Bump version: 0.30.0-beta.1 → 0.30.0-beta.2

## Summary
- Bump Lance Rust workspace dependencies to `v3.1.0-beta.2` via
`ci/set_lance_version.py`.
- Update Java `lance-core.version` to `3.1.0-beta.2`.

## Verification
- `cargo clippy --workspace --tests --all-features -- -D warnings`
- `cargo fmt --all`

## Release Reference
- refs/tags/v3.1.0-beta.2

.bumpversion.toml

@@ -1,5 +1,5 @@
 [tool.bumpversion]
-current_version = "0.26.2"
+current_version = "0.27.0-beta.1"
 parse = """(?x)
     (?P<major>0|[1-9]\\d*)\\.
     (?P<minor>0|[1-9]\\d*)\\.

.github/workflows/codex-fix-ci.yml

@@ -0,0 +1,173 @@
+name: Codex Fix CI
+
+on:
+  workflow_dispatch:
+    inputs:
+      workflow_run_url:
+        description: "Failing CI workflow run URL (e.g., https://github.com/lancedb/lancedb/actions/runs/12345678)"
+        required: true
+        type: string
+      branch:
+        description: "Branch to fix (e.g., main, release/v2.0, or feature-branch)"
+        required: true
+        type: string
+      guidelines:
+        description: "Additional guidelines for the fix (optional)"
+        required: false
+        type: string
+
+permissions:
+  contents: write
+  pull-requests: write
+  actions: read
+
+jobs:
+  fix-ci:
+    runs-on: warp-ubuntu-latest-x64-4x
+    timeout-minutes: 60
+    env:
+      CC: clang
+      CXX: clang++
+    steps:
+      - name: Show inputs
+        run: |
+          echo "workflow_run_url = ${{ inputs.workflow_run_url }}"
+          echo "branch = ${{ inputs.branch }}"
+          echo "guidelines = ${{ inputs.guidelines }}"
+
+      - name: Checkout Repo
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.branch }}
+          fetch-depth: 0
+          persist-credentials: true
+
+      - name: Set up Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Install Codex CLI
+        run: npm install -g @openai/codex
+
+      - name: Install Rust toolchain
+        uses: dtolnay/rust-toolchain@stable
+        with:
+          toolchain: stable
+          components: clippy, rustfmt
+
+      - uses: Swatinem/rust-cache@v2
+
+      - name: Install system dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y protobuf-compiler libssl-dev
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
+      - name: Install Python dependencies
+        run: |
+          pip install maturin ruff pytest pyarrow pandas polars
+
+      - name: Set up Java
+        uses: actions/setup-java@v4
+        with:
+          distribution: temurin
+          java-version: '11'
+          cache: maven
+
+      - name: Install Node.js dependencies for TypeScript bindings
+        run: |
+          cd nodejs
+          npm ci
+
+      - name: Configure git user
+        run: |
+          git config user.name "lancedb automation"
+          git config user.email "robot@lancedb.com"
+
+      - name: Run Codex to fix CI failure
+        env:
+          WORKFLOW_RUN_URL: ${{ inputs.workflow_run_url }}
+          BRANCH: ${{ inputs.branch }}
+          GUIDELINES: ${{ inputs.guidelines }}
+          GITHUB_TOKEN: ${{ secrets.ROBOT_TOKEN }}
+          GH_TOKEN: ${{ secrets.ROBOT_TOKEN }}
+          OPENAI_API_KEY: ${{ secrets.CODEX_TOKEN }}
+        run: |
+          set -euo pipefail
+
+          cat <<EOF >/tmp/codex-prompt.txt
+          You are running inside the lancedb repository on a GitHub Actions runner. Your task is to fix a CI failure.
+
+          Input parameters:
+          - Failing workflow run URL: ${WORKFLOW_RUN_URL}
+          - Branch to fix: ${BRANCH}
+          - Additional guidelines: ${GUIDELINES:-"None provided"}
+
+          Follow these steps exactly:
+
+          1. Extract the run ID from the workflow URL. The URL format is https://github.com/lancedb/lancedb/actions/runs/<run_id>.
+
+          2. Use "gh run view <run_id> --json jobs,conclusion,name" to get information about the failed run.
+
+          3. Identify which jobs failed. For each failed job, use "gh run view <run_id> --job <job_id> --log-failed" to get the failure logs.
+
+          4. Analyze the failure logs to understand what went wrong. Common failures include:
+             - Compilation errors
+             - Test failures
+             - Clippy warnings treated as errors
+             - Formatting issues
+             - Dependency issues
+
+          5. Based on the analysis, fix the issues in the codebase:
+             - For compilation errors: Fix the code that doesn't compile
+             - For test failures: Fix the failing tests or the code they test
+             - For clippy warnings: Apply the suggested fixes
+             - For formatting issues: Run "cargo fmt --all"
+             - For other issues: Apply appropriate fixes
+
+          6. After making fixes, verify them locally:
+             - Run "cargo fmt --all" to ensure formatting is correct
+             - Run "cargo clippy --workspace --tests --all-features -- -D warnings" to check for issues
+             - Run ONLY the specific failing tests to confirm they pass now:
+               - For Rust test failures: Run the specific test with "cargo test -p <crate> <test_name>"
+               - For Python test failures: Build with "cd python && maturin develop" then run "pytest <specific_test_file>::<test_name>"
+               - For Java test failures: Run "cd java && mvn test -Dtest=<TestClass>#<testMethod>"
+               - For TypeScript test failures: Run "cd nodejs && npm run build && npm test -- --testNamePattern='<test_name>'"
+               - Do NOT run the full test suite - only run the tests that were failing
+
+          7. If the additional guidelines are provided, follow them as well.
+
+          8. Inspect "git status --short" and "git diff" to review your changes.
+
+          9. Create a fix branch: "git checkout -b codex/fix-ci-<run_id>".
+
+          10. Stage all changes with "git add -A" and commit with message "fix: resolve CI failures from run <run_id>".
+
+          11. Push the branch: "git push origin codex/fix-ci-<run_id>". If the remote branch exists, delete it first with "gh api -X DELETE repos/lancedb/lancedb/git/refs/heads/codex/fix-ci-<run_id>" then push. Do NOT use "git push --force" or "git push -f".
+
+          12. Create a pull request targeting "${BRANCH}":
+              - Title: "ci: <short summary describing the fix>" (e.g., "ci: fix clippy warnings in lancedb" or "ci: resolve test flakiness in vector search")
+              - First, write the PR body to /tmp/pr-body.md using a heredoc (cat <<'PREOF' > /tmp/pr-body.md). The body should include:
+                - Link to the failing workflow run
+                - Summary of what failed
+                - Description of the fixes applied
+              - Then run "gh pr create --base ${BRANCH} --body-file /tmp/pr-body.md".
+
+          13. Display the new PR URL, "git status --short", and a summary of what was fixed.
+
+          Constraints:
+          - Use bash commands for all operations.
+          - Do not merge the PR.
+          - Do not modify GitHub workflow files unless they are the cause of the failure.
+          - If any command fails, diagnose and attempt to fix the issue instead of aborting immediately.
+          - If you cannot fix the issue automatically, create the PR anyway with a clear explanation of what you tried and what remains to be fixed.
+          - env "GH_TOKEN" is available, use "gh" tools for GitHub-related operations.
+          EOF
+
+          printenv OPENAI_API_KEY | codex login --with-api-key
+          codex --config shell_environment_policy.ignore_default_excludes=true exec --dangerously-bypass-approvals-and-sandbox "$(cat /tmp/codex-prompt.txt)"

.github/workflows/nodejs.yml

@@ -8,6 +8,7 @@ on:
     paths:
       - Cargo.toml
       - nodejs/**
+      - rust/**
       - docs/src/js/**
       - .github/workflows/nodejs.yml
       - docker-compose.yml
@@ -77,8 +78,11 @@ jobs:
         fetch-depth: 0
         lfs: true
     - uses: actions/setup-node@v3
+      name: Setup Node.js 20 for build
       with:
-        node-version: ${{ matrix.node-version }}
+        # @napi-rs/cli v3 requires Node >= 20.12 (via @inquirer/prompts@8).
+        # Build always on Node 20; tests run on the matrix version below.
+        node-version: 20
         cache: 'npm'
         cache-dependency-path: nodejs/package-lock.json
     - uses: Swatinem/rust-cache@v2
@@ -86,12 +90,16 @@ jobs:
       run: |
         sudo apt update
         sudo apt install -y protobuf-compiler libssl-dev
-        npm install -g @napi-rs/cli
     - name: Build
       run: |
         npm ci --include=optional
         npm run build:debug -- --profile ci
-        npm run tsc
+    - uses: actions/setup-node@v3
+      name: Setup Node.js ${{ matrix.node-version }} for test
+      with:
+        node-version: ${{ matrix.node-version }}
+    - name: Compile TypeScript
+      run: npm run tsc
     - name: Setup localstack
       working-directory: .
       run: docker compose up --detach --wait
@@ -144,7 +152,6 @@ jobs:
     - name: Install dependencies
       run: |
         brew install protobuf
-        npm install -g @napi-rs/cli
     - name: Build
       run: |
         npm ci --include=optional

.github/workflows/npm-publish.yml

@@ -128,16 +128,13 @@ jobs:
           - target: x86_64-unknown-linux-musl
             # This one seems to need some extra memory
             host: ubuntu-2404-8x-x64
-            # https://github.com/napi-rs/napi-rs/blob/main/alpine.Dockerfile
-            docker: ghcr.io/napi-rs/napi-rs/nodejs-rust:lts-alpine
             features: fp16kernels
             pre_build: |-
               set -e &&
-              apk add protobuf-dev curl &&
-              ln -s /usr/lib/gcc/x86_64-alpine-linux-musl/14.2.0/crtbeginS.o /usr/lib/crtbeginS.o &&
-              ln -s /usr/lib/libgcc_s.so /usr/lib/libgcc.so &&
-              CC=gcc &&
-              CXX=g++
+              sudo apt-get update &&
+              sudo apt-get install -y protobuf-compiler pkg-config &&
+              rustup target add x86_64-unknown-linux-musl &&
+              export EXTRA_ARGS="-x"
           - target: aarch64-unknown-linux-gnu
             host: ubuntu-2404-8x-x64
             # https://github.com/napi-rs/napi-rs/blob/main/debian-aarch64.Dockerfile
@@ -153,15 +150,13 @@ jobs:
               rustup target add aarch64-unknown-linux-gnu
           - target: aarch64-unknown-linux-musl
             host: ubuntu-2404-8x-x64
-            # https://github.com/napi-rs/napi-rs/blob/main/alpine.Dockerfile
-            docker: ghcr.io/napi-rs/napi-rs/nodejs-rust:lts-alpine
             features: ","
             pre_build: |-
               set -e &&
-              apk add protobuf-dev &&
+              sudo apt-get update &&
+              sudo apt-get install -y protobuf-compiler &&
               rustup target add aarch64-unknown-linux-musl &&
-              export CC_aarch64_unknown_linux_musl=aarch64-linux-musl-gcc &&
-              export CXX_aarch64_unknown_linux_musl=aarch64-linux-musl-g++
+              export EXTRA_ARGS="-x"
     name: build - ${{ matrix.settings.target }}
     runs-on: ${{ matrix.settings.host }}
     defaults:
@@ -192,12 +187,18 @@ jobs:
             .cargo-cache
             target/
           key: nodejs-${{ matrix.settings.target }}-cargo-${{ matrix.settings.host }}
-      - name: Setup toolchain
-        run: ${{ matrix.settings.setup }}
-        if: ${{ matrix.settings.setup }}
-        shell: bash
       - name: Install dependencies
         run: npm ci
+      - name: Install Zig
+        uses: mlugg/setup-zig@v2
+        if: ${{ contains(matrix.settings.target, 'musl') }}
+        with:
+          version: 0.14.1
+      - name: Install cargo-zigbuild
+        uses: taiki-e/install-action@v2
+        if: ${{ contains(matrix.settings.target, 'musl') }}
+        with:
+          tool: cargo-zigbuild
       - name: Build in docker
         uses: addnab/docker-run-action@v3
         if: ${{ matrix.settings.docker }}
@@ -210,24 +211,24 @@ jobs:
           run: |
             set -e
             ${{ matrix.settings.pre_build }}
-            npx napi build --platform  --release --no-const-enum \
+            npx napi build --platform --release \
               --features ${{ matrix.settings.features }} \
               --target ${{ matrix.settings.target }} \
               --dts ../lancedb/native.d.ts \
               --js ../lancedb/native.js \
               --strip \
-              dist/
+              --output-dir dist/
       - name: Build
         run: |
           ${{ matrix.settings.pre_build }}
-          npx napi build --platform  --release --no-const-enum \
+          npx napi build --platform --release \
               --features ${{ matrix.settings.features }} \
               --target ${{ matrix.settings.target }} \
               --dts ../lancedb/native.d.ts \
               --js ../lancedb/native.js \
               --strip \
               $EXTRA_ARGS \
-              dist/
+              --output-dir dist/
         if: ${{ !matrix.settings.docker }}
         shell: bash
       - name: Upload artifact

.github/workflows/python.yml

@@ -8,6 +8,7 @@ on:
     paths:
       - Cargo.toml
       - python/**
+      - rust/**
       - .github/workflows/python.yml
 
 concurrency:

.github/workflows/rust.yml

@@ -183,7 +183,7 @@ jobs:
     runs-on: ubuntu-24.04
     strategy:
       matrix:
-        msrv: ["1.88.0"] # This should match up with rust-version in Cargo.toml
+        msrv: ["1.91.0"] # This should match up with rust-version in Cargo.toml
     env:
       # Need up-to-date compilers for kernels
       CC: clang-18

Cargo.lock

@@ -128,15 +128,6 @@ version = "1.0.100"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a23eb6b1614318a8071c9b2521f36b424b2c83db5eb3a0fead4a6c0809af6e61"
 
-[[package]]
-name = "approx"
-version = "0.5.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cab112f0a86d568ea0e627cc1d6be74a1e9cd55214684db5561995f6dad897c6"
-dependencies = [
- "num-traits",
-]
-
 [[package]]
 name = "arbitrary"
 version = "1.4.2"
@@ -1389,9 +1380,9 @@ checksum = "1fd0f2584146f6f2ef48085050886acf353beff7305ebd1ae69500e27c67f64b"
 
 [[package]]
 name = "bytes"
-version = "1.10.1"
+version = "1.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d71b6127be86fdcfddb610f7182ac57211d4b18a3e9c82eb2d17662f2227ad6a"
+checksum = "1e748733b7cbc798e1434b6ac524f0c1ff2ab456fe201501e6497c8417a4fc33"
 
 [[package]]
 name = "bytes-utils"
@@ -1571,7 +1562,7 @@ checksum = "0b023947811758c97c59bf9d1c188fd619ad4718dcaa767947df1cadb14f39f4"
 dependencies = [
  "glob",
  "libc",
- "libloading",
+ "libloading 0.8.9",
 ]
 
 [[package]]
@@ -1658,9 +1649,9 @@ checksum = "7c74b8349d32d297c9134b8c88677813a227df8f779daa29bfc29c183fe3dca6"
 
 [[package]]
 name = "convert_case"
-version = "0.6.0"
+version = "0.11.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ec182b0ca2f35d8fc196cf3404988fd8b8c739a4d270ff118a398feb0cbec1ca"
+checksum = "affbf0190ed2caf063e3def54ff444b449371d55c58e513a95ab98eca50adb49"
 dependencies = [
  "unicode-segmentation",
 ]
@@ -1783,6 +1774,16 @@ dependencies = [
  "crossbeam-utils",
 ]
 
+[[package]]
+name = "crossbeam-skiplist"
+version = "0.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "df29de440c58ca2cc6e587ec3d22347551a32435fbde9d2bff64e78a9ffa151b"
+dependencies = [
+ "crossbeam-epoch",
+ "crossbeam-utils",
+]
+
 [[package]]
 name = "crossbeam-utils"
 version = "0.8.21"
@@ -1883,14 +1884,20 @@ dependencies = [
 
 [[package]]
 name = "ctor"
-version = "0.2.9"
+version = "0.6.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "32a2785755761f3ddc1492979ce1e48d2c00d09311c39e4466429188f3dd6501"
+checksum = "424e0138278faeb2b401f174ad17e715c829512d74f3d1e81eb43365c2e0590e"
 dependencies = [
- "quote",
- "syn 2.0.114",
+ "ctor-proc-macro",
+ "dtor",
 ]
 
+[[package]]
+name = "ctor-proc-macro"
+version = "0.0.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "52560adf09603e58c9a7ee1fe1dcb95a16927b17c127f0ac02d6e768a0e25bc1"
+
 [[package]]
 name = "darling"
 version = "0.20.11"
@@ -2727,6 +2734,21 @@ version = "2.0.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "117240f60069e65410b3ae1bb213295bd828f707b5bec6596a1afc8793ce0cbc"
 
