host)
+ORDER BY tmp.host, tmp.ts;
+
++---------------------+-------+--------------------------------------------------------+
+| ts                  | host  | min(tmp.val) ORDER BY [tmp.ts ASC NULLS LAST] RANGE 5s |
++---------------------+-------+--------------------------------------------------------+
+| 1970-01-01T00:00:00 | host1 | 3                                                      |
+| 1970-01-01T00:00:05 | host1 | 2                                                      |
+| 1970-01-01T00:00:10 | host1 | 1                                                      |
++---------------------+-------+--------------------------------------------------------+
+
+DROP TABLE host_union_0;
+
+Affected Rows: 0
+
+DROP TABLE host_union_1;
+
+Affected Rows: 0
+
 -- Test EXPLAIN and ANALYZE
 -- SQLNESS REPLACE (-+) -
 -- SQLNESS REPLACE (\s\s+) _
diff --git a/tests/cases/standalone/common/range/nest.sql b/tests/cases/standalone/common/range/nest.sql
index 4ee447cfbd..d92f9f1345 100644
--- a/tests/cases/standalone/common/range/nest.sql
+++ b/tests/cases/standalone/common/range/nest.sql
@@ -22,6 +22,54 @@ SELECT ts, host, foo FROM (SELECT ts, host, min(val) RANGE '5s' AS foo FROM host
 
 SELECT ts, b, min(c) RANGE '5s' FROM (SELECT ts, host AS b, val AS c FROM host WHERE host = 'host1') ALIGN '5s' BY (b) ORDER BY b, ts;
 
+CREATE TABLE host_union_0 (
+  ts timestamp(3) time index,
+  host STRING PRIMARY KEY,
+  val BIGINT,
+  ts2 timestamp(3),
+);
+
+CREATE TABLE host_union_1 (
+  ts timestamp(3) time index,
+  host STRING PRIMARY KEY,
+  val BIGINT,
+  ts2 timestamp(3),
+);
+
+INSERT INTO TABLE host_union_0 VALUES
+    (0,     'host1', 3, 0),
+    (5000,  'host1', 2, 5000),
+    (10000, 'host1', 1, 10000);
+
+INSERT INTO TABLE host_union_1 VALUES
+    (0,     'host1', 6, 0),
+    (5000,  'host1', 5, 5000),
+    (10000, 'host1', 4, 10000);
+
+SELECT ts, host, min(val ORDER BY ts ASC) RANGE '5s'
+FROM (
+  SELECT ts, host, val, ts2 FROM host_union_0
+  UNION ALL
+  SELECT ts, host, val, ts2 FROM host_union_1
+)
+WHERE ts >= '1970-01-01 00:00:00'
+ALIGN '5s' BY (host)
+ORDER BY host, ts;
+
+SELECT tmp.ts, tmp.host, min(tmp.val ORDER BY tmp.ts ASC) RANGE '5s'
+FROM (
+  SELECT ts, host, val, ts2 FROM host_union_0
+  UNION ALL
+  SELECT ts, host, val, ts2 FROM host_union_1
+) AS tmp
+WHERE tmp.ts >= '1970-01-01 00:00:00'
+ALIGN '5s' BY (tmp.host)
+ORDER BY tmp.host, tmp.ts;
+
+DROP TABLE host_union_0;
+
+DROP TABLE host_union_1;
+
 
 -- Test EXPLAIN and ANALYZE
 
diff --git a/tests/cases/standalone/common/system/information_schema.result b/tests/cases/standalone/common/system/information_schema.result
index c4a5a137aa..c0eb8e5319 100644
--- a/tests/cases/standalone/common/system/information_schema.result
+++ b/tests/cases/standalone/common/system/information_schema.result
@@ -201,6 +201,7 @@ select * from information_schema.columns order by table_schema, table_name, colu
 | greptime      | information_schema | flows                                 | flow_definition                   | 5                | 2147483647               | 2147483647             |                   |               |                    | utf8               | utf8_bin       |            |       | select,insert |                       | String               | string              | FIELD         |                | No          | string              |                |        |
 | greptime      | information_schema | flows                                 | flow_id                           | 2                |                          |                        | 10                | 0             |                    |                    |                |            |       | select,insert |                       | UInt32               | int unsigned        | FIELD         |                | No          | int unsigned        |                |        |
 | greptime      | information_schema | flows                                 | flow_name                         | 1                | 2147483647               | 2147483647             |                   |               |                    | utf8               | utf8_bin       |            |       | select,insert |                       | String               | string              | FIELD         |                | No          | string              |                |        |
+| greptime      | information_schema | flows                                 | flownode_addrs                    | 16               | 2147483647               | 2147483647             |                   |               |                    | utf8               | utf8_bin       |            |       | select,insert |                       | String               | string              | FIELD         |                | Yes         | string              |                |        |
 | greptime      | information_schema | flows                                 | flownode_ids                      | 10               | 2147483647               | 2147483647             |                   |               |                    | utf8               | utf8_bin       |            |       | select,insert |                       | String               | string              | FIELD         |                | Yes         | string              |                |        |
 | greptime      | information_schema | flows                                 | last_execution_time               | 14               |                          |                        |                   |               | 3                  |                    |                |            |       | select,insert |                       | TimestampMillisecond | timestamp(3)        | FIELD         |                | Yes         | timestamp(3)        |                |        |
 | greptime      | information_schema | flows                                 | options                           | 11               | 2147483647               | 2147483647             |                   |               |                    | utf8               | utf8_bin       |            |       | select,insert |                       | String               | string              | FIELD         |                | Yes         | string              |                |        |
diff --git a/tests/cases/standalone/common/system/pg_catalog.result b/tests/cases/standalone/common/system/pg_catalog.result
index ef0452e316..7a6b65c7c8 100644
--- a/tests/cases/standalone/common/system/pg_catalog.result
+++ b/tests/cases/standalone/common/system/pg_catalog.result
@@ -14,15 +14,12 @@ SELECT session_user is not null;
 +----------------------------+
 
 -- SQLNESS REPLACE PostgreSQL.* VERSION
+-- SQLNESS REPLACE [\s\-]+
 -- current_schema
 -- SQLNESS PROTOCOL POSTGRES
 select current_schema(), current_schemas(true), current_schemas(false), version(), current_database();
 
-+------------------+---------------------------------------------------------+---------------------------------+---------------------------------------+--------------------+
-| current_schema() | current_schemas(Boolean(true))                          | current_schemas(Boolean(false)) | version                               | current_database() |
-+------------------+---------------------------------------------------------+---------------------------------+---------------------------------------+--------------------+
-| public           | {public,information_schema,pg_catalog,greptime_private} | {public}                        | VERSION
-+------------------+---------------------------------------------------------+---------------------------------+---------------------------------------+--------------------+
+++++++|current_schema()|current_schemas(Boolean(true))|current_schemas(Boolean(false))|version|current_database()|++++++|public|{public,information_schema,pg_catalog,greptime_private}|{public}|VERSION++++++
 
 -- search_path for pg using schema for now FIXME when support real search_path
 -- SQLNESS PROTOCOL POSTGRES
diff --git a/tests/cases/standalone/common/system/pg_catalog.sql b/tests/cases/standalone/common/system/pg_catalog.sql
index ad59da372c..2e84ecd7ce 100644
--- a/tests/cases/standalone/common/system/pg_catalog.sql
+++ b/tests/cases/standalone/common/system/pg_catalog.sql
@@ -6,6 +6,7 @@ create database pg_catalog;
 SELECT session_user is not null;
 
 -- SQLNESS REPLACE PostgreSQL.* VERSION
+-- SQLNESS REPLACE [\s\-]+
 -- current_schema
 -- SQLNESS PROTOCOL POSTGRES
 select current_schema(), current_schemas(true), current_schemas(false), version(), current_database();
diff --git a/tests/cases/standalone/common/tql-explain-analyze/explain.result b/tests/cases/standalone/common/tql-explain-analyze/explain.result
index 1e4cf18b40..65532d738b 100644
--- a/tests/cases/standalone/common/tql-explain-analyze/explain.result
+++ b/tests/cases/standalone/common/tql-explain-analyze/explain.result
@@ -74,6 +74,7 @@ TQL EXPLAIN ('1970-01-01T00:00:00'::timestamp, '1970-01-01T00:00:00'::timestamp
 -- SQLNESS REPLACE (elapsed_compute.*) REDACTED
 -- SQLNESS REPLACE (peers.*) REDACTED
 -- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (RepartitionExec:.*) RepartitionExec: REDACTED
 TQL EXPLAIN VERBOSE (0, 10, '5s') test;
 
 +-+-+
@@ -90,6 +91,7 @@ TQL EXPLAIN VERBOSE (0, 10, '5s') test;
 | logical_plan after TranscribeAtatRule_| SAME TEXT AS ABOVE_|
 | logical_plan after resolve_grouping_function_| SAME TEXT AS ABOVE_|
 | logical_plan after type_coercion_| SAME TEXT AS ABOVE_|
+| logical_plan after ConstNormalizationRule_| SAME TEXT AS ABOVE_|
 | logical_plan after DistPlannerAnalyzer_| Projection: test.i, test.j, test.k_|
 |_|_MergeScan [is_placeholder=false, remote_input=[_|
 |_| PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
@@ -131,6 +133,7 @@ TQL EXPLAIN VERBOSE (0, 10, '5s') test;
 |_|_TableScan: test_|
 |_| ]]_|
 | logical_plan after ScanHintRule_| SAME TEXT AS ABOVE_|
+| logical_plan after JsonTypeConcretizeRule_| SAME TEXT AS ABOVE_|
 | logical_plan after rewrite_set_comparison_| SAME TEXT AS ABOVE_|
 | logical_plan after optimize_unions_| SAME TEXT AS ABOVE_|
 | logical_plan after simplify_expressions_| SAME TEXT AS ABOVE_|
@@ -156,6 +159,7 @@ TQL EXPLAIN VERBOSE (0, 10, '5s') test;
 | logical_plan after push_down_leaf_projections_| SAME TEXT AS ABOVE_|
 | logical_plan after optimize_projections_| SAME TEXT AS ABOVE_|
 | logical_plan after ScanHintRule_| SAME TEXT AS ABOVE_|
+| logical_plan after JsonTypeConcretizeRule_| SAME TEXT AS ABOVE_|
 | logical_plan_| MergeScan [is_placeholder=false, remote_input=[_|
 |_| PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
 |_|_PromSeriesDivide: tags=["k"]_|
@@ -216,6 +220,7 @@ TQL EXPLAIN VERBOSE (0, 10, '5s') test;
 -- SQLNESS REPLACE (elapsed_compute.*) REDACTED
 -- SQLNESS REPLACE (peers.*) REDACTED
 -- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (RepartitionExec:.*) RepartitionExec: REDACTED
 TQL EXPLAIN VERBOSE (0, 10, '5s') test AS series;
 
 +-+-+
@@ -233,6 +238,7 @@ TQL EXPLAIN VERBOSE (0, 10, '5s') test AS series;
 | logical_plan after TranscribeAtatRule_| SAME TEXT AS ABOVE_|
 | logical_plan after resolve_grouping_function_| SAME TEXT AS ABOVE_|
 | logical_plan after type_coercion_| SAME TEXT AS ABOVE_|
+| logical_plan after ConstNormalizationRule_| SAME TEXT AS ABOVE_|
 | logical_plan after DistPlannerAnalyzer_| Projection: series, test.k, test.j_|
 |_|_MergeScan [is_placeholder=false, remote_input=[_|
 |_| Projection: test.i AS series, test.k, test.j_|
@@ -276,6 +282,7 @@ TQL EXPLAIN VERBOSE (0, 10, '5s') test AS series;
 |_|_TableScan: test_|
 |_| ]]_|
 | logical_plan after ScanHintRule_| SAME TEXT AS ABOVE_|
+| logical_plan after JsonTypeConcretizeRule_| SAME TEXT AS ABOVE_|
 | logical_plan after rewrite_set_comparison_| SAME TEXT AS ABOVE_|
 | logical_plan after optimize_unions_| SAME TEXT AS ABOVE_|
 | logical_plan after simplify_expressions_| SAME TEXT AS ABOVE_|
@@ -301,6 +308,7 @@ TQL EXPLAIN VERBOSE (0, 10, '5s') test AS series;
 | logical_plan after push_down_leaf_projections_| SAME TEXT AS ABOVE_|
 | logical_plan after optimize_projections_| SAME TEXT AS ABOVE_|
 | logical_plan after ScanHintRule_| SAME TEXT AS ABOVE_|
+| logical_plan after JsonTypeConcretizeRule_| SAME TEXT AS ABOVE_|
 | logical_plan_| MergeScan [is_placeholder=false, remote_input=[_|
 |_| Projection: test.i AS series, test.k, test.j_|
 |_|_PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
@@ -356,6 +364,394 @@ TQL EXPLAIN VERBOSE (0, 10, '5s') test AS series;
 |_|_|
 +-+-+
 
