chore: add tests to comment column on information_schema (#7514 )

* feat: show comment on information_schema Signed-off-by: Ruihang Xia <waynestxia@gmail.com> * add to information schema for columns, add sqlness tests Signed-off-by: Ruihang Xia <waynestxia@gmail.com> * remove duplications Signed-off-by: Ruihang Xia <waynestxia@gmail.com> * fix typo Signed-off-by: Ruihang Xia <waynestxia@gmail.com> * update integration test Signed-off-by: Ruihang Xia <waynestxia@gmail.com> --------- Signed-off-by: Ruihang Xia <waynestxia@gmail.com>
feat(meta-srv): fallback to non-TLS connection when etcd TLS prefer mode fail (#7507 )
2026-01-04 20:32:56 +00:00 · 2026-01-04 09:05:50 +00:00 · 2025-12-31 10:03:34 +00:00 · 2025-12-31 04:55:22 +00:00 · 2025-12-31 03:31:08 +00:00 · 2025-12-30 09:42:16 +00:00
60 changed files with 4075 additions and 355 deletions
--- a/.github/actions/setup-greptimedb-cluster/action.yml
+++ b/.github/actions/setup-greptimedb-cluster/action.yml
@@ -70,19 +70,23 @@ runs:
        --wait \
        --wait-for-jobs
  - name: Wait for GreptimeDB
-    shell: bash
-    run: |
-      while true; do
-        PHASE=$(kubectl -n my-greptimedb get gtc my-greptimedb -o jsonpath='{.status.clusterPhase}')
-        if [ "$PHASE" == "Running" ]; then
-          echo "Cluster is ready"
-          break
-        else
-          echo "Cluster is not ready yet: Current phase: $PHASE"
-          kubectl get pods -n my-greptimedb
-          sleep 5 # wait for 5 seconds before check again.
-        fi
-      done
+    uses: nick-fields/retry@v3
+    with:
+      timeout_minutes: 3
+      max_attempts: 1
+      shell: bash
+      command: |
+        while true; do
+          PHASE=$(kubectl -n my-greptimedb get gtc my-greptimedb -o jsonpath='{.status.clusterPhase}')
+          if [ "$PHASE" == "Running" ]; then
+            echo "Cluster is ready"
+            break
+          else
+            echo "Cluster is not ready yet: Current phase: $PHASE"
+            kubectl get pods -n my-greptimedb
+            sleep 5 # wait for 5 seconds before check again.
+          fi
+        done
  - name: Print GreptimeDB info
    if: always()
    shell: bash
--- a/.github/workflows/develop.yml
+++ b/.github/workflows/develop.yml
@@ -755,7 +755,7 @@ jobs:
        run: ../../.github/scripts/pull-test-deps-images.sh && docker compose up -d --wait

      - name: Run nextest cases
-        run: cargo nextest run --workspace -F dashboard -F pg_kvbackend -F mysql_kvbackend
+        run: cargo nextest run --workspace -F dashboard -F pg_kvbackend -F mysql_kvbackend -F vector_index
        env:
          CARGO_BUILD_RUSTFLAGS: "-C link-arg=-fuse-ld=mold"
          RUST_BACKTRACE: 1
@@ -813,7 +813,7 @@ jobs:
        run: ../../.github/scripts/pull-test-deps-images.sh && docker compose up -d --wait

      - name: Run nextest cases
-        run: cargo llvm-cov nextest --workspace --lcov --output-path lcov.info -F dashboard -F pg_kvbackend -F mysql_kvbackend
+        run: cargo llvm-cov nextest --workspace --lcov --output-path lcov.info -F dashboard -F pg_kvbackend -F mysql_kvbackend -F vector_index
        env:
          CARGO_BUILD_RUSTFLAGS: "-C link-arg=-fuse-ld=mold"
          RUST_BACKTRACE: 1
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -212,7 +212,7 @@ checksum = "d301b3b94cb4b2f23d7917810addbbaff90738e0ca2be692bd027e70d7e0330c"

 [[package]]
 name = "api"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "arrow-schema",
 "common-base",
@@ -733,7 +733,7 @@ dependencies = [

 [[package]]
 name = "auth"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "async-trait",
@@ -1383,7 +1383,7 @@ dependencies = [

 [[package]]
 name = "cache"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "catalog",
 "common-error",
@@ -1418,7 +1418,7 @@ checksum = "37b2a672a2cb129a2e41c10b1224bb368f9f37a2b16b612598138befd7b37eb5"

 [[package]]
 name = "catalog"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "arrow",
@@ -1763,7 +1763,7 @@ checksum = "b94f61472cee1439c0b966b47e3aca9ae07e45d070759512cd390ea2bebc6675"

 [[package]]
 name = "cli"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "async-stream",
 "async-trait",
@@ -1817,7 +1817,7 @@ dependencies = [

 [[package]]
 name = "client"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "arc-swap",
@@ -1850,7 +1850,7 @@ dependencies = [
 "snafu 0.8.6",
 "store-api",
 "substrait 0.37.3",
- "substrait 1.0.0-beta.3",
+ "substrait 1.0.0-beta.4",
 "tokio",
 "tokio-stream",
 "tonic 0.13.1",
@@ -1890,7 +1890,7 @@ dependencies = [

 [[package]]
 name = "cmd"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "async-trait",
 "auth",
@@ -2024,7 +2024,7 @@ checksum = "55b672471b4e9f9e95499ea597ff64941a309b2cdbffcc46f2cc5e2d971fd335"

 [[package]]
 name = "common-base"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "anymap2",
 "async-trait",
@@ -2048,14 +2048,14 @@ dependencies = [

 [[package]]
 name = "common-catalog"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "const_format",
 ]

 [[package]]
 name = "common-config"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "common-base",
 "common-error",
@@ -2080,7 +2080,7 @@ dependencies = [

 [[package]]
 name = "common-datasource"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "arrow",
 "arrow-schema",
@@ -2115,7 +2115,7 @@ dependencies = [

 [[package]]
 name = "common-decimal"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "bigdecimal 0.4.8",
 "common-error",
@@ -2128,7 +2128,7 @@ dependencies = [

 [[package]]
 name = "common-error"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "common-macro",
 "http 1.3.1",
@@ -2139,7 +2139,7 @@ dependencies = [

 [[package]]
 name = "common-event-recorder"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "async-trait",
@@ -2161,7 +2161,7 @@ dependencies = [

 [[package]]
 name = "common-frontend"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "async-trait",
@@ -2183,13 +2183,14 @@ dependencies = [

 [[package]]
 name = "common-function"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "ahash 0.8.12",
 "api",
 "approx 0.5.1",
 "arc-swap",
 "arrow",
+ "arrow-cast",
 "arrow-schema",
 "async-trait",
 "bincode",
@@ -2220,6 +2221,7 @@ dependencies = [
 "h3o",
 "hyperloglogplus",
 "jsonb",
+ "jsonpath-rust 0.7.5",
 "memchr",
 "mito-codec",
 "nalgebra",
@@ -2243,7 +2245,7 @@ dependencies = [

 [[package]]
 name = "common-greptimedb-telemetry"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "async-trait",
 "common-runtime",
@@ -2260,7 +2262,7 @@ dependencies = [

 [[package]]
 name = "common-grpc"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "arrow-flight",
@@ -2295,7 +2297,7 @@ dependencies = [

 [[package]]
 name = "common-grpc-expr"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "common-base",
@@ -2315,7 +2317,7 @@ dependencies = [

 [[package]]
 name = "common-macro"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "greptime-proto",
 "once_cell",
@@ -2326,7 +2328,7 @@ dependencies = [

 [[package]]
 name = "common-mem-prof"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "anyhow",
 "common-error",
@@ -2342,7 +2344,7 @@ dependencies = [

 [[package]]
 name = "common-memory-manager"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "common-error",
 "common-macro",
@@ -2355,7 +2357,7 @@ dependencies = [

 [[package]]
 name = "common-meta"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "anymap2",
 "api",
@@ -2427,7 +2429,7 @@ dependencies = [

 [[package]]
 name = "common-options"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "common-grpc",
 "humantime-serde",
@@ -2436,11 +2438,11 @@ dependencies = [

 [[package]]
 name = "common-plugins"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"

 [[package]]
 name = "common-pprof"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "common-error",
 "common-macro",
@@ -2452,7 +2454,7 @@ dependencies = [

 [[package]]
 name = "common-procedure"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "async-stream",
@@ -2481,7 +2483,7 @@ dependencies = [

 [[package]]
 name = "common-procedure-test"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "async-trait",
 "common-procedure",
@@ -2491,7 +2493,7 @@ dependencies = [

 [[package]]
 name = "common-query"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "async-trait",
@@ -2517,7 +2519,7 @@ dependencies = [

 [[package]]
 name = "common-recordbatch"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "arc-swap",
 "common-base",
@@ -2541,7 +2543,7 @@ dependencies = [

 [[package]]
 name = "common-runtime"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "async-trait",
 "clap 4.5.40",
@@ -2570,7 +2572,7 @@ dependencies = [

 [[package]]
 name = "common-session"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "serde",
 "strum 0.27.1",
@@ -2578,7 +2580,7 @@ dependencies = [

 [[package]]
 name = "common-sql"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "arrow-schema",
 "common-base",
@@ -2598,7 +2600,7 @@ dependencies = [

 [[package]]
 name = "common-stat"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "common-base",
 "common-runtime",
@@ -2613,7 +2615,7 @@ dependencies = [

 [[package]]
 name = "common-telemetry"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "backtrace",
 "common-base",
@@ -2642,7 +2644,7 @@ dependencies = [

 [[package]]
 name = "common-test-util"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "client",
 "common-grpc",
@@ -2655,7 +2657,7 @@ dependencies = [

 [[package]]
 name = "common-time"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "arrow",
 "chrono",
@@ -2673,7 +2675,7 @@ dependencies = [

 [[package]]
 name = "common-version"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "build-data",
 "cargo-manifest",
@@ -2684,7 +2686,7 @@ dependencies = [

 [[package]]
 name = "common-wal"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "common-base",
 "common-error",
@@ -2707,7 +2709,7 @@ dependencies = [

 [[package]]
 name = "common-workload"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "common-telemetry",
 "serde",
@@ -4015,7 +4017,7 @@ dependencies = [

 [[package]]
 name = "datanode"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "arrow-flight",
@@ -4079,7 +4081,7 @@ dependencies = [

 [[package]]
 name = "datatypes"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "arrow",
 "arrow-array",
@@ -4754,7 +4756,7 @@ checksum = "37909eebbb50d72f9059c3b6d82c0463f2ff062c9e95845c43a6c9c0355411be"

 [[package]]
 name = "file-engine"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "async-trait",
@@ -4886,7 +4888,7 @@ checksum = "8bf7cc16383c4b8d58b9905a8509f02926ce3058053c056376248d958c9df1e8"

 [[package]]
 name = "flow"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "arrow",
@@ -4955,7 +4957,7 @@ dependencies = [
 "sql",
 "store-api",
 "strum 0.27.1",
- "substrait 1.0.0-beta.3",
+ "substrait 1.0.0-beta.4",
 "table",
 "tokio",
 "tonic 0.13.1",
@@ -5016,7 +5018,7 @@ checksum = "28dd6caf6059519a65843af8fe2a3ae298b14b80179855aeb4adc2c1934ee619"

 [[package]]
 name = "frontend"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "arc-swap",
@@ -5464,7 +5466,7 @@ dependencies = [
 [[package]]
 name = "greptime-proto"
 version = "0.1.0"
-source = "git+https://github.com/GreptimeTeam/greptime-proto.git?rev=520fa524f9d590752ea327683e82ffd65721b27c#520fa524f9d590752ea327683e82ffd65721b27c"
+source = "git+https://github.com/GreptimeTeam/greptime-proto.git?rev=a2e5099d72a1cfa8ba41fa4296101eb5f874074a#a2e5099d72a1cfa8ba41fa4296101eb5f874074a"
 dependencies = [
 "prost 0.13.5",
 "prost-types 0.13.5",
@@ -6232,7 +6234,7 @@ dependencies = [

 [[package]]
 name = "index"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "async-trait",
 "asynchronous-codec",
@@ -7173,7 +7175,7 @@ checksum = "13dc2df351e3202783a1fe0d44375f7295ffb4049267b0f3018346dc122a1d94"

 [[package]]
 name = "log-query"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "chrono",
 "common-error",
@@ -7185,7 +7187,7 @@ dependencies = [

 [[package]]
 name = "log-store"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "async-stream",
 "async-trait",
@@ -7486,7 +7488,7 @@ dependencies = [

 [[package]]
 name = "meta-client"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "async-trait",
@@ -7514,7 +7516,7 @@ dependencies = [

 [[package]]
 name = "meta-srv"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "async-trait",
@@ -7614,7 +7616,7 @@ dependencies = [

 [[package]]
 name = "metric-engine"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "aquamarine",
@@ -7712,7 +7714,7 @@ dependencies = [

 [[package]]
 name = "mito-codec"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "bytes",
@@ -7737,7 +7739,7 @@ dependencies = [

 [[package]]
 name = "mito2"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "aquamarine",
@@ -7777,7 +7779,6 @@ dependencies = [
 "either",
 "futures",
 "greptime-proto",
- "humantime",
 "humantime-serde",
 "index",
 "itertools 0.14.0",
@@ -7796,6 +7797,7 @@ dependencies = [
 "rand 0.9.1",
 "rayon",
 "regex",
+ "roaring",
 "rskafka",
 "rstest",
 "rstest_reuse",
@@ -7814,6 +7816,7 @@ dependencies = [
 "tokio-util",
 "toml 0.8.23",
 "tracing",
+ "usearch",
 "uuid",
 ]

@@ -8477,7 +8480,7 @@ dependencies = [

 [[package]]
 name = "object-store"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "anyhow",
 "bytes",
@@ -8762,7 +8765,7 @@ dependencies = [

 [[package]]
 name = "operator"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "ahash 0.8.12",
 "api",
@@ -8822,7 +8825,7 @@ dependencies = [
 "sql",
 "sqlparser",
 "store-api",
- "substrait 1.0.0-beta.3",
+ "substrait 1.0.0-beta.4",
 "table",
 "tokio",
 "tokio-util",
@@ -9108,7 +9111,7 @@ dependencies = [

 [[package]]
 name = "partition"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "async-trait",
@@ -9466,7 +9469,7 @@ checksum = "8b870d8c151b6f2fb93e84a13146138f05d02ed11c7e7c54f8826aaaf7c9f184"

 [[package]]
 name = "pipeline"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "ahash 0.8.12",
 "api",
@@ -9622,7 +9625,7 @@ dependencies = [

 [[package]]
 name = "plugins"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "auth",
 "catalog",
@@ -9924,7 +9927,7 @@ dependencies = [

 [[package]]
 name = "promql"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "ahash 0.8.12",
 "async-trait",
@@ -10207,7 +10210,7 @@ dependencies = [

 [[package]]
 name = "puffin"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "async-compression 0.4.19",
 "async-trait",
@@ -10249,7 +10252,7 @@ dependencies = [

 [[package]]
 name = "query"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "ahash 0.8.12",
 "api",
@@ -10316,7 +10319,7 @@ dependencies = [
 "sql",
 "sqlparser",
 "store-api",
- "substrait 1.0.0-beta.3",
+ "substrait 1.0.0-beta.4",
 "table",
 "tokio",
 "tokio-stream",
@@ -11668,7 +11671,7 @@ dependencies = [

 [[package]]
 name = "servers"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "ahash 0.8.12",
 "api",
@@ -11797,7 +11800,7 @@ dependencies = [

 [[package]]
 name = "session"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "ahash 0.8.12",
 "api",
@@ -12141,7 +12144,7 @@ dependencies = [

 [[package]]
 name = "sql"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "arrow-buffer",
@@ -12201,7 +12204,7 @@ dependencies = [

 [[package]]
 name = "sqlness-runner"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "async-trait",
 "clap 4.5.40",
@@ -12478,7 +12481,7 @@ dependencies = [

 [[package]]
 name = "standalone"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "async-trait",
 "catalog",
@@ -12520,7 +12523,7 @@ checksum = "a2eb9349b6444b326872e140eb1cf5e7c522154d69e7a0ffb0fb81c06b37543f"

 [[package]]
 name = "store-api"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "aquamarine",
@@ -12733,7 +12736,7 @@ dependencies = [

 [[package]]
 name = "substrait"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "async-trait",
 "bytes",
@@ -12856,7 +12859,7 @@ dependencies = [

 [[package]]
 name = "table"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "async-trait",
@@ -13125,7 +13128,7 @@ checksum = "8f50febec83f5ee1df3015341d8bd429f2d1cc62bcba7ea2076759d315084683"

 [[package]]
 name = "tests-fuzz"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "arbitrary",
 "async-trait",
@@ -13169,7 +13172,7 @@ dependencies = [

 [[package]]
 name = "tests-integration"
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 dependencies = [
 "api",
 "arrow-flight",
@@ -13245,7 +13248,7 @@ dependencies = [
 "sqlx",
 "standalone",
 "store-api",
- "substrait 1.0.0-beta.3",
+ "substrait 1.0.0-beta.4",
 "table",
 "tempfile",
 "time",
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -75,7 +75,7 @@ members = [
 resolver = "2"

 [workspace.package]
-version = "1.0.0-beta.3"
+version = "1.0.0-beta.4"
 edition = "2024"
 license = "Apache-2.0"

@@ -103,6 +103,7 @@ aquamarine = "0.6"
 arrow = { version = "56.2", features = ["prettyprint"] }
 arrow-array = { version = "56.2", default-features = false, features = ["chrono-tz"] }
 arrow-buffer = "56.2"
+arrow-cast = "56.2"
 arrow-flight = "56.2"
 arrow-ipc = { version = "56.2", default-features = false, features = ["lz4", "zstd"] }
 arrow-schema = { version = "56.2", features = ["serde"] }
@@ -150,7 +151,7 @@ etcd-client = { version = "0.16.1", features = [
 fst = "0.4.7"
 futures = "0.3"
 futures-util = "0.3"
-greptime-proto = { git = "https://github.com/GreptimeTeam/greptime-proto.git", rev = "520fa524f9d590752ea327683e82ffd65721b27c" }
+greptime-proto = { git = "https://github.com/GreptimeTeam/greptime-proto.git", rev = "a2e5099d72a1cfa8ba41fa4296101eb5f874074a" }
 hex = "0.4"
 http = "1"
 humantime = "2.1"
--- a/cliff.toml
+++ b/cliff.toml
@@ -17,7 +17,7 @@ Release date: {{ timestamp | date(format="%B %d, %Y") }}
 {%- set breakings = commits | filter(attribute="breaking", value=true) -%}
 {%- if breakings | length > 0 %}

-## Breaking changes
+### Breaking changes
    {% for commit in breakings %}
      * {{ commit.github.pr_title }}\
        {% if commit.github.username %} by \
--- a/cyborg/bin/bump-versions.ts
+++ b/cyborg/bin/bump-versions.ts
@@ -57,6 +57,20 @@ const REPO_CONFIGS: Record<string, RepoConfig> = {
        return ['bump-nightly-version.yml', version];
      }

+      // Check for prerelease versions (e.g., 1.0.0-beta.3, 1.0.0-rc.1)
+      const prereleaseMatch = version.match(/^(\d+)\.(\d+)\.(\d+)-(beta|rc)\.(\d+)$/);
+      if (prereleaseMatch) {
+        const [, major, minor, patch, prereleaseType, prereleaseNum] = prereleaseMatch;
+
+        // If it's beta.1 and patch version is 0, treat as major version
+        if (prereleaseType === 'beta' && prereleaseNum === '1' && patch === '0') {
+          return ['bump-version.yml', `${major}.${minor}`];
+        }
+
+        // Otherwise (beta.x where x > 1, or rc.x), treat as patch version
+        return ['bump-patch-version.yml', version];
+      }
+
      const parts = version.split('.');
      if (parts.length !== 3) {
        throw new Error('Invalid version format');
--- a/src/catalog/src/system_schema/information_schema/columns.rs
+++ b/src/catalog/src/system_schema/information_schema/columns.rs
@@ -399,8 +399,8 @@ impl InformationSchemaColumnsBuilder {
            self.is_nullables.push(Some("No"));
        }
        self.column_types.push(Some(&data_type));
-        self.column_comments
-            .push(column_schema.column_comment().map(|x| x.as_ref()));
+        let column_comment = column_schema.column_comment().map(|x| x.as_ref());
+        self.column_comments.push(column_comment);
    }

    fn finish(&mut self) -> Result<RecordBatch> {
--- a/src/cli/src/common/store.rs
+++ b/src/cli/src/common/store.rs
@@ -92,7 +92,7 @@ impl StoreConfig {
    pub fn tls_config(&self) -> Option<TlsOption> {
        if self.backend_tls_mode != TlsMode::Disable {
            Some(TlsOption {
-                mode: self.backend_tls_mode.clone(),
+                mode: self.backend_tls_mode,
                cert_path: self.backend_tls_cert_path.clone(),
                key_path: self.backend_tls_key_path.clone(),
                ca_cert_path: self.backend_tls_ca_cert_path.clone(),
--- a/src/cmd/Cargo.toml
+++ b/src/cmd/Cargo.toml
@@ -18,6 +18,7 @@ default = [
 ]
 enterprise = ["common-meta/enterprise", "frontend/enterprise", "meta-srv/enterprise"]
 tokio-console = ["common-telemetry/tokio-console"]
+vector_index = ["mito2/vector_index"]

 [lints]
 workspace = true
--- a/src/cmd/src/datanode/objbench.rs
+++ b/src/cmd/src/datanode/objbench.rs
@@ -233,6 +233,8 @@ impl ObjbenchCommand {
            inverted_index_config: MitoConfig::default().inverted_index,
            fulltext_index_config,
            bloom_filter_index_config: MitoConfig::default().bloom_filter_index,
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        };

        // Write SST
--- a/src/cmd/src/frontend.rs
+++ b/src/cmd/src/frontend.rs
@@ -236,7 +236,7 @@ impl StartCommand {
        };

        let tls_opts = TlsOption::new(
-            self.tls_mode.clone(),
+            self.tls_mode,
            self.tls_cert_path.clone(),
            self.tls_key_path.clone(),
            self.tls_watch,
--- a/src/cmd/src/standalone.rs
+++ b/src/cmd/src/standalone.rs
@@ -261,7 +261,7 @@ impl StartCommand {
        };

        let tls_opts = TlsOption::new(
-            self.tls_mode.clone(),
+            self.tls_mode,
            self.tls_cert_path.clone(),
            self.tls_key_path.clone(),
            self.tls_watch,
--- a/src/common/function/Cargo.toml
+++ b/src/common/function/Cargo.toml
@@ -17,6 +17,7 @@ ahash.workspace = true
 api.workspace = true
 arc-swap = "1.0"
 arrow.workspace = true
+arrow-cast.workspace = true
 arrow-schema.workspace = true
 async-trait.workspace = true
 bincode = "=1.3.3"
@@ -46,6 +47,7 @@ geohash = { version = "0.13", optional = true }
 h3o = { version = "0.6", optional = true }
 hyperloglogplus = "0.4"
 jsonb.workspace = true
+jsonpath-rust = "0.7.5"
 memchr = "2.7"
 mito-codec.workspace = true
 nalgebra.workspace = true
--- a/src/common/function/src/scalars/json/json_get.rs
+++ b/src/common/function/src/scalars/json/json_get.rs
@@ -13,17 +13,24 @@
 // limitations under the License.

 use std::fmt::{self, Display};
+use std::str::FromStr;
 use std::sync::Arc;

+use arrow::array::{ArrayRef, BinaryViewArray, StringViewArray, StructArray};
 use arrow::compute;
-use datafusion_common::DataFusionError;
+use arrow::datatypes::{Float64Type, Int64Type, UInt64Type};
 use datafusion_common::arrow::array::{
    Array, AsArray, BinaryViewBuilder, BooleanBuilder, Float64Builder, Int64Builder,
    StringViewBuilder,
 };
 use datafusion_common::arrow::datatypes::DataType;
+use datafusion_common::{DataFusionError, Result};
 use datafusion_expr::type_coercion::aggregates::STRINGS;
-use datafusion_expr::{ColumnarValue, ScalarFunctionArgs, Signature};
+use datafusion_expr::{ColumnarValue, ScalarFunctionArgs, Signature, Volatility};
+use datatypes::arrow_array::{int_array_value_at_index, string_array_value_at_index};
+use datatypes::json::JsonStructureSettings;
+use jsonpath_rust::JsonPath;
+use serde_json::Value;

 use crate::function::{Function, extract_args};
 use crate::helper;
@@ -124,13 +131,6 @@ macro_rules! json_get {
    };
 }

-json_get!(
-    JsonGetInt,
-    Int64,
-    i64,
-    "Get the value from the JSONB by the given path and return it as an integer."
-);
-
 json_get!(
    JsonGetFloat,
    Float64,
@@ -145,70 +145,356 @@ json_get!(
    "Get the value from the JSONB by the given path and return it as a boolean."
 );

-/// Get the value from the JSONB by the given path and return it as a string.
-#[derive(Clone, Debug)]
-pub struct JsonGetString {
+enum JsonResultValue<'a> {
+    Jsonb(Vec<u8>),
+    JsonStructByColumn(&'a ArrayRef, usize),
+    JsonStructByValue(&'a Value),
+}
+
+trait JsonGetResultBuilder {
+    fn append_value(&mut self, value: JsonResultValue<'_>) -> Result<()>;
+
+    fn append_null(&mut self);
+
+    fn build(&mut self) -> ArrayRef;
+}
+
+/// Common implementation for JSON get scalar functions.
+///
+/// `JsonGet` encapsulates the logic for extracting values from JSON inputs
+/// based on a path expression. Different JSON get functions reuse this
+/// implementation by supplying their own `JsonGetResultBuilder` to control
+/// how the resulting values are materialized into an Arrow array.
+struct JsonGet {
    signature: Signature,
 }

-impl JsonGetString {
-    pub const NAME: &'static str = "json_get_string";
+impl JsonGet {
+    fn invoke<F, B>(&self, args: ScalarFunctionArgs, builder_factory: F) -> Result<ColumnarValue>
+    where
+        F: Fn(usize) -> B,
+        B: JsonGetResultBuilder,
+    {
+        let [arg0, arg1] = extract_args("JSON_GET", &args)?;
+
+        let arg1 = compute::cast(&arg1, &DataType::Utf8View)?;
+        let paths = arg1.as_string_view();
+
+        let mut builder = (builder_factory)(arg0.len());
+        match arg0.data_type() {
+            DataType::Binary | DataType::LargeBinary | DataType::BinaryView => {
+                let arg0 = compute::cast(&arg0, &DataType::BinaryView)?;
+                let jsons = arg0.as_binary_view();
+                jsonb_get(jsons, paths, &mut builder)?;
+            }
+            DataType::Struct(_) => {
+                let jsons = arg0.as_struct();
+                json_struct_get(jsons, paths, &mut builder)?
+            }
+            _ => {
+                return Err(DataFusionError::Execution(format!(
+                    "JSON_GET not supported argument type {}",
+                    arg0.data_type(),
+                )));
+            }
+        };
+
+        Ok(ColumnarValue::Array(builder.build()))
+    }
 }

-impl Default for JsonGetString {
+impl Default for JsonGet {
    fn default() -> Self {
        Self {
-            // TODO(LFC): Use a more clear type here instead of "Binary" for Json input, once we have a "Json" type.
-            signature: helper::one_of_sigs2(
-                vec![DataType::Binary, DataType::BinaryView],
-                vec![DataType::Utf8, DataType::Utf8View],
-            ),
+            signature: Signature::any(2, Volatility::Immutable),
        }
    }
 }

+#[derive(Default)]
+pub struct JsonGetString(JsonGet);
+
+impl JsonGetString {
+    pub const NAME: &'static str = "json_get_string";
+}
+
 impl Function for JsonGetString {
    fn name(&self) -> &str {
        Self::NAME
    }

-    fn return_type(&self, _: &[DataType]) -> datafusion_common::Result<DataType> {
+    fn return_type(&self, _: &[DataType]) -> Result<DataType> {
        Ok(DataType::Utf8View)
    }

    fn signature(&self) -> &Signature {
-        &self.signature
+        &self.0.signature
    }

