ci: upload .pdb files too

* chore: add log for converting region to follower

* chore: apply suggestions from CR

.github/actions/build-greptime-binary/action.yml

@@ -54,7 +54,7 @@ runs:
         PROFILE_TARGET: ${{ inputs.cargo-profile == 'dev' && 'debug' || inputs.cargo-profile }}
       with:
         artifacts-dir: ${{ inputs.artifacts-dir }}
-        target-file: ./target/$PROFILE_TARGET/greptime
+        target-files: ./target/$PROFILE_TARGET/greptime
         version: ${{ inputs.version }}
         working-dir: ${{ inputs.working-dir }}
 
@@ -72,6 +72,6 @@ runs:
       if: ${{ inputs.build-android-artifacts == 'true' }}
       with:
         artifacts-dir: ${{ inputs.artifacts-dir }}
-        target-file: ./target/aarch64-linux-android/release/greptime
+        target-files: ./target/aarch64-linux-android/release/greptime
         version: ${{ inputs.version }}
         working-dir: ${{ inputs.working-dir }}

.github/actions/build-macos-artifacts/action.yml

@@ -90,5 +90,5 @@ runs:
       uses: ./.github/actions/upload-artifacts
       with:
         artifacts-dir: ${{ inputs.artifacts-dir }}
-        target-file: target/${{ inputs.arch }}/${{ inputs.cargo-profile }}/greptime
+        target-files: target/${{ inputs.arch }}/${{ inputs.cargo-profile }}/greptime
         version: ${{ inputs.version }}

.github/actions/build-windows-artifacts/action.yml

@@ -76,5 +76,5 @@ runs:
       uses: ./.github/actions/upload-artifacts
       with:
         artifacts-dir: ${{ inputs.artifacts-dir }}
-        target-file: target/${{ inputs.arch }}/${{ inputs.cargo-profile }}/greptime
+        target-files: target/${{ inputs.arch }}/${{ inputs.cargo-profile }}/greptime,target/${{ inputs.arch }}/${{ inputs.cargo-profile }}/greptime.pdb
         version: ${{ inputs.version }}

.github/actions/setup-greptimedb-cluster/with-minio-and-cache.yaml

@@ -5,7 +5,7 @@ meta:
 
     [datanode]
     [datanode.client]
-    timeout = "60s"
+    timeout = "120s"
 datanode:
   configData: |-
     [runtime]
@@ -21,7 +21,7 @@ frontend:
     global_rt_size = 4
 
     [meta_client]
-    ddl_timeout = "60s"
+    ddl_timeout = "120s"
 objectStorage:
   s3:
     bucket: default

.github/actions/setup-greptimedb-cluster/with-minio.yaml

@@ -5,7 +5,7 @@ meta:
     
     [datanode]
     [datanode.client]
-    timeout = "60s"
+    timeout = "120s"
 datanode:
   configData: |-
     [runtime]
@@ -17,7 +17,7 @@ frontend:
     global_rt_size = 4
 
     [meta_client]
-    ddl_timeout = "60s"
+    ddl_timeout = "120s"
 objectStorage:
   s3:
     bucket: default

.github/actions/setup-greptimedb-cluster/with-remote-wal.yaml

@@ -11,7 +11,7 @@ meta:
         
     [datanode]
     [datanode.client]
-    timeout = "60s"
+    timeout = "120s"
 datanode:
   configData: |-
     [runtime]
@@ -28,7 +28,7 @@ frontend:
     global_rt_size = 4
 
     [meta_client]
-    ddl_timeout = "60s"
+    ddl_timeout = "120s"
 objectStorage:
   s3:
     bucket: default

.github/actions/upload-artifacts/action.yml

@@ -4,8 +4,8 @@ inputs:
   artifacts-dir:
     description: Directory to store artifacts
     required: true
-  target-file:
+  target-files:
-    description: The path of the target artifact
+    description: The multiple target files to upload, separated by comma
     required: false
   version:
     description: Version of the artifact
@@ -18,12 +18,16 @@ runs:
   using: composite
   steps:
     - name: Create artifacts directory
-      if: ${{ inputs.target-file != '' }}
+      if: ${{ inputs.target-files != '' }}
       working-directory: ${{ inputs.working-dir }}
       shell: bash
       run: |
-        mkdir -p ${{ inputs.artifacts-dir }} && \
+        set -e
-        cp ${{ inputs.target-file }} ${{ inputs.artifacts-dir }}
+        mkdir -p ${{ inputs.artifacts-dir }}
+        IFS=',' read -ra FILES <<< "${{ inputs.target-files }}"
+        for file in "${FILES[@]}"; do
+          cp "$file" ${{ inputs.artifacts-dir }}/
+        done
 
     # The compressed artifacts will use the following layout:
     # greptime-linux-amd64-pyo3-v0.3.0sha256sum

.github/workflows/dev-build.yml

@@ -29,7 +29,7 @@ on:
       linux_arm64_runner:
         type: choice
         description: The runner uses to build linux-arm64 artifacts
-        default: ec2-c6g.8xlarge-arm64
+        default: ec2-c6g.4xlarge-arm64
         options:
           - ec2-c6g.xlarge-arm64 # 4C8G
           - ec2-c6g.2xlarge-arm64 # 8C16G

.github/workflows/nightly-build.yml

@@ -27,7 +27,7 @@ on:
       linux_arm64_runner:
         type: choice
         description: The runner uses to build linux-arm64 artifacts
-        default: ec2-c6g.8xlarge-arm64
+        default: ec2-c6g.4xlarge-arm64
         options:
           - ec2-c6g.xlarge-arm64 # 4C8G
           - ec2-c6g.2xlarge-arm64 # 8C16G

.github/workflows/nightly-ci.yml

@@ -117,7 +117,6 @@ jobs:
   cleanbuild-linux-nix:
     runs-on: ubuntu-latest-8-cores
     timeout-minutes: 60
-    needs: [coverage, fmt, clippy, check]
     steps:
       - uses: actions/checkout@v4
       - uses: cachix/install-nix-action@v27

.github/workflows/release.yml

@@ -91,7 +91,7 @@ env:
   # The scheduled version is '${{ env.NEXT_RELEASE_VERSION }}-nightly-YYYYMMDD', like v0.2.0-nigthly-20230313;
   NIGHTLY_RELEASE_PREFIX: nightly
   # Note: The NEXT_RELEASE_VERSION should be modified manually by every formal release.
-  NEXT_RELEASE_VERSION: v0.11.0
+  NEXT_RELEASE_VERSION: v0.12.0
 
 # Permission reference: https://docs.github.com/en/actions/using-jobs/assigning-permissions-to-jobs
 permissions:

Cargo.lock

@@ -188,7 +188,7 @@ checksum = "d301b3b94cb4b2f23d7917810addbbaff90738e0ca2be692bd027e70d7e0330c"
 
 [[package]]
 name = "api"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "common-base",
  "common-decimal",
@@ -773,7 +773,7 @@ dependencies = [
 
 [[package]]
 name = "auth"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "async-trait",
@@ -896,18 +896,6 @@ dependencies = [
  "rand",
 ]
 
-[[package]]
-name = "backon"
-version = "0.4.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d67782c3f868daa71d3533538e98a8e13713231969def7536e8039606fc46bf0"
-dependencies = [
- "fastrand",
- "futures-core",
- "pin-project",
- "tokio",
-]
-
 [[package]]
 name = "backon"
 version = "1.2.0"
@@ -1326,7 +1314,7 @@ dependencies = [
 
 [[package]]
 name = "cache"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "catalog",
  "common-error",
@@ -1360,7 +1348,7 @@ checksum = "37b2a672a2cb129a2e41c10b1224bb368f9f37a2b16b612598138befd7b37eb5"
 
 [[package]]
 name = "catalog"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "arrow",
@@ -1696,7 +1684,7 @@ checksum = "1462739cb27611015575c0c11df5df7601141071f07518d56fcc1be504cbec97"
 
 [[package]]
 name = "cli"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "async-trait",
  "auth",
@@ -1739,7 +1727,7 @@ dependencies = [
  "session",
  "snafu 0.8.5",
  "store-api",
- "substrait 0.11.1",
+ "substrait 0.12.0",
  "table",
  "tempfile",
  "tokio",
@@ -1748,7 +1736,7 @@ dependencies = [
 
 [[package]]
 name = "client"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "arc-swap",
@@ -1775,7 +1763,7 @@ dependencies = [
  "rand",
  "serde_json",
  "snafu 0.8.5",
- "substrait 0.11.1",
+ "substrait 0.12.0",
  "substrait 0.37.3",
  "tokio",
  "tokio-stream",
@@ -1816,7 +1804,7 @@ dependencies = [
 
 [[package]]
 name = "cmd"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "async-trait",
  "auth",
@@ -1876,7 +1864,7 @@ dependencies = [
  "similar-asserts",
  "snafu 0.8.5",
  "store-api",
- "substrait 0.11.1",
+ "substrait 0.12.0",
  "table",
  "temp-env",
  "tempfile",
@@ -1928,7 +1916,7 @@ checksum = "55b672471b4e9f9e95499ea597ff64941a309b2cdbffcc46f2cc5e2d971fd335"
 
 [[package]]
 name = "common-base"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "anymap2",
  "async-trait",
@@ -1950,11 +1938,11 @@ dependencies = [
 
 [[package]]
 name = "common-catalog"
-version = "0.11.1"
+version = "0.12.0"
 
 [[package]]
 name = "common-config"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "common-base",
  "common-error",
@@ -1977,7 +1965,7 @@ dependencies = [
 
 [[package]]
 name = "common-datasource"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "arrow",
  "arrow-schema",
@@ -2013,7 +2001,7 @@ dependencies = [
 
 [[package]]
 name = "common-decimal"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "bigdecimal 0.4.5",
  "common-error",
@@ -2026,7 +2014,7 @@ dependencies = [
 
 [[package]]
 name = "common-error"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "snafu 0.8.5",
  "strum 0.25.0",
@@ -2035,7 +2023,7 @@ dependencies = [
 
 [[package]]
 name = "common-frontend"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "async-trait",
  "common-error",
@@ -2045,7 +2033,7 @@ dependencies = [
 
 [[package]]
 name = "common-function"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "approx 0.5.1",
@@ -2089,7 +2077,7 @@ dependencies = [
 
 [[package]]
 name = "common-greptimedb-telemetry"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "async-trait",
  "common-runtime",
@@ -2106,7 +2094,7 @@ dependencies = [
 
 [[package]]
 name = "common-grpc"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "arrow-flight",
@@ -2132,7 +2120,7 @@ dependencies = [
 
 [[package]]
 name = "common-grpc-expr"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "common-base",
@@ -2151,7 +2139,7 @@ dependencies = [
 
 [[package]]
 name = "common-macro"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "arc-swap",
  "common-query",
@@ -2165,7 +2153,7 @@ dependencies = [
 
 [[package]]
 name = "common-mem-prof"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "common-error",
  "common-macro",
@@ -2178,7 +2166,7 @@ dependencies = [
 
 [[package]]
 name = "common-meta"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "anymap2",
  "api",
@@ -2235,7 +2223,7 @@ dependencies = [
 
 [[package]]
 name = "common-options"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "common-grpc",
  "humantime-serde",
@@ -2244,11 +2232,11 @@ dependencies = [
 
 [[package]]
 name = "common-plugins"
-version = "0.11.1"
+version = "0.12.0"
 
 [[package]]
 name = "common-pprof"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "common-error",
  "common-macro",
@@ -2260,11 +2248,11 @@ dependencies = [
 
 [[package]]
 name = "common-procedure"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "async-stream",
  "async-trait",
- "backon 1.2.0",
+ "backon",
  "common-base",
  "common-error",
  "common-macro",
@@ -2287,7 +2275,7 @@ dependencies = [
 
 [[package]]
 name = "common-procedure-test"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "async-trait",
  "common-procedure",
@@ -2295,7 +2283,7 @@ dependencies = [
 
 [[package]]
 name = "common-query"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "async-trait",
@@ -2321,7 +2309,7 @@ dependencies = [
 
 [[package]]
 name = "common-recordbatch"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "arc-swap",
  "common-error",
@@ -2340,7 +2328,7 @@ dependencies = [
 
 [[package]]
 name = "common-runtime"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "async-trait",
  "clap 4.5.19",
@@ -2368,7 +2356,7 @@ dependencies = [
 
 [[package]]
 name = "common-telemetry"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "atty",
  "backtrace",
@@ -2396,7 +2384,7 @@ dependencies = [
 
 [[package]]
 name = "common-test-util"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "client",
  "common-query",
@@ -2408,7 +2396,7 @@ dependencies = [
 
 [[package]]
 name = "common-time"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "arrow",
  "chrono",
@@ -2426,7 +2414,7 @@ dependencies = [
 
 [[package]]
 name = "common-version"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "build-data",
  "const_format",
@@ -2436,7 +2424,7 @@ dependencies = [
 
 [[package]]
 name = "common-wal"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "common-base",
  "common-error",
@@ -3235,7 +3223,7 @@ dependencies = [
 
 [[package]]
 name = "datanode"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "arrow-flight",
@@ -3286,7 +3274,7 @@ dependencies = [
  "session",
  "snafu 0.8.5",
  "store-api",
- "substrait 0.11.1",
+ "substrait 0.12.0",
  "table",
  "tokio",
  "toml 0.8.19",
@@ -3295,7 +3283,7 @@ dependencies = [
 
 [[package]]
 name = "datatypes"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "arrow",
  "arrow-array",
@@ -3919,7 +3907,7 @@ dependencies = [
 
 [[package]]
 name = "file-engine"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "async-trait",
@@ -4035,7 +4023,7 @@ checksum = "8bf7cc16383c4b8d58b9905a8509f02926ce3058053c056376248d958c9df1e8"
 
 [[package]]
 name = "flow"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "arrow",
@@ -4093,7 +4081,7 @@ dependencies = [
  "snafu 0.8.5",
  "store-api",
  "strum 0.25.0",
- "substrait 0.11.1",
+ "substrait 0.12.0",
  "table",
  "tokio",
  "tonic 0.11.0",
@@ -4131,7 +4119,7 @@ checksum = "6c2141d6d6c8512188a7891b4b01590a45f6dac67afb4f255c4124dbb86d4eaa"
 
 [[package]]
 name = "frontend"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "arc-swap",
@@ -4164,6 +4152,7 @@ dependencies = [
  "futures",
  "humantime-serde",
  "lazy_static",
+ "log-query",
  "log-store",
  "meta-client",
  "opentelemetry-proto 0.5.0",
@@ -5280,7 +5269,7 @@ dependencies = [
 
 [[package]]
 name = "index"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "async-trait",
  "asynchronous-codec",
@@ -6129,18 +6118,19 @@ checksum = "a7a70ba024b9dc04c27ea2f0c0548feb474ec5c54bba33a7f72f873a39d07b24"
 
 [[package]]
 name = "log-query"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "chrono",
  "common-error",
  "common-macro",
+ "serde",
  "snafu 0.8.5",
  "table",
 ]
 
 [[package]]
 name = "log-store"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "async-stream",
  "async-trait",
@@ -6484,7 +6474,7 @@ dependencies = [
 
 [[package]]
 name = "meta-client"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "async-trait",
@@ -6511,7 +6501,7 @@ dependencies = [
 
 [[package]]
 name = "meta-srv"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "async-trait",
@@ -6590,7 +6580,7 @@ dependencies = [
 
 [[package]]
 name = "metric-engine"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "aquamarine",
@@ -6684,7 +6674,7 @@ dependencies = [
 
 [[package]]
 name = "mito2"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "aquamarine",
@@ -7421,7 +7411,7 @@ dependencies = [
 
 [[package]]
 name = "object-store"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "anyhow",
  "bytes",
@@ -7481,13 +7471,12 @@ checksum = "b410bbe7e14ab526a0e86877eb47c6996a2bd7746f027ba551028c925390e4e9"
 
 [[package]]
 name = "opendal"
-version = "0.49.2"
+version = "0.50.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
+source = "git+https://github.com/GreptimeTeam/opendal.git?rev=c82605177f2feec83e49dcaa537c505639d94024#c82605177f2feec83e49dcaa537c505639d94024"
-checksum = "9b04d09b9822c2f75a1d2fc513a2c1279c70e91e7407936fffdf6a6976ec530a"
 dependencies = [
  "anyhow",
  "async-trait",
- "backon 0.4.4",
+ "backon",
  "base64 0.22.1",
  "bytes",
  "chrono",
@@ -7500,6 +7489,7 @@ dependencies = [
  "md-5",
  "once_cell",
  "percent-encoding",
+ "prometheus",
  "quick-xml 0.36.2",
  "reqsign",
  "reqwest",
@@ -7674,7 +7664,7 @@ dependencies = [
 
 [[package]]
 name = "operator"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "ahash 0.8.11",
  "api",
@@ -7722,7 +7712,7 @@ dependencies = [
  "sql",
  "sqlparser 0.45.0 (git+https://github.com/GreptimeTeam/sqlparser-rs.git?rev=54a267ac89c09b11c0c88934690530807185d3e7)",
  "store-api",
- "substrait 0.11.1",
+ "substrait 0.12.0",
  "table",
  "tokio",
  "tokio-util",
@@ -7972,7 +7962,7 @@ dependencies = [
 
 [[package]]
 name = "partition"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "async-trait",
@@ -8171,7 +8161,7 @@ dependencies = [
  "rand",
  "ring 0.17.8",
  "rust_decimal",
- "thiserror 2.0.4",
+ "thiserror 2.0.6",
  "tokio",
  "tokio-rustls 0.26.0",
  "tokio-util",
@@ -8258,7 +8248,7 @@ checksum = "8b870d8c151b6f2fb93e84a13146138f05d02ed11c7e7c54f8826aaaf7c9f184"
 
 [[package]]
 name = "pipeline"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "ahash 0.8.11",
  "api",
@@ -8420,7 +8410,7 @@ dependencies = [
 
 [[package]]
 name = "plugins"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "auth",
  "clap 4.5.19",
@@ -8708,7 +8698,7 @@ dependencies = [
 
 [[package]]
 name = "promql"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "ahash 0.8.11",
  "async-trait",
@@ -8943,7 +8933,7 @@ dependencies = [
 
 [[package]]
 name = "puffin"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "async-compression 0.4.13",
  "async-trait",
@@ -9068,7 +9058,7 @@ dependencies = [
 
 [[package]]
 name = "query"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "ahash 0.8.11",
  "api",
@@ -9109,6 +9099,7 @@ dependencies = [
  "humantime",
  "itertools 0.10.5",
  "lazy_static",
+ "log-query",
  "meter-core",
  "meter-macros",
  "num",
@@ -9131,7 +9122,7 @@ dependencies = [
  "sqlparser 0.45.0 (git+https://github.com/GreptimeTeam/sqlparser-rs.git?rev=54a267ac89c09b11c0c88934690530807185d3e7)",
  "statrs",
  "store-api",
- "substrait 0.11.1",
+ "substrait 0.12.0",
  "table",
  "tokio",
  "tokio-stream",
@@ -9515,9 +9506,9 @@ dependencies = [
 
 [[package]]
 name = "reqsign"
-version = "0.16.0"
+version = "0.16.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "03dd4ba7c3901dd43e6b8c7446a760d45bc1ea4301002e1a6fa48f97c3a796fa"
+checksum = "eb0075a66c8bfbf4cc8b70dca166e722e1f55a3ea9250ecbb85f4d92a5f64149"
 dependencies = [
  "anyhow",
  "async-trait",
@@ -10615,7 +10606,7 @@ checksum = "94143f37725109f92c262ed2cf5e59bce7498c01bcc1502d7b9afe439a4e9f49"
 
 [[package]]
 name = "script"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "arc-swap",
@@ -10907,7 +10898,7 @@ dependencies = [
 
 [[package]]
 name = "servers"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "ahash 0.8.11",
  "api",
@@ -10963,6 +10954,7 @@ dependencies = [
  "json5",
  "jsonb",
  "lazy_static",
+ "log-query",
  "loki-api",
  "mime_guess",
  "mysql_async",
@@ -11018,7 +11010,7 @@ dependencies = [
 
 [[package]]
 name = "session"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "arc-swap",
@@ -11372,7 +11364,7 @@ dependencies = [
 
 [[package]]
 name = "sql"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "chrono",
@@ -11436,7 +11428,7 @@ dependencies = [
 
 [[package]]
 name = "sqlness-runner"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "async-trait",
  "clap 4.5.19",
@@ -11654,7 +11646,7 @@ dependencies = [
 
 [[package]]
 name = "store-api"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "aquamarine",
@@ -11816,7 +11808,7 @@ dependencies = [
 
 [[package]]
 name = "substrait"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "async-trait",
  "bytes",
@@ -12015,7 +12007,7 @@ dependencies = [
 
 [[package]]
 name = "table"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "async-trait",
@@ -12292,7 +12284,7 @@ checksum = "3369f5ac52d5eb6ab48c6b4ffdc8efbcad6b89c765749064ba298f2c68a16a76"
 
 [[package]]
 name = "tests-fuzz"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "arbitrary",
  "async-trait",
@@ -12335,7 +12327,7 @@ dependencies = [
 
 [[package]]
 name = "tests-integration"
-version = "0.11.1"
+version = "0.12.0"
 dependencies = [
  "api",
  "arrow-flight",
@@ -12399,7 +12391,7 @@ dependencies = [
  "sql",
  "sqlx",
  "store-api",
- "substrait 0.11.1",
+ "substrait 0.12.0",
  "table",
  "tempfile",
  "time",
@@ -12445,11 +12437,11 @@ dependencies = [
 
 [[package]]
 name = "thiserror"
-version = "2.0.4"
+version = "2.0.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2f49a1853cf82743e3b7950f77e0f4d622ca36cf4317cba00c767838bac8d490"
+checksum = "8fec2a1820ebd077e2b90c4df007bebf344cd394098a13c563957d0afc83ea47"
 dependencies = [
- "thiserror-impl 2.0.4",
+ "thiserror-impl 2.0.6",
 ]
 
 [[package]]
@@ -12465,9 +12457,9 @@ dependencies = [
 
 [[package]]
 name = "thiserror-impl"
-version = "2.0.4"
+version = "2.0.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8381894bb3efe0c4acac3ded651301ceee58a15d47c2e34885ed1908ad667061"
+checksum = "d65750cab40f4ff1929fb1ba509e9914eb756131cef4210da8d5d700d26f6312"
 dependencies = [
  "proc-macro2",
  "quote",

Cargo.toml

@@ -68,7 +68,7 @@ members = [
 resolver = "2"
 
 [workspace.package]
-version = "0.11.1"
+version = "0.12.0"
 edition = "2021"
 license = "Apache-2.0"
 
@@ -238,6 +238,7 @@ file-engine = { path = "src/file-engine" }
 flow = { path = "src/flow" }
 frontend = { path = "src/frontend", default-features = false }
 index = { path = "src/index" }
+log-query = { path = "src/log-query" }
 log-store = { path = "src/log-store" }
 meta-client = { path = "src/meta-client" }
 meta-srv = { path = "src/meta-srv" }

README.md

@@ -70,23 +70,23 @@ Our core developers have been building time-series data platforms for years. Bas
 
 * **Unified Processing of Metrics, Logs, and Events**
 
-GreptimeDB unifies time series data processing by treating all data - whether metrics, logs, or events - as timestamped events with context. Users can analyze this data using either [SQL](https://docs.greptime.com/user-guide/query-data/sql) or [PromQL](https://docs.greptime.com/user-guide/query-data/promql) and leverage stream processing ([Flow](https://docs.greptime.com/user-guide/continuous-aggregation/overview)) to enable continuous aggregation. [Read more](https://docs.greptime.com/user-guide/concepts/data-model).
+  GreptimeDB unifies time series data processing by treating all data - whether metrics, logs, or events - as timestamped events with context. Users can analyze this data using either [SQL](https://docs.greptime.com/user-guide/query-data/sql) or [PromQL](https://docs.greptime.com/user-guide/query-data/promql) and leverage stream processing ([Flow](https://docs.greptime.com/user-guide/flow-computation/overview)) to enable continuous aggregation. [Read more](https://docs.greptime.com/user-guide/concepts/data-model).
 
 * **Cloud-native Distributed Database**
 
-Built for [Kubernetes](https://docs.greptime.com/user-guide/deployments/deploy-on-kubernetes/greptimedb-operator-management). GreptimeDB achieves seamless scalability with its [cloud-native architecture](https://docs.greptime.com/user-guide/concepts/architecture) of separated compute and storage, built on object storage (AWS S3, Azure Blob Storage, etc.) while enabling cross-cloud deployment through a unified data access layer.
+  Built for [Kubernetes](https://docs.greptime.com/user-guide/deployments/deploy-on-kubernetes/greptimedb-operator-management). GreptimeDB achieves seamless scalability with its [cloud-native architecture](https://docs.greptime.com/user-guide/concepts/architecture) of separated compute and storage, built on object storage (AWS S3, Azure Blob Storage, etc.) while enabling cross-cloud deployment through a unified data access layer.
 
 * **Performance and Cost-effective**
 
-Written in pure Rust for superior performance and reliability. GreptimeDB features a distributed query engine with intelligent indexing to handle high cardinality data efficiently. Its optimized columnar storage achieves 50x cost efficiency on cloud object storage through advanced compression. [Benchmark reports](https://www.greptime.com/blogs/2024-09-09-report-summary).
+  Written in pure Rust for superior performance and reliability. GreptimeDB features a distributed query engine with intelligent indexing to handle high cardinality data efficiently. Its optimized columnar storage achieves 50x cost efficiency on cloud object storage through advanced compression. [Benchmark reports](https://www.greptime.com/blogs/2024-09-09-report-summary).
 
 * **Cloud-Edge Collaboration**
 
-GreptimeDB seamlessly operates across cloud and edge (ARM/Android/Linux), providing consistent APIs and control plane for unified data management and efficient synchronization. [Learn how to run on Android](https://docs.greptime.com/user-guide/deployments/run-on-android/).
+  GreptimeDB seamlessly operates across cloud and edge (ARM/Android/Linux), providing consistent APIs and control plane for unified data management and efficient synchronization. [Learn how to run on Android](https://docs.greptime.com/user-guide/deployments/run-on-android/).
 
 * **Multi-protocol Ingestion, SQL & PromQL Ready**
 
-Widely adopted database protocols and APIs, including MySQL, PostgreSQL, InfluxDB, OpenTelemetry, Loki and Prometheus, etc.  Effortless Adoption & Seamless Migration. [Supported Protocols Overview](https://docs.greptime.com/user-guide/protocols/overview).
+  Widely adopted database protocols and APIs, including MySQL, PostgreSQL, InfluxDB, OpenTelemetry, Loki and Prometheus, etc.  Effortless Adoption & Seamless Migration. [Supported Protocols Overview](https://docs.greptime.com/user-guide/protocols/overview).
 
 For more detailed info please read  [Why GreptimeDB](https://docs.greptime.com/user-guide/concepts/why-greptimedb).
 
@@ -138,7 +138,7 @@ Check the prerequisite:
 
 * [Rust toolchain](https://www.rust-lang.org/tools/install) (nightly)
 * [Protobuf compiler](https://grpc.io/docs/protoc-installation/) (>= 3.15)
-* Python toolchain (optional): Required only if built with PyO3 backend. More detail for compiling with PyO3 can be found in its [documentation](https://pyo3.rs/v0.18.1/building_and_distribution#configuring-the-python-version).
+* Python toolchain (optional): Required only if built with PyO3 backend. More details for compiling with PyO3 can be found in its [documentation](https://pyo3.rs/v0.18.1/building_and_distribution#configuring-the-python-version).
 
 Build GreptimeDB binary:
 
@@ -154,6 +154,10 @@ cargo run -- standalone start
 
 ## Tools & Extensions
 
+### Kubernetes
+
+- [GreptimeDB Operator](https://github.com/GrepTimeTeam/greptimedb-operator)
+
 ### Dashboard
 
 - [The dashboard UI for GreptimeDB](https://github.com/GreptimeTeam/dashboard)
@@ -173,7 +177,7 @@ Our official Grafana dashboard for monitoring GreptimeDB is available at [grafan
 
 ## Project Status
 
-GreptimeDB is currently in Beta. We are targeting GA (General Availability) with v1.0 release by Early 2025. 
+GreptimeDB is currently in Beta. We are targeting GA (General Availability) with v1.0 release by Early 2025.
 