+CREATE TABLE test_nano(i DOUBLE, j TIMESTAMP(9) TIME INDEX, k STRING PRIMARY KEY);
+
+Affected Rows: 0
+
+INSERT INTO test_nano VALUES (1, 1000000, "a"), (1, 1000000, "b"), (2, 2000000, "a");
+
+Affected Rows: 3
+
+-- explain at 0s, 5s and 10s for a nanosecond time index.
+-- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (RepartitionExec:.*) RepartitionExec: REDACTED
+TQL EXPLAIN (0, 10, '5s') test_nano;
+
++---------------+-----------------------------------------------------------------------------------------------------------------------------------+
+| plan_type     | plan                                                                                                                              |
++---------------+-----------------------------------------------------------------------------------------------------------------------------------+
+| logical_plan  | PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]                                       |
+|               |   PromSeriesDivide: tags=["k"]                                                                                                    |
+|               |     Sort: test_nano.k ASC NULLS FIRST, test_nano.j ASC NULLS FIRST                                                                |
+|               |       Projection: test_nano.i, test_nano.k, CAST(test_nano.j AS Timestamp(ms)) AS j                                               |
+|               |         Projection: test_nano.i, test_nano.j, test_nano.k                                                                         |
+|               |           Filter: __common_expr_3 >= TimestampMillisecond(-299999, None) AND __common_expr_3 <= TimestampMillisecond(10000, None) |
+|               |             Projection: CAST(test_nano.j AS Timestamp(ms)) AS __common_expr_3, test_nano.i, test_nano.j, test_nano.k              |
+|               |               MergeScan [is_placeholder=false, remote_input=[                                                                     |
+|               | TableScan: test_nano                                                                                                              |
+|               | ]]                                                                                                                                |
+| physical_plan | PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]                                   |
+|               |   PromSeriesDivideExec: tags=["k"]                                                                                                |
+|               |     SortExec: expr=[k@1 ASC, j@2 ASC], preserve_partitioning=[true]                                                               |
+|               |       RepartitionExec: REDACTED
+|               |         ProjectionExec: expr=[i@0 as i, k@2 as k, CAST(j@1 AS Timestamp(ms)) as j]                                                |
+|               |           FilterExec: __common_expr_3@0 >= -299999 AND __common_expr_3@0 <= 10000, projection=[i@1, j@2, k@3]                     |
+|               |             ProjectionExec: expr=[CAST(j@1 AS Timestamp(ms)) as __common_expr_3, i@0 as i, j@1 as j, k@2 as k]                    |
+|               |               MergeScanExec: REDACTED
+|               |                                                                                                                                   |
++---------------+-----------------------------------------------------------------------------------------------------------------------------------+
+
+-- explain verbose at 0s, 5s and 10s for a nanosecond time index.
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (elapsed_compute.*) REDACTED
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (RepartitionExec:.*) RepartitionExec: REDACTED
+TQL EXPLAIN VERBOSE (0, 10, '5s') test_nano;
+
++-+-+
+| plan_type_| plan_|
++-+-+
+| initial_logical_plan_| PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivide: tags=["k"]_|
+|_|_Sort: test_nano.k ASC NULLS FIRST, test_nano.j ASC NULLS FIRST_|
+|_|_Filter: test_nano.j >= TimestampMillisecond(-299999, None) AND test_nano.j <= TimestampMillisecond(10000, None)_|
+|_|_Projection: test_nano.i, test_nano.k, CAST(test_nano.j AS Timestamp(ms)) AS j_|
+|_|_TableScan: test_nano_|
+| logical_plan after apply_function_rewrites_| SAME TEXT AS ABOVE_|
+| logical_plan after count_wildcard_to_time_index_rule_| SAME TEXT AS ABOVE_|
+| logical_plan after StringNormalizationRule_| SAME TEXT AS ABOVE_|
+| logical_plan after TranscribeAtatRule_| SAME TEXT AS ABOVE_|
+| logical_plan after resolve_grouping_function_| SAME TEXT AS ABOVE_|
+| logical_plan after type_coercion_| SAME TEXT AS ABOVE_|
+| logical_plan after ConstNormalizationRule_| SAME TEXT AS ABOVE_|
+| logical_plan after DistPlannerAnalyzer_| PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivide: tags=["k"]_|
+|_|_Sort: test_nano.k ASC NULLS FIRST, test_nano.j ASC NULLS FIRST_|
+|_|_Filter: test_nano.j >= TimestampMillisecond(-299999, None) AND test_nano.j <= TimestampMillisecond(10000, None)_|
+|_|_Projection: test_nano.i, test_nano.k, CAST(test_nano.j AS Timestamp(ms)) AS j_|
+|_|_Projection: test_nano.i, test_nano.j, test_nano.k_|
+|_|_MergeScan [is_placeholder=false, remote_input=[_|
+|_| TableScan: test_nano_|
+|_| ]]_|
+| logical_plan after FixStateUdafOrderingAnalyzer_| SAME TEXT AS ABOVE_|
+| analyzed_logical_plan_| SAME TEXT AS ABOVE_|
+| logical_plan after rewrite_set_comparison_| SAME TEXT AS ABOVE_|
+| logical_plan after optimize_unions_| SAME TEXT AS ABOVE_|
+| logical_plan after simplify_expressions_| SAME TEXT AS ABOVE_|
+| logical_plan after replace_distinct_aggregate_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_join_| SAME TEXT AS ABOVE_|
+| logical_plan after decorrelate_predicate_subquery_| SAME TEXT AS ABOVE_|
+| logical_plan after scalar_subquery_to_join_| SAME TEXT AS ABOVE_|
+| logical_plan after decorrelate_lateral_join_| SAME TEXT AS ABOVE_|
+| logical_plan after extract_equijoin_predicate_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_duplicated_expr_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_filter_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_cross_join_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_limit_| SAME TEXT AS ABOVE_|
+| logical_plan after propagate_empty_relation_| SAME TEXT AS ABOVE_|
+| logical_plan after filter_null_join_keys_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_outer_join_| SAME TEXT AS ABOVE_|
+| logical_plan after push_down_limit_| SAME TEXT AS ABOVE_|
+| logical_plan after push_down_filter_| PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivide: tags=["k"]_|
+|_|_Sort: test_nano.k ASC NULLS FIRST, test_nano.j ASC NULLS FIRST_|
+|_|_Projection: test_nano.i, test_nano.k, CAST(test_nano.j AS Timestamp(ms)) AS j_|
+|_|_Projection: test_nano.i, test_nano.j, test_nano.k_|
+|_|_Filter: CAST(test_nano.j AS Timestamp(ms)) >= TimestampMillisecond(-299999, None) AND CAST(test_nano.j AS Timestamp(ms)) <= TimestampMillisecond(10000, None)_|
+|_|_MergeScan [is_placeholder=false, remote_input=[_|
+|_| TableScan: test_nano_|
+|_| ]]_|
+| logical_plan after single_distinct_aggregation_to_group_by | SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_group_by_constant_| SAME TEXT AS ABOVE_|
+| logical_plan after common_sub_expression_eliminate_| PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivide: tags=["k"]_|
+|_|_Sort: test_nano.k ASC NULLS FIRST, test_nano.j ASC NULLS FIRST_|
+|_|_Projection: test_nano.i, test_nano.k, CAST(test_nano.j AS Timestamp(ms)) AS j_|
+|_|_Projection: test_nano.i, test_nano.j, test_nano.k_|
+|_|_Projection: test_nano.i, test_nano.j, test_nano.k_|
+|_|_Filter: __common_expr_1 >= TimestampMillisecond(-299999, None) AND __common_expr_1 <= TimestampMillisecond(10000, None)_|
+|_|_Projection: CAST(test_nano.j AS Timestamp(ms)) AS __common_expr_1, test_nano.i, test_nano.j, test_nano.k_|
+|_|_MergeScan [is_placeholder=false, remote_input=[_|
+|_| TableScan: test_nano_|
+|_| ]]_|
+| logical_plan after extract_leaf_expressions_| SAME TEXT AS ABOVE_|
+| logical_plan after push_down_leaf_projections_| SAME TEXT AS ABOVE_|
+| logical_plan after optimize_projections_| PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivide: tags=["k"]_|
+|_|_Sort: test_nano.k ASC NULLS FIRST, test_nano.j ASC NULLS FIRST_|
+|_|_Projection: test_nano.i, test_nano.k, CAST(test_nano.j AS Timestamp(ms)) AS j_|
+|_|_Projection: test_nano.i, test_nano.j, test_nano.k_|
+|_|_Filter: __common_expr_1 >= TimestampMillisecond(-299999, None) AND __common_expr_1 <= TimestampMillisecond(10000, None)_|
+|_|_Projection: CAST(test_nano.j AS Timestamp(ms)) AS __common_expr_1, test_nano.i, test_nano.j, test_nano.k_|
+|_|_MergeScan [is_placeholder=false, remote_input=[_|
+|_| TableScan: test_nano_|
+|_| ]]_|
+| logical_plan after ScanHintRule_| SAME TEXT AS ABOVE_|
+| logical_plan after JsonTypeConcretizeRule_| SAME TEXT AS ABOVE_|
+| logical_plan after rewrite_set_comparison_| SAME TEXT AS ABOVE_|
+| logical_plan after optimize_unions_| SAME TEXT AS ABOVE_|
+| logical_plan after simplify_expressions_| SAME TEXT AS ABOVE_|
+| logical_plan after replace_distinct_aggregate_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_join_| SAME TEXT AS ABOVE_|
+| logical_plan after decorrelate_predicate_subquery_| SAME TEXT AS ABOVE_|
+| logical_plan after scalar_subquery_to_join_| SAME TEXT AS ABOVE_|
+| logical_plan after decorrelate_lateral_join_| SAME TEXT AS ABOVE_|
+| logical_plan after extract_equijoin_predicate_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_duplicated_expr_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_filter_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_cross_join_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_limit_| SAME TEXT AS ABOVE_|
+| logical_plan after propagate_empty_relation_| SAME TEXT AS ABOVE_|
+| logical_plan after filter_null_join_keys_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_outer_join_| SAME TEXT AS ABOVE_|
+| logical_plan after push_down_limit_| SAME TEXT AS ABOVE_|
+| logical_plan after push_down_filter_| PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivide: tags=["k"]_|
+|_|_Sort: test_nano.k ASC NULLS FIRST, test_nano.j ASC NULLS FIRST_|
+|_|_Projection: test_nano.i, test_nano.k, CAST(test_nano.j AS Timestamp(ms)) AS j_|
+|_|_Projection: test_nano.i, test_nano.j, test_nano.k_|
+|_|_Projection: CAST(test_nano.j AS Timestamp(ms)) AS __common_expr_1, test_nano.i, test_nano.j, test_nano.k_|
+|_|_Filter: CAST(test_nano.j AS Timestamp(ms)) >= TimestampMillisecond(-299999, None) AND CAST(test_nano.j AS Timestamp(ms)) <= TimestampMillisecond(10000, None)_|
+|_|_MergeScan [is_placeholder=false, remote_input=[_|
+|_| TableScan: test_nano_|
+|_| ]]_|
+| logical_plan after single_distinct_aggregation_to_group_by | SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_group_by_constant_| SAME TEXT AS ABOVE_|
+| logical_plan after common_sub_expression_eliminate_| PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivide: tags=["k"]_|
+|_|_Sort: test_nano.k ASC NULLS FIRST, test_nano.j ASC NULLS FIRST_|
+|_|_Projection: test_nano.i, test_nano.k, CAST(test_nano.j AS Timestamp(ms)) AS j_|
+|_|_Projection: test_nano.i, test_nano.j, test_nano.k_|
+|_|_Projection: CAST(test_nano.j AS Timestamp(ms)) AS __common_expr_1, test_nano.i, test_nano.j, test_nano.k_|
+|_|_Projection: test_nano.i, test_nano.j, test_nano.k_|
+|_|_Filter: __common_expr_2 >= TimestampMillisecond(-299999, None) AND __common_expr_2 <= TimestampMillisecond(10000, None)_|
+|_|_Projection: CAST(test_nano.j AS Timestamp(ms)) AS __common_expr_2, test_nano.i, test_nano.j, test_nano.k_|
+|_|_MergeScan [is_placeholder=false, remote_input=[_|
+|_| TableScan: test_nano_|
+|_| ]]_|
+| logical_plan after extract_leaf_expressions_| SAME TEXT AS ABOVE_|
+| logical_plan after push_down_leaf_projections_| SAME TEXT AS ABOVE_|
+| logical_plan after optimize_projections_| PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivide: tags=["k"]_|
+|_|_Sort: test_nano.k ASC NULLS FIRST, test_nano.j ASC NULLS FIRST_|
+|_|_Projection: test_nano.i, test_nano.k, CAST(test_nano.j AS Timestamp(ms)) AS j_|
+|_|_Projection: test_nano.i, test_nano.j, test_nano.k_|
+|_|_Filter: __common_expr_2 >= TimestampMillisecond(-299999, None) AND __common_expr_2 <= TimestampMillisecond(10000, None)_|
+|_|_Projection: CAST(test_nano.j AS Timestamp(ms)) AS __common_expr_2, test_nano.i, test_nano.j, test_nano.k_|
+|_|_MergeScan [is_placeholder=false, remote_input=[_|
+|_| TableScan: test_nano_|
+|_| ]]_|
+| logical_plan after ScanHintRule_| SAME TEXT AS ABOVE_|
+| logical_plan after JsonTypeConcretizeRule_| SAME TEXT AS ABOVE_|
+| logical_plan after rewrite_set_comparison_| SAME TEXT AS ABOVE_|
+| logical_plan after optimize_unions_| SAME TEXT AS ABOVE_|
+| logical_plan after simplify_expressions_| SAME TEXT AS ABOVE_|
+| logical_plan after replace_distinct_aggregate_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_join_| SAME TEXT AS ABOVE_|
+| logical_plan after decorrelate_predicate_subquery_| SAME TEXT AS ABOVE_|
+| logical_plan after scalar_subquery_to_join_| SAME TEXT AS ABOVE_|
+| logical_plan after decorrelate_lateral_join_| SAME TEXT AS ABOVE_|
+| logical_plan after extract_equijoin_predicate_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_duplicated_expr_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_filter_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_cross_join_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_limit_| SAME TEXT AS ABOVE_|
+| logical_plan after propagate_empty_relation_| SAME TEXT AS ABOVE_|
+| logical_plan after filter_null_join_keys_| SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_outer_join_| SAME TEXT AS ABOVE_|
+| logical_plan after push_down_limit_| SAME TEXT AS ABOVE_|
+| logical_plan after push_down_filter_| PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivide: tags=["k"]_|
+|_|_Sort: test_nano.k ASC NULLS FIRST, test_nano.j ASC NULLS FIRST_|
+|_|_Projection: test_nano.i, test_nano.k, CAST(test_nano.j AS Timestamp(ms)) AS j_|
+|_|_Projection: test_nano.i, test_nano.j, test_nano.k_|
+|_|_Projection: CAST(test_nano.j AS Timestamp(ms)) AS __common_expr_2, test_nano.i, test_nano.j, test_nano.k_|
+|_|_Filter: CAST(test_nano.j AS Timestamp(ms)) >= TimestampMillisecond(-299999, None) AND CAST(test_nano.j AS Timestamp(ms)) <= TimestampMillisecond(10000, None)_|
+|_|_MergeScan [is_placeholder=false, remote_input=[_|
+|_| TableScan: test_nano_|
+|_| ]]_|
+| logical_plan after single_distinct_aggregation_to_group_by | SAME TEXT AS ABOVE_|
+| logical_plan after eliminate_group_by_constant_| SAME TEXT AS ABOVE_|
+| logical_plan after common_sub_expression_eliminate_| PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivide: tags=["k"]_|
+|_|_Sort: test_nano.k ASC NULLS FIRST, test_nano.j ASC NULLS FIRST_|
+|_|_Projection: test_nano.i, test_nano.k, CAST(test_nano.j AS Timestamp(ms)) AS j_|
+|_|_Projection: test_nano.i, test_nano.j, test_nano.k_|
+|_|_Projection: CAST(test_nano.j AS Timestamp(ms)) AS __common_expr_2, test_nano.i, test_nano.j, test_nano.k_|
+|_|_Projection: test_nano.i, test_nano.j, test_nano.k_|
+|_|_Filter: __common_expr_3 >= TimestampMillisecond(-299999, None) AND __common_expr_3 <= TimestampMillisecond(10000, None)_|
+|_|_Projection: CAST(test_nano.j AS Timestamp(ms)) AS __common_expr_3, test_nano.i, test_nano.j, test_nano.k_|
+|_|_MergeScan [is_placeholder=false, remote_input=[_|
+|_| TableScan: test_nano_|
+|_| ]]_|
+| logical_plan after extract_leaf_expressions_| SAME TEXT AS ABOVE_|
+| logical_plan after push_down_leaf_projections_| SAME TEXT AS ABOVE_|
+| logical_plan after optimize_projections_| PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivide: tags=["k"]_|
+|_|_Sort: test_nano.k ASC NULLS FIRST, test_nano.j ASC NULLS FIRST_|
+|_|_Projection: test_nano.i, test_nano.k, CAST(test_nano.j AS Timestamp(ms)) AS j_|
+|_|_Projection: test_nano.i, test_nano.j, test_nano.k_|
+|_|_Filter: __common_expr_3 >= TimestampMillisecond(-299999, None) AND __common_expr_3 <= TimestampMillisecond(10000, None)_|
+|_|_Projection: CAST(test_nano.j AS Timestamp(ms)) AS __common_expr_3, test_nano.i, test_nano.j, test_nano.k_|
+|_|_MergeScan [is_placeholder=false, remote_input=[_|
+|_| TableScan: test_nano_|
+|_| ]]_|
+| logical_plan after ScanHintRule_| SAME TEXT AS ABOVE_|
+| logical_plan after JsonTypeConcretizeRule_| SAME TEXT AS ABOVE_|
+| logical_plan_| PromInstantManipulate: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivide: tags=["k"]_|
+|_|_Sort: test_nano.k ASC NULLS FIRST, test_nano.j ASC NULLS FIRST_|
+|_|_Projection: test_nano.i, test_nano.k, CAST(test_nano.j AS Timestamp(ms)) AS j_|
+|_|_Projection: test_nano.i, test_nano.j, test_nano.k_|
+|_|_Filter: __common_expr_3 >= TimestampMillisecond(-299999, None) AND __common_expr_3 <= TimestampMillisecond(10000, None)_|
+|_|_Projection: CAST(test_nano.j AS Timestamp(ms)) AS __common_expr_3, test_nano.i, test_nano.j, test_nano.k_|
+|_|_MergeScan [is_placeholder=false, remote_input=[_|
+|_| TableScan: test_nano_|
+|_| ]]_|
+| initial_physical_plan_| PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivideExec: tags=["k"]_|
+|_|_SortExec: expr=[k@1 ASC, j@2 ASC], preserve_partitioning=[false]_|
+|_|_ProjectionExec: expr=[i@0 as i, k@2 as k, CAST(j@1 AS Timestamp(ms)) as j]_|
+|_|_ProjectionExec: expr=[i@1 as i, j@2 as j, k@3 as k]_|
+|_|_FilterExec: __common_expr_3@0 >= -299999 AND __common_expr_3@0 <= 10000_|
+|_|_ProjectionExec: expr=[CAST(j@1 AS Timestamp(ms)) as __common_expr_3, i@0 as i, j@1 as j, k@2 as k]_|
+|_|_MergeScanExec: REDACTED
+|_|_|
+| initial_physical_plan_with_stats_| PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[j], statistics=[Rows=Inexact(2), Bytes=Absent, [(Col[0]:),(Col[1]:),(Col[2]:)]]_|
+|_|_PromSeriesDivideExec: tags=["k"], statistics=[Rows=Absent, Bytes=Absent, [(Col[0]:),(Col[1]:),(Col[2]:)]]_|
+|_|_SortExec: expr=[k@1 ASC, j@2 ASC], preserve_partitioning=[false], statistics=[Rows=Absent, Bytes=Absent, [(Col[0]:),(Col[1]:),(Col[2]:)]]_|
+|_|_ProjectionExec: expr=[i@0 as i, k@2 as k, CAST(j@1 AS Timestamp(ms)) as j], statistics=[Rows=Absent, Bytes=Absent, [(Col[0]:),(Col[1]:),(Col[2]:)]]_|
+|_|_ProjectionExec: expr=[i@1 as i, j@2 as j, k@3 as k], statistics=[Rows=Absent, Bytes=Absent, [(Col[0]:),(Col[1]:),(Col[2]:)]]_|
+|_|_FilterExec: __common_expr_3@0 >= -299999 AND __common_expr_3@0 <= 10000, statistics=[Rows=Absent, Bytes=Absent, [(Col[0]:),(Col[1]:),(Col[2]:),(Col[3]:)]]_|
+|_|_ProjectionExec: expr=[CAST(j@1 AS Timestamp(ms)) as __common_expr_3, i@0 as i, j@1 as j, k@2 as k], statistics=[Rows=Absent, Bytes=Absent, [(Col[0]:),(Col[1]:),(Col[2]:),(Col[3]:)]] |
+|_|_MergeScanExec: REDACTED
+|_|_|
+| initial_physical_plan_with_schema_| PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[j], schema=[i:Float64;N, k:Utf8;N, j:Timestamp(ms)]_|
+|_|_PromSeriesDivideExec: tags=["k"], schema=[i:Float64;N, k:Utf8;N, j:Timestamp(ms)]_|
+|_|_SortExec: expr=[k@1 ASC, j@2 ASC], preserve_partitioning=[false], schema=[i:Float64;N, k:Utf8;N, j:Timestamp(ms)]_|
+|_|_ProjectionExec: expr=[i@0 as i, k@2 as k, CAST(j@1 AS Timestamp(ms)) as j], schema=[i:Float64;N, k:Utf8;N, j:Timestamp(ms)]_|
+|_|_ProjectionExec: expr=[i@1 as i, j@2 as j, k@3 as k], schema=[i:Float64;N, j:Timestamp(ns), k:Utf8;N]_|
+|_|_FilterExec: __common_expr_3@0 >= -299999 AND __common_expr_3@0 <= 10000, schema=[__common_expr_3:Timestamp(ms), i:Float64;N, j:Timestamp(ns), k:Utf8;N]_|
+|_|_ProjectionExec: expr=[CAST(j@1 AS Timestamp(ms)) as __common_expr_3, i@0 as i, j@1 as j, k@2 as k], schema=[__common_expr_3:Timestamp(ms), i:Float64;N, j:Timestamp(ns), k:Utf8;N]_|
+|_|_MergeScanExec: REDACTED
+|_|_|
+| physical_plan after OutputRequirements_| OutputRequirementExec: order_by=[], dist_by=Unspecified_|
+|_|_PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivideExec: tags=["k"]_|
+|_|_SortExec: expr=[k@1 ASC, j@2 ASC], preserve_partitioning=[false]_|
+|_|_ProjectionExec: expr=[i@0 as i, k@2 as k, CAST(j@1 AS Timestamp(ms)) as j]_|
+|_|_ProjectionExec: expr=[i@1 as i, j@2 as j, k@3 as k]_|
+|_|_FilterExec: __common_expr_3@0 >= -299999 AND __common_expr_3@0 <= 10000_|
+|_|_ProjectionExec: expr=[CAST(j@1 AS Timestamp(ms)) as __common_expr_3, i@0 as i, j@1 as j, k@2 as k]_|
+|_|_MergeScanExec: REDACTED
+|_|_|
+| physical_plan after aggregate_statistics_| SAME TEXT AS ABOVE_|
+| physical_plan after join_selection_| SAME TEXT AS ABOVE_|
+| physical_plan after LimitedDistinctAggregation_| SAME TEXT AS ABOVE_|
+| physical_plan after FilterPushdown_| SAME TEXT AS ABOVE_|
+| physical_plan after parallelize_scan_| SAME TEXT AS ABOVE_|
+| physical_plan after PassDistributionRule_| SAME TEXT AS ABOVE_|
+| physical_plan after EnforceSorting_| OutputRequirementExec: order_by=[], dist_by=Unspecified_|
+|_|_PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivideExec: tags=["k"]_|
+|_|_SortExec: expr=[k@1 ASC, j@2 ASC], preserve_partitioning=[true]_|
+|_|_ProjectionExec: expr=[i@0 as i, k@2 as k, CAST(j@1 AS Timestamp(ms)) as j]_|
+|_|_ProjectionExec: expr=[i@1 as i, j@2 as j, k@3 as k]_|
+|_|_FilterExec: __common_expr_3@0 >= -299999 AND __common_expr_3@0 <= 10000_|
+|_|_ProjectionExec: expr=[CAST(j@1 AS Timestamp(ms)) as __common_expr_3, i@0 as i, j@1 as j, k@2 as k]_|
+|_|_MergeScanExec: REDACTED
+|_|_|
+| physical_plan after EnforceDistribution_| OutputRequirementExec: order_by=[], dist_by=Unspecified_|
+|_|_PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivideExec: tags=["k"]_|
+|_|_SortExec: expr=[k@1 ASC, j@2 ASC], preserve_partitioning=[true]_|
+|_|_RepartitionExec: REDACTED
+|_|_SortExec: expr=[k@1 ASC, j@2 ASC], preserve_partitioning=[true]_|
+|_|_ProjectionExec: expr=[i@0 as i, k@2 as k, CAST(j@1 AS Timestamp(ms)) as j]_|
+|_|_ProjectionExec: expr=[i@1 as i, j@2 as j, k@3 as k]_|
+|_|_FilterExec: __common_expr_3@0 >= -299999 AND __common_expr_3@0 <= 10000_|
+|_|_ProjectionExec: expr=[CAST(j@1 AS Timestamp(ms)) as __common_expr_3, i@0 as i, j@1 as j, k@2 as k]_|
+|_|_MergeScanExec: REDACTED
+|_|_|
+| physical_plan after CombinePartialFinalAggregate_| SAME TEXT AS ABOVE_|
+| physical_plan after EnforceSorting_| OutputRequirementExec: order_by=[], dist_by=Unspecified_|
+|_|_PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivideExec: tags=["k"]_|
+|_|_SortExec: expr=[k@1 ASC, j@2 ASC], preserve_partitioning=[true]_|
+|_|_RepartitionExec: REDACTED
+|_|_ProjectionExec: expr=[i@0 as i, k@2 as k, CAST(j@1 AS Timestamp(ms)) as j]_|
+|_|_ProjectionExec: expr=[i@1 as i, j@2 as j, k@3 as k]_|
+|_|_FilterExec: __common_expr_3@0 >= -299999 AND __common_expr_3@0 <= 10000_|
+|_|_ProjectionExec: expr=[CAST(j@1 AS Timestamp(ms)) as __common_expr_3, i@0 as i, j@1 as j, k@2 as k]_|
+|_|_MergeScanExec: REDACTED
+|_|_|
+| physical_plan after OptimizeAggregateOrder_| SAME TEXT AS ABOVE_|
+| physical_plan after ProjectionPushdown_| OutputRequirementExec: order_by=[], dist_by=Unspecified_|
+|_|_PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivideExec: tags=["k"]_|
+|_|_SortExec: expr=[k@1 ASC, j@2 ASC], preserve_partitioning=[true]_|
+|_|_RepartitionExec: REDACTED
+|_|_ProjectionExec: expr=[i@0 as i, k@2 as k, CAST(j@1 AS Timestamp(ms)) as j]_|
+|_|_FilterExec: __common_expr_3@0 >= -299999 AND __common_expr_3@0 <= 10000, projection=[i@1, j@2, k@3]_|
+|_|_ProjectionExec: expr=[CAST(j@1 AS Timestamp(ms)) as __common_expr_3, i@0 as i, j@1 as j, k@2 as k]_|
+|_|_MergeScanExec: REDACTED
+|_|_|
+| physical_plan after OutputRequirements_| PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivideExec: tags=["k"]_|
+|_|_SortExec: expr=[k@1 ASC, j@2 ASC], preserve_partitioning=[true]_|
+|_|_RepartitionExec: REDACTED
+|_|_ProjectionExec: expr=[i@0 as i, k@2 as k, CAST(j@1 AS Timestamp(ms)) as j]_|
+|_|_FilterExec: __common_expr_3@0 >= -299999 AND __common_expr_3@0 <= 10000, projection=[i@1, j@2, k@3]_|
+|_|_ProjectionExec: expr=[CAST(j@1 AS Timestamp(ms)) as __common_expr_3, i@0 as i, j@1 as j, k@2 as k]_|
+|_|_MergeScanExec: REDACTED
+|_|_|
+| physical_plan after LimitAggregation_| SAME TEXT AS ABOVE_|
+| physical_plan after LimitPushPastWindows_| SAME TEXT AS ABOVE_|
+| physical_plan after LimitPushdown_| SAME TEXT AS ABOVE_|
+| physical_plan after ProjectionPushdown_| SAME TEXT AS ABOVE_|
+| physical_plan after PushdownSort_| SAME TEXT AS ABOVE_|
+| physical_plan after EnsureCooperative_| SAME TEXT AS ABOVE_|
+| physical_plan after FilterPushdown(Post)_| SAME TEXT AS ABOVE_|
+| physical_plan after WindowedSortRule_| SAME TEXT AS ABOVE_|
+| physical_plan after MatchesConstantTerm_| SAME TEXT AS ABOVE_|
+| physical_plan after RemoveDuplicateRule_| SAME TEXT AS ABOVE_|
+| physical_plan after SanityCheckPlan_| SAME TEXT AS ABOVE_|
+| physical_plan_| PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[j]_|
+|_|_PromSeriesDivideExec: tags=["k"]_|
+|_|_SortExec: expr=[k@1 ASC, j@2 ASC], preserve_partitioning=[true]_|
+|_|_RepartitionExec: REDACTED
+|_|_ProjectionExec: expr=[i@0 as i, k@2 as k, CAST(j@1 AS Timestamp(ms)) as j]_|
+|_|_FilterExec: __common_expr_3@0 >= -299999 AND __common_expr_3@0 <= 10000, projection=[i@1, j@2, k@3]_|
+|_|_ProjectionExec: expr=[CAST(j@1 AS Timestamp(ms)) as __common_expr_3, i@0 as i, j@1 as j, k@2 as k]_|
+|_|_MergeScanExec: REDACTED
+|_|_|
+| physical_plan_with_stats_| PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[j], statistics=[Rows=Inexact(2), Bytes=Absent, [(Col[0]:),(Col[1]:),(Col[2]:)]]_|
+|_|_PromSeriesDivideExec: tags=["k"], statistics=[Rows=Absent, Bytes=Absent, [(Col[0]:),(Col[1]:),(Col[2]:)]]_|
+|_|_SortExec: expr=[k@1 ASC, j@2 ASC], preserve_partitioning=[true], statistics=[Rows=Absent, Bytes=Absent, [(Col[0]:),(Col[1]:),(Col[2]:)]]_|
+|_|_RepartitionExec: REDACTED
+|_|_ProjectionExec: expr=[i@0 as i, k@2 as k, CAST(j@1 AS Timestamp(ms)) as j], statistics=[Rows=Absent, Bytes=Absent, [(Col[0]:),(Col[1]:),(Col[2]:)]]_|
+|_|_FilterExec: __common_expr_3@0 >= -299999 AND __common_expr_3@0 <= 10000, projection=[i@1, j@2, k@3], statistics=[Rows=Absent, Bytes=Absent, [(Col[0]:),(Col[1]:),(Col[2]:)]]_|
+|_|_ProjectionExec: expr=[CAST(j@1 AS Timestamp(ms)) as __common_expr_3, i@0 as i, j@1 as j, k@2 as k], statistics=[Rows=Absent, Bytes=Absent, [(Col[0]:),(Col[1]:),(Col[2]:),(Col[3]:)]] |
+|_|_MergeScanExec: REDACTED
+|_|_|
+| physical_plan_with_schema_| PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[j], schema=[i:Float64;N, k:Utf8;N, j:Timestamp(ms)]_|
+|_|_PromSeriesDivideExec: tags=["k"], schema=[i:Float64;N, k:Utf8;N, j:Timestamp(ms)]_|
+|_|_SortExec: expr=[k@1 ASC, j@2 ASC], preserve_partitioning=[true], schema=[i:Float64;N, k:Utf8;N, j:Timestamp(ms)]_|
+|_|_RepartitionExec: REDACTED
+|_|_ProjectionExec: expr=[i@0 as i, k@2 as k, CAST(j@1 AS Timestamp(ms)) as j], schema=[i:Float64;N, k:Utf8;N, j:Timestamp(ms)]_|
+|_|_FilterExec: __common_expr_3@0 >= -299999 AND __common_expr_3@0 <= 10000, projection=[i@1, j@2, k@3], schema=[i:Float64;N, j:Timestamp(ns), k:Utf8;N]_|
+|_|_ProjectionExec: expr=[CAST(j@1 AS Timestamp(ms)) as __common_expr_3, i@0 as i, j@1 as j, k@2 as k], schema=[__common_expr_3:Timestamp(ms), i:Float64;N, j:Timestamp(ns), k:Utf8;N]_|
+|_|_MergeScanExec: REDACTED
+|_|_|
++-+-+
+
+DROP TABLE test_nano;
+
+Affected Rows: 0
+
 DROP TABLE test;
 