-    fn invoke_with_args(
-        &self,
-        args: ScalarFunctionArgs,
-    ) -> datafusion_common::Result<ColumnarValue> {
-        let [arg0, arg1] = extract_args(self.name(), &args)?;
-        let arg0 = compute::cast(&arg0, &DataType::BinaryView)?;
-        let jsons = arg0.as_binary_view();
-        let arg1 = compute::cast(&arg1, &DataType::Utf8View)?;
-        let paths = arg1.as_string_view();
+    fn invoke_with_args(&self, args: ScalarFunctionArgs) -> Result<ColumnarValue> {
+        struct StringResultBuilder(StringViewBuilder);

-        let size = jsons.len();
-        let mut builder = StringViewBuilder::with_capacity(size);
-
-        for i in 0..size {
-            let json = jsons.is_valid(i).then(|| jsons.value(i));
-            let path = paths.is_valid(i).then(|| paths.value(i));
-            let result = match (json, path) {
-                (Some(json), Some(path)) => {
-                    get_json_by_path(json, path).and_then(|json| jsonb::to_str(&json).ok())
+        impl JsonGetResultBuilder for StringResultBuilder {
+            fn append_value(&mut self, value: JsonResultValue<'_>) -> Result<()> {
+                match value {
+                    JsonResultValue::Jsonb(value) => {
+                        self.0.append_option(jsonb::to_str(&value).ok())
+                    }
+                    JsonResultValue::JsonStructByColumn(column, i) => {
+                        if let Some(v) = string_array_value_at_index(column, i) {
+                            self.0.append_value(v);
+                        } else {
+                            self.0
+                                .append_value(arrow_cast::display::array_value_to_string(
+                                    column, i,
+                                )?);
+                        }
+                    }
+                    JsonResultValue::JsonStructByValue(value) => {
+                        if let Some(s) = value.as_str() {
+                            self.0.append_value(s)
+                        } else {
+                            self.0.append_value(value.to_string())
+                        }
+                    }
                }
-                _ => None,
-            };
-            builder.append_option(result);
+                Ok(())
+            }
+
+            fn append_null(&mut self) {
+                self.0.append_null();
+            }
+
+            fn build(&mut self) -> ArrayRef {
+                Arc::new(self.0.finish())
+            }
        }

-        Ok(ColumnarValue::Array(Arc::new(builder.finish())))
+        self.0.invoke(args, |len: usize| {
+            StringResultBuilder(StringViewBuilder::with_capacity(len))
+        })
    }
 }

+#[derive(Default)]
+pub struct JsonGetInt(JsonGet);
+
+impl JsonGetInt {
+    pub const NAME: &'static str = "json_get_int";
+}
+
+impl Function for JsonGetInt {
+    fn name(&self) -> &str {
+        Self::NAME
+    }
+
+    fn return_type(&self, _: &[DataType]) -> Result<DataType> {
+        Ok(DataType::Int64)
+    }
+
+    fn signature(&self) -> &Signature {
+        &self.0.signature
+    }
+
+    fn invoke_with_args(&self, args: ScalarFunctionArgs) -> Result<ColumnarValue> {
+        struct IntResultBuilder(Int64Builder);
+
+        impl JsonGetResultBuilder for IntResultBuilder {
+            fn append_value(&mut self, value: JsonResultValue<'_>) -> Result<()> {
+                match value {
+                    JsonResultValue::Jsonb(value) => {
+                        self.0.append_option(jsonb::to_i64(&value).ok())
+                    }
+                    JsonResultValue::JsonStructByColumn(column, i) => {
+                        self.0.append_option(int_array_value_at_index(column, i))
+                    }
+                    JsonResultValue::JsonStructByValue(value) => {
+                        self.0.append_option(value.as_i64())
+                    }
+                }
+                Ok(())
+            }
+
+            fn append_null(&mut self) {
+                self.0.append_null();
+            }
+
+            fn build(&mut self) -> ArrayRef {
+                Arc::new(self.0.finish())
+            }
+        }
+
+        self.0.invoke(args, |len: usize| {
+            IntResultBuilder(Int64Builder::with_capacity(len))
+        })
+    }
+}
+
+impl Display for JsonGetInt {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "{}", Self::NAME.to_ascii_uppercase())
+    }
+}
+
+fn jsonb_get(
+    jsons: &BinaryViewArray,
+    paths: &StringViewArray,
+    builder: &mut impl JsonGetResultBuilder,
+) -> Result<()> {
+    let size = jsons.len();
+    for i in 0..size {
+        let json = jsons.is_valid(i).then(|| jsons.value(i));
+        let path = paths.is_valid(i).then(|| paths.value(i));
+        let result = match (json, path) {
+            (Some(json), Some(path)) => get_json_by_path(json, path),
+            _ => None,
+        };
+        if let Some(v) = result {
+            builder.append_value(JsonResultValue::Jsonb(v))?;
+        } else {
+            builder.append_null();
+        }
+    }
+    Ok(())
+}
+
+fn json_struct_get(
+    jsons: &StructArray,
+    paths: &StringViewArray,
+    builder: &mut impl JsonGetResultBuilder,
+) -> Result<()> {
+    let size = jsons.len();
+    for i in 0..size {
+        if jsons.is_null(i) || paths.is_null(i) {
+            builder.append_null();
+            continue;
+        }
+        let path = paths.value(i);
+
+        // naively assume the JSON path is our kind of indexing to the field, by removing its "root"
+        let field_path = path.trim().replace("$.", "");
+        let column = jsons.column_by_name(&field_path);
+
+        if let Some(column) = column {
+            builder.append_value(JsonResultValue::JsonStructByColumn(column, i))?;
+        } else {
+            let Some(raw) = jsons
+                .column_by_name(JsonStructureSettings::RAW_FIELD)
+                .and_then(|x| string_array_value_at_index(x, i))
+            else {
+                builder.append_null();
+                continue;
+            };
+
+            let path: JsonPath<Value> = JsonPath::try_from(path).map_err(|e| {
+                DataFusionError::Execution(format!("{path} is not a valid JSON path: {e}"))
+            })?;
+            // the wanted field is not retrievable from the JSON struct columns directly, we have
+            // to combine everything (columns and the "_raw") into a complete JSON value to find it
+            let value = json_struct_to_value(raw, jsons, i)?;
+
+            match path.find(&value) {
+                Value::Null => builder.append_null(),
+                Value::Array(values) => match values.as_slice() {
+                    [] => builder.append_null(),
+                    [x] => builder.append_value(JsonResultValue::JsonStructByValue(x))?,
+                    _ => builder.append_value(JsonResultValue::JsonStructByValue(&value))?,
+                },
+                value => builder.append_value(JsonResultValue::JsonStructByValue(&value))?,
+            }
+        }
+    }
+
+    Ok(())
+}
+
+fn json_struct_to_value(raw: &str, jsons: &StructArray, i: usize) -> Result<Value> {
+    let Ok(mut json) = Value::from_str(raw) else {
+        return Err(DataFusionError::Internal(format!(
+            "inner field '{}' is not a valid JSON string",
+            JsonStructureSettings::RAW_FIELD
+        )));
+    };
+
+    for (column_name, column) in jsons.column_names().into_iter().zip(jsons.columns()) {
+        if column_name == JsonStructureSettings::RAW_FIELD {
+            continue;
+        }
+
+        let (json_pointer, field) = if let Some((json_object, field)) = column_name.rsplit_once(".")
+        {
+            let json_pointer = format!("/{}", json_object.replace(".", "/"));
+            (json_pointer, field)
+        } else {
+            ("".to_string(), column_name)
+        };
+        let Some(json_object) = json
+            .pointer_mut(&json_pointer)
+            .and_then(|x| x.as_object_mut())
+        else {
+            return Err(DataFusionError::Internal(format!(
+                "value at JSON pointer '{}' is not an object",
+                json_pointer
+            )));
+        };
+
+        macro_rules! insert {
+            ($column: ident, $i: ident, $json_object: ident, $field: ident) => {{
+                if let Some(value) = $column
+                    .is_valid($i)
+                    .then(|| serde_json::Value::from($column.value($i)))
+                {
+                    $json_object.insert($field.to_string(), value);
+                }
+            }};
+        }
+
+        match column.data_type() {
+            // boolean => Value::Bool
+            DataType::Boolean => {
+                let column = column.as_boolean();
+                insert!(column, i, json_object, field);
+            }
+            // int => Value::Number
+            DataType::Int64 => {
+                let column = column.as_primitive::<Int64Type>();
+                insert!(column, i, json_object, field);
+            }
+            DataType::UInt64 => {
+                let column = column.as_primitive::<UInt64Type>();
+                insert!(column, i, json_object, field);
+            }
+            DataType::Float64 => {
+                let column = column.as_primitive::<Float64Type>();
+                insert!(column, i, json_object, field);
+            }
+            // string => Value::String
+            DataType::Utf8 => {
+                let column = column.as_string::<i32>();
+                insert!(column, i, json_object, field);
+            }
+            DataType::LargeUtf8 => {
+                let column = column.as_string::<i64>();
+                insert!(column, i, json_object, field);
+            }
+            DataType::Utf8View => {
+                let column = column.as_string_view();
+                insert!(column, i, json_object, field);
+            }
+            // other => Value::Array and Value::Object
+            _ => {
+                return Err(DataFusionError::NotImplemented(format!(
+                    "{} is not yet supported to be executed with field {} of datatype {}",
+                    JsonGetString::NAME,
+                    column_name,
+                    column.data_type()
+                )));
+            }
+        }
+    }
+    Ok(json)
+}
+
 impl Display for JsonGetString {
    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
        write!(f, "{}", Self::NAME.to_ascii_uppercase())
@@ -296,14 +582,60 @@ impl Display for JsonGetObject {
 mod tests {
    use std::sync::Arc;

+    use arrow::array::{Float64Array, Int64Array, StructArray};
    use arrow_schema::Field;
    use datafusion_common::ScalarValue;
    use datafusion_common::arrow::array::{BinaryArray, BinaryViewArray, StringArray};
    use datafusion_common::arrow::datatypes::{Float64Type, Int64Type};
    use datatypes::types::parse_string_to_jsonb;
+    use serde_json::json;

    use super::*;

+    /// Create a JSON object like this (as a one element struct array for testing):
+    ///
+    /// ```JSON
+    /// {
+    ///     "kind": "foo",
+    ///     "payload": {
+    ///         "code": 404,
+    ///         "success": false,
+    ///         "result": {
+    ///             "error": "not found",
+    ///             "time_cost": 1.234
+    ///         }
+    ///     }
+    /// }
+    /// ```
+    fn test_json_struct() -> ArrayRef {
+        Arc::new(StructArray::new(
+            vec![
+                Field::new("kind", DataType::Utf8, true),
+                Field::new("payload.code", DataType::Int64, true),
+                Field::new("payload.result.time_cost", DataType::Float64, true),
+                Field::new(JsonStructureSettings::RAW_FIELD, DataType::Utf8View, true),
+            ]
+            .into(),
+            vec![
+                Arc::new(StringArray::from_iter([Some("foo")])) as ArrayRef,
+                Arc::new(Int64Array::from_iter([Some(404)])),
+                Arc::new(Float64Array::from_iter([Some(1.234)])),
+                Arc::new(StringViewArray::from_iter([Some(
+                    json! ({
+                        "payload": {
+                            "success": false,
+                            "result": {
+                                "error": "not found"
+                            }
+                        }
+                    })
+                    .to_string(),
+                )])),
+            ],
+            None,
+        ))
+    }
+
    #[test]
    fn test_json_get_int() {
        let json_get_int = JsonGetInt::default();
@@ -321,37 +653,55 @@ mod tests {
            r#"{"a": 4, "b": {"c": 6}, "c": 6}"#,
            r#"{"a": 7, "b": 8, "c": {"a": 7}}"#,
        ];
-        let paths = vec!["$.a.b", "$.a", "$.c"];
-        let results = [Some(2), Some(4), None];
+        let json_struct = test_json_struct();

-        let jsonbs = json_strings
+        let path_expects = vec![
+            ("$.a.b", Some(2)),
+            ("$.a", Some(4)),
+            ("$.c", None),
+            ("$.kind", None),
+            ("$.payload.code", Some(404)),
+            ("$.payload.success", None),
+            ("$.payload.result.time_cost", None),
+            ("$.payload.not-exists", None),
+            ("$.not-exists", None),
+            ("$", None),
+        ];
+
+        let mut jsons = json_strings
            .iter()
            .map(|s| {
                let value = jsonb::parse_value(s.as_bytes()).unwrap();
-                value.to_vec()
+                Arc::new(BinaryArray::from_iter_values([value.to_vec()])) as ArrayRef
            })
            .collect::<Vec<_>>();
+        let json_struct_arrays =
+            std::iter::repeat_n(json_struct, path_expects.len() - jsons.len()).collect::<Vec<_>>();
+        jsons.extend(json_struct_arrays);

-        let args = ScalarFunctionArgs {
-            args: vec![
-                ColumnarValue::Array(Arc::new(BinaryArray::from_iter_values(jsonbs))),
-                ColumnarValue::Array(Arc::new(StringArray::from_iter_values(paths))),
-            ],
-            arg_fields: vec![],
-            number_rows: 3,
-            return_field: Arc::new(Field::new("x", DataType::Int64, false)),
-            config_options: Arc::new(Default::default()),
-        };
-        let result = json_get_int
-            .invoke_with_args(args)
-            .and_then(|x| x.to_array(3))
-            .unwrap();
-        let vector = result.as_primitive::<Int64Type>();
+        for i in 0..jsons.len() {
+            let json = &jsons[i];
+            let (path, expect) = path_expects[i];

-        assert_eq!(3, vector.len());
-        for (i, gt) in results.iter().enumerate() {
-            let result = vector.is_valid(i).then(|| vector.value(i));
-            assert_eq!(*gt, result);
+            let args = ScalarFunctionArgs {
+                args: vec![
+                    ColumnarValue::Array(json.clone()),
+                    ColumnarValue::Scalar(path.into()),
+                ],
+                arg_fields: vec![],
+                number_rows: 1,
+                return_field: Arc::new(Field::new("x", DataType::Int64, false)),
+                config_options: Arc::new(Default::default()),
+            };
+            let result = json_get_int
+                .invoke_with_args(args)
+                .and_then(|x| x.to_array(1))
+                .unwrap();
+
+            let result = result.as_primitive::<Int64Type>();
+            assert_eq!(1, result.len());
+            let actual = result.is_valid(0).then(|| result.value(0));
+            assert_eq!(actual, expect);
        }
    }

@@ -474,42 +824,85 @@ mod tests {
            r#"{"a": "d", "b": {"c": "e"}, "c": "f"}"#,
            r#"{"a": "g", "b": "h", "c": {"a": "g"}}"#,
        ];
-        let paths = vec!["$.a.b", "$.a", ""];
-        let results = [Some("a"), Some("d"), None];
+        let json_struct = test_json_struct();

-        let jsonbs = json_strings
+        let paths = vec![
+            "$.a.b",
+            "$.a",
+            "",
+            "$.kind",
+            "$.payload.code",
+            "$.payload.result.time_cost",
+            "$.payload",
+            "$.payload.success",
+            "$.payload.result",
+            "$.payload.result.error",
+            "$.payload.result.not-exists",
+            "$.payload.not-exists",
+            "$.not-exists",
+            "$",
+        ];
+        let expects = [
+            Some("a"),
+            Some("d"),
+            None,
+            Some("foo"),
+            Some("404"),
+            Some("1.234"),
+            Some(
+                r#"{"code":404,"result":{"error":"not found","time_cost":1.234},"success":false}"#,
+            ),
+            Some("false"),
+            Some(r#"{"error":"not found","time_cost":1.234}"#),
+            Some("not found"),
+            None,
+            None,
+            None,
+            Some(
+                r#"{"kind":"foo","payload":{"code":404,"result":{"error":"not found","time_cost":1.234},"success":false}}"#,
+            ),
+        ];
+
+        let mut jsons = json_strings
            .iter()
            .map(|s| {
                let value = jsonb::parse_value(s.as_bytes()).unwrap();
-                value.to_vec()
+                Arc::new(BinaryArray::from_iter_values([value.to_vec()])) as ArrayRef
            })
            .collect::<Vec<_>>();
+        let json_struct_arrays =
+            std::iter::repeat_n(json_struct, expects.len() - jsons.len()).collect::<Vec<_>>();
+        jsons.extend(json_struct_arrays);

-        let args = ScalarFunctionArgs {
-            args: vec![
-                ColumnarValue::Array(Arc::new(BinaryArray::from_iter_values(jsonbs))),
-                ColumnarValue::Array(Arc::new(StringArray::from_iter_values(paths))),
-            ],
-            arg_fields: vec![],
-            number_rows: 3,
-            return_field: Arc::new(Field::new("x", DataType::Utf8View, false)),
-            config_options: Arc::new(Default::default()),
-        };
-        let result = json_get_string
-            .invoke_with_args(args)
-            .and_then(|x| x.to_array(3))
-            .unwrap();
-        let vector = result.as_string_view();
+        for i in 0..jsons.len() {
+            let json = &jsons[i];
+            let path = paths[i];
+            let expect = expects[i];

-        assert_eq!(3, vector.len());
-        for (i, gt) in results.iter().enumerate() {
-            let result = vector.is_valid(i).then(|| vector.value(i));
-            assert_eq!(*gt, result);
+            let args = ScalarFunctionArgs {
+                args: vec![
+                    ColumnarValue::Array(json.clone()),
+                    ColumnarValue::Scalar(path.into()),
+                ],
+                arg_fields: vec![],
+                number_rows: 1,
+                return_field: Arc::new(Field::new("x", DataType::Utf8View, false)),
+                config_options: Arc::new(Default::default()),
+            };
+            let result = json_get_string
+                .invoke_with_args(args)
+                .and_then(|x| x.to_array(1))
+                .unwrap();
+
+            let result = result.as_string_view();
+            assert_eq!(1, result.len());
+            let actual = result.is_valid(0).then(|| result.value(0));
+            assert_eq!(actual, expect);
        }
    }

    #[test]
-    fn test_json_get_object() -> datafusion_common::Result<()> {
+    fn test_json_get_object() -> Result<()> {
        let udf = JsonGetObject::default();
        assert_eq!("json_get_object", udf.name());
        assert_eq!(
--- a/src/common/meta/src/error.rs
+++ b/src/common/meta/src/error.rs
@@ -224,6 +224,13 @@ pub enum Error {
        location: Location,
    },

+    #[snafu(display("Failed to find table repartition metadata for table id {}", table_id))]
+    TableRepartNotFound {
+        table_id: TableId,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
    #[snafu(display("Failed to decode protobuf"))]
    DecodeProto {
        #[snafu(implicit)]
@@ -1091,6 +1098,7 @@ impl ErrorExt for Error {
            | DecodeProto { .. }
            | BuildTableMeta { .. }
            | TableRouteNotFound { .. }
+            | TableRepartNotFound { .. }
            | ConvertRawTableInfo { .. }
            | RegionOperatingRace { .. }
            | EncodeWalOptions { .. }
--- a/src/common/meta/src/key.rs
+++ b/src/common/meta/src/key.rs
@@ -106,6 +106,7 @@ mod schema_metadata_manager;
 pub mod schema_name;
 pub mod table_info;
 pub mod table_name;
+pub mod table_repart;
 pub mod table_route;
 #[cfg(any(test, feature = "testing"))]
 pub mod test_utils;
@@ -156,6 +157,7 @@ use crate::DatanodeId;
 use crate::error::{self, Result, SerdeJsonSnafu};
 use crate::key::flow::flow_state::FlowStateValue;
 use crate::key::node_address::NodeAddressValue;
+use crate::key::table_repart::{TableRepartKey, TableRepartManager};
 use crate::key::table_route::TableRouteKey;
 use crate::key::topic_region::TopicRegionValue;
 use crate::key::txn_helper::TxnOpGetResponseSet;
@@ -178,6 +180,7 @@ pub const TABLE_NAME_KEY_PREFIX: &str = "__table_name";
 pub const CATALOG_NAME_KEY_PREFIX: &str = "__catalog_name";
 pub const SCHEMA_NAME_KEY_PREFIX: &str = "__schema_name";
 pub const TABLE_ROUTE_PREFIX: &str = "__table_route";
+pub const TABLE_REPART_PREFIX: &str = "__table_repart";
 pub const NODE_ADDRESS_PREFIX: &str = "__node_address";
 pub const KAFKA_TOPIC_KEY_PREFIX: &str = "__topic_name/kafka";
 // The legacy topic key prefix is used to store the topic name in previous versions.
@@ -288,6 +291,11 @@ lazy_static! {
        Regex::new(&format!("^{TABLE_ROUTE_PREFIX}/([0-9]+)$")).unwrap();
 }

+lazy_static! {
+    pub(crate) static ref TABLE_REPART_KEY_PATTERN: Regex =
+        Regex::new(&format!("^{TABLE_REPART_PREFIX}/([0-9]+)$")).unwrap();
+}
+
 lazy_static! {
    static ref DATANODE_TABLE_KEY_PATTERN: Regex =
        Regex::new(&format!("^{DATANODE_TABLE_KEY_PREFIX}/([0-9]+)/([0-9]+)$")).unwrap();
@@ -386,6 +394,7 @@ pub struct TableMetadataManager {
    catalog_manager: CatalogManager,
    schema_manager: SchemaManager,
    table_route_manager: TableRouteManager,
+    table_repart_manager: TableRepartManager,
    tombstone_manager: TombstoneManager,
    topic_name_manager: TopicNameManager,
    topic_region_manager: TopicRegionManager,
@@ -538,6 +547,7 @@ impl TableMetadataManager {
            catalog_manager: CatalogManager::new(kv_backend.clone()),
            schema_manager: SchemaManager::new(kv_backend.clone()),
            table_route_manager: TableRouteManager::new(kv_backend.clone()),
+            table_repart_manager: TableRepartManager::new(kv_backend.clone()),
            tombstone_manager: TombstoneManager::new(kv_backend.clone()),
            topic_name_manager: TopicNameManager::new(kv_backend.clone()),
            topic_region_manager: TopicRegionManager::new(kv_backend.clone()),
@@ -558,6 +568,7 @@ impl TableMetadataManager {
            catalog_manager: CatalogManager::new(kv_backend.clone()),
            schema_manager: SchemaManager::new(kv_backend.clone()),
            table_route_manager: TableRouteManager::new(kv_backend.clone()),
+            table_repart_manager: TableRepartManager::new(kv_backend.clone()),
            tombstone_manager: TombstoneManager::new_with_prefix(
                kv_backend.clone(),
                tombstone_prefix,
@@ -616,6 +627,10 @@ impl TableMetadataManager {
        &self.table_route_manager
    }