 While in Beta, GreptimeDB is already:
 

src/auth/src/permission.rs

@@ -25,6 +25,7 @@ pub enum PermissionReq<'a> {
     GrpcRequest(&'a Request),
     SqlStatement(&'a Statement),
     PromQuery,
+    LogQuery,
     Opentsdb,
     LineProtocol,
     PromStoreWrite,

src/catalog/src/kvbackend/table_cache.rs

@@ -38,7 +38,7 @@ pub fn new_table_cache(
 ) -> TableCache {
     let init = init_factory(table_info_cache, table_name_cache);
 
-    CacheContainer::new(name, cache, Box::new(invalidator), init, Box::new(filter))
+    CacheContainer::new(name, cache, Box::new(invalidator), init, filter)
 }
 
 fn init_factory(

src/common/datasource/src/object_store/fs.rs

@@ -27,7 +27,7 @@ pub fn build_fs_backend(root: &str) -> Result<ObjectStore> {
             DefaultLoggingInterceptor,
         ))
         .layer(object_store::layers::TracingLayer)
-        .layer(object_store::layers::PrometheusMetricsLayer::new(true))
+        .layer(object_store::layers::build_prometheus_metrics_layer(true))
         .finish();
     Ok(object_store)
 }

src/common/datasource/src/object_store/s3.rs

@@ -89,7 +89,7 @@ pub fn build_s3_backend(
             DefaultLoggingInterceptor,
         ))
         .layer(object_store::layers::TracingLayer)
-        .layer(object_store::layers::PrometheusMetricsLayer::new(true))
+        .layer(object_store::layers::build_prometheus_metrics_layer(true))
         .finish())
 }
 

src/common/function/src/lib.rs

@@ -26,3 +26,4 @@ pub mod function_registry;
 pub mod handlers;
 pub mod helper;
 pub mod state;
+pub mod utils;

src/common/function/src/scalars/matches.rs

@@ -204,20 +204,10 @@ impl PatternAst {
     fn convert_literal(column: &str, pattern: &str) -> Expr {
         logical_expr::col(column).like(logical_expr::lit(format!(
             "%{}%",
-            Self::escape_pattern(pattern)
+            crate::utils::escape_like_pattern(pattern)
         )))
     }
 
-    fn escape_pattern(pattern: &str) -> String {
-        pattern
-            .chars()
-            .flat_map(|c| match c {
-                '\\' | '%' | '_' => vec!['\\', c],
-                _ => vec![c],
-            })
-            .collect::<String>()
-    }
-
     /// Transform this AST with preset rules to make it correct.
     fn transform_ast(self) -> Result<Self> {
         self.transform_up(Self::collapse_binary_branch_fn)

src/common/function/src/utils.rs

@@ -0,0 +1,58 @@
+// 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.
+
+/// Escapes special characters in the provided pattern string for `LIKE`.
+///
+/// Specifically, it prefixes the backslash (`\`), percent (`%`), and underscore (`_`)
+/// characters with an additional backslash to ensure they are treated literally.
+///
+/// # Examples
+///
+/// ```rust
+/// let escaped = escape_pattern("100%_some\\path");
+/// assert_eq!(escaped, "100\\%\\_some\\\\path");
+/// ```
+pub fn escape_like_pattern(pattern: &str) -> String {
+    pattern
+        .chars()
+        .flat_map(|c| match c {
+            '\\' | '%' | '_' => vec!['\\', c],
+            _ => vec![c],
+        })
+        .collect::<String>()
+}
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_escape_like_pattern() {
+        assert_eq!(
+            escape_like_pattern("100%_some\\path"),
+            "100\\%\\_some\\\\path"
+        );
+        assert_eq!(escape_like_pattern(""), "");
+        assert_eq!(escape_like_pattern("hello"), "hello");
+        assert_eq!(escape_like_pattern("\\%_"), "\\\\\\%\\_");
+        assert_eq!(escape_like_pattern("%%__\\\\"), "\\%\\%\\_\\_\\\\\\\\");
+        assert_eq!(escape_like_pattern("abc123"), "abc123");
+        assert_eq!(escape_like_pattern("%_\\"), "\\%\\_\\\\");
+        assert_eq!(
+            escape_like_pattern("%%__\\\\another%string"),
+            "\\%\\%\\_\\_\\\\\\\\another\\%string"
+        );
+        assert_eq!(escape_like_pattern("foo%bar_"), "foo\\%bar\\_");
+        assert_eq!(escape_like_pattern("\\_\\%"), "\\\\\\_\\\\\\%");
+    }
+}

src/common/meta/src/cache/container.rs

@@ -43,7 +43,7 @@ pub struct CacheContainer<K, V, CacheToken> {
     cache: Cache<K, V>,
     invalidator: Invalidator<K, V, CacheToken>,
     initializer: Initializer<K, V>,
-    token_filter: TokenFilter<CacheToken>,
+    token_filter: fn(&CacheToken) -> bool,
 }
 
 impl<K, V, CacheToken> CacheContainer<K, V, CacheToken>
@@ -58,7 +58,7 @@ where
         cache: Cache<K, V>,
         invalidator: Invalidator<K, V, CacheToken>,
         initializer: Initializer<K, V>,
-        token_filter: TokenFilter<CacheToken>,
+        token_filter: fn(&CacheToken) -> bool,
     ) -> Self {
         Self {
             name,
@@ -206,10 +206,13 @@ mod tests {
         name: &'a str,
     }
 
+    fn always_true_filter(_: &String) -> bool {
+        true
+    }
+
     #[tokio::test]
     async fn test_get() {
         let cache: Cache<NameKey, String> = CacheBuilder::new(128).build();
-        let filter: TokenFilter<String> = Box::new(|_| true);
         let counter = Arc::new(AtomicI32::new(0));
         let moved_counter = counter.clone();
         let init: Initializer<NameKey, String> = Arc::new(move |_| {
@@ -219,7 +222,13 @@ mod tests {
         let invalidator: Invalidator<NameKey, String, String> =
             Box::new(|_, _| Box::pin(async { Ok(()) }));
 
-        let adv_cache = CacheContainer::new("test".to_string(), cache, invalidator, init, filter);
+        let adv_cache = CacheContainer::new(
+            "test".to_string(),
+            cache,
+            invalidator,
+            init,
+            always_true_filter,
+        );
         let key = NameKey { name: "key" };
         let value = adv_cache.get(key).await.unwrap().unwrap();
         assert_eq!(value, "hi");
@@ -233,7 +242,6 @@ mod tests {
     #[tokio::test]
     async fn test_get_by_ref() {
         let cache: Cache<String, String> = CacheBuilder::new(128).build();
-        let filter: TokenFilter<String> = Box::new(|_| true);
         let counter = Arc::new(AtomicI32::new(0));
         let moved_counter = counter.clone();
         let init: Initializer<String, String> = Arc::new(move |_| {
@@ -243,7 +251,13 @@ mod tests {
         let invalidator: Invalidator<String, String, String> =
             Box::new(|_, _| Box::pin(async { Ok(()) }));
 
-        let adv_cache = CacheContainer::new("test".to_string(), cache, invalidator, init, filter);
+        let adv_cache = CacheContainer::new(
+            "test".to_string(),
+            cache,
+            invalidator,
+            init,
+            always_true_filter,
+        );
         let value = adv_cache.get_by_ref("foo").await.unwrap().unwrap();
         assert_eq!(value, "hi");
         let value = adv_cache.get_by_ref("foo").await.unwrap().unwrap();
@@ -257,13 +271,18 @@ mod tests {
     #[tokio::test]
     async fn test_get_value_not_exits() {
         let cache: Cache<String, String> = CacheBuilder::new(128).build();
-        let filter: TokenFilter<String> = Box::new(|_| true);
         let init: Initializer<String, String> =
             Arc::new(move |_| Box::pin(async { error::ValueNotExistSnafu {}.fail() }));
         let invalidator: Invalidator<String, String, String> =
             Box::new(|_, _| Box::pin(async { Ok(()) }));
 
-        let adv_cache = CacheContainer::new("test".to_string(), cache, invalidator, init, filter);
+        let adv_cache = CacheContainer::new(
+            "test".to_string(),
+            cache,
+            invalidator,
+            init,
+            always_true_filter,
+        );
         let value = adv_cache.get_by_ref("foo").await.unwrap();
         assert!(value.is_none());
     }
@@ -271,7 +290,6 @@ mod tests {
     #[tokio::test]
     async fn test_invalidate() {
         let cache: Cache<String, String> = CacheBuilder::new(128).build();
-        let filter: TokenFilter<String> = Box::new(|_| true);
         let counter = Arc::new(AtomicI32::new(0));
         let moved_counter = counter.clone();
         let init: Initializer<String, String> = Arc::new(move |_| {
@@ -285,7 +303,13 @@ mod tests {
             })
         });
 
-        let adv_cache = CacheContainer::new("test".to_string(), cache, invalidator, init, filter);
+        let adv_cache = CacheContainer::new(
+            "test".to_string(),
+            cache,
+            invalidator,
+            init,
+            always_true_filter,
+        );
         let value = adv_cache.get_by_ref("foo").await.unwrap().unwrap();
         assert_eq!(value, "hi");
         let value = adv_cache.get_by_ref("foo").await.unwrap().unwrap();

src/common/meta/src/cache/flow/table_flownode.rs

@@ -45,7 +45,7 @@ pub fn new_table_flownode_set_cache(
     let table_flow_manager = Arc::new(TableFlowManager::new(kv_backend));
     let init = init_factory(table_flow_manager);
 
-    CacheContainer::new(name, cache, Box::new(invalidator), init, Box::new(filter))
+    CacheContainer::new(name, cache, Box::new(invalidator), init, filter)
 }
 
 fn init_factory(table_flow_manager: TableFlowManagerRef) -> Initializer<TableId, FlownodeSet> {

src/common/meta/src/cache/registry.rs

@@ -151,12 +151,15 @@ mod tests {
     use crate::cache::*;
     use crate::instruction::CacheIdent;
 
+    fn always_true_filter(_: &CacheIdent) -> bool {
+        true
+    }
+
     fn test_cache(
         name: &str,
         invalidator: Invalidator<String, String, CacheIdent>,
     ) -> CacheContainer<String, String, CacheIdent> {
         let cache: Cache<String, String> = CacheBuilder::new(128).build();
-        let filter: TokenFilter<CacheIdent> = Box::new(|_| true);
         let counter = Arc::new(AtomicI32::new(0));
         let moved_counter = counter.clone();
         let init: Initializer<String, String> = Arc::new(move |_| {
@@ -164,7 +167,13 @@ mod tests {
             Box::pin(async { Ok(Some("hi".to_string())) })
         });
 
-        CacheContainer::new(name.to_string(), cache, invalidator, init, filter)
+        CacheContainer::new(
+            name.to_string(),
+            cache,
+            invalidator,
+            init,
+            always_true_filter,
+        )
     }
 
     fn test_i32_cache(
@@ -172,7 +181,6 @@ mod tests {
         invalidator: Invalidator<i32, String, CacheIdent>,
     ) -> CacheContainer<i32, String, CacheIdent> {
         let cache: Cache<i32, String> = CacheBuilder::new(128).build();
-        let filter: TokenFilter<CacheIdent> = Box::new(|_| true);
         let counter = Arc::new(AtomicI32::new(0));
         let moved_counter = counter.clone();
         let init: Initializer<i32, String> = Arc::new(move |_| {
@@ -180,7 +188,13 @@ mod tests {
             Box::pin(async { Ok(Some("foo".to_string())) })
         });
 
-        CacheContainer::new(name.to_string(), cache, invalidator, init, filter)
+        CacheContainer::new(
+            name.to_string(),
+            cache,
+            invalidator,
+            init,
+            always_true_filter,
+        )
     }
 
     #[tokio::test]

src/common/meta/src/cache/table/schema.rs

@@ -36,7 +36,7 @@ pub fn new_schema_cache(
     let schema_manager = SchemaManager::new(kv_backend.clone());
     let init = init_factory(schema_manager);
 
-    CacheContainer::new(name, cache, Box::new(invalidator), init, Box::new(filter))
+    CacheContainer::new(name, cache, Box::new(invalidator), init, filter)
 }
 
 fn init_factory(schema_manager: SchemaManager) -> Initializer<SchemaName, Arc<SchemaNameValue>> {

src/common/meta/src/cache/table/table_info.rs

@@ -41,7 +41,7 @@ pub fn new_table_info_cache(
     let table_info_manager = Arc::new(TableInfoManager::new(kv_backend));
     let init = init_factory(table_info_manager);
 
-    CacheContainer::new(name, cache, Box::new(invalidator), init, Box::new(filter))
+    CacheContainer::new(name, cache, Box::new(invalidator), init, filter)
 }
 
 fn init_factory(table_info_manager: TableInfoManagerRef) -> Initializer<TableId, Arc<TableInfo>> {

src/common/meta/src/cache/table/table_name.rs

@@ -41,7 +41,7 @@ pub fn new_table_name_cache(
     let table_name_manager = Arc::new(TableNameManager::new(kv_backend));
     let init = init_factory(table_name_manager);
 
-    CacheContainer::new(name, cache, Box::new(invalidator), init, Box::new(filter))
+    CacheContainer::new(name, cache, Box::new(invalidator), init, filter)
 }
 
 fn init_factory(table_name_manager: TableNameManagerRef) -> Initializer<TableName, TableId> {

src/common/meta/src/cache/table/table_route.rs

@@ -65,7 +65,7 @@ pub fn new_table_route_cache(
     let table_info_manager = Arc::new(TableRouteManager::new(kv_backend));
     let init = init_factory(table_info_manager);
 
-    CacheContainer::new(name, cache, Box::new(invalidator), init, Box::new(filter))
+    CacheContainer::new(name, cache, Box::new(invalidator), init, filter)
 }
 
 fn init_factory(

src/common/meta/src/cache/table/table_schema.rs

@@ -40,7 +40,7 @@ pub fn new_table_schema_cache(
     let table_info_manager = TableInfoManager::new(kv_backend);
     let init = init_factory(table_info_manager);
 
-    CacheContainer::new(name, cache, Box::new(invalidator), init, Box::new(filter))
+    CacheContainer::new(name, cache, Box::new(invalidator), init, filter)
 }
 
 fn init_factory(table_info_manager: TableInfoManager) -> Initializer<TableId, Arc<SchemaName>> {

src/common/meta/src/cache/table/view_info.rs

@@ -40,7 +40,7 @@ pub fn new_view_info_cache(
     let view_info_manager = Arc::new(ViewInfoManager::new(kv_backend));
     let init = init_factory(view_info_manager);
 
-    CacheContainer::new(name, cache, Box::new(invalidator), init, Box::new(filter))
+    CacheContainer::new(name, cache, Box::new(invalidator), init, filter)
 }
 
 fn init_factory(view_info_manager: ViewInfoManagerRef) -> Initializer<TableId, Arc<ViewInfoValue>> {

src/common/meta/src/key/catalog_name.rs

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 use std::fmt::Display;
-use std::sync::Arc;
 
 use common_catalog::consts::DEFAULT_CATALOG_NAME;
 use futures::stream::BoxStream;
@@ -146,7 +145,7 @@ impl CatalogManager {
             self.kv_backend.clone(),
             req,
             DEFAULT_PAGE_SIZE,
-            Arc::new(catalog_decoder),
+            catalog_decoder,
         )
         .into_stream();
 
@@ -156,6 +155,8 @@ impl CatalogManager {
 
 #[cfg(test)]
 mod tests {
+    use std::sync::Arc;
+
     use super::*;
     use crate::kv_backend::memory::MemoryKvBackend;
 

src/common/meta/src/key/datanode_table.rs

@@ -14,7 +14,6 @@
 
 use std::collections::HashMap;
 use std::fmt::Display;
-use std::sync::Arc;
 
 use futures::stream::BoxStream;
 use serde::{Deserialize, Serialize};
@@ -166,7 +165,7 @@ impl DatanodeTableManager {
             self.kv_backend.clone(),
             req,
             DEFAULT_PAGE_SIZE,
-            Arc::new(datanode_table_value_decoder),
+            datanode_table_value_decoder,
         )
         .into_stream();
 

src/common/meta/src/key/flow/flow_name.rs

@@ -12,8 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::sync::Arc;
-
 use futures::stream::BoxStream;
 use lazy_static::lazy_static;
 use regex::Regex;
@@ -201,7 +199,7 @@ impl FlowNameManager {
             self.kv_backend.clone(),
             req,
             DEFAULT_PAGE_SIZE,
-            Arc::new(flow_name_decoder),
+            flow_name_decoder,
         )
         .into_stream();
 

src/common/meta/src/key/flow/flow_route.rs

@@ -12,8 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::sync::Arc;
-
 use futures::stream::BoxStream;
 use lazy_static::lazy_static;
 use regex::Regex;
@@ -179,7 +177,7 @@ impl FlowRouteManager {
             self.kv_backend.clone(),
             req,
             DEFAULT_PAGE_SIZE,
-            Arc::new(flow_route_decoder),
+            flow_route_decoder,
         )
         .into_stream();
 

src/common/meta/src/key/flow/flownode_flow.rs

@@ -12,8 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::sync::Arc;
-
 use futures::stream::BoxStream;
 use futures::TryStreamExt;
 use lazy_static::lazy_static;
@@ -179,7 +177,7 @@ impl FlownodeFlowManager {
             self.kv_backend.clone(),
             req,
             DEFAULT_PAGE_SIZE,
-            Arc::new(flownode_flow_key_decoder),
+            flownode_flow_key_decoder,
         )
         .into_stream();
 

src/common/meta/src/key/flow/table_flow.rs

@@ -206,7 +206,7 @@ impl TableFlowManager {
             self.kv_backend.clone(),
             req,
             DEFAULT_PAGE_SIZE,
-            Arc::new(table_flow_decoder),
+            table_flow_decoder,
         )
         .into_stream();
 

src/common/meta/src/key/schema_name.rs

@@ -14,7 +14,6 @@
 
 use std::collections::HashMap;
 use std::fmt::Display;
-use std::sync::Arc;
 
 use common_catalog::consts::{DEFAULT_CATALOG_NAME, DEFAULT_SCHEMA_NAME};
 use common_time::DatabaseTimeToLive;
@@ -283,7 +282,7 @@ impl SchemaManager {
             self.kv_backend.clone(),
             req,
             DEFAULT_PAGE_SIZE,
-            Arc::new(schema_decoder),
+            schema_decoder,
         )
         .into_stream();
 
@@ -308,6 +307,7 @@ impl<'a> From<&'a SchemaName> for SchemaNameKey<'a> {
 
 #[cfg(test)]
 mod tests {
+    use std::sync::Arc;
     use std::time::Duration;
 
     use super::*;

src/common/meta/src/key/table_name.rs

@@ -269,7 +269,7 @@ impl TableNameManager {
             self.kv_backend.clone(),
             req,
             DEFAULT_PAGE_SIZE,
-            Arc::new(table_decoder),
+            table_decoder,
         )
         .into_stream();
 

src/common/meta/src/range_stream.rs

@@ -12,8 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::sync::Arc;
-
 use async_stream::try_stream;
 use common_telemetry::debug;
 use futures::Stream;
@@ -148,7 +146,7 @@ impl PaginationStreamFactory {
 }
 
 pub struct PaginationStream<T> {
-    decoder_fn: Arc<KeyValueDecoderFn<T>>,
+    decoder_fn: fn(KeyValue) -> Result<T>,
     factory: PaginationStreamFactory,
 }
 
@@ -158,7 +156,7 @@ impl<T> PaginationStream<T> {
         kv: KvBackendRef,
         req: RangeRequest,
         page_size: usize,
-        decoder_fn: Arc<KeyValueDecoderFn<T>>,
+        decoder_fn: fn(KeyValue) -> Result<T>,
     ) -> Self {
         Self {
             decoder_fn,
@@ -191,6 +189,7 @@ mod tests {
 
     use std::assert_matches::assert_matches;
     use std::collections::BTreeMap;
+    use std::sync::Arc;
 
     use futures::TryStreamExt;
 
@@ -250,7 +249,7 @@ mod tests {
                 ..Default::default()
             },
             DEFAULT_PAGE_SIZE,
-            Arc::new(decoder),
+            decoder,
         )
         .into_stream();
         let kv = stream.try_collect::<Vec<_>>().await.unwrap();
@@ -290,7 +289,7 @@ mod tests {
                 ..Default::default()
             },
             2,
-            Arc::new(decoder),
+            decoder,
         );
         let kv = stream
             .into_stream()

src/common/meta/src/state_store.rs

@@ -12,8 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::sync::Arc;
-
 use async_trait::async_trait;
 use common_error::ext::BoxedError;
 use common_procedure::error::{DeleteStatesSnafu, ListStateSnafu, PutStateSnafu};
@@ -171,7 +169,7 @@ impl StateStore for KvStateStore {
             self.kv_backend.clone(),
             req,
             self.max_num_per_range_request.unwrap_or_default(),
-            Arc::new(decode_kv),
+            decode_kv,
         )
         .into_stream();
 

src/common/procedure/src/local/runner.rs

@@ -544,7 +544,7 @@ mod tests {
     use common_test_util::temp_dir::create_temp_dir;
     use futures_util::future::BoxFuture;
     use futures_util::FutureExt;
-    use object_store::ObjectStore;
+    use object_store::{EntryMode, ObjectStore};
     use tokio::sync::mpsc;
 
     use super::*;
@@ -578,7 +578,11 @@ mod tests {
     ) {
         let dir = proc_path!(procedure_store, "{procedure_id}/");
         let lister = object_store.list(&dir).await.unwrap();
-        let mut files_in_dir: Vec<_> = lister.into_iter().map(|de| de.name().to_string()).collect();
+        let mut files_in_dir: Vec<_> = lister
+            .into_iter()
+            .filter(|x| x.metadata().mode() == EntryMode::FILE)
+            .map(|de| de.name().to_string())
+            .collect();
         files_in_dir.sort_unstable();
         assert_eq!(files, files_in_dir);
     }

src/datanode/src/error.rs

@@ -193,6 +193,14 @@ pub enum Error {
         location: Location,
     },
 
+    #[snafu(display("Failed to build http client"))]
+    BuildHttpClient {
+        #[snafu(implicit)]
+        location: Location,
+        #[snafu(source)]
+        error: reqwest::Error,
+    },
+
     #[snafu(display("Missing required field: {}", name))]
     MissingRequiredField {
         name: String,
@@ -406,9 +414,10 @@ impl ErrorExt for Error {
             | MissingKvBackend { .. }
             | TomlFormat { .. } => StatusCode::InvalidArguments,
 
-            PayloadNotExist { .. } | Unexpected { .. } | WatchAsyncTaskChange { .. } => {
+            PayloadNotExist { .. }
-                StatusCode::Unexpected
+            | Unexpected { .. }
-            }
+            | WatchAsyncTaskChange { .. }
+            | BuildHttpClient { .. } => StatusCode::Unexpected,
 
             AsyncTaskExecute { source, .. } => source.status_code(),
 

src/datanode/src/store.rs

@@ -32,7 +32,7 @@ use object_store::{Access, Error, HttpClient, ObjectStore, ObjectStoreBuilder, O
 use snafu::prelude::*;
 
 use crate::config::{HttpClientConfig, ObjectStoreConfig, DEFAULT_OBJECT_STORE_CACHE_SIZE};
-use crate::error::{self, CreateDirSnafu, Result};
+use crate::error::{self, BuildHttpClientSnafu, CreateDirSnafu, Result};
 
 pub(crate) async fn new_raw_object_store(
     store: &ObjectStoreConfig,
@@ -236,7 +236,8 @@ pub(crate) fn build_http_client(config: &HttpClientConfig) -> Result<HttpClient>
         builder.timeout(config.timeout)
     };
 
-    HttpClient::build(http_builder).context(error::InitBackendSnafu)
+    let client = http_builder.build().context(BuildHttpClientSnafu)?;
+    Ok(HttpClient::with(client))
 }
 struct PrintDetailedError;
 

src/file-engine/src/manifest.rs

@@ -46,7 +46,7 @@ impl FileRegionManifest {
     pub async fn store(&self, region_dir: &str, object_store: &ObjectStore) -> Result<()> {
         let path = &region_manifest_path(region_dir);
         let exist = object_store
-            .is_exist(path)
+            .exists(path)
             .await
             .context(CheckObjectSnafu { path })?;
         ensure!(!exist, ManifestExistsSnafu { path });

src/file-engine/src/region.rs

@@ -130,7 +130,7 @@ mod tests {
         assert_eq!(region.metadata.primary_key, vec![1]);
 
         assert!(object_store
-            .is_exist("create_region_dir/manifest/_file_manifest")
+            .exists("create_region_dir/manifest/_file_manifest")
             .await
             .unwrap());
 
@@ -198,13 +198,13 @@ mod tests {
             .unwrap();
 
         assert!(object_store
-            .is_exist("drop_region_dir/manifest/_file_manifest")
+            .exists("drop_region_dir/manifest/_file_manifest")
             .await
             .unwrap());
 
         FileRegion::drop(&region, &object_store).await.unwrap();
         assert!(!object_store
-            .is_exist("drop_region_dir/manifest/_file_manifest")
+            .exists("drop_region_dir/manifest/_file_manifest")
             .await
             .unwrap());
 

src/frontend/Cargo.toml

@@ -41,6 +41,7 @@ datafusion-expr.workspace = true
 datanode.workspace = true
 humantime-serde.workspace = true
 lazy_static.workspace = true
+log-query.workspace = true
 log-store.workspace = true
 meta-client.workspace = true
 opentelemetry-proto.workspace = true

src/frontend/src/instance.rs

@@ -16,6 +16,7 @@ pub mod builder;
 mod grpc;
 mod influxdb;
 mod log_handler;
+mod logs;
 mod opentsdb;
 mod otlp;
 mod prom_store;
@@ -64,8 +65,8 @@ use servers::prometheus_handler::PrometheusHandler;
 use servers::query_handler::grpc::GrpcQueryHandler;
 use servers::query_handler::sql::SqlQueryHandler;
 use servers::query_handler::{
-    InfluxdbLineProtocolHandler, OpenTelemetryProtocolHandler, OpentsdbProtocolHandler,
+    InfluxdbLineProtocolHandler, LogQueryHandler, OpenTelemetryProtocolHandler,
-    PipelineHandler, PromStoreProtocolHandler, ScriptHandler,
+    OpentsdbProtocolHandler, PipelineHandler, PromStoreProtocolHandler, ScriptHandler,
 };
 use servers::server::ServerHandlers;
 use session::context::QueryContextRef;
@@ -99,6 +100,7 @@ pub trait FrontendInstance:
     + ScriptHandler
     + PrometheusHandler
     + PipelineHandler
+    + LogQueryHandler
     + Send
     + Sync
     + 'static

src/frontend/src/instance/logs.rs

@@ -0,0 +1,67 @@
+// 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 auth::{PermissionChecker, PermissionCheckerRef, PermissionReq};
+use client::Output;
+use common_error::ext::BoxedError;
+use log_query::LogQuery;
+use server_error::Result as ServerResult;
+use servers::error::{self as server_error, AuthSnafu, ExecuteQuerySnafu};
+use servers::interceptor::{LogQueryInterceptor, LogQueryInterceptorRef};
+use servers::query_handler::LogQueryHandler;
+use session::context::QueryContextRef;
+use snafu::ResultExt;
+use tonic::async_trait;
+
+use super::Instance;
+
+#[async_trait]
+impl LogQueryHandler for Instance {
+    async fn query(&self, mut request: LogQuery, ctx: QueryContextRef) -> ServerResult<Output> {
+        let interceptor = self
+            .plugins
+            .get::<LogQueryInterceptorRef<server_error::Error>>();
+
+        self.plugins
+            .get::<PermissionCheckerRef>()
+            .as_ref()
+            .check_permission(ctx.current_user(), PermissionReq::LogQuery)
+            .context(AuthSnafu)?;
+
+        interceptor.as_ref().pre_query(&request, ctx.clone())?;
+
+        request
+            .time_filter
+            .canonicalize()
+            .map_err(BoxedError::new)
+            .context(ExecuteQuerySnafu)?;
+
+        let plan = self
+            .query_engine
+            .planner()
+            .plan_logs_query(request, ctx.clone())
+            .await
+            .map_err(BoxedError::new)
+            .context(ExecuteQuerySnafu)?;
+
+        let output = self
+            .statement_executor
+            .exec_plan(plan, ctx.clone())
+            .await
+            .map_err(BoxedError::new)
+            .context(ExecuteQuerySnafu)?;
+
+        Ok(interceptor.as_ref().post_query(output, ctx.clone())?)
+    }
+}

src/frontend/src/server.rs

@@ -87,6 +87,7 @@ where
         let ingest_interceptor = self.plugins.get::<LogIngestInterceptorRef<ServerError>>();
         builder =
             builder.with_log_ingest_handler(self.instance.clone(), validator, ingest_interceptor);
+        builder = builder.with_logs_handler(self.instance.clone());
 
         if let Some(user_provider) = self.plugins.get::<UserProviderRef>() {
             builder = builder.with_user_provider(user_provider);

src/index/src/bloom_filter.rs

@@ -15,11 +15,15 @@
 use serde::{Deserialize, Serialize};
 
 pub mod creator;
-mod error;
+pub mod error;
+pub mod reader;
 
 pub type Bytes = Vec<u8>;
 pub type BytesRef<'a> = &'a [u8];
 
