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" >> $GITHUB_OUTPUT
+  fi
 fi
diff --git a/.github/scripts/upload-artifacts-to-s3.sh b/.github/scripts/upload-artifacts-to-s3.sh
index 1ddf32044b..9c92859e3b 100755
--- a/.github/scripts/upload-artifacts-to-s3.sh
+++ b/.github/scripts/upload-artifacts-to-s3.sh
@@ -38,6 +38,11 @@ function upload_artifacts() {
     curl -X PUT \
       -u "$PROXY_USERNAME:$PROXY_PASSWORD" \
       -F "file=@$file" \
+      --max-time 3600 \
+      --connect-timeout 20 \
+      --retry 5 \
+      --retry-delay 10 \
+      --retry-max-time 3000 \
       "$TARGET_URL"
   done
 }
@@ -54,6 +59,11 @@ function update_version_info() {
       curl -X PUT \
         -u "$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
 
@@ -66,6 +76,11 @@ function update_version_info() {
       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
diff --git a/Cargo.lock b/Cargo.lock
index edb8ce04d4..d66fa2d686 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -212,7 +212,7 @@ checksum = "d301b3b94cb4b2f23d7917810addbbaff90738e0ca2be692bd027e70d7e0330c"
 
 [[package]]
 name = "api"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "arrow-schema 57.3.0",
  "common-base",
@@ -933,7 +933,7 @@ checksum = "1505bd5d3d116872e7271a6d4e16d81d0c8570876c8de68093a09ac269d8aac0"
 
 [[package]]
 name = "auth"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "async-trait",
@@ -1523,7 +1523,7 @@ dependencies = [
 
 [[package]]
 name = "cache"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "catalog",
  "common-error",
@@ -1559,7 +1559,7 @@ checksum = "37b2a672a2cb129a2e41c10b1224bb368f9f37a2b16b612598138befd7b37eb5"
 
 [[package]]
 name = "catalog"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "arrow 57.3.0",
@@ -1894,7 +1894,7 @@ checksum = "b94f61472cee1439c0b966b47e3aca9ae07e45d070759512cd390ea2bebc6675"
 
 [[package]]
 name = "cli"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "async-stream",
  "async-trait",
@@ -1951,7 +1951,7 @@ dependencies = [
 
 [[package]]
 name = "client"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "arc-swap",
@@ -1983,7 +1983,7 @@ dependencies = [
  "serde_json",
  "snafu 0.8.6",
  "store-api",
- "substrait 1.0.0",
+ "substrait 1.0.1",
  "tokio",
  "tokio-stream",
  "tonic 0.14.2",
@@ -2023,7 +2023,7 @@ dependencies = [
 
 [[package]]
 name = "cmd"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "async-trait",
@@ -2155,7 +2155,7 @@ dependencies = [
 
 [[package]]
 name = "common-base"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "ahash 0.8.12",
  "anymap2",
@@ -2175,14 +2175,14 @@ dependencies = [
 
 [[package]]
 name = "common-catalog"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "const_format",
 ]
 
 [[package]]
 name = "common-config"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "common-base",
  "common-error",
@@ -2206,7 +2206,7 @@ dependencies = [
 
 [[package]]
 name = "common-datasource"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "arrow 57.3.0",
  "arrow-schema 57.3.0",
@@ -2242,7 +2242,7 @@ dependencies = [
 
 [[package]]
 name = "common-decimal"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "bigdecimal 0.4.8",
  "common-error",
@@ -2255,7 +2255,7 @@ dependencies = [
 
 [[package]]
 name = "common-error"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "common-macro",
  "http 1.3.1",
@@ -2266,7 +2266,7 @@ dependencies = [
 
 [[package]]
 name = "common-event-recorder"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "async-trait",
@@ -2289,7 +2289,7 @@ dependencies = [
 
 [[package]]
 name = "common-frontend"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "async-trait",
@@ -2310,7 +2310,7 @@ dependencies = [
 
 [[package]]
 name = "common-function"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "ahash 0.8.12",
  "api",
@@ -2348,6 +2348,7 @@ dependencies = [
  "geohash",
  "h3o",
  "hyperloglogplus",
+ "icu_properties",
  "jsonb",
  "jsonpath-rust 0.7.5",
  "memchr",
@@ -2373,7 +2374,7 @@ dependencies = [
 
 [[package]]
 name = "common-greptimedb-telemetry"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "async-trait",
  "common-runtime",
@@ -2390,7 +2391,7 @@ dependencies = [
 
 [[package]]
 name = "common-grpc"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "arrow-flight",
@@ -2425,7 +2426,7 @@ dependencies = [
 
 [[package]]
 name = "common-grpc-expr"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "common-base",
@@ -2445,7 +2446,7 @@ dependencies = [
 
 [[package]]
 name = "common-macro"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "greptime-proto",
  "once_cell",
@@ -2456,7 +2457,7 @@ dependencies = [
 
 [[package]]
 name = "common-mem-prof"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "anyhow",
  "common-error",
@@ -2472,7 +2473,7 @@ dependencies = [
 
 [[package]]
 name = "common-memory-manager"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "common-error",
  "common-macro",
@@ -2484,7 +2485,7 @@ dependencies = [
 
 [[package]]
 name = "common-meta"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "anymap2",
  "api",
@@ -2555,7 +2556,7 @@ dependencies = [
 
 [[package]]
 name = "common-options"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "common-grpc",
  "humantime-serde",
@@ -2565,11 +2566,11 @@ dependencies = [
 
 [[package]]
 name = "common-plugins"
-version = "1.0.0"
+version = "1.0.1"
 
 [[package]]
 name = "common-pprof"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "common-error",
  "common-macro",
@@ -2580,7 +2581,7 @@ dependencies = [
 
 [[package]]
 name = "common-procedure"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "async-stream",
@@ -2609,7 +2610,7 @@ dependencies = [
 
 [[package]]
 name = "common-procedure-test"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "async-trait",
  "common-procedure",
@@ -2619,7 +2620,7 @@ dependencies = [
 
 [[package]]
 name = "common-query"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "async-trait",
@@ -2645,7 +2646,7 @@ dependencies = [
 
 [[package]]
 name = "common-recordbatch"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "arc-swap",
  "common-base",
@@ -2670,7 +2671,7 @@ dependencies = [
 
 [[package]]
 name = "common-runtime"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "async-trait",
  "clap",
@@ -2699,7 +2700,7 @@ dependencies = [
 
 [[package]]
 name = "common-session"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "serde",
  "strum 0.27.1",
@@ -2707,7 +2708,7 @@ dependencies = [
 
 [[package]]
 name = "common-sql"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "arrow-schema 57.3.0",
  "common-base",
@@ -2727,7 +2728,7 @@ dependencies = [
 
 [[package]]
 name = "common-stat"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "common-base",
  "common-runtime",
@@ -2742,7 +2743,7 @@ dependencies = [
 
 [[package]]
 name = "common-telemetry"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "backtrace",
  "common-base",
@@ -2771,7 +2772,7 @@ dependencies = [
 
 [[package]]
 name = "common-test-util"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "client",
  "common-grpc",
@@ -2784,7 +2785,7 @@ dependencies = [
 
 [[package]]
 name = "common-time"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "arrow 57.3.0",
  "chrono",
@@ -2802,7 +2803,7 @@ dependencies = [
 
 [[package]]
 name = "common-version"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "cargo-manifest",
  "const_format",
@@ -2812,7 +2813,7 @@ dependencies = [
 
 [[package]]
 name = "common-wal"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "common-base",
  "common-error",
@@ -2835,7 +2836,7 @@ dependencies = [
 
 [[package]]
 name = "common-workload"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "common-telemetry",
  "serde",
@@ -4197,7 +4198,7 @@ dependencies = [
 
 [[package]]
 name = "datanode"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "arrow-flight",
@@ -4265,7 +4266,7 @@ dependencies = [
 
 [[package]]
 name = "datatypes"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "arrow 57.3.0",
  "arrow-array 57.3.0",
@@ -4943,7 +4944,7 @@ checksum = "37909eebbb50d72f9059c3b6d82c0463f2ff062c9e95845c43a6c9c0355411be"
 
 [[package]]
 name = "file-engine"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "async-trait",
@@ -5075,7 +5076,7 @@ checksum = "8bf7cc16383c4b8d58b9905a8509f02926ce3058053c056376248d958c9df1e8"
 
 [[package]]
 name = "flow"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "arrow 57.3.0",
@@ -5144,7 +5145,7 @@ dependencies = [
  "sql",
  "store-api",
  "strum 0.27.1",
- "substrait 1.0.0",
+ "substrait 1.0.1",
  "table",
  "tokio",
  "tonic 0.14.2",
@@ -5205,7 +5206,7 @@ checksum = "28dd6caf6059519a65843af8fe2a3ae298b14b80179855aeb4adc2c1934ee619"
 