+    pub fn table_repart_manager(&self) -> &TableRepartManager {
+        &self.table_repart_manager
+    }
+
    pub fn topic_name_manager(&self) -> &TopicNameManager {
        &self.topic_name_manager
    }
@@ -923,6 +938,7 @@ impl TableMetadataManager {
        );
        let table_info_key = TableInfoKey::new(table_id);
        let table_route_key = TableRouteKey::new(table_id);
+        let table_repart_key = TableRepartKey::new(table_id);
        let datanode_table_keys = datanode_ids
            .into_iter()
            .map(|datanode_id| DatanodeTableKey::new(datanode_id, table_id))
@@ -937,6 +953,7 @@ impl TableMetadataManager {
        keys.push(table_name.to_bytes());
        keys.push(table_info_key.to_bytes());
        keys.push(table_route_key.to_bytes());
+        keys.push(table_repart_key.to_bytes());
        for key in &datanode_table_keys {
            keys.push(key.to_bytes());
        }
--- a/src/common/meta/src/key/table_repart.rs
+++ b/src/common/meta/src/key/table_repart.rs
@@ -0,0 +1,856 @@
+// Copyright 2023 Greptime Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::{BTreeMap, BTreeSet, HashMap};
+use std::fmt::Display;
+
+use serde::{Deserialize, Serialize};
+use snafu::{OptionExt as _, ResultExt, ensure};
+use store_api::storage::RegionId;
+use table::metadata::TableId;
+
+use crate::error::{InvalidMetadataSnafu, Result, SerdeJsonSnafu};
+use crate::key::txn_helper::TxnOpGetResponseSet;
+use crate::key::{
+    DeserializedValueWithBytes, MetadataKey, MetadataValue, TABLE_REPART_KEY_PATTERN,
+    TABLE_REPART_PREFIX,
+};
+use crate::kv_backend::KvBackendRef;
+use crate::kv_backend::txn::Txn;
+use crate::rpc::store::BatchGetRequest;
+
+/// The key stores table repartition metadata.
+/// Specifically, it records the relation between source and destination regions after a repartition operation is completed.
+/// This is distinct from the initial partitioning scheme of the table.
+/// For example, after repartition, a destination region may still hold files from a source region; this mapping should be updated once repartition is done.
+/// The GC scheduler uses this information to clean up those files (and removes this mapping if all files from the source region are cleaned).
+///
+/// The layout: `__table_repart/{table_id}`.
+#[derive(Debug, PartialEq)]
+pub struct TableRepartKey {
+    /// The unique identifier of the table whose re-partition information is stored in this key.
+    pub table_id: TableId,
+}
+
+impl TableRepartKey {
+    pub fn new(table_id: TableId) -> Self {
+        Self { table_id }
+    }
+
+    /// Returns the range prefix of the table repartition key.
+    pub fn range_prefix() -> Vec<u8> {
+        format!("{}/", TABLE_REPART_PREFIX).into_bytes()
+    }
+}
+
+impl MetadataKey<'_, TableRepartKey> for TableRepartKey {
+    fn to_bytes(&self) -> Vec<u8> {
+        self.to_string().into_bytes()
+    }
+
+    fn from_bytes(bytes: &[u8]) -> Result<TableRepartKey> {
+        let key = std::str::from_utf8(bytes).map_err(|e| {
+            InvalidMetadataSnafu {
+                err_msg: format!(
+                    "TableRepartKey '{}' is not a valid UTF8 string: {e}",
+                    String::from_utf8_lossy(bytes)
+                ),
+            }
+            .build()
+        })?;
+        let captures = TABLE_REPART_KEY_PATTERN
+            .captures(key)
+            .context(InvalidMetadataSnafu {
+                err_msg: format!("Invalid TableRepartKey '{key}'"),
+            })?;
+        // Safety: pass the regex check above
+        let table_id = captures[1].parse::<TableId>().unwrap();
+        Ok(TableRepartKey { table_id })
+    }
+}
+
+impl Display for TableRepartKey {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}/{}", TABLE_REPART_PREFIX, self.table_id)
+    }
+}
+
+#[derive(Debug, PartialEq, Serialize, Deserialize, Clone, Default)]
+pub struct TableRepartValue {
+    /// A mapping from source region IDs to sets of destination region IDs after repartition.
+    ///
+    /// Each key in the map is a `RegionId` representing a source region that has been repartitioned.
+    /// The corresponding value is a `BTreeSet<RegionId>` containing the IDs of destination regions
+    /// that currently hold files originally from the source region. This mapping is updated after
+    /// repartition and is used by the GC scheduler to track and clean up files that have been moved.
+    pub src_to_dst: BTreeMap<RegionId, BTreeSet<RegionId>>,
+}
+
+impl TableRepartValue {
+    /// Creates a new TableRepartValue with an empty src_to_dst map.
+    pub fn new() -> Self {
+        Default::default()
+    }
+    /// Update mapping from src region to dst regions. Should be called once repartition is done.
+    ///
+    /// If `dst` is empty, this method does nothing.
+    pub fn update_mappings(&mut self, src: RegionId, dst: &[RegionId]) {
+        if dst.is_empty() {
+            return;
+        }
+        self.src_to_dst.entry(src).or_default().extend(dst);
+    }
+
+    /// Remove mappings from src region to dst regions. Should be called once files from src region are cleaned up in dst regions.
+    pub fn remove_mappings(&mut self, src: RegionId, dsts: &[RegionId]) {
+        if let Some(dst_set) = self.src_to_dst.get_mut(&src) {
+            for dst in dsts {
+                dst_set.remove(dst);
+            }
+            if dst_set.is_empty() {
+                self.src_to_dst.remove(&src);
+            }
+        }
+    }
+}
+
+impl MetadataValue for TableRepartValue {
+    fn try_from_raw_value(raw_value: &[u8]) -> Result<Self> {
+        serde_json::from_slice::<TableRepartValue>(raw_value).context(SerdeJsonSnafu)
+    }
+
+    fn try_as_raw_value(&self) -> Result<Vec<u8>> {
+        serde_json::to_vec(self).context(SerdeJsonSnafu)
+    }
+}
+
+pub type TableRepartValueDecodeResult =
+    Result<Option<DeserializedValueWithBytes<TableRepartValue>>>;
+
+pub struct TableRepartManager {
+    kv_backend: KvBackendRef,
+}
+
+impl TableRepartManager {
+    pub fn new(kv_backend: KvBackendRef) -> Self {
+        Self { kv_backend }
+    }
+
+    /// Builds a create table repart transaction,
+    /// it expected the `__table_repart/{table_id}` wasn't occupied.
+    pub fn build_create_txn(
+        &self,
+        table_id: TableId,
+        table_repart_value: &TableRepartValue,
+    ) -> Result<(
+        Txn,
+        impl FnOnce(&mut TxnOpGetResponseSet) -> TableRepartValueDecodeResult + use<>,
+    )> {
+        let key = TableRepartKey::new(table_id);
+        let raw_key = key.to_bytes();
+
+        let txn = Txn::put_if_not_exists(raw_key.clone(), table_repart_value.try_as_raw_value()?);
+
+        Ok((
+            txn,
+            TxnOpGetResponseSet::decode_with(TxnOpGetResponseSet::filter(raw_key)),
+        ))
+    }
+
+    /// Builds a update table repart transaction,
+    /// it expected the remote value equals the `current_table_repart_value`.
+    /// It retrieves the latest value if the comparing failed.
+    pub fn build_update_txn(
+        &self,
+        table_id: TableId,
+        current_table_repart_value: &DeserializedValueWithBytes<TableRepartValue>,
+        new_table_repart_value: &TableRepartValue,
+    ) -> Result<(
+        Txn,
+        impl FnOnce(&mut TxnOpGetResponseSet) -> TableRepartValueDecodeResult + use<>,
+    )> {
+        let key = TableRepartKey::new(table_id);
+        let raw_key = key.to_bytes();
+        let raw_value = current_table_repart_value.get_raw_bytes();
+        let new_raw_value: Vec<u8> = new_table_repart_value.try_as_raw_value()?;
+
+        let txn = Txn::compare_and_put(raw_key.clone(), raw_value, new_raw_value);
+
+        Ok((
+            txn,
+            TxnOpGetResponseSet::decode_with(TxnOpGetResponseSet::filter(raw_key)),
+        ))
+    }
+
+    /// Returns the [`TableRepartValue`].
+    pub async fn get(&self, table_id: TableId) -> Result<Option<TableRepartValue>> {
+        self.get_inner(table_id).await
+    }
+
+    async fn get_inner(&self, table_id: TableId) -> Result<Option<TableRepartValue>> {
+        let key = TableRepartKey::new(table_id);
+        self.kv_backend
+            .get(&key.to_bytes())
+            .await?
+            .map(|kv| TableRepartValue::try_from_raw_value(&kv.value))
+            .transpose()
+    }
+
+    /// Returns the [`TableRepartValue`] wrapped with [`DeserializedValueWithBytes`].
+    pub async fn get_with_raw_bytes(
+        &self,
+        table_id: TableId,
+    ) -> Result<Option<DeserializedValueWithBytes<TableRepartValue>>> {
+        self.get_with_raw_bytes_inner(table_id).await
+    }
+
+    async fn get_with_raw_bytes_inner(
+        &self,
+        table_id: TableId,
+    ) -> Result<Option<DeserializedValueWithBytes<TableRepartValue>>> {
+        let key = TableRepartKey::new(table_id);
+        self.kv_backend
+            .get(&key.to_bytes())
+            .await?
+            .map(|kv| DeserializedValueWithBytes::from_inner_slice(&kv.value))
+            .transpose()
+    }
+
+    /// Returns batch of [`TableRepartValue`] that respects the order of `table_ids`.
+    pub async fn batch_get(&self, table_ids: &[TableId]) -> Result<Vec<Option<TableRepartValue>>> {
+        let raw_table_reparts = self.batch_get_inner(table_ids).await?;
+
+        Ok(raw_table_reparts
+            .into_iter()
+            .map(|v| v.map(|x| x.inner))
+            .collect())
+    }
+
+    /// Returns batch of [`TableRepartValue`] wrapped with [`DeserializedValueWithBytes`].
+    pub async fn batch_get_with_raw_bytes(
+        &self,
+        table_ids: &[TableId],
+    ) -> Result<Vec<Option<DeserializedValueWithBytes<TableRepartValue>>>> {
+        self.batch_get_inner(table_ids).await
+    }
+
+    async fn batch_get_inner(
+        &self,
+        table_ids: &[TableId],
+    ) -> Result<Vec<Option<DeserializedValueWithBytes<TableRepartValue>>>> {
+        let keys = table_ids
+            .iter()
+            .map(|id| TableRepartKey::new(*id).to_bytes())
+            .collect::<Vec<_>>();
+        let resp = self
+            .kv_backend
+            .batch_get(BatchGetRequest { keys: keys.clone() })
+            .await?;
+
+        let kvs = resp
+            .kvs
+            .into_iter()
+            .map(|kv| (kv.key, kv.value))
+            .collect::<HashMap<_, _>>();
+        keys.into_iter()
+            .map(|key| {
+                if let Some(value) = kvs.get(&key) {
+                    Ok(Some(DeserializedValueWithBytes::from_inner_slice(value)?))
+                } else {
+                    Ok(None)
+                }
+            })
+            .collect()
+    }
+
+    /// Updates mappings from src region to dst regions.
+    /// Should be called once repartition is done.
+    pub async fn update_mappings(&self, src: RegionId, dst: &[RegionId]) -> Result<()> {
+        let table_id = src.table_id();
+
+        // Get current table repart with raw bytes for CAS operation
+        let current_table_repart = self
+            .get_with_raw_bytes(table_id)
+            .await?
+            .context(crate::error::TableRepartNotFoundSnafu { table_id })?;
+
+        // Clone the current repart value and update mappings
+        let mut new_table_repart_value = current_table_repart.inner.clone();
+        new_table_repart_value.update_mappings(src, dst);
+
+        // Execute atomic update
+        let (txn, _) =
+            self.build_update_txn(table_id, &current_table_repart, &new_table_repart_value)?;
+
+        let result = self.kv_backend.txn(txn).await?;
+
+        ensure!(
+            result.succeeded,
+            crate::error::MetadataCorruptionSnafu {
+                err_msg: format!(
+                    "Failed to update mappings for table {}: CAS operation failed",
+                    table_id
+                ),
+            }
+        );
+
+        Ok(())
+    }
+
+    /// Removes mappings from src region to dst regions.
+    /// Should be called once files from src region are cleaned up in dst regions.
+    pub async fn remove_mappings(&self, src: RegionId, dsts: &[RegionId]) -> Result<()> {
+        let table_id = src.table_id();
+
+        // Get current table repart with raw bytes for CAS operation
+        let current_table_repart = self
+            .get_with_raw_bytes(table_id)
+            .await?
+            .context(crate::error::TableRepartNotFoundSnafu { table_id })?;
+
+        // Clone the current repart value and remove mappings
+        let mut new_table_repart_value = current_table_repart.inner.clone();
+        new_table_repart_value.remove_mappings(src, dsts);
+
+        // Execute atomic update
+        let (txn, _) =
+            self.build_update_txn(table_id, &current_table_repart, &new_table_repart_value)?;
+
+        let result = self.kv_backend.txn(txn).await?;
+
+        ensure!(
+            result.succeeded,
+            crate::error::MetadataCorruptionSnafu {
+                err_msg: format!(
+                    "Failed to remove mappings for table {}: CAS operation failed",
+                    table_id
+                ),
+            }
+        );
+
+        Ok(())
+    }
+
+    /// Returns the destination regions for a given source region.
+    pub async fn get_dst_regions(
+        &self,
+        src_region: RegionId,
+    ) -> Result<Option<BTreeSet<RegionId>>> {
+        let table_id = src_region.table_id();
+        let table_repart = self.get(table_id).await?;
+        Ok(table_repart.and_then(|repart| repart.src_to_dst.get(&src_region).cloned()))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::BTreeMap;
+    use std::sync::Arc;
+
+    use super::*;
+    use crate::kv_backend::TxnService;
+    use crate::kv_backend::memory::MemoryKvBackend;
+
+    #[test]
+    fn test_table_repart_key_serialization() {
+        let key = TableRepartKey::new(42);
+        let raw_key = key.to_bytes();
+        assert_eq!(raw_key, b"__table_repart/42");
+    }
+
+    #[test]
+    fn test_table_repart_key_deserialization() {
+        let expected = TableRepartKey::new(42);
+        let key = TableRepartKey::from_bytes(b"__table_repart/42").unwrap();
+        assert_eq!(key, expected);
+    }
+
+    #[test]
+    fn test_table_repart_key_deserialization_invalid_utf8() {
+        let result = TableRepartKey::from_bytes(b"__table_repart/\xff");
+        assert!(result.is_err());
+        assert!(
+            result
+                .unwrap_err()
+                .to_string()
+                .contains("not a valid UTF8 string")
+        );
+    }
+
+    #[test]
+    fn test_table_repart_key_deserialization_invalid_format() {
+        let result = TableRepartKey::from_bytes(b"invalid_key_format");
+        assert!(result.is_err());
+        assert!(
+            result
+                .unwrap_err()
+                .to_string()
+                .contains("Invalid TableRepartKey")
+        );
+    }
+
+    #[test]
+    fn test_table_repart_value_serialization_deserialization() {
+        let mut src_to_dst = BTreeMap::new();
+        let src_region = RegionId::new(1, 1);
+        let dst_regions = vec![RegionId::new(1, 2), RegionId::new(1, 3)];
+        src_to_dst.insert(src_region, dst_regions.into_iter().collect());
+
+        let value = TableRepartValue { src_to_dst };
+        let serialized = value.try_as_raw_value().unwrap();
+        let deserialized = TableRepartValue::try_from_raw_value(&serialized).unwrap();
+
+        assert_eq!(value, deserialized);
+    }
+
+    #[test]
+    fn test_table_repart_value_update_mappings_new_src() {
+        let mut value = TableRepartValue {
+            src_to_dst: BTreeMap::new(),
+        };
+
+        let src = RegionId::new(1, 1);
+        let dst = vec![RegionId::new(1, 2), RegionId::new(1, 3)];
+
+        value.update_mappings(src, &dst);
+
+        assert_eq!(value.src_to_dst.len(), 1);
+        assert!(value.src_to_dst.contains_key(&src));
+        assert_eq!(value.src_to_dst.get(&src).unwrap().len(), 2);
+        assert!(
+            value
+                .src_to_dst
+                .get(&src)
+                .unwrap()
+                .contains(&RegionId::new(1, 2))
+        );
+        assert!(
+            value
+                .src_to_dst
+                .get(&src)
+                .unwrap()
+                .contains(&RegionId::new(1, 3))
+        );
+    }
+
+    #[test]
+    fn test_table_repart_value_update_mappings_existing_src() {
+        let mut value = TableRepartValue {
+            src_to_dst: BTreeMap::new(),
+        };
+
+        let src = RegionId::new(1, 1);
+        let initial_dst = vec![RegionId::new(1, 2)];
+        let additional_dst = vec![RegionId::new(1, 3), RegionId::new(1, 4)];
+
+        // Initial mapping
+        value.update_mappings(src, &initial_dst);
+        // Update with additional destinations
+        value.update_mappings(src, &additional_dst);
+
+        assert_eq!(value.src_to_dst.len(), 1);
+        assert_eq!(value.src_to_dst.get(&src).unwrap().len(), 3);
+        assert!(
+            value
+                .src_to_dst
+                .get(&src)
+                .unwrap()
+                .contains(&RegionId::new(1, 2))
+        );
+        assert!(
+            value
+                .src_to_dst
+                .get(&src)
+                .unwrap()
+                .contains(&RegionId::new(1, 3))
+        );
+        assert!(
+            value
+                .src_to_dst
+                .get(&src)
+                .unwrap()
+                .contains(&RegionId::new(1, 4))
+        );
+    }
+
+    #[test]
+    fn test_table_repart_value_remove_mappings_existing() {
+        let mut value = TableRepartValue {
+            src_to_dst: BTreeMap::new(),
+        };
+
+        let src = RegionId::new(1, 1);
+        let dst_regions = vec![
+            RegionId::new(1, 2),
+            RegionId::new(1, 3),
+            RegionId::new(1, 4),
+        ];
+        value.update_mappings(src, &dst_regions);
+
+        // Remove some mappings
+        let to_remove = vec![RegionId::new(1, 2), RegionId::new(1, 3)];
+        value.remove_mappings(src, &to_remove);
+
+        assert_eq!(value.src_to_dst.len(), 1);
+        assert_eq!(value.src_to_dst.get(&src).unwrap().len(), 1);
+        assert!(
+            value
+                .src_to_dst
+                .get(&src)
+                .unwrap()
+                .contains(&RegionId::new(1, 4))
+        );
+    }
+
+    #[test]
+    fn test_table_repart_value_remove_mappings_all() {
+        let mut value = TableRepartValue {
+            src_to_dst: BTreeMap::new(),
+        };
+
+        let src = RegionId::new(1, 1);
+        let dst_regions = vec![RegionId::new(1, 2), RegionId::new(1, 3)];
+        value.update_mappings(src, &dst_regions);
+
+        // Remove all mappings
+        value.remove_mappings(src, &dst_regions);
+
+        assert_eq!(value.src_to_dst.len(), 0);
+    }
+
+    #[test]
+    fn test_table_repart_value_remove_mappings_nonexistent() {
+        let mut value = TableRepartValue {
+            src_to_dst: BTreeMap::new(),
+        };
+
+        let src = RegionId::new(1, 1);
+        let dst_regions = vec![RegionId::new(1, 2)];
+        value.update_mappings(src, &dst_regions);
+
+        // Try to remove non-existent mappings
+        let nonexistent_dst = vec![RegionId::new(1, 3), RegionId::new(1, 4)];
+        value.remove_mappings(src, &nonexistent_dst);
+
+        // Should remain unchanged
+        assert_eq!(value.src_to_dst.len(), 1);
+        assert_eq!(value.src_to_dst.get(&src).unwrap().len(), 1);
+        assert!(
+            value
+                .src_to_dst
+                .get(&src)
+                .unwrap()
+                .contains(&RegionId::new(1, 2))
+        );
+    }
+
+    #[test]
+    fn test_table_repart_value_remove_mappings_nonexistent_src() {
+        let mut value = TableRepartValue {
+            src_to_dst: BTreeMap::new(),
+        };
+
+        let src = RegionId::new(1, 1);
+        let dst_regions = vec![RegionId::new(1, 2)];
+
+        // Try to remove mappings for non-existent source
+        value.remove_mappings(src, &dst_regions);
+
+        // Should remain empty
+        assert_eq!(value.src_to_dst.len(), 0);
+    }
+
+    #[tokio::test]
+    async fn test_table_repart_manager_get_empty() {
+        let kv = Arc::new(MemoryKvBackend::default());
+        let manager = TableRepartManager::new(kv);
+        let result = manager.get(1024).await.unwrap();
+        assert!(result.is_none());
+    }
+
+    #[tokio::test]
+    async fn test_table_repart_manager_get_with_raw_bytes_empty() {
+        let kv = Arc::new(MemoryKvBackend::default());
+        let manager = TableRepartManager::new(kv);
+        let result = manager.get_with_raw_bytes(1024).await.unwrap();
+        assert!(result.is_none());
+    }
+
+    #[tokio::test]
+    async fn test_table_repart_manager_create_and_get() {
+        let kv = Arc::new(MemoryKvBackend::default());
+        let manager = TableRepartManager::new(kv.clone());
+
+        let mut src_to_dst = BTreeMap::new();
+        let src_region = RegionId::new(1, 1);
+        let dst_regions = vec![RegionId::new(1, 2), RegionId::new(1, 3)];
+        src_to_dst.insert(src_region, dst_regions.into_iter().collect());
+
+        let value = TableRepartValue { src_to_dst };
+
+        // Create the table repart
+        let (txn, _) = manager.build_create_txn(1024, &value).unwrap();
+        let result = kv.txn(txn).await.unwrap();
+        assert!(result.succeeded);
+
+        // Get the table repart
+        let retrieved = manager.get(1024).await.unwrap().unwrap();
+        assert_eq!(retrieved, value);
+    }
+
+    #[tokio::test]
+    async fn test_table_repart_manager_update_txn() {
+        let kv = Arc::new(MemoryKvBackend::default());
+        let manager = TableRepartManager::new(kv.clone());
+
+        let initial_value = TableRepartValue {
+            src_to_dst: BTreeMap::new(),
+        };
+
+        // Create initial table repart
+        let (create_txn, _) = manager.build_create_txn(1024, &initial_value).unwrap();
+        let result = kv.txn(create_txn).await.unwrap();
+        assert!(result.succeeded);
+
+        // Get current value with raw bytes
+        let current_value = manager.get_with_raw_bytes(1024).await.unwrap().unwrap();
+
+        // Create updated value
+        let mut updated_src_to_dst = BTreeMap::new();
+        let src_region = RegionId::new(1, 1);
+        let dst_regions = vec![RegionId::new(1, 2)];
+        updated_src_to_dst.insert(src_region, dst_regions.into_iter().collect());
+        let updated_value = TableRepartValue {
+            src_to_dst: updated_src_to_dst,
+        };
+
+        // Build update transaction
+        let (update_txn, _) = manager
+            .build_update_txn(1024, &current_value, &updated_value)
+            .unwrap();
+        let result = kv.txn(update_txn).await.unwrap();
+        assert!(result.succeeded);
+
+        // Verify update
+        let retrieved = manager.get(1024).await.unwrap().unwrap();
+        assert_eq!(retrieved, updated_value);
+    }
+
+    #[tokio::test]
+    async fn test_table_repart_manager_batch_get() {
+        let kv = Arc::new(MemoryKvBackend::default());
+        let manager = TableRepartManager::new(kv.clone());
+
+        // Create multiple table reparts
+        let table_reparts = vec![
+            (
+                1024,
+                TableRepartValue {
+                    src_to_dst: {
+                        let mut map = BTreeMap::new();
+                        map.insert(
+                            RegionId::new(1, 1),
+                            vec![RegionId::new(1, 2)].into_iter().collect(),
+                        );
+                        map
+                    },
+                },
+            ),
+            (
+                1025,
+                TableRepartValue {
+                    src_to_dst: {
+                        let mut map = BTreeMap::new();
+                        map.insert(
+                            RegionId::new(2, 1),
+                            vec![RegionId::new(2, 2), RegionId::new(2, 3)]
+                                .into_iter()
+                                .collect(),
+                        );
+                        map
+                    },
+                },
+            ),
+        ];
+
+        for (table_id, value) in &table_reparts {
+            let (txn, _) = manager.build_create_txn(*table_id, value).unwrap();
+            let result = kv.txn(txn).await.unwrap();
+            assert!(result.succeeded);
+        }
+
+        // Batch get
+        let results = manager.batch_get(&[1024, 1025, 1026]).await.unwrap();
+        assert_eq!(results.len(), 3);
+        assert_eq!(results[0].as_ref().unwrap(), &table_reparts[0].1);
+        assert_eq!(results[1].as_ref().unwrap(), &table_reparts[1].1);
+        assert!(results[2].is_none());
+    }
+
+    #[tokio::test]
+    async fn test_table_repart_manager_update_mappings() {
+        let kv = Arc::new(MemoryKvBackend::default());
+        let manager = TableRepartManager::new(kv.clone());
+
+        // Create initial table repart
+        let initial_value = TableRepartValue {
+            src_to_dst: BTreeMap::new(),
+        };
+        let (txn, _) = manager.build_create_txn(1024, &initial_value).unwrap();
+        let result = kv.txn(txn).await.unwrap();
+        assert!(result.succeeded);
+
+        // Update mappings
+        let src = RegionId::new(1024, 1);
+        let dst = vec![RegionId::new(1024, 2), RegionId::new(1024, 3)];
+        manager.update_mappings(src, &dst).await.unwrap();
+
+        // Verify update
+        let retrieved = manager.get(1024).await.unwrap().unwrap();
+        assert_eq!(retrieved.src_to_dst.len(), 1);
+        assert!(retrieved.src_to_dst.contains_key(&src));
+        assert_eq!(retrieved.src_to_dst.get(&src).unwrap().len(), 2);
+    }
+
+    #[tokio::test]
+    async fn test_table_repart_manager_remove_mappings() {
+        let kv = Arc::new(MemoryKvBackend::default());
+        let manager = TableRepartManager::new(kv.clone());
+
+        // Create initial table repart with mappings
+        let mut initial_src_to_dst = BTreeMap::new();
+        let src = RegionId::new(1024, 1);
+        let dst_regions = vec![
+            RegionId::new(1024, 2),
+            RegionId::new(1024, 3),
+            RegionId::new(1024, 4),
+        ];
+        initial_src_to_dst.insert(src, dst_regions.into_iter().collect());
+
+        let initial_value = TableRepartValue {
+            src_to_dst: initial_src_to_dst,
+        };
+        let (txn, _) = manager.build_create_txn(1024, &initial_value).unwrap();
+        let result = kv.txn(txn).await.unwrap();
+        assert!(result.succeeded);
+
+        // Remove some mappings
+        let to_remove = vec![RegionId::new(1024, 2), RegionId::new(1024, 3)];
+        manager.remove_mappings(src, &to_remove).await.unwrap();
+
+        // Verify removal
+        let retrieved = manager.get(1024).await.unwrap().unwrap();
+        assert_eq!(retrieved.src_to_dst.len(), 1);
+        assert_eq!(retrieved.src_to_dst.get(&src).unwrap().len(), 1);
+        assert!(
+            retrieved
+                .src_to_dst
+                .get(&src)
+                .unwrap()
+                .contains(&RegionId::new(1024, 4))
+        );
+    }
+
+    #[tokio::test]
+    async fn test_table_repart_manager_get_dst_regions() {
+        let kv = Arc::new(MemoryKvBackend::default());
+        let manager = TableRepartManager::new(kv.clone());
+
+        // Create initial table repart with mappings
+        let mut initial_src_to_dst = BTreeMap::new();
+        let src = RegionId::new(1024, 1);
+        let dst_regions = vec![RegionId::new(1024, 2), RegionId::new(1024, 3)];
+        initial_src_to_dst.insert(src, dst_regions.into_iter().collect());
+
+        let initial_value = TableRepartValue {
+            src_to_dst: initial_src_to_dst,
+        };
+        let (txn, _) = manager.build_create_txn(1024, &initial_value).unwrap();
+        let result = kv.txn(txn).await.unwrap();
+        assert!(result.succeeded);
+
+        // Get destination regions
+        let dst_regions = manager.get_dst_regions(src).await.unwrap();
+        assert!(dst_regions.is_some());
+        let dst_set = dst_regions.unwrap();
+        assert_eq!(dst_set.len(), 2);
+        assert!(dst_set.contains(&RegionId::new(1024, 2)));
+        assert!(dst_set.contains(&RegionId::new(1024, 3)));
+
+        // Test non-existent source region
+        let nonexistent_src = RegionId::new(1024, 99);
+        let result = manager.get_dst_regions(nonexistent_src).await.unwrap();
+        assert!(result.is_none());
+    }
+
+    #[tokio::test]
+    async fn test_table_repart_manager_operations_on_nonexistent_table() {
+        let kv = Arc::new(MemoryKvBackend::default());
+        let manager = TableRepartManager::new(kv);
+
+        let src = RegionId::new(1024, 1);
+        let dst = vec![RegionId::new(1024, 2)];
+
+        // Try to update mappings on non-existent table
+        let result = manager.update_mappings(src, &dst).await;
+        assert!(result.is_err());
+        let err_msg = result.unwrap_err().to_string();
+        assert!(
+            err_msg.contains("Failed to find table repartition metadata for table id 1024"),
+            "{err_msg}"
+        );
+
+        // Try to remove mappings on non-existent table
+        let result = manager.remove_mappings(src, &dst).await;
+        assert!(result.is_err());
+        let err_msg = result.unwrap_err().to_string();
+        assert!(
+            err_msg.contains("Failed to find table repartition metadata for table id 1024"),
+            "{err_msg}"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_table_repart_manager_batch_get_with_raw_bytes() {
+        let kv = Arc::new(MemoryKvBackend::default());
+        let manager = TableRepartManager::new(kv.clone());
+
+        // Create table repart
+        let value = TableRepartValue {
+            src_to_dst: {
+                let mut map = BTreeMap::new();
+                map.insert(
+                    RegionId::new(1, 1),
+                    vec![RegionId::new(1, 2)].into_iter().collect(),
+                );
+                map
+            },
+        };
+        let (txn, _) = manager.build_create_txn(1024, &value).unwrap();
+        let result = kv.txn(txn).await.unwrap();
+        assert!(result.succeeded);
+
+        // Batch get with raw bytes
+        let results = manager
+            .batch_get_with_raw_bytes(&[1024, 1025])
+            .await
+            .unwrap();
+        assert_eq!(results.len(), 2);
+        assert!(results[0].is_some());
+        assert!(results[1].is_none());
+
+        let retrieved = &results[0].as_ref().unwrap().inner;
+        assert_eq!(retrieved, &value);
+    }
+}
--- a/src/common/meta/src/kv_backend/rds/postgres.rs
+++ b/src/common/meta/src/kv_backend/rds/postgres.rs
@@ -868,6 +868,8 @@ impl PgStore {
        let client = match pool.get().await {
            Ok(client) => client,
            Err(e) => {
+                // We need to log the debug for the error to help diagnose the issue.
+                common_telemetry::error!(e; "Failed to get Postgres connection.");
                return GetPostgresConnectionSnafu {
                    reason: e.to_string(),
                }
--- a/src/datatypes/src/arrow_array.rs
+++ b/src/datatypes/src/arrow_array.rs
@@ -15,10 +15,12 @@
 use arrow::array::{ArrayRef, AsArray};
 use arrow::datatypes::{
    DataType, DurationMicrosecondType, DurationMillisecondType, DurationNanosecondType,
-    DurationSecondType, Time32MillisecondType, Time32SecondType, Time64MicrosecondType,
-    Time64NanosecondType, TimeUnit, TimestampMicrosecondType, TimestampMillisecondType,
-    TimestampNanosecondType, TimestampSecondType,
+    DurationSecondType, Int8Type, Int16Type, Int32Type, Int64Type, Time32MillisecondType,
+    Time32SecondType, Time64MicrosecondType, Time64NanosecondType, TimeUnit,
+    TimestampMicrosecondType, TimestampMillisecondType, TimestampNanosecondType,
+    TimestampSecondType, UInt8Type, UInt16Type, UInt32Type, UInt64Type,
 };
+use arrow_array::Array;
 use common_time::time::Time;
 use common_time::{Duration, Timestamp};