+/// The seed used for the Bloom filter.
+pub const SEED: u128 = 42;
+
 /// The Meta information of the bloom filter stored in the file.
 #[derive(Debug, Default, Serialize, Deserialize)]
 pub struct BloomFilterMeta {

src/index/src/bloom_filter/creator.rs

@@ -12,21 +12,23 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::collections::HashSet;
+mod finalize_segment;
+mod intermediate_codec;
 
-use fastbloom::BloomFilter;
+use std::collections::HashSet;
-use futures::{AsyncWrite, AsyncWriteExt};
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::sync::Arc;
+
+use finalize_segment::FinalizedBloomFilterStorage;
+use futures::{AsyncWrite, AsyncWriteExt, StreamExt};
 use snafu::ResultExt;
 
-use super::error::{IoSnafu, SerdeJsonSnafu};
+use crate::bloom_filter::error::{IoSnafu, Result, SerdeJsonSnafu};
-use crate::bloom_filter::error::Result;
+use crate::bloom_filter::{BloomFilterMeta, BloomFilterSegmentLocation, Bytes, SEED};
-use crate::bloom_filter::{BloomFilterMeta, BloomFilterSegmentLocation, Bytes};
+use crate::external_provider::ExternalTempFileProvider;
-
-/// The seed used for the Bloom filter.
-const SEED: u128 = 42;
 
 /// The false positive rate of the Bloom filter.
-const FALSE_POSITIVE_RATE: f64 = 0.01;
+pub const FALSE_POSITIVE_RATE: f64 = 0.01;
 
 /// `BloomFilterCreator` is responsible for creating and managing bloom filters
 /// for a set of elements. It divides the rows into segments and creates
@@ -58,6 +60,9 @@ pub struct BloomFilterCreator {
 
     /// Storage for finalized Bloom filters.
     finalized_bloom_filters: FinalizedBloomFilterStorage,
+
+    /// Global memory usage of the bloom filter creator.
+    global_memory_usage: Arc<AtomicUsize>,
 }
 
 impl BloomFilterCreator {
@@ -66,7 +71,12 @@ impl BloomFilterCreator {
     /// # PANICS
     ///
     /// `rows_per_segment` <= 0
-    pub fn new(rows_per_segment: usize) -> Self {
+    pub fn new(
+        rows_per_segment: usize,
+        intermediate_provider: Box<dyn ExternalTempFileProvider>,
+        global_memory_usage: Arc<AtomicUsize>,
+        global_memory_usage_threshold: Option<usize>,
+    ) -> Self {
         assert!(
             rows_per_segment > 0,
             "rows_per_segment must be greater than 0"
@@ -77,54 +87,67 @@ impl BloomFilterCreator {
             accumulated_row_count: 0,
             cur_seg_distinct_elems: HashSet::default(),
             cur_seg_distinct_elems_mem_usage: 0,
-            finalized_bloom_filters: FinalizedBloomFilterStorage::default(),
+            global_memory_usage: global_memory_usage.clone(),
+            finalized_bloom_filters: FinalizedBloomFilterStorage::new(
+                intermediate_provider,
+                global_memory_usage,
+                global_memory_usage_threshold,
+            ),
         }
     }
 
     /// Adds a row of elements to the bloom filter. If the number of accumulated rows
     /// reaches `rows_per_segment`, it finalizes the current segment.
-    pub fn push_row_elems(&mut self, elems: impl IntoIterator<Item = Bytes>) {
+    pub async fn push_row_elems(&mut self, elems: impl IntoIterator<Item = Bytes>) -> Result<()> {
         self.accumulated_row_count += 1;
+
+        let mut mem_diff = 0;
         for elem in elems.into_iter() {
             let len = elem.len();
             let is_new = self.cur_seg_distinct_elems.insert(elem);
             if is_new {
-                self.cur_seg_distinct_elems_mem_usage += len;
+                mem_diff += len;
             }
         }
+        self.cur_seg_distinct_elems_mem_usage += mem_diff;
+        self.global_memory_usage
+            .fetch_add(mem_diff, Ordering::Relaxed);
 
         if self.accumulated_row_count % self.rows_per_segment == 0 {
-            self.finalize_segment();
+            self.finalize_segment().await?;
         }
+
+        Ok(())
     }
 
     /// Finalizes any remaining segments and writes the bloom filters and metadata to the provided writer.
     pub async fn finish(&mut self, mut writer: impl AsyncWrite + Unpin) -> Result<()> {
         if !self.cur_seg_distinct_elems.is_empty() {
-            self.finalize_segment();
+            self.finalize_segment().await?;
         }
 
         let mut meta = BloomFilterMeta {
             rows_per_segment: self.rows_per_segment,
-            seg_count: self.finalized_bloom_filters.len(),
             row_count: self.accumulated_row_count,
             ..Default::default()
         };
 
-        let mut buf = Vec::new();
+        let mut segs = self.finalized_bloom_filters.drain().await?;
-        for segment in self.finalized_bloom_filters.drain() {
+        while let Some(segment) = segs.next().await {
-            let slice = segment.bloom_filter.as_slice();
+            let segment = segment?;
-            buf.clear();
+            writer
-            write_u64_slice(&mut buf, slice);
+                .write_all(&segment.bloom_filter_bytes)
-            writer.write_all(&buf).await.context(IoSnafu)?;
+                .await
+                .context(IoSnafu)?;
 
-            let size = buf.len();
+            let size = segment.bloom_filter_bytes.len();
             meta.bloom_filter_segments.push(BloomFilterSegmentLocation {
                 offset: meta.bloom_filter_segments_size as _,
                 size: size as _,
                 elem_count: segment.element_count,
             });
             meta.bloom_filter_segments_size += size;
+            meta.seg_count += 1;
         }
 
         let meta_bytes = serde_json::to_vec(&meta).context(SerdeJsonSnafu)?;
@@ -145,91 +168,29 @@ impl BloomFilterCreator {
         self.cur_seg_distinct_elems_mem_usage + self.finalized_bloom_filters.memory_usage()
     }
 
-    fn finalize_segment(&mut self) {
+    async fn finalize_segment(&mut self) -> Result<()> {
         let elem_count = self.cur_seg_distinct_elems.len();
         self.finalized_bloom_filters
-            .add(self.cur_seg_distinct_elems.drain(), elem_count);
+            .add(self.cur_seg_distinct_elems.drain(), elem_count)
+            .await?;
+
+        self.global_memory_usage
+            .fetch_sub(self.cur_seg_distinct_elems_mem_usage, Ordering::Relaxed);
         self.cur_seg_distinct_elems_mem_usage = 0;
-    }
+        Ok(())
-}
-
-/// Storage for finalized Bloom filters.
-///
-/// TODO(zhongzc): Add support for storing intermediate bloom filters on disk to control memory usage.
-#[derive(Debug, Default)]
-struct FinalizedBloomFilterStorage {
-    /// Bloom filters that are stored in memory.
-    in_memory: Vec<FinalizedBloomFilterSegment>,
-}
-
-impl FinalizedBloomFilterStorage {
-    fn memory_usage(&self) -> usize {
-        self.in_memory.iter().map(|s| s.size).sum()
-    }
-
-    /// Adds a new finalized Bloom filter to the storage.
-    ///
-    /// TODO(zhongzc): Add support for flushing to disk.
-    fn add(&mut self, elems: impl IntoIterator<Item = Bytes>, elem_count: usize) {
-        let mut bf = BloomFilter::with_false_pos(FALSE_POSITIVE_RATE)
-            .seed(&SEED)
-            .expected_items(elem_count);
-        for elem in elems.into_iter() {
-            bf.insert(&elem);
-        }
-
-        let cbf = FinalizedBloomFilterSegment::new(bf, elem_count);
-        self.in_memory.push(cbf);
-    }
-
-    fn len(&self) -> usize {
-        self.in_memory.len()
-    }
-
-    fn drain(&mut self) -> impl Iterator<Item = FinalizedBloomFilterSegment> + '_ {
-        self.in_memory.drain(..)
-    }
-}
-
-/// A finalized Bloom filter segment.
-#[derive(Debug)]
-struct FinalizedBloomFilterSegment {
-    /// The underlying Bloom filter.
-    bloom_filter: BloomFilter,
-
-    /// The number of elements in the Bloom filter.
-    element_count: usize,
-
-    /// The occupied memory size of the Bloom filter.
-    size: usize,
-}
-
-impl FinalizedBloomFilterSegment {
-    fn new(bloom_filter: BloomFilter, elem_count: usize) -> Self {
-        let memory_usage = std::mem::size_of_val(bloom_filter.as_slice());
-        Self {
-            bloom_filter,
-            element_count: elem_count,
-            size: memory_usage,
-        }
-    }
-}
-
-/// Writes a slice of `u64` to the buffer in little-endian order.
-fn write_u64_slice(buf: &mut Vec<u8>, slice: &[u64]) {
-    buf.reserve(std::mem::size_of_val(slice));
-    for &x in slice {
-        buf.extend_from_slice(&x.to_le_bytes());
     }
 }
 
 #[cfg(test)]
 mod tests {
+    use fastbloom::BloomFilter;
     use futures::io::Cursor;
 
     use super::*;
+    use crate::external_provider::MockExternalTempFileProvider;
 
-    fn u64_vec_from_bytes(bytes: &[u8]) -> Vec<u64> {
+    /// Converts a slice of bytes to a vector of `u64`.
+    pub fn u64_vec_from_bytes(bytes: &[u8]) -> Vec<u64> {
         bytes
             .chunks_exact(std::mem::size_of::<u64>())
             .map(|chunk| u64::from_le_bytes(chunk.try_into().unwrap()))
@@ -239,18 +200,32 @@ mod tests {
     #[tokio::test]
     async fn test_bloom_filter_creator() {
         let mut writer = Cursor::new(Vec::new());
-        let mut creator = BloomFilterCreator::new(2);
+        let mut creator = BloomFilterCreator::new(
+            2,
+            Box::new(MockExternalTempFileProvider::new()),
+            Arc::new(AtomicUsize::new(0)),
+            None,
+        );
 
-        creator.push_row_elems(vec![b"a".to_vec(), b"b".to_vec()]);
+        creator
+            .push_row_elems(vec![b"a".to_vec(), b"b".to_vec()])
+            .await
+            .unwrap();
         assert!(creator.cur_seg_distinct_elems_mem_usage > 0);
         assert!(creator.memory_usage() > 0);
 
-        creator.push_row_elems(vec![b"c".to_vec(), b"d".to_vec()]);
+        creator
+            .push_row_elems(vec![b"c".to_vec(), b"d".to_vec()])
+            .await
+            .unwrap();
         // Finalize the first segment
-        assert!(creator.cur_seg_distinct_elems_mem_usage == 0);
+        assert_eq!(creator.cur_seg_distinct_elems_mem_usage, 0);
         assert!(creator.memory_usage() > 0);
 
-        creator.push_row_elems(vec![b"e".to_vec(), b"f".to_vec()]);
+        creator
+            .push_row_elems(vec![b"e".to_vec(), b"f".to_vec()])
+            .await
+            .unwrap();
         assert!(creator.cur_seg_distinct_elems_mem_usage > 0);
         assert!(creator.memory_usage() > 0);
 

src/index/src/bloom_filter/creator/finalize_segment.rs

@@ -0,0 +1,293 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::pin::Pin;
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::sync::Arc;
+
+use asynchronous_codec::{FramedRead, FramedWrite};
+use fastbloom::BloomFilter;
+use futures::stream::StreamExt;
+use futures::{stream, AsyncWriteExt, Stream};
+use snafu::ResultExt;
+
+use super::intermediate_codec::IntermediateBloomFilterCodecV1;
+use crate::bloom_filter::creator::{FALSE_POSITIVE_RATE, SEED};
+use crate::bloom_filter::error::{IntermediateSnafu, IoSnafu, Result};
+use crate::bloom_filter::Bytes;
+use crate::external_provider::ExternalTempFileProvider;
+
+/// The minimum memory usage threshold for flushing in-memory Bloom filters to disk.
+const MIN_MEMORY_USAGE_THRESHOLD: usize = 1024 * 1024; // 1MB
+
+/// Storage for finalized Bloom filters.
+pub struct FinalizedBloomFilterStorage {
+    /// Bloom filters that are stored in memory.
+    in_memory: Vec<FinalizedBloomFilterSegment>,
+
+    /// Used to generate unique file IDs for intermediate Bloom filters.
+    intermediate_file_id_counter: usize,
+
+    /// Prefix for intermediate Bloom filter files.
+    intermediate_prefix: String,
+
+    /// The provider for intermediate Bloom filter files.
+    intermediate_provider: Box<dyn ExternalTempFileProvider>,
+
+    /// The memory usage of the in-memory Bloom filters.
+    memory_usage: usize,
+
+    /// The global memory usage provided by the user to track the
+    /// total memory usage of the creating Bloom filters.
+    global_memory_usage: Arc<AtomicUsize>,
+
+    /// The threshold of the global memory usage of the creating Bloom filters.
+    global_memory_usage_threshold: Option<usize>,
+}
+
+impl FinalizedBloomFilterStorage {
+    /// Creates a new `FinalizedBloomFilterStorage`.
+    pub fn new(
+        intermediate_provider: Box<dyn ExternalTempFileProvider>,
+        global_memory_usage: Arc<AtomicUsize>,
+        global_memory_usage_threshold: Option<usize>,
+    ) -> Self {
+        let external_prefix = format!("intm-bloom-filters-{}", uuid::Uuid::new_v4());
+        Self {
+            in_memory: Vec::new(),
+            intermediate_file_id_counter: 0,
+            intermediate_prefix: external_prefix,
+            intermediate_provider,
+            memory_usage: 0,
+            global_memory_usage,
+            global_memory_usage_threshold,
+        }
+    }
+
+    /// Returns the memory usage of the storage.
+    pub fn memory_usage(&self) -> usize {
+        self.memory_usage
+    }
+
+    /// Adds a new finalized Bloom filter to the storage.
+    ///
+    /// If the memory usage exceeds the threshold, flushes the in-memory Bloom filters to disk.
+    pub async fn add(
+        &mut self,
+        elems: impl IntoIterator<Item = Bytes>,
+        element_count: usize,
+    ) -> Result<()> {
+        let mut bf = BloomFilter::with_false_pos(FALSE_POSITIVE_RATE)
+            .seed(&SEED)
+            .expected_items(element_count);
+        for elem in elems.into_iter() {
+            bf.insert(&elem);
+        }
+
+        let fbf = FinalizedBloomFilterSegment::from(bf, element_count);
+
+        // Update memory usage.
+        let memory_diff = fbf.bloom_filter_bytes.len();
+        self.memory_usage += memory_diff;
+        self.global_memory_usage
+            .fetch_add(memory_diff, Ordering::Relaxed);
+
+        // Add the finalized Bloom filter to the in-memory storage.
+        self.in_memory.push(fbf);
+
+        // Flush to disk if necessary.
+
+        // Do not flush if memory usage is too low.
+        if self.memory_usage < MIN_MEMORY_USAGE_THRESHOLD {
+            return Ok(());
+        }
+
+        // Check if the global memory usage exceeds the threshold and flush to disk if necessary.
+        if let Some(threshold) = self.global_memory_usage_threshold {
+            let global = self.global_memory_usage.load(Ordering::Relaxed);
+
+            if global > threshold {
+                self.flush_in_memory_to_disk().await?;
+
+                self.global_memory_usage
+                    .fetch_sub(self.memory_usage, Ordering::Relaxed);
+                self.memory_usage = 0;
+            }
+        }
+
+        Ok(())
+    }
+
+    /// Drains the storage and returns a stream of finalized Bloom filter segments.
+    pub async fn drain(
+        &mut self,
+    ) -> Result<Pin<Box<dyn Stream<Item = Result<FinalizedBloomFilterSegment>> + '_>>> {
+        // FAST PATH: memory only
+        if self.intermediate_file_id_counter == 0 {
+            return Ok(Box::pin(stream::iter(self.in_memory.drain(..).map(Ok))));
+        }
+
+        // SLOW PATH: memory + disk
+        let mut on_disk = self
+            .intermediate_provider
+            .read_all(&self.intermediate_prefix)
+            .await
+            .context(IntermediateSnafu)?;
+        on_disk.sort_unstable_by(|x, y| x.0.cmp(&y.0));
+
+        let streams = on_disk
+            .into_iter()
+            .map(|(_, reader)| FramedRead::new(reader, IntermediateBloomFilterCodecV1::default()));
+
+        let in_memory_stream = stream::iter(self.in_memory.drain(..)).map(Ok);
+        Ok(Box::pin(
+            stream::iter(streams).flatten().chain(in_memory_stream),
+        ))
+    }
+
+    /// Flushes the in-memory Bloom filters to disk.
+    async fn flush_in_memory_to_disk(&mut self) -> Result<()> {
+        let file_id = self.intermediate_file_id_counter;
+        self.intermediate_file_id_counter += 1;
+
+        let file_id = format!("{:08}", file_id);
+        let mut writer = self
+            .intermediate_provider
+            .create(&self.intermediate_prefix, &file_id)
+            .await
+            .context(IntermediateSnafu)?;
+
+        let fw = FramedWrite::new(&mut writer, IntermediateBloomFilterCodecV1::default());
+        // `forward()` will flush and close the writer when the stream ends
+        if let Err(e) = stream::iter(self.in_memory.drain(..).map(Ok))
+            .forward(fw)
+            .await
+        {
+            writer.close().await.context(IoSnafu)?;
+            writer.flush().await.context(IoSnafu)?;
+            return Err(e);
+        }
+
+        Ok(())
+    }
+}
+
+/// A finalized Bloom filter segment.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct FinalizedBloomFilterSegment {
+    /// The underlying Bloom filter bytes.
+    pub bloom_filter_bytes: Vec<u8>,
+
+    /// The number of elements in the Bloom filter.
+    pub element_count: usize,
+}
+
+impl FinalizedBloomFilterSegment {
+    fn from(bf: BloomFilter, elem_count: usize) -> Self {
+        let bf_slice = bf.as_slice();
+        let mut bloom_filter_bytes = Vec::with_capacity(std::mem::size_of_val(bf_slice));
+        for &x in bf_slice {
+            bloom_filter_bytes.extend_from_slice(&x.to_le_bytes());
+        }
+
+        Self {
+            bloom_filter_bytes,
+            element_count: elem_count,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+    use std::sync::Mutex;
+
+    use futures::AsyncRead;
+    use tokio::io::duplex;
+    use tokio_util::compat::{TokioAsyncReadCompatExt, TokioAsyncWriteCompatExt};
+
+    use super::*;
+    use crate::bloom_filter::creator::tests::u64_vec_from_bytes;
+    use crate::external_provider::MockExternalTempFileProvider;
+
+    #[tokio::test]
+    async fn test_finalized_bloom_filter_storage() {
+        let mut mock_provider = MockExternalTempFileProvider::new();
+
+        let mock_files: Arc<Mutex<HashMap<String, Box<dyn AsyncRead + Unpin + Send>>>> =
+            Arc::new(Mutex::new(HashMap::new()));
+
+        mock_provider.expect_create().returning({
+            let files = Arc::clone(&mock_files);
+            move |file_group, file_id| {
+                assert!(file_group.starts_with("intm-bloom-filters-"));
+                let mut files = files.lock().unwrap();
+                let (writer, reader) = duplex(2 * 1024 * 1024);
+                files.insert(file_id.to_string(), Box::new(reader.compat()));
+                Ok(Box::new(writer.compat_write()))
+            }
+        });
+
+        mock_provider.expect_read_all().returning({
+            let files = Arc::clone(&mock_files);
+            move |file_group| {
+                assert!(file_group.starts_with("intm-bloom-filters-"));
+                let mut files = files.lock().unwrap();
+                Ok(files.drain().collect::<Vec<_>>())
+            }
+        });
+
+        let global_memory_usage = Arc::new(AtomicUsize::new(0));
+        let global_memory_usage_threshold = Some(1024 * 1024); // 1MB
+        let provider = Box::new(mock_provider);
+        let mut storage = FinalizedBloomFilterStorage::new(
+            provider,
+            global_memory_usage.clone(),
+            global_memory_usage_threshold,
+        );
+
+        let elem_count = 2000;
+        let batch = 1000;
+
+        for i in 0..batch {
+            let elems = (elem_count * i..elem_count * (i + 1)).map(|x| x.to_string().into_bytes());
+            storage.add(elems, elem_count).await.unwrap();
+        }
+
+        // Flush happens.
+        assert!(storage.intermediate_file_id_counter > 0);
+
+        // Drain the storage.
+        let mut stream = storage.drain().await.unwrap();
+
+        let mut i = 0;
+        while let Some(segment) = stream.next().await {
+            let segment = segment.unwrap();
+            assert_eq!(segment.element_count, elem_count);
+
+            let v = u64_vec_from_bytes(&segment.bloom_filter_bytes);
+
+            // Check the correctness of the Bloom filter.
+            let bf = BloomFilter::from_vec(v)
+                .seed(&SEED)
+                .expected_items(segment.element_count);
+            for elem in (elem_count * i..elem_count * (i + 1)).map(|x| x.to_string().into_bytes()) {
+                assert!(bf.contains(&elem));
+            }
+            i += 1;
+        }
+
+        assert_eq!(i, batch);
+    }
+}

src/index/src/bloom_filter/creator/intermediate_codec.rs

@@ -0,0 +1,248 @@
+// 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 asynchronous_codec::{BytesMut, Decoder, Encoder};
+use bytes::{Buf, BufMut};
+use snafu::{ensure, ResultExt};
+
+use crate::bloom_filter::creator::finalize_segment::FinalizedBloomFilterSegment;
+use crate::bloom_filter::error::{Error, InvalidIntermediateMagicSnafu, IoSnafu, Result};
+
+/// The magic number for the codec version 1 of the intermediate bloom filter.
+const CODEC_V1_MAGIC: &[u8; 4] = b"bi01";
+
+/// Codec of the intermediate finalized bloom filter segment.
+///
+/// # Format
+///
+/// [ magic ][ elem count ][    size    ][ bloom filter ][ elem count ][    size    ][ bloom filter ]...
+///    [4]       [8]            [8]           [size]         [8]            [8]           [size]
+#[derive(Debug, Default)]
+pub struct IntermediateBloomFilterCodecV1 {
+    handled_header_magic: bool,
+}
+
+impl Encoder for IntermediateBloomFilterCodecV1 {
+    type Item<'a> = FinalizedBloomFilterSegment;
+    type Error = Error;
+
+    fn encode(&mut self, item: FinalizedBloomFilterSegment, dst: &mut BytesMut) -> Result<()> {
+        if !self.handled_header_magic {
+            dst.extend_from_slice(CODEC_V1_MAGIC);
+            self.handled_header_magic = true;
+        }
+
+        let segment_bytes = item.bloom_filter_bytes;
+        let elem_count = item.element_count;
+
+        dst.reserve(2 * std::mem::size_of::<u64>() + segment_bytes.len());
+        dst.put_u64_le(elem_count as u64);
+        dst.put_u64_le(segment_bytes.len() as u64);
+        dst.extend_from_slice(&segment_bytes);
+        Ok(())
+    }
+}
+
+impl Decoder for IntermediateBloomFilterCodecV1 {
+    type Item = FinalizedBloomFilterSegment;
+    type Error = Error;
+
+    fn decode(&mut self, src: &mut BytesMut) -> Result<Option<Self::Item>> {
+        if !self.handled_header_magic {
+            let m_len = CODEC_V1_MAGIC.len();
+            if src.remaining() < m_len {
+                return Ok(None);
+            }
+            let magic_bytes = &src[..m_len];
+            ensure!(
+                magic_bytes == CODEC_V1_MAGIC,
+                InvalidIntermediateMagicSnafu {
+                    invalid: magic_bytes,
+                }
+            );
+            self.handled_header_magic = true;
+            src.advance(m_len);
+        }
+
+        let s = &src[..];
+
+        let u64_size = std::mem::size_of::<u64>();
+        let n_size = u64_size * 2;
+        if s.len() < n_size {
+            return Ok(None);
+        }
+
+        let element_count = u64::from_le_bytes(s[0..u64_size].try_into().unwrap()) as usize;
+        let segment_size = u64::from_le_bytes(s[u64_size..n_size].try_into().unwrap()) as usize;
+
+        if s.len() < n_size + segment_size {
+            return Ok(None);
+        }
+
+        let bloom_filter_bytes = s[n_size..n_size + segment_size].to_vec();
+        src.advance(n_size + segment_size);
+        Ok(Some(FinalizedBloomFilterSegment {
+            element_count,
+            bloom_filter_bytes,
+        }))
+    }
+}
+
+/// Required for [`Encoder`] and [`Decoder`] implementations.
+impl From<std::io::Error> for Error {
+    fn from(error: std::io::Error) -> Self {
+        Err::<(), std::io::Error>(error)
+            .context(IoSnafu)
+            .unwrap_err()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use asynchronous_codec::{FramedRead, FramedWrite};
+    use futures::io::Cursor;
+    use futures::{SinkExt, StreamExt};
+
+    use super::*;
+    use crate::bloom_filter::creator::finalize_segment::FinalizedBloomFilterSegment;
+
+    #[test]
+    fn test_intermediate_bloom_filter_codec_v1_basic() {
+        let mut encoder = IntermediateBloomFilterCodecV1::default();
+        let mut buf = BytesMut::new();
+
+        let item1 = FinalizedBloomFilterSegment {
+            element_count: 2,
+            bloom_filter_bytes: vec![1, 2, 3, 4],
+        };
+        let item2 = FinalizedBloomFilterSegment {
+            element_count: 3,
+            bloom_filter_bytes: vec![5, 6, 7, 8],
+        };
+        let item3 = FinalizedBloomFilterSegment {
+            element_count: 4,
+            bloom_filter_bytes: vec![9, 10, 11, 12],
+        };
+
+        encoder.encode(item1.clone(), &mut buf).unwrap();
+        encoder.encode(item2.clone(), &mut buf).unwrap();
+        encoder.encode(item3.clone(), &mut buf).unwrap();
+
+        let mut buf = buf.freeze().try_into_mut().unwrap();
+
+        let mut decoder = IntermediateBloomFilterCodecV1::default();
+        let decoded_item1 = decoder.decode(&mut buf).unwrap().unwrap();
+        let decoded_item2 = decoder.decode(&mut buf).unwrap().unwrap();
+        let decoded_item3 = decoder.decode(&mut buf).unwrap().unwrap();
+
+        assert_eq!(item1, decoded_item1);
+        assert_eq!(item2, decoded_item2);
+        assert_eq!(item3, decoded_item3);
+    }
+
+    #[tokio::test]
+    async fn test_intermediate_bloom_filter_codec_v1_frame_read_write() {
+        let item1 = FinalizedBloomFilterSegment {
+            element_count: 2,
+            bloom_filter_bytes: vec![1, 2, 3, 4],
+        };
+        let item2 = FinalizedBloomFilterSegment {
+            element_count: 3,
+            bloom_filter_bytes: vec![5, 6, 7, 8],
+        };
+        let item3 = FinalizedBloomFilterSegment {
+            element_count: 4,
+            bloom_filter_bytes: vec![9, 10, 11, 12],
+        };
+
+        let mut bytes = Cursor::new(vec![]);
+
+        let mut writer = FramedWrite::new(&mut bytes, IntermediateBloomFilterCodecV1::default());
+        writer.send(item1.clone()).await.unwrap();
+        writer.send(item2.clone()).await.unwrap();
+        writer.send(item3.clone()).await.unwrap();
+        writer.flush().await.unwrap();
+        writer.close().await.unwrap();
+
+        let bytes = bytes.into_inner();
+        let mut reader =
+            FramedRead::new(bytes.as_slice(), IntermediateBloomFilterCodecV1::default());
+        let decoded_item1 = reader.next().await.unwrap().unwrap();
+        let decoded_item2 = reader.next().await.unwrap().unwrap();
+        let decoded_item3 = reader.next().await.unwrap().unwrap();
+        assert!(reader.next().await.is_none());
+
+        assert_eq!(item1, decoded_item1);
+        assert_eq!(item2, decoded_item2);
+        assert_eq!(item3, decoded_item3);
+    }
+
+    #[tokio::test]
+    async fn test_intermediate_bloom_filter_codec_v1_frame_read_write_only_magic() {
+        let bytes = CODEC_V1_MAGIC.to_vec();
+        let mut reader =
+            FramedRead::new(bytes.as_slice(), IntermediateBloomFilterCodecV1::default());
+        assert!(reader.next().await.is_none());
+    }
+
+    #[tokio::test]
+    async fn test_intermediate_bloom_filter_codec_v1_frame_read_write_partial_magic() {
+        let bytes = CODEC_V1_MAGIC[..3].to_vec();
+        let mut reader =
+            FramedRead::new(bytes.as_slice(), IntermediateBloomFilterCodecV1::default());
+        let e = reader.next().await.unwrap();
+        assert!(e.is_err());
+    }
+
+    #[tokio::test]
+    async fn test_intermediate_bloom_filter_codec_v1_frame_read_write_partial_item() {
+        let mut bytes = vec![];
+        bytes.extend_from_slice(CODEC_V1_MAGIC);
+        bytes.extend_from_slice(&2u64.to_le_bytes());
+        bytes.extend_from_slice(&4u64.to_le_bytes());
+
+        let mut reader =
+            FramedRead::new(bytes.as_slice(), IntermediateBloomFilterCodecV1::default());
+        let e = reader.next().await.unwrap();
+        assert!(e.is_err());
+    }
+
+    #[tokio::test]
+    async fn test_intermediate_bloom_filter_codec_v1_frame_read_write_corrupted_magic() {
+        let mut bytes = vec![];
+        bytes.extend_from_slice(b"bi02");
+        bytes.extend_from_slice(&2u64.to_le_bytes());
+        bytes.extend_from_slice(&4u64.to_le_bytes());
+        bytes.extend_from_slice(&[1, 2, 3, 4]);
+
+        let mut reader =
+            FramedRead::new(bytes.as_slice(), IntermediateBloomFilterCodecV1::default());
+        let e = reader.next().await.unwrap();
+        assert!(e.is_err());
+    }
+
+    #[tokio::test]
+    async fn test_intermediate_bloom_filter_codec_v1_frame_read_write_corrupted_length() {
+        let mut bytes = vec![];
+        bytes.extend_from_slice(CODEC_V1_MAGIC);
+        bytes.extend_from_slice(&2u64.to_le_bytes());
+        bytes.extend_from_slice(&4u64.to_le_bytes());
+        bytes.extend_from_slice(&[1, 2, 3]);
+
+        let mut reader =
+            FramedRead::new(bytes.as_slice(), IntermediateBloomFilterCodecV1::default());
+        let e = reader.next().await.unwrap();
+        assert!(e.is_err());
+    }
+}

src/index/src/bloom_filter/error.rs

@@ -39,6 +39,43 @@ pub enum Error {
         location: Location,
     },
 
+    #[snafu(display("Failed to deserialize json"))]
+    DeserializeJson {
+        #[snafu(source)]
+        error: serde_json::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Intermediate error"))]
+    Intermediate {
+        source: crate::error::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("File size too small for bloom filter"))]
+    FileSizeTooSmall {
+        size: u64,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Unexpected bloom filter meta size"))]
+    UnexpectedMetaSize {
+        max_meta_size: u64,
+        actual_meta_size: u64,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[snafu(display("Invalid intermediate magic"))]
+    InvalidIntermediateMagic {
+        invalid: Vec<u8>,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
     #[snafu(display("External error"))]
     External {
         source: BoxedError,
@@ -52,8 +89,14 @@ impl ErrorExt for Error {
         use Error::*;
 
         match self {
-            Io { .. } | Self::SerdeJson { .. } => StatusCode::Unexpected,
+            Io { .. }
+            | SerdeJson { .. }
+            | FileSizeTooSmall { .. }
+            | UnexpectedMetaSize { .. }
+            | DeserializeJson { .. }
+            | InvalidIntermediateMagic { .. } => StatusCode::Unexpected,
 