 [[package]]
 name = "frontend"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "arc-swap",
@@ -5681,7 +5682,7 @@ 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=26a50f4069f50c37d65b45e0d39ae0cb42de5425#26a50f4069f50c37d65b45e0d39ae0cb42de5425"
 dependencies = [
  "prost 0.14.1",
  "prost-types 0.14.1",
@@ -5691,7 +5692,6 @@ dependencies = [
  "strum_macros 0.25.3",
  "tonic 0.14.2",
  "tonic-prost",
- "tonic-prost-build",
 ]
 
 [[package]]
@@ -6453,7 +6453,7 @@ dependencies = [
 
 [[package]]
 name = "index"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "async-trait",
  "asynchronous-codec",
@@ -7421,7 +7421,7 @@ checksum = "5e5032e24019045c762d3c0f28f5b6b8bbf38563a65908389bf7978758920897"
 
 [[package]]
 name = "log-query"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "chrono",
  "common-error",
@@ -7433,7 +7433,7 @@ dependencies = [
 
 [[package]]
 name = "log-store"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "async-stream",
  "async-trait",
@@ -7724,7 +7724,7 @@ dependencies = [
 
 [[package]]
 name = "meta-client"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "async-trait",
@@ -7755,7 +7755,7 @@ dependencies = [
 
 [[package]]
 name = "meta-srv"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "async-trait",
@@ -7855,7 +7855,7 @@ dependencies = [
 
 [[package]]
 name = "metric-engine"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "aquamarine",
@@ -7956,7 +7956,7 @@ dependencies = [
 
 [[package]]
 name = "mito-codec"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "bytes",
@@ -7981,7 +7981,7 @@ dependencies = [
 
 [[package]]
 name = "mito2"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "aquamarine",
@@ -8705,7 +8705,7 @@ dependencies = [
 
 [[package]]
 name = "object-store"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "anyhow",
  "bytes",
@@ -8883,7 +8883,7 @@ 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",
@@ -9032,7 +9032,7 @@ dependencies = [
 
 [[package]]
 name = "operator"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "ahash 0.8.12",
  "api",
@@ -9092,7 +9092,7 @@ dependencies = [
  "sql",
  "sqlparser",
  "store-api",
- "substrait 1.0.0",
+ "substrait 1.0.1",
  "table",
  "tokio",
  "tokio-util",
@@ -9368,7 +9368,7 @@ checksum = "e3c406c9e2aa74554e662d2c2ee11cd3e73756988800be7e6f5eddb16fed4699"
 
 [[package]]
 name = "partition"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "async-trait",
@@ -9724,7 +9724,7 @@ checksum = "8b870d8c151b6f2fb93e84a13146138f05d02ed11c7e7c54f8826aaaf7c9f184"
 
 [[package]]
 name = "pipeline"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "ahash 0.8.12",
  "api",
@@ -9881,7 +9881,7 @@ dependencies = [
 
 [[package]]
 name = "plugins"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "auth",
  "catalog",
@@ -10199,7 +10199,7 @@ dependencies = [
 
 [[package]]
 name = "promql"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "ahash 0.8.12",
  "async-trait",
@@ -10551,7 +10551,7 @@ dependencies = [
 
 [[package]]
 name = "puffin"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "async-compression",
  "async-trait",
@@ -10613,7 +10613,7 @@ dependencies = [
 
 [[package]]
 name = "query"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "ahash 0.8.12",
  "api",
@@ -10680,7 +10680,7 @@ dependencies = [
  "sql",
  "sqlparser",
  "store-api",
- "substrait 1.0.0",
+ "substrait 1.0.1",
  "table",
  "tokio",
  "tokio-stream",
@@ -11984,7 +11984,7 @@ dependencies = [
 
 [[package]]
 name = "servers"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "ahash 0.8.12",
  "api",
@@ -12034,6 +12034,7 @@ dependencies = [
  "datafusion-pg-catalog",
  "datatypes",
  "derive_builder 0.20.2",
+ "either",
  "futures",
  "futures-util",
  "headers",
@@ -12080,6 +12081,7 @@ dependencies = [
  "regex",
  "reqwest",
  "rust-embed",
+ "rust_decimal",
  "rustls",
  "rustls-pemfile",
  "rustls-pki-types",
@@ -12118,7 +12120,7 @@ dependencies = [
 
 [[package]]
 name = "session"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "ahash 0.8.12",
  "api",
@@ -12450,7 +12452,7 @@ dependencies = [
 
 [[package]]
 name = "sql"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "arrow-buffer 57.3.0",
@@ -12511,7 +12513,7 @@ dependencies = [
 
 [[package]]
 name = "sqlness-runner"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "async-trait",
  "clap",
@@ -12791,7 +12793,7 @@ dependencies = [
 
 [[package]]
 name = "standalone"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "async-trait",
  "catalog",
@@ -12835,7 +12837,7 @@ checksum = "a2eb9349b6444b326872e140eb1cf5e7c522154d69e7a0ffb0fb81c06b37543f"
 
 [[package]]
 name = "store-api"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "aquamarine",
@@ -13027,7 +13029,7 @@ dependencies = [
 
 [[package]]
 name = "substrait"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "async-trait",
  "bytes",
@@ -13149,7 +13151,7 @@ dependencies = [
 
 [[package]]
 name = "table"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "arc-swap",
@@ -13419,7 +13421,7 @@ checksum = "8f50febec83f5ee1df3015341d8bd429f2d1cc62bcba7ea2076759d315084683"
 
 [[package]]
 name = "tests-fuzz"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "arbitrary",
  "async-trait",
@@ -13463,7 +13465,7 @@ dependencies = [
 
 [[package]]
 name = "tests-integration"
-version = "1.0.0"
+version = "1.0.1"
 dependencies = [
  "api",
  "arrow-flight",
@@ -13501,6 +13503,7 @@ dependencies = [
  "datanode",
  "datatypes",
  "dotenv",
+ "either",
  "flate2",
  "flow",
  "frontend",
@@ -13516,7 +13519,6 @@ dependencies = [
  "meta-client",
  "meta-srv",
  "mito2",
- "moka",
  "mysql_async",
  "object-store",
  "opentelemetry-proto 0.31.0",
@@ -13540,7 +13542,7 @@ dependencies = [
  "sqlx",
  "standalone",
  "store-api",
- "substrait 1.0.0",
+ "substrait 1.0.1",
  "table",
  "tempfile",
  "time",
diff --git a/Cargo.toml b/Cargo.toml
index 227608bf64..aa3d3d0c6c 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -75,7 +75,7 @@ members = [
 resolver = "2"
 
 [workspace.package]
-version = "1.0.0"
+version = "1.0.1"
 edition = "2024"
 license = "Apache-2.0"
 
@@ -154,13 +154,14 @@ etcd-client = { version = "0.17", features = [
 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 = "26a50f4069f50c37d65b45e0d39ae0cb42de5425" }
 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"
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/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/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/data/export.rs b/src/cli/src/data/export.rs
index 051c07da35..148c27316e 100644
--- a/src/cli/src/data/export.rs
+++ b/src/cli/src/data/export.rs
@@ -1084,7 +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",
@@ -1092,30 +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",
@@ -1123,24 +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 ====================
@@ -1238,21 +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
@@ -1287,7 +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",
@@ -1295,15 +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;
@@ -1324,10 +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/cmd/Cargo.toml b/src/cmd/Cargo.toml
index 34619f4f1b..2a5a599a91 100644
--- a/src/cmd/Cargo.toml
+++ b/src/cmd/Cargo.toml
@@ -72,7 +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
diff --git a/src/cmd/src/cli.rs b/src/cmd/src/cli.rs
index 84e797c291..95c5f00b77 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,
+            "--rpc-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/standalone.rs b/src/cmd/src/standalone.rs
index 215bea0ec5..a95c1bad1f 100644
--- a/src/cmd/src/standalone.rs
+++ b/src/cmd/src/standalone.rs
@@ -42,6 +42,7 @@ use common_meta::region_keeper::MemoryRegionKeeper;
 use common_meta::region_registry::LeaderRegionRegistry;
 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;
@@ -369,6 +370,7 @@ impl StartCommand {
         creator: InstanceCreator,
     ) -> Result<(Instance, InstanceCreatorResult)> {
         let mut plugins = Plugins::new();
+        plugins.insert(StandaloneFlag);
         set_default_prefix(opts.default_column_prefix.as_deref())
             .map_err(BoxedError::new)
             .context(error::BuildCliSnafu)?;
diff --git a/src/common/function/Cargo.toml b/src/common/function/Cargo.toml
index d164b9285d..43ddf9ae0c 100644
--- a/src/common/function/Cargo.toml
+++ b/src/common/function/Cargo.toml
@@ -47,6 +47,7 @@ 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"
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/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/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/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/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/meta/Cargo.toml b/src/common/meta/Cargo.toml
index 8132a1fc4e..136bc268f0 100644
--- a/src/common/meta/Cargo.toml
+++ b/src/common/meta/Cargo.toml
@@ -59,7 +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
diff --git a/src/common/meta/src/datanode.rs b/src/common/meta/src/datanode.rs
index 8b521d8e43..d6c6229801 100644
--- a/src/common/meta/src/datanode.rs
+++ b/src/common/meta/src/datanode.rs
@@ -573,4 +573,27 @@ 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);
+    }
 }
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 4da83e367f..6b3e149877 100644
--- a/src/common/meta/src/ddl/drop_database/start.rs
+++ b/src/common/meta/src/ddl/drop_database/start.rs
@@ -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/test_util.rs b/src/common/meta/src/ddl/test_util.rs
index 36d422216f..7d8cdd895b 100644
--- a/src/common/meta/src/ddl/test_util.rs
+++ b/src/common/meta/src/ddl/test_util.rs
@@ -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/create_table.rs b/src/common/meta/src/ddl/tests/create_table.rs
index 5355ac8c7c..da91dedcaf 100644
--- a/src/common/meta/src/ddl/tests/create_table.rs
+++ b/src/common/meta/src/ddl/tests/create_table.rs
@@ -13,7 +13,7 @@
 // limitations under the License.
 