 Affected Rows: 0
diff --git a/tests/cases/standalone/common/tql-explain-analyze/explain.sql b/tests/cases/standalone/common/tql-explain-analyze/explain.sql
index 5e296d4b6b..62685b0952 100644
--- a/tests/cases/standalone/common/tql-explain-analyze/explain.sql
+++ b/tests/cases/standalone/common/tql-explain-analyze/explain.sql
@@ -25,6 +25,7 @@ TQL EXPLAIN ('1970-01-01T00:00:00'::timestamp, '1970-01-01T00:00:00'::timestamp
 -- SQLNESS REPLACE (elapsed_compute.*) REDACTED
 -- SQLNESS REPLACE (peers.*) REDACTED
 -- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (RepartitionExec:.*) RepartitionExec: REDACTED
 TQL EXPLAIN VERBOSE (0, 10, '5s') test;
 
 -- explain verbose at 0s, 5s and 10s. No point at 0s.
@@ -33,6 +34,28 @@ TQL EXPLAIN VERBOSE (0, 10, '5s') test;
 -- SQLNESS REPLACE (elapsed_compute.*) REDACTED
 -- SQLNESS REPLACE (peers.*) REDACTED
 -- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (RepartitionExec:.*) RepartitionExec: REDACTED
 TQL EXPLAIN VERBOSE (0, 10, '5s') test AS series;
 