+            Intermediate { source, .. } => source.status_code(),
             External { source, .. } => source.status_code(),
         }
     }

src/index/src/bloom_filter/reader.rs

@@ -0,0 +1,265 @@
+// 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::ops::Range;
+
+use async_trait::async_trait;
+use bytes::Bytes;
+use common_base::range_read::RangeReader;
+use fastbloom::BloomFilter;
+use snafu::{ensure, ResultExt};
+
+use crate::bloom_filter::error::{
+    DeserializeJsonSnafu, FileSizeTooSmallSnafu, IoSnafu, Result, UnexpectedMetaSizeSnafu,
+};
+use crate::bloom_filter::{BloomFilterMeta, BloomFilterSegmentLocation, SEED};
+
+/// Minimum size of the bloom filter, which is the size of the length of the bloom filter.
+const BLOOM_META_LEN_SIZE: u64 = 4;
+
+/// Default prefetch size of bloom filter meta.
+pub const DEFAULT_PREFETCH_SIZE: u64 = 1024; // 1KiB
+
+/// `BloomFilterReader` reads the bloom filter from the file.
+#[async_trait]
+pub trait BloomFilterReader {
+    /// Reads range of bytes from the file.
+    async fn range_read(&mut self, offset: u64, size: u32) -> Result<Bytes>;
+
+    /// Reads bunch of ranges from the file.
+    async fn read_vec(&mut self, ranges: &[Range<u64>]) -> Result<Vec<Bytes>>;
+
+    /// Reads the meta information of the bloom filter.
+    async fn metadata(&mut self) -> Result<BloomFilterMeta>;
+
+    /// Reads a bloom filter with the given location.
+    async fn bloom_filter(&mut self, loc: &BloomFilterSegmentLocation) -> Result<BloomFilter> {
+        let bytes = self.range_read(loc.offset, loc.size as _).await?;
+        let vec = bytes
+            .chunks_exact(std::mem::size_of::<u64>())
+            .map(|chunk| u64::from_le_bytes(chunk.try_into().unwrap()))
+            .collect();
+        let bm = BloomFilter::from_vec(vec)
+            .seed(&SEED)
+            .expected_items(loc.elem_count);
+        Ok(bm)
+    }
+}
+
+/// `BloomFilterReaderImpl` reads the bloom filter from the file.
+pub struct BloomFilterReaderImpl<R: RangeReader> {
+    /// The underlying reader.
+    reader: R,
+}
+
+impl<R: RangeReader> BloomFilterReaderImpl<R> {
+    /// Creates a new `BloomFilterReaderImpl` with the given reader.
+    pub fn new(reader: R) -> Self {
+        Self { reader }
+    }
+}
+
+#[async_trait]
+impl<R: RangeReader> BloomFilterReader for BloomFilterReaderImpl<R> {
+    async fn range_read(&mut self, offset: u64, size: u32) -> Result<Bytes> {
+        self.reader
+            .read(offset..offset + size as u64)
+            .await
+            .context(IoSnafu)
+    }
+
+    async fn read_vec(&mut self, ranges: &[Range<u64>]) -> Result<Vec<Bytes>> {
+        self.reader.read_vec(ranges).await.context(IoSnafu)
+    }
+
+    async fn metadata(&mut self) -> Result<BloomFilterMeta> {
+        let metadata = self.reader.metadata().await.context(IoSnafu)?;
+        let file_size = metadata.content_length;
+
+        let mut meta_reader =
+            BloomFilterMetaReader::new(&mut self.reader, file_size, Some(DEFAULT_PREFETCH_SIZE));
+        meta_reader.metadata().await
+    }
+}
+
+/// `BloomFilterMetaReader` reads the metadata of the bloom filter.
+struct BloomFilterMetaReader<R: RangeReader> {
+    reader: R,
+    file_size: u64,
+    prefetch_size: u64,
+}
+
+impl<R: RangeReader> BloomFilterMetaReader<R> {
+    pub fn new(reader: R, file_size: u64, prefetch_size: Option<u64>) -> Self {
+        Self {
+            reader,
+            file_size,
+            prefetch_size: prefetch_size
+                .unwrap_or(BLOOM_META_LEN_SIZE)
+                .max(BLOOM_META_LEN_SIZE),
+        }
+    }
+
+    /// Reads the metadata of the bloom filter.
+    ///
+    /// It will first prefetch some bytes from the end of the file,
+    /// then parse the metadata from the prefetch bytes.
+    pub async fn metadata(&mut self) -> Result<BloomFilterMeta> {
+        ensure!(
+            self.file_size >= BLOOM_META_LEN_SIZE,
+            FileSizeTooSmallSnafu {
+                size: self.file_size,
+            }
+        );
+
+        let meta_start = self.file_size.saturating_sub(self.prefetch_size);
+        let suffix = self
+            .reader
+            .read(meta_start..self.file_size)
+            .await
+            .context(IoSnafu)?;
+        let suffix_len = suffix.len();
+        let length = u32::from_le_bytes(Self::read_tailing_four_bytes(&suffix)?) as u64;
+        self.validate_meta_size(length)?;
+
+        if length > suffix_len as u64 - BLOOM_META_LEN_SIZE {
+            let metadata_start = self.file_size - length - BLOOM_META_LEN_SIZE;
+            let meta = self
+                .reader
+                .read(metadata_start..self.file_size - BLOOM_META_LEN_SIZE)
+                .await
+                .context(IoSnafu)?;
+            serde_json::from_slice(&meta).context(DeserializeJsonSnafu)
+        } else {
+            let metadata_start = self.file_size - length - BLOOM_META_LEN_SIZE - meta_start;
+            let meta = &suffix[metadata_start as usize..suffix_len - BLOOM_META_LEN_SIZE as usize];
+            serde_json::from_slice(meta).context(DeserializeJsonSnafu)
+        }
+    }
+
+    fn read_tailing_four_bytes(suffix: &[u8]) -> Result<[u8; 4]> {
+        let suffix_len = suffix.len();
+        ensure!(
+            suffix_len >= 4,
+            FileSizeTooSmallSnafu {
+                size: suffix_len as u64
+            }
+        );
+        let mut bytes = [0; 4];
+        bytes.copy_from_slice(&suffix[suffix_len - 4..suffix_len]);
+
+        Ok(bytes)
+    }
+
+    fn validate_meta_size(&self, length: u64) -> Result<()> {
+        let max_meta_size = self.file_size - BLOOM_META_LEN_SIZE;
+        ensure!(
+            length <= max_meta_size,
+            UnexpectedMetaSizeSnafu {
+                max_meta_size,
+                actual_meta_size: length,
+            }
+        );
+        Ok(())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::atomic::AtomicUsize;
+    use std::sync::Arc;
+
+    use futures::io::Cursor;
+
+    use super::*;
+    use crate::bloom_filter::creator::BloomFilterCreator;
+    use crate::external_provider::MockExternalTempFileProvider;
+
+    async fn mock_bloom_filter_bytes() -> Vec<u8> {
+        let mut writer = Cursor::new(vec![]);
+        let mut creator = BloomFilterCreator::new(
+            2,
+            Box::new(MockExternalTempFileProvider::new()),
+            Arc::new(AtomicUsize::new(0)),
+            None,
+        );
+
+        creator
+            .push_row_elems(vec![b"a".to_vec(), b"b".to_vec()])
+            .await
+            .unwrap();
+        creator
+            .push_row_elems(vec![b"c".to_vec(), b"d".to_vec()])
+            .await
+            .unwrap();
+        creator
+            .push_row_elems(vec![b"e".to_vec(), b"f".to_vec()])
+            .await
+            .unwrap();
+
+        creator.finish(&mut writer).await.unwrap();
+
+        writer.into_inner()
+    }
+
+    #[tokio::test]
+    async fn test_bloom_filter_meta_reader() {
+        let bytes = mock_bloom_filter_bytes().await;
+        let file_size = bytes.len() as u64;
+
+        for prefetch in [0u64, file_size / 2, file_size, file_size + 10] {
+            let mut reader =
+                BloomFilterMetaReader::new(bytes.clone(), file_size as _, Some(prefetch));
+            let meta = reader.metadata().await.unwrap();
+
+            assert_eq!(meta.rows_per_segment, 2);
+            assert_eq!(meta.seg_count, 2);
+            assert_eq!(meta.row_count, 3);
+            assert_eq!(meta.bloom_filter_segments.len(), 2);
+
+            assert_eq!(meta.bloom_filter_segments[0].offset, 0);
+            assert_eq!(meta.bloom_filter_segments[0].elem_count, 4);
+            assert_eq!(
+                meta.bloom_filter_segments[1].offset,
+                meta.bloom_filter_segments[0].size
+            );
+            assert_eq!(meta.bloom_filter_segments[1].elem_count, 2);
+        }
+    }
+
+    #[tokio::test]
+    async fn test_bloom_filter_reader() {
+        let bytes = mock_bloom_filter_bytes().await;
+
+        let mut reader = BloomFilterReaderImpl::new(bytes);
+        let meta = reader.metadata().await.unwrap();
+
+        assert_eq!(meta.bloom_filter_segments.len(), 2);
+        let bf = reader
+            .bloom_filter(&meta.bloom_filter_segments[0])
+            .await
+            .unwrap();
+        assert!(bf.contains(&b"a"));
+        assert!(bf.contains(&b"b"));
+        assert!(bf.contains(&b"c"));
+        assert!(bf.contains(&b"d"));
+
+        let bf = reader
+            .bloom_filter(&meta.bloom_filter_segments[1])
+            .await
+            .unwrap();
+        assert!(bf.contains(&b"e"));
+        assert!(bf.contains(&b"f"));
+    }
+}

src/index/src/error.rs

@@ -0,0 +1,48 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::any::Any;
+
+use common_error::ext::{BoxedError, ErrorExt};
+use common_error::status_code::StatusCode;
+use common_macro::stack_trace_debug;
+use snafu::{Location, Snafu};
+
+#[derive(Snafu)]
+#[snafu(visibility(pub))]
+#[stack_trace_debug]
+pub enum Error {
+    #[snafu(display("External error"))]
+    External {
+        source: BoxedError,
+        #[snafu(implicit)]
+        location: Location,
+    },
+}
+
+impl ErrorExt for Error {
+    fn status_code(&self) -> StatusCode {
+        use Error::*;
+
+        match self {
+            External { source, .. } => source.status_code(),
+        }
+    }
+
+    fn as_any(&self) -> &dyn Any {
+        self
+    }
+}
+
+pub type Result<T> = std::result::Result<T, Error>;

src/index/src/external_provider.rs

@@ -15,25 +15,24 @@
 use async_trait::async_trait;
 use futures::{AsyncRead, AsyncWrite};
 
-use crate::inverted_index::error::Result;
+use crate::error::Error;
 
-/// Trait for managing intermediate files during external sorting for a particular index.
+pub type Writer = Box<dyn AsyncWrite + Unpin + Send>;
+pub type Reader = Box<dyn AsyncRead + Unpin + Send>;
+
+/// Trait for managing intermediate files to control memory usage for a particular index.
 #[mockall::automock]
 #[async_trait]
 pub trait ExternalTempFileProvider: Send + Sync {
-    /// Creates and opens a new intermediate file associated with a specific index for writing.
+    /// Creates and opens a new intermediate file associated with a specific `file_group` for writing.
     /// The implementation should ensure that the file does not already exist.
     ///
-    /// - `index_name`: the name of the index for which the file will be associated
+    /// - `file_group`: a unique identifier for the group of files
     /// - `file_id`: a unique identifier for the new file
-    async fn create(
+    async fn create(&self, file_group: &str, file_id: &str) -> Result<Writer, Error>;
-        &self,
-        index_name: &str,
-        file_id: &str,
-    ) -> Result<Box<dyn AsyncWrite + Unpin + Send>>;
 
-    /// Retrieves all intermediate files associated with a specific index for an external sorting operation.
+    /// Retrieves all intermediate files and their associated file identifiers for a specific `file_group`.
     ///
-    /// `index_name`: the name of the index to retrieve intermediate files for
+    /// `file_group` is a unique identifier for the group of files.
-    async fn read_all(&self, index_name: &str) -> Result<Vec<Box<dyn AsyncRead + Unpin + Send>>>;
+    async fn read_all(&self, file_group: &str) -> Result<Vec<(String, Reader)>, Error>;
 }

src/index/src/inverted_index/create/sort.rs

@@ -12,7 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-pub mod external_provider;
 pub mod external_sort;
 mod intermediate_rw;
 mod merge_stream;

src/index/src/inverted_index/create/sort/external_sort.rs

@@ -23,15 +23,16 @@ use async_trait::async_trait;
 use common_base::BitVec;
 use common_telemetry::{debug, error};
 use futures::stream;
+use snafu::ResultExt;
 
-use crate::inverted_index::create::sort::external_provider::ExternalTempFileProvider;
+use crate::external_provider::ExternalTempFileProvider;
 use crate::inverted_index::create::sort::intermediate_rw::{
     IntermediateReader, IntermediateWriter,
 };
 use crate::inverted_index::create::sort::merge_stream::MergeSortedStream;
 use crate::inverted_index::create::sort::{SortOutput, SortedStream, Sorter};
 use crate::inverted_index::create::sort_create::SorterFactory;
-use crate::inverted_index::error::Result;
+use crate::inverted_index::error::{IntermediateSnafu, Result};
 use crate::inverted_index::{Bytes, BytesRef};
 
 /// `ExternalSorter` manages the sorting of data using both in-memory structures and external files.
@@ -107,7 +108,11 @@ impl Sorter for ExternalSorter {
     /// Finalizes the sorting operation, merging data from both in-memory buffer and external files
     /// into a sorted stream
     async fn output(&mut self) -> Result<SortOutput> {
-        let readers = self.temp_file_provider.read_all(&self.index_name).await?;
+        let readers = self
+            .temp_file_provider
+            .read_all(&self.index_name)
+            .await
+            .context(IntermediateSnafu)?;
 
         // TODO(zhongzc): k-way merge instead of 2-way merge
 
@@ -122,7 +127,7 @@ impl Sorter for ExternalSorter {
                     Ok((value, bitmap))
                 }),
         )));
-        for reader in readers {
+        for (_, reader) in readers {
             tree_nodes.push_back(IntermediateReader::new(reader).into_stream().await?);
         }
 
@@ -241,7 +246,11 @@ impl ExternalSorter {
 
         let file_id = &format!("{:012}", self.total_row_count);
         let index_name = &self.index_name;
-        let writer = self.temp_file_provider.create(index_name, file_id).await?;
+        let writer = self
+            .temp_file_provider
+            .create(index_name, file_id)
+            .await
+            .context(IntermediateSnafu)?;
 
         let values = mem::take(&mut self.values_buffer);
         self.global_memory_usage
@@ -302,7 +311,7 @@ mod tests {
     use tokio_util::compat::{TokioAsyncReadCompatExt, TokioAsyncWriteCompatExt};
 
     use super::*;
-    use crate::inverted_index::create::sort::external_provider::MockExternalTempFileProvider;
+    use crate::external_provider::MockExternalTempFileProvider;
 
     async fn test_external_sorter(
         current_memory_usage_threshold: Option<usize>,
@@ -332,7 +341,7 @@ mod tests {
             move |index_name| {
                 assert_eq!(index_name, "test");
                 let mut files = files.lock().unwrap();
-                Ok(files.drain().map(|f| f.1).collect::<Vec<_>>())
+                Ok(files.drain().collect::<Vec<_>>())
             }
         });
 

src/index/src/inverted_index/error.rs

@@ -213,6 +213,13 @@ pub enum Error {
         #[snafu(implicit)]
         location: Location,
     },
+
+    #[snafu(display("Intermediate error"))]
+    Intermediate {
+        source: crate::error::Error,
+        #[snafu(implicit)]
+        location: Location,
+    },
 }
 
 impl ErrorExt for Error {
@@ -245,6 +252,7 @@ impl ErrorExt for Error {
             | InconsistentRowCount { .. }
             | IndexNotFound { .. } => StatusCode::InvalidArguments,
 
+            Intermediate { source, .. } => source.status_code(),
             External { source, .. } => source.status_code(),
         }
     }

src/index/src/inverted_index/format/reader.rs

@@ -31,12 +31,21 @@ mod footer;
 /// InvertedIndexReader defines an asynchronous reader of inverted index data
 #[mockall::automock]
 #[async_trait]
-pub trait InvertedIndexReader: Send {
+pub trait InvertedIndexReader: Send + Sync {
     /// Seeks to given offset and reads data with exact size as provided.
     async fn range_read(&mut self, offset: u64, size: u32) -> Result<Vec<u8>>;
 
     /// Reads the bytes in the given ranges.
-    async fn read_vec(&mut self, ranges: &[Range<u64>]) -> Result<Vec<Bytes>>;
+    async fn read_vec(&mut self, ranges: &[Range<u64>]) -> Result<Vec<Bytes>> {
+        let mut result = Vec::with_capacity(ranges.len());
+        for range in ranges {
+            let data = self
+                .range_read(range.start, (range.end - range.start) as u32)
+                .await?;
+            result.push(Bytes::from(data));
+        }
+        Ok(result)
+    }
 
     /// Retrieves metadata of all inverted indices stored within the blob.
     async fn metadata(&mut self) -> Result<Arc<InvertedIndexMetas>>;

src/index/src/inverted_index/format/reader/blob.rs

@@ -51,7 +51,7 @@ impl<R> InvertedIndexBlobReader<R> {
 }
 
 #[async_trait]
-impl<R: RangeReader> InvertedIndexReader for InvertedIndexBlobReader<R> {
+impl<R: RangeReader + Sync> InvertedIndexReader for InvertedIndexBlobReader<R> {
     async fn range_read(&mut self, offset: u64, size: u32) -> Result<Vec<u8>> {
         let buf = self
             .source

src/index/src/lib.rs

@@ -16,5 +16,7 @@
 #![feature(assert_matches)]
 
 pub mod bloom_filter;
+pub mod error;
+pub mod external_provider;
 pub mod fulltext_index;
 pub mod inverted_index;

src/log-query/Cargo.toml

@@ -11,5 +11,6 @@ workspace = true
 chrono.workspace = true
 common-error.workspace = true
 common-macro.workspace = true
+serde.workspace = true
 snafu.workspace = true
 table.workspace = true

src/log-query/src/error.rs

@@ -15,6 +15,7 @@
 use std::any::Any;
 
 use common_error::ext::ErrorExt;
+use common_error::status_code::StatusCode;
 use common_macro::stack_trace_debug;
 use snafu::Snafu;
 
@@ -41,6 +42,15 @@ impl ErrorExt for Error {
     fn as_any(&self) -> &dyn Any {
         self
     }
+
+    fn status_code(&self) -> StatusCode {
+        match self {
+            Error::InvalidTimeFilter { .. }
+            | Error::InvalidDateFormat { .. }
+            | Error::InvalidSpanFormat { .. }
+            | Error::EndBeforeStart { .. } => StatusCode::InvalidArguments,
+        }
+    }
 }
 
 pub type Result<T> = std::result::Result<T, Error>;

src/log-query/src/log_query.rs

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 use chrono::{DateTime, Datelike, Duration, NaiveDate, NaiveTime, TimeZone, Utc};
+use serde::{Deserialize, Serialize};
 use table::table_name::TableName;
 
 use crate::error::{
@@ -21,9 +22,10 @@ use crate::error::{
 };
 
 /// GreptimeDB's log query request.
+#[derive(Debug, Serialize, Deserialize)]
 pub struct LogQuery {
     /// A fully qualified table name to query logs from.
-    pub table_name: TableName,
+    pub table: TableName,
     /// Specifies the time range for the log query. See [`TimeFilter`] for more details.
     pub time_filter: TimeFilter,
     /// Columns with filters to query.
@@ -34,6 +36,18 @@ pub struct LogQuery {
     pub context: Context,
 }
 
+impl Default for LogQuery {
+    fn default() -> Self {
+        Self {
+            table: TableName::new("", "", ""),
+            time_filter: Default::default(),
+            columns: vec![],
+            limit: None,
+            context: Default::default(),
+        }
+    }
+}
+
 /// Represents a time range for log query.
 ///
 /// This struct allows various formats to express a time range from the user side
@@ -58,7 +72,7 @@ pub struct LogQuery {
 ///
 /// This struct doesn't require a timezone to be presented. When the timezone is not
 /// provided, it will fill the default timezone with the same rules akin to other queries.
-#[derive(Debug, Clone)]
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
 pub struct TimeFilter {
     pub start: Option<String>,
     pub end: Option<String>,
@@ -69,8 +83,7 @@ impl TimeFilter {
     /// Validate and canonicalize the time filter.
     ///
     /// This function will try to fill the missing fields and convert all dates to timestamps
-    // false positive
+    #[allow(unused_assignments)] // false positive
-    #[allow(unused_assignments)]
     pub fn canonicalize(&mut self) -> Result<()> {
         let mut start_dt = None;
         let mut end_dt = None;
@@ -209,6 +222,7 @@ impl TimeFilter {
 }
 
 /// Represents a column with filters to query.
+#[derive(Debug, Serialize, Deserialize)]
 pub struct ColumnFilters {
     /// Case-sensitive column name to query.
     pub column_name: String,
@@ -216,6 +230,7 @@ pub struct ColumnFilters {
     pub filters: Vec<ContentFilter>,
 }
 
+#[derive(Debug, Serialize, Deserialize)]
 pub enum ContentFilter {
     /// Only match the exact content.
     ///
@@ -234,13 +249,16 @@ pub enum ContentFilter {
     Compound(Vec<ContentFilter>, BinaryOperator),
 }
 
+#[derive(Debug, Serialize, Deserialize)]
 pub enum BinaryOperator {
     And,
     Or,
 }
 
 /// Controls how many adjacent lines to return.
+#[derive(Debug, Default, Serialize, Deserialize)]
 pub enum Context {
+    #[default]
     None,
     /// Specify the number of lines before and after the matched line separately.
     Lines(usize, usize),

src/meta-client/src/client.rs

@@ -326,8 +326,8 @@ impl ClusterInfo for MetaClient {
         let cluster_kv_backend = Arc::new(self.cluster_client()?);
         let range_prefix = DatanodeStatKey::key_prefix_with_cluster_id(self.id.0);
         let req = RangeRequest::new().with_prefix(range_prefix);
-        let stream = PaginationStream::new(cluster_kv_backend, req, 256, Arc::new(decode_stats))
+        let stream =
-            .into_stream();
+            PaginationStream::new(cluster_kv_backend, req, 256, decode_stats).into_stream();
         let mut datanode_stats = stream
             .try_collect::<Vec<_>>()
             .await
@@ -994,8 +994,7 @@ mod tests {
 
         let req = RangeRequest::new().with_prefix(b"__prefix/");
         let stream =
-            PaginationStream::new(Arc::new(cluster_client), req, 10, Arc::new(mock_decoder))
+            PaginationStream::new(Arc::new(cluster_client), req, 10, mock_decoder).into_stream();
-                .into_stream();
 
         let res = stream.try_collect::<Vec<_>>().await.unwrap();
         assert_eq!(10, res.len());

src/meta-srv/src/service/store/cached_kv.rs

@@ -102,7 +102,7 @@ impl LeaderCachedKvBackend {
                 self.store.clone(),
                 RangeRequest::new().with_prefix(prefix.as_bytes()),
                 DEFAULT_PAGE_SIZE,
-                Arc::new(Ok),
+                Ok,
             )
             .into_stream();
 

src/metric-engine/src/test_util.rs

@@ -313,12 +313,12 @@ mod test {
         let region_dir = "test_metric_region";
         // assert metadata region's dir
         let metadata_region_dir = join_dir(region_dir, METADATA_REGION_SUBDIR);
-        let exist = object_store.is_exist(&metadata_region_dir).await.unwrap();
+        let exist = object_store.exists(&metadata_region_dir).await.unwrap();
         assert!(exist);
 
         // assert data region's dir
         let data_region_dir = join_dir(region_dir, DATA_REGION_SUBDIR);
-        let exist = object_store.is_exist(&data_region_dir).await.unwrap();
+        let exist = object_store.exists(&data_region_dir).await.unwrap();
         assert!(exist);
 
         // check mito engine

src/mito2/src/cache.rs

@@ -37,7 +37,7 @@ use store_api::storage::{ConcreteDataType, RegionId, TimeSeriesRowSelector};
 
 use crate::cache::cache_size::parquet_meta_size;
 use crate::cache::file_cache::{FileType, IndexKey};
-use crate::cache::index::{InvertedIndexCache, InvertedIndexCacheRef};
+use crate::cache::index::inverted_index::{InvertedIndexCache, InvertedIndexCacheRef};
 use crate::cache::write_cache::WriteCacheRef;
 use crate::metrics::{CACHE_BYTES, CACHE_EVICTION, CACHE_HIT, CACHE_MISS};
 use crate::read::Batch;

src/mito2/src/cache/file_cache.rs

@@ -286,7 +286,7 @@ impl FileCache {
     }
 
     async fn get_reader(&self, file_path: &str) -> object_store::Result<Option<Reader>> {
-        if self.local_store.is_exist(file_path).await? {
+        if self.local_store.exists(file_path).await? {
             Ok(Some(self.local_store.reader(file_path).await?))
         } else {
             Ok(None)
@@ -480,7 +480,7 @@ mod tests {
         cache.memory_index.run_pending_tasks().await;
 