+[[package]]
+name = "dtor"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "404d02eeb088a82cfd873006cb713fe411306c7d182c344905e101fb1167d301"
+dependencies = [
+ "dtor-proc-macro",
+]
+
+[[package]]
+name = "dtor-proc-macro"
+version = "0.0.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f678cf4a922c215c63e0de95eb1ff08a958a81d47e485cf9da1e27bf6305cfa5"
+
 [[package]]
 name = "dunce"
 version = "1.0.5"
@@ -2765,16 +2787,6 @@ version = "0.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "05dbec7076f432bb132db738df90d87a4f5789e99f59e7b1219a6b8ef61eaa68"
 
-[[package]]
-name = "earcutr"
-version = "0.4.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "79127ed59a85d7687c409e9978547cffb7dc79675355ed22da6b66fd5f6ead01"
-dependencies = [
- "itertools 0.11.0",
- "num-traits",
-]
-
 [[package]]
 name = "ecdsa"
 version = "0.14.8"
@@ -3021,12 +3033,6 @@ dependencies = [
  "miniz_oxide",
 ]
 
-[[package]]
-name = "float_next_after"
-version = "1.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8bf7cc16383c4b8d58b9905a8509f02926ce3058053c056376248d958c9df1e8"
-
 [[package]]
 name = "fnv"
 version = "1.0.7"
@@ -3072,9 +3078,8 @@ checksum = "42703706b716c37f96a77aea830392ad231f44c9e9a67872fa5548707e11b11c"
 
 [[package]]
 name = "fsst"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0f03a771ab914e207dd26bd2f12666839555ec8ecc7e1770e1ed6f9900d899a4"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow-array",
  "rand 0.9.2",
@@ -3451,128 +3456,6 @@ dependencies = [
  "version_check",
 ]
 
-[[package]]
-name = "geo"
-version = "0.31.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2fc1a1678e54befc9b4bcab6cd43b8e7f834ae8ea121118b0fd8c42747675b4a"
-dependencies = [
- "earcutr",
- "float_next_after",
- "geo-types",
- "geographiclib-rs",
- "i_overlay",
- "log",
- "num-traits",
- "robust",
- "rstar",
- "spade",
-]
-
-[[package]]
-name = "geo-traits"
-version = "0.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2e7c353d12a704ccfab1ba8bfb1a7fe6cb18b665bf89d37f4f7890edcd260206"
-dependencies = [
- "geo-types",
-]
-
-[[package]]
-name = "geo-types"
-version = "0.7.17"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "75a4dcd69d35b2c87a7c83bce9af69fd65c9d68d3833a0ded568983928f3fc99"
-dependencies = [
- "approx",
- "num-traits",
- "rayon",
- "rstar",
- "serde",
-]
-
-[[package]]
-name = "geoarrow-array"
-version = "0.7.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dc1cc4106ac0a0a512c398961ce95d8150475c84a84e17c4511c3643fa120a17"
-dependencies = [
- "arrow-array",
- "arrow-buffer",
- "arrow-schema",
- "geo-traits",
- "geoarrow-schema",
- "num-traits",
- "wkb",
- "wkt",
-]
-
-[[package]]
-name = "geoarrow-expr-geo"
-version = "0.7.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fa84300361ce57fb875bcaa6e32b95b0aff5c6b1af692b936bdd58ff343f4394"
-dependencies = [
- "arrow-array",
- "arrow-buffer",
- "geo",
- "geo-traits",
- "geoarrow-array",
- "geoarrow-schema",
-]
-
-[[package]]
-name = "geoarrow-schema"
-version = "0.7.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e97be4e9f523f92bd6a0e0458323f4b783d073d011664decd8dbf05651704f34"
-dependencies = [
- "arrow-schema",
- "geo-traits",
- "serde",
- "serde_json",
- "thiserror 1.0.69",
-]
-
-[[package]]
-name = "geodatafusion"
-version = "0.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "773cfa1fb0d7f7661b76b3fde00f3ffd8e0ff7b3635096f0ff6294fe5ca62a2b"
-dependencies = [
- "arrow-arith",
- "arrow-array",
- "arrow-schema",
- "datafusion",
- "geo",
- "geo-traits",
- "geoarrow-array",
- "geoarrow-expr-geo",
- "geoarrow-schema",
- "geohash",
- "thiserror 1.0.69",
- "wkt",
-]
-
-[[package]]
-name = "geographiclib-rs"
-version = "0.2.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f611040a2bb37eaa29a78a128d1e92a378a03e0b6e66ae27398d42b1ba9a7841"
-dependencies = [
- "libm",
-]
-
-[[package]]
-name = "geohash"
-version = "0.13.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0fb94b1a65401d6cbf22958a9040aa364812c26674f841bee538b12c135db1e6"
-dependencies = [
- "geo-types",
- "libm",
-]
-
 [[package]]
 name = "getrandom"
 version = "0.2.16"
@@ -3682,15 +3565,6 @@ dependencies = [
  "zerocopy",
 ]
 
-[[package]]
-name = "hash32"
-version = "0.3.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "47d60b12902ba28e2730cd37e95b8c9223af2808df9e902d4df49588d1470606"
-dependencies = [
- "byteorder",
-]
-
 [[package]]
 name = "hashbrown"
 version = "0.12.3"
@@ -3725,16 +3599,6 @@ version = "0.16.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "5419bdc4f6a9207fbeba6d11b604d481addf78ecd10c11ad51e76c2f6482748d"
 
-[[package]]
-name = "heapless"
-version = "0.8.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0bfb9eb618601c89945a70e254898da93b13be0388091d42117462b265bb3fad"
-dependencies = [
- "hash32",
- "stable_deref_trait",
-]
-
 [[package]]
 name = "heck"
 version = "0.4.1"
@@ -3996,49 +3860,6 @@ dependencies = [
  "serde",
 ]
 
-[[package]]
-name = "i_float"
-version = "1.15.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "010025c2c532c8d82e42d0b8bb5184afa449fa6f06c709ea9adcb16c49ae405b"
-dependencies = [
- "libm",
-]
-
-[[package]]
-name = "i_key_sort"
-version = "0.6.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9190f86706ca38ac8add223b2aed8b1330002b5cdbbce28fb58b10914d38fc27"
-
-[[package]]
-name = "i_overlay"
-version = "4.0.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0fcccbd4e4274e0f80697f5fbc6540fdac533cce02f2081b328e68629cce24f9"
-dependencies = [
- "i_float",
- "i_key_sort",
- "i_shape",
- "i_tree",
- "rayon",
-]
-
-[[package]]
-name = "i_shape"
-version = "1.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1ea154b742f7d43dae2897fcd5ead86bc7b5eefcedd305a7ebf9f69d44d61082"
-dependencies = [
- "i_float",
-]
-
-[[package]]
-name = "i_tree"
-version = "0.16.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "35e6d558e6d4c7b82bc51d9c771e7a927862a161a7d87bf2b0541450e0e20915"
-
 [[package]]
 name = "iana-time-zone"
 version = "0.1.64"
@@ -4405,9 +4226,8 @@ dependencies = [
 
 [[package]]
 name = "lance"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "47b685aca3f97ee02997c83ded16f59c747ccb69e74c8abbbae4aa3d22cf1301"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow",
  "arrow-arith",
@@ -4426,6 +4246,7 @@ dependencies = [
  "byteorder",
  "bytes",
  "chrono",
+ "crossbeam-skiplist",
  "dashmap",
  "datafusion",
  "datafusion-expr",
@@ -4443,7 +4264,6 @@ dependencies = [
  "lance-datafusion",
  "lance-encoding",
  "lance-file",
- "lance-geo",
  "lance-index",
  "lance-io",
  "lance-linalg",
@@ -4465,6 +4285,7 @@ dependencies = [
  "tantivy",
  "tokio",
  "tokio-stream",
+ "tokio-util",
  "tracing",
  "url",
  "uuid",
@@ -4472,9 +4293,8 @@ dependencies = [
 
 [[package]]
 name = "lance-arrow"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "daf00c7537df524cc518a089f0d156a036d95ca3f5bc2bc1f0a9f9293e9b62ef"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow-array",
  "arrow-buffer",
@@ -4493,9 +4313,8 @@ dependencies = [
 
 [[package]]
 name = "lance-bitpacking"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "46752e4ac8fc5590a445e780b63a8800adc7a770bd74770a8dc66963778e4e77"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrayref",
  "paste",
@@ -4504,9 +4323,8 @@ dependencies = [
 
 [[package]]
 name = "lance-core"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3d13d87d07305c6d4b4dc7780fb1107babf782a0e5b1dc7872e17ae1f8fd11ca"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow-array",
  "arrow-buffer",
@@ -4543,9 +4361,8 @@ dependencies = [
 
 [[package]]
 name = "lance-datafusion"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6451b5af876eaef8bec4b38a39dadac9d44621e1ecf85d0cdf6097a5d0aa8721"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow",
  "arrow-array",
@@ -4564,10 +4381,10 @@ dependencies = [
  "lance-arrow",
  "lance-core",
  "lance-datagen",
- "lance-geo",
  "log",
  "pin-project",
  "prost",
+ "prost-build",
  "snafu",
  "tokio",
  "tracing",
@@ -4575,9 +4392,8 @@ dependencies = [
 
 [[package]]
 name = "lance-datagen"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e1736708dd7867dfbab8fcc930b21c96717c6c00be73b7d9a240336a4ed80375"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow",
  "arrow-array",
@@ -4595,9 +4411,8 @@ dependencies = [
 
 [[package]]
 name = "lance-encoding"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d6b6ca4ff94833240d5ba4a94a742cba786d1949b3c3fa7e11d6f0050443432a"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow-arith",
  "arrow-array",
@@ -4634,9 +4449,8 @@ dependencies = [
 
 [[package]]
 name = "lance-file"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "55fbe959bffe185543aed3cbeb14484f1aa2e55886034fdb1ea3d8cc9b70aad8"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow-arith",
  "arrow-array",
@@ -4666,27 +4480,10 @@ dependencies = [
  "tracing",
 ]
 
-[[package]]
-name = "lance-geo"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a52b0adabc953d457f336a784a3b37353a180e6a79905f544949746e0d4c6483"
-dependencies = [
- "datafusion",
- "geo-traits",
- "geo-types",
- "geoarrow-array",
- "geoarrow-schema",
- "geodatafusion",
- "lance-core",
- "serde",
-]
-
 [[package]]
 name = "lance-index"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6b67654bf86fd942dd2cf08294ee7e91053427cd148225f49c9ff398ff9a40fd"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow",
  "arrow-arith",
@@ -4710,9 +4507,6 @@ dependencies = [
  "dirs",
  "fst",
  "futures",
- "geo-types",
- "geoarrow-array",
- "geoarrow-schema",
  "half",
  "itertools 0.13.0",
  "jsonb",
@@ -4722,7 +4516,6 @@ dependencies = [
  "lance-datagen",
  "lance-encoding",
  "lance-file",
- "lance-geo",
  "lance-io",
  "lance-linalg",
  "lance-table",
@@ -4753,9 +4546,8 @@ dependencies = [
 
 [[package]]
 name = "lance-io"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8eb0ccc1c414e31687d83992d546af0a0237c8d2f4bf2ae3d347d539fd0fc141"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow",
  "arrow-arith",
@@ -4786,8 +4578,8 @@ dependencies = [
  "prost",
  "rand 0.9.2",
  "serde",
- "shellexpand",
  "snafu",
+ "tempfile",
  "tokio",
  "tracing",
  "url",
@@ -4795,9 +4587,8 @@ dependencies = [
 
 [[package]]
 name = "lance-linalg"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "083404cf12dcdb1a7df98fb58f9daf626b6e43a2f794b37b6b89b4012a0e1f78"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow-array",
  "arrow-buffer",
@@ -4813,9 +4604,8 @@ dependencies = [
 
 [[package]]
 name = "lance-namespace"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c12778d2aabf9c2bfd16e2509ebe120e562a288d8ae630ec6b6b204868df41b2"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow",
  "async-trait",
@@ -4827,9 +4617,8 @@ dependencies = [
 
 [[package]]
 name = "lance-namespace-impls"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8863aababdd13a6d2c8d6179dc6981f4f8f49d8b66a00c5dd75115aec4cadc99"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow",
  "arrow-ipc",
@@ -4844,6 +4633,7 @@ dependencies = [
  "lance-index",
  "lance-io",
  "lance-namespace",
+ "lance-table",
  "log",
  "object_store",
  "rand 0.9.2",
@@ -4859,9 +4649,9 @@ dependencies = [
 
 [[package]]
 name = "lance-namespace-reqwest-client"
-version = "0.4.5"
+version = "0.5.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a2acdba67f84190067532fce07b51a435dd390d7cdc1129a05003e5cb3274cf0"
+checksum = "3ad4c947349acd6e37e984eba0254588bd894e6128434338b9e6904e56fb4633"
 dependencies = [
  "reqwest",
  "serde",
@@ -4872,9 +4662,8 @@ dependencies = [
 
 [[package]]
 name = "lance-table"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f0fcc83f197ce2000c4abe4f5e0873490ab1f41788fa76571c4209b87d4daf50"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow",
  "arrow-array",
@@ -4913,9 +4702,8 @@ dependencies = [
 
 [[package]]
 name = "lance-testing"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7fb1f7c7e06f91360e141ecee1cf2110f858c231705f69f2cd2fda9e30c1e9f4"
+version = "3.1.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v3.1.0-beta.2#ae3b1f413cc49d783f51abe62c8261c106c9b6cd"
 dependencies = [
  "arrow-array",
  "arrow-schema",
@@ -4926,7 +4714,7 @@ dependencies = [
 
 [[package]]
 name = "lancedb"
-version = "0.26.2"
+version = "0.27.0-beta.1"
 dependencies = [
  "ahash",
  "anyhow",
@@ -4956,8 +4744,10 @@ dependencies = [
  "datafusion-common",
  "datafusion-execution",
  "datafusion-expr",
+ "datafusion-functions",
  "datafusion-physical-expr",
  "datafusion-physical-plan",
+ "datafusion-sql",
  "futures",
  "half",
  "hf-hub",
@@ -5006,7 +4796,7 @@ dependencies = [
 
 [[package]]
 name = "lancedb-nodejs"
-version = "0.26.2"
+version = "0.27.0-beta.1"
 dependencies = [
  "arrow-array",
  "arrow-ipc",
@@ -5026,7 +4816,7 @@ dependencies = [
 
 [[package]]
 name = "lancedb-python"
-version = "0.29.2"
+version = "0.30.0-beta.1"
 dependencies = [
  "arrow",
  "async-trait",
@@ -5138,6 +4928,16 @@ dependencies = [
  "windows-link 0.2.1",
 ]
 
+[[package]]
+name = "libloading"
+version = "0.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "754ca22de805bb5744484a5b151a9e1a8e837d5dc232c2d7d8c2e3492edc8b60"
+dependencies = [
+ "cfg-if",
+ "windows-link 0.2.1",
+]
+
 [[package]]
 name = "libm"
 version = "0.2.15"
@@ -5485,32 +5285,34 @@ checksum = "2195bf6aa996a481483b29d62a7663eed3fe39600c460e323f8ff41e90bdd89b"
 
 [[package]]
 name = "napi"
-version = "2.16.17"
+version = "3.8.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "55740c4ae1d8696773c78fdafd5d0e5fe9bc9f1b071c7ba493ba5c413a9184f3"
+checksum = "e6944d0bf100571cd6e1a98a316cdca262deb6fccf8d93f5ae1502ca3fc88bd3"
 dependencies = [
  "bitflags 2.9.4",
  "ctor",
- "napi-derive",
+ "futures",
+ "napi-build",
  "napi-sys",
- "once_cell",
+ "nohash-hasher",
+ "rustc-hash 2.1.1",
  "tokio",
 ]
 
 [[package]]
 name = "napi-build"
-version = "2.2.3"
+version = "2.3.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dcae8ad5609d14afb3a3b91dee88c757016261b151e9dcecabf1b2a31a6cab14"
+checksum = "d376940fd5b723c6893cd1ee3f33abbfd86acb1cd1ec079f3ab04a2a3bc4d3b1"
 
 [[package]]
 name = "napi-derive"
-version = "2.16.13"
+version = "3.5.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7cbe2585d8ac223f7d34f13701434b9d5f4eb9c332cccce8dee57ea18ab8ab0c"
+checksum = "2c914b5e420182bfb73504e0607592cdb8e2e21437d450883077669fb72a114d"
 dependencies = [
- "cfg-if",
  "convert_case",
+ "ctor",
  "napi-derive-backend",
  "proc-macro2",
  "quote",
@@ -5519,26 +5321,24 @@ dependencies = [
 
 [[package]]
 name = "napi-derive-backend"
-version = "1.0.75"
+version = "5.0.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1639aaa9eeb76e91c6ae66da8ce3e89e921cd3885e99ec85f4abacae72fc91bf"
+checksum = "f0864cf6a82e2cfb69067374b64c9253d7e910e5b34db833ed7495dda56ccb18"
 dependencies = [
  "convert_case",
- "once_cell",
  "proc-macro2",
  "quote",
- "regex",
  "semver",
  "syn 2.0.114",
 ]
 