+CREATE TABLE test_nano(i DOUBLE, j TIMESTAMP(9) TIME INDEX, k STRING PRIMARY KEY);
+
+INSERT INTO test_nano VALUES (1, 1000000, "a"), (1, 1000000, "b"), (2, 2000000, "a");
+
+-- explain at 0s, 5s and 10s for a nanosecond time index.
+-- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (RepartitionExec:.*) RepartitionExec: REDACTED
+TQL EXPLAIN (0, 10, '5s') test_nano;
+
+-- explain verbose at 0s, 5s and 10s for a nanosecond time index.
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (elapsed_compute.*) REDACTED
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (RepartitionExec:.*) RepartitionExec: REDACTED
+TQL EXPLAIN VERBOSE (0, 10, '5s') test_nano;
+
+DROP TABLE test_nano;
+
 DROP TABLE test;
diff --git a/tests/cases/standalone/common/tql/tql-cte.result b/tests/cases/standalone/common/tql/tql-cte.result
index a8c0c45d5d..e8278e80bd 100644
--- a/tests/cases/standalone/common/tql/tql-cte.result
+++ b/tests/cases/standalone/common/tql/tql-cte.result
@@ -427,8 +427,8 @@ SELECT min(val) as min_computed, max(val) as max_computed FROM computed;
 |               |   Aggregate: groupBy=[[]], aggr=[[min(computed.val), max(computed.val)]]                                                    |
 |               |     SubqueryAlias: computed                                                                                                 |
 |               |       Projection: metric.ts AS ts, val * Float64(2) + Float64(1) AS val                                                     |
-|               |         Projection: metric.ts, val * Float64(2) + Float64(1) AS val * Float64(2) + Float64(1)                               |
-|               |           Projection: metric.ts, metric.val * Float64(2) AS val * Float64(2)                                                |
+|               |         Projection: metric.ts, CAST(val * Float64(2) AS Float64) + Float64(1) AS val * Float64(2) + Float64(1)              |
+|               |           Projection: metric.ts, CAST(metric.val AS Float64) * Float64(2) AS val * Float64(2)                               |
 |               |             PromInstantManipulate: range=[0..40000], lookback=[300000], interval=[10000], time index=[ts]                   |
 |               |               PromSeriesDivide: tags=[]                                                                                     |
 |               |                 Filter: metric.ts >= TimestampMillisecond(-299999, None) AND metric.ts <= TimestampMillisecond(40000, None) |
diff --git a/tests/cases/standalone/common/types/json/json-structured.result b/tests/cases/standalone/common/types/json/json-structured.result
deleted file mode 100644
index be04e2652d..0000000000
--- a/tests/cases/standalone/common/types/json/json-structured.result
+++ /dev/null
@@ -1,82 +0,0 @@
-CREATE TABLE t (ts TIMESTAMP TIME INDEX, j JSON(format = "structured") DEFAULT '{"foo": "bar"}');
-
-Error: 1001(Unsupported), Unsupported default constraint for column: 'j', reason: json column cannot have a default value
-
-CREATE TABLE t (ts TIMESTAMP TIME INDEX, j JSON(format = "structured"));
-
-Affected Rows: 0
-
-DESC TABLE t;
-
-+--------+----------------------+-----+------+---------+---------------+
-| Column | Type                 | Key | Null | Default | Semantic Type |
-+--------+----------------------+-----+------+---------+---------------+
-| ts     | TimestampMillisecond | PRI | NO   |         | TIMESTAMP     |
-| j      | Json<"<Null>">       |     | YES  |         | FIELD         |
-+--------+----------------------+-----+------+---------+---------------+
-
-INSERT INTO t VALUES
-(1762128001000, '{"int": 1}'),
-(1762128002000, '{"int": 2, "list": [0.1, 0.2, 0.3]}'),
-(1762128003000, '{"int": 3, "list": [0.4, 0.5, 0.6], "nested": {"a": {"x": "hello"}, "b": {"y": -1}}}');
-
-Affected Rows: 3
-
-DESC TABLE t;
-
-+--------+---------------------------------------------------------------------------------------------------+-----+------+---------+---------------+
-| Column | Type                                                                                              | Key | Null | Default | Semantic Type |
-+--------+---------------------------------------------------------------------------------------------------+-----+------+---------+---------------+
-| ts     | TimestampMillisecond                                                                              | PRI | NO   |         | TIMESTAMP     |
-| j      | Json<{"int":"<Number>","list":["<Number>"],"nested":{"a":{"x":"<String>"},"b":{"y":"<Number>"}}}> |     | YES  |         | FIELD         |
-+--------+---------------------------------------------------------------------------------------------------+-----+------+---------+---------------+
-
-INSERT INTO t VALUES
-(1762128004000, '{"int": 4, "bool": true, "nested": {"a": {"y": 1}}}'),
-(1762128005000, '{"int": 5, "bool": false, "nested": {"b": {"x": "world"}}}');
-
-Affected Rows: 2
-
-DESC TABLE t;
-
-+--------+-------------------------------------------------------------------------------------------------------------------------------------------------+-----+------+---------+---------------+
-| Column | Type                                                                                                                                            | Key | Null | Default | Semantic Type |
-+--------+-------------------------------------------------------------------------------------------------------------------------------------------------+-----+------+---------+---------------+
-| ts     | TimestampMillisecond                                                                                                                            | PRI | NO   |         | TIMESTAMP     |
-| j      | Json<{"bool":"<Bool>","int":"<Number>","list":["<Number>"],"nested":{"a":{"x":"<String>","y":"<Number>"},"b":{"x":"<String>","y":"<Number>"}}}> |     | YES  |         | FIELD         |
-+--------+-------------------------------------------------------------------------------------------------------------------------------------------------+-----+------+---------+---------------+
-
-INSERT INTO t VALUES (1762128006000, '{"int": 6, "list": [-6.0], "bool": true, "nested": {"a": {"x": "ax", "y": 66}, "b": {"y": -66, "x": "bx"}}}');
-
-Affected Rows: 1
-
-DESC TABLE t;
-
-+--------+-------------------------------------------------------------------------------------------------------------------------------------------------+-----+------+---------+---------------+
-| Column | Type                                                                                                                                            | Key | Null | Default | Semantic Type |
-+--------+-------------------------------------------------------------------------------------------------------------------------------------------------+-----+------+---------+---------------+
-| ts     | TimestampMillisecond                                                                                                                            | PRI | NO   |         | TIMESTAMP     |
-| j      | Json<{"bool":"<Bool>","int":"<Number>","list":["<Number>"],"nested":{"a":{"x":"<String>","y":"<Number>"},"b":{"x":"<String>","y":"<Number>"}}}> |     | YES  |         | FIELD         |
-+--------+-------------------------------------------------------------------------------------------------------------------------------------------------+-----+------+---------+---------------+
-
-INSERT INTO t VALUES (1762128011000, '{}');
-
-Error: 1004(InvalidArguments), Invalid InsertRequest, reason: empty json object is not supported, consider adding a dummy field
-
-SELECT ts, j FROM t order by ts;
-
-+---------------------+----------------------------------------------------------------------------------------+
-| ts                  | j                                                                                      |
-+---------------------+----------------------------------------------------------------------------------------+
-| 2025-11-03T00:00:01 | {bool: , int: 1, list: , nested: }                                                     |
-| 2025-11-03T00:00:02 | {bool: , int: 2, list: [0.1, 0.2, 0.3], nested: }                                      |
-| 2025-11-03T00:00:03 | {bool: , int: 3, list: [0.4, 0.5, 0.6], nested: {a: {x: hello, y: }, b: {x: , y: -1}}} |
-| 2025-11-03T00:00:04 | {bool: true, int: 4, list: , nested: {a: {x: , y: 1}, b: }}                            |
-| 2025-11-03T00:00:05 | {bool: false, int: 5, list: , nested: {a: , b: {x: world, y: }}}                       |
-| 2025-11-03T00:00:06 | {bool: true, int: 6, list: [-6.0], nested: {a: {x: ax, y: 66}, b: {x: bx, y: -66}}}    |
-+---------------------+----------------------------------------------------------------------------------------+
-
-DROP table t;
-
-Affected Rows: 0
-
diff --git a/tests/cases/standalone/common/types/json/json-structured.sql b/tests/cases/standalone/common/types/json/json-structured.sql
deleted file mode 100644
index 8bb10b4b0e..0000000000
--- a/tests/cases/standalone/common/types/json/json-structured.sql
+++ /dev/null
@@ -1,28 +0,0 @@
-CREATE TABLE t (ts TIMESTAMP TIME INDEX, j JSON(format = "structured") DEFAULT '{"foo": "bar"}');
-
-CREATE TABLE t (ts TIMESTAMP TIME INDEX, j JSON(format = "structured"));
-
-DESC TABLE t;
-
-INSERT INTO t VALUES
-(1762128001000, '{"int": 1}'),
-(1762128002000, '{"int": 2, "list": [0.1, 0.2, 0.3]}'),
-(1762128003000, '{"int": 3, "list": [0.4, 0.5, 0.6], "nested": {"a": {"x": "hello"}, "b": {"y": -1}}}');
-
-DESC TABLE t;
-
-INSERT INTO t VALUES
-(1762128004000, '{"int": 4, "bool": true, "nested": {"a": {"y": 1}}}'),
-(1762128005000, '{"int": 5, "bool": false, "nested": {"b": {"x": "world"}}}');
-
-DESC TABLE t;
-
-INSERT INTO t VALUES (1762128006000, '{"int": 6, "list": [-6.0], "bool": true, "nested": {"a": {"x": "ax", "y": 66}, "b": {"y": -66, "x": "bx"}}}');
-
-DESC TABLE t;
-
-INSERT INTO t VALUES (1762128011000, '{}');
-
-SELECT ts, j FROM t order by ts;
-
-DROP table t;
diff --git a/tests/cases/standalone/common/types/json/json.result b/tests/cases/standalone/common/types/json/json.result
index 8c4755f4ae..8fad9632b1 100644
--- a/tests/cases/standalone/common/types/json/json.result
+++ b/tests/cases/standalone/common/types/json/json.result
@@ -37,22 +37,23 @@ INSERT INTO jsons VALUES('[null]', 0),
             }
         ]
     }
-}}', 11);
+}}', 11),
+('{"a":"abc\u2028tom"}', 12);
 
-Affected Rows: 12
+Affected Rows: 13
 
-INSERT INTO jsons VALUES(parse_json('[null]'), 12),
-(parse_json('[true]'), 13),
-(parse_json('[false]'), 14),
-(parse_json('[0]'), 15),
-(parse_json('["foo"]'), 16),
-(parse_json('[]'), 17),
-(parse_json('{}'), 18),
-(parse_json('[0,1]'), 19),
-(parse_json('{"foo":"bar"}'), 20),
-(parse_json('{"a":null,"foo":"bar"}'), 21),
-(parse_json('[-1]'), 22),
-(parse_json('[-2147483648]'), 23),
+INSERT INTO jsons VALUES(parse_json('[null]'), 1000),
+(parse_json('[true]'), 1001),
+(parse_json('[false]'), 1002),
+(parse_json('[0]'), 1003),
+(parse_json('["foo"]'), 1004),
+(parse_json('[]'), 1005),
+(parse_json('{}'), 1006),
+(parse_json('[0,1]'), 1007),
+(parse_json('{"foo":"bar"}'), 1008),
+(parse_json('{"a":null,"foo":"bar"}'), 1009),
+(parse_json('[-1]'), 1010),
+(parse_json('[-2147483648]'), 1011),
 (parse_json('{"entities": {
             "description": {
                 "urls": [
@@ -76,9 +77,10 @@ INSERT INTO jsons VALUES(parse_json('[null]'), 12),
                     }
                 ]
             }
-        }}'), 24);
+        }}'), 1012),
+(parse_json('{"a":"abc\u2028tom"}'), 1013);
 
-Affected Rows: 13
+Affected Rows: 14
 
 SELECT json_to_string(j), t FROM jsons;
 
@@ -97,25 +99,27 @@ SELECT json_to_string(j), t FROM jsons;
 | {"a":null,"foo":"bar"}                                                                                                                                                                                                                                                                                                    | 1970-01-01T00:00:00.009 |
 | [-1]                                                                                                                                                                                                                                                                                                                      | 1970-01-01T00:00:00.010 |
 | {"entities":{"description":{"urls":[{"display_url":"pixiv.net/member.php?id=…","expanded_url":"http://www.pixiv.net/member.php?id=4776","indices":[58,80],"url":"http://t.co/QMLJeFmfMT"},{"display_url":"ask.fm/KATANA77","expanded_url":"http://ask.fm/KATANA77","indices":[95,117],"url":"http://t.co/LU8T7vmU3h"}]}}} | 1970-01-01T00:00:00.011 |
-| [null]                                                                                                                                                                                                                                                                                                                    | 1970-01-01T00:00:00.012 |
-| [true]                                                                                                                                                                                                                                                                                                                    | 1970-01-01T00:00:00.013 |
-| [false]                                                                                                                                                                                                                                                                                                                   | 1970-01-01T00:00:00.014 |
-| [0]                                                                                                                                                                                                                                                                                                                       | 1970-01-01T00:00:00.015 |
-| ["foo"]                                                                                                                                                                                                                                                                                                                   | 1970-01-01T00:00:00.016 |
-| []                                                                                                                                                                                                                                                                                                                        | 1970-01-01T00:00:00.017 |
-| {}                                                                                                                                                                                                                                                                                                                        | 1970-01-01T00:00:00.018 |
-| [0,1]                                                                                                                                                                                                                                                                                                                     | 1970-01-01T00:00:00.019 |
-| {"foo":"bar"}                                                                                                                                                                                                                                                                                                             | 1970-01-01T00:00:00.020 |
-| {"a":null,"foo":"bar"}                                                                                                                                                                                                                                                                                                    | 1970-01-01T00:00:00.021 |
-| [-1]                                                                                                                                                                                                                                                                                                                      | 1970-01-01T00:00:00.022 |
-| [-2147483648]                                                                                                                                                                                                                                                                                                             | 1970-01-01T00:00:00.023 |
-| {"entities":{"description":{"urls":[{"display_url":"pixiv.net/member.php?id=…","expanded_url":"http://www.pixiv.net/member.php?id=4776","indices":[58,80],"url":"http://t.co/QMLJeFmfMT"},{"display_url":"ask.fm/KATANA77","expanded_url":"http://ask.fm/KATANA77","indices":[95,117],"url":"http://t.co/LU8T7vmU3h"}]}}} | 1970-01-01T00:00:00.024 |
+| {"a":"abc\u2028tom"}                                                                                                                                                                                                                                                                                                      | 1970-01-01T00:00:00.012 |
+| [null]                                                                                                                                                                                                                                                                                                                    | 1970-01-01T00:00:01     |
+| [true]                                                                                                                                                                                                                                                                                                                    | 1970-01-01T00:00:01.001 |
+| [false]                                                                                                                                                                                                                                                                                                                   | 1970-01-01T00:00:01.002 |
+| [0]                                                                                                                                                                                                                                                                                                                       | 1970-01-01T00:00:01.003 |
+| ["foo"]                                                                                                                                                                                                                                                                                                                   | 1970-01-01T00:00:01.004 |
+| []                                                                                                                                                                                                                                                                                                                        | 1970-01-01T00:00:01.005 |
+| {}                                                                                                                                                                                                                                                                                                                        | 1970-01-01T00:00:01.006 |
+| [0,1]                                                                                                                                                                                                                                                                                                                     | 1970-01-01T00:00:01.007 |
+| {"foo":"bar"}                                                                                                                                                                                                                                                                                                             | 1970-01-01T00:00:01.008 |
+| {"a":null,"foo":"bar"}                                                                                                                                                                                                                                                                                                    | 1970-01-01T00:00:01.009 |
+| [-1]                                                                                                                                                                                                                                                                                                                      | 1970-01-01T00:00:01.010 |
+| [-2147483648]                                                                                                                                                                                                                                                                                                             | 1970-01-01T00:00:01.011 |
+| {"entities":{"description":{"urls":[{"display_url":"pixiv.net/member.php?id=…","expanded_url":"http://www.pixiv.net/member.php?id=4776","indices":[58,80],"url":"http://t.co/QMLJeFmfMT"},{"display_url":"ask.fm/KATANA77","expanded_url":"http://ask.fm/KATANA77","indices":[95,117],"url":"http://t.co/LU8T7vmU3h"}]}}} | 1970-01-01T00:00:01.012 |
+| {"a":"abc\u2028tom"}                                                                                                                                                                                                                                                                                                      | 1970-01-01T00:00:01.013 |
 +---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+-------------------------+
 
 --Insert invalid json strings--
 DELETE FROM jsons;
 
-Affected Rows: 25
+Affected Rows: 27
 
 INSERT INTO jsons VALUES(parse_json('{"a":1, "b":2, "c":3'), 4);
 
diff --git a/tests/cases/standalone/common/types/json/json.sql b/tests/cases/standalone/common/types/json/json.sql
index 868edc59e8..5a521ee1c6 100644
--- a/tests/cases/standalone/common/types/json/json.sql
+++ b/tests/cases/standalone/common/types/json/json.sql
@@ -35,20 +35,21 @@ INSERT INTO jsons VALUES('[null]', 0),
             }
         ]
     }
-}}', 11);
+}}', 11),
+('{"a":"abc\u2028tom"}', 12);
 