         // The file also not exists.
-        assert!(!local_store.is_exist(&file_path).await.unwrap());
+        assert!(!local_store.exists(&file_path).await.unwrap());
         assert_eq!(0, cache.memory_index.weighted_size());
     }
 

src/mito2/src/cache/index.rs

@@ -12,168 +12,29 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+pub mod inverted_index;
+
+use std::future::Future;
+use std::hash::Hash;
 use std::ops::Range;
 use std::sync::Arc;
 
-use api::v1::index::InvertedIndexMetas;
-use async_trait::async_trait;
 use bytes::Bytes;
-use common_base::BitVec;
-use index::inverted_index::error::DecodeFstSnafu;
-use index::inverted_index::format::reader::InvertedIndexReader;
-use index::inverted_index::FstMap;
 use object_store::Buffer;
-use prost::Message;
-use snafu::ResultExt;
 
 use crate::metrics::{CACHE_BYTES, CACHE_HIT, CACHE_MISS};
-use crate::sst::file::FileId;
 
 /// Metrics for index metadata.
 const INDEX_METADATA_TYPE: &str = "index_metadata";
 /// Metrics for index content.
 const INDEX_CONTENT_TYPE: &str = "index_content";
 
-/// Inverted index blob reader with cache.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
-pub struct CachedInvertedIndexBlobReader<R> {
+pub struct PageKey {
-    file_id: FileId,
-    file_size: u64,
-    inner: R,
-    cache: InvertedIndexCacheRef,
-}
-
-impl<R> CachedInvertedIndexBlobReader<R> {
-    pub fn new(file_id: FileId, file_size: u64, inner: R, cache: InvertedIndexCacheRef) -> Self {
-        Self {
-            file_id,
-            file_size,
-            inner,
-            cache,
-        }
-    }
-}
-
-impl<R> CachedInvertedIndexBlobReader<R>
-where
-    R: InvertedIndexReader,
-{
-    /// Gets given range of index data from cache, and loads from source if the file
-    /// is not already cached.
-    async fn get_or_load(
-        &mut self,
-        offset: u64,
-        size: u32,
-    ) -> index::inverted_index::error::Result<Vec<u8>> {
-        let keys =
-            IndexDataPageKey::generate_page_keys(self.file_id, offset, size, self.cache.page_size);
-        // Size is 0, return empty data.
-        if keys.is_empty() {
-            return Ok(Vec::new());
-        }
-        let mut data = Vec::with_capacity(keys.len());
-        data.resize(keys.len(), Bytes::new());
-        let mut cache_miss_range = vec![];
-        let mut cache_miss_idx = vec![];
-        let last_index = keys.len() - 1;
-        // TODO: Avoid copy as much as possible.
-        for (i, index) in keys.iter().enumerate() {
-            match self.cache.get_index(index) {
-                Some(page) => {
-                    CACHE_HIT.with_label_values(&[INDEX_CONTENT_TYPE]).inc();
-                    data[i] = page;
-                }
-                None => {
-                    CACHE_MISS.with_label_values(&[INDEX_CONTENT_TYPE]).inc();
-                    let base_offset = index.page_id * self.cache.page_size;
-                    let pruned_size = if i == last_index {
-                        prune_size(&keys, self.file_size, self.cache.page_size)
-                    } else {
-                        self.cache.page_size
-                    };
-                    cache_miss_range.push(base_offset..base_offset + pruned_size);
-                    cache_miss_idx.push(i);
-                }
-            }
-        }
-        if !cache_miss_range.is_empty() {
-            let pages = self.inner.read_vec(&cache_miss_range).await?;
-            for (i, page) in cache_miss_idx.into_iter().zip(pages.into_iter()) {
-                let key = keys[i].clone();
-                data[i] = page.clone();
-                self.cache.put_index(key, page.clone());
-            }
-        }
-        let buffer = Buffer::from_iter(data.into_iter());
-        Ok(buffer
-            .slice(IndexDataPageKey::calculate_range(
-                offset,
-                size,
-                self.cache.page_size,
-            ))
-            .to_vec())
-    }
-}
-
-#[async_trait]
-impl<R: InvertedIndexReader> InvertedIndexReader for CachedInvertedIndexBlobReader<R> {
-    async fn range_read(
-        &mut self,
-        offset: u64,
-        size: u32,
-    ) -> index::inverted_index::error::Result<Vec<u8>> {
-        self.inner.range_read(offset, size).await
-    }
-
-    async fn read_vec(
-        &mut self,
-        ranges: &[Range<u64>],
-    ) -> index::inverted_index::error::Result<Vec<Bytes>> {
-        self.inner.read_vec(ranges).await
-    }
-
-    async fn metadata(&mut self) -> index::inverted_index::error::Result<Arc<InvertedIndexMetas>> {
-        if let Some(cached) = self.cache.get_index_metadata(self.file_id) {
-            CACHE_HIT.with_label_values(&[INDEX_METADATA_TYPE]).inc();
-            Ok(cached)
-        } else {
-            let meta = self.inner.metadata().await?;
-            self.cache.put_index_metadata(self.file_id, meta.clone());
-            CACHE_MISS.with_label_values(&[INDEX_METADATA_TYPE]).inc();
-            Ok(meta)
-        }
-    }
-
-    async fn fst(
-        &mut self,
-        offset: u64,
-        size: u32,
-    ) -> index::inverted_index::error::Result<FstMap> {
-        self.get_or_load(offset, size)
-            .await
-            .and_then(|r| FstMap::new(r).context(DecodeFstSnafu))
-    }
-
-    async fn bitmap(
-        &mut self,
-        offset: u64,
-        size: u32,
-    ) -> index::inverted_index::error::Result<BitVec> {
-        self.get_or_load(offset, size).await.map(BitVec::from_vec)
-    }
-}
-
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-pub struct IndexMetadataKey {
-    file_id: FileId,
-}
-
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-pub struct IndexDataPageKey {
-    file_id: FileId,
     page_id: u64,
 }
 
-impl IndexDataPageKey {
+impl PageKey {
     /// Converts an offset to a page ID based on the page size.
     fn calculate_page_id(offset: u64, page_size: u64) -> u64 {
         offset / page_size
@@ -199,49 +60,60 @@ impl IndexDataPageKey {
         start..end
     }
 
-    /// Generates a vector of IndexKey instances for the pages that a given offset and size span.
+    /// Generates a iterator of `IndexKey` for the pages that a given offset and size span.
-    fn generate_page_keys(file_id: FileId, offset: u64, size: u32, page_size: u64) -> Vec<Self> {
+    fn generate_page_keys(offset: u64, size: u32, page_size: u64) -> impl Iterator<Item = Self> {
         let start_page = Self::calculate_page_id(offset, page_size);
         let total_pages = Self::calculate_page_count(offset, size, page_size);
-        (0..total_pages)
+        (0..total_pages).map(move |i| Self {
-            .map(|i| Self {
+            page_id: start_page + i as u64,
-                file_id,
+        })
-                page_id: start_page + i as u64,
-            })
-            .collect()
     }
 }
 
-pub type InvertedIndexCacheRef = Arc<InvertedIndexCache>;
+/// Cache for index metadata and content.
-
+pub struct IndexCache<K, M> {
-pub struct InvertedIndexCache {
+    /// Cache for index metadata
-    /// Cache for inverted index metadata
+    index_metadata: moka::sync::Cache<K, Arc<M>>,
-    index_metadata: moka::sync::Cache<IndexMetadataKey, Arc<InvertedIndexMetas>>,
+    /// Cache for index content.
-    /// Cache for inverted index content.
+    index: moka::sync::Cache<(K, PageKey), Bytes>,
-    index: moka::sync::Cache<IndexDataPageKey, Bytes>,
     // Page size for index content.
     page_size: u64,
+
+    /// Weighter for metadata.
+    weight_of_metadata: fn(&K, &Arc<M>) -> u32,
+    /// Weighter for content.
+    weight_of_content: fn(&(K, PageKey), &Bytes) -> u32,
 }
 
-impl InvertedIndexCache {
+impl<K, M> IndexCache<K, M>
-    /// Creates `InvertedIndexCache` with provided `index_metadata_cap` and `index_content_cap`.
+where
-    pub fn new(index_metadata_cap: u64, index_content_cap: u64, page_size: u64) -> Self {
+    K: Hash + Eq + Send + Sync + 'static,
-        common_telemetry::debug!("Building InvertedIndexCache with metadata size: {index_metadata_cap}, content size: {index_content_cap}");
+    M: Send + Sync + 'static,
+{
+    pub fn new_with_weighter(
+        index_metadata_cap: u64,
+        index_content_cap: u64,
+        page_size: u64,
+        index_type: &'static str,
+        weight_of_metadata: fn(&K, &Arc<M>) -> u32,
+        weight_of_content: fn(&(K, PageKey), &Bytes) -> u32,
+    ) -> Self {
+        common_telemetry::debug!("Building IndexCache with metadata size: {index_metadata_cap}, content size: {index_content_cap}, page size: {page_size}, index type: {index_type}");
         let index_metadata = moka::sync::CacheBuilder::new(index_metadata_cap)
-            .name("inverted_index_metadata")
+            .name(&format!("index_metadata_{}", index_type))
-            .weigher(index_metadata_weight)
+            .weigher(weight_of_metadata)
-            .eviction_listener(|k, v, _cause| {
+            .eviction_listener(move |k, v, _cause| {
-                let size = index_metadata_weight(&k, &v);
+                let size = weight_of_metadata(&k, &v);
                 CACHE_BYTES
                     .with_label_values(&[INDEX_METADATA_TYPE])
                     .sub(size.into());
             })
             .build();
         let index_cache = moka::sync::CacheBuilder::new(index_content_cap)
-            .name("inverted_index_content")
+            .name(&format!("index_content_{}", index_type))
-            .weigher(index_content_weight)
+            .weigher(weight_of_content)
-            .eviction_listener(|k, v, _cause| {
+            .eviction_listener(move |k, v, _cause| {
-                let size = index_content_weight(&k, &v);
+                let size = weight_of_content(&k, &v);
                 CACHE_BYTES
                     .with_label_values(&[INDEX_CONTENT_TYPE])
                     .sub(size.into());
@@ -251,259 +123,109 @@ impl InvertedIndexCache {
             index_metadata,
             index: index_cache,
             page_size,
+            weight_of_content,
+            weight_of_metadata,
         }
     }
 }
 
-impl InvertedIndexCache {
+impl<K, M> IndexCache<K, M>
-    pub fn get_index_metadata(&self, file_id: FileId) -> Option<Arc<InvertedIndexMetas>> {
+where
-        self.index_metadata.get(&IndexMetadataKey { file_id })
+    K: Hash + Eq + Clone + Copy + Send + Sync + 'static,
+    M: Send + Sync + 'static,
+{
+    pub fn get_metadata(&self, key: K) -> Option<Arc<M>> {
+        self.index_metadata.get(&key)
     }
 
-    pub fn put_index_metadata(&self, file_id: FileId, metadata: Arc<InvertedIndexMetas>) {
+    pub fn put_metadata(&self, key: K, metadata: Arc<M>) {
-        let key = IndexMetadataKey { file_id };
         CACHE_BYTES
             .with_label_values(&[INDEX_METADATA_TYPE])
-            .add(index_metadata_weight(&key, &metadata).into());
+            .add((self.weight_of_metadata)(&key, &metadata).into());
         self.index_metadata.insert(key, metadata)
     }
 
-    pub fn get_index(&self, key: &IndexDataPageKey) -> Option<Bytes> {
+    /// Gets given range of index data from cache, and loads from source if the file
-        self.index.get(key)
+    /// is not already cached.
+    async fn get_or_load<F, Fut, E>(
+        &self,
+        key: K,
+        file_size: u64,
+        offset: u64,
+        size: u32,
+        load: F,
+    ) -> Result<Vec<u8>, E>
+    where
+        F: FnOnce(Vec<Range<u64>>) -> Fut,
+        Fut: Future<Output = Result<Vec<Bytes>, E>>,
+        E: std::error::Error,
+    {
+        let page_keys =
+            PageKey::generate_page_keys(offset, size, self.page_size).collect::<Vec<_>>();
+        // Size is 0, return empty data.
+        if page_keys.is_empty() {
+            return Ok(Vec::new());
+        }
+        let mut data = Vec::with_capacity(page_keys.len());
+        data.resize(page_keys.len(), Bytes::new());
+        let mut cache_miss_range = vec![];
+        let mut cache_miss_idx = vec![];
+        let last_index = page_keys.len() - 1;
+        // TODO: Avoid copy as much as possible.
+        for (i, page_key) in page_keys.iter().enumerate() {
+            match self.get_page(key, *page_key) {
+                Some(page) => {
+                    CACHE_HIT.with_label_values(&[INDEX_CONTENT_TYPE]).inc();
+                    data[i] = page;
+                }
+                None => {
+                    CACHE_MISS.with_label_values(&[INDEX_CONTENT_TYPE]).inc();
+                    let base_offset = page_key.page_id * self.page_size;
+                    let pruned_size = if i == last_index {
+                        prune_size(page_keys.iter(), file_size, self.page_size)
+                    } else {
+                        self.page_size
+                    };
+                    cache_miss_range.push(base_offset..base_offset + pruned_size);
+                    cache_miss_idx.push(i);
+                }
+            }
+        }
+        if !cache_miss_range.is_empty() {
+            let pages = load(cache_miss_range).await?;
+            for (i, page) in cache_miss_idx.into_iter().zip(pages.into_iter()) {
+                let page_key = page_keys[i];
+                data[i] = page.clone();
+                self.put_page(key, page_key, page.clone());
+            }
+        }
+        let buffer = Buffer::from_iter(data.into_iter());
+        Ok(buffer
+            .slice(PageKey::calculate_range(offset, size, self.page_size))
+            .to_vec())
     }
 
-    pub fn put_index(&self, key: IndexDataPageKey, value: Bytes) {
+    fn get_page(&self, key: K, page_key: PageKey) -> Option<Bytes> {
+        self.index.get(&(key, page_key))
+    }
+
+    fn put_page(&self, key: K, page_key: PageKey, value: Bytes) {
         CACHE_BYTES
             .with_label_values(&[INDEX_CONTENT_TYPE])
-            .add(index_content_weight(&key, &value).into());
+            .add((self.weight_of_content)(&(key, page_key), &value).into());
-        self.index.insert(key, value);
+        self.index.insert((key, page_key), value);
     }
 }
 
-/// Calculates weight for index metadata.
-fn index_metadata_weight(k: &IndexMetadataKey, v: &Arc<InvertedIndexMetas>) -> u32 {
-    (k.file_id.as_bytes().len() + v.encoded_len()) as u32
-}
-
-/// Calculates weight for index content.
-fn index_content_weight(k: &IndexDataPageKey, v: &Bytes) -> u32 {
-    (k.file_id.as_bytes().len() + v.len()) as u32
-}
-
 /// Prunes the size of the last page based on the indexes.
 /// We have following cases:
 /// 1. The rest file size is less than the page size, read to the end of the file.
 /// 2. Otherwise, read the page size.
-fn prune_size(indexes: &[IndexDataPageKey], file_size: u64, page_size: u64) -> u64 {
+fn prune_size<'a>(
+    indexes: impl Iterator<Item = &'a PageKey>,
+    file_size: u64,
+    page_size: u64,
+) -> u64 {
     let last_page_start = indexes.last().map(|i| i.page_id * page_size).unwrap_or(0);
     page_size.min(file_size - last_page_start)
 }
-
-#[cfg(test)]
-mod test {
-    use std::num::NonZeroUsize;
-
-    use common_base::BitVec;
-    use futures::stream;
-    use index::inverted_index::format::reader::{InvertedIndexBlobReader, InvertedIndexReader};
-    use index::inverted_index::format::writer::{InvertedIndexBlobWriter, InvertedIndexWriter};
-    use index::inverted_index::Bytes;
-    use prometheus::register_int_counter_vec;
-    use rand::{Rng, RngCore};
-
-    use super::*;
-    use crate::sst::index::store::InstrumentedStore;
-    use crate::test_util::TestEnv;
-
-    // Repeat times for following little fuzz tests.
-    const FUZZ_REPEAT_TIMES: usize = 100;
-
-    // Fuzz test for index data page key
-    #[test]
-    fn fuzz_index_calculation() {
-        // randomly generate a large u8 array
-        let mut rng = rand::thread_rng();
-        let mut data = vec![0u8; 1024 * 1024];
-        rng.fill_bytes(&mut data);
-        let file_id = FileId::random();
-
-        for _ in 0..FUZZ_REPEAT_TIMES {
-            let offset = rng.gen_range(0..data.len() as u64);
-            let size = rng.gen_range(0..data.len() as u32 - offset as u32);
-            let page_size: usize = rng.gen_range(1..1024);
-
-            let indexes =
-                IndexDataPageKey::generate_page_keys(file_id, offset, size, page_size as u64);
-            let page_num = indexes.len();
-            let mut read = Vec::with_capacity(size as usize);
-            for key in indexes.into_iter() {
-                let start = key.page_id as usize * page_size;
-                let page = if start + page_size < data.len() {
-                    &data[start..start + page_size]
-                } else {
-                    &data[start..]
-                };
-                read.extend_from_slice(page);
-            }
-            let expected_range = offset as usize..(offset + size as u64 as u64) as usize;
-            let read =
-                read[IndexDataPageKey::calculate_range(offset, size, page_size as u64)].to_vec();
-            if read != data.get(expected_range).unwrap() {
-                panic!(
-                    "fuzz_read_index failed, offset: {}, size: {}, page_size: {}\nread len: {}, expected len: {}\nrange: {:?}, page num: {}",
-                    offset, size, page_size, read.len(), size as usize,
-                    IndexDataPageKey::calculate_range(offset, size, page_size as u64),
-                    page_num
-                );
-            }
-        }
-    }
-
-    fn unpack(fst_value: u64) -> [u32; 2] {
-        bytemuck::cast::<u64, [u32; 2]>(fst_value)
-    }
-
-    async fn create_inverted_index_blob() -> Vec<u8> {
-        let mut blob = Vec::new();
-        let mut writer = InvertedIndexBlobWriter::new(&mut blob);
-        writer
-            .add_index(
-                "tag0".to_string(),
-                BitVec::from_slice(&[0b0000_0001, 0b0000_0000]),
-                Box::new(stream::iter(vec![
-                    Ok((Bytes::from("a"), BitVec::from_slice(&[0b0000_0001]))),
-                    Ok((Bytes::from("b"), BitVec::from_slice(&[0b0010_0000]))),
-                    Ok((Bytes::from("c"), BitVec::from_slice(&[0b0000_0001]))),
-                ])),
-            )
-            .await
-            .unwrap();
-        writer
-            .add_index(
-                "tag1".to_string(),
-                BitVec::from_slice(&[0b0000_0001, 0b0000_0000]),
-                Box::new(stream::iter(vec![
-                    Ok((Bytes::from("x"), BitVec::from_slice(&[0b0000_0001]))),
-                    Ok((Bytes::from("y"), BitVec::from_slice(&[0b0010_0000]))),
-                    Ok((Bytes::from("z"), BitVec::from_slice(&[0b0000_0001]))),
-                ])),
-            )
-            .await
-            .unwrap();
-        writer
-            .finish(8, NonZeroUsize::new(1).unwrap())
-            .await
-            .unwrap();
-
-        blob
-    }
-
-    #[tokio::test]
-    async fn test_inverted_index_cache() {
-        let blob = create_inverted_index_blob().await;
-
-        // Init a test range reader in local fs.
-        let mut env = TestEnv::new();
-        let file_size = blob.len() as u64;
-        let store = env.init_object_store_manager();
-        let temp_path = "data";
-        store.write(temp_path, blob).await.unwrap();
-        let store = InstrumentedStore::new(store);
-        let metric =
-            register_int_counter_vec!("test_bytes", "a counter for test", &["test"]).unwrap();
-        let counter = metric.with_label_values(&["test"]);
-        let range_reader = store
-            .range_reader("data", &counter, &counter)
-            .await
-            .unwrap();
-
-        let reader = InvertedIndexBlobReader::new(range_reader);
-        let mut cached_reader = CachedInvertedIndexBlobReader::new(
-            FileId::random(),
-            file_size,
-            reader,
-            Arc::new(InvertedIndexCache::new(8192, 8192, 50)),
-        );
-        let metadata = cached_reader.metadata().await.unwrap();
-        assert_eq!(metadata.total_row_count, 8);
-        assert_eq!(metadata.segment_row_count, 1);
-        assert_eq!(metadata.metas.len(), 2);
-        // tag0
-        let tag0 = metadata.metas.get("tag0").unwrap();
-        let stats0 = tag0.stats.as_ref().unwrap();
-        assert_eq!(stats0.distinct_count, 3);
-        assert_eq!(stats0.null_count, 1);
-        assert_eq!(stats0.min_value, Bytes::from("a"));
-        assert_eq!(stats0.max_value, Bytes::from("c"));
-        let fst0 = cached_reader
-            .fst(
-                tag0.base_offset + tag0.relative_fst_offset as u64,
-                tag0.fst_size,
-            )
-            .await
-            .unwrap();
-        assert_eq!(fst0.len(), 3);
-        let [offset, size] = unpack(fst0.get(b"a").unwrap());
-        let bitmap = cached_reader
-            .bitmap(tag0.base_offset + offset as u64, size)
-            .await
-            .unwrap();
-        assert_eq!(bitmap, BitVec::from_slice(&[0b0000_0001]));
-        let [offset, size] = unpack(fst0.get(b"b").unwrap());
-        let bitmap = cached_reader
-            .bitmap(tag0.base_offset + offset as u64, size)
-            .await
-            .unwrap();
-        assert_eq!(bitmap, BitVec::from_slice(&[0b0010_0000]));
-        let [offset, size] = unpack(fst0.get(b"c").unwrap());
-        let bitmap = cached_reader
-            .bitmap(tag0.base_offset + offset as u64, size)
-            .await
-            .unwrap();
-        assert_eq!(bitmap, BitVec::from_slice(&[0b0000_0001]));
-
-        // tag1
-        let tag1 = metadata.metas.get("tag1").unwrap();
-        let stats1 = tag1.stats.as_ref().unwrap();
-        assert_eq!(stats1.distinct_count, 3);
-        assert_eq!(stats1.null_count, 1);
-        assert_eq!(stats1.min_value, Bytes::from("x"));
-        assert_eq!(stats1.max_value, Bytes::from("z"));
-        let fst1 = cached_reader
-            .fst(
-                tag1.base_offset + tag1.relative_fst_offset as u64,
-                tag1.fst_size,
-            )
-            .await
-            .unwrap();
-        assert_eq!(fst1.len(), 3);
-        let [offset, size] = unpack(fst1.get(b"x").unwrap());
-        let bitmap = cached_reader
-            .bitmap(tag1.base_offset + offset as u64, size)
-            .await
-            .unwrap();
-        assert_eq!(bitmap, BitVec::from_slice(&[0b0000_0001]));
-        let [offset, size] = unpack(fst1.get(b"y").unwrap());
-        let bitmap = cached_reader
-            .bitmap(tag1.base_offset + offset as u64, size)
-            .await
-            .unwrap();
-        assert_eq!(bitmap, BitVec::from_slice(&[0b0010_0000]));
-        let [offset, size] = unpack(fst1.get(b"z").unwrap());
-        let bitmap = cached_reader
-            .bitmap(tag1.base_offset + offset as u64, size)
-            .await
-            .unwrap();
-        assert_eq!(bitmap, BitVec::from_slice(&[0b0000_0001]));
-
-        // fuzz test
-        let mut rng = rand::thread_rng();
-        for _ in 0..FUZZ_REPEAT_TIMES {
-            let offset = rng.gen_range(0..file_size);
-            let size = rng.gen_range(0..file_size as u32 - offset as u32);
-            let expected = cached_reader.range_read(offset, size).await.unwrap();
-            let read = cached_reader.get_or_load(offset, size).await.unwrap();
-            assert_eq!(read, expected);
-        }
-    }
-}