@@ -126,3 +128,87 @@ pub fn duration_array_value(array: &ArrayRef, i: usize) -> Duration {
    };
    Duration::new(v, time_unit.into())
 }
+
+/// Get the string value at index `i` for `Utf8`, `LargeUtf8`, or `Utf8View` arrays.
+///
+/// Returns `None` when the array type is not a string type or the value is null.
+///
+/// # Panics
+///
+/// If index `i` is out of bounds.
+pub fn string_array_value_at_index(array: &ArrayRef, i: usize) -> Option<&str> {
+    match array.data_type() {
+        DataType::Utf8 => {
+            let array = array.as_string::<i32>();
+            array.is_valid(i).then(|| array.value(i))
+        }
+        DataType::LargeUtf8 => {
+            let array = array.as_string::<i64>();
+            array.is_valid(i).then(|| array.value(i))
+        }
+        DataType::Utf8View => {
+            let array = array.as_string_view();
+            array.is_valid(i).then(|| array.value(i))
+        }
+        _ => None,
+    }
+}
+
+/// Get the integer value (`i64`) at index `i` for any integer array.
+///
+/// Returns `None` when:
+///
+/// - the array type is not an integer type;
+/// - the value is larger than `i64::MAX`;
+/// - the value is null.
+///
+/// # Panics
+///
+/// If index `i` is out of bounds.
+pub fn int_array_value_at_index(array: &ArrayRef, i: usize) -> Option<i64> {
+    match array.data_type() {
+        DataType::Int8 => {
+            let array = array.as_primitive::<Int8Type>();
+            array.is_valid(i).then(|| array.value(i) as i64)
+        }
+        DataType::Int16 => {
+            let array = array.as_primitive::<Int16Type>();
+            array.is_valid(i).then(|| array.value(i) as i64)
+        }
+        DataType::Int32 => {
+            let array = array.as_primitive::<Int32Type>();
+            array.is_valid(i).then(|| array.value(i) as i64)
+        }
+        DataType::Int64 => {
+            let array = array.as_primitive::<Int64Type>();
+            array.is_valid(i).then(|| array.value(i))
+        }
+        DataType::UInt8 => {
+            let array = array.as_primitive::<UInt8Type>();
+            array.is_valid(i).then(|| array.value(i) as i64)
+        }
+        DataType::UInt16 => {
+            let array = array.as_primitive::<UInt16Type>();
+            array.is_valid(i).then(|| array.value(i) as i64)
+        }
+        DataType::UInt32 => {
+            let array = array.as_primitive::<UInt32Type>();
+            array.is_valid(i).then(|| array.value(i) as i64)
+        }
+        DataType::UInt64 => {
+            let array = array.as_primitive::<UInt64Type>();
+            array
+                .is_valid(i)
+                .then(|| {
+                    let i = array.value(i);
+                    if i <= i64::MAX as u64 {
+                        Some(i as i64)
+                    } else {
+                        None
+                    }
+                })
+                .flatten()
+        }
+        _ => None,
+    }
+}
--- a/src/meta-srv/src/metasrv.rs
+++ b/src/meta-srv/src/metasrv.rs
@@ -299,7 +299,7 @@ impl Default for MetasrvOptions {
            #[allow(deprecated)]
            server_addr: String::new(),
            store_addrs: vec!["127.0.0.1:2379".to_string()],
-            backend_tls: None,
+            backend_tls: Some(TlsOption::prefer()),
            selector: SelectorType::default(),
            enable_region_failover: false,
            heartbeat_interval: distributed_time_constants::BASE_HEARTBEAT_INTERVAL,
--- a/src/meta-srv/src/utils/etcd.rs
+++ b/src/meta-srv/src/utils/etcd.rs
@@ -13,6 +13,7 @@
 // limitations under the License.

 use common_meta::kv_backend::etcd::create_etcd_tls_options;
+use common_telemetry::warn;
 use etcd_client::{Client, ConnectOptions};
 use servers::tls::{TlsMode, TlsOption};
 use snafu::ResultExt;
@@ -38,9 +39,12 @@ pub async fn create_etcd_client_with_tls(
            client_options.keep_alive_interval,
            client_options.keep_alive_timeout,
        );
+
+    let all_endpoints_use_https = etcd_endpoints.iter().all(|e| e.starts_with("https"));
    if let Some(tls_config) = tls_config
-        && let Some(tls_options) = create_etcd_tls_options(&convert_tls_option(tls_config))
-            .context(BuildTlsOptionsSnafu)?
+        && let Some(tls_options) =
+            create_etcd_tls_options(&convert_tls_option(all_endpoints_use_https, tls_config))
+                .context(BuildTlsOptionsSnafu)?
    {
        connect_options = connect_options.with_tls(tls_options);
    }
@@ -50,9 +54,22 @@ pub async fn create_etcd_client_with_tls(
        .context(error::ConnectEtcdSnafu)
 }

-fn convert_tls_option(tls_option: &TlsOption) -> common_meta::kv_backend::etcd::TlsOption {
+fn convert_tls_option(
+    all_endpoints_use_https: bool,
+    tls_option: &TlsOption,
+) -> common_meta::kv_backend::etcd::TlsOption {
    let mode = match tls_option.mode {
        TlsMode::Disable => common_meta::kv_backend::etcd::TlsMode::Disable,
+        TlsMode::Prefer => {
+            if all_endpoints_use_https {
+                common_meta::kv_backend::etcd::TlsMode::Require
+            } else {
+                warn!(
+                    "All endpoints use HTTP, TLS prefer mode is not supported, using disable mode"
+                );
+                common_meta::kv_backend::etcd::TlsMode::Disable
+            }
+        }
        _ => common_meta::kv_backend::etcd::TlsMode::Require,
    };
    common_meta::kv_backend::etcd::TlsOption {
--- a/src/mito2/Cargo.toml
+++ b/src/mito2/Cargo.toml
@@ -8,6 +8,7 @@ license.workspace = true
 default = []
 test = ["common-test-util", "rstest", "rstest_reuse", "rskafka"]
 enterprise = []
+vector_index = ["dep:usearch", "dep:roaring", "index/vector_index"]

 [lints]
 workspace = true
@@ -28,9 +29,10 @@ common-datasource.workspace = true
 common-decimal.workspace = true
 common-error.workspace = true
 common-grpc.workspace = true
+common-function.workspace = true
 common-macro.workspace = true
-common-meta.workspace = true
 common-memory-manager.workspace = true
+common-meta.workspace = true
 common-query.workspace = true
 common-recordbatch.workspace = true
 common-runtime.workspace = true
@@ -49,7 +51,6 @@ dotenv.workspace = true
 either.workspace = true
 futures.workspace = true
 humantime-serde.workspace = true
-humantime.workspace = true
 index.workspace = true
 itertools.workspace = true
 greptime-proto.workspace = true
@@ -67,6 +68,7 @@ partition.workspace = true
 puffin.workspace = true
 rand.workspace = true
 rayon = "1.10"
+roaring = { version = "0.10", optional = true }
 regex.workspace = true
 rskafka = { workspace = true, optional = true }
 rstest = { workspace = true, optional = true }
@@ -84,6 +86,7 @@ tokio.workspace = true
 tokio-stream.workspace = true
 tokio-util.workspace = true
 tracing.workspace = true
+usearch = { version = "2.21", default-features = false, features = ["fp16lib"], optional = true }
 uuid.workspace = true

 [dev-dependencies]
--- a/src/mito2/src/access_layer.rs
+++ b/src/mito2/src/access_layer.rs
@@ -313,6 +313,8 @@ impl AccessLayer {
                inverted_index_config: request.inverted_index_config,
                fulltext_index_config: request.fulltext_index_config,
                bloom_filter_index_config: request.bloom_filter_index_config,
+                #[cfg(feature = "vector_index")]
+                vector_index_config: request.vector_index_config,
            };
            // We disable write cache on file system but we still use atomic write.
            // TODO(yingwen): If we support other non-fs stores without the write cache, then
@@ -467,6 +469,8 @@ pub struct SstWriteRequest {
    pub inverted_index_config: InvertedIndexConfig,
    pub fulltext_index_config: FulltextIndexConfig,
    pub bloom_filter_index_config: BloomFilterConfig,
+    #[cfg(feature = "vector_index")]
+    pub vector_index_config: crate::config::VectorIndexConfig,
 }

 /// Cleaner to remove temp files on the atomic write dir.
--- a/src/mito2/src/cache/write_cache.rs
+++ b/src/mito2/src/cache/write_cache.rs
@@ -227,6 +227,8 @@ impl WriteCache {
            inverted_index_config: write_request.inverted_index_config,
            fulltext_index_config: write_request.fulltext_index_config,
            bloom_filter_index_config: write_request.bloom_filter_index_config,
+            #[cfg(feature = "vector_index")]
+            vector_index_config: write_request.vector_index_config,
        };

        let cleaner = TempFileCleaner::new(region_id, store.clone());
@@ -520,6 +522,8 @@ mod tests {
            inverted_index_config: Default::default(),
            fulltext_index_config: Default::default(),
            bloom_filter_index_config: Default::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        };

        let upload_request = SstUploadRequest {
@@ -620,6 +624,8 @@ mod tests {
            inverted_index_config: Default::default(),
            fulltext_index_config: Default::default(),
            bloom_filter_index_config: Default::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        };
        let write_opts = WriteOptions {
            row_group_size: 512,
@@ -701,6 +707,8 @@ mod tests {
            inverted_index_config: Default::default(),
            fulltext_index_config: Default::default(),
            bloom_filter_index_config: Default::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        };
        let write_opts = WriteOptions {
            row_group_size: 512,
--- a/src/mito2/src/compaction/compactor.rs
+++ b/src/mito2/src/compaction/compactor.rs
@@ -332,6 +332,8 @@ impl DefaultCompactor {
        let inverted_index_config = compaction_region.engine_config.inverted_index.clone();
        let fulltext_index_config = compaction_region.engine_config.fulltext_index.clone();
        let bloom_filter_index_config = compaction_region.engine_config.bloom_filter_index.clone();
+        #[cfg(feature = "vector_index")]
+        let vector_index_config = compaction_region.engine_config.vector_index.clone();

        let input_file_names = output
            .inputs
@@ -378,6 +380,8 @@ impl DefaultCompactor {
                    inverted_index_config,
                    fulltext_index_config,
                    bloom_filter_index_config,
+                    #[cfg(feature = "vector_index")]
+                    vector_index_config,
                },
                &write_opts,
                &mut metrics,
--- a/src/mito2/src/config.rs
+++ b/src/mito2/src/config.rs
@@ -158,6 +158,9 @@ pub struct MitoConfig {
    pub fulltext_index: FulltextIndexConfig,
    /// Bloom filter index configs.
    pub bloom_filter_index: BloomFilterConfig,
+    /// Vector index configs (HNSW).
+    #[cfg(feature = "vector_index")]
+    pub vector_index: VectorIndexConfig,

    /// Memtable config
    pub memtable: MemtableConfig,
@@ -214,6 +217,8 @@ impl Default for MitoConfig {
            inverted_index: InvertedIndexConfig::default(),
            fulltext_index: FulltextIndexConfig::default(),
            bloom_filter_index: BloomFilterConfig::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index: VectorIndexConfig::default(),
            memtable: MemtableConfig::default(),
            min_compaction_interval: Duration::from_secs(0),
            default_experimental_flat_format: false,
@@ -643,6 +648,51 @@ impl BloomFilterConfig {
    }
 }

+/// Configuration options for the vector index (HNSW).
+#[cfg(feature = "vector_index")]
+#[serde_as]
+#[derive(Debug, Serialize, Deserialize, Clone, PartialEq, Eq)]
+#[serde(default)]
+pub struct VectorIndexConfig {
+    /// Whether to create the index on flush: automatically or never.
+    pub create_on_flush: Mode,
+    /// Whether to create the index on compaction: automatically or never.
+    pub create_on_compaction: Mode,
+    /// Whether to apply the index on query: automatically or never.
+    pub apply_on_query: Mode,
+    /// Memory threshold for creating the index.
+    pub mem_threshold_on_create: MemoryThreshold,
+}
+
+#[cfg(feature = "vector_index")]
+impl Default for VectorIndexConfig {
+    fn default() -> Self {
+        Self {
+            create_on_flush: Mode::Auto,
+            create_on_compaction: Mode::Auto,
+            apply_on_query: Mode::Auto,
+            mem_threshold_on_create: MemoryThreshold::Auto,
+        }
+    }
+}
+
+#[cfg(feature = "vector_index")]
+impl VectorIndexConfig {
+    pub fn mem_threshold_on_create(&self) -> Option<usize> {
+        match self.mem_threshold_on_create {
+            MemoryThreshold::Auto => {
+                if let Some(sys_memory) = get_total_memory_readable() {
+                    Some((sys_memory / INDEX_CREATE_MEM_THRESHOLD_FACTOR).as_bytes() as usize)
+                } else {
+                    Some(ReadableSize::mb(64).as_bytes() as usize)
+                }
+            }
+            MemoryThreshold::Unlimited => None,
+            MemoryThreshold::Size(size) => Some(size.as_bytes() as usize),
+        }
+    }
+}
+
 /// Divide cpu num by a non-zero `divisor` and returns at least 1.
 fn divide_num_cpus(divisor: usize) -> usize {
    debug_assert!(divisor > 0);
--- a/src/mito2/src/engine.rs
+++ b/src/mito2/src/engine.rs
@@ -126,7 +126,7 @@ use crate::config::MitoConfig;
 use crate::engine::puffin_index::{IndexEntryContext, collect_index_entries_from_puffin};
 use crate::error::{
    InvalidRequestSnafu, JoinSnafu, MitoManifestInfoSnafu, RecvSnafu, RegionNotFoundSnafu, Result,
-    SerdeJsonSnafu, SerializeColumnMetadataSnafu, SerializeManifestSnafu,
+    SerdeJsonSnafu, SerializeColumnMetadataSnafu,
 };
 #[cfg(feature = "enterprise")]
 use crate::extension::BoxedExtensionRangeProviderFactory;
@@ -1057,19 +1057,8 @@ impl EngineInner {
        let region_id = request.region_id;
        let (request, receiver) = WorkerRequest::try_from_remap_manifests_request(request)?;
        self.workers.submit_to_worker(region_id, request).await?;
-        let manifests = receiver.await.context(RecvSnafu)??;
-
-        let new_manifests = manifests
-            .into_iter()
-            .map(|(region_id, manifest)| {
-                Ok((
-                    region_id,
-                    serde_json::to_string(&manifest)
-                        .context(SerializeManifestSnafu { region_id })?,
-                ))
-            })
-            .collect::<Result<HashMap<_, _>>>()?;
-        Ok(RemapManifestsResponse { new_manifests })
+        let manifest_paths = receiver.await.context(RecvSnafu)??;
+        Ok(RemapManifestsResponse { manifest_paths })
    }

    async fn copy_region_from(
--- a/src/mito2/src/engine/apply_staging_manifest_test.rs
+++ b/src/mito2/src/engine/apply_staging_manifest_test.rs
@@ -69,7 +69,8 @@ async fn test_apply_staging_manifest_invalid_region_state_with_format(flat_forma
            region_id,
            RegionRequest::ApplyStagingManifest(ApplyStagingManifestRequest {
                partition_expr: range_expr("x", 0, 100).as_json_str().unwrap(),
-                files_to_add: vec![],
+                central_region_id: RegionId::new(1, 0),
+                manifest_path: "manifest.json".to_string(),
            }),
        )
        .await
@@ -88,7 +89,8 @@ async fn test_apply_staging_manifest_invalid_region_state_with_format(flat_forma
            region_id,
            RegionRequest::ApplyStagingManifest(ApplyStagingManifestRequest {
                partition_expr: range_expr("x", 0, 100).as_json_str().unwrap(),
-                files_to_add: vec![],
+                central_region_id: RegionId::new(1, 0),
+                manifest_path: "manifest.json".to_string(),
            }),
        )
        .await
@@ -136,7 +138,8 @@ async fn test_apply_staging_manifest_mismatched_partition_expr_with_format(flat_
            region_id,
            RegionRequest::ApplyStagingManifest(ApplyStagingManifestRequest {
                partition_expr: range_expr("x", 0, 100).as_json_str().unwrap(),
-                files_to_add: vec![],
+                central_region_id: RegionId::new(1, 0),
+                manifest_path: "dummy".to_string(),
            }),
        )
        .await
@@ -144,7 +147,36 @@ async fn test_apply_staging_manifest_mismatched_partition_expr_with_format(flat_
    assert_matches!(
        err.into_inner().as_any().downcast_ref::<Error>().unwrap(),
        Error::StagingPartitionExprMismatch { .. }
-    )
+    );
+
+    // If staging manifest's partition expr is different from the request.
+    let result = engine
+        .remap_manifests(RemapManifestsRequest {
+            region_id,
+            input_regions: vec![region_id],
+            region_mapping: [(region_id, vec![region_id])].into_iter().collect(),
+            new_partition_exprs: [(region_id, range_expr("x", 0, 49).as_json_str().unwrap())]
+                .into_iter()
+                .collect(),
+        })
+        .await
+        .unwrap();
+
+    let err = engine
+        .handle_request(
+            region_id,
+            RegionRequest::ApplyStagingManifest(ApplyStagingManifestRequest {
+                partition_expr: range_expr("x", 0, 50).as_json_str().unwrap(),
+                central_region_id: region_id,
+                manifest_path: result.manifest_paths[&region_id].clone(),
+            }),
+        )
+        .await
+        .unwrap_err();
+    assert_matches!(
+        err.into_inner().as_any().downcast_ref::<Error>().unwrap(),
+        Error::StagingPartitionExprMismatch { .. }
+    );
 }

 #[tokio::test]
@@ -216,13 +248,26 @@ async fn test_apply_staging_manifest_success_with_format(flat_format: bool) {
        })
        .await
        .unwrap();
-    assert_eq!(result.new_manifests.len(), 2);
-    let new_manifest_1 =
-        serde_json::from_str::<RegionManifest>(&result.new_manifests[&new_region_id_1]).unwrap();
-    let new_manifest_2 =
-        serde_json::from_str::<RegionManifest>(&result.new_manifests[&new_region_id_2]).unwrap();
+    let region = engine.get_region(region_id).unwrap();
+    let manager = region.manifest_ctx.manifest_manager.write().await;
+    let manifest_storage = manager.store();
+    let blob_store = manifest_storage.staging_storage().blob_storage();
+
+    assert_eq!(result.manifest_paths.len(), 2);
+    common_telemetry::debug!("manifest paths: {:?}", result.manifest_paths);
+    let new_manifest_1 = blob_store
+        .get(&result.manifest_paths[&new_region_id_1])
+        .await
+        .unwrap();
+    let new_manifest_2 = blob_store
+        .get(&result.manifest_paths[&new_region_id_2])
+        .await
+        .unwrap();
+    let new_manifest_1 = serde_json::from_slice::<RegionManifest>(&new_manifest_1).unwrap();
+    let new_manifest_2 = serde_json::from_slice::<RegionManifest>(&new_manifest_2).unwrap();
    assert_eq!(new_manifest_1.files.len(), 3);
    assert_eq!(new_manifest_2.files.len(), 3);
+    drop(manager);

    let request = CreateRequestBuilder::new().build();
    engine
@@ -238,7 +283,6 @@ async fn test_apply_staging_manifest_success_with_format(flat_format: bool) {
        )
        .await
        .unwrap();
-    let mut files_to_add = new_manifest_1.files.values().cloned().collect::<Vec<_>>();
    // Before apply staging manifest, the files should be empty
    let region = engine.get_region(new_region_id_1).unwrap();
    let manifest = region.manifest_ctx.manifest().await;
@@ -251,7 +295,8 @@ async fn test_apply_staging_manifest_success_with_format(flat_format: bool) {
            new_region_id_1,
            RegionRequest::ApplyStagingManifest(ApplyStagingManifestRequest {
                partition_expr: range_expr("tag_0", 0, 50).as_json_str().unwrap(),
-                files_to_add: serde_json::to_vec(&files_to_add).unwrap(),
+                central_region_id: region_id,
+                manifest_path: result.manifest_paths[&new_region_id_1].clone(),
            }),
        )
        .await
@@ -277,23 +322,52 @@ async fn test_apply_staging_manifest_success_with_format(flat_format: bool) {
    let region_dir = format!("{}/data/test/1_0000000001", data_home.display());
    let staging_manifest_dir = format!("{}/staging/manifest", region_dir);
    let staging_files = fs::read_dir(&staging_manifest_dir)
-        .map(|entries| entries.collect::<Result<Vec<_>, _>>().unwrap_or_default())
+        .map(|entries| {
+            entries
+                .filter(|e| e.as_ref().unwrap().metadata().unwrap().is_file())
+                .collect::<Result<Vec<_>, _>>()
+                .unwrap_or_default()
+        })
        .unwrap_or_default();
-    assert_eq!(staging_files.len(), 0);
+    assert_eq!(staging_files.len(), 0, "staging_files: {:?}", staging_files);
+
+    let region = engine.get_region(region_id).unwrap();
+    let manager = region.manifest_ctx.manifest_manager.write().await;
+    let manifest_storage = manager.store();
+    let blob_store = manifest_storage.staging_storage().blob_storage();
+
+    let new_manifest_1 = blob_store
+        .get(&result.manifest_paths[&new_region_id_1])
+        .await
+        .unwrap();
+    let mut new_manifest_1 = serde_json::from_slice::<RegionManifest>(&new_manifest_1).unwrap();

    // Try to modify the file sequence.
-    files_to_add.push(FileMeta {
-        region_id,
-        file_id: FileId::random(),
-        ..Default::default()
-    });
+    let file_id = FileId::random();
+    new_manifest_1.files.insert(
+        file_id,
+        FileMeta {
+            region_id,
+            file_id,
+            ..Default::default()
+        },
+    );
+    blob_store
+        .put(
+            &result.manifest_paths[&new_region_id_1],
+            serde_json::to_vec(&new_manifest_1).unwrap(),
+        )
+        .await
+        .unwrap();
+    drop(manager);
    // This request will be ignored.
    engine
        .handle_request(
            new_region_id_1,
            RegionRequest::ApplyStagingManifest(ApplyStagingManifestRequest {
                partition_expr: range_expr("tag_0", 0, 50).as_json_str().unwrap(),
-                files_to_add: serde_json::to_vec(&files_to_add).unwrap(),
+                central_region_id: region_id,
+                manifest_path: result.manifest_paths[&new_region_id_1].clone(),
            }),
        )
        .await
@@ -334,12 +408,40 @@ async fn test_apply_staging_manifest_invalid_files_to_add_with_format(flat_forma
        )
        .await
        .unwrap();
+    // Apply staging manifest with not exists manifest path.
    let err = engine
        .handle_request(
            region_id,
            RegionRequest::ApplyStagingManifest(ApplyStagingManifestRequest {
                partition_expr: range_expr("tag_0", 0, 50).as_json_str().unwrap(),
-                files_to_add: b"invalid".to_vec(),
+                central_region_id: RegionId::new(1, 0),
+                manifest_path: "dummy".to_string(),
+            }),
+        )
+        .await
+        .unwrap_err();
+    assert_matches!(
+        err.into_inner().as_any().downcast_ref::<Error>().unwrap(),
+        Error::OpenDal { .. }
+    );
+
+    // Apply staging manifest with invalid bytes.
+    let region = engine.get_region(region_id).unwrap();
+    let manager = region.manifest_ctx.manifest_manager.write().await;
+    let manifest_storage = manager.store();
+    let blob_store = manifest_storage.staging_storage().blob_storage();
+    blob_store
+        .put("invalid_bytes", b"invalid_bytes".to_vec())
+        .await
+        .unwrap();
+    drop(manager);
+    let err = engine
+        .handle_request(
+            region_id,
+            RegionRequest::ApplyStagingManifest(ApplyStagingManifestRequest {
+                partition_expr: range_expr("tag_0", 0, 50).as_json_str().unwrap(),
+                central_region_id: region_id,
+                manifest_path: "invalid_bytes".to_string(),
            }),
        )
        .await
@@ -349,52 +451,3 @@ async fn test_apply_staging_manifest_invalid_files_to_add_with_format(flat_forma
        Error::SerdeJson { .. }
    );
 }
-
-#[tokio::test]
-async fn test_apply_staging_manifest_empty_files() {
-    common_telemetry::init_default_ut_logging();
-    test_apply_staging_manifest_empty_files_with_format(false).await;
-    test_apply_staging_manifest_empty_files_with_format(true).await;
-}
-
-async fn test_apply_staging_manifest_empty_files_with_format(flat_format: bool) {
-    let mut env = TestEnv::with_prefix("empty-files").await;
-    let engine = env
-        .create_engine(MitoConfig {
-            default_experimental_flat_format: flat_format,
-            ..Default::default()
-        })
-        .await;
-    let region_id = RegionId::new(1, 1);
-    let request = CreateRequestBuilder::new().build();
-    engine
-        .handle_request(region_id, RegionRequest::Create(request))
-        .await
-        .unwrap();
-    engine
-        .handle_request(
-            region_id,
-            RegionRequest::EnterStaging(EnterStagingRequest {
-                partition_expr: range_expr("tag_0", 0, 50).as_json_str().unwrap(),
-            }),
-        )
-        .await
-        .unwrap();
-    engine
-        .handle_request(
-            region_id,
-            RegionRequest::ApplyStagingManifest(ApplyStagingManifestRequest {
-                partition_expr: range_expr("tag_0", 0, 50).as_json_str().unwrap(),
-                files_to_add: serde_json::to_vec::<Vec<FileMeta>>(&vec![]).unwrap(),
-            }),
-        )
-        .await
-        .unwrap();
-    let region = engine.get_region(region_id).unwrap();
-    let manifest = region.manifest_ctx.manifest().await;
-    assert_eq!(manifest.files.len(), 0);
-    let staging_manifest = region.manifest_ctx.staging_manifest().await;
-    assert!(staging_manifest.is_none());
-    let staging_partition_expr = region.staging_partition_expr.lock().unwrap();
-    assert!(staging_partition_expr.is_none());
-}
--- a/src/mito2/src/engine/remap_manifests_test.rs
+++ b/src/mito2/src/engine/remap_manifests_test.rs
@@ -229,11 +229,23 @@ async fn test_remap_manifests_success_with_format(flat_format: bool) {
        })
        .await
        .unwrap();
-    assert_eq!(result.new_manifests.len(), 2);
-    let new_manifest_1 =
-        serde_json::from_str::<RegionManifest>(&result.new_manifests[&new_region_id_1]).unwrap();
-    let new_manifest_2 =
-        serde_json::from_str::<RegionManifest>(&result.new_manifests[&new_region_id_2]).unwrap();
+    let region = engine.get_region(region_id).unwrap();
+    let manager = region.manifest_ctx.manifest_manager.write().await;
+    let manifest_storage = manager.store();
+    let blob_store = manifest_storage.staging_storage().blob_storage();
+
+    assert_eq!(result.manifest_paths.len(), 2);
+    common_telemetry::debug!("manifest paths: {:?}", result.manifest_paths);
+    let new_manifest_1 = blob_store
+        .get(&result.manifest_paths[&new_region_id_1])
+        .await
+        .unwrap();
+    let new_manifest_2 = blob_store
+        .get(&result.manifest_paths[&new_region_id_2])
+        .await
+        .unwrap();
+    let new_manifest_1 = serde_json::from_slice::<RegionManifest>(&new_manifest_1).unwrap();
+    let new_manifest_2 = serde_json::from_slice::<RegionManifest>(&new_manifest_2).unwrap();
    assert_eq!(new_manifest_1.files.len(), 3);
    assert_eq!(new_manifest_2.files.len(), 3);
 }
--- a/src/mito2/src/error.rs
+++ b/src/mito2/src/error.rs
@@ -1039,6 +1039,22 @@ pub enum Error {
        location: Location,
    },

+    #[cfg(feature = "vector_index")]
+    #[snafu(display("Failed to build vector index: {}", reason))]
+    VectorIndexBuild {
+        reason: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
+    #[cfg(feature = "vector_index")]
+    #[snafu(display("Failed to finish vector index: {}", reason))]
+    VectorIndexFinish {
+        reason: String,
+        #[snafu(implicit)]
+        location: Location,
+    },
+
    #[snafu(display("Manual compaction is override by following operations."))]
    ManualCompactionOverride {},

@@ -1345,6 +1361,9 @@ impl ErrorExt for Error {
                source.status_code()
            }

+            #[cfg(feature = "vector_index")]
+            VectorIndexBuild { .. } | VectorIndexFinish { .. } => StatusCode::Internal,
+
            ManualCompactionOverride {} => StatusCode::Cancelled,

            CompactionMemoryExhausted { source, .. } => source.status_code(),
--- a/src/mito2/src/flush.rs
+++ b/src/mito2/src/flush.rs
@@ -669,6 +669,8 @@ impl RegionFlushTask {
            inverted_index_config: self.engine_config.inverted_index.clone(),
            fulltext_index_config: self.engine_config.fulltext_index.clone(),
            bloom_filter_index_config: self.engine_config.bloom_filter_index.clone(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: self.engine_config.vector_index.clone(),
        }
    }

--- a/src/mito2/src/manifest/storage.rs
+++ b/src/mito2/src/manifest/storage.rs
@@ -378,6 +378,11 @@ impl ManifestObjectStore {
    pub async fn clear_staging_manifests(&mut self) -> Result<()> {
        self.staging_storage.clear().await
    }
+
+    /// Returns the staging storage.
+    pub(crate) fn staging_storage(&self) -> &StagingStorage {
+        &self.staging_storage
+    }
 }

 #[cfg(test)]
--- a/src/mito2/src/manifest/storage/staging.rs
+++ b/src/mito2/src/manifest/storage/staging.rs
@@ -26,20 +26,104 @@ use crate::manifest::storage::size_tracker::NoopTracker;
 use crate::manifest::storage::utils::sort_manifests;
 use crate::manifest::storage::{file_version, is_delta_file};