 [[package]]
 name = "napi-sys"
-version = "2.4.0"
+version = "3.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "427802e8ec3a734331fec1035594a210ce1ff4dc5bc1950530920ab717964ea3"
+checksum = "8eb602b84d7c1edae45e50bbf1374696548f36ae179dfa667f577e384bb90c2b"
 dependencies = [
- "libloading",
+ "libloading 0.9.0",
 ]
 
 [[package]]
@@ -5556,6 +5356,12 @@ dependencies = [
  "rawpointer",
 ]
 
+[[package]]
+name = "nohash-hasher"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2bf50223579dc7cdcfb3bfcacf7069ff68243f8c363f62ffa99cf000a6b9c451"
+
 [[package]]
 name = "nom"
 version = "7.1.3"
@@ -5628,11 +5434,10 @@ dependencies = [
 
 [[package]]
 name = "num-bigint-dig"
-version = "0.8.4"
+version = "0.8.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dc84195820f291c7697304f3cbdadd1cb7199c0efc917ff5eafd71225c136151"
+checksum = "e661dda6640fad38e827a6d4a310ff4763082116fe217f279885c97f511bb0b7"
 dependencies = [
- "byteorder",
  "lazy_static",
  "libm",
  "num-integer",
@@ -7274,20 +7079,14 @@ dependencies = [
 
 [[package]]
 name = "roaring"
-version = "0.10.12"
+version = "0.11.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "19e8d2cfa184d94d0726d650a9f4a1be7f9b76ac9fdb954219878dc00c1c1e7b"
+checksum = "8ba9ce64a8f45d7fc86358410bb1a82e8c987504c0d4900e9141d69a9f26c885"
 dependencies = [
  "bytemuck",
  "byteorder",
 ]
 
-[[package]]
-name = "robust"
-version = "1.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4e27ee8bb91ca0adcf0ecb116293afa12d393f9c2b9b9cd54d33e8078fe19839"
-
 [[package]]
 name = "rsa"
 version = "0.9.8"
@@ -7309,17 +7108,6 @@ dependencies = [
  "zeroize",
 ]
 
-[[package]]
-name = "rstar"
-version = "0.12.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "421400d13ccfd26dfa5858199c30a5d76f9c54e0dba7575273025b43c5175dbb"
-dependencies = [
- "heapless",
- "num-traits",
- "smallvec",
-]
-
 [[package]]
 name = "rstest"
 version = "0.23.0"
@@ -7874,15 +7662,6 @@ dependencies = [
  "lazy_static",
 ]
 
-[[package]]
-name = "shellexpand"
-version = "3.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8b1fdf65dd6331831494dd616b30351c38e96e45921a27745cf98490458b90bb"
-dependencies = [
- "dirs",
-]
-
 [[package]]
 name = "shlex"
 version = "1.3.0"
@@ -8032,18 +7811,6 @@ dependencies = [
  "winapi",
 ]
 
-[[package]]
-name = "spade"
-version = "2.15.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fb313e1c8afee5b5647e00ee0fe6855e3d529eb863a0fdae1d60006c4d1e9990"
-dependencies = [
- "hashbrown 0.15.5",
- "num-traits",
- "robust",
- "smallvec",
-]
-
 [[package]]
 name = "spin"
 version = "0.9.8"
@@ -8938,7 +8705,7 @@ checksum = "90b70b37e9074642bc5f60bb23247fd072a84314ca9e71cdf8527593406a0dd3"
 dependencies = [
  "gemm 0.18.2",
  "half",
- "libloading",
+ "libloading 0.8.9",
  "memmap2 0.9.8",
  "num",
  "num-traits",
@@ -9638,31 +9405,6 @@ version = "0.46.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f17a85883d4e6d00e8a97c586de764dabcc06133f7f1d55dce5cdc070ad7fe59"
 
-[[package]]
-name = "wkb"
-version = "0.9.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a120b336c7ad17749026d50427c23d838ecb50cd64aaea6254b5030152f890a9"
-dependencies = [
- "byteorder",
- "geo-traits",
- "num_enum",
- "thiserror 1.0.69",
-]
-
-[[package]]
-name = "wkt"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "efb2b923ccc882312e559ffaa832a055ba9d1ac0cc8e86b3e25453247e4b81d7"
-dependencies = [
- "geo-traits",
- "geo-types",
- "log",
- "num-traits",
- "thiserror 1.0.69",
-]
-
 [[package]]
 name = "writeable"
 version = "0.6.1"

Cargo.toml

@@ -12,23 +12,23 @@ repository = "https://github.com/lancedb/lancedb"
 description = "Serverless, low-latency vector database for AI applications"
 keywords = ["lancedb", "lance", "database", "vector", "search"]
 categories = ["database-implementations"]
-rust-version = "1.88.0"
+rust-version = "1.91.0"
 
 [workspace.dependencies]
-lance = { "version" = "=2.0.0", default-features = false }
-lance-core = "=2.0.0"
-lance-datagen = "=2.0.0"
-lance-file = "=2.0.0"
-lance-io = { "version" = "=2.0.0", default-features = false }
-lance-index = "=2.0.0"
-lance-linalg = "=2.0.0"
-lance-namespace = "=2.0.0"
-lance-namespace-impls = { "version" = "=2.0.0", default-features = false }
-lance-table = "=2.0.0"
-lance-testing = "=2.0.0"
-lance-datafusion = "=2.0.0"
-lance-encoding = "=2.0.0"
-lance-arrow = "=2.0.0"
+lance = { "version" = "=3.1.0-beta.2", default-features = false, "tag" = "v3.1.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-core = { "version" = "=3.1.0-beta.2", "tag" = "v3.1.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-datagen = { "version" = "=3.1.0-beta.2", "tag" = "v3.1.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-file = { "version" = "=3.1.0-beta.2", "tag" = "v3.1.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-io = { "version" = "=3.1.0-beta.2", default-features = false, "tag" = "v3.1.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-index = { "version" = "=3.1.0-beta.2", "tag" = "v3.1.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-linalg = { "version" = "=3.1.0-beta.2", "tag" = "v3.1.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-namespace = { "version" = "=3.1.0-beta.2", "tag" = "v3.1.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-namespace-impls = { "version" = "=3.1.0-beta.2", default-features = false, "tag" = "v3.1.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-table = { "version" = "=3.1.0-beta.2", "tag" = "v3.1.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-testing = { "version" = "=3.1.0-beta.2", "tag" = "v3.1.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-datafusion = { "version" = "=3.1.0-beta.2", "tag" = "v3.1.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-encoding = { "version" = "=3.1.0-beta.2", "tag" = "v3.1.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-arrow = { "version" = "=3.1.0-beta.2", "tag" = "v3.1.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
 ahash = "0.8"
 # Note that this one does not include pyarrow
 arrow = { version = "57.2", optional = false }