src/mito2/src/cache/index/inverted_index.rs

@@ -0,0 +1,322 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::sync::Arc;
+
+use api::v1::index::InvertedIndexMetas;
+use async_trait::async_trait;
+use bytes::Bytes;
+use index::inverted_index::error::Result;
+use index::inverted_index::format::reader::InvertedIndexReader;
+use prost::Message;
+
+use crate::cache::index::{IndexCache, PageKey, INDEX_METADATA_TYPE};
+use crate::metrics::{CACHE_HIT, CACHE_MISS};
+use crate::sst::file::FileId;
+
+const INDEX_TYPE_INVERTED_INDEX: &str = "inverted_index";
+
+/// Cache for inverted index.
+pub type InvertedIndexCache = IndexCache<FileId, InvertedIndexMetas>;
+pub type InvertedIndexCacheRef = Arc<InvertedIndexCache>;
+
+impl InvertedIndexCache {
+    /// Creates a new inverted index cache.
+    pub fn new(index_metadata_cap: u64, index_content_cap: u64, page_size: u64) -> Self {
+        Self::new_with_weighter(
+            index_metadata_cap,
+            index_content_cap,
+            page_size,
+            INDEX_TYPE_INVERTED_INDEX,
+            inverted_index_metadata_weight,
+            inverted_index_content_weight,
+        )
+    }
+}
+
+/// Calculates weight for inverted index metadata.
+fn inverted_index_metadata_weight(k: &FileId, v: &Arc<InvertedIndexMetas>) -> u32 {
+    (k.as_bytes().len() + v.encoded_len()) as u32
+}
+
+/// Calculates weight for inverted index content.
+fn inverted_index_content_weight((k, _): &(FileId, PageKey), v: &Bytes) -> u32 {
+    (k.as_bytes().len() + v.len()) as u32
+}
+
+/// Inverted index blob reader with cache.
+pub struct CachedInvertedIndexBlobReader<R> {
+    file_id: FileId,
+    file_size: u64,
+    inner: R,
+    cache: InvertedIndexCacheRef,
+}
+
+impl<R> CachedInvertedIndexBlobReader<R> {
+    /// Creates a new inverted index blob reader with cache.
+    pub fn new(file_id: FileId, file_size: u64, inner: R, cache: InvertedIndexCacheRef) -> Self {
+        Self {
+            file_id,
+            file_size,
+            inner,
+            cache,
+        }
+    }
+}
+
+#[async_trait]
+impl<R: InvertedIndexReader> InvertedIndexReader for CachedInvertedIndexBlobReader<R> {
+    async fn range_read(&mut self, offset: u64, size: u32) -> Result<Vec<u8>> {
+        let inner = &mut self.inner;
+        self.cache
+            .get_or_load(
+                self.file_id,
+                self.file_size,
+                offset,
+                size,
+                move |ranges| async move { inner.read_vec(&ranges).await },
+            )
+            .await
+    }
+
+    async fn metadata(&mut self) -> Result<Arc<InvertedIndexMetas>> {
+        if let Some(cached) = self.cache.get_metadata(self.file_id) {
+            CACHE_HIT.with_label_values(&[INDEX_METADATA_TYPE]).inc();
+            Ok(cached)
+        } else {
+            let meta = self.inner.metadata().await?;
+            self.cache.put_metadata(self.file_id, meta.clone());
+            CACHE_MISS.with_label_values(&[INDEX_METADATA_TYPE]).inc();
+            Ok(meta)
+        }
+    }
+}
+
+#[cfg(test)]
+mod test {
+    use std::num::NonZeroUsize;
+
+    use common_base::BitVec;
+    use futures::stream;
+    use index::inverted_index::format::reader::{InvertedIndexBlobReader, InvertedIndexReader};
+    use index::inverted_index::format::writer::{InvertedIndexBlobWriter, InvertedIndexWriter};
+    use index::inverted_index::Bytes;
+    use prometheus::register_int_counter_vec;
+    use rand::{Rng, RngCore};
+
+    use super::*;
+    use crate::sst::index::store::InstrumentedStore;
+    use crate::test_util::TestEnv;
+
+    // Repeat times for following little fuzz tests.
+    const FUZZ_REPEAT_TIMES: usize = 100;
+
+    // Fuzz test for index data page key
+    #[test]
+    fn fuzz_index_calculation() {
+        // randomly generate a large u8 array
+        let mut rng = rand::thread_rng();
+        let mut data = vec![0u8; 1024 * 1024];
+        rng.fill_bytes(&mut data);
+
+        for _ in 0..FUZZ_REPEAT_TIMES {
+            let offset = rng.gen_range(0..data.len() as u64);
+            let size = rng.gen_range(0..data.len() as u32 - offset as u32);
+            let page_size: usize = rng.gen_range(1..1024);
+
+            let indexes =
+                PageKey::generate_page_keys(offset, size, page_size as u64).collect::<Vec<_>>();
+            let page_num = indexes.len();
+            let mut read = Vec::with_capacity(size as usize);
+            for key in indexes.into_iter() {
+                let start = key.page_id as usize * page_size;
+                let page = if start + page_size < data.len() {
+                    &data[start..start + page_size]
+                } else {
+                    &data[start..]
+                };
+                read.extend_from_slice(page);
+            }
+            let expected_range = offset as usize..(offset + size as u64 as u64) as usize;
+            let read = read[PageKey::calculate_range(offset, size, page_size as u64)].to_vec();
+            if read != data.get(expected_range).unwrap() {
+                panic!(
+                    "fuzz_read_index failed, offset: {}, size: {}, page_size: {}\nread len: {}, expected len: {}\nrange: {:?}, page num: {}",
+                    offset, size, page_size, read.len(), size as usize,
+                    PageKey::calculate_range(offset, size, page_size as u64),
+                    page_num
+                );
+            }
+        }
+    }
+
+    fn unpack(fst_value: u64) -> [u32; 2] {
+        bytemuck::cast::<u64, [u32; 2]>(fst_value)
+    }
+
+    async fn create_inverted_index_blob() -> Vec<u8> {
+        let mut blob = Vec::new();
+        let mut writer = InvertedIndexBlobWriter::new(&mut blob);
+        writer
+            .add_index(
+                "tag0".to_string(),
+                BitVec::from_slice(&[0b0000_0001, 0b0000_0000]),
+                Box::new(stream::iter(vec![
+                    Ok((Bytes::from("a"), BitVec::from_slice(&[0b0000_0001]))),
+                    Ok((Bytes::from("b"), BitVec::from_slice(&[0b0010_0000]))),
+                    Ok((Bytes::from("c"), BitVec::from_slice(&[0b0000_0001]))),
+                ])),
+            )
+            .await
+            .unwrap();
+        writer
+            .add_index(
+                "tag1".to_string(),
+                BitVec::from_slice(&[0b0000_0001, 0b0000_0000]),
+                Box::new(stream::iter(vec![
+                    Ok((Bytes::from("x"), BitVec::from_slice(&[0b0000_0001]))),
+                    Ok((Bytes::from("y"), BitVec::from_slice(&[0b0010_0000]))),
+                    Ok((Bytes::from("z"), BitVec::from_slice(&[0b0000_0001]))),
+                ])),
+            )
+            .await
+            .unwrap();
+        writer
+            .finish(8, NonZeroUsize::new(1).unwrap())
+            .await
+            .unwrap();
+
+        blob
+    }
+
+    #[tokio::test]
+    async fn test_inverted_index_cache() {
+        let blob = create_inverted_index_blob().await;
+
+        // Init a test range reader in local fs.
+        let mut env = TestEnv::new();
+        let file_size = blob.len() as u64;
+        let store = env.init_object_store_manager();
+        let temp_path = "data";
+        store.write(temp_path, blob).await.unwrap();
+        let store = InstrumentedStore::new(store);
+        let metric =
+            register_int_counter_vec!("test_bytes", "a counter for test", &["test"]).unwrap();
+        let counter = metric.with_label_values(&["test"]);
+        let range_reader = store
+            .range_reader("data", &counter, &counter)
+            .await
+            .unwrap();
+
+        let reader = InvertedIndexBlobReader::new(range_reader);
+        let mut cached_reader = CachedInvertedIndexBlobReader::new(
+            FileId::random(),
+            file_size,
+            reader,
+            Arc::new(InvertedIndexCache::new(8192, 8192, 50)),
+        );
+        let metadata = cached_reader.metadata().await.unwrap();
+        assert_eq!(metadata.total_row_count, 8);
+        assert_eq!(metadata.segment_row_count, 1);
+        assert_eq!(metadata.metas.len(), 2);
+        // tag0
+        let tag0 = metadata.metas.get("tag0").unwrap();
+        let stats0 = tag0.stats.as_ref().unwrap();
+        assert_eq!(stats0.distinct_count, 3);
+        assert_eq!(stats0.null_count, 1);
+        assert_eq!(stats0.min_value, Bytes::from("a"));
+        assert_eq!(stats0.max_value, Bytes::from("c"));
+        let fst0 = cached_reader
+            .fst(
+                tag0.base_offset + tag0.relative_fst_offset as u64,
+                tag0.fst_size,
+            )
+            .await
+            .unwrap();
+        assert_eq!(fst0.len(), 3);
+        let [offset, size] = unpack(fst0.get(b"a").unwrap());
+        let bitmap = cached_reader
+            .bitmap(tag0.base_offset + offset as u64, size)
+            .await
+            .unwrap();
+        assert_eq!(bitmap, BitVec::from_slice(&[0b0000_0001]));
+        let [offset, size] = unpack(fst0.get(b"b").unwrap());
+        let bitmap = cached_reader
+            .bitmap(tag0.base_offset + offset as u64, size)
+            .await
+            .unwrap();
+        assert_eq!(bitmap, BitVec::from_slice(&[0b0010_0000]));
+        let [offset, size] = unpack(fst0.get(b"c").unwrap());
+        let bitmap = cached_reader
+            .bitmap(tag0.base_offset + offset as u64, size)
+            .await
+            .unwrap();
+        assert_eq!(bitmap, BitVec::from_slice(&[0b0000_0001]));
+
+        // tag1
+        let tag1 = metadata.metas.get("tag1").unwrap();
+        let stats1 = tag1.stats.as_ref().unwrap();
+        assert_eq!(stats1.distinct_count, 3);
+        assert_eq!(stats1.null_count, 1);
+        assert_eq!(stats1.min_value, Bytes::from("x"));
+        assert_eq!(stats1.max_value, Bytes::from("z"));
+        let fst1 = cached_reader
+            .fst(
+                tag1.base_offset + tag1.relative_fst_offset as u64,
+                tag1.fst_size,
+            )
+            .await
+            .unwrap();
+        assert_eq!(fst1.len(), 3);
+        let [offset, size] = unpack(fst1.get(b"x").unwrap());
+        let bitmap = cached_reader
+            .bitmap(tag1.base_offset + offset as u64, size)
+            .await
+            .unwrap();
+        assert_eq!(bitmap, BitVec::from_slice(&[0b0000_0001]));
+        let [offset, size] = unpack(fst1.get(b"y").unwrap());
+        let bitmap = cached_reader
+            .bitmap(tag1.base_offset + offset as u64, size)
+            .await
+            .unwrap();
+        assert_eq!(bitmap, BitVec::from_slice(&[0b0010_0000]));
+        let [offset, size] = unpack(fst1.get(b"z").unwrap());
+        let bitmap = cached_reader
+            .bitmap(tag1.base_offset + offset as u64, size)
+            .await
+            .unwrap();
+        assert_eq!(bitmap, BitVec::from_slice(&[0b0000_0001]));
+
+        // fuzz test
+        let mut rng = rand::thread_rng();
+        for _ in 0..FUZZ_REPEAT_TIMES {
+            let offset = rng.gen_range(0..file_size);
+            let size = rng.gen_range(0..file_size as u32 - offset as u32);
+            let expected = cached_reader.range_read(offset, size).await.unwrap();
+            let inner = &mut cached_reader.inner;
+            let read = cached_reader
+                .cache
+                .get_or_load(
+                    cached_reader.file_id,
+                    file_size,
+                    offset,
+                    size,
+                    |ranges| async move { inner.read_vec(&ranges).await },
+                )
+                .await
+                .unwrap();
+            assert_eq!(read, expected);
+        }
+    }
+}

src/mito2/src/engine/create_test.rs

@@ -192,12 +192,12 @@ async fn test_engine_create_with_custom_store() {
     assert!(object_store_manager
         .find("Gcs")
         .unwrap()
-        .is_exist(region_dir)
+        .exists(region_dir)
         .await
         .unwrap());
     assert!(!object_store_manager
         .default_object_store()
-        .is_exist(region_dir)
+        .exists(region_dir)
         .await
         .unwrap());
 }

src/mito2/src/engine/drop_test.rs

@@ -71,7 +71,7 @@ async fn test_engine_drop_region() {
     assert!(!env
         .get_object_store()
         .unwrap()
-        .is_exist(&join_path(&region_dir, DROPPING_MARKER_FILE))
+        .exists(&join_path(&region_dir, DROPPING_MARKER_FILE))
         .await
         .unwrap());
 
@@ -93,7 +93,7 @@ async fn test_engine_drop_region() {
     listener.wait().await;
 
     let object_store = env.get_object_store().unwrap();
-    assert!(!object_store.is_exist(&region_dir).await.unwrap());
+    assert!(!object_store.exists(&region_dir).await.unwrap());
 }
 
 #[tokio::test]
@@ -167,13 +167,13 @@ async fn test_engine_drop_region_for_custom_store() {
     assert!(object_store_manager
         .find("Gcs")
         .unwrap()
-        .is_exist(&custom_region_dir)
+        .exists(&custom_region_dir)
         .await
         .unwrap());
     assert!(object_store_manager
         .find("default")
         .unwrap()
-        .is_exist(&global_region_dir)
+        .exists(&global_region_dir)
         .await
         .unwrap());
 
@@ -190,13 +190,13 @@ async fn test_engine_drop_region_for_custom_store() {
     assert!(!object_store_manager
         .find("Gcs")
         .unwrap()
-        .is_exist(&custom_region_dir)
+        .exists(&custom_region_dir)
         .await
         .unwrap());
     assert!(object_store_manager
         .find("default")
         .unwrap()
-        .is_exist(&global_region_dir)
+        .exists(&global_region_dir)
         .await
         .unwrap());
 }

src/mito2/src/engine/open_test.rs

@@ -228,13 +228,13 @@ async fn test_engine_region_open_with_custom_store() {
     let object_store_manager = env.get_object_store_manager().unwrap();
     assert!(!object_store_manager
         .default_object_store()
-        .is_exist(region.access_layer.region_dir())
+        .exists(region.access_layer.region_dir())
         .await
         .unwrap());
     assert!(object_store_manager
         .find("Gcs")
         .unwrap()
-        .is_exist(region.access_layer.region_dir())
+        .exists(region.access_layer.region_dir())
         .await
         .unwrap());
 }

src/mito2/src/error.rs

@@ -723,10 +723,20 @@ pub enum Error {
 
     #[snafu(display("Failed to iter data part"))]
     ReadDataPart {
+        #[snafu(implicit)]
+        location: Location,
         #[snafu(source)]
         error: parquet::errors::ParquetError,
     },
 
+    #[snafu(display("Failed to read row group in memtable"))]
+    DecodeArrowRowGroup {
+        #[snafu(source)]
+        error: ArrowError,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
     #[snafu(display("Invalid region options, {}", reason))]
     InvalidRegionOptions {
         reason: String,
@@ -1029,6 +1039,7 @@ impl ErrorExt for Error {
             RegionBusy { .. } => StatusCode::RegionBusy,
             GetSchemaMetadata { source, .. } => source.status_code(),
             Timeout { .. } => StatusCode::Cancelled,
+            DecodeArrowRowGroup { .. } => StatusCode::Internal,
         }
     }
 

src/mito2/src/manifest/tests/checkpoint.rs

@@ -84,6 +84,7 @@ async fn manager_without_checkpoint() {
 
     // check files
     let mut expected = vec![
+        "/",
         "00000000000000000010.json",
         "00000000000000000009.json",
         "00000000000000000008.json",
@@ -130,6 +131,7 @@ async fn manager_with_checkpoint_distance_1() {
 
     // check files
     let mut expected = vec![
+        "/",
         "00000000000000000009.checkpoint",
         "00000000000000000010.checkpoint",
         "00000000000000000010.json",

src/mito2/src/memtable/bulk.rs

@@ -27,8 +27,12 @@ use crate::memtable::{
     BoxedBatchIterator, KeyValues, Memtable, MemtableId, MemtableRanges, MemtableRef, MemtableStats,
 };
 
+#[allow(unused)]
+mod context;
 #[allow(unused)]
 pub(crate) mod part;
+mod part_reader;
+mod row_group_reader;
 
 #[derive(Debug)]
 pub struct BulkMemtable {

src/mito2/src/memtable/bulk/context.rs

@@ -0,0 +1,117 @@
+// 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.
+
+//! Context for iterating bulk memtable.
+
+use std::collections::VecDeque;
+use std::sync::Arc;
+
+use parquet::file::metadata::ParquetMetaData;
+use store_api::metadata::RegionMetadataRef;
+use store_api::storage::ColumnId;
+use table::predicate::Predicate;
+
+use crate::row_converter::McmpRowCodec;
+use crate::sst::parquet::file_range::RangeBase;
+use crate::sst::parquet::format::ReadFormat;
+use crate::sst::parquet::reader::SimpleFilterContext;
+use crate::sst::parquet::stats::RowGroupPruningStats;
+
+pub(crate) type BulkIterContextRef = Arc<BulkIterContext>;
+
+pub(crate) struct BulkIterContext {
+    pub(crate) base: RangeBase,
+    pub(crate) predicate: Option<Predicate>,
+}
+
+impl BulkIterContext {
+    pub(crate) fn new(
+        region_metadata: RegionMetadataRef,
+        projection: &Option<&[ColumnId]>,
+        predicate: Option<Predicate>,
+    ) -> Self {
+        let codec = McmpRowCodec::new_with_primary_keys(&region_metadata);
+
+        let simple_filters = predicate
+            .as_ref()
+            .iter()
+            .flat_map(|predicate| {
+                predicate
+                    .exprs()
+                    .iter()
+                    .filter_map(|expr| SimpleFilterContext::new_opt(&region_metadata, None, expr))
+            })
+            .collect();
+
+        let read_format = build_read_format(region_metadata, projection);
+
+        Self {
+            base: RangeBase {
+                filters: simple_filters,
+                read_format,
+                codec,
+                // we don't need to compat batch since all batch in memtable have the same schema.
+                compat_batch: None,
+            },
+            predicate,
+        }
+    }
+
+    /// Prunes row groups by stats.
+    pub(crate) fn row_groups_to_read(&self, file_meta: &Arc<ParquetMetaData>) -> VecDeque<usize> {
+        let region_meta = self.base.read_format.metadata();
+        let row_groups = file_meta.row_groups();
+        // expected_metadata is set to None since we always expect region metadata of memtable is up-to-date.
+        let stats = RowGroupPruningStats::new(row_groups, &self.base.read_format, None);
+        if let Some(predicate) = self.predicate.as_ref() {
+            predicate
+                .prune_with_stats(&stats, region_meta.schema.arrow_schema())
+                .iter()
+                .zip(0..file_meta.num_row_groups())
+                .filter_map(|(selected, row_group)| {
+                    if !*selected {
+                        return None;
+                    }
+                    Some(row_group)
+                })
+                .collect::<VecDeque<_>>()
+        } else {
+            (0..file_meta.num_row_groups()).collect()
+        }
+    }
+
+    pub(crate) fn read_format(&self) -> &ReadFormat {
+        &self.base.read_format
+    }
+}
+
+fn build_read_format(
+    region_metadata: RegionMetadataRef,
+    projection: &Option<&[ColumnId]>,
+) -> ReadFormat {
+    let read_format = if let Some(column_ids) = &projection {
+        ReadFormat::new(region_metadata, column_ids.iter().copied())
+    } else {
+        // No projection, lists all column ids to read.
+        ReadFormat::new(
+            region_metadata.clone(),
+            region_metadata
+                .column_metadatas
+                .iter()
+                .map(|col| col.column_id),
+        )
+    };
+
+    read_format
+}

src/mito2/src/memtable/bulk/part.rs

@@ -13,10 +13,12 @@
 // limitations under the License.
 
 //! Bulk part encoder/decoder.
+
 use std::collections::VecDeque;
 use std::sync::Arc;
 
 use api::v1::Mutation;
+use bytes::Bytes;
 use common_time::timestamp::TimeUnit;
 use datafusion::arrow::array::{TimestampNanosecondArray, UInt64Builder};
 use datatypes::arrow;
@@ -26,93 +28,145 @@ use datatypes::arrow::array::{
     UInt8Builder,
 };
 use datatypes::arrow::compute::TakeOptions;
-use datatypes::arrow::datatypes::{DataType as ArrowDataType, SchemaRef};
+use datatypes::arrow::datatypes::SchemaRef;
 use datatypes::arrow_array::BinaryArray;
 use datatypes::data_type::DataType;
 use datatypes::prelude::{MutableVector, ScalarVectorBuilder, Vector};
-use datatypes::types::TimestampType;
+use parquet::arrow::arrow_reader::{ArrowReaderMetadata, ArrowReaderOptions};
 use parquet::arrow::ArrowWriter;
 use parquet::data_type::AsBytes;
+use parquet::file::metadata::ParquetMetaData;
+use parquet::file::properties::WriterProperties;
 use snafu::ResultExt;
 use store_api::metadata::RegionMetadataRef;
+use store_api::storage::ColumnId;
+use table::predicate::Predicate;
 
+use crate::error;
 use crate::error::{ComputeArrowSnafu, EncodeMemtableSnafu, NewRecordBatchSnafu, Result};
+use crate::memtable::bulk::context::BulkIterContextRef;
+use crate::memtable::bulk::part_reader::BulkPartIter;
 use crate::memtable::key_values::KeyValuesRef;
-use crate::read::Batch;
+use crate::memtable::BoxedBatchIterator;
 use crate::row_converter::{McmpRowCodec, RowCodec};
-use crate::sst::parquet::format::PrimaryKeyArray;
+use crate::sst::parquet::format::{PrimaryKeyArray, ReadFormat};
+use crate::sst::parquet::helper::parse_parquet_metadata;
 use crate::sst::to_sst_arrow_schema;
 
 #[derive(Debug)]
 pub struct BulkPart {
-    data: Vec<u8>,
+    data: Bytes,
     metadata: BulkPartMeta,
 }
 
 impl BulkPart {
-    pub fn new(data: Vec<u8>, metadata: BulkPartMeta) -> Self {
+    pub fn new(data: Bytes, metadata: BulkPartMeta) -> Self {
         Self { data, metadata }
     }
 
     pub(crate) fn metadata(&self) -> &BulkPartMeta {
         &self.metadata
     }
+
+    pub(crate) fn read(&self, context: BulkIterContextRef) -> Result<Option<BoxedBatchIterator>> {
+        // use predicate to find row groups to read.
+        let row_groups_to_read = context.row_groups_to_read(&self.metadata.parquet_metadata);
+
+        if row_groups_to_read.is_empty() {
+            // All row groups are filtered.
+            return Ok(None);
+        }
+
+        let iter = BulkPartIter::try_new(
+            context,
+            row_groups_to_read,
+            self.metadata.parquet_metadata.clone(),
+            self.data.clone(),
+        )?;
+        Ok(Some(Box::new(iter) as BoxedBatchIterator))
+    }
 }
 
 #[derive(Debug)]
 pub struct BulkPartMeta {
+    /// Total rows in part.
     pub num_rows: usize,
+    /// Max timestamp in part.
     pub max_timestamp: i64,
+    /// Min timestamp in part.
     pub min_timestamp: i64,
-}
+    /// Part file metadata.
-
+    pub parquet_metadata: Arc<ParquetMetaData>,
-impl Default for BulkPartMeta {
+    /// Part region schema.
-    fn default() -> Self {
+    pub region_metadata: RegionMetadataRef,
-        Self {
-            num_rows: 0,
-            max_timestamp: i64::MIN,
-            min_timestamp: i64::MAX,
-        }
-    }
 }
 
 pub struct BulkPartEncoder {
     metadata: RegionMetadataRef,
-    arrow_schema: SchemaRef,
     pk_encoder: McmpRowCodec,
+    row_group_size: usize,
     dedup: bool,
+    writer_props: Option<WriterProperties>,
+}
+
+impl BulkPartEncoder {
+    pub(crate) fn new(
+        metadata: RegionMetadataRef,
+        dedup: bool,
+        row_group_size: usize,
+    ) -> BulkPartEncoder {
+        let codec = McmpRowCodec::new_with_primary_keys(&metadata);
+        let writer_props = Some(
+            WriterProperties::builder()
+                .set_write_batch_size(row_group_size)
+                .set_max_row_group_size(row_group_size)
+                .build(),
+        );
+        Self {
+            metadata,
+            pk_encoder: codec,
+            row_group_size,
+            dedup,
+            writer_props,
+        }
+    }
 }
 
 impl BulkPartEncoder {
     /// Encodes mutations to a [BulkPart], returns true if encoded data has been written to `dest`.
-    fn encode_mutations(&self, mutations: &[Mutation], dest: &mut BulkPart) -> Result<bool> {
+    fn encode_mutations(&self, mutations: &[Mutation]) -> Result<Option<BulkPart>> {
         let Some((arrow_record_batch, min_ts, max_ts)) =
-            mutations_to_record_batch(mutations, &self.metadata, &self.pk_encoder, false)?
+            mutations_to_record_batch(mutations, &self.metadata, &self.pk_encoder, self.dedup)?
         else {
-            return Ok(false);
+            return Ok(None);
         };
 
+        let mut buf = Vec::with_capacity(4096);
         let arrow_schema = arrow_record_batch.schema();
-        {
+
-            let mut writer = ArrowWriter::try_new(&mut dest.data, arrow_schema, None)
+        let file_metadata = {
-                .context(EncodeMemtableSnafu)?;
+            let mut writer =
+                ArrowWriter::try_new(&mut buf, arrow_schema, self.writer_props.clone())
+                    .context(EncodeMemtableSnafu)?;
             writer
                 .write(&arrow_record_batch)
                 .context(EncodeMemtableSnafu)?;
-            let _metadata = writer.finish().context(EncodeMemtableSnafu)?;
+            writer.finish().context(EncodeMemtableSnafu)?
-        }
-
-        dest.metadata = BulkPartMeta {
-            num_rows: arrow_record_batch.num_rows(),
-            max_timestamp: max_ts,
-            min_timestamp: min_ts,
         };
-        Ok(true)
-    }
 
-    /// Decodes [BulkPart] to [Batch]es.
+        let buf = Bytes::from(buf);
-    fn decode_to_batches(&self, _part: &BulkPart, _dest: &mut VecDeque<Batch>) -> Result<()> {
+        let parquet_metadata = Arc::new(parse_parquet_metadata(file_metadata)?);
-        todo!()
+
+        Ok(Some(BulkPart {
+            data: buf,
+            metadata: BulkPartMeta {
+                num_rows: arrow_record_batch.num_rows(),
+                max_timestamp: max_ts,
+                min_timestamp: min_ts,
+                parquet_metadata,
+                region_metadata: self.metadata.clone(),
+            },
+        }))
     }
 }
 
@@ -379,10 +433,12 @@ fn binary_array_to_dictionary(input: &BinaryArray) -> Result<PrimaryKeyArray> {
 mod tests {
     use std::collections::VecDeque;
 
+    use datafusion_common::ScalarValue;
     use datatypes::prelude::{ScalarVector, Value};
     use datatypes::vectors::{Float64Vector, TimestampMillisecondVector};
 
     use super::*;
+    use crate::memtable::bulk::context::BulkIterContext;
     use crate::sst::parquet::format::ReadFormat;
     use crate::test_util::memtable_util::{build_key_values_with_ts_seq_values, metadata_for_test};
 
@@ -444,7 +500,7 @@ mod tests {
         k0: &'a str,
         k1: u32,
         timestamps: &'a [i64],
-        v0: &'a [Option<f64>],
+        v1: &'a [Option<f64>],
         sequence: u64,
     }
 
@@ -452,7 +508,7 @@ mod tests {
     struct BatchOutput<'a> {
         pk_values: &'a [Value],
         timestamps: &'a [i64],
-        v0: &'a [Option<f64>],
+        v1: &'a [Option<f64>],
     }
 
     fn check_mutations_to_record_batches(
@@ -470,7 +526,7 @@ mod tests {
                     m.k0.to_string(),
                     m.k1,
                     m.timestamps.iter().copied(),
-                    m.v0.iter().copied(),
+                    m.v1.iter().copied(),
                     m.sequence,
                 )
                 .mutation
@@ -526,7 +582,7 @@ mod tests {
         for idx in 0..expected.len() {
             assert_eq!(expected[idx].pk_values, &batch_values[idx].0);
             assert_eq!(expected[idx].timestamps, &batch_values[idx].1);
-            assert_eq!(expected[idx].v0, &batch_values[idx].2);
+            assert_eq!(expected[idx].v1, &batch_values[idx].2);
         }
     }
 