+/// A simple blob storage for arbitrary binary data in the staging directory.
+///
+/// This is primarily used during repartition operations to store generated
+/// manifests that will be consumed by other regions via [`ApplyStagingManifestRequest`](store_api::region_request::ApplyStagingManifestRequest).
+/// The blobs are stored in `{region_dir}/staging/blob/` directory.
+#[derive(Debug, Clone)]
+pub(crate) struct StagingBlobStorage {
+    object_store: ObjectStore,
+    path: String,
+}
+
+/// Returns the staging path from the blob path.
+///
+/// # Example
+/// - Input: `"data/table/region_0001/manifest/"`
+/// - Output: `"data/table/region_0001/staging/blob/"`
+pub fn staging_blob_path(manifest_path: &str) -> String {
+    let parent_dir = manifest_path
+        .trim_end_matches("manifest/")
+        .trim_end_matches('/');
+    util::normalize_dir(&format!("{}/staging/blob", parent_dir))
+}
+
+impl StagingBlobStorage {
+    pub fn new(path: String, object_store: ObjectStore) -> Self {
+        let path = util::normalize_dir(&path);
+        common_telemetry::debug!(
+            "Staging blob storage path: {}, root: {}",
+            path,
+            object_store.info().root()
+        );
+        Self { object_store, path }
+    }
+
+    /// Put the bytes to the blob storage.
+    pub async fn put(&self, path: &str, bytes: Vec<u8>) -> Result<()> {
+        let path = format!("{}{}", self.path, path);
+        common_telemetry::debug!(
+            "Putting blob to staging blob storage, path: {}, root: {}, bytes: {}",
+            path,
+            self.object_store.info().root(),
+            bytes.len()
+        );
+        self.object_store
+            .write(&path, bytes)
+            .await
+            .context(OpenDalSnafu)?;
+        Ok(())
+    }
+
+    /// Get the bytes from the blob storage.
+    pub async fn get(&self, path: &str) -> Result<Vec<u8>> {
+        let path = format!("{}{}", self.path, path);
+        common_telemetry::debug!(
+            "Reading blob from staging blob storage, path: {}, root: {}",
+            path,
+            self.object_store.info().root()
+        );
+        let bytes = self.object_store.read(&path).await.context(OpenDalSnafu)?;
+
+        Ok(bytes.to_vec())
+    }
+}
+
+/// Storage for staging manifest files and blobs used during repartition operations.
+///
+/// Fields:
+/// - `delta_storage`: Manages incremental manifest delta files specific to the staging region.
+/// - `blob_storage`: Manages arbitrary blobs, such as generated manifests for regions.
+///
+/// Directory structure:
+/// - `{region_dir}/staging/manifest/` — for incremental manifest delta files for the staging region.
+/// - `{region_dir}/staging/blob/` — for arbitrary blobs (e.g., generated region manifests).
 #[derive(Debug, Clone)]
 pub(crate) struct StagingStorage {
    delta_storage: DeltaStorage<NoopTracker>,
+    blob_storage: StagingBlobStorage,
+}
+
+/// Returns the staging path from the manifest path.
+///
+/// # Example
+/// - Input: `"data/table/region_0001/manifest/"`
+/// - Output: `"data/table/region_0001/staging/manifest/"`
+pub fn staging_manifest_path(manifest_path: &str) -> String {
+    let parent_dir = manifest_path
+        .trim_end_matches("manifest/")
+        .trim_end_matches('/');
+    util::normalize_dir(&format!("{}/staging/manifest", parent_dir))
 }

 impl StagingStorage {
    pub fn new(path: String, object_store: ObjectStore, compress_type: CompressionType) -> Self {
-        let staging_path = {
-            // Convert "region_dir/manifest/" to "region_dir/staging/manifest/"
-            let parent_dir = path.trim_end_matches("manifest/").trim_end_matches('/');
-            util::normalize_dir(&format!("{}/staging/manifest", parent_dir))
-        };
+        let staging_blob_path = staging_blob_path(&path);
+        let blob_storage = StagingBlobStorage::new(staging_blob_path, object_store.clone());
+        let staging_manifest_path = staging_manifest_path(&path);
        let delta_storage = DeltaStorage::new(
-            staging_path.clone(),
+            staging_manifest_path.clone(),
            object_store.clone(),
            compress_type,
            // StagingStorage does not use a manifest cache; set to None.
@@ -48,7 +132,16 @@ impl StagingStorage {
            // deleted after exiting staging mode.
            Arc::new(NoopTracker),
        );
-        Self { delta_storage }
+
+        Self {
+            delta_storage,
+            blob_storage,
+        }
+    }
+
+    /// Returns the blob storage.
+    pub(crate) fn blob_storage(&self) -> &StagingBlobStorage {
+        &self.blob_storage
    }

    /// Returns an iterator of manifests from staging directory.
@@ -107,3 +200,22 @@ impl StagingStorage {
        self.delta_storage.set_compress_type(compress_type);
    }
 }
+
+#[cfg(test)]
+mod tests {
+    use crate::manifest::storage::staging::{staging_blob_path, staging_manifest_path};
+
+    #[test]
+    fn test_staging_path() {
+        let path = "/data/table/region_0001/manifest/";
+        let expected = "/data/table/region_0001/staging/manifest/";
+        assert_eq!(staging_manifest_path(path), expected);
+    }
+
+    #[test]
+    fn test_staging_blob_path() {
+        let path = "/data/table/region_0001/manifest/";
+        let expected = "/data/table/region_0001/staging/blob/";
+        assert_eq!(staging_blob_path(path), expected);
+    }
+}
--- a/src/mito2/src/request.rs
+++ b/src/mito2/src/request.rs
@@ -50,7 +50,7 @@ use crate::error::{
    FlushRegionSnafu, InvalidPartitionExprSnafu, InvalidRequestSnafu, MissingPartitionExprSnafu,
    Result, UnexpectedSnafu,
 };
-use crate::manifest::action::{RegionEdit, RegionManifest, TruncateKind};
+use crate::manifest::action::{RegionEdit, TruncateKind};
 use crate::memtable::MemtableId;
 use crate::memtable::bulk::part::BulkPart;
 use crate::metrics::COMPACTION_ELAPSED_TOTAL;
@@ -796,10 +796,7 @@ impl WorkerRequest {
            region_mapping,
            new_partition_exprs,
        }: store_api::region_engine::RemapManifestsRequest,
-    ) -> Result<(
-        WorkerRequest,
-        Receiver<Result<HashMap<RegionId, RegionManifest>>>,
-    )> {
+    ) -> Result<(WorkerRequest, Receiver<Result<HashMap<RegionId, String>>>)> {
        let (sender, receiver) = oneshot::channel();
        let new_partition_exprs = new_partition_exprs
            .into_iter()
@@ -1116,8 +1113,10 @@ pub(crate) struct RemapManifestsRequest {
    pub(crate) region_mapping: HashMap<RegionId, Vec<RegionId>>,
    /// New partition expressions for the new regions.
    pub(crate) new_partition_exprs: HashMap<RegionId, PartitionExpr>,
-    /// Result sender.
-    pub(crate) sender: Sender<Result<HashMap<RegionId, RegionManifest>>>,
+    /// Sender for the result of the remap operation.
+    ///
+    /// The result is a map from region IDs to their corresponding staging manifest paths.
+    pub(crate) sender: Sender<Result<HashMap<RegionId, String>>>,
 }

 #[derive(Debug)]
--- a/src/mito2/src/sst/file.rs
+++ b/src/mito2/src/sst/file.rs
@@ -287,6 +287,9 @@ pub enum IndexType {
    FulltextIndex,
    /// Bloom Filter index
    BloomFilterIndex,
+    /// Vector index (HNSW).
+    #[cfg(feature = "vector_index")]
+    VectorIndex,
 }

 /// Metadata of indexes created for a specific column in an SST file.
--- a/src/mito2/src/sst/index.rs
+++ b/src/mito2/src/sst/index.rs
@@ -20,6 +20,8 @@ pub(crate) mod inverted_index;
 pub mod puffin_manager;
 mod statistics;
 pub(crate) mod store;
+#[cfg(feature = "vector_index")]
+pub(crate) mod vector_index;

 use std::cmp::Ordering;
 use std::collections::{BinaryHeap, HashMap, HashSet};
@@ -41,10 +43,14 @@ use store_api::metadata::RegionMetadataRef;
 use store_api::storage::{ColumnId, FileId, RegionId};
 use strum::IntoStaticStr;
 use tokio::sync::mpsc::Sender;
+#[cfg(feature = "vector_index")]
+use vector_index::creator::VectorIndexer;

 use crate::access_layer::{AccessLayerRef, FilePathProvider, OperationType, RegionFilePathFactory};
 use crate::cache::file_cache::{FileCacheRef, FileType, IndexKey};
 use crate::cache::write_cache::{UploadTracker, WriteCacheRef};
+#[cfg(feature = "vector_index")]
+use crate::config::VectorIndexConfig;
 use crate::config::{BloomFilterConfig, FulltextIndexConfig, InvertedIndexConfig};
 use crate::error::{
    BuildIndexAsyncSnafu, DecodeSnafu, Error, InvalidRecordBatchSnafu, RegionClosedSnafu,
@@ -76,6 +82,8 @@ use crate::worker::WorkerListener;
 pub(crate) const TYPE_INVERTED_INDEX: &str = "inverted_index";
 pub(crate) const TYPE_FULLTEXT_INDEX: &str = "fulltext_index";
 pub(crate) const TYPE_BLOOM_FILTER_INDEX: &str = "bloom_filter_index";
+#[cfg(feature = "vector_index")]
+pub(crate) const TYPE_VECTOR_INDEX: &str = "vector_index";

 /// Triggers background download of an index file to the local cache.
 pub(crate) fn trigger_index_background_download(
@@ -114,6 +122,9 @@ pub struct IndexOutput {
    pub fulltext_index: FulltextIndexOutput,
    /// Bloom filter output.
    pub bloom_filter: BloomFilterOutput,
+    /// Vector index output.
+    #[cfg(feature = "vector_index")]
+    pub vector_index: VectorIndexOutput,
 }

 impl IndexOutput {
@@ -128,6 +139,10 @@ impl IndexOutput {
        if self.bloom_filter.is_available() {
            indexes.push(IndexType::BloomFilterIndex);
        }
+        #[cfg(feature = "vector_index")]
+        if self.vector_index.is_available() {
+            indexes.push(IndexType::VectorIndex);
+        }
        indexes
    }

@@ -151,6 +166,12 @@ impl IndexOutput {
                    .push(IndexType::BloomFilterIndex);
            }
        }
+        #[cfg(feature = "vector_index")]
+        if self.vector_index.is_available() {
+            for &col in &self.vector_index.columns {
+                map.entry(col).or_default().push(IndexType::VectorIndex);
+            }
+        }

        map.into_iter()
            .map(|(column_id, created_indexes)| ColumnIndexMetadata {
@@ -184,6 +205,9 @@ pub type InvertedIndexOutput = IndexBaseOutput;
 pub type FulltextIndexOutput = IndexBaseOutput;
 /// Output of the bloom filter creation.
 pub type BloomFilterOutput = IndexBaseOutput;
+/// Output of the vector index creation.
+#[cfg(feature = "vector_index")]
+pub type VectorIndexOutput = IndexBaseOutput;

 /// The index creator that hides the error handling details.
 #[derive(Default)]
@@ -199,6 +223,10 @@ pub struct Indexer {
    last_mem_fulltext_index: usize,
    bloom_filter_indexer: Option<BloomFilterIndexer>,
    last_mem_bloom_filter: usize,
+    #[cfg(feature = "vector_index")]
+    vector_indexer: Option<VectorIndexer>,
+    #[cfg(feature = "vector_index")]
+    last_mem_vector_index: usize,
    intermediate_manager: Option<IntermediateManager>,
 }

@@ -259,6 +287,18 @@ impl Indexer {
            .with_label_values(&[TYPE_BLOOM_FILTER_INDEX])
            .add(bloom_filter_mem as i64 - self.last_mem_bloom_filter as i64);
        self.last_mem_bloom_filter = bloom_filter_mem;
+
+        #[cfg(feature = "vector_index")]
+        {
+            let vector_mem = self
+                .vector_indexer
+                .as_ref()
+                .map_or(0, |creator| creator.memory_usage());
+            INDEX_CREATE_MEMORY_USAGE
+                .with_label_values(&[TYPE_VECTOR_INDEX])
+                .add(vector_mem as i64 - self.last_mem_vector_index as i64);
+            self.last_mem_vector_index = vector_mem;
+        }
    }
 }

@@ -279,6 +319,8 @@ pub(crate) struct IndexerBuilderImpl {
    pub(crate) inverted_index_config: InvertedIndexConfig,
    pub(crate) fulltext_index_config: FulltextIndexConfig,
    pub(crate) bloom_filter_index_config: BloomFilterConfig,
+    #[cfg(feature = "vector_index")]
+    pub(crate) vector_index_config: VectorIndexConfig,
 }

 #[async_trait::async_trait]
@@ -296,11 +338,23 @@ impl IndexerBuilder for IndexerBuilderImpl {
        indexer.inverted_indexer = self.build_inverted_indexer(file_id);
        indexer.fulltext_indexer = self.build_fulltext_indexer(file_id).await;
        indexer.bloom_filter_indexer = self.build_bloom_filter_indexer(file_id);
-        indexer.intermediate_manager = Some(self.intermediate_manager.clone());
-        if indexer.inverted_indexer.is_none()
-            && indexer.fulltext_indexer.is_none()
-            && indexer.bloom_filter_indexer.is_none()
+        #[cfg(feature = "vector_index")]
        {
+            indexer.vector_indexer = self.build_vector_indexer(file_id);
+        }
+        indexer.intermediate_manager = Some(self.intermediate_manager.clone());
+
+        #[cfg(feature = "vector_index")]
+        let has_any_indexer = indexer.inverted_indexer.is_some()
+            || indexer.fulltext_indexer.is_some()
+            || indexer.bloom_filter_indexer.is_some()
+            || indexer.vector_indexer.is_some();
+        #[cfg(not(feature = "vector_index"))]
+        let has_any_indexer = indexer.inverted_indexer.is_some()
+            || indexer.fulltext_indexer.is_some()
+            || indexer.bloom_filter_indexer.is_some();
+
+        if !has_any_indexer {
            indexer.abort().await;
            return Indexer::default();
        }
@@ -476,6 +530,69 @@ impl IndexerBuilderImpl {

        None
    }
+
+    #[cfg(feature = "vector_index")]
+    fn build_vector_indexer(&self, file_id: FileId) -> Option<VectorIndexer> {
+        let create = match self.build_type {
+            IndexBuildType::Flush => self.vector_index_config.create_on_flush.auto(),
+            IndexBuildType::Compact => self.vector_index_config.create_on_compaction.auto(),
+            _ => true,
+        };
+
+        if !create {
+            debug!(
+                "Skip creating vector index due to config, region_id: {}, file_id: {}",
+                self.metadata.region_id, file_id,
+            );
+            return None;
+        }
+
+        // Get vector index column IDs and options from metadata
+        let vector_index_options = self.metadata.vector_indexed_column_ids();
+        if vector_index_options.is_empty() {
+            debug!(
+                "No vector columns to index, skip creating vector index, region_id: {}, file_id: {}",
+                self.metadata.region_id, file_id,
+            );
+            return None;
+        }
+
+        let mem_limit = self.vector_index_config.mem_threshold_on_create();
+        let indexer = VectorIndexer::new(
+            file_id,
+            &self.metadata,
+            self.intermediate_manager.clone(),
+            mem_limit,
+            &vector_index_options,
+        );
+
+        let err = match indexer {
+            Ok(indexer) => {
+                if indexer.is_none() {
+                    debug!(
+                        "Skip creating vector index due to no columns require indexing, region_id: {}, file_id: {}",
+                        self.metadata.region_id, file_id,
+                    );
+                }
+                return indexer;
+            }
+            Err(err) => err,
+        };
+
+        if cfg!(any(test, feature = "test")) {
+            panic!(
+                "Failed to create vector index, region_id: {}, file_id: {}, err: {:?}",
+                self.metadata.region_id, file_id, err
+            );
+        } else {
+            warn!(
+                err; "Failed to create vector index, region_id: {}, file_id: {}",
+                self.metadata.region_id, file_id,
+            );
+        }
+
+        None
+    }
 }

 /// Type of an index build task.
@@ -1115,6 +1232,8 @@ mod tests {
        with_inverted: bool,
        with_fulltext: bool,
        with_skipping_bloom: bool,
+        #[cfg(feature = "vector_index")]
+        with_vector: bool,
    }

    fn mock_region_metadata(
@@ -1122,6 +1241,8 @@ mod tests {
            with_inverted,
            with_fulltext,
            with_skipping_bloom,
+            #[cfg(feature = "vector_index")]
+            with_vector,
        }: MetaConfig,
    ) -> RegionMetadataRef {
        let mut builder = RegionMetadataBuilder::new(RegionId::new(1, 2));
@@ -1187,6 +1308,24 @@ mod tests {
            builder.push_column_metadata(column);
        }

+        #[cfg(feature = "vector_index")]
+        if with_vector {
+            use index::vector::VectorIndexOptions;
+
+            let options = VectorIndexOptions::default();
+            let column_schema =
+                ColumnSchema::new("vec", ConcreteDataType::vector_datatype(4), true)
+                    .with_vector_index_options(&options)
+                    .unwrap();
+            let column = ColumnMetadata {
+                column_schema,
+                semantic_type: SemanticType::Field,
+                column_id: 6,
+            };
+
+            builder.push_column_metadata(column);
+        }
+
        Arc::new(builder.build().unwrap())
    }

@@ -1237,6 +1376,8 @@ mod tests {
            inverted_index_config: Default::default(),
            fulltext_index_config: Default::default(),
            bloom_filter_index_config: Default::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        };
        let mut metrics = Metrics::new(WriteType::Flush);
        env.access_layer
@@ -1287,6 +1428,8 @@ mod tests {
            inverted_index_config: InvertedIndexConfig::default(),
            fulltext_index_config: FulltextIndexConfig::default(),
            bloom_filter_index_config: BloomFilterConfig::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        })
    }

@@ -1300,6 +1443,8 @@ mod tests {
            with_inverted: true,
            with_fulltext: true,
            with_skipping_bloom: true,
+            #[cfg(feature = "vector_index")]
+            with_vector: false,
        });
        let indexer = IndexerBuilderImpl {
            build_type: IndexBuildType::Flush,
@@ -1312,6 +1457,8 @@ mod tests {
            inverted_index_config: InvertedIndexConfig::default(),
            fulltext_index_config: FulltextIndexConfig::default(),
            bloom_filter_index_config: BloomFilterConfig::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        }
        .build(FileId::random(), 0)
        .await;
@@ -1331,6 +1478,8 @@ mod tests {
            with_inverted: true,
            with_fulltext: true,
            with_skipping_bloom: true,
+            #[cfg(feature = "vector_index")]
+            with_vector: false,
        });
        let indexer = IndexerBuilderImpl {
            build_type: IndexBuildType::Flush,
@@ -1346,6 +1495,8 @@ mod tests {
            },
            fulltext_index_config: FulltextIndexConfig::default(),
            bloom_filter_index_config: BloomFilterConfig::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        }
        .build(FileId::random(), 0)
        .await;
@@ -1368,6 +1519,8 @@ mod tests {
                ..Default::default()
            },
            bloom_filter_index_config: BloomFilterConfig::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        }
        .build(FileId::random(), 0)
        .await;
@@ -1390,6 +1543,8 @@ mod tests {
                create_on_compaction: Mode::Disable,
                ..Default::default()
            },
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        }
        .build(FileId::random(), 0)
        .await;
@@ -1409,6 +1564,8 @@ mod tests {
            with_inverted: false,
            with_fulltext: true,
            with_skipping_bloom: true,
+            #[cfg(feature = "vector_index")]
+            with_vector: false,
        });
        let indexer = IndexerBuilderImpl {
            build_type: IndexBuildType::Flush,
@@ -1421,6 +1578,8 @@ mod tests {
            inverted_index_config: InvertedIndexConfig::default(),
            fulltext_index_config: FulltextIndexConfig::default(),
            bloom_filter_index_config: BloomFilterConfig::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        }
        .build(FileId::random(), 0)
        .await;
@@ -1433,6 +1592,8 @@ mod tests {
            with_inverted: true,
            with_fulltext: false,
            with_skipping_bloom: true,
+            #[cfg(feature = "vector_index")]
+            with_vector: false,
        });
        let indexer = IndexerBuilderImpl {
            build_type: IndexBuildType::Flush,
@@ -1445,6 +1606,8 @@ mod tests {
            inverted_index_config: InvertedIndexConfig::default(),
            fulltext_index_config: FulltextIndexConfig::default(),
            bloom_filter_index_config: BloomFilterConfig::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        }
        .build(FileId::random(), 0)
        .await;
@@ -1457,6 +1620,8 @@ mod tests {
            with_inverted: true,
            with_fulltext: true,
            with_skipping_bloom: false,
+            #[cfg(feature = "vector_index")]
+            with_vector: false,
        });
        let indexer = IndexerBuilderImpl {
            build_type: IndexBuildType::Flush,
@@ -1469,6 +1634,8 @@ mod tests {
            inverted_index_config: InvertedIndexConfig::default(),
            fulltext_index_config: FulltextIndexConfig::default(),
            bloom_filter_index_config: BloomFilterConfig::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        }
        .build(FileId::random(), 0)
        .await;
@@ -1488,6 +1655,8 @@ mod tests {
            with_inverted: true,
            with_fulltext: true,
            with_skipping_bloom: true,
+            #[cfg(feature = "vector_index")]
+            with_vector: false,
        });
        let indexer = IndexerBuilderImpl {
            build_type: IndexBuildType::Flush,
@@ -1500,6 +1669,8 @@ mod tests {
            inverted_index_config: InvertedIndexConfig::default(),
            fulltext_index_config: FulltextIndexConfig::default(),
            bloom_filter_index_config: BloomFilterConfig::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        }
        .build(FileId::random(), 0)
        .await;
@@ -1507,6 +1678,82 @@ mod tests {
        assert!(indexer.inverted_indexer.is_none());
    }

+    #[cfg(feature = "vector_index")]
+    #[tokio::test]
+    async fn test_update_flat_builds_vector_index() {
+        use datatypes::arrow::array::BinaryBuilder;
+        use datatypes::arrow::datatypes::{DataType, Field, Schema};
+
+        struct TestPathProvider;
+
+        impl FilePathProvider for TestPathProvider {
+            fn build_index_file_path(&self, file_id: RegionFileId) -> String {
+                format!("index/{}.puffin", file_id)
+            }
+
+            fn build_index_file_path_with_version(&self, index_id: RegionIndexId) -> String {
+                format!("index/{}.puffin", index_id)
+            }
+
+            fn build_sst_file_path(&self, file_id: RegionFileId) -> String {
+                format!("sst/{}.parquet", file_id)
+            }
+        }
+
+        fn f32s_to_bytes(values: &[f32]) -> Vec<u8> {
+            let mut bytes = Vec::with_capacity(values.len() * 4);
+            for v in values {
+                bytes.extend_from_slice(&v.to_le_bytes());
+            }
+            bytes
+        }
+
+        let (dir, factory) =
+            PuffinManagerFactory::new_for_test_async("test_update_flat_builds_vector_index_").await;
+        let intm_manager = mock_intm_mgr(dir.path().to_string_lossy()).await;
+
+        let metadata = mock_region_metadata(MetaConfig {
+            with_inverted: false,
+            with_fulltext: false,
+            with_skipping_bloom: false,
+            with_vector: true,
+        });
+
+        let mut indexer = IndexerBuilderImpl {
+            build_type: IndexBuildType::Flush,
+            metadata,
+            row_group_size: 1024,
+            puffin_manager: factory.build(mock_object_store(), TestPathProvider),
+            write_cache_enabled: false,
+            intermediate_manager: intm_manager,
+            index_options: IndexOptions::default(),
+            inverted_index_config: InvertedIndexConfig::default(),
+            fulltext_index_config: FulltextIndexConfig::default(),
+            bloom_filter_index_config: BloomFilterConfig::default(),
+            vector_index_config: Default::default(),
+        }
+        .build(FileId::random(), 0)
+        .await;
+
+        assert!(indexer.vector_indexer.is_some());
+
+        let vec1 = f32s_to_bytes(&[1.0, 0.0, 0.0, 0.0]);
+        let vec2 = f32s_to_bytes(&[0.0, 1.0, 0.0, 0.0]);
+
+        let mut builder = BinaryBuilder::with_capacity(2, vec1.len() + vec2.len());
+        builder.append_value(&vec1);
+        builder.append_value(&vec2);
+
+        let schema = Arc::new(Schema::new(vec![Field::new("vec", DataType::Binary, true)]));
+        let batch = RecordBatch::try_new(schema, vec![Arc::new(builder.finish())]).unwrap();
+
+        indexer.update_flat(&batch).await;
+        let output = indexer.finish().await;
+
+        assert!(output.vector_index.is_available());
+        assert!(output.vector_index.columns.contains(&6));
+    }
+
    #[tokio::test]
    async fn test_index_build_task_sst_not_exist() {
        let env = SchedulerEnv::new().await;
@@ -1839,6 +2086,8 @@ mod tests {
            inverted_index_config: InvertedIndexConfig::default(),
            fulltext_index_config: FulltextIndexConfig::default(),
            bloom_filter_index_config: BloomFilterConfig::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        });

        let sst_info = mock_sst_file(metadata.clone(), &env, IndexBuildMode::Async).await;
--- a/src/mito2/src/sst/index/indexer/abort.rs
+++ b/src/mito2/src/sst/index/indexer/abort.rs
@@ -23,6 +23,8 @@ impl Indexer {
        self.do_abort_inverted_index().await;
        self.do_abort_fulltext_index().await;
        self.do_abort_bloom_filter().await;
+        #[cfg(feature = "vector_index")]
+        self.do_abort_vector_index().await;
        self.do_prune_intm_sst_dir().await;
        if self.write_cache_enabled {
            self.do_abort_clean_fs_temp_dir().await;
@@ -106,4 +108,26 @@ impl Indexer {
        .to_string();
        TempFileCleaner::clean_atomic_dir_files(fs_accessor.store().store(), &[&fs_handle]).await;
    }
+
+    #[cfg(feature = "vector_index")]
+    async fn do_abort_vector_index(&mut self) {
+        let Some(mut indexer) = self.vector_indexer.take() else {
+            return;
+        };
+        let Err(err) = indexer.abort().await else {
+            return;
+        };
+
+        if cfg!(any(test, feature = "test")) {
+            panic!(
+                "Failed to abort vector index, region_id: {}, file_id: {}, err: {:?}",
+                self.region_id, self.file_id, err
+            );
+        } else {
+            warn!(
+                err; "Failed to abort vector index, region_id: {}, file_id: {}",
+                self.region_id, self.file_id,
+            );
+        }
+    }
 }
--- a/src/mito2/src/sst/index/indexer/finish.rs
+++ b/src/mito2/src/sst/index/indexer/finish.rs
@@ -17,6 +17,8 @@ use puffin::puffin_manager::{PuffinManager, PuffinWriter};
 use store_api::storage::ColumnId;

 use crate::sst::file::{RegionFileId, RegionIndexId};
+#[cfg(feature = "vector_index")]
+use crate::sst::index::VectorIndexOutput;
 use crate::sst::index::puffin_manager::SstPuffinWriter;
 use crate::sst::index::statistics::{ByteCount, RowCount};
 use crate::sst::index::{
@@ -54,6 +56,15 @@ impl Indexer {
            return IndexOutput::default();
        }

+        #[cfg(feature = "vector_index")]
+        {
+            let success = self.do_finish_vector_index(&mut writer, &mut output).await;
+            if !success {
+                self.do_abort().await;
+                return IndexOutput::default();
+            }
+        }
+
        self.do_prune_intm_sst_dir().await;
        output.file_size = self.do_finish_puffin_writer(writer).await;
        output.version = self.index_version;
@@ -276,6 +287,63 @@ impl Indexer {
        output.columns = column_ids;
    }