 use std::assert_matches;
-use std::collections::HashMap;
+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/drop_table.rs b/src/common/meta/src/ddl/tests/drop_table.rs
index fb2c882da0..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,
@@ -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/common/meta/src/key.rs b/src/common/meta/src/key.rs
index 332c60f225..3a9217e710 100644
--- a/src/common/meta/src/key.rs
+++ b/src/common/meta/src/key.rs
@@ -663,7 +663,7 @@ 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))
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/kv_backend/rds/mysql.rs b/src/common/meta/src/kv_backend/rds/mysql.rs
index ee776239bf..ea8bab4704 100644
--- a/src/common/meta/src/kv_backend/rds/mysql.rs
+++ b/src/common/meta/src/kv_backend/rds/mysql.rs
@@ -15,14 +15,18 @@
 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,
@@ -37,6 +41,18 @@ 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 +177,11 @@ 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 +206,8 @@ 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,
 }
@@ -534,6 +556,68 @@ 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> {
         let pool = MySqlPool::connect(url)
@@ -558,6 +642,7 @@ 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 +659,7 @@ 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 +671,45 @@ 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();
+        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 +726,156 @@ 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!();
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/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/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/local/runner.rs b/src/common/procedure/src/local/runner.rs
index 2a974de889..ca3e221f43 100644
--- a/src/common/procedure/src/local/runner.rs
+++ b/src/common/procedure/src/local/runner.rs
@@ -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/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/heartbeat.rs b/src/datanode/src/heartbeat.rs
index be662dfe94..fe8866b7f9 100644
--- a/src/datanode/src/heartbeat.rs
+++ b/src/datanode/src/heartbeat.rs
@@ -148,9 +148,9 @@ 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,
                         }
                     }
diff --git a/src/datanode/src/region_server.rs b/src/datanode/src/region_server.rs
index ec10691bea..aa3ffbfe3a 100644
--- a/src/datanode/src/region_server.rs
+++ b/src/datanode/src/region_server.rs
@@ -360,6 +360,7 @@ impl RegionServer {
             engine.role(region_id).map(|role| match role {
                 RegionRole::Follower => false,
                 RegionRole::Leader => true,
+                RegionRole::StagingLeader => true,
                 RegionRole::DowngradingLeader => true,
             })
         })
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/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/meta-srv/src/handler/collect_cluster_info_handler.rs b/src/meta-srv/src/handler/collect_cluster_info_handler.rs
index c96229f9cf..3fc785a1cb 100644
--- a/src/meta-srv/src/handler/collect_cluster_info_handler.rs
+++ b/src/meta-srv/src/handler/collect_cluster_info_handler.rs
@@ -129,7 +129,7 @@ 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;
 
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/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/procedure/region_migration/open_candidate_region.rs b/src/meta-srv/src/procedure/region_migration/open_candidate_region.rs
index 189ba89449..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,
@@ -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/update_metadata/upgrade_candidate_region.rs b/src/meta-srv/src/procedure/region_migration/update_metadata/upgrade_candidate_region.rs
index d5aa8699ec..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
@@ -231,6 +231,7 @@ mod tests {
     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/repartition.rs b/src/meta-srv/src/procedure/repartition.rs
index db8bfeadc5..f314a40080 100644
--- a/src/meta-srv/src/procedure/repartition.rs
+++ b/src/meta-srv/src/procedure/repartition.rs
@@ -42,7 +42,7 @@ use common_meta::lock_key::{CatalogLock, SchemaLock, TableLock, TableNameLock};
 use common_meta::node_manager::NodeManagerRef;
 use common_meta::region_keeper::{MemoryRegionKeeperRef, OperatingRegionGuard};
 use common_meta::region_registry::LeaderRegionRegistryRef;
-use common_meta::rpc::router::{RegionRoute, operating_leader_regions};
+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,
@@ -59,11 +59,13 @@ 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)]
@@ -76,11 +78,17 @@ pub struct PersistentContext {
     pub table_name: String,
     pub table_id: TableId,
     pub plans: Vec<RepartitionPlanEntry>,
-    /// Records failed sub-procedures for metadata rollback.
     #[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 metadata rollback.
+    /// 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")]
@@ -409,9 +417,9 @@ impl Context {
         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) {
+        for (region_id, datanode_id, role) in operating_leader_region_roles(region_routes) {
             let guard = memory_region_keeper
-                .register(datanode_id, region_id)
+                .register_with_role(datanode_id, region_id, role)
                 .context(error::RegionOperatingRaceSnafu {
                     peer_id: datanode_id,
                     region_id,
@@ -506,6 +514,23 @@ impl RepartitionProcedure {
             || 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>()
@@ -519,13 +544,9 @@ impl RepartitionProcedure {
                 .collect();
         }
 
-        self.context
-            .persistent_ctx
-            .failed_procedures
-            .iter()
-            .chain(self.context.persistent_ctx.unknown_procedures.iter())
-            .flat_map(|procedure_meta| {
-                let plan_index = procedure_meta.plan_index;
+        self.rollback_plan_indices()
+            .into_iter()
+            .flat_map(|plan_index| {
                 self.context.persistent_ctx.plans[plan_index]
                     .allocated_region_ids
                     .iter()
@@ -534,15 +555,35 @@ impl RepartitionProcedure {
             .collect()
     }
 
-    fn filter_allocated_region_routes(
-        region_routes: &[RegionRoute],
-        allocated_region_ids: &HashSet<store_api::storage::RegionId>,
-    ) -> Vec<RegionRoute> {
-        region_routes
-            .iter()
-            .filter(|route| !allocated_region_ids.contains(&route.region.id))
-            .cloned()
-            .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<()> {
@@ -552,17 +593,17 @@ impl RepartitionProcedure {
 
         let table_id = self.context.persistent_ctx.table_id;
         let allocated_region_ids = self.rollback_allocated_region_ids();
-        if allocated_region_ids.is_empty() {
-            return Ok(());
-        }
 
         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 current_region_routes = table_route_value.region_routes().unwrap();
+        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,
+            &current_region_routes,
             &allocated_region_ids,
         );
         if !allocated_region_routes.is_empty() {
@@ -587,9 +628,9 @@ impl RepartitionProcedure {
         }
 
         let new_region_routes =
-            Self::filter_allocated_region_routes(current_region_routes, &allocated_region_ids);
+            DeallocateRegion::generate_region_routes(&current_region_routes, &allocated_region_ids);
 
-        if new_region_routes.len() != current_region_routes.len() {
+        if new_region_routes != *original_region_routes {
             self.context
                 .update_table_route(&table_route_value, new_region_routes, HashMap::new())
                 .await
@@ -796,9 +837,11 @@ mod tests {
     };
     use common_meta::error;
     use common_meta::peer::Peer;
-    use common_meta::rpc::router::{Region, RegionRoute};
+    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;
@@ -809,6 +852,7 @@ mod tests {
     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::{
@@ -837,9 +881,52 @@ mod tests {
             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 }
     }
@@ -870,10 +957,8 @@ mod tests {
         ];
         let allocated_region_ids = HashSet::from([RegionId::new(table_id, 2)]);
 
-        let new_region_routes = RepartitionProcedure::filter_allocated_region_routes(
-            &region_routes,
-            &allocated_region_ids,
-        );
+        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));
@@ -910,6 +995,59 @@ mod tests {
         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();
@@ -939,7 +1077,16 @@ mod tests {
             table_id,
             None,
         );
-        persistent_ctx.plans = vec![test_plan(table_id)];
+        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(),
@@ -1050,6 +1197,16 @@ mod tests {
             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 {
@@ -1069,6 +1226,13 @@ mod tests {
             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 {
@@ -1100,6 +1264,212 @@ mod tests {
         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();
@@ -1129,7 +1499,16 @@ mod tests {
             table_id,
             None,
         );
-        persistent_ctx.plans = vec![test_plan(table_id)];
+        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(),
@@ -1164,6 +1543,147 @@ mod tests {
         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();
diff --git a/src/meta-srv/src/procedure/repartition/allocate_region.rs b/src/meta-srv/src/procedure/repartition/allocate_region.rs
index 12ffac9918..c1f3ca2503 100644
--- a/src/meta-srv/src/procedure/repartition/allocate_region.rs
+++ b/src/meta-srv/src/procedure/repartition/allocate_region.rs
@@ -25,8 +25,8 @@ use common_meta::rpc::router::RegionRoute;
 use common_procedure::{Context as ProcedureContext, Status};
 use common_telemetry::{debug, info};
 use serde::{Deserialize, Deserializer, Serialize};
-use snafu::ResultExt;
-use store_api::storage::{RegionNumber, TableId};
+use snafu::{OptionExt, ResultExt};
+use store_api::storage::{RegionId, RegionNumber, TableId};
 use table::metadata::TableInfo;
 use table::table_reference::TableReference;
 use tokio::time::Instant;
@@ -104,7 +104,8 @@ impl BuildPlan {
             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!(
@@ -258,25 +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],
@@ -357,6 +397,8 @@ 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;
 