-INSERT INTO jsons VALUES(parse_json('[null]'), 12),
-(parse_json('[true]'), 13),
-(parse_json('[false]'), 14),
-(parse_json('[0]'), 15),
-(parse_json('["foo"]'), 16),
-(parse_json('[]'), 17),
-(parse_json('{}'), 18),
-(parse_json('[0,1]'), 19),
-(parse_json('{"foo":"bar"}'), 20),
-(parse_json('{"a":null,"foo":"bar"}'), 21),
-(parse_json('[-1]'), 22),
-(parse_json('[-2147483648]'), 23),
+INSERT INTO jsons VALUES(parse_json('[null]'), 1000),
+(parse_json('[true]'), 1001),
+(parse_json('[false]'), 1002),
+(parse_json('[0]'), 1003),
+(parse_json('["foo"]'), 1004),
+(parse_json('[]'), 1005),
+(parse_json('{}'), 1006),
+(parse_json('[0,1]'), 1007),
+(parse_json('{"foo":"bar"}'), 1008),
+(parse_json('{"a":null,"foo":"bar"}'), 1009),
+(parse_json('[-1]'), 1010),
+(parse_json('[-2147483648]'), 1011),
 (parse_json('{"entities": {
             "description": {
                 "urls": [
@@ -72,7 +73,8 @@ INSERT INTO jsons VALUES(parse_json('[null]'), 12),
                     }
                 ]
             }
-        }}'), 24);
+        }}'), 1012),
+(parse_json('{"a":"abc\u2028tom"}'), 1013);
 
 SELECT json_to_string(j), t FROM jsons;
 
diff --git a/tests/cases/standalone/common/types/json/json2.result b/tests/cases/standalone/common/types/json/json2.result
new file mode 100644
index 0000000000..5bdc641433
--- /dev/null
+++ b/tests/cases/standalone/common/types/json/json2.result
@@ -0,0 +1,149 @@
+create table json2_table (
+    ts timestamp time index,
+    j  json2
+) with (
+    'append_mode' = 'true',
+    'sst_format' = 'flat',
+);
+
+Affected Rows: 0
+
+insert into json2_table (ts, j)
+values (1, '{"a": {"b": 1}, "c": "s1", "d": [{"e": {"f": 0.1}}]}'),
+       (2, '{"a": {"b": -2}, "c": "s2", "d": [{"e": {"f": 0.2}}]}');
+
+Affected Rows: 2
+
+admin flush_table('json2_table');
+
++----------------------------------+
+| ADMIN flush_table('json2_table') |
++----------------------------------+
+| 0                                |
++----------------------------------+
+
+insert into json2_table (ts, j)
+values (3, '{"a": {"b": 3}, "c": "s3"}');
+
+Affected Rows: 1
+
+insert into json2_table
+values (4, '{"a": {"b": -4}, "d": [{"e": {"g": -0.4}}]}'),
+       (5, '{"a": {}, "c": "s5"}'),
+       (6, '{"c": "s6"}');
+
+Affected Rows: 3
+
+admin flush_table('json2_table');
+
++----------------------------------+
+| ADMIN flush_table('json2_table') |
++----------------------------------+
+| 0                                |
++----------------------------------+
+
+insert into json2_table
+values (7, '{"a": {"b": "s7"}, "c": [1], "d": [{"e": {"g": -0.7}}]}'),
+       (8, '{"a": {"b": 8}, "c": "s8"}');
+
+Affected Rows: 2
+
+admin flush_table('json2_table');
+
++----------------------------------+
+| ADMIN flush_table('json2_table') |
++----------------------------------+
+| 0                                |
++----------------------------------+
+
+insert into json2_table
+values (9, '{"a": {"x": true}, "c": "s9", "d": [{"e": {"g": -0.9}}]}'),
+       (10, '{"a": {"b": 10}, "y": false}');
+
+Affected Rows: 2
+
+-- SQLNESS REPLACE (peers.*) REDACTED
+explain select j.a.b from json2_table;
+
++---------------+---------------------------------------------------+
+| plan_type     | plan                                              |
++---------------+---------------------------------------------------+
+| logical_plan  | MergeScan [is_placeholder=false, remote_input=[   |
+|               | Projection: json_get(json2_table.j, Utf8("a.b"))  |
+|               |   TableScan: json2_table                          |
+|               | ]]                                                |
+| physical_plan | CooperativeExec                                   |
+|               |   MergeScanExec: REDACTED
+|               |                                                   |
++---------------+---------------------------------------------------+
+
+-- SQLNESS REPLACE (peers.*) REDACTED
+explain select j.a.x::bool from json2_table;
+
++---------------+------------------------------------------------------------------------------------------------------------------------------------+
+| plan_type     | plan                                                                                                                               |
++---------------+------------------------------------------------------------------------------------------------------------------------------------+
+| logical_plan  | MergeScan [is_placeholder=false, remote_input=[                                                                                    |
+|               | Projection: json_get(json2_table.j, Utf8("a.x"), Boolean(NULL)) AS arrow_cast(json_get(json2_table.j,Utf8("a.x")),Utf8("Boolean")) |
+|               |   TableScan: json2_table                                                                                                           |
+|               | ]]                                                                                                                                 |
+| physical_plan | CooperativeExec                                                                                                                    |
+|               |   MergeScanExec: REDACTED
+|               |                                                                                                                                    |
++---------------+------------------------------------------------------------------------------------------------------------------------------------+
+
+select j.a.b from json2_table order by ts;
+
++-------------------------------------+
+| json_get(json2_table.j,Utf8("a.b")) |
++-------------------------------------+
+| 1                                   |
+| -2                                  |
+| 3                                   |
+| -4                                  |
+|                                     |
+|                                     |
+| "s7"                                |
+| 8                                   |
+|                                     |
+| 10                                  |
++-------------------------------------+
+
+select j.c, j.y from json2_table order by ts;
+
++-----------------------------------+-----------------------------------+
+| json_get(json2_table.j,Utf8("c")) | json_get(json2_table.j,Utf8("y")) |
++-----------------------------------+-----------------------------------+
+| s1                                |                                   |
+| s2                                |                                   |
+| s3                                |                                   |
+|                                   |                                   |
+| s5                                |                                   |
+| s6                                |                                   |
+| [1]                               |                                   |
+| "s8"                              |                                   |
+| s9                                |                                   |
+|                                   | false                             |
++-----------------------------------+-----------------------------------+
+
+select j.d from json2_table order by ts;
+
++-----------------------------------+
+| json_get(json2_table.j,Utf8("d")) |
++-----------------------------------+
+| [{e: {f: 0.1}}]                   |
+| [{e: {f: 0.2}}]                   |
+|                                   |
+| [{e: {g: -0.4}}]                  |
+|                                   |
+|                                   |
+| [{e: {g: -0.7}}]                  |
+|                                   |
+| [{e: {g: -0.9}}]                  |
+|                                   |
++-----------------------------------+
+
+drop table json2_table;
+
+Affected Rows: 0
+
diff --git a/tests/cases/standalone/common/types/json/json2.sql b/tests/cases/standalone/common/types/json/json2.sql
new file mode 100644
index 0000000000..75c7b46b41
--- /dev/null
+++ b/tests/cases/standalone/common/types/json/json2.sql
@@ -0,0 +1,47 @@
+create table json2_table (
+    ts timestamp time index,
+    j  json2
+) with (
+    'append_mode' = 'true',
+    'sst_format' = 'flat',
+);
+
+insert into json2_table (ts, j)
+values (1, '{"a": {"b": 1}, "c": "s1", "d": [{"e": {"f": 0.1}}]}'),
+       (2, '{"a": {"b": -2}, "c": "s2", "d": [{"e": {"f": 0.2}}]}');
+
+admin flush_table('json2_table');
+
+insert into json2_table (ts, j)
+values (3, '{"a": {"b": 3}, "c": "s3"}');
+
+insert into json2_table
+values (4, '{"a": {"b": -4}, "d": [{"e": {"g": -0.4}}]}'),
+       (5, '{"a": {}, "c": "s5"}'),
+       (6, '{"c": "s6"}');
+
+admin flush_table('json2_table');
+
+insert into json2_table
+values (7, '{"a": {"b": "s7"}, "c": [1], "d": [{"e": {"g": -0.7}}]}'),
+       (8, '{"a": {"b": 8}, "c": "s8"}');
+
+admin flush_table('json2_table');
+
+insert into json2_table
+values (9, '{"a": {"x": true}, "c": "s9", "d": [{"e": {"g": -0.9}}]}'),
+       (10, '{"a": {"b": 10}, "y": false}');
+
+-- SQLNESS REPLACE (peers.*) REDACTED
+explain select j.a.b from json2_table;
+
+-- SQLNESS REPLACE (peers.*) REDACTED
+explain select j.a.x::bool from json2_table;
+
+select j.a.b from json2_table order by ts;
+
+select j.c, j.y from json2_table order by ts;
+
+select j.d from json2_table order by ts;
+
+drop table json2_table;
diff --git a/tests/cases/standalone/common/view/create.result b/tests/cases/standalone/common/view/create.result
index 1c6e0ee50b..76b9838628 100644
--- a/tests/cases/standalone/common/view/create.result
+++ b/tests/cases/standalone/common/view/create.result
@@ -30,6 +30,10 @@ CREATE VIEW test_view as SELECT * FROM public.numbers;
 
 Affected Rows: 0
 
+CREATE VIEW test_view2 as SELECT * FROM test_view;
+
+Affected Rows: 0
+
 --- View already exists ----
 CREATE VIEW test_view as SELECT * FROM public.numbers;
 
@@ -51,6 +55,7 @@ SHOW TABLES;
 | numbers          |
 | test_table       |
 | test_view        |
+| test_view2       |
 +------------------+
 
 SHOW FULL TABLES;
@@ -61,6 +66,7 @@ SHOW FULL TABLES;
 | numbers          | LOCAL TEMPORARY |
 | test_table       | BASE TABLE      |
 | test_view        | VIEW            |
+| test_view2       | VIEW            |
 +------------------+-----------------+
 
 -- psql: \dv
@@ -124,17 +130,19 @@ SELECT * FROM INFORMATION_SCHEMA.TABLES ORDER BY TABLE_NAME, TABLE_TYPE;
 |greptime|information_schema|tables|LOCALTEMPORARY|ID|ID|ID|ID|ID|ID||ID|Fixed|ID|ID|ID|DATETIME|DATETIME||utf8_bin|ID|||Y|
 |greptime|public|test_table|BASETABLE|ID|ID|ID|ID|ID|ID|mito|ID|Fixed|ID|ID|ID|DATETIME|DATETIME||utf8_bin|ID|||N|
 |greptime|public|test_view|VIEW|ID|ID|ID|ID|ID|ID||ID|Fixed|ID|ID|ID|DATETIME|DATETIME||utf8_bin|ID|||N|
+|greptime|public|test_view2|VIEW|ID|ID|ID|ID|ID|ID||ID|Fixed|ID|ID|ID|DATETIME|DATETIME||utf8_bin|ID|||N|
 |greptime|information_schema|views|LOCALTEMPORARY|ID|ID|ID|ID|ID|ID||ID|Fixed|ID|ID|ID|DATETIME|DATETIME||utf8_bin|ID|||Y|
 +++++++++++++++++++++++++
 
 -- SQLNESS REPLACE (\s\d+\s) ID
 -- SQLNESS REPLACE (\s[\-0-9T:\.]{15,}) DATETIME
-SELECT * FROM INFORMATION_SCHEMA.TABLES WHERE TABLE_TYPE = 'VIEW';
+SELECT * FROM INFORMATION_SCHEMA.TABLES WHERE TABLE_TYPE = 'VIEW' ORDER BY TABLE_NAME;
 
 +---------------+--------------+------------+------------+----------+-------------+-----------------+--------------+------------------+----------------+--------+---------+------------+------------+-----------+----------------+---------------------+---------------------+------------+-----------------+----------+----------------+---------------+-----------+
 | table_catalog | table_schema | table_name | table_type | table_id | data_length | max_data_length | index_length | max_index_length | avg_row_length | engine | version | row_format | table_rows | data_free | auto_increment | create_time         | update_time         | check_time | table_collation | checksum | create_options | table_comment | temporary |
 +---------------+--------------+------------+------------+----------+-------------+-----------------+--------------+------------------+----------------+--------+---------+------------+------------+-----------+----------------+---------------------+---------------------+------------+-----------------+----------+----------------+---------------+-----------+
 | greptime      | public       | test_view  | VIEW       |ID    |ID          |ID              |ID           |ID               |ID             |        |ID     | Fixed      |ID         |ID        |ID             |DATETIME |DATETIME |            | utf8_bin        |ID       |                |               | N         |
+| greptime      | public       | test_view2 | VIEW       |ID    |ID          |ID              |ID           |ID               |ID             |        |ID     | Fixed      |ID         |ID        |ID             |DATETIME |DATETIME |            | utf8_bin        |ID       |                |               | N         |
 +---------------+--------------+------------+------------+----------+-------------+-----------------+--------------+------------------+----------------+--------+---------+------------+------------+-----------+----------------+---------------------+---------------------+------------+-----------------+----------+----------------+---------------+-----------+
 
 SHOW COLUMNS FROM test_view;
@@ -169,10 +177,31 @@ SELECT * FROM test_view LIMIT 10;
 | 9      |
 +--------+
 
+SELECT * FROM test_view2 LIMIT 10;
+
++--------+
+| number |
++--------+
+| 0      |
+| 1      |
+| 2      |
+| 3      |
+| 4      |
+| 5      |
+| 6      |
+| 7      |
+| 8      |
+| 9      |
++--------+
+
 DROP VIEW test_view;
 
 Affected Rows: 0
 
+DROP VIEW test_view2;
+
+Affected Rows: 0
+
 DROP TABLE test_table;
 
 Affected Rows: 0
diff --git a/tests/cases/standalone/common/view/create.sql b/tests/cases/standalone/common/view/create.sql
index b82704d3a9..91149f44f4 100644
--- a/tests/cases/standalone/common/view/create.sql
+++ b/tests/cases/standalone/common/view/create.sql
@@ -16,6 +16,8 @@ CREATE OR REPLACE VIEW test_table as SELECT * FROM public.numbers;
 
 CREATE VIEW test_view as SELECT * FROM public.numbers;
 
+CREATE VIEW test_view2 as SELECT * FROM test_view;
+
 --- View already exists ----
 CREATE VIEW test_view as SELECT * FROM public.numbers;
 
@@ -48,7 +50,7 @@ SELECT * FROM INFORMATION_SCHEMA.TABLES ORDER BY TABLE_NAME, TABLE_TYPE;
 
 -- SQLNESS REPLACE (\s\d+\s) ID
 -- SQLNESS REPLACE (\s[\-0-9T:\.]{15,}) DATETIME
-SELECT * FROM INFORMATION_SCHEMA.TABLES WHERE TABLE_TYPE = 'VIEW';
+SELECT * FROM INFORMATION_SCHEMA.TABLES WHERE TABLE_TYPE = 'VIEW' ORDER BY TABLE_NAME;
 
 SHOW COLUMNS FROM test_view;
 
@@ -58,8 +60,12 @@ SELECT * FROM INFORMATION_SCHEMA.COLUMNS WHERE TABLE_NAME = 'test_view';
 
 SELECT * FROM test_view LIMIT 10;
 
+SELECT * FROM test_view2 LIMIT 10;
+
 DROP VIEW test_view;
 
+DROP VIEW test_view2;
+
 DROP TABLE test_table;
 