+    #[cfg(feature = "vector_index")]
+    async fn do_finish_vector_index(
+        &mut self,
+        puffin_writer: &mut SstPuffinWriter,
+        index_output: &mut IndexOutput,
+    ) -> bool {
+        let Some(mut indexer) = self.vector_indexer.take() else {
+            return true;
+        };
+
+        let column_ids = indexer.column_ids().collect();
+        let err = match indexer.finish(puffin_writer).await {
+            Ok((row_count, byte_count)) => {
+                self.fill_vector_index_output(
+                    &mut index_output.vector_index,
+                    row_count,
+                    byte_count,
+                    column_ids,
+                );
+                return true;
+            }
+            Err(err) => err,
+        };
+
+        if cfg!(any(test, feature = "test")) {
+            panic!(
+                "Failed to finish vector index, region_id: {}, file_id: {}, err: {:?}",
+                self.region_id, self.file_id, err
+            );
+        } else {
+            warn!(
+                err; "Failed to finish vector index, region_id: {}, file_id: {}",
+                self.region_id, self.file_id,
+            );
+        }
+
+        false
+    }
+
+    #[cfg(feature = "vector_index")]
+    fn fill_vector_index_output(
+        &mut self,
+        output: &mut VectorIndexOutput,
+        row_count: RowCount,
+        byte_count: ByteCount,
+        column_ids: Vec<ColumnId>,
+    ) {
+        debug!(
+            "Vector index created, region_id: {}, file_id: {}, written_bytes: {}, written_rows: {}, columns: {:?}",
+            self.region_id, self.file_id, byte_count, row_count, column_ids
+        );
+
+        output.index_size = byte_count;
+        output.row_count = row_count;
+        output.columns = column_ids;
+    }
+
    pub(crate) async fn do_prune_intm_sst_dir(&mut self) {
        if let Some(manager) = self.intermediate_manager.take()
            && let Err(e) = manager.prune_sst_dir(&self.region_id, &self.file_id).await
--- a/src/mito2/src/sst/index/indexer/update.rs
+++ b/src/mito2/src/sst/index/indexer/update.rs
@@ -33,6 +33,10 @@ impl Indexer {
        if !self.do_update_bloom_filter(batch).await {
            self.do_abort().await;
        }
+        #[cfg(feature = "vector_index")]
+        if !self.do_update_vector_index(batch).await {
+            self.do_abort().await;
+        }
    }

    /// Returns false if the update failed.
@@ -110,6 +114,32 @@ impl Indexer {
        false
    }

+    /// Returns false if the update failed.
+    #[cfg(feature = "vector_index")]
+    async fn do_update_vector_index(&mut self, batch: &mut Batch) -> bool {
+        let Some(creator) = self.vector_indexer.as_mut() else {
+            return true;
+        };
+
+        let Err(err) = creator.update(batch).await else {
+            return true;
+        };
+
+        if cfg!(any(test, feature = "test")) {
+            panic!(
+                "Failed to update vector index, region_id: {}, file_id: {}, err: {:?}",
+                self.region_id, self.file_id, err
+            );
+        } else {
+            warn!(
+                err; "Failed to update vector index, region_id: {}, file_id: {}",
+                self.region_id, self.file_id,
+            );
+        }
+
+        false
+    }
+
    pub(crate) async fn do_update_flat(&mut self, batch: &RecordBatch) {
        if batch.num_rows() == 0 {
            return;
@@ -124,6 +154,10 @@ impl Indexer {
        if !self.do_update_flat_bloom_filter(batch).await {
            self.do_abort().await;
        }
+        #[cfg(feature = "vector_index")]
+        if !self.do_update_flat_vector_index(batch).await {
+            self.do_abort().await;
+        }
    }

    /// Returns false if the update failed.
@@ -200,4 +234,30 @@ impl Indexer {

        false
    }
+
+    /// Returns false if the update failed.
+    #[cfg(feature = "vector_index")]
+    async fn do_update_flat_vector_index(&mut self, batch: &RecordBatch) -> bool {
+        let Some(creator) = self.vector_indexer.as_mut() else {
+            return true;
+        };
+
+        let Err(err) = creator.update_flat(batch).await else {
+            return true;
+        };
+
+        if cfg!(any(test, feature = "test")) {
+            panic!(
+                "Failed to update vector index with flat format, region_id: {}, file_id: {}, err: {:?}",
+                self.region_id, self.file_id, err
+            );
+        } else {
+            warn!(
+                err; "Failed to update vector index with flat format, region_id: {}, file_id: {}",
+                self.region_id, self.file_id,
+            );
+        }
+
+        false
+    }
 }
--- a/src/mito2/src/sst/index/vector_index/creator.rs
+++ b/src/mito2/src/sst/index/vector_index/creator.rs
@@ -0,0 +1,920 @@
+// 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.
+
+//! Vector index creator using pluggable vector index engines.
+
+use std::collections::HashMap;
+use std::sync::Arc;
+use std::sync::atomic::AtomicUsize;
+
+use common_telemetry::warn;
+use datatypes::arrow::array::{Array, BinaryArray};
+use datatypes::arrow::record_batch::RecordBatch;
+use datatypes::data_type::ConcreteDataType;
+use datatypes::prelude::ValueRef;
+use index::vector::{VectorDistanceMetric, VectorIndexOptions, distance_metric_to_usearch};
+use puffin::puffin_manager::{PuffinWriter, PutOptions};
+use roaring::RoaringBitmap;
+use snafu::{ResultExt, ensure};
+use store_api::metadata::RegionMetadataRef;
+use store_api::storage::{ColumnId, FileId, VectorIndexEngine, VectorIndexEngineType};
+use tokio_util::compat::TokioAsyncReadCompatExt;
+use usearch::MetricKind;
+
+use crate::error::{
+    BiErrorsSnafu, OperateAbortedIndexSnafu, PuffinAddBlobSnafu, Result, VectorIndexBuildSnafu,
+    VectorIndexFinishSnafu,
+};
+use crate::metrics::{INDEX_CREATE_BYTES_TOTAL, INDEX_CREATE_ROWS_TOTAL};
+use crate::read::Batch;
+use crate::sst::index::TYPE_VECTOR_INDEX;
+use crate::sst::index::intermediate::{
+    IntermediateLocation, IntermediateManager, TempFileProvider,
+};
+use crate::sst::index::puffin_manager::SstPuffinWriter;
+use crate::sst::index::statistics::{ByteCount, RowCount, Statistics};
+use crate::sst::index::vector_index::util::bytes_to_f32_slice;
+use crate::sst::index::vector_index::{INDEX_BLOB_TYPE, engine};
+
+/// The buffer size for the pipe used to send index data to the puffin blob.
+const PIPE_BUFFER_SIZE_FOR_SENDING_BLOB: usize = 8192;
+
+/// Configuration for a single column's vector index.
+#[derive(Debug, Clone)]
+pub struct VectorIndexConfig {
+    /// The vector index engine type.
+    pub engine: VectorIndexEngineType,
+    /// The dimension of vectors in this column.
+    pub dim: usize,
+    /// The distance metric to use (e.g., L2, Cosine, IP) - usearch format.
+    pub metric: MetricKind,
+    /// The original distance metric (for serialization).
+    pub distance_metric: VectorDistanceMetric,
+    /// HNSW connectivity parameter (M in the paper).
+    /// Higher values give better recall but use more memory.
+    pub connectivity: usize,
+    /// Expansion factor during index construction (ef_construction).
+    pub expansion_add: usize,
+    /// Expansion factor during search (ef_search).
+    pub expansion_search: usize,
+}
+
+impl VectorIndexConfig {
+    /// Creates a new vector index config from VectorIndexOptions.
+    pub fn new(dim: usize, options: &VectorIndexOptions) -> Self {
+        Self {
+            engine: options.engine,
+            dim,
+            metric: distance_metric_to_usearch(options.metric),
+            distance_metric: options.metric,
+            connectivity: options.connectivity as usize,
+            expansion_add: options.expansion_add as usize,
+            expansion_search: options.expansion_search as usize,
+        }
+    }
+}
+
+/// Creator for a single column's vector index.
+struct VectorIndexCreator {
+    /// The vector index engine (e.g., USearch HNSW).
+    engine: Box<dyn VectorIndexEngine>,
+    /// Configuration for this index.
+    config: VectorIndexConfig,
+    /// Bitmap tracking which row offsets have NULL vectors.
+    /// HNSW keys are sequential (0, 1, 2...) but row offsets may have gaps due to NULLs.
+    null_bitmap: RoaringBitmap,
+    /// Current row offset (including NULLs).
+    current_row_offset: u64,
+    /// Next HNSW key to assign (only for non-NULL vectors).
+    next_hnsw_key: u64,
+    /// Memory usage estimation.
+    memory_usage: usize,
+}
+
+impl VectorIndexCreator {
+    /// Creates a new vector index creator.
+    fn new(config: VectorIndexConfig) -> Result<Self> {
+        let engine_instance = engine::create_engine(config.engine, &config)?;
+
+        Ok(Self {
+            engine: engine_instance,
+            config,
+            null_bitmap: RoaringBitmap::new(),
+            current_row_offset: 0,
+            next_hnsw_key: 0,
+            memory_usage: 0,
+        })
+    }
+
+    /// Reserves capacity for the expected number of vectors.
+    #[allow(dead_code)]
+    fn reserve(&mut self, capacity: usize) -> Result<()> {
+        self.engine.reserve(capacity).map_err(|e| {
+            VectorIndexBuildSnafu {
+                reason: format!("Failed to reserve capacity: {}", e),
+            }
+            .build()
+        })
+    }
+
+    /// Adds a vector to the index.
+    /// Returns the HNSW key assigned to this vector.
+    fn add_vector(&mut self, vector: &[f32]) -> Result<u64> {
+        let key = self.next_hnsw_key;
+        self.engine.add(key, vector).map_err(|e| {
+            VectorIndexBuildSnafu {
+                reason: e.to_string(),
+            }
+            .build()
+        })?;
+        self.next_hnsw_key += 1;
+        self.current_row_offset += 1;
+        self.memory_usage = self.engine.memory_usage();
+        Ok(key)
+    }
+
+    /// Records a NULL vector at the current row offset.
+    fn add_null(&mut self) {
+        self.null_bitmap.insert(self.current_row_offset as u32);
+        self.current_row_offset += 1;
+    }
+
+    /// Records multiple NULL vectors starting at the current row offset.
+    fn add_nulls(&mut self, n: usize) {
+        let start = self.current_row_offset as u32;
+        let end = start + n as u32;
+        self.null_bitmap.insert_range(start..end);
+        self.current_row_offset += n as u64;
+    }
+
+    /// Returns the serialized size of the index.
+    fn serialized_length(&self) -> usize {
+        self.engine.serialized_length()
+    }
+
+    /// Serializes the index to a buffer.
+    fn save_to_buffer(&self, buffer: &mut [u8]) -> Result<()> {
+        self.engine.save_to_buffer(buffer).map_err(|e| {
+            VectorIndexFinishSnafu {
+                reason: format!("Failed to serialize index: {}", e),
+            }
+            .build()
+        })
+    }
+
+    /// Returns the memory usage of this creator.
+    fn memory_usage(&self) -> usize {
+        self.memory_usage + self.null_bitmap.serialized_size()
+    }
+
+    /// Returns the number of vectors in the index (excluding NULLs).
+    fn size(&self) -> usize {
+        self.engine.size()
+    }
+
+    /// Returns the engine type.
+    fn engine_type(&self) -> VectorIndexEngineType {
+        self.config.engine
+    }
+
+    /// Returns the distance metric.
+    fn metric(&self) -> VectorDistanceMetric {
+        self.config.distance_metric
+    }
+}
+
+/// The indexer for vector indexes across multiple columns.
+pub struct VectorIndexer {
+    /// Per-column vector index creators.
+    creators: HashMap<ColumnId, VectorIndexCreator>,
+    /// Provider for intermediate files.
+    temp_file_provider: Arc<TempFileProvider>,
+    /// Whether the indexing process has been aborted.
+    aborted: bool,
+    /// Statistics for this indexer.
+    stats: Statistics,
+    /// Global memory usage tracker.
+    #[allow(dead_code)]
+    global_memory_usage: Arc<AtomicUsize>,
+    /// Region metadata for column lookups.
+    #[allow(dead_code)]
+    metadata: RegionMetadataRef,
+    /// Memory usage threshold.
+    memory_usage_threshold: Option<usize>,
+}
+
+impl VectorIndexer {
+    /// Creates a new vector indexer.
+    ///
+    /// Returns `None` if there are no vector columns that need indexing.
+    pub fn new(
+        sst_file_id: FileId,
+        metadata: &RegionMetadataRef,
+        intermediate_manager: IntermediateManager,
+        memory_usage_threshold: Option<usize>,
+        vector_index_options: &HashMap<ColumnId, VectorIndexOptions>,
+    ) -> Result<Option<Self>> {
+        let mut creators = HashMap::new();
+
+        let temp_file_provider = Arc::new(TempFileProvider::new(
+            IntermediateLocation::new(&metadata.region_id, &sst_file_id),
+            intermediate_manager,
+        ));
+        let global_memory_usage = Arc::new(AtomicUsize::new(0));
+
+        // Find all vector columns that have vector index enabled
+        for column in &metadata.column_metadatas {
+            // Check if this column has vector index options configured
+            let Some(options) = vector_index_options.get(&column.column_id) else {
+                continue;
+            };
+
+            // Verify the column is a vector type
+            let ConcreteDataType::Vector(vector_type) = &column.column_schema.data_type else {
+                continue;
+            };
+
+            let config = VectorIndexConfig::new(vector_type.dim as usize, options);
+            let creator = VectorIndexCreator::new(config)?;
+            creators.insert(column.column_id, creator);
+        }
+
+        if creators.is_empty() {
+            return Ok(None);
+        }
+
+        let indexer = Self {
+            creators,
+            temp_file_provider,
+            aborted: false,
+            stats: Statistics::new(TYPE_VECTOR_INDEX),
+            global_memory_usage,
+            metadata: metadata.clone(),
+            memory_usage_threshold,
+        };
+
+        Ok(Some(indexer))
+    }
+
+    /// Updates index with a batch of rows.
+    /// Garbage will be cleaned up if failed to update.
+    pub async fn update(&mut self, batch: &mut Batch) -> Result<()> {
+        ensure!(!self.aborted, OperateAbortedIndexSnafu);
+
+        if self.creators.is_empty() {
+            return Ok(());
+        }
+
+        if let Err(update_err) = self.do_update(batch).await {
+            // Clean up garbage if failed to update
+            if let Err(err) = self.do_cleanup().await {
+                if cfg!(any(test, feature = "test")) {
+                    panic!("Failed to clean up vector index creator, err: {err:?}");
+                } else {
+                    warn!(err; "Failed to clean up vector index creator");
+                }
+            }
+            return Err(update_err);
+        }
+
+        Ok(())
+    }
+
+    /// Updates index with a flat format `RecordBatch`.
+    /// Garbage will be cleaned up if failed to update.
+    pub async fn update_flat(&mut self, batch: &RecordBatch) -> Result<()> {
+        ensure!(!self.aborted, OperateAbortedIndexSnafu);
+
+        if self.creators.is_empty() || batch.num_rows() == 0 {
+            return Ok(());
+        }
+
+        if let Err(update_err) = self.do_update_flat(batch).await {
+            // Clean up garbage if failed to update
+            if let Err(err) = self.do_cleanup().await {
+                if cfg!(any(test, feature = "test")) {
+                    panic!("Failed to clean up vector index creator, err: {err:?}");
+                } else {
+                    warn!(err; "Failed to clean up vector index creator");
+                }
+            }
+            return Err(update_err);
+        }
+
+        Ok(())
+    }
+
+    /// Internal update implementation.
+    async fn do_update(&mut self, batch: &mut Batch) -> Result<()> {
+        let mut guard = self.stats.record_update();
+        let n = batch.num_rows();
+        guard.inc_row_count(n);
+
+        for (col_id, creator) in &mut self.creators {
+            let Some(values) = batch.field_col_value(*col_id) else {
+                continue;
+            };
+
+            // Process each row in the batch
+            for i in 0..n {
+                let value = values.data.get_ref(i);
+                if value.is_null() {
+                    creator.add_null();
+                } else {
+                    // Extract the vector bytes and convert to f32 slice
+                    if let ValueRef::Binary(bytes) = value {
+                        let floats = bytes_to_f32_slice(bytes);
+                        if floats.len() != creator.config.dim {
+                            return VectorIndexBuildSnafu {
+                                reason: format!(
+                                    "Vector dimension mismatch: expected {}, got {}",
+                                    creator.config.dim,
+                                    floats.len()
+                                ),
+                            }
+                            .fail();
+                        }
+                        creator.add_vector(&floats)?;
+                    } else {
+                        creator.add_null();
+                    }
+                }
+            }
+
+            // Check memory limit - abort index creation if exceeded
+            if let Some(threshold) = self.memory_usage_threshold {
+                let current_usage = creator.memory_usage();
+                if current_usage > threshold {
+                    warn!(
+                        "Vector index memory usage {} exceeds threshold {}, aborting index creation, region_id: {}",
+                        current_usage, threshold, self.metadata.region_id
+                    );
+                    return VectorIndexBuildSnafu {
+                        reason: format!(
+                            "Memory usage {} exceeds threshold {}",
+                            current_usage, threshold
+                        ),
+                    }
+                    .fail();
+                }
+            }
+        }
+
+        Ok(())
+    }
+
+    /// Internal flat update implementation.
+    async fn do_update_flat(&mut self, batch: &RecordBatch) -> Result<()> {
+        let mut guard = self.stats.record_update();
+        let n = batch.num_rows();
+        guard.inc_row_count(n);
+
+        for (col_id, creator) in &mut self.creators {
+            // This should never happen: creator exists but column not in metadata
+            let column_meta = self.metadata.column_by_id(*col_id).ok_or_else(|| {
+                VectorIndexBuildSnafu {
+                    reason: format!(
+                        "Column {} not found in region metadata, this is a bug",
+                        col_id
+                    ),
+                }
+                .build()
+            })?;
+
+            let column_name = &column_meta.column_schema.name;
+            // Column not in batch is normal for flat format - treat as NULLs
+            let Some(column_array) = batch.column_by_name(column_name) else {
+                creator.add_nulls(n);
+                continue;
+            };
+
+            // Vector type must be stored as binary array
+            let binary_array = column_array
+                .as_any()
+                .downcast_ref::<BinaryArray>()
+                .ok_or_else(|| {
+                    VectorIndexBuildSnafu {
+                        reason: format!(
+                            "Column {} is not a binary array, got {:?}",
+                            column_name,
+                            column_array.data_type()
+                        ),
+                    }
+                    .build()
+                })?;
+
+            for i in 0..n {
+                if !binary_array.is_valid(i) {
+                    creator.add_null();
+                } else {
+                    let bytes = binary_array.value(i);
+                    let floats = bytes_to_f32_slice(bytes);
+                    if floats.len() != creator.config.dim {
+                        return VectorIndexBuildSnafu {
+                            reason: format!(
+                                "Vector dimension mismatch: expected {}, got {}",
+                                creator.config.dim,
+                                floats.len()
+                            ),
+                        }
+                        .fail();
+                    }
+                    creator.add_vector(&floats)?;
+                }
+            }
+
+            if let Some(threshold) = self.memory_usage_threshold {
+                let current_usage = creator.memory_usage();
+                if current_usage > threshold {
+                    warn!(
+                        "Vector index memory usage {} exceeds threshold {}, aborting index creation, region_id: {}",
+                        current_usage, threshold, self.metadata.region_id
+                    );
+                    return VectorIndexBuildSnafu {
+                        reason: format!(
+                            "Memory usage {} exceeds threshold {}",
+                            current_usage, threshold
+                        ),
+                    }
+                    .fail();
+                }
+            }
+        }
+
+        Ok(())
+    }
+
+    /// Finishes index creation and writes to puffin.
+    /// Returns the number of rows and bytes written.
+    pub async fn finish(
+        &mut self,
+        puffin_writer: &mut SstPuffinWriter,
+    ) -> Result<(RowCount, ByteCount)> {
+        ensure!(!self.aborted, OperateAbortedIndexSnafu);
+
+        if self.stats.row_count() == 0 {
+            // No IO is performed, no garbage to clean up
+            return Ok((0, 0));
+        }
+
+        let finish_res = self.do_finish(puffin_writer).await;
+        // Clean up garbage no matter finish successfully or not
+        if let Err(err) = self.do_cleanup().await {
+            if cfg!(any(test, feature = "test")) {
+                panic!("Failed to clean up vector index creator, err: {err:?}");
+            } else {
+                warn!(err; "Failed to clean up vector index creator");
+            }
+        }
+
+        // Report metrics on successful finish
+        if finish_res.is_ok() {
+            INDEX_CREATE_ROWS_TOTAL
+                .with_label_values(&[TYPE_VECTOR_INDEX])
+                .inc_by(self.stats.row_count() as u64);
+            INDEX_CREATE_BYTES_TOTAL
+                .with_label_values(&[TYPE_VECTOR_INDEX])
+                .inc_by(self.stats.byte_count());
+        }
+
+        finish_res.map(|_| (self.stats.row_count(), self.stats.byte_count()))
+    }
+
+    /// Internal finish implementation.
+    async fn do_finish(&mut self, puffin_writer: &mut SstPuffinWriter) -> Result<()> {
+        let mut guard = self.stats.record_finish();
+
+        for (id, creator) in &mut self.creators {
+            if creator.size() == 0 {
+                // No vectors to index
+                continue;
+            }
+
+            let written_bytes = Self::do_finish_single_creator(*id, creator, puffin_writer).await?;
+            guard.inc_byte_count(written_bytes);
+        }
+
+        Ok(())
+    }
+
+    /// Finishes a single column's vector index.
+    ///
+    /// The blob format v1 (header = 33 bytes):
+    /// ```text
+    /// +------------------+
+    /// | Version          | 1 byte (u8, = 1)
+    /// +------------------+
+    /// | Engine type      | 1 byte (u8, engine identifier)
+    /// +------------------+
+    /// | Dimension        | 4 bytes (u32, little-endian)
+    /// +------------------+
+    /// | Metric           | 1 byte (u8, distance metric)
+    /// +------------------+
+    /// | Connectivity     | 2 bytes (u16, little-endian, HNSW M parameter)
+    /// +------------------+
+    /// | Expansion add    | 2 bytes (u16, little-endian, ef_construction)
+    /// +------------------+
+    /// | Expansion search | 2 bytes (u16, little-endian, ef_search)
+    /// +------------------+
+    /// | Total rows       | 8 bytes (u64, little-endian, total rows in SST)
+    /// +------------------+
+    /// | Indexed rows     | 8 bytes (u64, little-endian, non-NULL rows indexed)
+    /// +------------------+
+    /// | NULL bitmap len  | 4 bytes (u32, little-endian)
+    /// +------------------+
+    /// | NULL bitmap      | variable length (serialized RoaringBitmap)
+    /// +------------------+
+    /// | Vector index     | variable length (engine-specific serialized format)
+    /// +------------------+
+    /// ```
+    async fn do_finish_single_creator(
+        col_id: ColumnId,
+        creator: &mut VectorIndexCreator,
+        puffin_writer: &mut SstPuffinWriter,
+    ) -> Result<ByteCount> {
+        // Serialize the NULL bitmap
+        let mut null_bitmap_bytes = Vec::new();
+        creator
+            .null_bitmap
+            .serialize_into(&mut null_bitmap_bytes)
+            .map_err(|e| {
+                VectorIndexFinishSnafu {
+                    reason: format!("Failed to serialize NULL bitmap: {}", e),
+                }
+                .build()
+            })?;
+
+        // Serialize the vector index
+        let index_size = creator.serialized_length();
+        let mut index_bytes = vec![0u8; index_size];
+        creator.save_to_buffer(&mut index_bytes)?;
+
+        // Header size: version(1) + engine(1) + dim(4) + metric(1) +
+        //              connectivity(2) + expansion_add(2) + expansion_search(2) +
+        //              total_rows(8) + indexed_rows(8) + bitmap_len(4) = 33 bytes
+        /// Size of the vector index blob header in bytes.
+        /// Header format: version(1) + engine(1) + dim(4) + metric(1) +
+        /// connectivity(2) + expansion_add(2) + expansion_search(2) +
+        /// total_rows(8) + indexed_rows(8) + bitmap_len(4) = 33 bytes
+        const VECTOR_INDEX_BLOB_HEADER_SIZE: usize = 33;
+        let total_size =
+            VECTOR_INDEX_BLOB_HEADER_SIZE + null_bitmap_bytes.len() + index_bytes.len();
+        let mut blob_data = Vec::with_capacity(total_size);
+
+        // Write version (1 byte)
+        blob_data.push(1u8);
+        // Write engine type (1 byte)
+        blob_data.push(creator.engine_type().as_u8());
+        // Write dimension (4 bytes, little-endian)
+        blob_data.extend_from_slice(&(creator.config.dim as u32).to_le_bytes());
+        // Write metric (1 byte)
+        blob_data.push(creator.metric().as_u8());
+        // Write connectivity/M (2 bytes, little-endian)
+        blob_data.extend_from_slice(&(creator.config.connectivity as u16).to_le_bytes());
+        // Write expansion_add/ef_construction (2 bytes, little-endian)
+        blob_data.extend_from_slice(&(creator.config.expansion_add as u16).to_le_bytes());
+        // Write expansion_search/ef_search (2 bytes, little-endian)
+        blob_data.extend_from_slice(&(creator.config.expansion_search as u16).to_le_bytes());
+        // Write total_rows (8 bytes, little-endian)
+        blob_data.extend_from_slice(&creator.current_row_offset.to_le_bytes());
+        // Write indexed_rows (8 bytes, little-endian)
+        blob_data.extend_from_slice(&creator.next_hnsw_key.to_le_bytes());
+        // Write NULL bitmap length (4 bytes, little-endian)
+        let bitmap_len: u32 = null_bitmap_bytes.len().try_into().map_err(|_| {
+            VectorIndexBuildSnafu {
+                reason: format!(
+                    "NULL bitmap size {} exceeds maximum allowed size {}",
+                    null_bitmap_bytes.len(),
+                    u32::MAX
+                ),
+            }
+            .build()
+        })?;
+        blob_data.extend_from_slice(&bitmap_len.to_le_bytes());
+        // Write NULL bitmap
+        blob_data.extend_from_slice(&null_bitmap_bytes);
+        // Write vector index
+        blob_data.extend_from_slice(&index_bytes);
+
+        // Create blob name following the same pattern as bloom filter
+        let blob_name = format!("{}-{}", INDEX_BLOB_TYPE, col_id);
+
+        // Write to puffin using a pipe
+        let (tx, rx) = tokio::io::duplex(PIPE_BUFFER_SIZE_FOR_SENDING_BLOB);
+
+        // Writer task writes the blob data to the pipe
+        let write_index = async move {
+            use tokio::io::AsyncWriteExt;
+            let mut writer = tx;
+            writer.write_all(&blob_data).await?;
+            writer.shutdown().await?;
+            Ok::<(), std::io::Error>(())
+        };
+
+        let (index_write_result, puffin_add_blob) = futures::join!(
+            write_index,
+            puffin_writer.put_blob(
+                &blob_name,
+                rx.compat(),
+                PutOptions::default(),
+                Default::default()
+            )
+        );
+
+        match (
+            puffin_add_blob.context(PuffinAddBlobSnafu),
+            index_write_result.map_err(|e| {
+                VectorIndexFinishSnafu {
+                    reason: format!("Failed to write blob data: {}", e),
+                }
+                .build()
+            }),
+        ) {
+            (Err(e1), Err(e2)) => BiErrorsSnafu {
+                first: Box::new(e1),
+                second: Box::new(e2),
+            }
+            .fail()?,
+
+            (Ok(_), e @ Err(_)) => e?,
+            (e @ Err(_), Ok(_)) => e.map(|_| ())?,
+            (Ok(written_bytes), Ok(_)) => {
+                return Ok(written_bytes);
+            }
+        }
+
+        Ok(0)
+    }
+
+    /// Aborts index creation and cleans up garbage.
+    pub async fn abort(&mut self) -> Result<()> {
+        if self.aborted {
+            return Ok(());
+        }
+        self.aborted = true;
+        self.do_cleanup().await
+    }
+
+    /// Cleans up temporary files.
+    async fn do_cleanup(&mut self) -> Result<()> {
+        let mut _guard = self.stats.record_cleanup();
+        self.creators.clear();
+        self.temp_file_provider.cleanup().await
+    }
+
+    /// Returns the memory usage of the indexer.
+    pub fn memory_usage(&self) -> usize {
+        self.creators.values().map(|c| c.memory_usage()).sum()
+    }
+
+    /// Returns the column IDs being indexed.
+    pub fn column_ids(&self) -> impl Iterator<Item = ColumnId> + '_ {
+        self.creators.keys().copied()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_vector_index_creator() {
+        let options = VectorIndexOptions::default();
+        let config = VectorIndexConfig::new(4, &options);
+        let mut creator = VectorIndexCreator::new(config).unwrap();
+
+        creator.reserve(10).unwrap();
+
+        // Add some vectors
+        let v1 = vec![1.0f32, 0.0, 0.0, 0.0];
+        let v2 = vec![0.0f32, 1.0, 0.0, 0.0];
+
+        creator.add_vector(&v1).unwrap();
+        creator.add_null();
+        creator.add_vector(&v2).unwrap();
+
+        assert_eq!(creator.size(), 2); // 2 vectors (excluding NULL)
+        assert_eq!(creator.current_row_offset, 3); // 3 rows total
+        assert!(creator.null_bitmap.contains(1)); // Row 1 is NULL
+    }
+
+    #[test]
+    fn test_vector_index_creator_serialization() {
+        let options = VectorIndexOptions::default();
+        let config = VectorIndexConfig::new(4, &options);
+        let mut creator = VectorIndexCreator::new(config).unwrap();
+
+        creator.reserve(10).unwrap();
+
+        // Add vectors
+        let vectors = vec![
+            vec![1.0f32, 0.0, 0.0, 0.0],
+            vec![0.0f32, 1.0, 0.0, 0.0],
+            vec![0.0f32, 0.0, 1.0, 0.0],
+        ];
+
+        for v in &vectors {
+            creator.add_vector(v).unwrap();
+        }
+
+        // Test serialization
+        let size = creator.serialized_length();
+        assert!(size > 0);
+
+        let mut buffer = vec![0u8; size];
+        creator.save_to_buffer(&mut buffer).unwrap();
+
+        // Verify buffer is not empty and starts with some data
+        assert!(!buffer.iter().all(|&b| b == 0));
+    }
+
+    #[test]
+    fn test_vector_index_creator_null_bitmap_serialization() {
+        let options = VectorIndexOptions::default();
+        let config = VectorIndexConfig::new(4, &options);
+        let mut creator = VectorIndexCreator::new(config).unwrap();
+
+        creator.reserve(10).unwrap();
+
+        // Add pattern: vector, null, vector, null, null, vector
+        creator.add_vector(&[1.0, 0.0, 0.0, 0.0]).unwrap();
+        creator.add_null();
+        creator.add_vector(&[0.0, 1.0, 0.0, 0.0]).unwrap();
+        creator.add_nulls(2);
+        creator.add_vector(&[0.0, 0.0, 1.0, 0.0]).unwrap();
+
+        assert_eq!(creator.size(), 3); // 3 vectors
+        assert_eq!(creator.current_row_offset, 6); // 6 rows total
+        assert!(!creator.null_bitmap.contains(0));
+        assert!(creator.null_bitmap.contains(1));
+        assert!(!creator.null_bitmap.contains(2));
+        assert!(creator.null_bitmap.contains(3));
+        assert!(creator.null_bitmap.contains(4));
+        assert!(!creator.null_bitmap.contains(5));
+
+        // Test NULL bitmap serialization
+        let mut bitmap_bytes = Vec::new();
+        creator
+            .null_bitmap
+            .serialize_into(&mut bitmap_bytes)
+            .unwrap();
+
+        // Deserialize and verify
+        let restored = RoaringBitmap::deserialize_from(&bitmap_bytes[..]).unwrap();
+        assert_eq!(restored.len(), 3); // 3 NULLs
+        assert!(restored.contains(1));
+        assert!(restored.contains(3));
+        assert!(restored.contains(4));
+    }
+
+    #[test]
+    fn test_vector_index_config() {
+        use index::vector::VectorDistanceMetric;
+
+        let options = VectorIndexOptions {
+            engine: VectorIndexEngineType::default(),
+            metric: VectorDistanceMetric::Cosine,
+            connectivity: 32,
+            expansion_add: 256,
+            expansion_search: 128,
+        };
+        let config = VectorIndexConfig::new(128, &options);
+        assert_eq!(config.engine, VectorIndexEngineType::Usearch);
+        assert_eq!(config.dim, 128);
+        assert_eq!(config.metric, MetricKind::Cos);
+        assert_eq!(config.connectivity, 32);
+        assert_eq!(config.expansion_add, 256);
+        assert_eq!(config.expansion_search, 128);
+    }
+
+    #[test]
+    fn test_vector_index_header_format() {
+        use index::vector::VectorDistanceMetric;
+
+        // Create config with specific HNSW parameters
+        let options = VectorIndexOptions {
+            engine: VectorIndexEngineType::Usearch,
+            metric: VectorDistanceMetric::L2sq,
+            connectivity: 24,
+            expansion_add: 200,
+            expansion_search: 100,
+        };
+        let config = VectorIndexConfig::new(4, &options);
+        let mut creator = VectorIndexCreator::new(config).unwrap();
+
+        creator.reserve(10).unwrap();
+
+        // Add pattern: vector, null, vector, null, vector
+        creator.add_vector(&[1.0, 0.0, 0.0, 0.0]).unwrap();
+        creator.add_null();
+        creator.add_vector(&[0.0, 1.0, 0.0, 0.0]).unwrap();
+        creator.add_null();
+        creator.add_vector(&[0.0, 0.0, 1.0, 0.0]).unwrap();
+
+        // Verify counts
+        assert_eq!(creator.current_row_offset, 5); // total_rows
+        assert_eq!(creator.next_hnsw_key, 3); // indexed_rows
+
+        // Build blob data manually (simulating write_to_puffin header writing)
+        let mut null_bitmap_bytes = Vec::new();
+        creator
+            .null_bitmap
+            .serialize_into(&mut null_bitmap_bytes)
+            .unwrap();
+
+        let index_size = creator.serialized_length();
+        let mut index_bytes = vec![0u8; index_size];
+        creator.save_to_buffer(&mut index_bytes).unwrap();
+
+        // Header: 33 bytes
+        let header_size = 33;
+        let total_size = header_size + null_bitmap_bytes.len() + index_bytes.len();
+        let mut blob_data = Vec::with_capacity(total_size);
+
+        // Write header fields
+        blob_data.push(1u8); // version
+        blob_data.push(creator.engine_type().as_u8()); // engine type
+        blob_data.extend_from_slice(&(creator.config.dim as u32).to_le_bytes()); // dimension
+        blob_data.push(creator.metric().as_u8()); // metric
+        blob_data.extend_from_slice(&(creator.config.connectivity as u16).to_le_bytes());
+        blob_data.extend_from_slice(&(creator.config.expansion_add as u16).to_le_bytes());
+        blob_data.extend_from_slice(&(creator.config.expansion_search as u16).to_le_bytes());
+        blob_data.extend_from_slice(&creator.current_row_offset.to_le_bytes()); // total_rows
+        blob_data.extend_from_slice(&creator.next_hnsw_key.to_le_bytes()); // indexed_rows
+        let bitmap_len: u32 = null_bitmap_bytes.len().try_into().unwrap();
+        blob_data.extend_from_slice(&bitmap_len.to_le_bytes());
+        blob_data.extend_from_slice(&null_bitmap_bytes);
+        blob_data.extend_from_slice(&index_bytes);
+
+        // Verify header size
+        assert_eq!(blob_data.len(), total_size);
+
+        // Parse header and verify values
+        assert_eq!(blob_data[0], 1); // version
+        assert_eq!(blob_data[1], VectorIndexEngineType::Usearch.as_u8()); // engine
+
+        let dim = u32::from_le_bytes([blob_data[2], blob_data[3], blob_data[4], blob_data[5]]);
+        assert_eq!(dim, 4);
+
+        let metric = blob_data[6];
+        assert_eq!(
+            metric,
+            datatypes::schema::VectorDistanceMetric::L2sq.as_u8()
+        );
+
+        let connectivity = u16::from_le_bytes([blob_data[7], blob_data[8]]);
+        assert_eq!(connectivity, 24);
+
+        let expansion_add = u16::from_le_bytes([blob_data[9], blob_data[10]]);
+        assert_eq!(expansion_add, 200);
+
+        let expansion_search = u16::from_le_bytes([blob_data[11], blob_data[12]]);
+        assert_eq!(expansion_search, 100);
+
+        let total_rows = u64::from_le_bytes([
+            blob_data[13],
+            blob_data[14],
+            blob_data[15],
+            blob_data[16],
+            blob_data[17],
+            blob_data[18],
+            blob_data[19],
+            blob_data[20],
+        ]);
+        assert_eq!(total_rows, 5);
+
+        let indexed_rows = u64::from_le_bytes([
+            blob_data[21],
+            blob_data[22],
+            blob_data[23],
+            blob_data[24],
+            blob_data[25],
+            blob_data[26],
+            blob_data[27],
+            blob_data[28],
+        ]);
+        assert_eq!(indexed_rows, 3);
+
+        let null_bitmap_len =
+            u32::from_le_bytes([blob_data[29], blob_data[30], blob_data[31], blob_data[32]]);
+        assert_eq!(null_bitmap_len as usize, null_bitmap_bytes.len());
+
+        // Verify null bitmap can be deserialized
+        let null_bitmap_data = &blob_data[header_size..header_size + null_bitmap_len as usize];
+        let restored_bitmap = RoaringBitmap::deserialize_from(null_bitmap_data).unwrap();
+        assert_eq!(restored_bitmap.len(), 2); // 2 nulls
+        assert!(restored_bitmap.contains(1));
+        assert!(restored_bitmap.contains(3));
+    }
+}
--- a/src/mito2/src/sst/index/vector_index/engine/mod.rs
+++ b/src/mito2/src/sst/index/vector_index/engine/mod.rs
@@ -0,0 +1,45 @@
+// 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.
+
+//! Pluggable vector index engine implementations.
+
+mod usearch_impl;
+
+use store_api::storage::{VectorIndexEngine, VectorIndexEngineType};
+pub use usearch_impl::UsearchEngine;
+
+use crate::error::Result;
+use crate::sst::index::vector_index::creator::VectorIndexConfig;
+
+/// Creates a new vector index engine based on the engine type.
+pub fn create_engine(
+    engine_type: VectorIndexEngineType,
+    config: &VectorIndexConfig,
+) -> Result<Box<dyn VectorIndexEngine>> {
+    match engine_type {
+        VectorIndexEngineType::Usearch => Ok(Box::new(UsearchEngine::create(config)?)),
+    }
+}
+
+/// Loads a vector index engine from serialized data.
+#[allow(unused)]
+pub fn load_engine(
+    engine_type: VectorIndexEngineType,
+    config: &VectorIndexConfig,
+    data: &[u8],
+) -> Result<Box<dyn VectorIndexEngine>> {
+    match engine_type {
+        VectorIndexEngineType::Usearch => Ok(Box::new(UsearchEngine::load(config, data)?)),
+    }
+}
--- a/src/mito2/src/sst/index/vector_index/engine/usearch_impl.rs
+++ b/src/mito2/src/sst/index/vector_index/engine/usearch_impl.rs
@@ -0,0 +1,231 @@
+// 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.
+
+//! USearch HNSW implementation of VectorIndexEngine.
+
+use common_error::ext::BoxedError;
+use store_api::storage::{VectorIndexEngine, VectorSearchMatches};
+use usearch::{Index, IndexOptions, ScalarKind};
+
+use crate::error::{Result, VectorIndexBuildSnafu};
+use crate::sst::index::vector_index::creator::VectorIndexConfig;
+
+/// USearch-based vector index engine using HNSW algorithm.
+pub struct UsearchEngine {
+    index: Index,
+}
+
+impl UsearchEngine {
+    /// Creates a new USearch engine with the given configuration.
+    pub fn create(config: &VectorIndexConfig) -> Result<Self> {
+        let options = IndexOptions {
+            dimensions: config.dim,
+            metric: config.metric,
+            quantization: ScalarKind::F32,
+            connectivity: config.connectivity,
+            expansion_add: config.expansion_add,
+            expansion_search: config.expansion_search,
+            multi: false,
+        };
+
+        let index = Index::new(&options).map_err(|e| {
+            VectorIndexBuildSnafu {
+                reason: format!("Failed to create USearch index: {}", e),
+            }
+            .build()
+        })?;
+
+        Ok(Self { index })
+    }
+
+    /// Loads a USearch engine from serialized data.
+    #[allow(unused)]
+    pub fn load(config: &VectorIndexConfig, data: &[u8]) -> Result<Self> {
+        let options = IndexOptions {
+            dimensions: config.dim,
+            metric: config.metric,
+            quantization: ScalarKind::F32,
+            // These will be loaded from serialized data
+            connectivity: 0,
+            expansion_add: 0,
+            expansion_search: 0,
+            multi: false,
+        };
+
+        let index = Index::new(&options).map_err(|e| {
+            VectorIndexBuildSnafu {
+                reason: format!("Failed to create USearch index for loading: {}", e),
+            }
+            .build()
+        })?;
+
+        index.load_from_buffer(data).map_err(|e| {
+            VectorIndexBuildSnafu {
+                reason: format!("Failed to load USearch index from buffer: {}", e),
+            }
+            .build()
+        })?;
+
+        Ok(Self { index })
+    }
+}
+
+impl VectorIndexEngine for UsearchEngine {
+    fn add(&mut self, key: u64, vector: &[f32]) -> Result<(), BoxedError> {
+        // Reserve capacity if needed
+        if self.index.size() >= self.index.capacity() {
+            let new_capacity = std::cmp::max(1, self.index.capacity() * 2);
+            self.index.reserve(new_capacity).map_err(|e| {
+                BoxedError::new(
+                    VectorIndexBuildSnafu {
+                        reason: format!("Failed to reserve capacity: {}", e),
+                    }
+                    .build(),
+                )
+            })?;
+        }
+
+        self.index.add(key, vector).map_err(|e| {
+            BoxedError::new(
+                VectorIndexBuildSnafu {
+                    reason: format!("Failed to add vector: {}", e),
+                }
+                .build(),
+            )
+        })
+    }
+
+    fn search(&self, query: &[f32], k: usize) -> Result<VectorSearchMatches, BoxedError> {
+        let matches = self.index.search(query, k).map_err(|e| {
+            BoxedError::new(
+                VectorIndexBuildSnafu {
+                    reason: format!("Failed to search: {}", e),
+                }
+                .build(),
+            )
+        })?;
+
+        Ok(VectorSearchMatches {
+            keys: matches.keys,
+            distances: matches.distances,
+        })
+    }
+
+    fn serialized_length(&self) -> usize {
+        self.index.serialized_length()
+    }
+
+    fn save_to_buffer(&self, buffer: &mut [u8]) -> Result<(), BoxedError> {
+        self.index.save_to_buffer(buffer).map_err(|e| {
+            BoxedError::new(
+                VectorIndexBuildSnafu {
+                    reason: format!("Failed to save to buffer: {}", e),
+                }
+                .build(),
+            )
+        })
+    }
+
+    fn reserve(&mut self, capacity: usize) -> Result<(), BoxedError> {
+        self.index.reserve(capacity).map_err(|e| {
+            BoxedError::new(
+                VectorIndexBuildSnafu {
+                    reason: format!("Failed to reserve: {}", e),
+                }
+                .build(),
+            )
+        })
+    }
+
+    fn size(&self) -> usize {
+        self.index.size()
+    }
+
+    fn capacity(&self) -> usize {
+        self.index.capacity()
+    }
+
+    fn memory_usage(&self) -> usize {
+        self.index.memory_usage()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use index::vector::VectorDistanceMetric;
+    use store_api::storage::VectorIndexEngineType;
+    use usearch::MetricKind;
+
+    use super::*;
+
+    fn test_config() -> VectorIndexConfig {
+        VectorIndexConfig {
+            engine: VectorIndexEngineType::Usearch,
+            dim: 4,
+            metric: MetricKind::L2sq,
+            distance_metric: VectorDistanceMetric::L2sq,
+            connectivity: 16,
+            expansion_add: 128,
+            expansion_search: 64,
+        }
+    }
+
+    #[test]
+    fn test_usearch_engine_create() {
+        let config = test_config();
+        let engine = UsearchEngine::create(&config).unwrap();
+        assert_eq!(engine.size(), 0);
+    }
+
+    #[test]
+    fn test_usearch_engine_add_and_search() {
+        let config = test_config();
+        let mut engine = UsearchEngine::create(&config).unwrap();
+
+        // Add some vectors
+        engine.add(0, &[1.0, 0.0, 0.0, 0.0]).unwrap();
+        engine.add(1, &[0.0, 1.0, 0.0, 0.0]).unwrap();
+        engine.add(2, &[0.0, 0.0, 1.0, 0.0]).unwrap();
+
+        assert_eq!(engine.size(), 3);
+
+        // Search
+        let matches = engine.search(&[1.0, 0.0, 0.0, 0.0], 2).unwrap();
+        assert_eq!(matches.keys.len(), 2);
+        // First result should be the exact match (key 0)
+        assert_eq!(matches.keys[0], 0);
+    }
+
+    #[test]
+    fn test_usearch_engine_serialization() {
+        let config = test_config();
+        let mut engine = UsearchEngine::create(&config).unwrap();
+
+        engine.add(0, &[1.0, 0.0, 0.0, 0.0]).unwrap();
+        engine.add(1, &[0.0, 1.0, 0.0, 0.0]).unwrap();
+
+        // Serialize
+        let len = engine.serialized_length();
+        let mut buffer = vec![0u8; len];
+        engine.save_to_buffer(&mut buffer).unwrap();
+
+        // Load
+        let loaded = UsearchEngine::load(&config, &buffer).unwrap();
+        assert_eq!(loaded.size(), 2);
+
+        // Verify search works on loaded index
+        let matches = loaded.search(&[1.0, 0.0, 0.0, 0.0], 1).unwrap();
+        assert_eq!(matches.keys[0], 0);
+    }
+}
--- a/src/mito2/src/sst/index/vector_index/mod.rs
+++ b/src/mito2/src/sst/index/vector_index/mod.rs
@@ -0,0 +1,22 @@
+// 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.
+
+//! Vector index module for HNSW-based approximate nearest neighbor search.
+
+pub(crate) mod creator;
+pub(crate) mod engine;
+pub(crate) mod util;
+
+/// The blob type identifier for vector index in puffin files.
+pub(crate) const INDEX_BLOB_TYPE: &str = "greptime-vector-index-v1";
--- a/src/mito2/src/sst/index/vector_index/util.rs
+++ b/src/mito2/src/sst/index/vector_index/util.rs
@@ -0,0 +1,108 @@
+// 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.
+
+//! Utility functions for vector index operations.
+
+use std::borrow::Cow;
+
+/// Converts a byte slice (little-endian format) to f32 slice, handling unaligned data gracefully.
+/// Returns `Cow::Borrowed` for aligned data on little-endian systems (zero-copy)
+/// or `Cow::Owned` for unaligned data or big-endian systems.
+///
+/// # Panics
+///
+/// Panics if the byte slice length is not a multiple of 4.
+pub fn bytes_to_f32_slice(bytes: &[u8]) -> Cow<'_, [f32]> {
+    assert!(
+        bytes.len().is_multiple_of(4),
+        "Vector bytes length {} is not a multiple of 4",
+        bytes.len()
+    );
+
+    if bytes.is_empty() {
+        return Cow::Borrowed(&[]);
+    }
+
+    let ptr = bytes.as_ptr();
+    // Fast path: zero-copy only when data is aligned AND we're on little-endian system
+    // (since vector data is stored in little-endian format)
+    #[cfg(target_endian = "little")]
+    if (ptr as usize).is_multiple_of(std::mem::align_of::<f32>()) {
+        // Safety: We've verified alignment and length requirements,
+        // and on little-endian systems the byte representation matches f32 layout
+        return Cow::Borrowed(unsafe {
+            std::slice::from_raw_parts(ptr as *const f32, bytes.len() / 4)
+        });
+    }
+
+    // Slow path: data is not aligned or we're on big-endian system
+    let floats: Vec<f32> = bytes
+        .chunks_exact(4)
+        .map(|chunk| f32::from_le_bytes([chunk[0], chunk[1], chunk[2], chunk[3]]))
+        .collect();
+    Cow::Owned(floats)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_bytes_to_f32_slice() {
+        let floats = [1.0f32, 2.0, 3.0, 4.0];
+        let bytes: Vec<u8> = floats.iter().flat_map(|f| f.to_le_bytes()).collect();
+
+        let result = bytes_to_f32_slice(&bytes);
+        assert_eq!(result.len(), 4);
+        assert_eq!(result[0], 1.0);
+        assert_eq!(result[1], 2.0);
+        assert_eq!(result[2], 3.0);
+        assert_eq!(result[3], 4.0);
+    }
+
+    #[test]
+    fn test_bytes_to_f32_slice_unaligned() {
+        // Create a buffer with an extra byte at the start to force misalignment
+        let floats = [1.0f32, 2.0, 3.0, 4.0];
+        let mut bytes: Vec<u8> = vec![0u8]; // padding byte
+        bytes.extend(floats.iter().flat_map(|f| f.to_le_bytes()));
+
+        // Take a slice starting at offset 1 (unaligned)
+        let unaligned_bytes = &bytes[1..];
+
+        // Verify it's actually unaligned
+        let ptr = unaligned_bytes.as_ptr();
+        let is_aligned = (ptr as usize).is_multiple_of(std::mem::align_of::<f32>());
+
+        // The function should work regardless of alignment
+        let result = bytes_to_f32_slice(unaligned_bytes);
+        assert_eq!(result.len(), 4);
+        assert_eq!(result[0], 1.0);
+        assert_eq!(result[1], 2.0);
+        assert_eq!(result[2], 3.0);
+        assert_eq!(result[3], 4.0);
+
+        // If it was unaligned, it should return an owned Vec (Cow::Owned)
+        if !is_aligned {
+            assert!(matches!(result, Cow::Owned(_)));
+        }
+    }
+
+    #[test]
+    fn test_bytes_to_f32_slice_empty() {
+        let bytes: &[u8] = &[];
+        let result = bytes_to_f32_slice(bytes);
+        assert!(result.is_empty());
+    }
+}
--- a/src/mito2/src/sst/parquet.rs
+++ b/src/mito2/src/sst/parquet.rs
@@ -742,6 +742,8 @@ mod tests {
            inverted_index_config: Default::default(),
            fulltext_index_config: Default::default(),
            bloom_filter_index_config: Default::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        };