@@ -405,6 +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;
@@ -421,7 +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);
@@ -446,7 +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);
@@ -479,7 +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);
@@ -504,7 +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);
@@ -524,7 +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);
diff --git a/src/meta-srv/src/procedure/repartition/deallocate_region.rs b/src/meta-srv/src/procedure/repartition/deallocate_region.rs
index 3f5dc5bd8e..f3102221fb 100644
--- a/src/meta-srv/src/procedure/repartition/deallocate_region.rs
+++ b/src/meta-srv/src/procedure/repartition/deallocate_region.rs
@@ -267,6 +267,7 @@ mod tests {
             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;
diff --git a/src/meta-srv/src/procedure/repartition/group.rs b/src/meta-srv/src/procedure/repartition/group.rs
index e5a06f79a8..12374e8ada 100644
--- a/src/meta-srv/src/procedure/repartition/group.rs
+++ b/src/meta-srv/src/procedure/repartition/group.rs
@@ -41,18 +41,14 @@ use common_procedure::{
     Context as ProcedureContext, Error as ProcedureError, LockKey, Procedure,
     Result as ProcedureResult, Status, StringKey, UserMetadata,
 };
-use common_telemetry::{error, info, warn};
+use common_telemetry::{error, info};
 use serde::{Deserialize, Serialize};
 use snafu::{OptionExt, ResultExt};
 use store_api::storage::{RegionId, TableId};
 use uuid::Uuid;
 
 use crate::error::{self, Result};
-use crate::procedure::repartition::group::apply_staging_manifest::ApplyStagingManifest;
-use crate::procedure::repartition::group::enter_staging_region::EnterStagingRegion;
-use crate::procedure::repartition::group::remap_manifest::RemapManifest;
 use crate::procedure::repartition::group::repartition_start::RepartitionStart;
-use crate::procedure::repartition::group::update_metadata::UpdateMetadata;
 use crate::procedure::repartition::plan::RegionDescriptor;
 use crate::procedure::repartition::utils::get_datanode_table_value;
 use crate::procedure::repartition::{self};
@@ -196,62 +192,6 @@ impl RepartitionGroupProcedure {
 
         Ok(Self { state, context })
     }
-
-    async fn rollback_inner(&mut self, procedure_ctx: &ProcedureContext) -> Result<()> {
-        if !self.should_rollback_metadata() {
-            return Ok(());
-        }
-
-        let table_lock =
-            common_meta::lock_key::TableLock::Write(self.context.persistent_ctx.table_id).into();
-        let _guard = procedure_ctx.provider.acquire_lock(&table_lock).await;
-        UpdateMetadata::RollbackStaging
-            .rollback_staging_regions(&mut self.context)
-            .await?;
-
-        if let Err(err) = self.context.invalidate_table_cache().await {
-            warn!(
-                err;
-                "Failed to broadcast the invalidate table cache message during repartition group rollback"
-            );
-        }
-
-        Ok(())
-    }
-
-    /// Returns whether group rollback should revert staging metadata.
-    ///
-    /// This uses an "after metadata apply, before exit staging" semantic.
-    /// Once execution reaches `UpdateMetadata::ApplyStaging` or any later staging state,
-    /// rollback must restore table-route metadata back to the pre-apply view.
-    ///
-    /// State flow:
-    /// `RepartitionStart -> SyncRegion -> UpdateMetadata::ApplyStaging -> EnterStagingRegion`
-    /// `                 -> RemapManifest -> ApplyStagingManifest -> UpdateMetadata::ExitStaging -> RepartitionEnd`
-    /// `                               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^`
-    /// `                               rollback staging metadata`
-    ///
-    /// Notes:
-    /// - `RepartitionStart` / `SyncRegion`: no-op, metadata has not been staged yet.
-    /// - `UpdateMetadata::ApplyStaging` / `EnterStagingRegion` / `RemapManifest` /
-    ///   `ApplyStagingManifest` / `UpdateMetadata::RollbackStaging`: rollback-active.
-    /// - `UpdateMetadata::ExitStaging` / `RepartitionEnd`: excluded, because metadata has
-    ///   already moved into the post-commit exit path.
-    fn should_rollback_metadata(&self) -> bool {
-        self.state.as_any().is::<EnterStagingRegion>()
-            || self.state.as_any().is::<RemapManifest>()
-            || self.state.as_any().is::<ApplyStagingManifest>()
-            || self
-                .state
-                .as_any()
-                .downcast_ref::<UpdateMetadata>()
-                .is_some_and(|state| {
-                    matches!(
-                        state,
-                        UpdateMetadata::ApplyStaging | UpdateMetadata::RollbackStaging
-                    )
-                })
-    }
 }
 
 #[async_trait::async_trait]
@@ -260,10 +200,10 @@ impl Procedure for RepartitionGroupProcedure {
         Self::TYPE_NAME
     }
 
-    async fn rollback(&mut self, ctx: &ProcedureContext) -> ProcedureResult<()> {
-        self.rollback_inner(ctx)
-            .await
-            .map_err(ProcedureError::external)
+    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(
@@ -304,7 +244,9 @@ impl Procedure for RepartitionGroupProcedure {
     }
 
     fn rollback_supported(&self) -> bool {
-        true
+        // Parent repartition owns rollback and recovery because subprocedures are
+        // not relied on as durable rollback units across metasrv restarts.
+        false
     }
 
     fn dump(&self) -> ProcedureResult<String> {
@@ -665,149 +607,12 @@ pub(crate) trait State: Sync + Send + Debug {
 mod tests {
     use std::assert_matches;
     use std::sync::Arc;
-    use std::time::Duration;
 
     use common_meta::key::TableMetadataManager;
     use common_meta::kv_backend::test_util::MockKvBackendBuilder;
-    use common_meta::peer::Peer;
-    use common_meta::rpc::router::{Region, RegionRoute};
-    use common_procedure::{Context as ProcedureContext, Procedure, ProcedureId};
-    use common_procedure_test::MockContextProvider;
-    use partition::expr::PartitionExpr;
-    use store_api::storage::RegionId;
 