docs/src/java/java.md

@@ -14,7 +14,7 @@ Add the following dependency to your `pom.xml`:
 <dependency>
     <groupId>com.lancedb</groupId>
     <artifactId>lancedb-core</artifactId>
-    <version>0.26.2</version>
+    <version>0.27.0-beta.1</version>
 </dependency>
 ```
 

java/lancedb-core/pom.xml

@@ -8,7 +8,7 @@
     <parent>
       <groupId>com.lancedb</groupId>
       <artifactId>lancedb-parent</artifactId>
-      <version>0.26.2-final.0</version>
+      <version>0.27.0-beta.1</version>
       <relativePath>../pom.xml</relativePath>
     </parent>
 

java/pom.xml

@@ -6,7 +6,7 @@
 
     <groupId>com.lancedb</groupId>
     <artifactId>lancedb-parent</artifactId>
-    <version>0.26.2-final.0</version>
+    <version>0.27.0-beta.1</version>
     <packaging>pom</packaging>
     <name>${project.artifactId}</name>
     <description>LanceDB Java SDK Parent POM</description>
@@ -28,7 +28,7 @@
     <properties>
         <project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
         <arrow.version>15.0.0</arrow.version>
-        <lance-core.version>2.0.0</lance-core.version>
+        <lance-core.version>3.1.0-beta.2</lance-core.version>
         <spotless.skip>false</spotless.skip>
         <spotless.version>2.30.0</spotless.version>
         <spotless.java.googlejavaformat.version>1.7</spotless.java.googlejavaformat.version>

nodejs/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
 name = "lancedb-nodejs"
 edition.workspace = true
-version = "0.26.2"
+version = "0.27.0-beta.1"
 license.workspace = true
 description.workspace = true
 repository.workspace = true
@@ -19,11 +19,11 @@ arrow-schema.workspace = true
 env_logger.workspace = true
 futures.workspace = true
 lancedb = { path = "../rust/lancedb", default-features = false }
-napi = { version = "2.16.8", default-features = false, features = [
+napi = { version = "3.8.3", default-features = false, features = [
     "napi9",
     "async"
 ] }
-napi-derive = "2.16.4"
+napi-derive = "3.5.2"
 # Prevent dynamic linking of lzma, which comes from datafusion
 lzma-sys = { version = "*", features = ["static"] }
 log.workspace = true
@@ -33,7 +33,7 @@ aws-lc-sys = "=0.28.0"
 aws-lc-rs = "=1.13.0"
 
 [build-dependencies]
-napi-build = "2.1"
+napi-build = "2.3.1"
 
 [features]
 default = ["remote", "lancedb/aws", "lancedb/gcs", "lancedb/azure", "lancedb/dynamodb", "lancedb/oss", "lancedb/huggingface"]

nodejs/lancedb/index.ts

@@ -273,7 +273,9 @@ export async function connect(
   let nativeProvider: NativeJsHeaderProvider | undefined;
   if (finalHeaderProvider) {
     if (typeof finalHeaderProvider === "function") {
-      nativeProvider = new NativeJsHeaderProvider(finalHeaderProvider);
+      nativeProvider = new NativeJsHeaderProvider(async () =>
+        finalHeaderProvider(),
+      );
     } else if (
       finalHeaderProvider &&
       typeof finalHeaderProvider.getHeaders === "function"

nodejs/lancedb/query.ts

@@ -684,19 +684,17 @@ export class VectorQuery extends StandardQueryBase<NativeVectorQuery> {
 
   rerank(reranker: Reranker): VectorQuery {
     super.doCall((inner) =>
-      inner.rerank({
-        rerankHybrid: async (_, args) => {
-          const vecResults = await fromBufferToRecordBatch(args.vecResults);
-          const ftsResults = await fromBufferToRecordBatch(args.ftsResults);
-          const result = await reranker.rerankHybrid(
-            args.query,
-            vecResults as RecordBatch,
-            ftsResults as RecordBatch,
-          );
+      inner.rerank(async (args) => {
+        const vecResults = await fromBufferToRecordBatch(args.vecResults);
+        const ftsResults = await fromBufferToRecordBatch(args.ftsResults);
+        const result = await reranker.rerankHybrid(
+          args.query,
+          vecResults as RecordBatch,
+          ftsResults as RecordBatch,
+        );
 
-          const buffer = fromRecordBatchToBuffer(result);
-          return buffer;
-        },
+        const buffer = fromRecordBatchToBuffer(result);
+        return buffer;
       }),
     );
 

nodejs/npm/darwin-arm64/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-darwin-arm64",
-	"version": "0.26.2",
+	"version": "0.27.0-beta.1",
 	"os": ["darwin"],
 	"cpu": ["arm64"],
 	"main": "lancedb.darwin-arm64.node",

nodejs/npm/linux-arm64-gnu/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-arm64-gnu",
-	"version": "0.26.2",
+	"version": "0.27.0-beta.1",
 	"os": ["linux"],
 	"cpu": ["arm64"],
 	"main": "lancedb.linux-arm64-gnu.node",

nodejs/npm/linux-arm64-musl/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-arm64-musl",
-	"version": "0.26.2",
+	"version": "0.27.0-beta.1",
 	"os": ["linux"],
 	"cpu": ["arm64"],
 	"main": "lancedb.linux-arm64-musl.node",

nodejs/npm/linux-x64-gnu/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-x64-gnu",
-	"version": "0.26.2",
+	"version": "0.27.0-beta.1",
 	"os": ["linux"],
 	"cpu": ["x64"],
 	"main": "lancedb.linux-x64-gnu.node",

nodejs/npm/linux-x64-musl/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-x64-musl",
-	"version": "0.26.2",
+	"version": "0.27.0-beta.1",
 	"os": ["linux"],
 	"cpu": ["x64"],
 	"main": "lancedb.linux-x64-musl.node",

nodejs/npm/win32-arm64-msvc/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lancedb/lancedb-win32-arm64-msvc",
-  "version": "0.26.2",
+  "version": "0.27.0-beta.1",
   "os": [
     "win32"
   ],

nodejs/npm/win32-x64-msvc/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-win32-x64-msvc",
-	"version": "0.26.2",
+	"version": "0.27.0-beta.1",
 	"os": ["win32"],
 	"cpu": ["x64"],
 	"main": "lancedb.win32-x64-msvc.node",

nodejs/package.json

@@ -11,7 +11,7 @@
     "ann"
   ],
   "private": false,
-  "version": "0.26.2",
+  "version": "0.27.0-beta.1",
   "main": "dist/index.js",
   "exports": {
     ".": "./dist/index.js",
@@ -21,19 +21,16 @@
   },
   "types": "dist/index.d.ts",
   "napi": {
-    "name": "lancedb",
-    "triples": {
-      "defaults": false,
-      "additional": [
-        "aarch64-apple-darwin",
-        "x86_64-unknown-linux-gnu",
-        "aarch64-unknown-linux-gnu",
-        "x86_64-unknown-linux-musl",
-        "aarch64-unknown-linux-musl",
-        "x86_64-pc-windows-msvc",
-        "aarch64-pc-windows-msvc"
-      ]
-    }
+    "binaryName": "lancedb",
+    "targets": [
+      "aarch64-apple-darwin",
+      "x86_64-unknown-linux-gnu",
+      "aarch64-unknown-linux-gnu",
+      "x86_64-unknown-linux-musl",
+      "aarch64-unknown-linux-musl",
+      "x86_64-pc-windows-msvc",
+      "aarch64-pc-windows-msvc"
+    ]
   },
   "license": "Apache-2.0",
   "repository": {
@@ -46,7 +43,7 @@
     "@aws-sdk/client-s3": "^3.33.0",
     "@biomejs/biome": "^1.7.3",
     "@jest/globals": "^29.7.0",
-    "@napi-rs/cli": "^2.18.3",
+    "@napi-rs/cli": "^3.5.1",
     "@types/axios": "^0.14.0",
     "@types/jest": "^29.1.2",
     "@types/node": "^22.7.4",
@@ -75,9 +72,9 @@
   "os": ["darwin", "linux", "win32"],
   "scripts": {
     "artifacts": "napi artifacts",
-    "build:debug": "napi build --platform --no-const-enum --dts ../lancedb/native.d.ts --js ../lancedb/native.js lancedb",
+    "build:debug": "napi build --platform --dts ../lancedb/native.d.ts --js ../lancedb/native.js --output-dir lancedb",
     "postbuild:debug": "shx mkdir -p dist && shx cp lancedb/*.node dist/",
-    "build:release": "napi build --platform --no-const-enum --release --dts ../lancedb/native.d.ts --js ../lancedb/native.js dist/",
+    "build:release": "napi build --platform --release --dts ../lancedb/native.d.ts --js ../lancedb/native.js --output-dir dist",
     "postbuild:release": "shx mkdir -p dist && shx cp lancedb/*.node dist/",
     "build": "npm run build:debug && npm run tsc",
     "build-release": "npm run build:release && npm run tsc",
@@ -91,7 +88,7 @@
     "prepublishOnly": "napi prepublish -t npm",
     "test": "jest --verbose",
     "integration": "S3_TEST=1 npm run test",
-    "universal": "napi universal",
+    "universal": "napi universalize",
     "version": "napi version"
   },
   "dependencies": {

nodejs/src/connection.rs

@@ -13,6 +13,7 @@ use crate::header::JsHeaderProvider;
 use crate::table::Table;
 use crate::ConnectionOptions;
 use lancedb::connection::{ConnectBuilder, Connection as LanceDBConnection};
+
 use lancedb::ipc::{ipc_file_to_batches, ipc_file_to_schema};
 
 #[napi]

nodejs/src/header.rs

@@ -1,20 +1,19 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
-use napi::{
-    bindgen_prelude::*,
-    threadsafe_function::{ErrorStrategy, ThreadsafeFunction},
-};
+use napi::{bindgen_prelude::*, threadsafe_function::ThreadsafeFunction};
 use napi_derive::napi;
 use std::collections::HashMap;
 use std::sync::Arc;
 
+type GetHeadersFn = ThreadsafeFunction<(), Promise<HashMap<String, String>>, (), Status, false>;
+
 /// JavaScript HeaderProvider implementation that wraps a JavaScript callback.
 /// This is the only native header provider - all header provider implementations
 /// should provide a JavaScript function that returns headers.
 #[napi]
 pub struct JsHeaderProvider {
-    get_headers_fn: Arc<ThreadsafeFunction<(), ErrorStrategy::CalleeHandled>>,
+    get_headers_fn: Arc<GetHeadersFn>,
 }
 
 impl Clone for JsHeaderProvider {
@@ -29,9 +28,12 @@ impl Clone for JsHeaderProvider {
 impl JsHeaderProvider {
     /// Create a new JsHeaderProvider from a JavaScript callback
     #[napi(constructor)]
-    pub fn new(get_headers_callback: JsFunction) -> Result<Self> {
+    pub fn new(
+        get_headers_callback: Function<(), Promise<HashMap<String, String>>>,
+    ) -> Result<Self> {
         let get_headers_fn = get_headers_callback
-            .create_threadsafe_function(0, |ctx| Ok(vec![ctx.value]))
+            .build_threadsafe_function()
+            .build()
             .map_err(|e| {
                 Error::new(
                     Status::GenericFailure,
@@ -51,7 +53,7 @@ impl lancedb::remote::HeaderProvider for JsHeaderProvider {
     async fn get_headers(&self) -> lancedb::error::Result<HashMap<String, String>> {
         // Call the JavaScript function asynchronously
         let promise: Promise<HashMap<String, String>> =
-            self.get_headers_fn.call_async(Ok(())).await.map_err(|e| {
+            self.get_headers_fn.call_async(()).await.map_err(|e| {
                 lancedb::error::Error::Runtime {
                     message: format!("Failed to call JavaScript get_headers: {}", e),
                 }

nodejs/src/lib.rs

@@ -60,7 +60,7 @@ pub struct OpenTableOptions {
     pub storage_options: Option<HashMap<String, String>>,
 }
 
-#[napi::module_init]
+#[napi_derive::module_init]
 fn init() {
     let env = Env::new()
         .filter_or("LANCEDB_LOG", "warn")

nodejs/src/query.rs

@@ -20,8 +20,8 @@ use napi_derive::napi;
 use crate::error::convert_error;
 use crate::error::NapiErrorExt;
 use crate::iterator::RecordBatchIterator;
+use crate::rerankers::RerankHybridCallbackArgs;
 use crate::rerankers::Reranker;
-use crate::rerankers::RerankerCallbacks;
 use crate::util::{parse_distance_type, schema_to_buffer};
 
 #[napi]
@@ -42,7 +42,7 @@ impl Query {
     }
 
     #[napi]
-    pub fn full_text_search(&mut self, query: napi::JsObject) -> napi::Result<()> {
+    pub fn full_text_search(&mut self, query: Object) -> napi::Result<()> {
         let query = parse_fts_query(query)?;
         self.inner = self.inner.clone().full_text_search(query);
         Ok(())
@@ -235,7 +235,7 @@ impl VectorQuery {
     }
 
     #[napi]
-    pub fn full_text_search(&mut self, query: napi::JsObject) -> napi::Result<()> {
+    pub fn full_text_search(&mut self, query: Object) -> napi::Result<()> {
         let query = parse_fts_query(query)?;
         self.inner = self.inner.clone().full_text_search(query);
         Ok(())
@@ -272,11 +272,13 @@ impl VectorQuery {
     }
 
     #[napi]
-    pub fn rerank(&mut self, callbacks: RerankerCallbacks) {
-        self.inner = self
-            .inner
-            .clone()
-            .rerank(Arc::new(Reranker::new(callbacks)));
+    pub fn rerank(
+        &mut self,
+        rerank_hybrid: Function<RerankHybridCallbackArgs, Promise<Buffer>>,
+    ) -> napi::Result<()> {
+        let reranker = Reranker::new(rerank_hybrid)?;
+        self.inner = self.inner.clone().rerank(Arc::new(reranker));
+        Ok(())
     }
 
     #[napi(catch_unwind)]
@@ -523,12 +525,12 @@ impl JsFullTextQuery {
     }
 }
 
-fn parse_fts_query(query: napi::JsObject) -> napi::Result<FullTextSearchQuery> {
-    if let Ok(Some(query)) = query.get::<_, &JsFullTextQuery>("query") {
+fn parse_fts_query(query: Object) -> napi::Result<FullTextSearchQuery> {
+    if let Ok(Some(query)) = query.get::<&JsFullTextQuery>("query") {
         Ok(FullTextSearchQuery::new_query(query.inner.clone()))
-    } else if let Ok(Some(query_text)) = query.get::<_, String>("query") {
+    } else if let Ok(Some(query_text)) = query.get::<String>("query") {
         let mut query_text = query_text;
-        let columns = query.get::<_, Option<Vec<String>>>("columns")?.flatten();
+        let columns = query.get::<Option<Vec<String>>>("columns")?.flatten();
 
         let is_phrase =
             query_text.len() >= 2 && query_text.starts_with('"') && query_text.ends_with('"');

nodejs/src/rerankers.rs

@@ -3,10 +3,7 @@
 
 use arrow_array::RecordBatch;
 use async_trait::async_trait;
-use napi::{
-    bindgen_prelude::*,
-    threadsafe_function::{ErrorStrategy, ThreadsafeFunction},
-};
+use napi::{bindgen_prelude::*, threadsafe_function::ThreadsafeFunction};
 use napi_derive::napi;
 
 use lancedb::ipc::batches_to_ipc_file;
@@ -15,27 +12,28 @@ use lancedb::{error::Error, ipc::ipc_file_to_batches};
 
 use crate::error::NapiErrorExt;
 
+type RerankHybridFn = ThreadsafeFunction<
+    RerankHybridCallbackArgs,
+    Promise<Buffer>,
+    RerankHybridCallbackArgs,
+    Status,
+    false,
+>;
+
 /// Reranker implementation that "wraps" a NodeJS Reranker implementation.
 /// This contains references to the callbacks that can be used to invoke the
 /// reranking methods on the NodeJS implementation and handles serializing the
 /// record batches to Arrow IPC buffers.
-#[napi]
 pub struct Reranker {
-    /// callback to the Javascript which will call the rerankHybrid method of
-    /// some Reranker implementation
-    rerank_hybrid: ThreadsafeFunction<RerankHybridCallbackArgs, ErrorStrategy::CalleeHandled>,
+    rerank_hybrid: RerankHybridFn,
 }
 
-#[napi]
 impl Reranker {
-    #[napi]
-    pub fn new(callbacks: RerankerCallbacks) -> Self {
-        let rerank_hybrid = callbacks
-            .rerank_hybrid
-            .create_threadsafe_function(0, move |ctx| Ok(vec![ctx.value]))
-            .unwrap();
-
-        Self { rerank_hybrid }
+    pub fn new(
+        rerank_hybrid: Function<RerankHybridCallbackArgs, Promise<Buffer>>,
+    ) -> napi::Result<Self> {
+        let rerank_hybrid = rerank_hybrid.build_threadsafe_function().build()?;
+        Ok(Self { rerank_hybrid })
     }
 }
 
@@ -49,16 +47,16 @@ impl lancedb::rerankers::Reranker for Reranker {
     ) -> lancedb::error::Result<RecordBatch> {
         let callback_args = RerankHybridCallbackArgs {
             query: query.to_string(),
-            vec_results: batches_to_ipc_file(&[vector_results])?,
-            fts_results: batches_to_ipc_file(&[fts_results])?,
+            vec_results: Buffer::from(batches_to_ipc_file(&[vector_results])?.as_ref()),
+            fts_results: Buffer::from(batches_to_ipc_file(&[fts_results])?.as_ref()),
         };
         let promised_buffer: Promise<Buffer> = self
             .rerank_hybrid
-            .call_async(Ok(callback_args))
+            .call_async(callback_args)
             .await
             .map_err(|e| Error::Runtime {
-                message: format!("napi error status={}, reason={}", e.status, e.reason),
-            })?;
+            message: format!("napi error status={}, reason={}", e.status, e.reason),
+        })?;
         let buffer = promised_buffer.await.map_err(|e| Error::Runtime {
             message: format!("napi error status={}, reason={}", e.status, e.reason),
         })?;
@@ -77,16 +75,11 @@ impl std::fmt::Debug for Reranker {
     }
 }
 
-#[napi(object)]
-pub struct RerankerCallbacks {
-    pub rerank_hybrid: JsFunction,
-}
-
 #[napi(object)]
 pub struct RerankHybridCallbackArgs {
     pub query: String,
-    pub vec_results: Vec<u8>,
-    pub fts_results: Vec<u8>,
+    pub vec_results: Buffer,
+    pub fts_results: Buffer,
 }
 
 fn buffer_to_record_batch(buffer: Buffer) -> Result<RecordBatch> {

nodejs/src/session.rs

@@ -96,7 +96,6 @@ impl napi::bindgen_prelude::FromNapiValue for Session {
     ) -> napi::Result<Self> {
         let object: napi::bindgen_prelude::ClassInstance<Self> =
             napi::bindgen_prelude::ClassInstance::from_napi_value(env, napi_val)?;
-        let copy = object.clone();
-        Ok(copy)
+        Ok((*object).clone())
     }
 }

nodejs/src/table.rs

@@ -71,6 +71,17 @@ impl Table {
     pub async fn add(&self, buf: Buffer, mode: String) -> napi::Result<AddResult> {
         let batches = ipc_file_to_batches(buf.to_vec())
             .map_err(|e| napi::Error::from_reason(format!("Failed to read IPC file: {}", e)))?;
+        let batches = batches
+            .into_iter()
+            .map(|batch| {
+                batch.map_err(|e| {
+                    napi::Error::from_reason(format!(
+                        "Failed to read record batch from IPC file: {}",
+                        e
+                    ))
+                })
+            })
+            .collect::<Result<Vec<_>>>()?;
         let mut op = self.inner_ref()?.add(batches);
 
         op = if mode == "append" {

python/.bumpversion.toml

@@ -1,5 +1,5 @@
 [tool.bumpversion]
-current_version = "0.29.2"
+current_version = "0.30.0-beta.2"
 parse = """(?x)
     (?P<major>0|[1-9]\\d*)\\.
     (?P<minor>0|[1-9]\\d*)\\.

python/Cargo.toml

@@ -1,13 +1,13 @@
 [package]
 name = "lancedb-python"
-version = "0.29.2"
+version = "0.30.0-beta.2"
 edition.workspace = true
 description = "Python bindings for LanceDB"
 license.workspace = true
 repository.workspace = true
 keywords.workspace = true
 categories.workspace = true
-rust-version = "1.88.0"
+rust-version = "1.91.0"
 
 [lib]
 name = "_lancedb"

python/pyproject.toml

@@ -61,7 +61,7 @@ tests = [
     "pyarrow-stubs",
     "pylance>=1.0.0b14",
     "requests",
-    "datafusion",
+    "datafusion<52",
 ]
 dev = [
     "ruff",

python/python/lancedb/arrow.py

@@ -1,8 +1,10 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
+from functools import singledispatch
 from typing import List, Optional, Tuple, Union
 