@@ -537,13 +593,13 @@ mod tests {
                 k0: "a",
                 k1: 0,
                 timestamps: &[0],
-                v0: &[Some(0.1)],
+                v1: &[Some(0.1)],
                 sequence: 0,
             }],
             &[BatchOutput {
                 pk_values: &[Value::String("a".into()), Value::UInt32(0)],
                 timestamps: &[0],
-                v0: &[Some(0.1)],
+                v1: &[Some(0.1)],
             }],
             (0, 0),
             true,
@@ -555,28 +611,28 @@ mod tests {
                     k0: "a",
                     k1: 0,
                     timestamps: &[0],
-                    v0: &[Some(0.1)],
+                    v1: &[Some(0.1)],
                     sequence: 0,
                 },
                 MutationInput {
                     k0: "b",
                     k1: 0,
                     timestamps: &[0],
-                    v0: &[Some(0.0)],
+                    v1: &[Some(0.0)],
                     sequence: 0,
                 },
                 MutationInput {
                     k0: "a",
                     k1: 0,
                     timestamps: &[1],
-                    v0: &[Some(0.2)],
+                    v1: &[Some(0.2)],
                     sequence: 1,
                 },
                 MutationInput {
                     k0: "a",
                     k1: 1,
                     timestamps: &[1],
-                    v0: &[Some(0.3)],
+                    v1: &[Some(0.3)],
                     sequence: 2,
                 },
             ],
@@ -584,17 +640,17 @@ mod tests {
                 BatchOutput {
                     pk_values: &[Value::String("a".into()), Value::UInt32(0)],
                     timestamps: &[0, 1],
-                    v0: &[Some(0.1), Some(0.2)],
+                    v1: &[Some(0.1), Some(0.2)],
                 },
                 BatchOutput {
                     pk_values: &[Value::String("a".into()), Value::UInt32(1)],
                     timestamps: &[1],
-                    v0: &[Some(0.3)],
+                    v1: &[Some(0.3)],
                 },
                 BatchOutput {
                     pk_values: &[Value::String("b".into()), Value::UInt32(0)],
                     timestamps: &[0],
-                    v0: &[Some(0.0)],
+                    v1: &[Some(0.0)],
                 },
             ],
             (0, 1),
@@ -607,21 +663,21 @@ mod tests {
                     k0: "a",
                     k1: 0,
                     timestamps: &[0],
-                    v0: &[Some(0.1)],
+                    v1: &[Some(0.1)],
                     sequence: 0,
                 },
                 MutationInput {
                     k0: "b",
                     k1: 0,
                     timestamps: &[0],
-                    v0: &[Some(0.0)],
+                    v1: &[Some(0.0)],
                     sequence: 0,
                 },
                 MutationInput {
                     k0: "a",
                     k1: 0,
                     timestamps: &[0],
-                    v0: &[Some(0.2)],
+                    v1: &[Some(0.2)],
                     sequence: 1,
                 },
             ],
@@ -629,12 +685,12 @@ mod tests {
                 BatchOutput {
                     pk_values: &[Value::String("a".into()), Value::UInt32(0)],
                     timestamps: &[0],
-                    v0: &[Some(0.2)],
+                    v1: &[Some(0.2)],
                 },
                 BatchOutput {
                     pk_values: &[Value::String("b".into()), Value::UInt32(0)],
                     timestamps: &[0],
-                    v0: &[Some(0.0)],
+                    v1: &[Some(0.0)],
                 },
             ],
             (0, 0),
@@ -646,21 +702,21 @@ mod tests {
                     k0: "a",
                     k1: 0,
                     timestamps: &[0],
-                    v0: &[Some(0.1)],
+                    v1: &[Some(0.1)],
                     sequence: 0,
                 },
                 MutationInput {
                     k0: "b",
                     k1: 0,
                     timestamps: &[0],
-                    v0: &[Some(0.0)],
+                    v1: &[Some(0.0)],
                     sequence: 0,
                 },
                 MutationInput {
                     k0: "a",
                     k1: 0,
                     timestamps: &[0],
-                    v0: &[Some(0.2)],
+                    v1: &[Some(0.2)],
                     sequence: 1,
                 },
             ],
@@ -668,16 +724,194 @@ mod tests {
                 BatchOutput {
                     pk_values: &[Value::String("a".into()), Value::UInt32(0)],
                     timestamps: &[0, 0],
-                    v0: &[Some(0.2), Some(0.1)],
+                    v1: &[Some(0.2), Some(0.1)],
                 },
                 BatchOutput {
                     pk_values: &[Value::String("b".into()), Value::UInt32(0)],
                     timestamps: &[0],
-                    v0: &[Some(0.0)],
+                    v1: &[Some(0.0)],
                 },
             ],
             (0, 0),
             false,
         );
     }
+
+    fn encode(input: &[MutationInput]) -> BulkPart {
+        let metadata = metadata_for_test();
+        let mutations = input
+            .iter()
+            .map(|m| {
+                build_key_values_with_ts_seq_values(
+                    &metadata,
+                    m.k0.to_string(),
+                    m.k1,
+                    m.timestamps.iter().copied(),
+                    m.v1.iter().copied(),
+                    m.sequence,
+                )
+                .mutation
+            })
+            .collect::<Vec<_>>();
+        let encoder = BulkPartEncoder::new(metadata, true, 1024);
+        encoder.encode_mutations(&mutations).unwrap().unwrap()
+    }
+
+    #[test]
+    fn test_write_and_read_part_projection() {
+        let part = encode(&[
+            MutationInput {
+                k0: "a",
+                k1: 0,
+                timestamps: &[1],
+                v1: &[Some(0.1)],
+                sequence: 0,
+            },
+            MutationInput {
+                k0: "b",
+                k1: 0,
+                timestamps: &[1],
+                v1: &[Some(0.0)],
+                sequence: 0,
+            },
+            MutationInput {
+                k0: "a",
+                k1: 0,
+                timestamps: &[2],
+                v1: &[Some(0.2)],
+                sequence: 1,
+            },
+        ]);
+
+        let projection = &[4u32];
+
+        let mut reader = part
+            .read(Arc::new(BulkIterContext::new(
+                part.metadata.region_metadata.clone(),
+                &Some(projection.as_slice()),
+                None,
+            )))
+            .unwrap()
+            .expect("expect at least one row group");
+
+        let mut total_rows_read = 0;
+        let mut field = vec![];
+        for res in reader {
+            let batch = res.unwrap();
+            assert_eq!(1, batch.fields().len());
+            assert_eq!(4, batch.fields()[0].column_id);
+            field.extend(
+                batch.fields()[0]
+                    .data
+                    .as_any()
+                    .downcast_ref::<Float64Vector>()
+                    .unwrap()
+                    .iter_data()
+                    .map(|v| v.unwrap()),
+            );
+            total_rows_read += batch.num_rows();
+        }
+        assert_eq!(3, total_rows_read);
+        assert_eq!(vec![0.1, 0.2, 0.0], field);
+    }
+
+    fn prepare(key_values: Vec<(&str, u32, (i64, i64), u64)>) -> BulkPart {
+        let metadata = metadata_for_test();
+        let mutations = key_values
+            .into_iter()
+            .map(|(k0, k1, (start, end), sequence)| {
+                let ts = (start..end);
+                let v1 = (start..end).map(|_| None);
+                build_key_values_with_ts_seq_values(&metadata, k0.to_string(), k1, ts, v1, sequence)
+                    .mutation
+            })
+            .collect::<Vec<_>>();
+        let encoder = BulkPartEncoder::new(metadata, true, 100);
+        encoder.encode_mutations(&mutations).unwrap().unwrap()
+    }
+
+    fn check_prune_row_group(part: &BulkPart, predicate: Option<Predicate>, expected_rows: usize) {
+        let context = Arc::new(BulkIterContext::new(
+            part.metadata.region_metadata.clone(),
+            &None,
+            predicate,
+        ));
+        let mut reader = part
+            .read(context)
+            .unwrap()
+            .expect("expect at least one row group");
+        let mut total_rows_read = 0;
+        for res in reader {
+            let batch = res.unwrap();
+            total_rows_read += batch.num_rows();
+        }
+        // Should only read row group 1.
+        assert_eq!(expected_rows, total_rows_read);
+    }
+
+    #[test]
+    fn test_prune_row_groups() {
+        let part = prepare(vec![
+            ("a", 0, (0, 40), 1),
+            ("a", 1, (0, 60), 1),
+            ("b", 0, (0, 100), 2),
+            ("b", 1, (100, 180), 3),
+            ("b", 1, (180, 210), 4),
+        ]);
+
+        let context = Arc::new(BulkIterContext::new(
+            part.metadata.region_metadata.clone(),
+            &None,
+            Some(Predicate::new(vec![datafusion_expr::col("ts").eq(
+                datafusion_expr::lit(ScalarValue::TimestampMillisecond(Some(300), None)),
+            )])),
+        ));
+        assert!(part.read(context).unwrap().is_none());
+
+        check_prune_row_group(&part, None, 310);
+
+        check_prune_row_group(
+            &part,
+            Some(Predicate::new(vec![
+                datafusion_expr::col("k0").eq(datafusion_expr::lit("a")),
+                datafusion_expr::col("k1").eq(datafusion_expr::lit(0u32)),
+            ])),
+            40,
+        );
+
+        check_prune_row_group(
+            &part,
+            Some(Predicate::new(vec![
+                datafusion_expr::col("k0").eq(datafusion_expr::lit("a")),
+                datafusion_expr::col("k1").eq(datafusion_expr::lit(1u32)),
+            ])),
+            60,
+        );
+
+        check_prune_row_group(
+            &part,
+            Some(Predicate::new(vec![
+                datafusion_expr::col("k0").eq(datafusion_expr::lit("a"))
+            ])),
+            100,
+        );
+
+        check_prune_row_group(
+            &part,
+            Some(Predicate::new(vec![
+                datafusion_expr::col("k0").eq(datafusion_expr::lit("b")),
+                datafusion_expr::col("k1").eq(datafusion_expr::lit(0u32)),
+            ])),
+            100,
+        );
+
+        /// Predicates over field column can do precise filtering.
+        check_prune_row_group(
+            &part,
+            Some(Predicate::new(vec![
+                datafusion_expr::col("v0").eq(datafusion_expr::lit(150i64))
+            ])),
+            1,
+        );
+    }
 }

src/mito2/src/memtable/bulk/part_reader.rs

@@ -0,0 +1,149 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::VecDeque;
+use std::sync::Arc;
+
+use bytes::Bytes;
+use parquet::arrow::ProjectionMask;
+use parquet::file::metadata::ParquetMetaData;
+
+use crate::error;
+use crate::memtable::bulk::context::BulkIterContextRef;
+use crate::memtable::bulk::row_group_reader::{
+    MemtableRowGroupReader, MemtableRowGroupReaderBuilder,
+};
+use crate::read::Batch;
+
+/// Iterator for reading data inside a bulk part.
+pub struct BulkPartIter {
+    row_groups_to_read: VecDeque<usize>,
+    current_reader: Option<PruneReader>,
+    builder: MemtableRowGroupReaderBuilder,
+}
+
+impl BulkPartIter {
+    /// Creates a new [BulkPartIter].
+    pub(crate) fn try_new(
+        context: BulkIterContextRef,
+        mut row_groups_to_read: VecDeque<usize>,
+        parquet_meta: Arc<ParquetMetaData>,
+        data: Bytes,
+    ) -> error::Result<Self> {
+        let projection_mask = ProjectionMask::roots(
+            parquet_meta.file_metadata().schema_descr(),
+            context.read_format().projection_indices().iter().copied(),
+        );
+
+        let builder = MemtableRowGroupReaderBuilder::try_new(
+            context.clone(),
+            projection_mask,
+            parquet_meta,
+            data,
+        )?;
+
+        let init_reader = row_groups_to_read
+            .pop_front()
+            .map(|first_row_group| builder.build_row_group_reader(first_row_group, None))
+            .transpose()?
+            .map(|r| PruneReader::new(context, r));
+        Ok(Self {
+            row_groups_to_read,
+            current_reader: init_reader,
+            builder,
+        })
+    }
+
+    pub(crate) fn next_batch(&mut self) -> error::Result<Option<Batch>> {
+        let Some(current) = &mut self.current_reader else {
+            // All row group exhausted.
+            return Ok(None);
+        };
+
+        if let Some(batch) = current.next_batch()? {
+            return Ok(Some(batch));
+        }
+
+        // Previous row group exhausted, read next row group
+        while let Some(next_row_group) = self.row_groups_to_read.pop_front() {
+            current.reset(self.builder.build_row_group_reader(next_row_group, None)?);
+            if let Some(next_batch) = current.next_batch()? {
+                return Ok(Some(next_batch));
+            }
+        }
+        Ok(None)
+    }
+}
+
+impl Iterator for BulkPartIter {
+    type Item = error::Result<Batch>;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        self.next_batch().transpose()
+    }
+}
+
+struct PruneReader {
+    context: BulkIterContextRef,
+    row_group_reader: MemtableRowGroupReader,
+}
+
+//todo(hl): maybe we also need to support lastrow mode here.
+impl PruneReader {
+    fn new(context: BulkIterContextRef, reader: MemtableRowGroupReader) -> Self {
+        Self {
+            context,
+            row_group_reader: reader,
+        }
+    }
+
+    /// Iterates current inner reader until exhausted.
+    fn next_batch(&mut self) -> error::Result<Option<Batch>> {
+        while let Some(b) = self.row_group_reader.next_inner()? {
+            match self.prune(b)? {
+                Some(b) => {
+                    return Ok(Some(b));
+                }
+                None => {
+                    continue;
+                }
+            }
+        }
+        Ok(None)
+    }
+
+    /// Prunes batch according to filters.
+    fn prune(&mut self, batch: Batch) -> error::Result<Option<Batch>> {
+        //todo(hl): add metrics.
+
+        // fast path
+        if self.context.base.filters.is_empty() {
+            return Ok(Some(batch));
+        }
+
+        let Some(batch_filtered) = self.context.base.precise_filter(batch)? else {
+            // the entire batch is filtered out
+            return Ok(None);
+        };
+        if !batch_filtered.is_empty() {
+            Ok(Some(batch_filtered))
+        } else {
+            Ok(None)
+        }
+    }
+
+    fn reset(&mut self, reader: MemtableRowGroupReader) {
+        self.row_group_reader = reader;
+    }
+}

src/mito2/src/memtable/bulk/row_group_reader.rs

@@ -0,0 +1,189 @@
+// 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::ops::Range;
+use std::sync::Arc;
+
+use bytes::Bytes;
+use datatypes::arrow::array::RecordBatch;
+use datatypes::arrow::error::ArrowError;
+use parquet::arrow::arrow_reader::{ParquetRecordBatchReader, RowGroups, RowSelection};
+use parquet::arrow::{parquet_to_arrow_field_levels, FieldLevels, ProjectionMask};
+use parquet::column::page::{PageIterator, PageReader};
+use parquet::file::metadata::ParquetMetaData;
+use snafu::ResultExt;
+
+use crate::error;
+use crate::error::ReadDataPartSnafu;
+use crate::memtable::bulk::context::BulkIterContextRef;
+use crate::sst::parquet::format::ReadFormat;
+use crate::sst::parquet::reader::{RowGroupReaderBase, RowGroupReaderContext};
+use crate::sst::parquet::row_group::{ColumnChunkIterator, RowGroupBase};
+use crate::sst::parquet::DEFAULT_READ_BATCH_SIZE;
+
+/// Helper for reading specific row group inside Memtable Parquet parts.
+// This is similar to [mito2::sst::parquet::row_group::InMemoryRowGroup] since
+// it's a workaround for lacking of keyword generics.
+pub struct MemtableRowGroupPageFetcher<'a> {
+    /// Shared structs for reading row group.
+    base: RowGroupBase<'a>,
+    bytes: Bytes,
+}
+
+impl<'a> MemtableRowGroupPageFetcher<'a> {
+    pub(crate) fn create(
+        row_group_idx: usize,
+        parquet_meta: &'a ParquetMetaData,
+        bytes: Bytes,
+    ) -> Self {
+        let metadata = parquet_meta.row_group(row_group_idx);
+        let row_count = metadata.num_rows() as usize;
+        let page_locations = parquet_meta
+            .offset_index()
+            .map(|x| x[row_group_idx].as_slice());
+
+        Self {
+            base: RowGroupBase {
+                metadata,
+                page_locations,
+                row_count,
+                column_chunks: vec![None; metadata.columns().len()],
+                // the cached `column_uncompressed_pages` would never be used in Memtable readers.
+                column_uncompressed_pages: vec![None; metadata.columns().len()],
+            },
+            bytes,
+        }
+    }
+
+    /// Fetches column pages from memory file.
+    pub(crate) fn fetch(&mut self, projection: &ProjectionMask, selection: Option<&RowSelection>) {
+        if let Some((selection, page_locations)) = selection.zip(self.base.page_locations) {
+            // Selection provided.
+            let (fetch_ranges, page_start_offsets) =
+                self.base
+                    .calc_sparse_read_ranges(projection, page_locations, selection);
+            if fetch_ranges.is_empty() {
+                return;
+            }
+            let chunk_data = self.fetch_bytes(&fetch_ranges);
+
+            self.base
+                .assign_sparse_chunk(projection, chunk_data, page_start_offsets);
+        } else {
+            let fetch_ranges = self.base.calc_dense_read_ranges(projection);
+            if fetch_ranges.is_empty() {
+                // Nothing to fetch.
+                return;
+            }
+            let chunk_data = self.fetch_bytes(&fetch_ranges);
+            self.base.assign_dense_chunk(projection, chunk_data);
+        }
+    }
+
+    fn fetch_bytes(&self, ranges: &[Range<u64>]) -> Vec<Bytes> {
+        ranges
+            .iter()
+            .map(|range| self.bytes.slice(range.start as usize..range.end as usize))
+            .collect()
+    }
+
+    /// Creates a page reader to read column at `i`.
+    fn column_page_reader(&self, i: usize) -> parquet::errors::Result<Box<dyn PageReader>> {
+        let reader = self.base.column_reader(i)?;
+        Ok(Box::new(reader))
+    }
+}
+
+impl RowGroups for MemtableRowGroupPageFetcher<'_> {
+    fn num_rows(&self) -> usize {
+        self.base.row_count
+    }
+
+    fn column_chunks(&self, i: usize) -> parquet::errors::Result<Box<dyn PageIterator>> {
+        Ok(Box::new(ColumnChunkIterator {
+            reader: Some(self.column_page_reader(i)),
+        }))
+    }
+}
+
+impl RowGroupReaderContext for BulkIterContextRef {
+    fn map_result(
+        &self,
+        result: Result<Option<RecordBatch>, ArrowError>,
+    ) -> error::Result<Option<RecordBatch>> {
+        result.context(error::DecodeArrowRowGroupSnafu)
+    }
+
+    fn read_format(&self) -> &ReadFormat {
+        self.as_ref().read_format()
+    }
+}
+
+pub(crate) type MemtableRowGroupReader = RowGroupReaderBase<BulkIterContextRef>;
+
+pub(crate) struct MemtableRowGroupReaderBuilder {
+    context: BulkIterContextRef,
+    projection: ProjectionMask,
+    parquet_metadata: Arc<ParquetMetaData>,
+    field_levels: FieldLevels,
+    data: Bytes,
+}
+
+impl MemtableRowGroupReaderBuilder {
+    pub(crate) fn try_new(
+        context: BulkIterContextRef,
+        projection: ProjectionMask,
+        parquet_metadata: Arc<ParquetMetaData>,
+        data: Bytes,
+    ) -> error::Result<Self> {
+        let parquet_schema_desc = parquet_metadata.file_metadata().schema_descr();
+        let hint = Some(context.read_format().arrow_schema().fields());
+        let field_levels =
+            parquet_to_arrow_field_levels(parquet_schema_desc, projection.clone(), hint)
+                .context(ReadDataPartSnafu)?;
+        Ok(Self {
+            context,
+            projection,
+            parquet_metadata,
+            field_levels,
+            data,
+        })
+    }
+
+    /// Builds a reader to read the row group at `row_group_idx` from memory.
+    pub(crate) fn build_row_group_reader(
+        &self,
+        row_group_idx: usize,
+        row_selection: Option<RowSelection>,
+    ) -> error::Result<MemtableRowGroupReader> {
+        let mut row_group = MemtableRowGroupPageFetcher::create(
+            row_group_idx,
+            &self.parquet_metadata,
+            self.data.clone(),
+        );
+        // Fetches data from memory part. Currently, row selection is not supported.
+        row_group.fetch(&self.projection, row_selection.as_ref());
+
+        // Builds the parquet reader.
+        // Now the row selection is None.
+        let reader = ParquetRecordBatchReader::try_new_with_row_groups(
+            &self.field_levels,
+            &row_group,
+            DEFAULT_READ_BATCH_SIZE,
+            row_selection,
+        )
+        .context(ReadDataPartSnafu)?;
+        Ok(MemtableRowGroupReader::create(self.context.clone(), reader))
+    }
+}

src/mito2/src/read/last_row.rs

@@ -99,11 +99,8 @@ impl RowGroupLastRowCachedReader {
             return Self::new_miss(key, row_group_reader, None);
         };
         if let Some(value) = cache_manager.get_selector_result(&key) {
-            let schema_matches = value.projection
+            let schema_matches =
-                == row_group_reader
+                value.projection == row_group_reader.read_format().projection_indices();
-                    .context()
-                    .read_format()
-                    .projection_indices();
             if schema_matches {
                 // Schema matches, use cache batches.
                 Self::new_hit(value)
@@ -218,29 +215,23 @@ impl RowGroupLastRowReader {
         };
 
         // All last rows in row group are yielded, update cache.
-        self.update_cache();
+        self.maybe_update_cache();
         Ok(last_batch)
     }
 
     /// Updates row group's last row cache if cache manager is present.
-    fn update_cache(&mut self) {
+    fn maybe_update_cache(&mut self) {
-        if self.yielded_batches.is_empty() {
+        if let Some(cache) = &self.cache_manager {
-            // we always expect that row groups yields batches.
+            if self.yielded_batches.is_empty() {
-            return;
+                // we always expect that row groups yields batches.
+                return;
+            }
+            let value = Arc::new(SelectorResultValue {
+                result: std::mem::take(&mut self.yielded_batches),
+                projection: self.reader.read_format().projection_indices().to_vec(),
+            });
+            cache.put_selector_result(self.key, value)
         }
-        let Some(cache) = &self.cache_manager else {
-            return;
-        };
-        let value = Arc::new(SelectorResultValue {
-            result: std::mem::take(&mut self.yielded_batches),
-            projection: self
-                .reader
-                .context()
-                .read_format()
-                .projection_indices()
-                .to_vec(),
-        });
-        cache.put_selector_result(self.key, value);
     }
 
     fn metrics(&self) -> &ReaderMetrics {

src/mito2/src/read/scan_region.rs

@@ -399,7 +399,7 @@ impl ScanRegion {
         });
     }
 
-    /// Use the latest schema to build the inveretd index applier.
+    /// Use the latest schema to build the inverted index applier.
     fn build_invereted_index_applier(&self) -> Option<InvertedIndexApplierRef> {
         if self.ignore_inverted_index {
             return None;

src/mito2/src/region.rs

@@ -447,7 +447,26 @@ impl ManifestContext {
     pub(crate) fn set_role(&self, next_role: RegionRole, region_id: RegionId) {
         match next_role {
             RegionRole::Follower => {
-                self.state.store(RegionRoleState::Follower);
+                match self.state.fetch_update(|state| {
+                    if !matches!(state, RegionRoleState::Follower) {
+                        Some(RegionRoleState::Follower)
+                    } else {
+                        None
+                    }
+                }) {
+                    Ok(state) => info!(
+                        "Convert region {} to follower, previous role state: {:?}",
+                        region_id, state
+                    ),
+                    Err(state) => {
+                        if state != RegionRoleState::Follower {
+                            warn!(
+                                "Failed to convert region {} to follower, current role state: {:?}",
+                                region_id, state
+                            )
+                        }
+                    }
+                }
             }
             RegionRole::Leader => {
                 match self.state.fetch_update(|state| {

src/mito2/src/sst/file_purger.rs

@@ -185,7 +185,7 @@ mod tests {
 
         scheduler.stop(true).await.unwrap();
 
-        assert!(!object_store.is_exist(&path).await.unwrap());
+        assert!(!object_store.exists(&path).await.unwrap());
     }
 
     #[tokio::test]
@@ -247,7 +247,7 @@ mod tests {
 
         scheduler.stop(true).await.unwrap();
 
-        assert!(!object_store.is_exist(&path).await.unwrap());
+        assert!(!object_store.exists(&path).await.unwrap());
-        assert!(!object_store.is_exist(&index_path).await.unwrap());
+        assert!(!object_store.exists(&index_path).await.unwrap());
     }
 }

src/mito2/src/sst/index/intermediate.rs

@@ -104,16 +104,28 @@ impl IntermediateLocation {
         &self.files_dir
     }
 
-    /// Returns the path of the directory for intermediate files associated with a column:
+    /// Returns the path of the directory for intermediate files associated with the `file_group`:
-    /// `__intm/{region_id}/{sst_file_id}/{uuid}/{column_id}/`
+    /// `__intm/{region_id}/{sst_file_id}/{uuid}/{file_group}/`
-    pub fn column_path(&self, column_id: &str) -> String {
+    pub fn file_group_path(&self, file_group: &str) -> String {
-        util::join_path(&self.files_dir, &format!("{column_id}/"))
+        util::join_path(&self.files_dir, &format!("{file_group}/"))
     }
 
-    /// Returns the path of the intermediate file with the given id for a column:
+    /// Returns the path of the intermediate file with the given `file_group` and `im_file_id`:
-    /// `__intm/{region_id}/{sst_file_id}/{uuid}/{column_id}/{im_file_id}.im`
+    /// `__intm/{region_id}/{sst_file_id}/{uuid}/{file_group}/{im_file_id}.im`
-    pub fn file_path(&self, column_id: &str, im_file_id: &str) -> String {
+    pub fn file_path(&self, file_group: &str, im_file_id: &str) -> String {
-        util::join_path(&self.column_path(column_id), &format!("{im_file_id}.im"))
+        util::join_path(
+            &self.file_group_path(file_group),
+            &format!("{im_file_id}.im"),
+        )
+    }
+
+    /// Returns the intermediate file id from the path.
+    pub fn im_file_id_from_path(&self, path: &str) -> String {
+        path.rsplit('/')
+            .next()
+            .and_then(|s| s.strip_suffix(".im"))
+            .unwrap_or_default()
+            .to_string()
     }
 }
 
@@ -161,17 +173,20 @@ mod tests {
 
         let uuid = location.files_dir.split('/').nth(3).unwrap();
 
-        let column_id = "1";
+        let file_group = "1";
         assert_eq!(
-            location.column_path(column_id),
+            location.file_group_path(file_group),
-            format!("{INTERMEDIATE_DIR}/0/{sst_file_id}/{uuid}/{column_id}/")
+            format!("{INTERMEDIATE_DIR}/0/{sst_file_id}/{uuid}/{file_group}/")
         );
 
         let im_file_id = "000000000010";
+        let file_path = location.file_path(file_group, im_file_id);
         assert_eq!(
-            location.file_path(column_id, im_file_id),
+            file_path,
-            format!("{INTERMEDIATE_DIR}/0/{sst_file_id}/{uuid}/{column_id}/{im_file_id}.im")
+            format!("{INTERMEDIATE_DIR}/0/{sst_file_id}/{uuid}/{file_group}/{im_file_id}.im")
         );
+
+        assert_eq!(location.im_file_id_from_path(&file_path), im_file_id);
     }
 
     #[tokio::test]

src/mito2/src/sst/index/inverted_index/applier.rs

@@ -29,7 +29,7 @@ use snafu::ResultExt;
 use store_api::storage::RegionId;
 
 use crate::cache::file_cache::{FileCacheRef, FileType, IndexKey};
-use crate::cache::index::{CachedInvertedIndexBlobReader, InvertedIndexCacheRef};
+use crate::cache::index::inverted_index::{CachedInvertedIndexBlobReader, InvertedIndexCacheRef};
 use crate::error::{
     ApplyInvertedIndexSnafu, MetadataSnafu, PuffinBuildReaderSnafu, PuffinReadBlobSnafu, Result,
 };

src/mito2/src/sst/index/inverted_index/applier/builder.rs

@@ -34,7 +34,7 @@ use store_api::metadata::RegionMetadata;
 use store_api::storage::ColumnId;
 
 use crate::cache::file_cache::FileCacheRef;
-use crate::cache::index::InvertedIndexCacheRef;
+use crate::cache::index::inverted_index::InvertedIndexCacheRef;
 use crate::error::{BuildIndexApplierSnafu, ColumnNotFoundSnafu, ConvertValueSnafu, Result};
 use crate::row_converter::SortField;
 use crate::sst::index::inverted_index::applier::InvertedIndexApplier;

src/mito2/src/sst/index/inverted_index/creator.rs

@@ -316,7 +316,7 @@ mod tests {
     use store_api::storage::RegionId;
 
     use super::*;
-    use crate::cache::index::InvertedIndexCache;
+    use crate::cache::index::inverted_index::InvertedIndexCache;
     use crate::metrics::CACHE_BYTES;
     use crate::read::BatchColumn;
     use crate::row_converter::{McmpRowCodec, RowCodec, SortField};

src/mito2/src/sst/index/inverted_index/creator/temp_provider.rs

@@ -16,9 +16,9 @@ use async_trait::async_trait;
 use common_error::ext::BoxedError;
 use common_telemetry::warn;
 use futures::{AsyncRead, AsyncWrite};
-use index::inverted_index::create::sort::external_provider::ExternalTempFileProvider;
+use index::error as index_error;
-use index::inverted_index::error as index_error;
+use index::error::Result as IndexResult;
-use index::inverted_index::error::Result as IndexResult;
+use index::external_provider::ExternalTempFileProvider;
 use snafu::ResultExt;
 
 use crate::error::Result;
@@ -42,10 +42,10 @@ pub(crate) struct TempFileProvider {
 impl ExternalTempFileProvider for TempFileProvider {
     async fn create(
         &self,
-        column_id: &str,
+        file_group: &str,
         file_id: &str,
     ) -> IndexResult<Box<dyn AsyncWrite + Unpin + Send>> {
-        let path = self.location.file_path(column_id, file_id);
+        let path = self.location.file_path(file_group, file_id);
         let writer = self
             .manager
             .store()
@@ -63,13 +63,13 @@ impl ExternalTempFileProvider for TempFileProvider {
 
     async fn read_all(
         &self,
-        column_id: &str,
+        file_group: &str,
-    ) -> IndexResult<Vec<Box<dyn AsyncRead + Unpin + Send>>> {
+    ) -> IndexResult<Vec<(String, Box<dyn AsyncRead + Unpin + Send>)>> {
-        let column_path = self.location.column_path(column_id);
+        let file_group_path = self.location.file_group_path(file_group);
         let entries = self
             .manager
             .store()
-            .list(&column_path)
+            .list(&file_group_path)
             .await
             .map_err(BoxedError::new)
             .context(index_error::ExternalSnafu)?;
@@ -81,6 +81,8 @@ impl ExternalTempFileProvider for TempFileProvider {
                 continue;
             }
 
+            let im_file_id = self.location.im_file_id_from_path(entry.path());
+
             let reader = self
                 .manager
                 .store()
@@ -93,7 +95,7 @@ impl ExternalTempFileProvider for TempFileProvider {
                 .await
                 .map_err(BoxedError::new)
                 .context(index_error::ExternalSnafu)?;
-            readers.push(Box::new(reader) as _);
+            readers.push((im_file_id, Box::new(reader) as _));
         }
 
         Ok(readers)
@@ -133,36 +135,36 @@ mod tests {
         let store = IntermediateManager::init_fs(path).await.unwrap();
         let provider = TempFileProvider::new(location.clone(), store);
 
-        let column_name = "tag0";
+        let file_group = "tag0";
         let file_id = "0000000010";
-        let mut writer = provider.create(column_name, file_id).await.unwrap();
+        let mut writer = provider.create(file_group, file_id).await.unwrap();
         writer.write_all(b"hello").await.unwrap();
         writer.flush().await.unwrap();
         writer.close().await.unwrap();
 
         let file_id = "0000000100";
-        let mut writer = provider.create(column_name, file_id).await.unwrap();
+        let mut writer = provider.create(file_group, file_id).await.unwrap();
         writer.write_all(b"world").await.unwrap();
         writer.flush().await.unwrap();
         writer.close().await.unwrap();
 