 SELECT * FROM test_view LIMIT 10;
diff --git a/tests/cases/standalone/flow-tql/flow_tql_flownode_addrs.result b/tests/cases/standalone/flow-tql/flow_tql_flownode_addrs.result
new file mode 100644
index 0000000000..716fbd7d95
--- /dev/null
+++ b/tests/cases/standalone/flow-tql/flow_tql_flownode_addrs.result
@@ -0,0 +1,37 @@
+CREATE TABLE http_requests (
+  ts timestamp(3) time index,
+  host STRING,
+  idc STRING,
+  val DOUBLE,
+  PRIMARY KEY(host, idc),
+);
+
+Affected Rows: 0
+
+CREATE FLOW calc_reqs SINK TO cnt_reqs EVAL INTERVAL '1m' AS
+TQL EVAL (now() - '1m'::interval, now(), '5s') count_values("status_code", http_requests);
+
+Affected Rows: 0
+
+SELECT flownode_addrs IS NULL AS flownode_addrs_is_null
+FROM information_schema.flows
+WHERE flow_name = 'calc_reqs';
+
++------------------------+
+| flownode_addrs_is_null |
++------------------------+
+| true                   |
++------------------------+
+
+DROP FLOW calc_reqs;
+
+Affected Rows: 0
+
+DROP TABLE http_requests;
+
+Affected Rows: 0
+
+DROP TABLE cnt_reqs;
+
+Affected Rows: 0
+
diff --git a/tests/cases/standalone/flow-tql/flow_tql_flownode_addrs.sql b/tests/cases/standalone/flow-tql/flow_tql_flownode_addrs.sql
new file mode 100644
index 0000000000..7aae20aa43
--- /dev/null
+++ b/tests/cases/standalone/flow-tql/flow_tql_flownode_addrs.sql
@@ -0,0 +1,18 @@
+CREATE TABLE http_requests (
+  ts timestamp(3) time index,
+  host STRING,
+  idc STRING,
+  val DOUBLE,
+  PRIMARY KEY(host, idc),
+);
+
+CREATE FLOW calc_reqs SINK TO cnt_reqs EVAL INTERVAL '1m' AS
+TQL EVAL (now() - '1m'::interval, now(), '5s') count_values("status_code", http_requests);
+
+SELECT flownode_addrs IS NULL AS flownode_addrs_is_null
+FROM information_schema.flows
+WHERE flow_name = 'calc_reqs';
+
+DROP FLOW calc_reqs;
+DROP TABLE http_requests;
+DROP TABLE cnt_reqs;
diff --git a/tests/cases/standalone/information_schema/cluster_info.result b/tests/cases/standalone/information_schema/cluster_info.result
index 07fc4bd5c6..04567ff721 100644
--- a/tests/cases/standalone/information_schema/cluster_info.result
+++ b/tests/cases/standalone/information_schema/cluster_info.result
@@ -24,9 +24,9 @@ DESC TABLE CLUSTER_INFO;
 +----------------------+----------------------+-----+------+---------+---------------+
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
 -- SQLNESS REPLACE (\s[\-0-9T:\.]{15,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
 -- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO;
@@ -34,9 +34,9 @@ SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, a
 +++++++++|peer_id|peer_type|peer_addr|node_version|git_commit|start_time|uptime|active_time|+++++++++|0|STANDALONE||Version|Hash|Start_time|Duration||+++++++++
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
 -- SQLNESS REPLACE (\s[\-0-9T:\.]{15,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
 -- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO WHERE PEER_TYPE = 'STANDALONE';
@@ -49,9 +49,9 @@ SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, a
 ++
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
 -- SQLNESS REPLACE (\s[\-0-9T:\.]{15,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
 -- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO WHERE PEER_ID = 0;
diff --git a/tests/cases/standalone/information_schema/cluster_info.sql b/tests/cases/standalone/information_schema/cluster_info.sql
index 5e253fc43d..798e9eff28 100644
--- a/tests/cases/standalone/information_schema/cluster_info.sql
+++ b/tests/cases/standalone/information_schema/cluster_info.sql
@@ -3,17 +3,17 @@ USE INFORMATION_SCHEMA;
 DESC TABLE CLUSTER_INFO;
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
 -- SQLNESS REPLACE (\s[\-0-9T:\.]{15,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
 -- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO;
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
 -- SQLNESS REPLACE (\s[\-0-9T:\.]{15,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
 -- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO WHERE PEER_TYPE = 'STANDALONE';
@@ -21,9 +21,9 @@ SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, a
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO WHERE PEER_TYPE != 'STANDALONE';
 
 -- SQLNESS REPLACE version node_version
--- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*-[a-zA-Z0-9.-]+) Version
--- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
 -- SQLNESS REPLACE (\s[\-0-9T:\.]{15,}) Start_time
+-- SQLNESS REPLACE (\d+\.\d+(?:\.\d+)*(?:-[a-zA-Z0-9.-]+)?) Version
+-- SQLNESS REPLACE (\s[a-z0-9]{7,10}\s) Hash
 -- SQLNESS REPLACE ((\d+(s|ms|m)\s)+) Duration
 -- SQLNESS REPLACE [\s\-]+
 SELECT peer_id, peer_type, peer_addr, version, git_commit, start_time, uptime, active_time FROM CLUSTER_INFO WHERE PEER_ID = 0;
diff --git a/tests/cases/standalone/optimizer/order_by.result b/tests/cases/standalone/optimizer/order_by.result
index 8bc4c14816..06b06ae442 100644
--- a/tests/cases/standalone/optimizer/order_by.result
+++ b/tests/cases/standalone/optimizer/order_by.result
@@ -142,8 +142,7 @@ EXPLAIN ANALYZE SELECT i, t AS alias_ts FROM test_pk ORDER BY t DESC LIMIT 5;
 |_|_|_|
 | 1_| 0_|_ProjectionExec: expr=[i@0 as i, alias_ts@1 as alias_ts] REDACTED
 |_|_|_SortPreservingMergeExec: [t@2 DESC], fetch=5 REDACTED
-|_|_|_WindowedSortExec: expr=alias_ts@1 DESC num_ranges=REDACTED fetch=5 REDACTED
-|_|_|_PartSortExec: expr=alias_ts@1 DESC num_ranges=REDACTED limit=5 REDACTED
+|_|_|_SortExec: TopK(fetch=5), expr=[alias_ts@1 DESC], preserve_partitioning=[true], filter=[alias_ts@1 IS NULL OR alias_ts@1 > 2] REDACTED
 |_|_|_ProjectionExec: expr=[i@0 as i, t@1 as alias_ts, t@1 as t] REDACTED
 |_|_|_SeqScan: region=REDACTED, "partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0} REDACTED
 |_|_|_|
@@ -165,8 +164,7 @@ EXPLAIN ANALYZE SELECT i, t AS alias_ts FROM test_pk ORDER BY alias_ts DESC LIMI
 |_|_|_MergeScanExec: REDACTED
 |_|_|_|
 | 1_| 0_|_SortPreservingMergeExec: [alias_ts@1 DESC], fetch=5 REDACTED
-|_|_|_WindowedSortExec: expr=alias_ts@1 DESC num_ranges=REDACTED fetch=5 REDACTED
-|_|_|_PartSortExec: expr=alias_ts@1 DESC num_ranges=REDACTED limit=5 REDACTED
+|_|_|_SortExec: TopK(fetch=5), expr=[alias_ts@1 DESC], preserve_partitioning=[true], filter=[alias_ts@1 IS NULL OR alias_ts@1 > 2] REDACTED
 |_|_|_ProjectionExec: expr=[i@0 as i, t@1 as alias_ts] REDACTED
 |_|_|_SeqScan: region=REDACTED, "partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0} REDACTED
 |_|_|_|
diff --git a/tests/cases/standalone/optimizer/time_index_filter_pushdown.result b/tests/cases/standalone/optimizer/time_index_filter_pushdown.result
index ea10cfc416..1f281c684a 100644
--- a/tests/cases/standalone/optimizer/time_index_filter_pushdown.result
+++ b/tests/cases/standalone/optimizer/time_index_filter_pushdown.result
@@ -4,6 +4,7 @@ CREATE TABLE
     IF NOT EXISTS `cpu` (
         `rack` STRING NULL,
         `os` STRING NULL,
+        `usage_small` SMALLINT NULL,
         `usage_user` BIGINT NULL,
         `greptime_timestamp` TIMESTAMP(9) NOT NULL,
         TIME INDEX (`greptime_timestamp`),
@@ -25,11 +26,11 @@ Affected Rows: 0
 INSERT INTO
     cpu
 VALUES
-    ("1", "linux", 10, "2023-06-12 01:04:49"),
-    ("1", "linux", 15, "2023-06-12 01:04:50"),
-    ("3", "windows", 25, "2023-06-12 01:05:00"),
-    ("5", "mac", 30, "2023-06-12 01:03:00"),
-    ("7", "linux", 45, "2023-06-12 02:00:00");
+    ("1", "linux", 10, 10, "2023-06-12 01:04:49"),
+    ("1", "linux", 15, 15, "2023-06-12 01:04:50"),
+    ("3", "windows", 25, 25, "2023-06-12 01:05:00"),
+    ("5", "mac", 30, 30, "2023-06-12 01:03:00"),
+    ("7", "linux", 45, 45, "2023-06-12 02:00:00");
 
 Affected Rows: 5
 
@@ -44,14 +45,108 @@ ADMIN FLUSH_TABLE ('cpu');
 INSERT INTO
     cpu
 VALUES
-    ("2", "linux", 20, "2023-06-12 01:04:51"),
-    ("2", "windows", 22, "2023-06-12 01:06:00"),
-    ("4", "mac", 12, "2023-06-12 00:59:00"),
-    ("6", "linux", 35, "2023-06-12 01:04:55"),
-    ("8", "windows", 50, "2023-06-12 02:10:00");
+    ("2", "linux", 20, 20, "2023-06-12 01:04:51"),
+    ("2", "windows", 22, 22, "2023-06-12 01:06:00"),
+    ("4", "mac", 12, 12, "2023-06-12 00:59:00"),
+    ("6", "linux", 35, 35, "2023-06-12 01:04:55"),
+    ("8", "windows", 50, 50, "2023-06-12 02:10:00");
 
 Affected Rows: 5
 
+-- SQLNESS REPLACE (peers.*) REDACTED
+EXPLAIN SELECT
+    rack
+FROM
+    cpu
+WHERE
+    usage_small IN (10, 20);
+
++---------------+---------------------------------------------------------------------------------------------------------------------------------------------------+
+| plan_type     | plan                                                                                                                                              |
++---------------+---------------------------------------------------------------------------------------------------------------------------------------------------+
+| logical_plan  | MergeScan [is_placeholder=false, remote_input=[                                                                                                   |
+|               | Projection: cpu.rack                                                                                                                              |
+|               |   Filter: cpu.usage_small = Int16(10) OR cpu.usage_small = Int16(20)                                                                              |
+|               |     TableScan: cpu                                                                                                                                |
+|               | ]]                                                                                                                                                |
+| physical_plan | CooperativeExec                                                                                                                                   |
+|               |   MergeScanExec: REDACTED
+|               |                                                                                                                                                   |
++---------------+---------------------------------------------------------------------------------------------------------------------------------------------------+
+
+-- SQLNESS REPLACE (peers.*) REDACTED
+EXPLAIN SELECT
+    rack
+FROM
+    cpu
+WHERE
+    usage_small BETWEEN 10 AND 20;
+
++---------------+---------------------------------------------------------------------------------------------------------------------------------------------------+
+| plan_type     | plan                                                                                                                                              |
++---------------+---------------------------------------------------------------------------------------------------------------------------------------------------+
+| logical_plan  | MergeScan [is_placeholder=false, remote_input=[                                                                                                   |
+|               | Projection: cpu.rack                                                                                                                              |
+|               |   Filter: cpu.usage_small >= Int16(10) AND cpu.usage_small <= Int16(20)                                                                           |
+|               |     TableScan: cpu                                                                                                                                |
+|               | ]]                                                                                                                                                |
+| physical_plan | CooperativeExec                                                                                                                                   |
+|               |   MergeScanExec: REDACTED
+|               |                                                                                                                                                   |
++---------------+---------------------------------------------------------------------------------------------------------------------------------------------------+
+
+CREATE TABLE
+    IF NOT EXISTS `cpu_single` (
+        `rack` STRING NULL,
+        `usage_small` SMALLINT NULL,
+        `greptime_timestamp` TIMESTAMP(9) NOT NULL,
+        TIME INDEX (`greptime_timestamp`),
+    ) ENGINE = mito
+WITH
+    (append_mode = 'true', sst_format = 'flat');
+
+Affected Rows: 0
+
+INSERT INTO
+    cpu_single
+VALUES
+    ("1", 10, "2023-06-12 01:04:49"),
+    ("2", 20, "2023-06-12 01:04:50"),
+    ("3", 25, "2023-06-12 01:05:00");
+
+Affected Rows: 3
+
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE "partition_count":\{(.*?)\} "partition_count":REDACTED
+-- SQLNESS REPLACE "flat_format":\s\w+, "flat_format": REDACTED,
+EXPLAIN ANALYZE VERBOSE SELECT
+    rack
+FROM
+    cpu_single
+WHERE
+    10 <= usage_small;
+
++-+-+-+
+| stage | node | plan_|
++-+-+-+
+| 0_| 0_|_CooperativeExec REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_|
+| 1_| 0_|_FilterExec: usage_small@1 >= 10, projection=[rack@0] REDACTED
+|_|_|_CooperativeExec REDACTED
+|_|_|_UnorderedScan: region=REDACTED, {"partition_count":REDACTED, "projection": ["rack", "usage_small"], "filters": ["usage_small >= Int16(10)"], "flat_format": REDACTED, "REDACTED
+|_|_|_|
+|_|_| Total rows: 3_|
++-+-+-+
+
+drop table cpu_single;
+
+Affected Rows: 0
+
 -- SQLNESS SORT_RESULT 3 1
 select
     count(*)
diff --git a/tests/cases/standalone/optimizer/time_index_filter_pushdown.sql b/tests/cases/standalone/optimizer/time_index_filter_pushdown.sql
index 28f4180a2c..c0636e26a4 100644
--- a/tests/cases/standalone/optimizer/time_index_filter_pushdown.sql
+++ b/tests/cases/standalone/optimizer/time_index_filter_pushdown.sql
@@ -4,6 +4,7 @@ CREATE TABLE
     IF NOT EXISTS `cpu` (
         `rack` STRING NULL,
         `os` STRING NULL,
+        `usage_small` SMALLINT NULL,
         `usage_user` BIGINT NULL,
         `greptime_timestamp` TIMESTAMP(9) NOT NULL,
         TIME INDEX (`greptime_timestamp`),
@@ -23,22 +24,71 @@ WITH
 INSERT INTO
     cpu
 VALUES
-    ("1", "linux", 10, "2023-06-12 01:04:49"),
-    ("1", "linux", 15, "2023-06-12 01:04:50"),
-    ("3", "windows", 25, "2023-06-12 01:05:00"),
-    ("5", "mac", 30, "2023-06-12 01:03:00"),
-    ("7", "linux", 45, "2023-06-12 02:00:00");
+    ("1", "linux", 10, 10, "2023-06-12 01:04:49"),
+    ("1", "linux", 15, 15, "2023-06-12 01:04:50"),
+    ("3", "windows", 25, 25, "2023-06-12 01:05:00"),
+    ("5", "mac", 30, 30, "2023-06-12 01:03:00"),
+    ("7", "linux", 45, 45, "2023-06-12 02:00:00");
 
 ADMIN FLUSH_TABLE ('cpu');
 
 INSERT INTO
     cpu
 VALUES
-    ("2", "linux", 20, "2023-06-12 01:04:51"),
-    ("2", "windows", 22, "2023-06-12 01:06:00"),
-    ("4", "mac", 12, "2023-06-12 00:59:00"),
-    ("6", "linux", 35, "2023-06-12 01:04:55"),
-    ("8", "windows", 50, "2023-06-12 02:10:00");
+    ("2", "linux", 20, 20, "2023-06-12 01:04:51"),
+    ("2", "windows", 22, 22, "2023-06-12 01:06:00"),
+    ("4", "mac", 12, 12, "2023-06-12 00:59:00"),
+    ("6", "linux", 35, 35, "2023-06-12 01:04:55"),
+    ("8", "windows", 50, 50, "2023-06-12 02:10:00");
+
+-- SQLNESS REPLACE (peers.*) REDACTED
+EXPLAIN SELECT
+    rack
+FROM
+    cpu
+WHERE
+    usage_small IN (10, 20);
+
+-- SQLNESS REPLACE (peers.*) REDACTED
+EXPLAIN SELECT
+    rack
+FROM
+    cpu
+WHERE
+    usage_small BETWEEN 10 AND 20;
+
+CREATE TABLE
+    IF NOT EXISTS `cpu_single` (
+        `rack` STRING NULL,
+        `usage_small` SMALLINT NULL,
+        `greptime_timestamp` TIMESTAMP(9) NOT NULL,
+        TIME INDEX (`greptime_timestamp`),
+    ) ENGINE = mito
+WITH
+    (append_mode = 'true', sst_format = 'flat');
+
+INSERT INTO
+    cpu_single
+VALUES
+    ("1", 10, "2023-06-12 01:04:49"),
+    ("2", 20, "2023-06-12 01:04:50"),
+    ("3", 25, "2023-06-12 01:05:00");
+
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE "partition_count":\{(.*?)\} "partition_count":REDACTED
+-- SQLNESS REPLACE "flat_format":\s\w+, "flat_format": REDACTED,
+EXPLAIN ANALYZE VERBOSE SELECT
+    rack
+FROM
+    cpu_single
+WHERE
+    10 <= usage_small;
+
+drop table cpu_single;
 