+from lancedb.pydantic import LanceModel, model_to_dict
 import pyarrow as pa
 
 from ._lancedb import RecordBatchStream
@@ -80,3 +82,32 @@ def peek_reader(
         yield from reader
 
     return batch, pa.RecordBatchReader.from_batches(batch.schema, all_batches())
+
+
+@singledispatch
+def to_arrow(data) -> pa.Table:
+    """Convert a single data object to a pa.Table."""
+    raise NotImplementedError(f"to_arrow not implemented for type {type(data)}")
+
+
+@to_arrow.register(pa.RecordBatch)
+def _arrow_from_batch(data: pa.RecordBatch) -> pa.Table:
+    return pa.Table.from_batches([data])
+
+
+@to_arrow.register(pa.Table)
+def _arrow_from_table(data: pa.Table) -> pa.Table:
+    return data
+
+
+@to_arrow.register(list)
+def _arrow_from_list(data: list) -> pa.Table:
+    if not data:
+        raise ValueError("Cannot create table from empty list without a schema")
+
+    if isinstance(data[0], LanceModel):
+        schema = data[0].__class__.to_arrow_schema()
+        dicts = [model_to_dict(d) for d in data]
+        return pa.Table.from_pylist(dicts, schema=schema)
+
+    return pa.Table.from_pylist(data)

python/python/lancedb/embeddings/gte.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
 
+import warnings
 from typing import List, Union
 
 import numpy as np
@@ -15,6 +16,8 @@ from .utils import weak_lru
 @register("gte-text")
 class GteEmbeddings(TextEmbeddingFunction):
     """
+    Deprecated: GTE embeddings should be used through sentence-transformers.
+
     An embedding function that uses GTE-LARGE MLX format(for Apple silicon devices only)
     as well as the standard cpu/gpu version from: https://huggingface.co/thenlper/gte-large.
 
@@ -61,6 +64,13 @@ class GteEmbeddings(TextEmbeddingFunction):
 
     def __init__(self, **kwargs):
         super().__init__(**kwargs)
+        warnings.warn(
+            "GTE embeddings as a standalone embedding function are deprecated. "
+            "Use the 'sentence-transformers' embedding function with a GTE model "
+            "instead.",
+            DeprecationWarning,
+            stacklevel=3,
+        )
         self._ndims = None
         if kwargs:
             self.mlx = kwargs.get("mlx", False)

python/python/lancedb/embeddings/openai.py

@@ -110,6 +110,9 @@ class OpenAIEmbeddings(TextEmbeddingFunction):
             valid_embeddings = {
                 idx: v.embedding for v, idx in zip(rs.data, valid_indices)
             }
+        except openai.AuthenticationError:
+            logging.error("Authentication failed: Invalid API key provided")
+            raise
         except openai.BadRequestError:
             logging.exception("Bad request: %s", texts)
             return [None] * len(texts)

python/python/lancedb/embeddings/siglip.py

@@ -6,6 +6,7 @@ import io
 import os
 from typing import TYPE_CHECKING, List, Union
 import urllib.parse as urlparse
+import warnings
 
 import numpy as np
 import pyarrow as pa
@@ -24,6 +25,7 @@ if TYPE_CHECKING:
 
 @register("siglip")
 class SigLipEmbeddings(EmbeddingFunction):
+    # Deprecated: prefer CLIP embeddings via `open-clip`.
     model_name: str = "google/siglip-base-patch16-224"
     device: str = "cpu"
     batch_size: int = 64
@@ -36,6 +38,12 @@ class SigLipEmbeddings(EmbeddingFunction):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
+        warnings.warn(
+            "SigLip embeddings are deprecated. Use CLIP embeddings via the "
+            "'open-clip' embedding function instead.",
+            DeprecationWarning,
+            stacklevel=3,
+        )
         transformers = attempt_import_or_raise("transformers")
         self._torch = attempt_import_or_raise("torch")
 

python/python/lancedb/embeddings/utils.py

@@ -269,6 +269,11 @@ def retry_with_exponential_backoff(
             # and say that it is assumed that if this portion errors out, it's due
             # to rate limit but the user should check the error message to be sure.
             except Exception as e:  # noqa: PERF203
+                # Don't retry on authentication errors (e.g., OpenAI 401)
+                # These are permanent failures that won't be fixed by retrying
+                if _is_non_retryable_error(e):
+                    raise
+
                 num_retries += 1
 
                 if num_retries > max_retries:
@@ -289,6 +294,29 @@ def retry_with_exponential_backoff(
     return wrapper
 
 
+def _is_non_retryable_error(error: Exception) -> bool:
+    """Check if an error should not be retried.
+
+    Args:
+        error: The exception to check
+
+    Returns:
+        True if the error should not be retried, False otherwise
+    """
+    # Check for OpenAI authentication errors
+    error_type = type(error).__name__
+    if error_type == "AuthenticationError":
+        return True
+
+    # Check for other common non-retryable HTTP status codes
+    # 401 Unauthorized, 403 Forbidden
+    if hasattr(error, "status_code"):
+        if error.status_code in (401, 403):
+            return True
+
+    return False
+
+
 def url_retrieve(url: str):
     """
     Parameters

python/python/lancedb/namespace.py

@@ -44,7 +44,7 @@ from lance_namespace import (
     ListNamespacesRequest,
     CreateNamespaceRequest,
     DropNamespaceRequest,
-    CreateEmptyTableRequest,
+    DeclareTableRequest,
 )
 from lancedb.table import AsyncTable, LanceTable, Table
 from lancedb.util import validate_table_name
@@ -318,20 +318,20 @@ class LanceNamespaceDBConnection(DBConnection):
 
         if location is None:
             # Table doesn't exist or mode is "create", reserve a new location
-            create_empty_request = CreateEmptyTableRequest(
+            declare_request = DeclareTableRequest(
                 id=table_id,
                 location=None,
                 properties=self.storage_options if self.storage_options else None,
             )
-            create_empty_response = self._ns.create_empty_table(create_empty_request)
+            declare_response = self._ns.declare_table(declare_request)
 
-            if not create_empty_response.location:
+            if not declare_response.location:
                 raise ValueError(
-                    "Table location is missing from create_empty_table response"
+                    "Table location is missing from declare_table response"
                 )
 
-            location = create_empty_response.location
-            namespace_storage_options = create_empty_response.storage_options
+            location = declare_response.location
+            namespace_storage_options = declare_response.storage_options
 
         # Merge storage options: self.storage_options < user options < namespace options
         merged_storage_options = dict(self.storage_options)
@@ -759,20 +759,20 @@ class AsyncLanceNamespaceDBConnection:
 
         if location is None:
             # Table doesn't exist or mode is "create", reserve a new location
-            create_empty_request = CreateEmptyTableRequest(
+            declare_request = DeclareTableRequest(
                 id=table_id,
                 location=None,
                 properties=self.storage_options if self.storage_options else None,
             )
-            create_empty_response = self._ns.create_empty_table(create_empty_request)
+            declare_response = self._ns.declare_table(declare_request)
 
-            if not create_empty_response.location:
+            if not declare_response.location:
                 raise ValueError(
-                    "Table location is missing from create_empty_table response"
+                    "Table location is missing from declare_table response"
                 )
 
-            location = create_empty_response.location
-            namespace_storage_options = create_empty_response.storage_options
+            location = declare_response.location
+            namespace_storage_options = declare_response.storage_options
 
         # Merge storage options: self.storage_options < user options < namespace options
         merged_storage_options = dict(self.storage_options)

python/python/lancedb/query.py

@@ -1782,6 +1782,26 @@ class LanceHybridQueryBuilder(LanceQueryBuilder):
             vector_results = LanceHybridQueryBuilder._rank(vector_results, "_distance")
             fts_results = LanceHybridQueryBuilder._rank(fts_results, "_score")
 
+        # If both result sets are empty (e.g. after hard filtering),
+        # return early to avoid errors in reranking or score restoration.
+        if vector_results.num_rows == 0 and fts_results.num_rows == 0:
+            # Build a minimal empty table with the _relevance_score column
+            combined_schema = pa.unify_schemas(
+                [vector_results.schema, fts_results.schema],
+            )
+            empty = pa.table(
+                {
+                    col: pa.array([], type=combined_schema.field(col).type)
+                    for col in combined_schema.names
+                }
+            )
+            empty = empty.append_column(
+                "_relevance_score", pa.array([], type=pa.float32())
+            )
+            if not with_row_ids and "_rowid" in empty.column_names:
+                empty = empty.drop(["_rowid"])
+            return empty
+
         original_distances = None
         original_scores = None
         original_distance_row_ids = None

python/python/lancedb/scannable.py

@@ -0,0 +1,214 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+from dataclasses import dataclass
+from functools import singledispatch
+import sys
+from typing import Callable, Iterator, Optional
+from lancedb.arrow import to_arrow
+import pyarrow as pa
+import pyarrow.dataset as ds
+
+from .pydantic import LanceModel
+
+
+@dataclass
+class Scannable:
+    schema: pa.Schema
+    num_rows: Optional[int]
+    # Factory function to create a new reader each time (supports re-scanning)
+    reader: Callable[[], pa.RecordBatchReader]
+    # Whether reader can be called more than once. For example, an iterator can
+    # only be consumed once, while a DataFrame can be converted to a new reader
+    # each time.
+    rescannable: bool = True
+
+
+@singledispatch
+def to_scannable(data) -> Scannable:
+    # Fallback: try iterable protocol
+    if hasattr(data, "__iter__"):
+        return _from_iterable(iter(data))
+    raise NotImplementedError(f"to_scannable not implemented for type {type(data)}")
+
+
+@to_scannable.register(pa.RecordBatchReader)
+def _from_reader(data: pa.RecordBatchReader) -> Scannable:
+    # RecordBatchReader can only be consumed once - not rescannable
+    return Scannable(
+        schema=data.schema, num_rows=None, reader=lambda: data, rescannable=False
+    )
+
+
+@to_scannable.register(pa.RecordBatch)
+def _from_batch(data: pa.RecordBatch) -> Scannable:
+    return Scannable(
+        schema=data.schema,
+        num_rows=data.num_rows,
+        reader=lambda: pa.RecordBatchReader.from_batches(data.schema, [data]),
+    )
+
+
+@to_scannable.register(pa.Table)
+def _from_table(data: pa.Table) -> Scannable:
+    return Scannable(schema=data.schema, num_rows=data.num_rows, reader=data.to_reader)
+
+
+@to_scannable.register(ds.Dataset)
+def _from_dataset(data: ds.Dataset) -> Scannable:
+    return Scannable(
+        schema=data.schema,
+        num_rows=data.count_rows(),
+        reader=lambda: data.scanner().to_reader(),
+    )
+
+
+@to_scannable.register(ds.Scanner)
+def _from_scanner(data: ds.Scanner) -> Scannable:
+    # Scanner can only be consumed once - not rescannable
+    return Scannable(
+        schema=data.projected_schema,
+        num_rows=None,
+        reader=data.to_reader,
+        rescannable=False,
+    )
+
+
+@to_scannable.register(list)
+def _from_list(data: list) -> Scannable:
+    if not data:
+        raise ValueError("Cannot create table from empty list without a schema")
+    table = to_arrow(data)
+    return Scannable(
+        schema=table.schema, num_rows=table.num_rows, reader=table.to_reader
+    )
+
+
+@to_scannable.register(dict)
+def _from_dict(data: dict) -> Scannable:
+    raise ValueError("Cannot add a single dictionary to a table. Use a list.")
+
+
+@to_scannable.register(LanceModel)
+def _from_lance_model(data: LanceModel) -> Scannable:
+    raise ValueError("Cannot add a single LanceModel to a table. Use a list.")
+
+
+def _from_iterable(data: Iterator) -> Scannable:
+    first_item = next(data, None)
+    if first_item is None:
+        raise ValueError("Cannot create table from empty iterator")
+    first = to_arrow(first_item)
+    schema = first.schema
+
+    def iter():
+        yield from first.to_batches()
+        for item in data:
+            batch = to_arrow(item)
+            if batch.schema != schema:
+                try:
+                    batch = batch.cast(schema)
+                except pa.lib.ArrowInvalid:
+                    raise ValueError(
+                        f"Input iterator yielded a batch with schema that "
+                        f"does not match the schema of other batches.\n"
+                        f"Expected:\n{schema}\nGot:\n{batch.schema}"
+                    )
+            yield from batch.to_batches()
+
+    reader = pa.RecordBatchReader.from_batches(schema, iter())
+    return to_scannable(reader)
+
+
+_registered_modules: set[str] = set()
+
+
+def _register_optional_converters():
+    """Register converters for optional dependencies that are already imported."""
+
+    if "pandas" in sys.modules and "pandas" not in _registered_modules:
+        _registered_modules.add("pandas")
+        import pandas as pd
+
+        @to_arrow.register(pd.DataFrame)
+        def _arrow_from_pandas(data: pd.DataFrame) -> pa.Table:
+            table = pa.Table.from_pandas(data, preserve_index=False)
+            return table.replace_schema_metadata(None)
+
+        @to_scannable.register(pd.DataFrame)
+        def _from_pandas(data: pd.DataFrame) -> Scannable:
+            return to_scannable(_arrow_from_pandas(data))
+
+    if "polars" in sys.modules and "polars" not in _registered_modules:
+        _registered_modules.add("polars")
+        import polars as pl
+
+        @to_arrow.register(pl.DataFrame)
+        def _arrow_from_polars(data: pl.DataFrame) -> pa.Table:
+            return data.to_arrow()
+
+        @to_scannable.register(pl.DataFrame)
+        def _from_polars(data: pl.DataFrame) -> Scannable:
+            arrow = data.to_arrow()
+            return Scannable(
+                schema=arrow.schema, num_rows=len(data), reader=arrow.to_reader
+            )
+
+        @to_scannable.register(pl.LazyFrame)
+        def _from_polars_lazy(data: pl.LazyFrame) -> Scannable:
+            arrow = data.collect().to_arrow()
+            return Scannable(
+                schema=arrow.schema, num_rows=arrow.num_rows, reader=arrow.to_reader
+            )
+
+    if "datasets" in sys.modules and "datasets" not in _registered_modules:
+        _registered_modules.add("datasets")
+        from datasets import Dataset as HFDataset
+        from datasets import DatasetDict as HFDatasetDict
+
+        @to_scannable.register(HFDataset)
+        def _from_hf_dataset(data: HFDataset) -> Scannable:
+            table = data.data.table  # Access underlying Arrow table
+            return Scannable(
+                schema=table.schema, num_rows=len(data), reader=table.to_reader
+            )
+
+        @to_scannable.register(HFDatasetDict)
+        def _from_hf_dataset_dict(data: HFDatasetDict) -> Scannable:
+            # HuggingFace DatasetDict: combine all splits with a 'split' column
+            schema = data[list(data.keys())[0]].features.arrow_schema
+            if "split" not in schema.names:
+                schema = schema.append(pa.field("split", pa.string()))
+
+            def gen():
+                for split_name, dataset in data.items():
+                    for batch in dataset.data.to_batches():
+                        split_arr = pa.array(
+                            [split_name] * len(batch), type=pa.string()
+                        )
+                        yield pa.RecordBatch.from_arrays(
+                            list(batch.columns) + [split_arr], schema=schema
+                        )
+
+            total_rows = sum(len(dataset) for dataset in data.values())
+            return Scannable(
+                schema=schema,
+                num_rows=total_rows,
+                reader=lambda: pa.RecordBatchReader.from_batches(schema, gen()),
+            )
+
+    if "lance" in sys.modules and "lance" not in _registered_modules:
+        _registered_modules.add("lance")
+        import lance
+
+        @to_scannable.register(lance.LanceDataset)
+        def _from_lance(data: lance.LanceDataset) -> Scannable:
+            return Scannable(
+                schema=data.schema,
+                num_rows=data.count_rows(),
+                reader=lambda: data.scanner().to_reader(),
+            )
+
+
+# Register on module load
+_register_optional_converters()

python/python/lancedb/table.py

@@ -25,6 +25,8 @@ from typing import (
 )
 from urllib.parse import urlparse
 
+from lancedb.scannable import _register_optional_converters, to_scannable
+
 from . import __version__
 from lancedb.arrow import peek_reader
 from lancedb.background_loop import LOOP
@@ -3727,18 +3729,31 @@ class AsyncTable:
             on_bad_vectors = "error"
         if fill_value is None:
             fill_value = 0.0
-        data = _sanitize_data(
-            data,
-            schema,
-            metadata=schema.metadata,
-            on_bad_vectors=on_bad_vectors,
-            fill_value=fill_value,
-            allow_subschema=True,
-        )
-        if isinstance(data, pa.Table):
-            data = data.to_reader()
 
-        return await self._inner.add(data, mode or "append")
+        # _santitize_data is an old code path, but we will use it until the
+        # new code path is ready.
+        if on_bad_vectors != "error" or (
+            schema.metadata is not None and b"embedding_functions" in schema.metadata
+        ):
+            data = _sanitize_data(
+                data,
+                schema,
+                metadata=schema.metadata,
+                on_bad_vectors=on_bad_vectors,
+                fill_value=fill_value,
+                allow_subschema=True,
+            )
+        _register_optional_converters()
+        data = to_scannable(data)
+        try:
+            return await self._inner.add(data, mode or "append")
+        except RuntimeError as e:
+            if "Cast error" in str(e):
+                raise ValueError(e)
+            elif "Vector column contains NaN" in str(e):
+                raise ValueError(e)
+            else:
+                raise
 
     def merge_insert(self, on: Union[str, Iterable[str]]) -> LanceMergeInsertBuilder:
         """

python/python/tests/test_embeddings.py

@@ -515,3 +515,34 @@ def test_openai_propagates_api_key(monkeypatch):
     query = "greetings"
     actual = table.search(query).limit(1).to_pydantic(Words)[0]
     assert len(actual.text) > 0
+
+
+@patch("time.sleep")
+def test_openai_no_retry_on_401(mock_sleep):
+    """
+    Test that OpenAI embedding function does not retry on 401 authentication
+    errors.
+    """
+    from lancedb.embeddings.utils import retry_with_exponential_backoff
+
+    # Create a mock that raises an AuthenticationError
+    class MockAuthenticationError(Exception):
+        """Mock OpenAI AuthenticationError"""
+
+        pass
+
+    MockAuthenticationError.__name__ = "AuthenticationError"
+
+    mock_func = MagicMock(side_effect=MockAuthenticationError("Invalid API key"))
+
+    # Wrap the function with retry logic
+    wrapped_func = retry_with_exponential_backoff(mock_func, max_retries=3)
+
+    # Should raise without retrying
+    with pytest.raises(MockAuthenticationError):
+        wrapped_func()
+
+    # Verify that the function was only called once (no retries)
+    assert mock_func.call_count == 1
+    # Verify that sleep was never called (no retries)
+    assert mock_sleep.call_count == 0

python/python/tests/test_rerankers.py

@@ -531,6 +531,78 @@ def test_empty_result_reranker():
         )
 
 
+def test_empty_hybrid_result_reranker():
+    """Test that hybrid search with empty results after filtering doesn't crash.
+
+    Regression test for https://github.com/lancedb/lancedb/issues/2425
+    """
+    from lancedb.query import LanceHybridQueryBuilder
+
+    # Simulate empty vector and FTS results with the expected schema
+    vector_schema = pa.schema(
+        [
+            ("text", pa.string()),
+            ("vector", pa.list_(pa.float32(), 4)),
+            ("_rowid", pa.uint64()),
+            ("_distance", pa.float32()),
+        ]
+    )
+    fts_schema = pa.schema(
+        [
+            ("text", pa.string()),
+            ("vector", pa.list_(pa.float32(), 4)),
+            ("_rowid", pa.uint64()),
+            ("_score", pa.float32()),
+        ]
+    )
+    empty_vector = pa.table(
+        {
+            "text": pa.array([], type=pa.string()),
+            "vector": pa.array([], type=pa.list_(pa.float32(), 4)),
+            "_rowid": pa.array([], type=pa.uint64()),
+            "_distance": pa.array([], type=pa.float32()),
+        },
+        schema=vector_schema,
+    )
+    empty_fts = pa.table(
+        {
+            "text": pa.array([], type=pa.string()),
+            "vector": pa.array([], type=pa.list_(pa.float32(), 4)),
+            "_rowid": pa.array([], type=pa.uint64()),
+            "_score": pa.array([], type=pa.float32()),
+        },
+        schema=fts_schema,
+    )
+
+    for reranker in [LinearCombinationReranker(), RRFReranker()]:
+        result = LanceHybridQueryBuilder._combine_hybrid_results(
+            fts_results=empty_fts,
+            vector_results=empty_vector,
+            norm="score",
+            fts_query="nonexistent query",
+            reranker=reranker,
+            limit=10,
+            with_row_ids=False,
+        )
+        assert len(result) == 0
+        assert "_relevance_score" in result.column_names
+        assert "_rowid" not in result.column_names
+
+    # Also test with with_row_ids=True
+    result = LanceHybridQueryBuilder._combine_hybrid_results(
+        fts_results=empty_fts,
+        vector_results=empty_vector,
+        norm="score",
+        fts_query="nonexistent query",
+        reranker=LinearCombinationReranker(),
+        limit=10,
+        with_row_ids=True,
+    )
+    assert len(result) == 0
+    assert "_relevance_score" in result.column_names
+    assert "_rowid" in result.column_names
+
+
 @pytest.mark.parametrize("use_tantivy", [True, False])
 def test_cross_encoder_reranker_return_all(tmp_path, use_tantivy):
     pytest.importorskip("sentence_transformers")

python/python/tests/test_table.py

@@ -810,7 +810,7 @@ def test_create_index_name_and_train_parameters(
     )
 
 
-def test_add_with_nans(mem_db: DBConnection):
+def test_create_with_nans(mem_db: DBConnection):
     # by default we raise an error on bad input vectors
     bad_data = [
         {"vector": [np.nan], "item": "bar", "price": 20.0},
@@ -854,6 +854,57 @@ def test_add_with_nans(mem_db: DBConnection):
     assert np.allclose(v, np.array([0.0, 0.0]))
 
 
+def test_add_with_nans(mem_db: DBConnection):
+    schema = pa.schema(
+        [
+            pa.field("vector", pa.list_(pa.float32(), 2), nullable=True),
+            pa.field("item", pa.string(), nullable=True),
+            pa.field("price", pa.float64(), nullable=False),
+        ],
+    )
+    table = mem_db.create_table("test", schema=schema)
+    # by default we raise an error on bad input vectors
+    bad_data = [
+        {"vector": [np.nan], "item": "bar", "price": 20.0},
+        {"vector": [5], "item": "bar", "price": 20.0},
+        {"vector": [np.nan, np.nan], "item": "bar", "price": 20.0},
+        {"vector": [np.nan, 5.0], "item": "bar", "price": 20.0},
+    ]
+    for row in bad_data:
+        with pytest.raises(ValueError):
+            table.add(
+                data=[row],
+            )
+
+    table.add(
+        [
+            {"vector": [3.1, 4.1], "item": "foo", "price": 10.0},
+            {"vector": [2.1, 4.1], "item": "foo", "price": 9.0},
+            {"vector": [np.nan], "item": "bar", "price": 20.0},
+            {"vector": [5], "item": "bar", "price": 20.0},
+            {"vector": [np.nan, np.nan], "item": "bar", "price": 20.0},
+        ],
+        on_bad_vectors="drop",
+    )
+    assert len(table) == 2
+    table.delete("true")
+
+    # We can fill bad input with some value
+    table.add(
+        data=[
+            {"vector": [3.1, 4.1], "item": "foo", "price": 10.0},
+            {"vector": [np.nan], "item": "bar", "price": 20.0},
+            {"vector": [np.nan, np.nan], "item": "bar", "price": 20.0},
+        ],
+        on_bad_vectors="fill",
+        fill_value=0.0,
+    )
+    assert len(table) == 3
+    arrow_tbl = table.search().where("item == 'bar'").to_arrow()
+    v = arrow_tbl["vector"].to_pylist()[0]
+    assert np.allclose(v, np.array([0.0, 0.0]))
+
+
 def test_restore(mem_db: DBConnection):
     table = mem_db.create_table(
         "my_table",

python/python/tests/test_util.py

@@ -292,18 +292,14 @@ class TestModel(lancedb.pydantic.LanceModel):
         lambda: pa.table({"a": [1], "b": [2]}),
         lambda: pa.table({"a": [1], "b": [2]}).to_reader(),
         lambda: iter(pa.table({"a": [1], "b": [2]}).to_batches()),
-        lambda: (
-            lance.write_dataset(
-                pa.table({"a": [1], "b": [2]}),
-                "memory://test",
-            )
-        ),
-        lambda: (
-            lance.write_dataset(
-                pa.table({"a": [1], "b": [2]}),
-                "memory://test",
-            ).scanner()
+        lambda: lance.write_dataset(
+            pa.table({"a": [1], "b": [2]}),
+            "memory://test",
         ),
+        lambda: lance.write_dataset(
+            pa.table({"a": [1], "b": [2]}),
+            "memory://test",
+        ).scanner(),
         lambda: pd.DataFrame({"a": [1], "b": [2]}),
         lambda: pl.DataFrame({"a": [1], "b": [2]}),
         lambda: pl.LazyFrame({"a": [1], "b": [2]}),

python/src/connection.rs

@@ -121,7 +121,8 @@ impl Connection {
 
         let mode = Self::parse_create_mode_str(mode)?;
 
-        let batches = ArrowArrayStreamReader::from_pyarrow_bound(&data)?;
+        let batches: Box<dyn arrow::array::RecordBatchReader + Send> =
+            Box::new(ArrowArrayStreamReader::from_pyarrow_bound(&data)?);
 
         let mut builder = inner.create_table(name, batches).mode(mode);
 

python/src/permutation.rs

@@ -23,10 +23,25 @@ use pyo3::{
 };
 use pyo3_async_runtimes::tokio::future_into_py;
 
+fn table_from_py<'a>(table: Bound<'a, PyAny>) -> PyResult<Bound<'a, Table>> {
+    if table.hasattr("_inner")? {
+        Ok(table.getattr("_inner")?.downcast_into::<Table>()?)
+    } else if table.hasattr("_table")? {
+        Ok(table
+            .getattr("_table")?
+            .getattr("_inner")?
+            .downcast_into::<Table>()?)
+    } else {
+        Err(PyRuntimeError::new_err(
+            "Provided table does not appear to be a Table or RemoteTable instance",
+        ))
+    }
+}
+
 /// Create a permutation builder for the given table
 #[pyo3::pyfunction]
 pub fn async_permutation_builder(table: Bound<'_, PyAny>) -> PyResult<PyAsyncPermutationBuilder> {
-    let table = table.getattr("_inner")?.downcast_into::<Table>()?;
+    let table = table_from_py(table)?;
     let inner_table = table.borrow().inner_ref()?.clone();
     let inner_builder = LancePermutationBuilder::new(inner_table);
 
@@ -250,10 +265,8 @@ impl PyPermutationReader {
         permutation_table: Option<Bound<'py, PyAny>>,
         split: u64,
     ) -> PyResult<Bound<'py, PyAny>> {
-        let base_table = base_table.getattr("_inner")?.downcast_into::<Table>()?;
-        let permutation_table = permutation_table
-            .map(|p| PyResult::Ok(p.getattr("_inner")?.downcast_into::<Table>()?))
-            .transpose()?;
+        let base_table = table_from_py(base_table)?;
+        let permutation_table = permutation_table.map(table_from_py).transpose()?;
 
         let base_table = base_table.borrow().inner_ref()?.base_table().clone();
         let permutation_table = permutation_table

python/src/table.rs

@@ -7,6 +7,7 @@ use crate::{
     error::PythonErrorExt,
     index::{extract_index_params, IndexConfig},
     query::{Query, TakeQuery},
+    table::scannable::PyScannable,
 };
 use arrow::{
     datatypes::{DataType, Schema},
@@ -25,6 +26,8 @@ use pyo3::{
 };
 use pyo3_async_runtimes::tokio::future_into_py;
 
+mod scannable;
+
 /// Statistics about a compaction operation.
 #[pyclass(get_all)]
 #[derive(Clone, Debug)]
@@ -293,11 +296,10 @@ impl Table {
 
     pub fn add<'a>(
         self_: PyRef<'a, Self>,
-        data: Bound<'_, PyAny>,
+        data: PyScannable,
         mode: String,
     ) -> PyResult<Bound<'a, PyAny>> {
-        let batches = ArrowArrayStreamReader::from_pyarrow_bound(&data)?;
-        let mut op = self_.inner_ref()?.add(batches);
+        let mut op = self_.inner_ref()?.add(data);
         if mode == "append" {
             op = op.mode(AddDataMode::Append);
         } else if mode == "overwrite" {

python/src/table/scannable.rs

@@ -0,0 +1,145 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+use std::sync::Arc;
+
+use arrow::{
+    datatypes::{Schema, SchemaRef},
+    ffi_stream::ArrowArrayStreamReader,
+    pyarrow::{FromPyArrow, PyArrowType},
+};
+use futures::StreamExt;
+use lancedb::{
+    arrow::{SendableRecordBatchStream, SimpleRecordBatchStream},
+    data::scannable::Scannable,
+    Error,
+};
+use pyo3::{types::PyAnyMethods, FromPyObject, Py, PyAny, Python};
+
+/// Adapter that implements Scannable for a Python reader factory callable.
+///
+/// This holds a Python callable that returns a RecordBatchReader when called.
+/// For rescannable sources, the callable can be invoked multiple times to
+/// get fresh readers.
+pub struct PyScannable {
+    /// Python callable that returns a RecordBatchReader
+    reader_factory: Py<PyAny>,
+    schema: SchemaRef,
+    num_rows: Option<usize>,
+    rescannable: bool,
+}
+
+impl std::fmt::Debug for PyScannable {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("PyScannable")
+            .field("schema", &self.schema)
+            .field("num_rows", &self.num_rows)
+            .field("rescannable", &self.rescannable)
+            .finish()
+    }
+}
+
+impl Scannable for PyScannable {
+    fn schema(&self) -> SchemaRef {
+        self.schema.clone()
+    }
+
+    fn scan_as_stream(&mut self) -> SendableRecordBatchStream {
+        let reader: Result<ArrowArrayStreamReader, Error> = {
+            Python::attach(|py| {
+                let result =
+                    self.reader_factory
+                        .call0(py)
+                        .map_err(|e| lancedb::Error::Runtime {
+                            message: format!("Python reader factory failed: {}", e),
+                        })?;
+                ArrowArrayStreamReader::from_pyarrow_bound(result.bind(py)).map_err(|e| {
+                    lancedb::Error::Runtime {
+                        message: format!("Failed to create Arrow reader from Python: {}", e),
+                    }
+                })
+            })
+        };
+
+        // Reader is blocking but stream is non-blocking, so we need to spawn a task to pull.
+        let (tx, rx) = tokio::sync::mpsc::channel(1);
+
+        let join_handle = tokio::task::spawn_blocking(move || {
+            let reader = match reader {
+                Ok(reader) => reader,
+                Err(e) => {
+                    let _ = tx.blocking_send(Err(e));
+                    return;
+                }
+            };
+            for batch in reader {
+                match batch {
+                    Ok(batch) => {
+                        if tx.blocking_send(Ok(batch)).is_err() {
+                            // Receiver dropped, stop processing
+                            break;
+                        }
+                    }
+                    Err(source) => {
+                        let _ = tx.blocking_send(Err(Error::Arrow { source }));
+                        break;
+                    }
+                }
+            }
+        });
+
+        let schema = self.schema.clone();
+        let stream = futures::stream::unfold(
+            (rx, Some(join_handle)),
+            |(mut rx, join_handle)| async move {
+                match rx.recv().await {
+                    Some(Ok(batch)) => Some((Ok(batch), (rx, join_handle))),
+                    Some(Err(e)) => Some((Err(e), (rx, join_handle))),
+                    None => {
+                        // Channel closed. Check if the task panicked — a panic
+                        // drops the sender without sending an error, so without
+                        // this check we'd silently return a truncated stream.
+                        if let Some(handle) = join_handle {
+                            if let Err(join_err) = handle.await {
+                                return Some((
+                                    Err(Error::Runtime {
+                                        message: format!("Reader task panicked: {}", join_err),
+                                    }),
+                                    (rx, None),
+                                ));
+                            }
+                        }
+                        None
+                    }
+                }
+            },
+        );
+        Box::pin(SimpleRecordBatchStream::new(stream.fuse(), schema))
+    }
+
+    fn num_rows(&self) -> Option<usize> {
+        self.num_rows
+    }
+
+    fn rescannable(&self) -> bool {
+        self.rescannable
+    }
+}
+
+impl<'py> FromPyObject<'py> for PyScannable {
+    fn extract_bound(ob: &pyo3::Bound<'py, PyAny>) -> pyo3::PyResult<Self> {
+        // Convert from Scannable dataclass.
+        let schema: PyArrowType<Schema> = ob.getattr("schema")?.extract()?;
+        let schema = Arc::new(schema.0);
+        let num_rows: Option<usize> = ob.getattr("num_rows")?.extract()?;
+        let rescannable: bool = ob.getattr("rescannable")?.extract()?;
+        let reader_factory: Py<PyAny> = ob.getattr("reader")?.unbind();
+
+        Ok(Self {
+            schema,
+            reader_factory,
+            num_rows,
+            rescannable,
+        })
+    }
+}

rust-toolchain.toml

@@ -1,2 +1,2 @@
 [toolchain]
-channel = "1.90.0"
+channel = "1.91.0"

rust/lancedb/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "lancedb"
-version = "0.26.2"
+version = "0.27.0-beta.1"
 edition.workspace = true
 description = "LanceDB: A serverless, low-latency vector database for AI applications"
 license.workspace = true
@@ -25,7 +25,9 @@ datafusion-catalog.workspace = true
 datafusion-common.workspace = true
 datafusion-execution.workspace = true
 datafusion-expr.workspace = true
+datafusion-functions = "51.0"
 datafusion-physical-expr.workspace = true
+datafusion-sql = "51.0"
 datafusion-physical-plan.workspace = true
 datafusion.workspace = true
 object_store = { workspace = true }

rust/lancedb/examples/bedrock.rs

@@ -3,13 +3,12 @@
 
 use std::{iter::once, sync::Arc};
 
-use arrow_array::{Float64Array, Int32Array, RecordBatch, RecordBatchIterator, StringArray};
+use arrow_array::{Float64Array, Int32Array, RecordBatch, StringArray};
 use arrow_schema::{DataType, Field, Schema};
 use aws_config::Region;
 use aws_sdk_bedrockruntime::Client;
 use futures::StreamExt;
 use lancedb::{
-    arrow::IntoArrow,
     connect,
     embeddings::{bedrock::BedrockEmbeddingFunction, EmbeddingDefinition, EmbeddingFunction},
     query::{ExecutableQuery, QueryBase},
@@ -67,7 +66,7 @@ async fn main() -> Result<()> {
     Ok(())
 }
 
-fn make_data() -> impl IntoArrow {
+fn make_data() -> RecordBatch {
     let schema = Schema::new(vec![
         Field::new("id", DataType::Int32, true),
         Field::new("text", DataType::Utf8, false),
@@ -83,10 +82,9 @@ fn make_data() -> impl IntoArrow {
     ]);
     let price = Float64Array::from(vec![10.0, 50.0, 100.0, 30.0]);
     let schema = Arc::new(schema);
-    let rb = RecordBatch::try_new(
+    RecordBatch::try_new(
         schema.clone(),
         vec![Arc::new(id), Arc::new(text), Arc::new(price)],
     )
-    .unwrap();
-    Box::new(RecordBatchIterator::new(vec![Ok(rb)], schema))
+    .unwrap()
 }

rust/lancedb/examples/full_text_search.rs

@@ -3,12 +3,13 @@
 
 use std::sync::Arc;
 
-use arrow_array::{Int32Array, RecordBatch, RecordBatchIterator, RecordBatchReader, StringArray};
+use arrow_array::{Int32Array, RecordBatch, RecordBatchIterator, StringArray};
 use arrow_schema::{DataType, Field, Schema};
 
 use futures::TryStreamExt;
 use lance_index::scalar::FullTextSearchQuery;
 use lancedb::connection::Connection;
+
 use lancedb::index::scalar::FtsIndexBuilder;
 use lancedb::index::Index;
 use lancedb::query::{ExecutableQuery, QueryBase};
@@ -29,7 +30,7 @@ async fn main() -> Result<()> {