        let mut metrics = Metrics::new(WriteType::Flush);
@@ -1152,6 +1154,8 @@ mod tests {
            inverted_index_config: Default::default(),
            fulltext_index_config: Default::default(),
            bloom_filter_index_config: Default::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        }
    }

@@ -1271,6 +1275,8 @@ mod tests {
            inverted_index_config: Default::default(),
            fulltext_index_config: Default::default(),
            bloom_filter_index_config: Default::default(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: Default::default(),
        };

        let mut metrics = Metrics::new(WriteType::Flush);
--- a/src/mito2/src/worker/handle_apply_staging.rs
+++ b/src/mito2/src/worker/handle_apply_staging.rs
@@ -21,12 +21,14 @@ use store_api::storage::RegionId;
 use tokio::sync::oneshot;

 use crate::error::{
-    RegionStateSnafu, SerdeJsonSnafu, StagingPartitionExprMismatchSnafu, UnexpectedSnafu,
+    RegionStateSnafu, Result, SerdeJsonSnafu, StagingPartitionExprMismatchSnafu, UnexpectedSnafu,
 };
-use crate::manifest::action::RegionEdit;
-use crate::region::{RegionLeaderState, RegionRoleState};
-use crate::request::{OptionOutputTx, RegionEditRequest};
-use crate::sst::file::FileMeta;
+use crate::manifest::action::{RegionEdit, RegionManifest};
+use crate::manifest::storage::manifest_dir;
+use crate::manifest::storage::staging::{StagingBlobStorage, staging_blob_path};
+use crate::region::{MitoRegionRef, RegionLeaderState, RegionRoleState};
+use crate::request::{OptionOutputTx, RegionEditRequest, WorkerRequest, WorkerRequestWithTime};
+use crate::sst::location::region_dir_from_table_dir;
 use crate::worker::RegionWorkerLoop;

 impl<S: LogStore> RegionWorkerLoop<S> {
@@ -86,21 +88,32 @@ impl<S: LogStore> RegionWorkerLoop<S> {
            return;
        }

-        let (tx, rx) = oneshot::channel();
-        let files_to_add = match serde_json::from_slice::<Vec<FileMeta>>(&request.files_to_add)
-            .context(SerdeJsonSnafu)
-        {
-            Ok(files_to_add) => files_to_add,
-            Err(e) => {
-                sender.send(Err(e));
+        let worker_sender = self.sender.clone();
+        common_runtime::spawn_global(async move {
+            let staging_manifest = match Self::fetch_staging_manifest(
+                &region,
+                request.central_region_id,
+                &request.manifest_path,
+            )
+            .await
+            {
+                Ok(staging_manifest) => staging_manifest,
+                Err(e) => {
+                    sender.send(Err(e));
+                    return;
+                }
+            };
+            if staging_manifest.metadata.partition_expr.as_ref() != Some(&request.partition_expr) {
+                sender.send(Err(StagingPartitionExprMismatchSnafu {
+                    manifest_expr: staging_manifest.metadata.partition_expr.clone(),
+                    request_expr: request.partition_expr,
+                }
+                .build()));
                return;
            }
-        };

-        info!("Applying staging manifest request to region {}", region_id);
-        self.handle_region_edit(RegionEditRequest {
-            region_id,
-            edit: RegionEdit {
+            let files_to_add = staging_manifest.files.values().cloned().collect::<Vec<_>>();
+            let edit = RegionEdit {
                files_to_add,
                files_to_remove: vec![],
                timestamp_ms: Some(Utc::now().timestamp_millis()),
@@ -108,11 +121,23 @@ impl<S: LogStore> RegionWorkerLoop<S> {
                flushed_entry_id: None,
                flushed_sequence: None,
                committed_sequence: None,
-            },
-            tx,
-        });
+            };
+
+            let (tx, rx) = oneshot::channel();
+            info!(
+                "Applying staging manifest request to region {}",
+                region.region_id,
+            );
+            let _ = worker_sender
+                .send(WorkerRequestWithTime::new(WorkerRequest::EditRegion(
+                    RegionEditRequest {
+                        region_id: region.region_id,
+                        edit,
+                        tx,
+                    },
+                )))
+                .await;

-        common_runtime::spawn_global(async move {
            // Await the result from the region edit and forward the outcome to the original sender.
            // If the operation completes successfully, respond with Ok(0); otherwise, respond with an appropriate error.
            if let Ok(result) = rx.await {
@@ -137,4 +162,25 @@ impl<S: LogStore> RegionWorkerLoop<S> {
            }
        });
    }
+
+    /// Fetches the staging manifest from the central region's staging blob storage.
+    ///
+    /// The `central_region_id` is used to locate the staging directory because the staging
+    /// manifest was created by the central region during `remap_manifests` operation.
+    async fn fetch_staging_manifest(
+        region: &MitoRegionRef,
+        central_region_id: RegionId,
+        manifest_path: &str,
+    ) -> Result<RegionManifest> {
+        let region_dir =
+            region_dir_from_table_dir(region.table_dir(), central_region_id, region.path_type());
+        let staging_blob_path = staging_blob_path(&manifest_dir(&region_dir));
+        let staging_blob_storage = StagingBlobStorage::new(
+            staging_blob_path,
+            region.access_layer().object_store().clone(),
+        );
+        let staging_manifest = staging_blob_storage.get(manifest_path).await?;
+
+        serde_json::from_slice::<RegionManifest>(&staging_manifest).context(SerdeJsonSnafu)
+    }
 }
--- a/src/mito2/src/worker/handle_rebuild_index.rs
+++ b/src/mito2/src/worker/handle_rebuild_index.rs
@@ -64,6 +64,8 @@ impl<S> RegionWorkerLoop<S> {
            inverted_index_config: self.config.inverted_index.clone(),
            fulltext_index_config: self.config.fulltext_index.clone(),
            bloom_filter_index_config: self.config.bloom_filter_index.clone(),
+            #[cfg(feature = "vector_index")]
+            vector_index_config: self.config.vector_index.clone(),
            index_options: version.options.index_options.clone(),
            row_group_size: WriteOptions::default().row_group_size,
            intermediate_manager,
--- a/src/mito2/src/worker/handle_remap.rs
+++ b/src/mito2/src/worker/handle_remap.rs
@@ -16,14 +16,13 @@ use std::collections::HashMap;
 use std::time::Instant;

 use common_error::ext::BoxedError;
-use common_telemetry::info;
+use common_telemetry::{debug, info};
 use futures::future::try_join_all;
 use partition::expr::PartitionExpr;
 use snafu::{OptionExt, ResultExt};
 use store_api::storage::RegionId;