-    use super::{
-        Context, PersistentContext, RepartitionGroupProcedure, RepartitionStart, State,
-        region_routes,
-    };
     use crate::error::Error;
-    use crate::procedure::repartition::dispatch::build_region_mapping;
-    use crate::procedure::repartition::group::apply_staging_manifest::ApplyStagingManifest;
-    use crate::procedure::repartition::group::enter_staging_region::EnterStagingRegion;
-    use crate::procedure::repartition::group::remap_manifest::RemapManifest;
-    use crate::procedure::repartition::group::repartition_start::RepartitionStart as GroupRepartitionStart;
-    use crate::procedure::repartition::group::sync_region::SyncRegion;
-    use crate::procedure::repartition::group::update_metadata::UpdateMetadata;
-    use crate::procedure::repartition::plan;
-    use crate::procedure::repartition::repartition_start::RepartitionStart as ParentRepartitionStart;
-    use crate::procedure::repartition::test_util::{
-        TestingEnv, new_persistent_context, range_expr,
-    };
-
-    struct GroupRollbackFixture {
-        context: Context,
-        original_region_routes: Vec<RegionRoute>,
-        next_state: Option<Box<dyn State>>,
-    }
-
-    async fn new_group_rollback_fixture(
-        original_region_routes: Vec<RegionRoute>,
-        from_exprs: Vec<PartitionExpr>,
-        to_exprs: Vec<PartitionExpr>,
-        sync_region: bool,
-    ) -> GroupRollbackFixture {
-        let env = TestingEnv::new();
-        let procedure_ctx = TestingEnv::procedure_context();
-        let table_id = 1024;
-        let mut next_region_number = 10;
-
-        env.create_physical_table_metadata(table_id, original_region_routes.clone())
-            .await;
-
-        let (_, physical_route) = env
-            .table_metadata_manager
-            .table_route_manager()
-            .get_physical_table_route(table_id)
-            .await
-            .unwrap();
-        let allocation_plans =
-            ParentRepartitionStart::build_plan(&physical_route, &from_exprs, &to_exprs).unwrap();
-        assert_eq!(allocation_plans.len(), 1);
-
-        let repartition_plan = plan::convert_allocation_plan_to_repartition_plan(
-            table_id,
-            &mut next_region_number,
-            &allocation_plans[0],
-        );
-        let region_mapping = build_region_mapping(
-            &repartition_plan.source_regions,
-            &repartition_plan.target_regions,
-            &repartition_plan.transition_map,
-        );
-        let persistent_context = PersistentContext::new(
-            repartition_plan.group_id,
-            table_id,
-            "test_catalog".to_string(),
-            "test_schema".to_string(),
-            repartition_plan.source_regions,
-            repartition_plan.target_regions,
-            region_mapping,
-            sync_region,
-            repartition_plan.allocated_region_ids,
-            repartition_plan.pending_deallocate_region_ids,
-            Duration::from_secs(120),
-        );
-        let mut context = env.create_context(persistent_context);
-        let (next_state, _) = GroupRepartitionStart
-            .next(&mut context, &procedure_ctx)
-            .await
-            .unwrap();
-
-        GroupRollbackFixture {
-            context,
-            original_region_routes,
-            next_state: Some(next_state),
-        }
-    }
-
-    async fn new_split_group_rollback_fixture(sync_region: bool) -> GroupRollbackFixture {
-        new_group_rollback_fixture(
-            vec![
-                new_region_route(RegionId::new(1024, 1), Some(range_expr("x", 0, 100))),
-                new_region_route(RegionId::new(1024, 2), Some(range_expr("x", 100, 200))),
-                new_region_route(RegionId::new(1024, 10), None),
-            ],
-            vec![range_expr("x", 0, 100)],
-            vec![range_expr("x", 0, 50), range_expr("x", 50, 100)],
-            sync_region,
-        )
-        .await
-    }
-
-    async fn new_merge_group_rollback_fixture(sync_region: bool) -> GroupRollbackFixture {
-        new_group_rollback_fixture(
-            vec![
-                new_region_route(RegionId::new(1024, 1), Some(range_expr("x", 0, 100))),
-                new_region_route(RegionId::new(1024, 2), Some(range_expr("x", 100, 200))),
-                new_region_route(RegionId::new(1024, 3), Some(range_expr("x", 200, 300))),
-            ],
-            vec![range_expr("x", 0, 100), range_expr("x", 100, 200)],
-            vec![range_expr("x", 0, 200)],
-            sync_region,
-        )
-        .await
-    }
-
-    async fn stage_metadata(context: &mut Context) {
-        UpdateMetadata::ApplyStaging
-            .apply_staging_regions(context)
-            .await
-            .unwrap();
-    }
-
-    fn new_region_route(region_id: RegionId, partition_expr: Option<PartitionExpr>) -> RegionRoute {
-        RegionRoute {
-            region: Region {
-                id: region_id,
-                partition_expr: partition_expr
-                    .map(|expr| expr.as_json_str().unwrap())
-                    .unwrap_or_default(),
-                ..Default::default()
-            },
-            leader_peer: Some(Peer::empty(1)),
-            ..Default::default()
-        }
-    }
+    use crate::procedure::repartition::test_util::{TestingEnv, new_persistent_context};
 