     Ok(())
 }
 
-fn create_some_records() -> Result<Box<dyn RecordBatchReader + Send>> {
+fn create_some_records() -> Result<Box<dyn arrow_array::RecordBatchReader + Send>> {
     const TOTAL: usize = 1000;
 
     let schema = Arc::new(Schema::new(vec![
@@ -66,7 +67,7 @@ fn create_some_records() -> Result<Box<dyn RecordBatchReader + Send>> {
 }
 
 async fn create_table(db: &Connection) -> Result<Table> {
-    let initial_data: Box<dyn RecordBatchReader + Send> = create_some_records()?;
+    let initial_data = create_some_records()?;
     let tbl = db.create_table("my_table", initial_data).execute().await?;
     Ok(tbl)
 }

rust/lancedb/examples/hybrid_search.rs

@@ -1,14 +1,13 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
-use arrow_array::{RecordBatch, RecordBatchIterator, StringArray};
+use arrow_array::{RecordBatch, StringArray};
 use arrow_schema::{DataType, Field, Schema};
 use futures::TryStreamExt;
 use lance_index::scalar::FullTextSearchQuery;
 use lancedb::index::scalar::FtsIndexBuilder;
 use lancedb::index::Index;
 use lancedb::{
-    arrow::IntoArrow,
     connect,
     embeddings::{
         sentence_transformers::SentenceTransformersEmbeddings, EmbeddingDefinition,
@@ -70,7 +69,7 @@ async fn main() -> Result<()> {
     Ok(())
 }
 
-fn make_data() -> impl IntoArrow {
+fn make_data() -> RecordBatch {
     let schema = Schema::new(vec![Field::new("facts", DataType::Utf8, false)]);
 
     let facts = StringArray::from_iter_values(vec![
@@ -101,8 +100,7 @@ fn make_data() -> impl IntoArrow {
         "The first chatbot was ELIZA, created in the 1960s.",
     ]);
     let schema = Arc::new(schema);
-    let rb = RecordBatch::try_new(schema.clone(), vec![Arc::new(facts)]).unwrap();
-    Box::new(RecordBatchIterator::new(vec![Ok(rb)], schema))
+    RecordBatch::try_new(schema.clone(), vec![Arc::new(facts)]).unwrap()
 }
 
 async fn create_index(table: &Table) -> Result<()> {

rust/lancedb/examples/ivf_pq.rs

@@ -8,13 +8,12 @@
 use std::sync::Arc;
 
 use arrow_array::types::Float32Type;
-use arrow_array::{
-    FixedSizeListArray, Int32Array, RecordBatch, RecordBatchIterator, RecordBatchReader,
-};
+use arrow_array::{FixedSizeListArray, Int32Array, RecordBatch, RecordBatchIterator};
 use arrow_schema::{DataType, Field, Schema};
 
 use futures::TryStreamExt;
 use lancedb::connection::Connection;
+
 use lancedb::index::vector::IvfPqIndexBuilder;
 use lancedb::index::Index;
 use lancedb::query::{ExecutableQuery, QueryBase};
@@ -34,7 +33,7 @@ async fn main() -> Result<()> {
     Ok(())
 }
 
-fn create_some_records() -> Result<Box<dyn RecordBatchReader + Send>> {
+fn create_some_records() -> Result<Box<dyn arrow_array::RecordBatchReader + Send>> {
     const TOTAL: usize = 1000;
     const DIM: usize = 128;
 
@@ -73,9 +72,9 @@ fn create_some_records() -> Result<Box<dyn RecordBatchReader + Send>> {
 }
 
 async fn create_table(db: &Connection) -> Result<Table> {
-    let initial_data: Box<dyn RecordBatchReader + Send> = create_some_records()?;
+    let initial_data = create_some_records()?;
     let tbl = db
-        .create_table("my_table", Box::new(initial_data))
+        .create_table("my_table", initial_data)
         .execute()
         .await
         .unwrap();

rust/lancedb/examples/openai.rs

@@ -5,11 +5,9 @@
 
 use std::{iter::once, sync::Arc};
 
-use arrow_array::{Float64Array, Int32Array, RecordBatch, RecordBatchIterator, StringArray};
-use arrow_schema::{DataType, Field, Schema};
+use arrow_array::{RecordBatch, StringArray};
 use futures::StreamExt;
 use lancedb::{
-    arrow::IntoArrow,
     connect,
     embeddings::{openai::OpenAIEmbeddingFunction, EmbeddingDefinition, EmbeddingFunction},
     query::{ExecutableQuery, QueryBase},
@@ -64,26 +62,20 @@ async fn main() -> Result<()> {
 }
 // --8<-- [end:openai_embeddings]
 
-fn make_data() -> impl IntoArrow {
-    let schema = Schema::new(vec![
-        Field::new("id", DataType::Int32, true),
-        Field::new("text", DataType::Utf8, false),
-        Field::new("price", DataType::Float64, false),
-    ]);
-
-    let id = Int32Array::from(vec![1, 2, 3, 4]);
-    let text = StringArray::from_iter_values(vec![
-        "Black T-Shirt",
-        "Leather Jacket",
-        "Winter Parka",
-        "Hooded Sweatshirt",
-    ]);
-    let price = Float64Array::from(vec![10.0, 50.0, 100.0, 30.0]);
-    let schema = Arc::new(schema);
-    let rb = RecordBatch::try_new(
-        schema.clone(),
-        vec![Arc::new(id), Arc::new(text), Arc::new(price)],
+fn make_data() -> RecordBatch {
+    arrow_array::record_batch!(
+        ("id", Int32, [1, 2, 3, 4]),
+        (
+            "text",
+            Utf8,
+            [
+                "Black T-Shirt",
+                "Leather Jacket",
+                "Winter Parka",
+                "Hooded Sweatshirt"
+            ]
+        ),
+        ("price", Float64, [10.0, 50.0, 100.0, 30.0])
     )
-    .unwrap();
-    Box::new(RecordBatchIterator::new(vec![Ok(rb)], schema))
+    .unwrap()
 }

rust/lancedb/examples/sentence_transformers.rs

@@ -3,11 +3,10 @@
 
 use std::{iter::once, sync::Arc};
 
-use arrow_array::{RecordBatch, RecordBatchIterator, StringArray};
+use arrow_array::{RecordBatch, StringArray};
 use arrow_schema::{DataType, Field, Schema};
 use futures::StreamExt;
 use lancedb::{
-    arrow::IntoArrow,
     connect,
     embeddings::{
         sentence_transformers::SentenceTransformersEmbeddings, EmbeddingDefinition,
@@ -59,7 +58,7 @@ async fn main() -> Result<()> {
     Ok(())
 }
 
-fn make_data() -> impl IntoArrow {
+fn make_data() -> RecordBatch {
     let schema = Schema::new(vec![Field::new("facts", DataType::Utf8, false)]);
 
     let facts = StringArray::from_iter_values(vec![
@@ -90,6 +89,5 @@ fn make_data() -> impl IntoArrow {
         "The first chatbot was ELIZA, created in the 1960s.",
     ]);
     let schema = Arc::new(schema);
-    let rb = RecordBatch::try_new(schema.clone(), vec![Arc::new(facts)]).unwrap();
-    Box::new(RecordBatchIterator::new(vec![Ok(rb)], schema))
+    RecordBatch::try_new(schema.clone(), vec![Arc::new(facts)]).unwrap()
 }

rust/lancedb/examples/simple.rs

@@ -8,11 +8,9 @@
 use std::sync::Arc;
 
 use arrow_array::types::Float32Type;
-use arrow_array::{FixedSizeListArray, Int32Array, RecordBatch, RecordBatchIterator};
+use arrow_array::{FixedSizeListArray, Int32Array, RecordBatch};
 use arrow_schema::{DataType, Field, Schema};
 use futures::TryStreamExt;
-
-use lancedb::arrow::IntoArrow;
 use lancedb::connection::Connection;
 use lancedb::index::Index;
 use lancedb::query::{ExecutableQuery, QueryBase};
@@ -59,7 +57,7 @@ async fn open_with_existing_tbl() -> Result<()> {
     Ok(())
 }
 
-fn create_some_records() -> Result<impl IntoArrow> {
+fn create_some_records() -> Result<RecordBatch> {
     const TOTAL: usize = 1000;
     const DIM: usize = 128;
 
@@ -76,25 +74,18 @@ fn create_some_records() -> Result<impl IntoArrow> {
     ]));
 
     // Create a RecordBatch stream.
-    let batches = RecordBatchIterator::new(
-        vec![RecordBatch::try_new(
-            schema.clone(),
-            vec![
-                Arc::new(Int32Array::from_iter_values(0..TOTAL as i32)),
-                Arc::new(
-                    FixedSizeListArray::from_iter_primitive::<Float32Type, _, _>(
-                        (0..TOTAL).map(|_| Some(vec![Some(1.0); DIM])),
-                        DIM as i32,
-                    ),
-                ),
-            ],
-        )
-        .unwrap()]
-        .into_iter()
-        .map(Ok),
+    Ok(RecordBatch::try_new(
         schema.clone(),
-    );
-    Ok(Box::new(batches))
+        vec![
+            Arc::new(Int32Array::from_iter_values(0..TOTAL as i32)),
+            Arc::new(
+                FixedSizeListArray::from_iter_primitive::<Float32Type, _, _>(
+                    (0..TOTAL).map(|_| Some(vec![Some(1.0); DIM])),
+                    DIM as i32,
+                ),
+            ),
+        ],
+    )?)
 }
 
 async fn create_table(db: &Connection) -> Result<LanceDbTable> {

rust/lancedb/src/arrow.rs

@@ -155,9 +155,7 @@ impl IntoArrowStream for SendableRecordBatchStream {
 impl IntoArrowStream for datafusion_physical_plan::SendableRecordBatchStream {
     fn into_arrow(self) -> Result<SendableRecordBatchStream> {
         let schema = self.schema();
-        let stream = self.map_err(|df_err| Error::Runtime {
-            message: df_err.to_string(),
-        });
+        let stream = self.map_err(|df_err| df_err.into());
         Ok(Box::pin(SimpleRecordBatchStream::new(stream, schema)))
     }
 }

rust/lancedb/src/connection.rs

@@ -6,8 +6,8 @@
 use std::collections::HashMap;
 use std::sync::Arc;
 
-use arrow_array::RecordBatchReader;
-use arrow_schema::{Field, SchemaRef};
+use arrow_array::RecordBatch;
+use arrow_schema::SchemaRef;
 use lance::dataset::ReadParams;
 use lance_namespace::models::{
     CreateNamespaceRequest, CreateNamespaceResponse, DescribeNamespaceRequest,
@@ -17,24 +17,20 @@ use lance_namespace::models::{
 #[cfg(feature = "aws")]
 use object_store::aws::AwsCredential;
 
-use crate::arrow::{IntoArrow, IntoArrowStream, SendableRecordBatchStream};
-use crate::database::listing::{
-    ListingDatabase, OPT_NEW_TABLE_STORAGE_VERSION, OPT_NEW_TABLE_V2_MANIFEST_PATHS,
-};
+use crate::connection::create_table::CreateTableBuilder;
+use crate::data::scannable::Scannable;
+use crate::database::listing::ListingDatabase;
 use crate::database::{
-    CloneTableRequest, CreateTableData, CreateTableMode, CreateTableRequest, Database,
-    DatabaseOptions, OpenTableRequest, ReadConsistency, TableNamesRequest,
-};
-use crate::embeddings::{
-    EmbeddingDefinition, EmbeddingFunction, EmbeddingRegistry, MemoryRegistry, WithEmbeddings,
+    CloneTableRequest, Database, DatabaseOptions, OpenTableRequest, ReadConsistency,
+    TableNamesRequest,
 };
+use crate::embeddings::{EmbeddingRegistry, MemoryRegistry};
 use crate::error::{Error, Result};
 #[cfg(feature = "remote")]
 use crate::remote::{
     client::ClientConfig,
     db::{OPT_REMOTE_API_KEY, OPT_REMOTE_HOST_OVERRIDE, OPT_REMOTE_REGION},
 };
-use crate::table::{TableDefinition, WriteOptions};
 use crate::Table;
 use lance::io::ObjectStoreParams;
 pub use lance_encoding::version::LanceFileVersion;
@@ -42,6 +38,8 @@ pub use lance_encoding::version::LanceFileVersion;
 use lance_io::object_store::StorageOptions;
 use lance_io::object_store::{StorageOptionsAccessor, StorageOptionsProvider};
 
+mod create_table;
+
 fn merge_storage_options(
     store_params: &mut ObjectStoreParams,
     pairs: impl IntoIterator<Item = (String, String)>,
@@ -116,337 +114,6 @@ impl TableNamesBuilder {
     }
 }
 
-pub struct NoData {}
-
-impl IntoArrow for NoData {
-    fn into_arrow(self) -> Result<Box<dyn arrow_array::RecordBatchReader + Send>> {
-        unreachable!("NoData should never be converted to Arrow")
-    }
-}
-
-// Stores the value given from the initial CreateTableBuilder::new call
-// and defers errors until `execute` is called
-enum CreateTableBuilderInitialData {
-    None,
-    Iterator(Result<Box<dyn RecordBatchReader + Send>>),
-    Stream(Result<SendableRecordBatchStream>),
-}
-
-/// A builder for configuring a [`Connection::create_table`] operation
-pub struct CreateTableBuilder<const HAS_DATA: bool> {
-    parent: Arc<dyn Database>,
-    embeddings: Vec<(EmbeddingDefinition, Arc<dyn EmbeddingFunction>)>,
-    embedding_registry: Arc<dyn EmbeddingRegistry>,
-    request: CreateTableRequest,
-    // This is a bit clumsy but we defer errors until `execute` is called
-    // to maintain backwards compatibility
-    data: CreateTableBuilderInitialData,
-}
-
-// Builder methods that only apply when we have initial data
-impl CreateTableBuilder<true> {
-    fn new<T: IntoArrow>(
-        parent: Arc<dyn Database>,
-        name: String,
-        data: T,
-        embedding_registry: Arc<dyn EmbeddingRegistry>,
-    ) -> Self {
-        let dummy_schema = Arc::new(arrow_schema::Schema::new(Vec::<Field>::default()));
-        Self {
-            parent,
-            request: CreateTableRequest::new(
-                name,
-                CreateTableData::Empty(TableDefinition::new_from_schema(dummy_schema)),
-            ),
-            embeddings: Vec::new(),
-            embedding_registry,
-            data: CreateTableBuilderInitialData::Iterator(data.into_arrow()),
-        }
-    }
-
-    fn new_streaming<T: IntoArrowStream>(
-        parent: Arc<dyn Database>,
-        name: String,
-        data: T,
-        embedding_registry: Arc<dyn EmbeddingRegistry>,
-    ) -> Self {
-        let dummy_schema = Arc::new(arrow_schema::Schema::new(Vec::<Field>::default()));
-        Self {
-            parent,
-            request: CreateTableRequest::new(
-                name,
-                CreateTableData::Empty(TableDefinition::new_from_schema(dummy_schema)),
-            ),
-            embeddings: Vec::new(),
-            embedding_registry,
-            data: CreateTableBuilderInitialData::Stream(data.into_arrow()),
-        }
-    }
-
-    /// Execute the create table operation
-    pub async fn execute(self) -> Result<Table> {
-        let embedding_registry = self.embedding_registry.clone();
-        let parent = self.parent.clone();
-        let request = self.into_request()?;
-        Ok(Table::new_with_embedding_registry(
-            parent.create_table(request).await?,
-            parent,
-            embedding_registry,
-        ))
-    }
-
-    fn into_request(self) -> Result<CreateTableRequest> {
-        if self.embeddings.is_empty() {
-            match self.data {
-                CreateTableBuilderInitialData::Iterator(maybe_iter) => {
-                    let data = maybe_iter?;
-                    Ok(CreateTableRequest {
-                        data: CreateTableData::Data(data),
-                        ..self.request
-                    })
-                }
-                CreateTableBuilderInitialData::None => {
-                    unreachable!("No data provided for CreateTableBuilder<true>")
-                }
-                CreateTableBuilderInitialData::Stream(maybe_stream) => {
-                    let data = maybe_stream?;
-                    Ok(CreateTableRequest {
-                        data: CreateTableData::StreamingData(data),
-                        ..self.request
-                    })
-                }
-            }
-        } else {
-            let CreateTableBuilderInitialData::Iterator(maybe_iter) = self.data else {
-                return Err(Error::NotSupported { message: "Creating a table with embeddings is currently not support when the input is streaming".to_string() });
-            };
-            let data = maybe_iter?;
-            let data = Box::new(WithEmbeddings::new(data, self.embeddings));
-            Ok(CreateTableRequest {
-                data: CreateTableData::Data(data),
-                ..self.request
-            })
-        }
-    }
-}
-
-// Builder methods that only apply when we do not have initial data
-impl CreateTableBuilder<false> {
-    fn new(
-        parent: Arc<dyn Database>,
-        name: String,
-        schema: SchemaRef,
-        embedding_registry: Arc<dyn EmbeddingRegistry>,
-    ) -> Self {
-        let table_definition = TableDefinition::new_from_schema(schema);
-        Self {
-            parent,
-            request: CreateTableRequest::new(name, CreateTableData::Empty(table_definition)),
-            data: CreateTableBuilderInitialData::None,
-            embeddings: Vec::default(),
-            embedding_registry,
-        }
-    }
-
-    /// Execute the create table operation
-    pub async fn execute(self) -> Result<Table> {
-        let parent = self.parent.clone();
-        let embedding_registry = self.embedding_registry.clone();
-        let request = self.into_request()?;
-        Ok(Table::new_with_embedding_registry(
-            parent.create_table(request).await?,
-            parent,
-            embedding_registry,
-        ))
-    }
-
-    fn into_request(self) -> Result<CreateTableRequest> {
-        if self.embeddings.is_empty() {
-            return Ok(self.request);
-        }
-
-        let CreateTableData::Empty(table_def) = self.request.data else {
-            unreachable!("CreateTableBuilder<false> should always have Empty data")
-        };
-
-        let schema = table_def.schema.clone();
-        let empty_batch = arrow_array::RecordBatch::new_empty(schema.clone());
-
-        let reader = Box::new(std::iter::once(Ok(empty_batch)).collect::<Vec<_>>());
-        let reader = arrow_array::RecordBatchIterator::new(reader.into_iter(), schema);
-        let with_embeddings = WithEmbeddings::new(reader, self.embeddings);
-        let table_definition = with_embeddings.table_definition()?;
-
-        Ok(CreateTableRequest {
-            data: CreateTableData::Empty(table_definition),
-            ..self.request
-        })
-    }
-}
-
-impl<const HAS_DATA: bool> CreateTableBuilder<HAS_DATA> {
-    /// Set the mode for creating the table
-    ///
-    /// This controls what happens if a table with the given name already exists
-    pub fn mode(mut self, mode: CreateTableMode) -> Self {
-        self.request.mode = mode;
-        self
-    }
-
-    /// Apply the given write options when writing the initial data
-    pub fn write_options(mut self, write_options: WriteOptions) -> Self {
-        self.request.write_options = write_options;
-        self
-    }
-
-    /// Set an option for the storage layer.
-    ///
-    /// Options already set on the connection will be inherited by the table,
-    /// but can be overridden here.
-    ///
-    /// See available options at <https://lancedb.com/docs/storage/>
-    pub fn storage_option(mut self, key: impl Into<String>, value: impl Into<String>) -> Self {
-        let store_params = self
-            .request
-            .write_options
-            .lance_write_params
-            .get_or_insert(Default::default())
-            .store_params
-            .get_or_insert(Default::default());
-        merge_storage_options(store_params, [(key.into(), value.into())]);
-        self
-    }
-
-    /// Set multiple options for the storage layer.
-    ///
-    /// Options already set on the connection will be inherited by the table,
-    /// but can be overridden here.
-    ///
-    /// See available options at <https://lancedb.com/docs/storage/>
-    pub fn storage_options(
-        mut self,
-        pairs: impl IntoIterator<Item = (impl Into<String>, impl Into<String>)>,
-    ) -> Self {
-        let store_params = self