-use crate::error::{FetchManifestsSnafu, InvalidRequestSnafu, MissingManifestSnafu, Result};
-use crate::manifest::action::RegionManifest;
+use crate::error::{self, FetchManifestsSnafu, InvalidRequestSnafu, MissingManifestSnafu, Result};
 use crate::region::{MitoRegionRef, RegionMetadataLoader};
 use crate::remap_manifest::RemapManifest;
 use crate::request::RemapManifestsRequest;
@@ -75,13 +74,17 @@ impl<S> RegionWorkerLoop<S> {
        });
    }

+    // Fetches manifests for input regions, remaps them according to the provided
+    // mapping and partition expressions.
+    //
+    // Returns a map from each new region to its relative staging manifest path.
    async fn fetch_and_remap_manifests(
        region: MitoRegionRef,
        region_metadata_loader: RegionMetadataLoader,
        input_regions: Vec<RegionId>,
        new_partition_exprs: HashMap<RegionId, PartitionExpr>,
        region_mapping: HashMap<RegionId, Vec<RegionId>>,
-    ) -> Result<HashMap<RegionId, RegionManifest>> {
+    ) -> Result<HashMap<RegionId, String>> {
        let mut tasks = Vec::with_capacity(input_regions.len());
        let region_options = region.version().options.clone();
        let table_dir = region.table_dir();
@@ -97,7 +100,6 @@ impl<S> RegionWorkerLoop<S> {
                    .await
            });
        }
-
        let results = try_join_all(tasks)
            .await
            .map_err(BoxedError::new)
@@ -112,12 +114,38 @@ impl<S> RegionWorkerLoop<S> {
            .collect::<Result<HashMap<_, _>>>()?;
        let mut mapper = RemapManifest::new(manifests, new_partition_exprs, region_mapping);
        let remap_result = mapper.remap_manifests()?;
+
+        // Write new manifests to staging blob storage.
+        let manifest_manager = region.manifest_ctx.manifest_manager.write().await;
+        let manifest_storage = manifest_manager.store();
+        let staging_blob_storage = manifest_storage.staging_storage().blob_storage().clone();
+        let mut tasks = Vec::with_capacity(remap_result.new_manifests.len());
+
+        for (remap_region_id, manifest) in &remap_result.new_manifests {
+            let bytes = serde_json::to_vec(&manifest).context(error::SerializeManifestSnafu {
+                region_id: *remap_region_id,
+            })?;
+            let key = remap_manifest_key(remap_region_id);
+            tasks.push(async {
+                debug!(
+                    "Putting manifest to staging blob storage, region_id: {}, key: {}",
+                    *remap_region_id, key
+                );
+                staging_blob_storage.put(&key, bytes).await?;
+                Ok((*remap_region_id, key))
+            });
+        }
+        let r = try_join_all(tasks).await?;
        info!(
            "Remap manifests cost: {:?}, region: {}",
            now.elapsed(),
            region.region_id
        );

-        Ok(remap_result.new_manifests)
+        Ok(r.into_iter().collect::<HashMap<_, _>>())
    }
 }
+
+fn remap_manifest_key(region_id: &RegionId) -> String {
+    format!("remap_manifest_{}", region_id.as_u64())
+}
--- a/src/query/src/dist_plan/commutativity.rs
+++ b/src/query/src/dist_plan/commutativity.rs
@@ -144,7 +144,7 @@ impl Categorizer {
                    }
                }
                // all group by expressions are partition columns can push down, unless
-                // another push down(including `Limit` or `Sort`) is already in progress(which will then prvent next cond commutative node from being push down).
+                // another push down(including `Limit` or `Sort`) is already in progress(which will then prevent next cond commutative node from being push down).
                // TODO(discord9): This is a temporary solution(that works), a better description of
                // commutativity is needed under this situation.
                Commutativity::ConditionalCommutative(None)
--- a/src/query/src/query_engine/state.rs
+++ b/src/query/src/query_engine/state.rs
@@ -234,7 +234,7 @@ impl QueryEngineState {
        rules.retain(|rule| rule.name() != name);
    }

-    /// Optimize the logical plan by the extension anayzer rules.
+    /// Optimize the logical plan by the extension analyzer rules.
    pub fn optimize_by_extension_rules(
        &self,
        plan: DfLogicalPlan,
--- a/src/servers/src/tls.rs
+++ b/src/servers/src/tls.rs
@@ -29,7 +29,7 @@ use strum::EnumString;
 use crate::error::{InternalIoSnafu, Result};

 /// TlsMode is used for Mysql and Postgres server start up.
-#[derive(Debug, Default, Serialize, Deserialize, Clone, PartialEq, Eq, EnumString)]
+#[derive(Debug, Default, Serialize, Deserialize, Clone, Copy, PartialEq, Eq, EnumString)]
 #[serde(rename_all = "snake_case")]
 pub enum TlsMode {
    #[default]
@@ -91,6 +91,17 @@ impl TlsOption {
        tls_option
    }

+    /// Creates a new TLS option with the prefer mode.
+    pub fn prefer() -> Self {
+        Self {
+            mode: TlsMode::Prefer,
+            cert_path: String::new(),
+            key_path: String::new(),
+            ca_cert_path: String::new(),
+            watch: false,
+        }
+    }
+
    /// Validates the TLS configuration.
    ///
    /// Returns an error if:
--- a/src/store-api/src/metadata.rs
+++ b/src/store-api/src/metadata.rs
@@ -29,7 +29,7 @@ use common_error::status_code::StatusCode;
 use common_macro::stack_trace_debug;
 use datatypes::arrow;
 use datatypes::arrow::datatypes::FieldRef;
-use datatypes::schema::{ColumnSchema, FulltextOptions, Schema, SchemaRef};
+use datatypes::schema::{ColumnSchema, FulltextOptions, Schema, SchemaRef, VectorIndexOptions};
 use datatypes::types::TimestampType;
 use itertools::Itertools;
 use serde::de::Error;
@@ -384,6 +384,22 @@ impl RegionMetadata {
        inverted_index
    }

+    /// Gets the column IDs that have vector indexes along with their options.
+    /// Returns a map from column ID to the vector index options.
+    pub fn vector_indexed_column_ids(&self) -> HashMap<ColumnId, VectorIndexOptions> {
+        self.column_metadatas
+            .iter()
+            .filter_map(|column| {
+                column
+                    .column_schema
+                    .vector_index_options()
+                    .ok()
+                    .flatten()
+                    .map(|options| (column.column_id, options))
+            })
+            .collect()
+    }
+
    /// Checks whether the metadata is valid.
    fn validate(&self) -> Result<()> {
        // Id to name.
--- a/src/store-api/src/region_engine.rs
+++ b/src/store-api/src/region_engine.rs
@@ -759,8 +759,11 @@ pub struct RemapManifestsRequest {
 /// Response to remap manifests from old regions to new regions.
 #[derive(Debug, Clone)]
 pub struct RemapManifestsResponse {
-    /// The new manifests for the new regions.
-    pub new_manifests: HashMap<RegionId, String>,
+    /// Maps region id to its staging manifest path.
+    ///
+    /// These paths are relative paths within the central region's staging blob storage,
+    /// and should be passed to [`ApplyStagingManifestRequest`](RegionRequest::ApplyStagingManifest) to finalize the repartition.
+    pub manifest_paths: HashMap<RegionId, String>,
 }

 /// Request to copy files from a source region to a target region.
--- a/src/store-api/src/region_request.rs
+++ b/src/store-api/src/region_request.rs
@@ -421,20 +421,17 @@ fn make_region_apply_staging_manifest(
    api::v1::region::ApplyStagingManifestRequest {
        region_id,
        partition_expr,
-        files_to_add,
+        central_region_id,
+        manifest_path,
    }: api::v1::region::ApplyStagingManifestRequest,
 ) -> Result<Vec<(RegionId, RegionRequest)>> {
    let region_id = region_id.into();
-    let files_to_add = files_to_add
-        .context(UnexpectedSnafu {
-            reason: "'files_to_add' field is missing",
-        })?
-        .data;
    Ok(vec![(
        region_id,
        RegionRequest::ApplyStagingManifest(ApplyStagingManifestRequest {
            partition_expr,
-            files_to_add,
+            central_region_id: central_region_id.into(),
+            manifest_path,
        }),
    )])
 }
@@ -1464,8 +1461,10 @@ pub struct EnterStagingRequest {
 /// In practice, this means:
 /// - The `partition_expr` identifies the staging region rule that the manifest
 ///   was generated for.
-/// - `files_to_add` carries the serialized metadata (such as file manifests or
-///   file lists) that should be attached to the region under the new rule.
+/// - `central_region_id` specifies which region holds the staging blob storage
+///   where the manifest was written during the `remap_manifests` operation.
+/// - `manifest_path` is the relative path within the central region's staging
+///   blob storage to fetch the generated manifest.
 ///
 /// It should typically be called **after** the staging region has been
 /// initialized by [`EnterStagingRequest`] and the new file layout has been
@@ -1474,8 +1473,11 @@ pub struct EnterStagingRequest {
 pub struct ApplyStagingManifestRequest {
    /// The partition expression of the staging region.
    pub partition_expr: String,
-    /// The files to add to the region.
-    pub files_to_add: Vec<u8>,
+    /// The region that stores the staging manifests in its staging blob storage.
+    pub central_region_id: RegionId,
+    /// The relative path to the staging manifest within the central region's
+    /// staging blob storage.
+    pub manifest_path: String,
 }

 impl fmt::Display for RegionRequest {
--- a/tests-integration/tests/http.rs
+++ b/tests-integration/tests/http.rs
@@ -1546,6 +1546,12 @@ create_on_compaction = "auto"
 apply_on_query = "auto"
 mem_threshold_on_create = "auto"

+[region_engine.mito.vector_index]
+create_on_flush = "auto"
+create_on_compaction = "auto"
+apply_on_query = "auto"
+mem_threshold_on_create = "auto"
+
 [region_engine.mito.memtable]
 type = "time_series"

--- a/tests-integration/tests/jsonbench.rs
+++ b/tests-integration/tests/jsonbench.rs
@@ -56,6 +56,121 @@ async fn query_data(frontend: &Arc<Instance>) -> io::Result<()> {
    ))?;
    execute_sql_and_expect(frontend, sql, &expected).await;

+    // query 1:
+    let sql = "\
+SELECT \
+    json_get_string(data, '$.commit.collection') AS event, count() AS count \
+FROM bluesky \
+GROUP BY event \
+ORDER BY count DESC, event ASC";
+    let expected = r#"
+-----------------------+-------+
+| event                 | count |
+-----------------------+-------+
+| app.bsky.feed.like    | 3     |
+| app.bsky.feed.post    | 3     |
+| app.bsky.graph.follow | 3     |
+| app.bsky.feed.repost  | 1     |
+-----------------------+-------+"#;
+    execute_sql_and_expect(frontend, sql, expected).await;
+
+    // query 2:
+    let sql = "\
+SELECT \
+    json_get_string(data, '$.commit.collection') AS event, \
+    count() AS count, \
+    count(DISTINCT json_get_string(data, '$.did')) AS users \
+FROM bluesky \
+WHERE \
+    (json_get_string(data, '$.kind') = 'commit') AND \
+    (json_get_string(data, '$.commit.operation') = 'create') \
+GROUP BY event \
+ORDER BY count DESC, event ASC";
+    let expected = r#"
+-----------------------+-------+-------+
+| event                 | count | users |
+-----------------------+-------+-------+
+| app.bsky.feed.like    | 3     | 3     |
+| app.bsky.feed.post    | 3     | 3     |
+| app.bsky.graph.follow | 3     | 3     |
+| app.bsky.feed.repost  | 1     | 1     |
+-----------------------+-------+-------+"#;
+    execute_sql_and_expect(frontend, sql, expected).await;
+
+    // query 3:
+    let sql = "\
+SELECT \
+    json_get_string(data, '$.commit.collection') AS event, \
+    date_part('hour', to_timestamp_micros(json_get_int(data, '$.time_us'))) as hour_of_day, \
+    count() AS count \
+FROM bluesky \
+WHERE \
+    (json_get_string(data, '$.kind') = 'commit') AND \
+    (json_get_string(data, '$.commit.operation') = 'create') AND \
+    json_get_string(data, '$.commit.collection') IN \
+        ('app.bsky.feed.post', 'app.bsky.feed.repost', 'app.bsky.feed.like') \
+GROUP BY event, hour_of_day \
+ORDER BY hour_of_day, event";
+    let expected = r#"
+----------------------+-------------+-------+
+| event                | hour_of_day | count |
+----------------------+-------------+-------+
+| app.bsky.feed.like   | 16          | 3     |
+| app.bsky.feed.post   | 16          | 3     |
+| app.bsky.feed.repost | 16          | 1     |
+----------------------+-------------+-------+"#;
+    execute_sql_and_expect(frontend, sql, expected).await;
+
+    // query 4:
+    let sql = "\
+SELECT
+    json_get_string(data, '$.did') as user_id,
+    min(to_timestamp_micros(json_get_int(data, '$.time_us'))) AS first_post_ts
+FROM bluesky
+WHERE
+    (json_get_string(data, '$.kind') = 'commit') AND
+    (json_get_string(data, '$.commit.operation') = 'create') AND
+    (json_get_string(data, '$.commit.collection') = 'app.bsky.feed.post')
+GROUP BY user_id
+ORDER BY first_post_ts ASC, user_id DESC
+LIMIT 3";
+    let expected = r#"
+----------------------------------+----------------------------+
+| user_id                          | first_post_ts              |
+----------------------------------+----------------------------+
+| did:plc:yj3sjq3blzpynh27cumnp5ks | 2024-11-21T16:25:49.000167 |
+| did:plc:l5o3qjrmfztir54cpwlv2eme | 2024-11-21T16:25:49.001905 |
+| did:plc:s4bwqchfzm6gjqfeb6mexgbu | 2024-11-21T16:25:49.003907 |
+----------------------------------+----------------------------+"#;
+    execute_sql_and_expect(frontend, sql, expected).await;
+
+    // query 5:
+    let sql = "
+SELECT
+    json_get_string(data, '$.did') as user_id,
+    date_part(
+        'epoch',
+        max(to_timestamp_micros(json_get_int(data, '$.time_us'))) -
+        min(to_timestamp_micros(json_get_int(data, '$.time_us')))
+    ) AS activity_span
+FROM bluesky
+WHERE
+    (json_get_string(data, '$.kind') = 'commit') AND
+    (json_get_string(data, '$.commit.operation') = 'create') AND
+    (json_get_string(data, '$.commit.collection') = 'app.bsky.feed.post')
+GROUP BY user_id
+ORDER BY activity_span DESC, user_id DESC
+LIMIT 3";
+    let expected = r#"
+----------------------------------+---------------+
+| user_id                          | activity_span |
+----------------------------------+---------------+
+| did:plc:yj3sjq3blzpynh27cumnp5ks | 0.0           |
+| did:plc:s4bwqchfzm6gjqfeb6mexgbu | 0.0           |
+| did:plc:l5o3qjrmfztir54cpwlv2eme | 0.0           |
+----------------------------------+---------------+"#;
+    execute_sql_and_expect(frontend, sql, expected).await;
+
    Ok(())
 }

--- a/tests/cases/standalone/common/comment.result
+++ b/tests/cases/standalone/common/comment.result
@@ -32,6 +32,17 @@ SHOW CREATE TABLE comment_table_test;
 |                    | )                                                 |
 +--------------------+---------------------------------------------------+

+SELECT table_comment
+FROM information_schema.tables
+WHERE table_schema = 'public'
+  AND table_name = 'comment_table_test';
+
+-------------------------+
+| table_comment           |
+-------------------------+
+| table level description |
+-------------------------+
+
 -- Remove table comment
 COMMENT ON TABLE comment_table_test IS NULL;

@@ -54,6 +65,17 @@ SHOW CREATE TABLE comment_table_test;
 |                    |                                                   |
 +--------------------+---------------------------------------------------+

+SELECT table_comment
+FROM information_schema.tables
+WHERE table_schema = 'public'
+  AND table_name = 'comment_table_test';
+
+---------------+
+| table_comment |
+---------------+
+|               |
+---------------+
+
 DROP TABLE comment_table_test;

 Affected Rows: 0
@@ -90,6 +112,18 @@ SHOW CREATE TABLE comment_column_test;
 |                     |                                                         |
 +---------------------+---------------------------------------------------------+

+SELECT column_comment
+FROM information_schema.columns
+WHERE table_schema = 'public'
+  AND table_name = 'comment_column_test'
+  AND column_name = 'val';
+
+--------------------------+
+| column_comment           |
+--------------------------+
+| value column description |
+--------------------------+
+
 -- Remove column comment
 COMMENT ON COLUMN comment_column_test.val IS NULL;

@@ -112,6 +146,18 @@ SHOW CREATE TABLE comment_column_test;
 |                     |                                                    |
 +---------------------+----------------------------------------------------+

+SELECT column_comment
+FROM information_schema.columns
+WHERE table_schema = 'public'
+  AND table_name = 'comment_column_test'
+  AND column_name = 'val';
+
+----------------+
+| column_comment |
+----------------+
+|                |
+----------------+
+
 DROP TABLE comment_column_test;

 Affected Rows: 0
@@ -155,6 +201,16 @@ SHOW CREATE FLOW flow_comment_test;
 |                   | AS SELECT desc_str, ts FROM flow_source_comment_test |
 +-------------------+------------------------------------------------------+

+SELECT comment
+FROM information_schema.flows
+WHERE flow_name = 'flow_comment_test';
+
+------------------------+
+| comment                |
+------------------------+
+| flow level description |
+------------------------+
+
 -- Remove flow comment
 COMMENT ON FLOW flow_comment_test IS NULL;

@@ -170,6 +226,16 @@ SHOW CREATE FLOW flow_comment_test;
 |                   | AS SELECT desc_str, ts FROM flow_source_comment_test |
 +-------------------+------------------------------------------------------+

+SELECT comment
+FROM information_schema.flows
+WHERE flow_name = 'flow_comment_test';
+
+---------+
+| comment |
+---------+
+|         |
+---------+
+
 DROP FLOW flow_comment_test;

 Affected Rows: 0
--- a/tests/cases/standalone/common/comment.sql
+++ b/tests/cases/standalone/common/comment.sql
@@ -9,10 +9,18 @@ CREATE TABLE comment_table_test (
 -- Add table comment
 COMMENT ON TABLE comment_table_test IS 'table level description';
 SHOW CREATE TABLE comment_table_test;
+SELECT table_comment
+FROM information_schema.tables
+WHERE table_schema = 'public'
+  AND table_name = 'comment_table_test';

 -- Remove table comment
 COMMENT ON TABLE comment_table_test IS NULL;
 SHOW CREATE TABLE comment_table_test;
+SELECT table_comment
+FROM information_schema.tables
+WHERE table_schema = 'public'
+  AND table_name = 'comment_table_test';

 DROP TABLE comment_table_test;

@@ -27,10 +35,20 @@ CREATE TABLE comment_column_test (
 -- Add column comment
 COMMENT ON COLUMN comment_column_test.val IS 'value column description';
 SHOW CREATE TABLE comment_column_test;
+SELECT column_comment
+FROM information_schema.columns
+WHERE table_schema = 'public'
+  AND table_name = 'comment_column_test'
+  AND column_name = 'val';

 -- Remove column comment
 COMMENT ON COLUMN comment_column_test.val IS NULL;
 SHOW CREATE TABLE comment_column_test;
+SELECT column_comment
+FROM information_schema.columns
+WHERE table_schema = 'public'
+  AND table_name = 'comment_column_test'
+  AND column_name = 'val';

 DROP TABLE comment_column_test;

@@ -54,12 +72,17 @@ SELECT desc_str, ts FROM flow_source_comment_test;
 -- Add flow comment
 COMMENT ON FLOW flow_comment_test IS 'flow level description';
 SHOW CREATE FLOW flow_comment_test;
+SELECT comment
+FROM information_schema.flows
+WHERE flow_name = 'flow_comment_test';

 -- Remove flow comment
 COMMENT ON FLOW flow_comment_test IS NULL;
 SHOW CREATE FLOW flow_comment_test;
+SELECT comment
+FROM information_schema.flows
+WHERE flow_name = 'flow_comment_test';

 DROP FLOW flow_comment_test;
 DROP TABLE flow_source_comment_test;
 DROP TABLE flow_sink_comment_test;
-
--- a/tests/cases/standalone/common/join/join_with_aggregates.result
+++ b/tests/cases/standalone/common/join/join_with_aggregates.result
@@ -96,7 +96,7 @@ FROM (
    s."location",
    COUNT(DISTINCT s.sensor_id) as sensor_count,
    COUNT(r.reading_id) / COUNT(DISTINCT s.sensor_id) as avg_readings_per_sensor,
-    AVG(r."value") as location_avg_value
+    ROUND(AVG(r."value"), 6) as location_avg_value
  FROM sensors s
  INNER JOIN readings r ON s.sensor_id = r.sensor_id
  GROUP BY s."location"
@@ -107,7 +107,7 @@ ORDER BY location_summary.location_avg_value DESC;
 | location | sensor_count | avg_readings_per_sensor | location_avg_value |
 +----------+--------------+-------------------------+--------------------+
 | Room B   | 2            | 2                       | 35.88              |
-| Room A   | 2            | 2                       | 31.880000000000003 |
+| Room A   | 2            | 2                       | 31.88              |
 +----------+--------------+-------------------------+--------------------+

 -- Join with aggregated conditions
--- a/tests/cases/standalone/common/join/join_with_aggregates.sql
+++ b/tests/cases/standalone/common/join/join_with_aggregates.sql
@@ -62,7 +62,7 @@ FROM (
    s."location",
    COUNT(DISTINCT s.sensor_id) as sensor_count,
    COUNT(r.reading_id) / COUNT(DISTINCT s.sensor_id) as avg_readings_per_sensor,
-    AVG(r."value") as location_avg_value
+    ROUND(AVG(r."value"), 6) as location_avg_value
  FROM sensors s
  INNER JOIN readings r ON s.sensor_id = r.sensor_id
  GROUP BY s."location"
Author	SHA1	Message	Date
Ruihang Xia	3556eb4476	chore: add tests to comment column on information_schema (#7514 ) * feat: show comment on information_schema Signed-off-by: Ruihang Xia <waynestxia@gmail.com> * add to information schema for columns, add sqlness tests Signed-off-by: Ruihang Xia <waynestxia@gmail.com> * remove duplications Signed-off-by: Ruihang Xia <waynestxia@gmail.com> * fix typo Signed-off-by: Ruihang Xia <waynestxia@gmail.com> * update integration test Signed-off-by: Ruihang Xia <waynestxia@gmail.com> --------- Signed-off-by: Ruihang Xia <waynestxia@gmail.com>	2026-01-04 09:05:50 +00:00
Weny Xu	9343da7fe8	feat(meta-srv): fallback to non-TLS connection when etcd TLS prefer mode fail (#7507 ) * feat(meta-srv): fallback to non-TLS connection when etcd TLS prefer mode fail Signed-off-by: WenyXu <wenymedia@gmail.com> * chore(ci): set timeout for deploy cluster Signed-off-by: WenyXu <wenymedia@gmail.com> * refactor: simplify etcd TLS prefer mode handling Signed-off-by: WenyXu <wenymedia@gmail.com> --------- Signed-off-by: WenyXu <wenymedia@gmail.com>	2025-12-31 10:03:34 +00:00
Alan Tang	8a07dbf605	fix: fix sqlness test error about double precision (#7476 ) * fix: fix sqlness test error about double precision Signed-off-by: StandingMan <jmtangcs@gmail.com> * fix: use round method to truncate the result Signed-off-by: StandingMan <jmtangcs@gmail.com> --------- Signed-off-by: StandingMan <jmtangcs@gmail.com>	2025-12-31 04:55:22 +00:00
Weny Xu	83932c8c9e	fix: align backend_tls default value with example config (#7496 ) * fix: align backend_tls default value with example config Signed-off-by: WenyXu <wenymedia@gmail.com> * Update src/common/meta/src/kv_backend/rds/postgres.rs Co-authored-by: dennis zhuang <killme2008@gmail.com> --------- Signed-off-by: WenyXu <wenymedia@gmail.com> Co-authored-by: dennis zhuang <killme2008@gmail.com>	2025-12-31 03:31:08 +00:00
LFC	dc9fc582a0	feat: impl `json_get_int` for new json type (#7495 ) Update src/common/function/src/scalars/json/json_get.rs impl `json_get_int` for new json type Signed-off-by: luofucong <luofc@foxmail.com>	2025-12-30 09:42:16 +00:00
Weny Xu	b1d81913f5	feat: update `ApplyStagingManifestRequest` to fetch manifest from central region (#7493 ) * feat: update ApplyStagingManifestRequest to fetch manifest from central region Signed-off-by: WenyXu <wenymedia@gmail.com> * chore: refine comments Signed-off-by: WenyXu <wenymedia@gmail.com> * refactor(mito2): rename `StagingDataStorage` to `StagingBlobStorage` Signed-off-by: WenyXu <wenymedia@gmail.com> * chore: apply suggestions Signed-off-by: WenyXu <wenymedia@gmail.com> * chore: update proto Signed-off-by: WenyXu <wenymedia@gmail.com> --------- Signed-off-by: WenyXu <wenymedia@gmail.com>	2025-12-30 07:29:56 +00:00
Yingwen	554f3943b6	ci: update breaking change title level (#7497 ) Signed-off-by: evenyag <realevenyag@gmail.com>	2025-12-30 06:17:51 +00:00
dennis zhuang	e4b5ef275f	feat: impl vector index building (#7468 ) * feat: impl vector index building Signed-off-by: Dennis Zhuang <killme2008@gmail.com> * feat: supports flat format Signed-off-by: Dennis Zhuang <killme2008@gmail.com> * ci: add vector_index feature to test Signed-off-by: Dennis Zhuang <killme2008@gmail.com> * chore: apply suggestions Signed-off-by: Dennis Zhuang <killme2008@gmail.com> * chore: apply suggestions from copilot Signed-off-by: Dennis Zhuang <killme2008@gmail.com> --------- Signed-off-by: Dennis Zhuang <killme2008@gmail.com>	2025-12-30 03:38:51 +00:00
Yingwen	f2a9d50071	ci: handle prerelease version (#7492 ) Signed-off-by: evenyag <realevenyag@gmail.com>	2025-12-29 08:21:05 +00:00
LFC	0c54e70e1f	feat: impl `json_get_string` with new json type (#7489 ) * impl `json_get_string` with new json type Signed-off-by: luofucong <luofc@foxmail.com> * resolve PR comments Signed-off-by: luofucong <luofc@foxmail.com> * fix ci Signed-off-by: luofucong <luofc@foxmail.com> --------- Signed-off-by: luofucong <luofc@foxmail.com>	2025-12-29 04:35:53 +00:00
Yingwen	b51f62c3c2	feat: bump version to beta.4 (#7490 ) Signed-off-by: evenyag <realevenyag@gmail.com>	2025-12-29 04:20:00 +00:00
discord9	1ddc535b52	feat: repartition map kv (#7420 ) * table partition key Signed-off-by: discord9 <discord9@163.com> * feat: table part key Signed-off-by: discord9 <discord9@163.com> * ut Signed-off-by: discord9 <discord9@163.com> * stuff Signed-off-by: discord9 <discord9@163.com> * feat: add Default trait to TablePartValue struct Signed-off-by: discord9 <discord9@163.com> * rename to Rep Signed-off-by: discord9 <discord9@163.com> * rename file Signed-off-by: discord9 <discord9@163.com> * more rename Signed-off-by: discord9 <discord9@163.com> * pcr Signed-off-by: discord9 <discord9@163.com> * test: update err msg Signed-off-by: discord9 <discord9@163.com> * feat: add TableRepartKey to TableMetadataManager Signed-off-by: discord9 <discord9@163.com> * feat: add TableRepartManager to TableMetadataManager Signed-off-by: discord9 <discord9@163.com> * docs: udpate Signed-off-by: discord9 <discord9@163.com> * c Signed-off-by: discord9 <discord9@163.com> --------- Signed-off-by: discord9 <discord9@163.com>	2025-12-29 02:45:35 +00:00