     #[tokio::test]
     async fn test_get_table_route_value_not_found_error() {
@@ -856,198 +661,4 @@ mod tests {
         let err = ctx.get_datanode_table_value(1024, 1).await.unwrap_err();
         assert!(err.is_retryable());
     }
-
-    #[tokio::test]
-    async fn test_group_rollback_supported() {
-        let env = TestingEnv::new();
-        let persistent_context = new_persistent_context(1024, vec![], vec![]);
-        let procedure = RepartitionGroupProcedure {
-            state: Box::new(RepartitionStart),
-            context: env.create_context(persistent_context),
-        };
-
-        assert!(procedure.rollback_supported());
-    }
-
-    #[tokio::test]
-    async fn test_group_rollback_is_noop_before_apply_staging() {
-        let env = TestingEnv::new();
-        let persistent_context = new_persistent_context(1024, vec![], vec![]);
-        let ctx = env.create_context(persistent_context.clone());
-        let mut procedure = RepartitionGroupProcedure {
-            state: Box::new(RepartitionStart),
-            context: ctx,
-        };
-        let provider = Arc::new(MockContextProvider::new(Default::default()));
-        let procedure_ctx = ProcedureContext {
-            procedure_id: ProcedureId::random(),
-            provider,
-        };
-
-        procedure.rollback(&procedure_ctx).await.unwrap();
-
-        assert!(procedure.state.as_any().is::<RepartitionStart>());
-        assert_eq!(procedure.context.persistent_ctx, persistent_context);
-    }
-
-    async fn assert_noop_rollback(
-        fixture: GroupRollbackFixture,
-        state: Box<dyn State>,
-        assert_state: impl FnOnce(&dyn State),
-    ) {
-        let original_region_routes = fixture.original_region_routes.clone();
-        let procedure_ctx = TestingEnv::procedure_context();
-        let mut procedure = RepartitionGroupProcedure {
-            state,
-            context: fixture.context,
-        };
-
-        procedure.rollback(&procedure_ctx).await.unwrap();
-
-        assert_state(&*procedure.state);
-        let table_route_value = procedure
-            .context
-            .get_table_route_value()
-            .await
-            .unwrap()
-            .into_inner();
-        let region_routes = region_routes(
-            procedure.context.persistent_ctx.table_id,
-            &table_route_value,
-        )
-        .unwrap();
-        assert_eq!(region_routes.clone(), original_region_routes);
-    }
-
-    async fn assert_metadata_rollback_restores_table_route(
-        mut fixture: GroupRollbackFixture,
-        state: Box<dyn State>,
-    ) {
-        let original_region_routes = fixture.original_region_routes.clone();
-        let procedure_ctx = TestingEnv::procedure_context();
-        stage_metadata(&mut fixture.context).await;
-        let mut procedure = RepartitionGroupProcedure {
-            state,
-            context: fixture.context,
-        };
-
-        procedure.rollback(&procedure_ctx).await.unwrap();
-
-        let table_route_value = procedure
-            .context
-            .get_table_route_value()
-            .await
-            .unwrap()
-            .into_inner();
-        let region_routes = region_routes(
-            procedure.context.persistent_ctx.table_id,
-            &table_route_value,
-        )
-        .unwrap();
-        assert_eq!(region_routes.clone(), original_region_routes);
-    }
-
-    #[tokio::test]
-    async fn test_group_rollback_is_noop_in_sync_region() {
-        let mut fixture = new_split_group_rollback_fixture(true).await;
-        assert!(
-            fixture
-                .next_state
-                .as_ref()
-                .unwrap()
-                .as_any()
-                .is::<SyncRegion>()
-        );
-        let state = fixture.next_state.take().unwrap();
-
-        assert_noop_rollback(fixture, state, |state| {
-            assert!(state.as_any().is::<SyncRegion>());
-        })
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_group_rollback_is_noop_in_exit_staging() {
-        let fixture = new_split_group_rollback_fixture(false).await;
-
-        assert_noop_rollback(fixture, Box::new(UpdateMetadata::ExitStaging), |state| {
-            assert!(state.as_any().is::<UpdateMetadata>());
-            assert!(matches!(
-                state.as_any().downcast_ref::<UpdateMetadata>(),
-                Some(UpdateMetadata::ExitStaging)
-            ));
-        })
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_group_rollback_restores_split_routes_from_apply_staging() {
-        let fixture = new_split_group_rollback_fixture(false).await;
-        assert_metadata_rollback_restores_table_route(
-            fixture,
-            Box::new(UpdateMetadata::ApplyStaging),
-        )
-        .await;
-    }
-
-    #[tokio::test]
-    async fn test_group_rollback_restores_split_routes_from_enter_staging_region() {
-        let fixture = new_split_group_rollback_fixture(false).await;
-        assert_metadata_rollback_restores_table_route(fixture, Box::new(EnterStagingRegion)).await;
-    }
-
-    #[tokio::test]
-    async fn test_group_rollback_restores_split_routes_from_remap_manifest() {
-        let fixture = new_split_group_rollback_fixture(false).await;
-        assert_metadata_rollback_restores_table_route(fixture, Box::new(RemapManifest)).await;
-    }
-
-    #[tokio::test]
-    async fn test_group_rollback_restores_split_routes_from_apply_staging_manifest() {
-        let fixture = new_split_group_rollback_fixture(false).await;
-        assert_metadata_rollback_restores_table_route(fixture, Box::new(ApplyStagingManifest))
-            .await;
-    }
-
-    #[tokio::test]
-    async fn test_group_rollback_restores_merge_routes_and_is_idempotent() {
-        let mut fixture = new_merge_group_rollback_fixture(false).await;
-        let original_region_routes = fixture.original_region_routes.clone();
-        let procedure_ctx = TestingEnv::procedure_context();
-        stage_metadata(&mut fixture.context).await;
-        let mut procedure = RepartitionGroupProcedure {
-            state: Box::new(UpdateMetadata::ApplyStaging),
-            context: fixture.context,
-        };
-
-        procedure.rollback(&procedure_ctx).await.unwrap();
-        let table_route_value = procedure
-            .context
-            .get_table_route_value()
-            .await
-            .unwrap()
-            .into_inner();
-        let once = region_routes(
-            procedure.context.persistent_ctx.table_id,
-            &table_route_value,
-        )
-        .unwrap()
-        .clone();
-        procedure.rollback(&procedure_ctx).await.unwrap();
-        let table_route_value = procedure
-            .context
-            .get_table_route_value()
-            .await
-            .unwrap()
-            .into_inner();
-        let twice = region_routes(
-            procedure.context.persistent_ctx.table_id,
-            &table_route_value,
-        )
-        .unwrap()
-        .clone();
-
-        assert_eq!(once, original_region_routes);
-        assert_eq!(once, twice);
-    }
 }
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/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 4e6bf67fc8..0000000000
--- a/src/meta-srv/src/procedure/repartition/group/update_metadata/rollback_staging_region.rs
+++ /dev/null
@@ -1,301 +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 store_api::storage::RegionId;
-
-use crate::error::{self, Result};
-use crate::procedure::repartition::group::update_metadata::UpdateMetadata;
-use crate::procedure::repartition::group::{Context, GroupId, region_routes};
-use crate::procedure::repartition::plan::RegionDescriptor;
-
-impl UpdateMetadata {
-    /// Rolls back the staging regions.
-    ///
-    /// Abort:
-    /// - Source region not found.
-    /// - Target region not found.
-    fn rollback_staging_region_routes(
-        group_id: GroupId,
-        sources: &[RegionDescriptor],
-        original_target_routes: &[RegionRoute],
-        pending_deallocate_region_ids: &[RegionId],
-        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 sources {
-            let region_route = region_routes_map.get_mut(&source.region_id).context(
-                error::RepartitionSourceRegionMissingSnafu {
-                    group_id,
-                    region_id: source.region_id,
-                },
-            )?;
-            // Clean leader staging state for source regions.
-            region_route.clear_leader_staging();
-            if pending_deallocate_region_ids.contains(&source.region_id) {
-                // Clean ignore all writes state for source regions if it's pending to be deallocated,
-                // which means the source region is merged into the target region.
-                region_route.clear_ignore_all_writes();
-            }
-        }
-
-        for target in original_target_routes {
-            let region_route = region_routes_map.get_mut(&target.region.id).context(
-                error::RepartitionTargetRegionMissingSnafu {
-                    group_id,
-                    region_id: target.region.id,
-                },
-            )?;
-
-            // Revert the partition expression and write route policy to the original value for the target region.
-            region_route.region.partition_expr = target.region.partition_expr.clone();
-            region_route.write_route_policy = target.write_route_policy;
-
-            // Clean leader staging state for target regions.
-            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,
-            &ctx.persistent_ctx.sources,
-            &prepare_result.target_routes,
-            &ctx.persistent_ctx.pending_deallocate_region_ids,
-            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 std::collections::HashSet;
-
-    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::plan::RegionDescriptor;
-    use crate::procedure::repartition::test_util::range_expr;
-
-    fn new_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_rollback_staging_region_routes_split_case() {
-        let group_id = Uuid::new_v4();
-        let table_id = 1024;
-        let original_region_routes = vec![
-            new_region_route(
-                RegionId::new(table_id, 1),
-                &range_expr("x", 0, 100).as_json_str().unwrap(),
-                None,
-                false,
-            ),
-            new_region_route(
-                RegionId::new(table_id, 2),
-                &range_expr("x", 100, 200).as_json_str().unwrap(),
-                None,
-                false,
-            ),
-            new_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 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);
-        let new_region_routes = UpdateMetadata::rollback_staging_region_routes(
-            group_id,
-            &sources,
-            &target_routes,
-            &[],
-            &applied_region_routes,
-        )
-        .unwrap();
-
-        assert_eq!(new_region_routes, original_region_routes);
-    }
-
-    #[test]
-    fn test_rollback_staging_region_routes_merge_case_is_idempotent() {
-        let group_id = Uuid::new_v4();
-        let table_id = 1024;
-        let original_region_routes = vec![
-            new_region_route(
-                RegionId::new(table_id, 1),
-                &range_expr("x", 0, 100).as_json_str().unwrap(),
-                None,
-                false,
-            ),
-            new_region_route(
-                RegionId::new(table_id, 2),
-                &range_expr("x", 100, 200).as_json_str().unwrap(),
-                None,
-                false,
-            ),
-            new_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 applied_region_routes = UpdateMetadata::apply_staging_region_routes(
-            group_id,
-            &sources,
-            &targets,
-            &[RegionId::new(table_id, 2)],
-            &original_region_routes,
-        )
-        .unwrap();
-
-        let once = UpdateMetadata::rollback_staging_region_routes(
-            group_id,
-            &sources,
-            &target_routes,
-            &[RegionId::new(table_id, 2)],
-            &applied_region_routes,
-        )
-        .unwrap();
-        let twice = UpdateMetadata::rollback_staging_region_routes(
-            group_id,
-            &sources,
-            &target_routes,
-            &[RegionId::new(table_id, 2)],
-            &once,
-        )
-        .unwrap();
-
-        assert_eq!(once, original_region_routes);
-        assert_eq!(once, twice);
-    }
-}
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/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/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/metric-engine/Cargo.toml b/src/metric-engine/Cargo.toml
index 5b561997ab..0005a4dc54 100644
--- a/src/metric-engine/Cargo.toml
+++ b/src/metric-engine/Cargo.toml
@@ -33,7 +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/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/compaction.rs b/src/mito2/src/compaction.rs
index 944c51ebd6..981d4f1d11 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};
@@ -53,9 +54,9 @@ 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::BoxedRecordBatchStream;
@@ -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 {
@@ -703,6 +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 {
@@ -719,9 +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() {
@@ -764,6 +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.
@@ -1362,6 +1516,58 @@ 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();
@@ -1542,6 +1748,11 @@ 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 {
@@ -1558,6 +1769,142 @@ 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;
@@ -1713,6 +2060,11 @@ 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 {
@@ -1752,6 +2104,7 @@ 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),
@@ -1827,6 +2180,7 @@ 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),
@@ -1873,6 +2227,11 @@ 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)
@@ -1910,6 +2269,11 @@ 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 ff4317331f..a4a3b9b145 100644
--- a/src/mito2/src/compaction/compactor.rs
+++ b/src/mito2/src/compaction/compactor.rs
@@ -17,8 +17,9 @@ 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;
@@ -38,11 +39,10 @@ use crate::compaction::picker::{PickerOutput, new_picker};
 use crate::compaction::{CompactionOutput, CompactionSstReaderBuilder, find_dynamic_options};
 use crate::config::MitoConfig;
 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;
 use crate::region::options::RegionOptions;
 use crate::region::version::VersionRef;
@@ -56,6 +56,7 @@ use crate::sst::index::puffin_manager::PuffinManagerFactory;
 use crate::sst::location::region_dir_from_table_dir;
 use crate::sst::parquet::WriteOptions;
 use crate::sst::version::{SstVersion, SstVersionRef};
+use crate::{error, metrics};
 
 /// Region version for compaction that does not hold memtables.
 #[derive(Clone)]
@@ -299,12 +300,28 @@ pub trait Compactor: Send + Sync + 'static {
     ) -> 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,
@@ -424,54 +441,145 @@ 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()
@@ -480,7 +588,7 @@ 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),
         })
     }
@@ -558,3 +666,328 @@ impl Compactor for DefaultCompactor {
         Ok(())
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::sync::atomic::{AtomicUsize, Ordering};
+    use std::sync::{Arc, Mutex};
+    use std::time::Duration;
+
+    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 = 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);
+
+        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,
+        };
+
+        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/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/engine.rs b/src/mito2/src/engine.rs
index d006067f0d..5bd1002581 100644
--- a/src/mito2/src/engine.rs
+++ b/src/mito2/src/engine.rs
@@ -1114,13 +1114,9 @@ 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())
     }
 }
 