-        let column_name = "tag1";
+        let file_group = "tag1";
         let file_id = "0000000010";
-        let mut writer = provider.create(column_name, file_id).await.unwrap();
+        let mut writer = provider.create(file_group, file_id).await.unwrap();
         writer.write_all(b"foo").await.unwrap();
         writer.flush().await.unwrap();
         writer.close().await.unwrap();
 
         let readers = provider.read_all("tag0").await.unwrap();
         assert_eq!(readers.len(), 2);
-        for mut reader in readers {
+        for (_, mut reader) in readers {
             let mut buf = Vec::new();
             reader.read_to_end(&mut buf).await.unwrap();
             assert!(matches!(buf.as_slice(), b"hello" | b"world"));
         }
         let readers = provider.read_all("tag1").await.unwrap();
         assert_eq!(readers.len(), 1);
-        let mut reader = readers.into_iter().next().unwrap();
+        let mut reader = readers.into_iter().map(|x| x.1).next().unwrap();
         let mut buf = Vec::new();
         reader.read_to_end(&mut buf).await.unwrap();
         assert_eq!(buf, b"foo");

src/mito2/src/sst/parquet.rs

@@ -27,11 +27,11 @@ pub(crate) mod file_range;
 pub mod format;
 pub(crate) mod helper;
 pub(crate) mod metadata;
-mod page_reader;
+pub(crate) mod page_reader;
 pub mod reader;
 pub mod row_group;
 mod row_selection;
-mod stats;
+pub(crate) mod stats;
 pub mod writer;
 
 /// Key of metadata in parquet SST.

src/mito2/src/sst/parquet/reader.rs

@@ -24,6 +24,7 @@ use async_trait::async_trait;
 use common_recordbatch::filter::SimpleFilterEvaluator;
 use common_telemetry::{debug, warn};
 use datafusion_expr::Expr;
+use datatypes::arrow::error::ArrowError;
 use datatypes::arrow::record_batch::RecordBatch;
 use datatypes::data_type::ConcreteDataType;
 use itertools::Itertools;
@@ -39,7 +40,8 @@ use table::predicate::Predicate;
 
 use crate::cache::CacheManagerRef;
 use crate::error::{
-    ArrowReaderSnafu, InvalidMetadataSnafu, InvalidParquetSnafu, ReadParquetSnafu, Result,
+    ArrowReaderSnafu, InvalidMetadataSnafu, InvalidParquetSnafu, ReadDataPartSnafu,
+    ReadParquetSnafu, Result,
 };
 use crate::metrics::{
     PRECISE_FILTER_ROWS_TOTAL, READ_ROWS_IN_ROW_GROUP_TOTAL, READ_ROWS_TOTAL,
@@ -207,8 +209,7 @@ impl ParquetReaderBuilder {
         let hint = Some(read_format.arrow_schema().fields());
         let field_levels =
             parquet_to_arrow_field_levels(parquet_schema_desc, projection_mask.clone(), hint)
-                .context(ReadParquetSnafu { path: &file_path })?;
+                .context(ReadDataPartSnafu)?;
-
         let row_groups = self
             .row_groups_to_read(&read_format, &parquet_meta, &mut metrics.filter_metrics)
             .await;
@@ -871,7 +872,7 @@ impl SimpleFilterContext {
     ///
     /// Returns None if the column to filter doesn't exist in the SST metadata or the
     /// expected metadata.
-    fn new_opt(
+    pub(crate) fn new_opt(
         sst_meta: &RegionMetadataRef,
         expected_meta: Option<&RegionMetadata>,
         expr: &Expr,
@@ -1035,10 +1036,51 @@ impl ParquetReader {
     }
 }
 
+/// RowGroupReaderContext represents the fields that cannot be shared
+/// between different `RowGroupReader`s.
+pub(crate) trait RowGroupReaderContext: Send {
+    fn map_result(
+        &self,
+        result: std::result::Result<Option<RecordBatch>, ArrowError>,
+    ) -> Result<Option<RecordBatch>>;
+
+    fn read_format(&self) -> &ReadFormat;
+}
+
+impl RowGroupReaderContext for FileRangeContextRef {
+    fn map_result(
+        &self,
+        result: std::result::Result<Option<RecordBatch>, ArrowError>,
+    ) -> Result<Option<RecordBatch>> {
+        result.context(ArrowReaderSnafu {
+            path: self.file_path(),
+        })
+    }
+
+    fn read_format(&self) -> &ReadFormat {
+        self.as_ref().read_format()
+    }
+}
+
+/// [RowGroupReader] that reads from [FileRange].
+pub(crate) type RowGroupReader = RowGroupReaderBase<FileRangeContextRef>;
+
+impl RowGroupReader {
+    /// Creates a new reader from file range.
+    pub(crate) fn new(context: FileRangeContextRef, reader: ParquetRecordBatchReader) -> Self {
+        Self {
+            context,
+            reader,
+            batches: VecDeque::new(),
+            metrics: ReaderMetrics::default(),
+        }
+    }
+}
+
 /// Reader to read a row group of a parquet file.
-pub struct RowGroupReader {
+pub(crate) struct RowGroupReaderBase<T> {
-    /// Context for file ranges.
+    /// Context of [RowGroupReader] so adapts to different underlying implementation.
-    context: FileRangeContextRef,
+    context: T,
     /// Inner parquet reader.
     reader: ParquetRecordBatchReader,
     /// Buffered batches to return.
@@ -1047,9 +1089,12 @@ pub struct RowGroupReader {
     metrics: ReaderMetrics,
 }
 
-impl RowGroupReader {
+impl<T> RowGroupReaderBase<T>
+where
+    T: RowGroupReaderContext,
+{
     /// Creates a new reader.
-    pub(crate) fn new(context: FileRangeContextRef, reader: ParquetRecordBatchReader) -> Self {
+    pub(crate) fn create(context: T, reader: ParquetRecordBatchReader) -> Self {
         Self {
             context,
             reader,
@@ -1062,21 +1107,19 @@ impl RowGroupReader {
     pub(crate) fn metrics(&self) -> &ReaderMetrics {
         &self.metrics
     }
-    pub(crate) fn context(&self) -> &FileRangeContextRef {
+
-        &self.context
+    /// Gets [ReadFormat] of underlying reader.
+    pub(crate) fn read_format(&self) -> &ReadFormat {
+        self.context.read_format()
     }
 
     /// Tries to fetch next [RecordBatch] from the reader.
     fn fetch_next_record_batch(&mut self) -> Result<Option<RecordBatch>> {
-        self.reader.next().transpose().context(ArrowReaderSnafu {
+        self.context.map_result(self.reader.next().transpose())
-            path: self.context.file_path(),
-        })
     }
-}
 
-#[async_trait::async_trait]
+    /// Returns the next [Batch].
-impl BatchReader for RowGroupReader {
+    pub(crate) fn next_inner(&mut self) -> Result<Option<Batch>> {
-    async fn next_batch(&mut self) -> Result<Option<Batch>> {
         let scan_start = Instant::now();
         if let Some(batch) = self.batches.pop_front() {
             self.metrics.num_rows += batch.num_rows();
@@ -1104,6 +1147,16 @@ impl BatchReader for RowGroupReader {
     }
 }
 
+#[async_trait::async_trait]
+impl<T> BatchReader for RowGroupReaderBase<T>
+where
+    T: RowGroupReaderContext,
+{
+    async fn next_batch(&mut self) -> Result<Option<Batch>> {
+        self.next_inner()
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use parquet::arrow::arrow_reader::RowSelector;

src/mito2/src/sst/parquet/row_group.rs

@@ -38,25 +38,196 @@ use crate::sst::file::FileId;
 use crate::sst::parquet::helper::fetch_byte_ranges;
 use crate::sst::parquet::page_reader::RowGroupCachedReader;
 
-/// An in-memory collection of column chunks
+pub(crate) struct RowGroupBase<'a> {
-pub struct InMemoryRowGroup<'a> {
+    pub(crate) metadata: &'a RowGroupMetaData,
-    metadata: &'a RowGroupMetaData,
+    pub(crate) page_locations: Option<&'a [Vec<PageLocation>]>,
-    page_locations: Option<&'a [Vec<PageLocation>]>,
     /// Compressed page of each column.
-    column_chunks: Vec<Option<Arc<ColumnChunkData>>>,
+    pub(crate) column_chunks: Vec<Option<Arc<ColumnChunkData>>>,
-    row_count: usize,
+    pub(crate) row_count: usize,
-    region_id: RegionId,
-    file_id: FileId,
-    row_group_idx: usize,
-    cache_manager: Option<CacheManagerRef>,
     /// Row group level cached pages for each column.
     ///
     /// These pages are uncompressed pages of a row group.
     /// `column_uncompressed_pages.len()` equals to `column_chunks.len()`.
-    column_uncompressed_pages: Vec<Option<Arc<PageValue>>>,
+    pub(crate) column_uncompressed_pages: Vec<Option<Arc<PageValue>>>,
+}
+
+impl<'a> RowGroupBase<'a> {
+    pub(crate) fn new(parquet_meta: &'a ParquetMetaData, row_group_idx: usize) -> Self {
+        let metadata = parquet_meta.row_group(row_group_idx);
+        // `page_locations` is always `None` if we don't set
+        // [with_page_index()](https://docs.rs/parquet/latest/parquet/arrow/arrow_reader/struct.ArrowReaderOptions.html#method.with_page_index)
+        // to `true`.
+        let page_locations = parquet_meta
+            .offset_index()
+            .map(|x| x[row_group_idx].as_slice());
+
+        Self {
+            metadata,
+            page_locations,
+            column_chunks: vec![None; metadata.columns().len()],
+            row_count: metadata.num_rows() as usize,
+            column_uncompressed_pages: vec![None; metadata.columns().len()],
+        }
+    }
+
+    pub(crate) fn calc_sparse_read_ranges(
+        &self,
+        projection: &ProjectionMask,
+        page_locations: &[Vec<PageLocation>],
+        selection: &RowSelection,
+    ) -> (Vec<Range<u64>>, Vec<Vec<usize>>) {
+        // If we have a `RowSelection` and an `OffsetIndex` then only fetch pages required for the
+        // `RowSelection`
+        let mut page_start_offsets: Vec<Vec<usize>> = vec![];
+        let ranges = self
+            .column_chunks
+            .iter()
+            .zip(self.metadata.columns())
+            .enumerate()
+            .filter(|&(idx, (chunk, _chunk_meta))| chunk.is_none() && projection.leaf_included(idx))
+            .flat_map(|(idx, (_chunk, chunk_meta))| {
+                // If the first page does not start at the beginning of the column,
+                // then we need to also fetch a dictionary page.
+                let mut ranges = vec![];
+                let (start, _len) = chunk_meta.byte_range();
+                match page_locations[idx].first() {
+                    Some(first) if first.offset as u64 != start => {
+                        ranges.push(start..first.offset as u64);
+                    }
+                    _ => (),
+                }
+
+                ranges.extend(
+                    selection
+                        .scan_ranges(&page_locations[idx])
+                        .iter()
+                        .map(|range| range.start as u64..range.end as u64),
+                );
+                page_start_offsets.push(ranges.iter().map(|range| range.start as usize).collect());
+
+                ranges
+            })
+            .collect::<Vec<_>>();
+        (ranges, page_start_offsets)
+    }
+
+    pub(crate) fn assign_sparse_chunk(
+        &mut self,
+        projection: &ProjectionMask,
+        data: Vec<Bytes>,
+        page_start_offsets: Vec<Vec<usize>>,
+    ) {
+        let mut page_start_offsets = page_start_offsets.into_iter();
+        let mut chunk_data = data.into_iter();
+
+        for (idx, chunk) in self.column_chunks.iter_mut().enumerate() {
+            if chunk.is_some() || !projection.leaf_included(idx) {
+                continue;
+            }
+
+            if let Some(offsets) = page_start_offsets.next() {
+                let mut chunks = Vec::with_capacity(offsets.len());
+                for _ in 0..offsets.len() {
+                    chunks.push(chunk_data.next().unwrap());
+                }
+
+                *chunk = Some(Arc::new(ColumnChunkData::Sparse {
+                    length: self.metadata.column(idx).byte_range().1 as usize,
+                    data: offsets.into_iter().zip(chunks).collect(),
+                }))
+            }
+        }
+    }
+
+    pub(crate) fn calc_dense_read_ranges(&self, projection: &ProjectionMask) -> Vec<Range<u64>> {
+        self.column_chunks
+            .iter()
+            .zip(&self.column_uncompressed_pages)
+            .enumerate()
+            .filter(|&(idx, (chunk, uncompressed_pages))| {
+                // Don't need to fetch column data if we already cache the column's pages.
+                chunk.is_none() && projection.leaf_included(idx) && uncompressed_pages.is_none()
+            })
+            .map(|(idx, (_chunk, _pages))| {
+                let column = self.metadata.column(idx);
+                let (start, length) = column.byte_range();
+                start..(start + length)
+            })
+            .collect::<Vec<_>>()
+    }
+
+    /// Assigns uncompressed chunk binary data to [RowGroupBase::column_chunks]
+    /// and returns the chunk offset and binary data assigned.
+    pub(crate) fn assign_dense_chunk(
+        &mut self,
+        projection: &ProjectionMask,
+        chunk_data: Vec<Bytes>,
+    ) -> Vec<(usize, Bytes)> {
+        let mut chunk_data = chunk_data.into_iter();
+        let mut res = vec![];
+
+        for (idx, (chunk, row_group_pages)) in self
+            .column_chunks
+            .iter_mut()
+            .zip(&self.column_uncompressed_pages)
+            .enumerate()
+        {
+            if chunk.is_some() || !projection.leaf_included(idx) || row_group_pages.is_some() {
+                continue;
+            }
+
+            // Get the fetched page.
+            let Some(data) = chunk_data.next() else {
+                continue;
+            };
+
+            let column = self.metadata.column(idx);
+            res.push((idx, data.clone()));
+            *chunk = Some(Arc::new(ColumnChunkData::Dense {
+                offset: column.byte_range().0 as usize,
+                data,
+            }));
+        }
+        res
+    }
+
+    /// Create [PageReader] from [RowGroupBase::column_chunks]
+    pub(crate) fn column_reader(
+        &self,
+        col_idx: usize,
+    ) -> Result<SerializedPageReader<ColumnChunkData>> {
+        let page_reader = match &self.column_chunks[col_idx] {
+            None => {
+                return Err(ParquetError::General(format!(
+                    "Invalid column index {col_idx}, column was not fetched"
+                )))
+            }
+            Some(data) => {
+                let page_locations = self.page_locations.map(|index| index[col_idx].clone());
+                SerializedPageReader::new(
+                    data.clone(),
+                    self.metadata.column(col_idx),
+                    self.row_count,
+                    page_locations,
+                )?
+            }
+        };
+
+        // This column don't cache uncompressed pages.
+        Ok(page_reader)
+    }
+}
+
+/// An in-memory collection of column chunks
+pub struct InMemoryRowGroup<'a> {
+    region_id: RegionId,
+    file_id: FileId,
+    row_group_idx: usize,
+    cache_manager: Option<CacheManagerRef>,
     file_path: &'a str,
     /// Object store.
     object_store: ObjectStore,
+    base: RowGroupBase<'a>,
 }
 
 impl<'a> InMemoryRowGroup<'a> {
@@ -73,24 +244,12 @@ impl<'a> InMemoryRowGroup<'a> {
         file_path: &'a str,
         object_store: ObjectStore,
     ) -> Self {
-        let metadata = parquet_meta.row_group(row_group_idx);
-        // `page_locations` is always `None` if we don't set
-        // [with_page_index()](https://docs.rs/parquet/latest/parquet/arrow/arrow_reader/struct.ArrowReaderOptions.html#method.with_page_index)
-        // to `true`.
-        let page_locations = parquet_meta
-            .offset_index()
-            .map(|x| x[row_group_idx].as_slice());
-
         Self {
-            metadata,
+            base: RowGroupBase::new(parquet_meta, row_group_idx),
-            row_count: metadata.num_rows() as usize,
-            column_chunks: vec![None; metadata.columns().len()],
-            page_locations,
             region_id,
             file_id,
             row_group_idx,
             cache_manager,
-            column_uncompressed_pages: vec![None; metadata.columns().len()],
             file_path,
             object_store,
         }
@@ -102,65 +261,15 @@ impl<'a> InMemoryRowGroup<'a> {
         projection: &ProjectionMask,
         selection: Option<&RowSelection>,
     ) -> Result<()> {
-        if let Some((selection, page_locations)) = selection.zip(self.page_locations) {
+        if let Some((selection, page_locations)) = selection.zip(self.base.page_locations) {
-            // If we have a `RowSelection` and an `OffsetIndex` then only fetch pages required for the
+            let (fetch_ranges, page_start_offsets) =
-            // `RowSelection`
+                self.base
-            let mut page_start_offsets: Vec<Vec<usize>> = vec![];
+                    .calc_sparse_read_ranges(projection, page_locations, selection);
 
-            let fetch_ranges = self
+            let chunk_data = self.fetch_bytes(&fetch_ranges).await?;
-                .column_chunks
+            // Assign sparse chunk data to base.
-                .iter()
+            self.base
-                .zip(self.metadata.columns())
+                .assign_sparse_chunk(projection, chunk_data, page_start_offsets);
-                .enumerate()
-                .filter(|&(idx, (chunk, _chunk_meta))| {
-                    chunk.is_none() && projection.leaf_included(idx)
-                })
-                .flat_map(|(idx, (_chunk, chunk_meta))| {
-                    // If the first page does not start at the beginning of the column,
-                    // then we need to also fetch a dictionary page.
-                    let mut ranges = vec![];
-                    let (start, _len) = chunk_meta.byte_range();
-                    match page_locations[idx].first() {
-                        Some(first) if first.offset as u64 != start => {
-                            ranges.push(start..first.offset as u64);
-                        }
-                        _ => (),
-                    }
-
-                    ranges.extend(
-                        selection
-                            .scan_ranges(&page_locations[idx])
-                            .iter()
-                            .map(|range| range.start as u64..range.end as u64),
-                    );
-                    page_start_offsets
-                        .push(ranges.iter().map(|range| range.start as usize).collect());
-
-                    ranges
-                })
-                .collect::<Vec<_>>();
-
-            let mut chunk_data = self.fetch_bytes(&fetch_ranges).await?.into_iter();
-
-            let mut page_start_offsets = page_start_offsets.into_iter();
-
-            for (idx, chunk) in self.column_chunks.iter_mut().enumerate() {
-                if chunk.is_some() || !projection.leaf_included(idx) {
-                    continue;
-                }
-
-                if let Some(offsets) = page_start_offsets.next() {
-                    let mut chunks = Vec::with_capacity(offsets.len());
-                    for _ in 0..offsets.len() {
-                        chunks.push(chunk_data.next().unwrap());
-                    }
-
-                    *chunk = Some(Arc::new(ColumnChunkData::Sparse {
-                        length: self.metadata.column(idx).byte_range().1 as usize,
-                        data: offsets.into_iter().zip(chunks).collect(),
-                    }))
-                }
-            }
         } else {
             // Now we only use cache in dense chunk data.
             self.fetch_pages_from_cache(projection);
@@ -169,46 +278,24 @@ impl<'a> InMemoryRowGroup<'a> {
             // is a synchronous, CPU-bound operation.
             yield_now().await;
 
-            let fetch_ranges = self
+            // Calculate ranges to read.
-                .column_chunks
+            let fetch_ranges = self.base.calc_dense_read_ranges(projection);
-                .iter()
-                .zip(&self.column_uncompressed_pages)
-                .enumerate()
-                .filter(|&(idx, (chunk, uncompressed_pages))| {
-                    // Don't need to fetch column data if we already cache the column's pages.
-                    chunk.is_none() && projection.leaf_included(idx) && uncompressed_pages.is_none()
-                })
-                .map(|(idx, (_chunk, _pages))| {
-                    let column = self.metadata.column(idx);
-                    let (start, length) = column.byte_range();
-                    start..(start + length)
-                })
-                .collect::<Vec<_>>();
 
             if fetch_ranges.is_empty() {
                 // Nothing to fetch.
                 return Ok(());
             }
 
-            let mut chunk_data = self.fetch_bytes(&fetch_ranges).await?.into_iter();
+            // Fetch data with ranges
+            let chunk_data = self.fetch_bytes(&fetch_ranges).await?;
 
-            for (idx, (chunk, row_group_pages)) in self
+            // Assigns fetched data to base.
-                .column_chunks
+            let assigned_columns = self.base.assign_dense_chunk(projection, chunk_data);
-                .iter_mut()
-                .zip(&self.column_uncompressed_pages)
-                .enumerate()
-            {
-                if chunk.is_some() || !projection.leaf_included(idx) || row_group_pages.is_some() {
-                    continue;
-                }
 
-                // Get the fetched page.
+            // Put fetched data to cache if necessary.
-                let Some(data) = chunk_data.next() else {
+            if let Some(cache) = &self.cache_manager {
-                    continue;
+                for (col_idx, data) in assigned_columns {
-                };
+                    let column = self.base.metadata.column(col_idx);
-
-                let column = self.metadata.column(idx);
-                if let Some(cache) = &self.cache_manager {
                     if !cache_uncompressed_pages(column) {
                         // For columns that have multiple uncompressed pages, we only cache the compressed page
                         // to save memory.
@@ -216,17 +303,12 @@ impl<'a> InMemoryRowGroup<'a> {
                             self.region_id,
                             self.file_id,
                             self.row_group_idx,
-                            idx,
+                            col_idx,
                         );
                         cache
                             .put_pages(page_key, Arc::new(PageValue::new_compressed(data.clone())));
                     }
                 }
-
-                *chunk = Some(Arc::new(ColumnChunkData::Dense {
-                    offset: column.byte_range().0 as usize,
-                    data,
-                }));
             }
         }
 
@@ -237,7 +319,8 @@ impl<'a> InMemoryRowGroup<'a> {
     /// If the page is in the cache, sets the column chunk or `column_uncompressed_pages` for the column.
     fn fetch_pages_from_cache(&mut self, projection: &ProjectionMask) {
         let _timer = READ_STAGE_FETCH_PAGES.start_timer();
-        self.column_chunks
+        self.base
+            .column_chunks
             .iter_mut()
             .enumerate()
             .filter(|(idx, chunk)| chunk.is_none() && projection.leaf_included(*idx))
@@ -245,7 +328,7 @@ impl<'a> InMemoryRowGroup<'a> {
                 let Some(cache) = &self.cache_manager else {
                     return;
                 };
-                let column = self.metadata.column(idx);
+                let column = self.base.metadata.column(idx);
                 if cache_uncompressed_pages(column) {
                     // Fetches uncompressed pages for the row group.
                     let page_key = PageKey::new_uncompressed(
@@ -254,7 +337,7 @@ impl<'a> InMemoryRowGroup<'a> {
                         self.row_group_idx,
                         idx,
                     );
-                    self.column_uncompressed_pages[idx] = cache.get_pages(&page_key);
+                    self.base.column_uncompressed_pages[idx] = cache.get_pages(&page_key);
                 } else {
                     // Fetches the compressed page from the cache.
                     let page_key = PageKey::new_compressed(
@@ -308,34 +391,19 @@ impl<'a> InMemoryRowGroup<'a> {
 
     /// Creates a page reader to read column at `i`.
     fn column_page_reader(&self, i: usize) -> Result<Box<dyn PageReader>> {
-        if let Some(cached_pages) = &self.column_uncompressed_pages[i] {
+        if let Some(cached_pages) = &self.base.column_uncompressed_pages[i] {
             debug_assert!(!cached_pages.row_group.is_empty());
             // Hits the row group level page cache.
             return Ok(Box::new(RowGroupCachedReader::new(&cached_pages.row_group)));
         }
 
-        let page_reader = match &self.column_chunks[i] {
+        let page_reader = self.base.column_reader(i)?;
-            None => {
-                return Err(ParquetError::General(format!(
-                    "Invalid column index {i}, column was not fetched"
-                )))
-            }
-            Some(data) => {
-                let page_locations = self.page_locations.map(|index| index[i].clone());
-                SerializedPageReader::new(
-                    data.clone(),
-                    self.metadata.column(i),
-                    self.row_count,
-                    page_locations,
-                )?
-            }
-        };
 
         let Some(cache) = &self.cache_manager else {
             return Ok(Box::new(page_reader));
         };
 
-        let column = self.metadata.column(i);
+        let column = self.base.metadata.column(i);
         if cache_uncompressed_pages(column) {
             // This column use row group level page cache.
             // We collect all pages and put them into the cache.
@@ -362,7 +430,7 @@ fn cache_uncompressed_pages(column: &ColumnChunkMetaData) -> bool {
 
 impl RowGroups for InMemoryRowGroup<'_> {
     fn num_rows(&self) -> usize {
-        self.row_count
+        self.base.row_count
     }
 
     fn column_chunks(&self, i: usize) -> Result<Box<dyn PageIterator>> {
@@ -430,8 +498,8 @@ impl ChunkReader for ColumnChunkData {
 }
 
 /// Implements [`PageIterator`] for a single column chunk, yielding a single [`PageReader`]
-struct ColumnChunkIterator {
+pub(crate) struct ColumnChunkIterator {
-    reader: Option<Result<Box<dyn PageReader>>>,
+    pub(crate) reader: Option<Result<Box<dyn PageReader>>>,
 }
 
 impl Iterator for ColumnChunkIterator {

src/mito2/src/test_util/memtable_util.rs

@@ -14,7 +14,6 @@
 
 //! Memtable test utilities.
 
-use std::collections::BTreeMap;
 use std::sync::Arc;
 
 use api::helper::ColumnDataTypeWrapper;
@@ -34,8 +33,8 @@ use crate::error::Result;
 use crate::memtable::key_values::KeyValue;
 use crate::memtable::partition_tree::data::{timestamp_array_to_i64_slice, DataBatch, DataBuffer};
 use crate::memtable::{
-    BoxedBatchIterator, BulkPart, KeyValues, Memtable, MemtableBuilder, MemtableId, MemtableRange,
+    BoxedBatchIterator, BulkPart, KeyValues, Memtable, MemtableBuilder, MemtableId, MemtableRanges,
-    MemtableRanges, MemtableRef, MemtableStats,
+    MemtableRef, MemtableStats,
 };
 use crate::row_converter::{McmpRowCodec, RowCodec, SortField};
 

src/mito2/src/worker/handle_open.rs

@@ -51,7 +51,7 @@ impl<S: LogStore> RegionWorkerLoop<S> {
         // Check if this region is pending drop. And clean the entire dir if so.
         if !self.dropping_regions.is_region_exists(region_id)
             && object_store
-                .is_exist(&join_path(&request.region_dir, DROPPING_MARKER_FILE))
+                .exists(&join_path(&request.region_dir, DROPPING_MARKER_FILE))
                 .await
                 .context(OpenDalSnafu)?
         {

src/object-store/Cargo.toml

@@ -17,8 +17,9 @@ futures.workspace = true
 lazy_static.workspace = true
 md5 = "0.7"
 moka = { workspace = true, features = ["future"] }
-opendal = { version = "0.49", features = [
+opendal = { git = "https://github.com/GreptimeTeam/opendal.git", rev = "c82605177f2feec83e49dcaa537c505639d94024", features = [
     "layers-tracing",
+    "layers-prometheus",
     "services-azblob",
     "services-fs",
     "services-gcs",

src/object-store/src/layers.rs

@@ -13,8 +13,37 @@
 // limitations under the License.
 
 mod lru_cache;
-mod prometheus;
 
 pub use lru_cache::*;
 pub use opendal::layers::*;
-pub use prometheus::PrometheusMetricsLayer;
+pub use prometheus::build_prometheus_metrics_layer;
+
+mod prometheus {
+    use std::sync::{Mutex, OnceLock};
+
+    use opendal::layers::PrometheusLayer;
+
+    static PROMETHEUS_LAYER: OnceLock<Mutex<PrometheusLayer>> = OnceLock::new();
+
+    pub fn build_prometheus_metrics_layer(with_path_label: bool) -> PrometheusLayer {
+        PROMETHEUS_LAYER
+            .get_or_init(|| {
+                // This logical tries to extract parent path from the object storage operation
+                // the function also relies on assumption that the region path is built from
+                // pattern `<data|index>/catalog/schema/table_id/....`
+                //
+                // We'll get the data/catalog/schema from path.
+                let path_level = if with_path_label { 3 } else { 0 };
+
+                let layer = PrometheusLayer::builder()
+                    .path_label(path_level)
+                    .register_default()
+                    .unwrap();
+
+                Mutex::new(layer)
+            })
+            .lock()
+            .unwrap()
+            .clone()
+    }
+}

src/object-store/src/layers/lru_cache/read_cache.rs

@@ -156,9 +156,12 @@ impl<C: Access> ReadCache<C> {
             let size = entry.metadata().content_length();
             OBJECT_STORE_LRU_CACHE_ENTRIES.inc();
             OBJECT_STORE_LRU_CACHE_BYTES.add(size as i64);
-            self.mem_cache
+            // ignore root path
-                .insert(read_key.to_string(), ReadResult::Success(size as u32))
+            if entry.path() != "/" {
-                .await;
+                self.mem_cache
+                    .insert(read_key.to_string(), ReadResult::Success(size as u32))
+                    .await;
+            }
         }
 
         Ok(self.cache_stat().await)