 -- SQLNESS SORT_RESULT 3 1
 select
@@ -70,4 +120,4 @@ where
 group by
     os;
 
-drop table cpu;
\ No newline at end of file
+drop table cpu;
diff --git a/tests/cases/standalone/optimizer/windowed_sort_advance.result b/tests/cases/standalone/optimizer/windowed_sort_advance.result
new file mode 100644
index 0000000000..c0be1bece9
--- /dev/null
+++ b/tests/cases/standalone/optimizer/windowed_sort_advance.result
@@ -0,0 +1,117 @@
+create table `a` (`value` double, `status` bigint, ts timestamp(9) time index);
+
+Affected Rows: 0
+
+INSERT INTO `a` (`value`, `status`, `ts`) VALUES
+(46.82, 200, '2026-03-12T08:00:05.000000000+08:00'),
+(46.84, 200, '2026-03-12T08:00:15.000000000+08:00'),
+(46.85, 200, '2026-03-12T08:00:25.000000000+08:00'),
+(46.86, 200, '2026-03-12T08:00:35.000000000+08:00'),
+(46.88, 200, '2026-03-12T08:00:45.000000000+08:00'),
+(46.89, 200, '2026-03-12T08:00:55.000000000+08:00'),
+(46.91, 200, '2026-03-12T08:01:05.000000000+08:00'),
+(46.90, 200, '2026-03-12T08:01:15.000000000+08:00'),
+(46.87, 200, '2026-03-12T08:01:25.000000000+08:00'),
+(46.85, 200, '2026-03-12T08:01:35.000000000+08:00');
+
+Affected Rows: 10
+
+select ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
++---------------------+--------+-------+
+| ts                  | status | value |
++---------------------+--------+-------+
+| 2026-03-12T00:00:05 | 200    | 46.82 |
+| 2026-03-12T00:00:15 | 200    | 46.84 |
+| 2026-03-12T00:00:25 | 200    | 46.85 |
+| 2026-03-12T00:00:35 | 200    | 46.86 |
+| 2026-03-12T00:00:45 | 200    | 46.88 |
+| 2026-03-12T00:00:55 | 200    | 46.89 |
+| 2026-03-12T00:01:05 | 200    | 46.91 |
+| 2026-03-12T00:01:15 | 200    | 46.9  |
+| 2026-03-12T00:01:25 | 200    | 46.87 |
+| 2026-03-12T00:01:35 | 200    | 46.85 |
++---------------------+--------+-------+
+
+select ts as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
++---------------------+--------+-------+
+| ts                  | status | value |
++---------------------+--------+-------+
+| 2026-03-12T00:00:05 | 200    | 46.82 |
+| 2026-03-12T00:00:15 | 200    | 46.84 |
+| 2026-03-12T00:00:25 | 200    | 46.85 |
+| 2026-03-12T00:00:35 | 200    | 46.86 |
+| 2026-03-12T00:00:45 | 200    | 46.88 |
+| 2026-03-12T00:00:55 | 200    | 46.89 |
+| 2026-03-12T00:01:05 | 200    | 46.91 |
+| 2026-03-12T00:01:15 | 200    | 46.9  |
+| 2026-03-12T00:01:25 | 200    | 46.87 |
+| 2026-03-12T00:01:35 | 200    | 46.85 |
++---------------------+--------+-------+
+
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE num_ranges=\d+ num_ranges=REDACTED
+EXPLAIN ANALYZE select ts as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
++-+-+-+
+| stage | node | plan_|
++-+-+-+
+| 0_| 0_|_CooperativeExec REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_|
+| 1_| 0_|_SortPreservingMergeExec: [ts@0 ASC NULLS LAST] REDACTED
+|_|_|_WindowedSortExec: expr=ts@0 ASC NULLS LAST num_ranges=REDACTED REDACTED
+|_|_|_ProjectionExec: expr=[ts@2 as ts, status@1 as status, value@0 as value] REDACTED
+|_|_|_SeqScan: region=REDACTED, "partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0} REDACTED
+|_|_|_|
+|_|_| Total rows: 10_|
++-+-+-+
+
+select to_timestamp_millis(ts) as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
++---------------------+--------+-------+
+| ts                  | status | value |
++---------------------+--------+-------+
+| 2026-03-12T00:00:05 | 200    | 46.82 |
+| 2026-03-12T00:00:15 | 200    | 46.84 |
+| 2026-03-12T00:00:25 | 200    | 46.85 |
+| 2026-03-12T00:00:35 | 200    | 46.86 |
+| 2026-03-12T00:00:45 | 200    | 46.88 |
+| 2026-03-12T00:00:55 | 200    | 46.89 |
+| 2026-03-12T00:01:05 | 200    | 46.91 |
+| 2026-03-12T00:01:15 | 200    | 46.9  |
+| 2026-03-12T00:01:25 | 200    | 46.87 |
+| 2026-03-12T00:01:35 | 200    | 46.85 |
++---------------------+--------+-------+
+
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE num_ranges=\d+ num_ranges=REDACTED
+EXPLAIN ANALYZE select to_timestamp_millis(ts) as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
++-+-+-+
+| stage | node | plan_|
++-+-+-+
+| 0_| 0_|_CooperativeExec REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_|
+| 1_| 0_|_SortPreservingMergeExec: [ts@0 ASC NULLS LAST] REDACTED
+|_|_|_WindowedSortExec: expr=ts@0 ASC NULLS LAST num_ranges=REDACTED REDACTED
+|_|_|_ProjectionExec: expr=[to_timestamp_millis(ts@2) as ts, status@1 as status, value@0 as value] REDACTED
+|_|_|_SeqScan: region=REDACTED, "partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0} REDACTED
+|_|_|_|
+|_|_| Total rows: 10_|
++-+-+-+
+
+DROP TABLE `a`;
+
+Affected Rows: 0
+
diff --git a/tests/cases/standalone/optimizer/windowed_sort_advance.sql b/tests/cases/standalone/optimizer/windowed_sort_advance.sql
new file mode 100644
index 0000000000..fa73b97933
--- /dev/null
+++ b/tests/cases/standalone/optimizer/windowed_sort_advance.sql
@@ -0,0 +1,37 @@
+create table `a` (`value` double, `status` bigint, ts timestamp(9) time index);
+
+INSERT INTO `a` (`value`, `status`, `ts`) VALUES
+(46.82, 200, '2026-03-12T08:00:05.000000000+08:00'),
+(46.84, 200, '2026-03-12T08:00:15.000000000+08:00'),
+(46.85, 200, '2026-03-12T08:00:25.000000000+08:00'),
+(46.86, 200, '2026-03-12T08:00:35.000000000+08:00'),
+(46.88, 200, '2026-03-12T08:00:45.000000000+08:00'),
+(46.89, 200, '2026-03-12T08:00:55.000000000+08:00'),
+(46.91, 200, '2026-03-12T08:01:05.000000000+08:00'),
+(46.90, 200, '2026-03-12T08:01:15.000000000+08:00'),
+(46.87, 200, '2026-03-12T08:01:25.000000000+08:00'),
+(46.85, 200, '2026-03-12T08:01:35.000000000+08:00');
+
+select ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
+select ts as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE num_ranges=\d+ num_ranges=REDACTED
+EXPLAIN ANALYZE select ts as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
+select to_timestamp_millis(ts) as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE num_ranges=\d+ num_ranges=REDACTED
+EXPLAIN ANALYZE select to_timestamp_millis(ts) as ts, status, value from `a` where ts >= '2026-03-12T08:00:00+08:00' and ts < '2026-03-12T08:02:01+08:00' order by ts asc;
+
+DROP TABLE `a`;
diff --git a/tests/cases/standalone/optimizer/windowed_sort_nanos.result b/tests/cases/standalone/optimizer/windowed_sort_nanos.result
new file mode 100644
index 0000000000..f673636307
--- /dev/null
+++ b/tests/cases/standalone/optimizer/windowed_sort_nanos.result
@@ -0,0 +1,78 @@
+create table `a_ms` (`value` double, `status` bigint, ts timestamp(3) time index);
+
+Affected Rows: 0
+
+INSERT INTO `a_ms` (`value`, `status`, `ts`) VALUES
+(46.82, 200, '2026-03-12T08:00:05.123+08:00'),
+(46.84, 200, '2026-03-12T08:00:15.234+08:00'),
+(46.85, 200, '2026-03-12T08:00:25.345+08:00'),
+(46.86, 200, '2026-03-12T08:00:35.456+08:00'),
+(46.88, 200, '2026-03-12T08:00:45.567+08:00'),
+(46.89, 200, '2026-03-12T08:00:55.678+08:00'),
+(46.91, 200, '2026-03-12T08:01:05.789+08:00'),
+(46.90, 200, '2026-03-12T08:01:15.890+08:00'),
+(46.87, 200, '2026-03-12T08:01:25.901+08:00'),
+(46.85, 200, '2026-03-12T08:01:35.999+08:00');
+
+Affected Rows: 10
+
+select ts, status, value from `a_ms` where ts >= '2026-03-12T08:00:00.000+08:00' and ts < '2026-03-12T08:02:01.000+08:00' order by ts asc;
+
++-------------------------+--------+-------+
+| ts                      | status | value |
++-------------------------+--------+-------+
+| 2026-03-12T00:00:05.123 | 200    | 46.82 |
+| 2026-03-12T00:00:15.234 | 200    | 46.84 |
+| 2026-03-12T00:00:25.345 | 200    | 46.85 |
+| 2026-03-12T00:00:35.456 | 200    | 46.86 |
+| 2026-03-12T00:00:45.567 | 200    | 46.88 |
+| 2026-03-12T00:00:55.678 | 200    | 46.89 |
+| 2026-03-12T00:01:05.789 | 200    | 46.91 |
+| 2026-03-12T00:01:15.890 | 200    | 46.9  |
+| 2026-03-12T00:01:25.901 | 200    | 46.87 |
+| 2026-03-12T00:01:35.999 | 200    | 46.85 |
++-------------------------+--------+-------+
+
+select to_timestamp_nanos(ts) as ts, status, value from `a_ms` where ts >= '2026-03-12T08:00:00.000+08:00' and ts < '2026-03-12T08:02:01.000+08:00' order by ts asc;
+
++-------------------------+--------+-------+
+| ts                      | status | value |
++-------------------------+--------+-------+
+| 2026-03-12T00:00:05.123 | 200    | 46.82 |
+| 2026-03-12T00:00:15.234 | 200    | 46.84 |
+| 2026-03-12T00:00:25.345 | 200    | 46.85 |
+| 2026-03-12T00:00:35.456 | 200    | 46.86 |
+| 2026-03-12T00:00:45.567 | 200    | 46.88 |
+| 2026-03-12T00:00:55.678 | 200    | 46.89 |
+| 2026-03-12T00:01:05.789 | 200    | 46.91 |
+| 2026-03-12T00:01:15.890 | 200    | 46.9  |
+| 2026-03-12T00:01:25.901 | 200    | 46.87 |
+| 2026-03-12T00:01:35.999 | 200    | 46.85 |
++-------------------------+--------+-------+
+
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE num_ranges=\d+ num_ranges=REDACTED
+EXPLAIN ANALYZE select to_timestamp_nanos(ts) as ts, status, value from `a_ms` where ts >= '2026-03-12T08:00:00.000+08:00' and ts < '2026-03-12T08:02:01.000+08:00' order by ts asc;
+
++-+-+-+
+| stage | node | plan_|
++-+-+-+
+| 0_| 0_|_CooperativeExec REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_|
+| 1_| 0_|_SortPreservingMergeExec: [ts@0 ASC NULLS LAST] REDACTED
+|_|_|_WindowedSortExec: expr=ts@0 ASC NULLS LAST num_ranges=REDACTED REDACTED
+|_|_|_ProjectionExec: expr=[to_timestamp_nanos(ts@2) as ts, status@1 as status, value@0 as value] REDACTED
+|_|_|_SeqScan: region=REDACTED, "partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0} REDACTED
+|_|_|_|
+|_|_| Total rows: 10_|
++-+-+-+
+
+DROP TABLE `a_ms`;
+
+Affected Rows: 0
+
diff --git a/tests/cases/standalone/optimizer/windowed_sort_nanos.sql b/tests/cases/standalone/optimizer/windowed_sort_nanos.sql
new file mode 100644
index 0000000000..6a6cf51c4b
--- /dev/null
+++ b/tests/cases/standalone/optimizer/windowed_sort_nanos.sql
@@ -0,0 +1,27 @@
+create table `a_ms` (`value` double, `status` bigint, ts timestamp(3) time index);
+
+INSERT INTO `a_ms` (`value`, `status`, `ts`) VALUES
+(46.82, 200, '2026-03-12T08:00:05.123+08:00'),
+(46.84, 200, '2026-03-12T08:00:15.234+08:00'),
+(46.85, 200, '2026-03-12T08:00:25.345+08:00'),
+(46.86, 200, '2026-03-12T08:00:35.456+08:00'),
+(46.88, 200, '2026-03-12T08:00:45.567+08:00'),
+(46.89, 200, '2026-03-12T08:00:55.678+08:00'),
+(46.91, 200, '2026-03-12T08:01:05.789+08:00'),
+(46.90, 200, '2026-03-12T08:01:15.890+08:00'),
+(46.87, 200, '2026-03-12T08:01:25.901+08:00'),
+(46.85, 200, '2026-03-12T08:01:35.999+08:00');
+
+select ts, status, value from `a_ms` where ts >= '2026-03-12T08:00:00.000+08:00' and ts < '2026-03-12T08:02:01.000+08:00' order by ts asc;
+
+select to_timestamp_nanos(ts) as ts, status, value from `a_ms` where ts >= '2026-03-12T08:00:00.000+08:00' and ts < '2026-03-12T08:02:01.000+08:00' order by ts asc;
+
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE num_ranges=\d+ num_ranges=REDACTED
+EXPLAIN ANALYZE select to_timestamp_nanos(ts) as ts, status, value from `a_ms` where ts >= '2026-03-12T08:00:00.000+08:00' and ts < '2026-03-12T08:02:01.000+08:00' order by ts asc;
+
+DROP TABLE `a_ms`;
diff --git a/tests/cases/standalone/tql-explain-analyze/tsid_column.result b/tests/cases/standalone/tql-explain-analyze/tsid_column.result
index 84544b1655..4a7a875060 100644
--- a/tests/cases/standalone/tql-explain-analyze/tsid_column.result
+++ b/tests/cases/standalone/tql-explain-analyze/tsid_column.result
@@ -112,10 +112,63 @@ TQL ANALYZE (0, 10, '5s')  sum(irate(tsid_metric[1h])) / scalar(count(count(tsid
 |_|_|_AggregateExec: mode=FinalPartitioned, gby=[ts@0 as ts], aggr=[count(count(tsid_metric.val))] REDACTED
 |_|_|_RepartitionExec: partitioning=REDACTED
 |_|_|_AggregateExec: mode=Partial, gby=[ts@0 as ts], aggr=[count(count(tsid_metric.val))] REDACTED
-|_|_|_ProjectionExec: expr=[ts@1 as ts, count(tsid_metric.val)@2 as count(tsid_metric.val)] REDACTED
-|_|_|_AggregateExec: mode=FinalPartitioned, gby=[job@0 as job, ts@1 as ts], aggr=[count(tsid_metric.val)] REDACTED
+|_|_|_ProjectionExec: expr=[ts@0 as ts] REDACTED
+|_|_|_AggregateExec: mode=FinalPartitioned, gby=[ts@0 as ts, job@1 as job], aggr=[] REDACTED
 |_|_|_RepartitionExec: partitioning=REDACTED
-|_|_|_AggregateExec: mode=Partial, gby=[job@1 as job, ts@2 as ts], aggr=[count(tsid_metric.val)] REDACTED
+|_|_|_AggregateExec: mode=Partial, gby=[ts@0 as ts, job@1 as job], aggr=[] REDACTED
+|_|_|_ProjectionExec: expr=[ts@3 as ts, job@1 as job] REDACTED
+|_|_|_PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[ts] REDACTED
+|_|_|_PromSeriesDivideExec: tags=["__tsid"] REDACTED
+|_|_|_ProjectionExec: expr=[val@1 as val, job@3 as job, __tsid@2 as __tsid, ts@0 as ts] REDACTED
+|_|_|_SeriesScan: region=REDACTED, "partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0}, "distribution":"PerSeries" REDACTED
+|_|_|_|
+| 1_| 0_|_SortPreservingMergeExec: [ts@0 ASC NULLS LAST] REDACTED
+|_|_|_SortExec: expr=[ts@0 ASC NULLS LAST], preserve_partitioning=[true] REDACTED
+|_|_|_AggregateExec: mode=FinalPartitioned, gby=[ts@0 as ts], aggr=[sum(prom_irate(ts_range,val))] REDACTED
+|_|_|_RepartitionExec: partitioning=REDACTED
+|_|_|_AggregateExec: mode=Partial, gby=[ts@0 as ts], aggr=[sum(prom_irate(ts_range,val))] REDACTED
+|_|_|_FilterExec: prom_irate(ts_range,val)@1 IS NOT NULL REDACTED
+|_|_|_ProjectionExec: expr=[ts@2 as ts, prom_irate(ts_range@3, val@0) as prom_irate(ts_range,val)] REDACTED
+|_|_|_PromRangeManipulateExec: req range=[0..10000], interval=[5000], eval range=[3600000], time index=[ts] REDACTED
+|_|_|_PromSeriesNormalizeExec: offset=[0], time index=[ts], filter NaN: [true] REDACTED
+|_|_|_PromSeriesDivideExec: tags=["__tsid"] REDACTED
+|_|_|_ProjectionExec: expr=[val@1 as val, __tsid@2 as __tsid, ts@0 as ts] REDACTED
+|_|_|_SeriesScan: region=REDACTED, "partition_count":{"count":1, "mem_ranges":1, "files":0, "file_ranges":0}, "distribution":"PerSeries" REDACTED
+|_|_|_|
+|_|_| Total rows: 2_|
++-+-+-+
+
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE (Hash.*) REDACTED
+TQL ANALYZE (0, 10, '5s')  sum(irate(tsid_metric[1h])) / scalar(count(sum(tsid_metric) by (job)));
+
++-+-+-+
+| stage | node | plan_|
++-+-+-+
+| 0_| 0_|_ProjectionExec: expr=[ts@1 as ts, sum(prom_irate(ts_range,val))@2 / scalar(count(sum(tsid_metric.val)))@0 as lhs.sum(prom_irate(ts_range,val)) / rhs.scalar(count(sum(tsid_metric.val)))] REDACTED
+|_|_|_REDACTED
+|_|_|_ScalarCalculateExec: tags=[] REDACTED
+|_|_|_CoalescePartitionsExec REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_CooperativeExec REDACTED
+|_|_|_MergeScanExec: REDACTED
+|_|_|_|
+| 1_| 0_|_SortPreservingMergeExec: [ts@0 ASC NULLS LAST] REDACTED
+|_|_|_SortExec: expr=[ts@0 ASC NULLS LAST], preserve_partitioning=[true] REDACTED
+|_|_|_AggregateExec: mode=FinalPartitioned, gby=[ts@0 as ts], aggr=[count(sum(tsid_metric.val))] REDACTED
+|_|_|_RepartitionExec: partitioning=REDACTED
+|_|_|_AggregateExec: mode=Partial, gby=[ts@0 as ts], aggr=[count(sum(tsid_metric.val))] REDACTED
+|_|_|_ProjectionExec: expr=[ts@0 as ts] REDACTED
+|_|_|_AggregateExec: mode=FinalPartitioned, gby=[ts@0 as ts, job@1 as job], aggr=[] REDACTED
+|_|_|_RepartitionExec: partitioning=REDACTED
+|_|_|_AggregateExec: mode=Partial, gby=[ts@0 as ts, job@1 as job], aggr=[] REDACTED
+|_|_|_ProjectionExec: expr=[ts@1 as ts, job@0 as job] REDACTED
+|_|_|_FilterExec: val@0 IS NOT NULL, projection=[job@1, ts@2] REDACTED
 |_|_|_ProjectionExec: expr=[val@0 as val, job@1 as job, ts@3 as ts] REDACTED
 |_|_|_PromInstantManipulateExec: range=[0..10000], lookback=[300000], interval=[5000], time index=[ts] REDACTED
 |_|_|_PromSeriesDivideExec: tags=["__tsid"] REDACTED
diff --git a/tests/cases/standalone/tql-explain-analyze/tsid_column.sql b/tests/cases/standalone/tql-explain-analyze/tsid_column.sql
index 7b3de23f33..dedce2dfb1 100644
--- a/tests/cases/standalone/tql-explain-analyze/tsid_column.sql
+++ b/tests/cases/standalone/tql-explain-analyze/tsid_column.sql
@@ -51,6 +51,14 @@ TQL ANALYZE (0, 10, '5s') sum by (job, instance) (tsid_metric);
 -- SQLNESS REPLACE (Hash.*) REDACTED
 TQL ANALYZE (0, 10, '5s')  sum(irate(tsid_metric[1h])) / scalar(count(count(tsid_metric) by (job)));
 