diff --git a/src/mito2/src/engine/alter_test.rs b/src/mito2/src/engine/alter_test.rs
index 05ba5dae25..a7798a7678 100644
--- a/src/mito2/src/engine/alter_test.rs
+++ b/src/mito2/src/engine/alter_test.rs
@@ -19,10 +19,11 @@ 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,
@@ -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,
diff --git a/src/mito2/src/engine/apply_staging_manifest_test.rs b/src/mito2/src/engine/apply_staging_manifest_test.rs
index a82fcfe049..efa0713cfc 100644
--- a/src/mito2/src/engine/apply_staging_manifest_test.rs
+++ b/src/mito2/src/engine/apply_staging_manifest_test.rs
@@ -333,7 +333,7 @@ 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());
diff --git a/src/mito2/src/engine/compaction_test.rs b/src/mito2/src/engine/compaction_test.rs
index cbcad3a58a..dc8aa94fdf 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;
@@ -650,7 +652,7 @@ 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 +708,91 @@ 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]
diff --git a/src/mito2/src/engine/filter_deleted_test.rs b/src/mito2/src/engine/filter_deleted_test.rs
index 497583b8bc..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;
@@ -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/prune_test.rs b/src/mito2/src/engine/prune_test.rs
index 599547ec8d..f217151034 100644
--- a/src/mito2/src/engine/prune_test.rs
+++ b/src/mito2/src/engine/prune_test.rs
@@ -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())))),
@@ -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/set_role_state_test.rs b/src/mito2/src/engine/set_role_state_test.rs
index 4fb15ab7fe..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;
@@ -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
@@ -325,6 +331,264 @@ 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
diff --git a/src/mito2/src/engine/staging_test.rs b/src/mito2/src/engine/staging_test.rs
index bd90779e0b..9846933d1f 100644
--- a/src/mito2/src/engine/staging_test.rs
+++ b/src/mito2/src/engine/staging_test.rs
@@ -547,7 +547,7 @@ 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()
diff --git a/src/mito2/src/error.rs b/src/mito2/src/error.rs
index c6b69fe607..eb802d50b7 100644
--- a/src/mito2/src/error.rs
+++ b/src/mito2/src/error.rs
@@ -1073,6 +1073,9 @@ 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,
@@ -1389,7 +1392,7 @@ 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/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..95ad00fb8c 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);
diff --git a/src/mito2/src/manifest/tests/checkpoint.rs b/src/mito2/src/manifest/tests/checkpoint.rs
index 64547f45aa..89f53e5627 100644
--- a/src/mito2/src/manifest/tests/checkpoint.rs
+++ b/src/mito2/src/manifest/tests/checkpoint.rs
@@ -14,9 +14,13 @@
 
 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();
@@ -414,7 +419,7 @@ 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 +494,96 @@ 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 `flush`, simulating the S3 batch-delete failure
+/// described in issue #7986.
+struct FailingDeleter {
+    inner: oio::Deleter,
+    flush_calls: Arc<AtomicUsize>,
+}
+
+impl oio::Delete for FailingDeleter {
+    fn delete(&mut self, path: &str, args: OpDelete) -> MockResult<()> {
+        self.inner.delete(path, args)
+    }
+
+    async fn flush(&mut self) -> MockResult<usize> {
+        self.flush_calls.fetch_add(1, Ordering::Relaxed);
+        Err(MockError::new(
+            ErrorKind::Unexpected,
+            "mock manifest delete flush failure",
+        ))
+    }
+}
+
+#[tokio::test]
+async fn checkpoint_advances_and_recovery_works_when_delete_fails() {
+    common_telemetry::init_default_ut_logging();
+
+    let flush_calls = Arc::new(AtomicUsize::new(0));
+    let factory_flush_calls = flush_calls.clone();
+    let mock_layer = MockLayerBuilder::default()
+        .deleter_factory(Arc::new(move |inner| {
+            Box::new(FailingDeleter {
+                inner,
+                flush_calls: factory_flush_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!(flush_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/bulk/part.rs b/src/mito2/src/memtable/bulk/part.rs
index 986e9409ee..826474df05 100644
--- a/src/mito2/src/memtable/bulk/part.rs
+++ b/src/mito2/src/memtable/bulk/part.rs
@@ -59,7 +59,6 @@ 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};
 use crate::sst::parquet::{PARQUET_METADATA_KEY, SstInfo};
@@ -1028,9 +1027,8 @@ 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 =
@@ -1050,20 +1048,6 @@ 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
diff --git a/src/mito2/src/memtable/bulk/part_reader.rs b/src/mito2/src/memtable/bulk/part_reader.rs
index a9caeef08c..cea3304154 100644
--- a/src/mito2/src/memtable/bulk/part_reader.rs
+++ b/src/mito2/src/memtable/bulk/part_reader.rs
@@ -29,7 +29,7 @@ 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::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};
 
@@ -78,7 +78,7 @@ impl EncodedBulkPartIter {
 
         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)
             }
@@ -140,9 +140,7 @@ 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);
@@ -299,11 +297,7 @@ 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,
diff --git a/src/mito2/src/memtable/bulk/row_group_reader.rs b/src/mito2/src/memtable/bulk/row_group_reader.rs
index 40a5b2f85d..36e8e8aac1 100644
--- a/src/mito2/src/memtable/bulk/row_group_reader.rs
+++ b/src/mito2/src/memtable/bulk/row_group_reader.rs
@@ -31,7 +31,6 @@ use crate::sst::parquet::DEFAULT_READ_BATCH_SIZE;
 
 pub(crate) struct MemtableRowGroupReaderBuilder {
     projection: ProjectionMask,
-    parquet_metadata: Arc<ParquetMetaData>,
     arrow_metadata: ArrowReaderMetadata,
     data: Bytes,
 }
@@ -51,7 +50,6 @@ impl MemtableRowGroupReaderBuilder {
                 .context(ReadDataPartSnafu)?;
         Ok(Self {
             projection,
-            parquet_metadata,
             arrow_metadata,
             data,
         })
@@ -79,23 +77,4 @@ impl MemtableRowGroupReaderBuilder {
 
         builder.build().context(ReadDataPartSnafu)
     }
-
-    /// 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)
-            }
-        }
-    }
 }
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/scan_region.rs b/src/mito2/src/read/scan_region.rs
index c447685822..cbda7265b6 100644
--- a/src/mito2/src/read/scan_region.rs
+++ b/src/mito2/src/read/scan_region.rs
@@ -1101,10 +1101,10 @@ 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 reader = self
             .access_layer
@@ -1125,7 +1125,7 @@ impl ScanInput {
         let res = reader
             .expected_metadata(Some(self.mapper.metadata().clone()))
             .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;
@@ -1330,6 +1330,18 @@ 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")]
@@ -1372,7 +1384,7 @@ 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,
     }
 }
@@ -1533,6 +1545,13 @@ 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
@@ -2095,4 +2114,24 @@ 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/seq_scan.rs b/src/mito2/src/read/seq_scan.rs
index 15ab435425..8e560fbc1f 100644
--- a/src/mito2/src/read/seq_scan.rs
+++ b/src/mito2/src/read/seq_scan.rs
@@ -208,12 +208,18 @@ impl SeqScan {
         }
 
         let mapper = stream_ctx.input.mapper.as_flat().unwrap();
-        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?;
+        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 dedup = !stream_ctx.input.append_mode;
         let dedup_metrics_reporter = part_metrics.map(|m| m.dedup_metrics_reporter());
@@ -221,7 +227,7 @@ impl SeqScan {
             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,
                     )
@@ -229,7 +235,7 @@ 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,
@@ -240,7 +246,7 @@ impl SeqScan {
                 ) as _,
             }
         } else {
-            Box::pin(reader.into_stream()) as _
+            reader
         };
 
         let reader = match &stream_ctx.input.series_row_selector {
diff --git a/src/mito2/src/region.rs b/src/mito2/src/region.rs
index 26ab96c779..3804b28afb 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,
 }
@@ -333,6 +328,17 @@ 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 +399,8 @@ 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),
         )
     }
@@ -819,7 +824,7 @@ 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 +842,8 @@ 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 +857,8 @@ 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 +878,11 @@ 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 +890,46 @@ 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()
@@ -1028,27 +1075,50 @@ 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)
@@ -1071,6 +1141,20 @@ 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,
@@ -1096,7 +1180,27 @@ 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),
@@ -1438,8 +1542,8 @@ 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;
@@ -1512,7 +1616,6 @@ mod tests {
             topic_latest_entry_id: Default::default(),
             written_bytes: Arc::new(AtomicU64::new(0)),
             stats: ManifestStats::default(),
-            staging_partition_info: Mutex::new(None),
         }
     }
 
@@ -1684,6 +1787,13 @@ 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!(
@@ -1825,7 +1935,6 @@ 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 b23e73557d..c1240c3829 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};
@@ -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);
diff --git a/src/mito2/src/request.rs b/src/mito2/src/request.rs
index 4f8f1b0d51..cc965f48df 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};
@@ -895,6 +895,8 @@ 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 +993,24 @@ 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 +1169,13 @@ 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 +1239,21 @@ 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/parquet/file_range.rs b/src/mito2/src/sst/parquet/file_range.rs
index bf86e4a764..5863833df4 100644
--- a/src/mito2/src/sst/parquet/file_range.rs
+++ b/src/mito2/src/sst/parquet/file_range.rs
@@ -146,7 +146,7 @@ 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
@@ -158,22 +158,6 @@ 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].
     #[allow(dead_code)]
     pub(crate) async fn reader(
@@ -185,7 +169,7 @@ impl FileRange {
             return Ok(None);
         }
         // Compute skip_fields once for this row group
-        let skip_fields = self.context.should_skip_fields(self.row_group_idx);
+        let skip_fields = self.context.base.pre_filter_mode.skip_fields();
         let parquet_reader = self
             .context
             .reader_builder
@@ -247,7 +231,7 @@ impl FileRange {
             return Ok(None);
         }
         // Compute skip_fields once for this row group
-        let skip_fields = self.context.should_skip_fields(self.row_group_idx);
+        let skip_fields = self.context.base.pre_filter_mode.skip_fields();
         let parquet_reader = self
             .context
             .reader_builder
@@ -404,16 +388,8 @@ impl FileRangeContext {
         )
     }
 