+-- SQLNESS REPLACE (metrics.*) REDACTED
+-- SQLNESS REPLACE (RoundRobinBatch.*) REDACTED
+-- SQLNESS REPLACE (-+) -
+-- SQLNESS REPLACE (\s\s+) _
+-- SQLNESS REPLACE (peers.*) REDACTED
+-- SQLNESS REPLACE region=\d+\(\d+,\s+\d+\) region=REDACTED
+-- SQLNESS REPLACE (Hash.*) REDACTED
+TQL ANALYZE (0, 10, '5s')  sum(irate(tsid_metric[1h])) / scalar(count(sum(tsid_metric) by (job)));
+
 DROP TABLE tsid_metric;
 DROP TABLE tsid_physical;
-
diff --git a/tests/conf/datanode-test.toml.template b/tests/conf/datanode-test.toml.template
index 4cb0423c72..e68a76cc9a 100644
--- a/tests/conf/datanode-test.toml.template
+++ b/tests/conf/datanode-test.toml.template
@@ -6,7 +6,7 @@ rpc_runtime_size = 8
 [[region_engine]]
 [region_engine.mito]
 {{ if enable_flat_format }}
-default_experimental_flat_format = true
+default_flat_format = true
 {{ endif }}
 
 [wal]
@@ -28,7 +28,7 @@ type = 'File'
 data_home = '{data_home}'
 
 [meta_client_options]
-metasrv_addrs = ['{metasrv_addr}']
+metasrv_addrs = ['{addrs.metasrv_addr}']
 timeout_millis = 3000
 connect_timeout_millis = 5000
 tcp_nodelay = false
diff --git a/tests/conf/frontend-test.toml.template b/tests/conf/frontend-test.toml.template
index de4ce86adc..25d44ff6e4 100644
--- a/tests/conf/frontend-test.toml.template
+++ b/tests/conf/frontend-test.toml.template
@@ -1,3 +1,3 @@
 [grpc]
-bind_addr = "{grpc_addr}"
-server_addr = "{grpc_addr}"
+bind_addr = "{addrs.grpc_addr}"
+server_addr = "{addrs.grpc_addr}"
diff --git a/tests/conf/standalone-test.toml.template b/tests/conf/standalone-test.toml.template
index 509eac7ca6..bcd263d0b5 100644
--- a/tests/conf/standalone-test.toml.template
+++ b/tests/conf/standalone-test.toml.template
@@ -5,7 +5,7 @@ require_lease_before_startup = true
 [[region_engine]]
 [region_engine.mito]
 {{ if enable_flat_format }}
-default_experimental_flat_format = true
+default_flat_format = true
 {{ endif }}
 
 [wal]
@@ -26,12 +26,12 @@ type = 'File'
 data_home = '{data_home}'
 
 [grpc]
-bind_addr = '{grpc_addr}'
+bind_addr = '{addrs.grpc_addr}'
 runtime_size = 8
 
 [mysql]
 enable = true
-addr = "{mysql_addr}"
+addr = "{addrs.mysql_addr}"
 runtime_size = 2
 prepared_stmt_cache_size= 10000
 
@@ -40,7 +40,7 @@ mode = "disable"
 
 [postgres]
 enable = true
-addr = "{postgres_addr}"
+addr = "{addrs.postgres_addr}"
 runtime_size = 2
 
 [procedure]
diff --git a/tests/runner/Cargo.toml b/tests/runner/Cargo.toml
index db1d1402aa..6b41e5a61e 100644
--- a/tests/runner/Cargo.toml
+++ b/tests/runner/Cargo.toml
@@ -19,9 +19,9 @@ datatypes = { workspace = true }
 flate2 = "1.0"
 hex = "0.4"
 local-ip-address = "0.6"
-mysql = { version = "26", default-features = false, features = ["minimal", "rustls-tls-ring"] }
+mysql = { version = "26", default-features = false, features = ["minimal", "rustls-tls"] }
 num_cpus = "1.16"
-reqwest = { version = "0.12", default-features = false, features = ["rustls-tls"] }
+reqwest = { version = "0.12", default-features = false, features = ["rustls-tls", "stream"] }
 serde.workspace = true
 serde_json.workspace = true
 sha2 = "0.10"
diff --git a/tests/runner/src/cmd/bare.rs b/tests/runner/src/cmd/bare.rs
index e9a4ff8b79..58199f959e 100644
--- a/tests/runner/src/cmd/bare.rs
+++ b/tests/runner/src/cmd/bare.rs
@@ -103,7 +103,7 @@ pub struct BareCommand {
     #[clap(long)]
     extra_args: Vec<String>,
 
-    /// Enable flat format for storage engine (sets default_experimental_flat_format = true).
+    /// Enable flat format for storage engine (sets default_flat_format = true).
     #[clap(long, default_value = "false")]
     enable_flat_format: bool,
 }
diff --git a/tests/runner/src/server_mode.rs b/tests/runner/src/server_mode.rs
index 172baf32ff..3bfb7959fd 100644
--- a/tests/runner/src/server_mode.rs
+++ b/tests/runner/src/server_mode.rs
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::collections::HashSet;
+use std::collections::{HashMap, HashSet};
 use std::path::Path;
 use std::sync::{Mutex, OnceLock};
 
@@ -96,15 +96,7 @@ struct ConfigContext {
     use_etcd: bool,
     store_addrs: String,
     instance_id: usize,
-    // for following addrs, leave it empty if not needed
-    // required for datanode
-    metasrv_addr: String,
-    // for frontend and standalone
-    grpc_addr: String,
-    // for standalone
-    mysql_addr: String,
-    // for standalone
-    postgres_addr: String,
+    addrs: HashMap<String, String>,
     // enable flat format for storage engine
     enable_flat_format: bool,
 }
@@ -275,40 +267,26 @@ impl ServerMode {
         let procedure_dir = data_home.join("procedure").display().to_string();
 
         // Get the required addresses based on server mode
-        let (metasrv_addr, grpc_addr, mysql_addr, postgres_addr) = match self {
+        let addrs: HashMap<String, String> = match self {
             ServerMode::Standalone {
                 rpc_bind_addr,
                 mysql_addr,
                 postgres_addr,
-                ..
-            } => (
-                String::new(),
-                rpc_bind_addr.clone(),
-                mysql_addr.clone(),
-                postgres_addr.clone(),
-            ),
-            ServerMode::Frontend {
-                rpc_bind_addr,
-                mysql_addr,
-                postgres_addr,
-                ..
-            } => (
-                String::new(),
-                rpc_bind_addr.clone(),
-                mysql_addr.clone(),
-                postgres_addr.clone(),
-            ),
-            ServerMode::Datanode {
-                rpc_bind_addr,
-                metasrv_addr,
-                ..
-            } => (
-                metasrv_addr.clone(),
-                rpc_bind_addr.clone(),
-                String::new(),
-                String::new(),
-            ),
-            _ => (String::new(), String::new(), String::new(), String::new()),
+                http_addr,
+            } => [
+                ("http_addr".to_string(), http_addr.clone()),
+                ("grpc_addr".to_string(), rpc_bind_addr.clone()),
+                ("mysql_addr".to_string(), mysql_addr.clone()),
+                ("postgres_addr".to_string(), postgres_addr.clone()),
+            ]
+            .into(),
+            ServerMode::Frontend { rpc_bind_addr, .. } => {
+                [("grpc_addr".to_string(), rpc_bind_addr.clone())].into()
+            }
+            ServerMode::Datanode { metasrv_addr, .. } => {
+                [("metasrv_addr".to_string(), metasrv_addr.clone())].into()
+            }
+            _ => HashMap::new(),
         };
 
         let ctx = ConfigContext {
@@ -326,10 +304,7 @@ impl ServerMode {
                 .collect::<Vec<_>>()
                 .join(","),
             instance_id: id,
-            metasrv_addr,
-            grpc_addr,
-            mysql_addr,
-            postgres_addr,
+            addrs,
             enable_flat_format: db_ctx.store_config().enable_flat_format,
         };
 
@@ -380,7 +355,7 @@ impl ServerMode {
                     "-c".to_string(),
                     self.generate_config_file(sqlness_home, db_ctx, id),
                     format!("--http-addr={http_addr}"),
-                    format!("--rpc-addr={rpc_bind_addr}"),
+                    format!("--grpc-bind-addr={rpc_bind_addr}"),
                     format!("--mysql-addr={mysql_addr}"),
                     format!("--postgres-addr={postgres_addr}"),
                 ]);
@@ -395,10 +370,10 @@ impl ServerMode {
                 args.extend([
                     format!("--metasrv-addrs={metasrv_addr}"),
                     format!("--http-addr={http_addr}"),
-                    format!("--rpc-addr={rpc_bind_addr}"),
+                    format!("--grpc-bind-addr={rpc_bind_addr}"),
                     // since sqlness run on local, bind addr is the same as server addr
                     // this is needed so that `cluster_info`'s server addr column can be correct
-                    format!("--rpc-server-addr={rpc_bind_addr}"),
+                    format!("--grpc-server-addr={rpc_bind_addr}"),
                     format!("--mysql-addr={mysql_addr}"),
                     format!("--postgres-addr={postgres_addr}"),
                     format!(
@@ -416,9 +391,9 @@ impl ServerMode {
                 http_addr,
             } => {
                 args.extend([
-                    "--bind-addr".to_string(),
+                    "--grpc-bind-addr".to_string(),
                     rpc_bind_addr.clone(),
-                    "--server-addr".to_string(),
+                    "--grpc-server-addr".to_string(),
                     rpc_server_addr.clone(),
                     "--enable-region-failover".to_string(),
                     "false".to_string(),
@@ -501,8 +476,8 @@ impl ServerMode {
                     db_ctx.time()
                 ));
                 args.extend([
-                    format!("--rpc-addr={rpc_bind_addr}"),
-                    format!("--rpc-server-addr={rpc_server_addr}"),
+                    format!("--grpc-bind-addr={rpc_bind_addr}"),
+                    format!("--grpc-server-addr={rpc_server_addr}"),
                     format!("--http-addr={http_addr}"),
                     format!("--data-home={}", data_home.display()),
                     format!("--log-dir={}/logs", data_home.display()),
@@ -520,8 +495,8 @@ impl ServerMode {
                 node_id,
             } => {
                 args.extend([
-                    format!("--rpc-addr={rpc_bind_addr}"),
-                    format!("--rpc-server-addr={rpc_server_addr}"),
+                    format!("--grpc-bind-addr={rpc_bind_addr}"),
+                    format!("--grpc-server-addr={rpc_server_addr}"),
                     format!("--node-id={node_id}"),
                     format!(
                         "--log-dir={}/greptimedb-{}-flownode/logs",