-    /// 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.
@@ -447,17 +423,20 @@ impl FileRangeContext {
 }
 
 /// Mode to pre-filter columns in a range.
-#[derive(Debug, Clone, Copy)]
+#[derive(Debug, Clone, Copy, PartialEq)]
 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>,
@@ -514,7 +493,7 @@ impl RangeBase {
     ///
     /// # Arguments
     /// * `input` - The batch 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(
         &self,
         mut input: Batch,
@@ -626,7 +605,7 @@ impl RangeBase {
     ///
     /// # 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,
@@ -679,7 +658,7 @@ impl RangeBase {
     ///
     /// # 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,
diff --git a/src/mito2/src/sst/parquet/reader.rs b/src/mito2/src/sst/parquet/reader.rs
index 73ca7748e9..34fb12a1a7 100644
--- a/src/mito2/src/sst/parquet/reader.rs
+++ b/src/mito2/src/sst/parquet/reader.rs
@@ -72,7 +72,6 @@ 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::metadata::MetadataLoader;
@@ -676,7 +675,7 @@ 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,
@@ -1112,25 +1111,6 @@ 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,
@@ -1955,7 +1935,7 @@ impl ParquetReader {
                 return Ok(None);
             };
 
-            let skip_fields = self.context.should_skip_fields(row_group_idx);
+            let skip_fields = self.context.pre_filter_mode().skip_fields();
             let parquet_reader = self
                 .context
                 .reader_builder()
@@ -1988,7 +1968,7 @@ impl ParquetReader {
         debug_assert!(context.read_format().as_flat().is_some());
         let fetch_metrics = ParquetFetchMetrics::default();
         let reader = if let Some((row_group_idx, row_selection)) = selection.pop_first() {
-            let skip_fields = context.should_skip_fields(row_group_idx);
+            let skip_fields = context.pre_filter_mode().skip_fields();
             let parquet_reader = context
                 .reader_builder()
                 .build(context.build_context(
diff --git a/src/mito2/src/test_util.rs b/src/mito2/src/test_util.rs
index 350195bfa9..15785d8ac7 100644
--- a/src/mito2/src/test_util.rs
+++ b/src/mito2/src/test_util.rs
@@ -610,9 +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
diff --git a/src/mito2/src/worker.rs b/src/mito2/src/worker.rs
index fd5ad82f3f..3fb6a63184 100644
--- a/src/mito2/src/worker.rs
+++ b/src/mito2/src/worker.rs
@@ -1196,6 +1196,9 @@ 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_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_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_enter_staging.rs b/src/mito2/src/worker/handle_enter_staging.rs
index 8b75fdd24f..75e7071468 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);
@@ -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/object-store/tests/object_store_test.rs b/src/object-store/tests/object_store_test.rs
index 5df6ab04bd..3dac15da46 100644
--- a/src/object-store/tests/object_store_test.rs
+++ b/src/object-store/tests/object_store_test.rs
@@ -17,6 +17,7 @@ use std::env;
 use anyhow::Result;
 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;
@@ -103,6 +104,109 @@ 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();
@@ -129,6 +233,7 @@ 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 +263,7 @@ 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 +293,7 @@ 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 +323,7 @@ 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 +352,7 @@ 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..139d08378e 100644
--- a/src/operator/Cargo.toml
+++ b/src/operator/Cargo.toml
@@ -53,7 +53,7 @@ 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
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/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/servers/Cargo.toml b/src/servers/Cargo.toml
index 2d68f17699..d4366e5d1f 100644
--- a/src/servers/Cargo.toml
+++ b/src/servers/Cargo.toml
@@ -63,6 +63,7 @@ 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"
@@ -84,7 +85,8 @@ 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
@@ -107,6 +109,7 @@ rand.workspace = true
 regex.workspace = true
 reqwest.workspace = true
 rust-embed = { version = "6.6", optional = true, features = ["debug-embed"] }
+rust_decimal = { workspace = true, features = ["db-postgres"] }
 rustls = { workspace = true, default-features = false, features = ["ring", "logging", "std", "tls12"] }
 rustls-pemfile = "2.0"
 rustls-pki-types = "1.0"
diff --git a/src/servers/src/http.rs b/src/servers/src/http.rs
index d25be0485f..6e3209cc8a 100644
--- a/src/servers/src/http.rs
+++ b/src/servers/src/http.rs
@@ -129,7 +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 {
@@ -764,6 +765,7 @@ impl HttpServer {
         };
 
         router = router
+            .route("/", routing::get(handler::index))
             .route(
                 "/health",
                 routing::get(handler::health).post(handler::health),
diff --git a/src/servers/src/http/authorize.rs b/src/servers/src/http/authorize.rs
index d2bfd9eba2..ef3a74b38e 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;
         }
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/mysql/handler.rs b/src/servers/src/mysql/handler.rs
index 2ce229fea4..9bd54a2de4 100644
--- a/src/servers/src/mysql/handler.rs
+++ b/src/servers/src/mysql/handler.rs
@@ -455,9 +455,17 @@ 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
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/postgres/handler.rs b/src/servers/src/postgres/handler.rs
index 2b84b3aa30..94363b06eb 100644
--- a/src/servers/src/postgres/handler.rs
+++ b/src/servers/src/postgres/handler.rs
@@ -456,16 +456,13 @@ impl ExtendedQueryHandler for PostgresServerHandlerInner {
                 .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)?);
-            }
-
+            // 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(&sql, query_ctx.clone())
+                .do_query(&sql_plan.query, query_ctx.clone())
                 .await
                 .remove(0)
         };
diff --git a/src/servers/src/postgres/types.rs b/src/servers/src/postgres/types.rs
index d4d15ef64a..203e477c6f 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;
 
@@ -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 +352,17 @@ 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())
+    }
+}
+
 pub(super) fn parameters_to_scalar_values(
     plan: &LogicalPlan,
     portal: &Portal<PgSqlPlan>,
@@ -442,7 +398,7 @@ pub(super) fn parameters_to_scalar_values(
         };
 
         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 {
@@ -558,6 +514,24 @@ 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 {
@@ -837,7 +811,67 @@ 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<_>>();
@@ -1098,6 +1132,7 @@ 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::{
@@ -1516,4 +1551,26 @@ 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));
+    }
 }
diff --git a/src/servers/tests/mod.rs b/src/servers/tests/mod.rs
index c4f83c5e6c..3ecf2acf73 100644
--- a/src/servers/tests/mod.rs
+++ b/src/servers/tests/mod.rs
@@ -19,18 +19,21 @@ use api::v1::query_request::Query;
 use async_trait::async_trait;
 use catalog::memory::MemoryCatalogManager;
 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;
 
@@ -52,15 +55,29 @@ 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(
diff --git a/src/servers/tests/mysql/mysql_server_test.rs b/src/servers/tests/mysql/mysql_server_test.rs
index c53ff34d45..94fcebf13c 100644
--- a/src/servers/tests/mysql/mysql_server_test.rs
+++ b/src/servers/tests/mysql/mysql_server_test.rs
@@ -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(())
 }
 
diff --git a/src/store-api/src/region_engine.rs b/src/store-api/src/region_engine.rs
index 287f64d225..b235fcffc7 100644
--- a/src/store-api/src/region_engine.rs
+++ b/src/store-api/src/region_engine.rs
@@ -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,
         }
diff --git a/tests-integration/Cargo.toml b/tests-integration/Cargo.toml
index 10b7097f4f..b9bda50c0b 100644
--- a/tests-integration/Cargo.toml
+++ b/tests-integration/Cargo.toml
@@ -62,7 +62,6 @@ loki-proto.workspace = true
 meta-client.workspace = true
 meta-srv = { workspace = true, features = ["mock"] }
 mito2.workspace = true
-moka.workspace = true
 object-store.workspace = true
 operator = { workspace = true, features = ["testing"] }
 prost.workspace = true
@@ -96,6 +95,7 @@ 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
diff --git a/tests-integration/tests/main.rs b/tests-integration/tests/main.rs
index 01aed6c8e9..95f605ff11 100644
--- a/tests-integration/tests/main.rs
+++ b/tests-integration/tests/main.rs
@@ -27,6 +27,7 @@ mod region_migration;
 mod repartition;
 #[macro_use]
 mod repartition_expr_version;
+mod mysql;
 
 grpc_tests!(File, S3, S3WithCache, Oss, Azblob, Gcs);
 
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..aa0062e4fd 100644
--- a/tests-integration/tests/sql.rs
+++ b/tests-integration/tests/sql.rs
@@ -82,6 +82,7 @@ macro_rules! sql_tests {
                 test_postgres_datestyle,
                 test_postgres_intervalstyle,
                 test_postgres_parameter_inference,
+                test_postgres_uint64_parameter,
                 test_postgres_array_types,
                 test_mysql_prepare_stmt_insert_timestamp,
                 test_declare_fetch_close_cursor,
@@ -1300,6 +1301,57 @@ 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_mysql_async_timestamp(store_type: StorageType) {
     use mysql_async::prelude::*;
     use time::PrimitiveDateTime;
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/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/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;