-            .request
-            .write_options
-            .lance_write_params
-            .get_or_insert(Default::default())
-            .store_params
-            .get_or_insert(Default::default());
-        let updates = pairs
-            .into_iter()
-            .map(|(key, value)| (key.into(), value.into()));
-        merge_storage_options(store_params, updates);
-        self
-    }
-
-    /// Add an embedding definition to the table.
-    ///
-    /// The `embedding_name` must match the name of an embedding function that
-    /// was previously registered with the connection's [`EmbeddingRegistry`].
-    pub fn add_embedding(mut self, definition: EmbeddingDefinition) -> Result<Self> {
-        // Early verification of the embedding name
-        let embedding_func = self
-            .embedding_registry
-            .get(&definition.embedding_name)
-            .ok_or_else(|| Error::EmbeddingFunctionNotFound {
-                name: definition.embedding_name.clone(),
-                reason: "No embedding function found in the connection's embedding_registry"
-                    .to_string(),
-            })?;
-
-        self.embeddings.push((definition, embedding_func));
-        Ok(self)
-    }
-
-    /// Set whether to use V2 manifest paths for the table. (default: false)
-    ///
-    /// These paths provide more efficient opening of tables with many
-    /// versions on object stores.
-    ///
-    /// <div class="warning">Turning this on will make the dataset unreadable
-    /// for older versions of LanceDB (prior to 0.10.0).</div>
-    ///
-    /// To migrate an existing dataset, instead use the
-    /// [[NativeTable::migrate_manifest_paths_v2]].
-    ///
-    /// This has no effect in LanceDB Cloud.
-    #[deprecated(since = "0.15.1", note = "Use `database_options` instead")]
-    pub fn enable_v2_manifest_paths(mut self, use_v2_manifest_paths: bool) -> Self {
-        let store_params = self
-            .request
-            .write_options
-            .lance_write_params
-            .get_or_insert_with(Default::default)
-            .store_params
-            .get_or_insert_with(Default::default);
-        let value = if use_v2_manifest_paths {
-            "true".to_string()
-        } else {
-            "false".to_string()
-        };
-        merge_storage_options(
-            store_params,
-            [(OPT_NEW_TABLE_V2_MANIFEST_PATHS.to_string(), value)],
-        );
-        self
-    }
-
-    /// Set the data storage version.
-    ///
-    /// The default is `LanceFileVersion::Stable`.
-    #[deprecated(since = "0.15.1", note = "Use `database_options` instead")]
-    pub fn data_storage_version(mut self, data_storage_version: LanceFileVersion) -> Self {
-        let store_params = self
-            .request
-            .write_options
-            .lance_write_params
-            .get_or_insert_with(Default::default)
-            .store_params
-            .get_or_insert_with(Default::default);
-        merge_storage_options(
-            store_params,
-            [(
-                OPT_NEW_TABLE_STORAGE_VERSION.to_string(),
-                data_storage_version.to_string(),
-            )],
-        );
-        self
-    }
-
-    /// Set the namespace for the table
-    pub fn namespace(mut self, namespace: Vec<String>) -> Self {
-        self.request.namespace = namespace;
-        self
-    }
-
-    /// Set a custom location for the table.
-    ///
-    /// If not set, the database will derive a location from its URI and the table name.
-    /// This is useful when integrating with namespace systems that manage table locations.
-    pub fn location(mut self, location: impl Into<String>) -> Self {
-        self.request.location = Some(location.into());
-        self
-    }
-
-    /// Set a storage options provider for automatic credential refresh.
-    ///
-    /// This allows tables to automatically refresh cloud storage credentials
-    /// when they expire, enabling long-running operations on remote storage.
-    pub fn storage_options_provider(mut self, provider: Arc<dyn StorageOptionsProvider>) -> Self {
-        let store_params = self
-            .request
-            .write_options
-            .lance_write_params
-            .get_or_insert(Default::default())
-            .store_params
-            .get_or_insert(Default::default());
-        set_storage_options_provider(store_params, provider);
-        self
-    }
-}
-
 #[derive(Clone, Debug)]
 pub struct OpenTableBuilder {
     parent: Arc<dyn Database>,
@@ -684,35 +351,17 @@ impl Connection {
     ///
     /// * `name` - The name of the table
     /// * `initial_data` - The initial data to write to the table
-    pub fn create_table<T: IntoArrow>(
+    pub fn create_table<T: Scannable + 'static>(
         &self,
         name: impl Into<String>,
         initial_data: T,
-    ) -> CreateTableBuilder<true> {
-        CreateTableBuilder::<true>::new(
+    ) -> CreateTableBuilder {
+        let initial_data = Box::new(initial_data);
+        CreateTableBuilder::new(
             self.internal.clone(),
+            self.embedding_registry.clone(),
             name.into(),
             initial_data,
-            self.embedding_registry.clone(),
-        )
-    }
-
-    /// Create a new table from a stream of data
-    ///
-    /// # Parameters
-    ///
-    /// * `name` - The name of the table
-    /// * `initial_data` - The initial data to write to the table
-    pub fn create_table_streaming<T: IntoArrowStream>(
-        &self,
-        name: impl Into<String>,
-        initial_data: T,
-    ) -> CreateTableBuilder<true> {
-        CreateTableBuilder::<true>::new_streaming(
-            self.internal.clone(),
-            name.into(),
-            initial_data,
-            self.embedding_registry.clone(),
         )
     }
 
@@ -726,13 +375,9 @@ impl Connection {
         &self,
         name: impl Into<String>,
         schema: SchemaRef,
-    ) -> CreateTableBuilder<false> {
-        CreateTableBuilder::<false>::new(
-            self.internal.clone(),
-            name.into(),
-            schema,
-            self.embedding_registry.clone(),
-        )
+    ) -> CreateTableBuilder {
+        let empty_batch = RecordBatch::new_empty(schema);
+        self.create_table(name, empty_batch)
     }
 
     /// Open an existing table in the database
@@ -1349,20 +994,11 @@ mod test_utils {
 
 #[cfg(test)]
 mod tests {
-    use crate::database::listing::{ListingDatabaseOptions, NewTableConfig};
-    use crate::query::QueryBase;
-    use crate::query::{ExecutableQuery, QueryExecutionOptions};
-    use crate::test_utils::connection::new_test_connection;
-    use arrow::compute::concat_batches;
-    use arrow_array::RecordBatchReader;
     use arrow_schema::{DataType, Field, Schema};
-    use datafusion_physical_plan::stream::RecordBatchStreamAdapter;
-    use futures::{stream, TryStreamExt};
-    use lance_core::error::{ArrowResult, DataFusionResult};
     use lance_testing::datagen::{BatchGenerator, IncrementingInt32};
     use tempfile::tempdir;
 
-    use crate::arrow::SimpleRecordBatchStream;
+    use crate::test_utils::connection::new_test_connection;
 
     use super::*;
 
@@ -1478,139 +1114,6 @@ mod tests {
         assert_eq!(tables, vec!["table1".to_owned()]);
     }
 
-    fn make_data() -> Box<dyn RecordBatchReader + Send + 'static> {
-        let id = Box::new(IncrementingInt32::new().named("id".to_string()));
-        Box::new(BatchGenerator::new().col(id).batches(10, 2000))
-    }
-
-    #[tokio::test]
-    async fn test_create_table_v2() {
-        let tmp_dir = tempdir().unwrap();
-        let uri = tmp_dir.path().to_str().unwrap();
-        let db = connect(uri)
-            .database_options(&ListingDatabaseOptions {
-                new_table_config: NewTableConfig {
-                    data_storage_version: Some(LanceFileVersion::Legacy),
-                    ..Default::default()
-                },
-                ..Default::default()
-            })
-            .execute()
-            .await
-            .unwrap();
-
-        let tbl = db
-            .create_table("v1_test", make_data())
-            .execute()
-            .await
-            .unwrap();
-
-        // In v1 the row group size will trump max_batch_length
-        let batches = tbl
-            .query()
-            .limit(20000)
-            .execute_with_options(QueryExecutionOptions {
-                max_batch_length: 50000,
-                ..Default::default()
-            })
-            .await
-            .unwrap()
-            .try_collect::<Vec<_>>()
-            .await
-            .unwrap();
-        assert_eq!(batches.len(), 20);
-
-        let db = connect(uri)
-            .database_options(&ListingDatabaseOptions {
-                new_table_config: NewTableConfig {
-                    data_storage_version: Some(LanceFileVersion::Stable),
-                    ..Default::default()
-                },
-                ..Default::default()
-            })
-            .execute()
-            .await
-            .unwrap();
-
-        let tbl = db
-            .create_table("v2_test", make_data())
-            .execute()
-            .await
-            .unwrap();
-
-        // In v2 the page size is much bigger than 50k so we should get a single batch
-        let batches = tbl
-            .query()
-            .execute_with_options(QueryExecutionOptions {
-                max_batch_length: 50000,
-                ..Default::default()
-            })
-            .await
-            .unwrap()
-            .try_collect::<Vec<_>>()
-            .await
-            .unwrap();
-
-        assert_eq!(batches.len(), 1);
-    }
-
-    #[tokio::test]
-    async fn test_create_table_streaming() {
-        let tmp_dir = tempdir().unwrap();
-
-        let uri = tmp_dir.path().to_str().unwrap();
-        let db = connect(uri).execute().await.unwrap();
-
-        let batches = make_data().collect::<ArrowResult<Vec<_>>>().unwrap();
-
-        let schema = batches.first().unwrap().schema();
-        let one_batch = concat_batches(&schema, batches.iter()).unwrap();
-
-        let ldb_stream = stream::iter(batches.clone().into_iter().map(Result::Ok));
-        let ldb_stream: SendableRecordBatchStream =
-            Box::pin(SimpleRecordBatchStream::new(ldb_stream, schema.clone()));
-
-        let tbl1 = db
-            .create_table_streaming("one", ldb_stream)
-            .execute()
-            .await
-            .unwrap();
-
-        let df_stream = stream::iter(batches.into_iter().map(DataFusionResult::Ok));
-        let df_stream: datafusion_physical_plan::SendableRecordBatchStream =
-            Box::pin(RecordBatchStreamAdapter::new(schema.clone(), df_stream));
-
-        let tbl2 = db
-            .create_table_streaming("two", df_stream)
-            .execute()
-            .await
-            .unwrap();
-
-        let tbl1_data = tbl1
-            .query()
-            .execute()
-            .await
-            .unwrap()
-            .try_collect::<Vec<_>>()
-            .await
-            .unwrap();
-
-        let tbl1_data = concat_batches(&schema, tbl1_data.iter()).unwrap();
-        assert_eq!(tbl1_data, one_batch);
-
-        let tbl2_data = tbl2
-            .query()
-            .execute()
-            .await
-            .unwrap()
-            .try_collect::<Vec<_>>()
-            .await
-            .unwrap();
-
-        let tbl2_data = concat_batches(&schema, tbl2_data.iter()).unwrap();
-        assert_eq!(tbl2_data, one_batch);
-    }
-
     #[tokio::test]
     async fn drop_table() {
         let tc = new_test_connection().await.unwrap();
@@ -1640,41 +1143,6 @@ mod tests {
         assert_eq!(tables.len(), 0);
     }
 
-    #[tokio::test]
-    async fn test_create_table_already_exists() {
-        let tmp_dir = tempdir().unwrap();
-        let uri = tmp_dir.path().to_str().unwrap();
-        let db = connect(uri).execute().await.unwrap();
-        let schema = Arc::new(Schema::new(vec![Field::new("x", DataType::Int32, false)]));
-        db.create_empty_table("test", schema.clone())
-            .execute()
-            .await
-            .unwrap();
-        // TODO: None of the open table options are "inspectable" right now but once one is we
-        // should assert we are passing these options in correctly
-        db.create_empty_table("test", schema)
-            .mode(CreateTableMode::exist_ok(|mut req| {
-                req.index_cache_size = Some(16);
-                req
-            }))
-            .execute()
-            .await
-            .unwrap();
-        let other_schema = Arc::new(Schema::new(vec![Field::new("y", DataType::Int32, false)]));
-        assert!(db
-            .create_empty_table("test", other_schema.clone())
-            .execute()
-            .await
-            .is_err());
-        let overwritten = db
-            .create_empty_table("test", other_schema.clone())
-            .mode(CreateTableMode::Overwrite)
-            .execute()
-            .await
-            .unwrap();
-        assert_eq!(other_schema, overwritten.schema().await.unwrap());
-    }
-
     #[tokio::test]
     async fn test_clone_table() {
         let tmp_dir = tempdir().unwrap();
@@ -1685,7 +1153,8 @@ mod tests {
         let mut batch_gen = BatchGenerator::new()
             .col(Box::new(IncrementingInt32::new().named("id")))
             .col(Box::new(IncrementingInt32::new().named("value")));
-        let reader = batch_gen.batches(5, 100);
+        let reader: Box<dyn arrow_array::RecordBatchReader + Send> =
+            Box::new(batch_gen.batches(5, 100));
 
         let source_table = db
             .create_table("source_table", reader)
@@ -1720,128 +1189,4 @@ mod tests {
         let cloned_count = cloned_table.count_rows(None).await.unwrap();
         assert_eq!(source_count, cloned_count);
     }
-
-    #[tokio::test]
-    async fn test_create_empty_table_with_embeddings() {
-        use crate::embeddings::{EmbeddingDefinition, EmbeddingFunction};
-        use arrow_array::{
-            Array, FixedSizeListArray, Float32Array, RecordBatch, RecordBatchIterator, StringArray,
-        };
-        use std::borrow::Cow;
-
-        #[derive(Debug, Clone)]
-        struct MockEmbedding {
-            dim: usize,
-        }
-
-        impl EmbeddingFunction for MockEmbedding {
-            fn name(&self) -> &str {
-                "test_embedding"
-            }
-
-            fn source_type(&self) -> Result<Cow<'_, DataType>> {
-                Ok(Cow::Owned(DataType::Utf8))
-            }
-
-            fn dest_type(&self) -> Result<Cow<'_, DataType>> {
-                Ok(Cow::Owned(DataType::new_fixed_size_list(
-                    DataType::Float32,
-                    self.dim as i32,
-                    true,
-                )))
-            }
-
-            fn compute_source_embeddings(&self, source: Arc<dyn Array>) -> Result<Arc<dyn Array>> {
-                let len = source.len();
-                let values = vec![1.0f32; len * self.dim];
-                let values = Arc::new(Float32Array::from(values));
-                let field = Arc::new(Field::new("item", DataType::Float32, true));
-                Ok(Arc::new(FixedSizeListArray::new(
-                    field,
-                    self.dim as i32,
-                    values,
-                    None,
-                )))
-            }
-
-            fn compute_query_embeddings(&self, _input: Arc<dyn Array>) -> Result<Arc<dyn Array>> {
-                unimplemented!()
-            }
-        }
-
-        let tmp_dir = tempdir().unwrap();
-        let uri = tmp_dir.path().to_str().unwrap();
-        let db = connect(uri).execute().await.unwrap();
-
-        let embed_func = Arc::new(MockEmbedding { dim: 128 });
-        db.embedding_registry()
-            .register("test_embedding", embed_func.clone())
-            .unwrap();
-
-        let schema = Arc::new(Schema::new(vec![Field::new("name", DataType::Utf8, true)]));
-        let ed = EmbeddingDefinition {
-            source_column: "name".to_owned(),
-            dest_column: Some("name_embedding".to_owned()),
-            embedding_name: "test_embedding".to_owned(),
-        };
-
-        let table = db
-            .create_empty_table("test", schema)
-            .mode(CreateTableMode::Overwrite)
-            .add_embedding(ed)
-            .unwrap()
-            .execute()
-            .await
-            .unwrap();
-
-        let table_schema = table.schema().await.unwrap();
-        assert!(table_schema.column_with_name("name").is_some());
-        assert!(table_schema.column_with_name("name_embedding").is_some());
-
-        let embedding_field = table_schema.field_with_name("name_embedding").unwrap();
-        assert_eq!(
-            embedding_field.data_type(),
-            &DataType::new_fixed_size_list(DataType::Float32, 128, true)
-        );
-
-        let input_schema = Arc::new(Schema::new(vec![Field::new("name", DataType::Utf8, true)]));
-        let input_batch = RecordBatch::try_new(
-            input_schema.clone(),
-            vec![Arc::new(StringArray::from(vec![
-                Some("Alice"),
-                Some("Bob"),
-                Some("Charlie"),
-            ]))],
-        )
-        .unwrap();
-
-        let input_reader = Box::new(RecordBatchIterator::new(
-            vec![Ok(input_batch)].into_iter(),
-            input_schema,
-        ));
-
-        table.add(input_reader).execute().await.unwrap();
-
-        let results = table
-            .query()
-            .execute()
-            .await
-            .unwrap()
-            .try_collect::<Vec<_>>()
-            .await
-            .unwrap();
-
-        assert_eq!(results.len(), 1);
-        let batch = &results[0];
-        assert_eq!(batch.num_rows(), 3);
-        assert!(batch.column_by_name("name_embedding").is_some());
-
-        let embedding_col = batch
-            .column_by_name("name_embedding")
-            .unwrap()
-            .as_any()
-            .downcast_ref::<FixedSizeListArray>()
-            .unwrap();
-        assert_eq!(embedding_col.len(), 3);
-    }
 }

rust/lancedb/src/connection/create_table.rs

@@ -0,0 +1,612 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+use std::sync::Arc;
+
+use lance_io::object_store::StorageOptionsProvider;
+
+use crate::{
+    connection::{merge_storage_options, set_storage_options_provider},
+    data::scannable::{Scannable, WithEmbeddingsScannable},
+    database::{CreateTableMode, CreateTableRequest, Database},
+    embeddings::{EmbeddingDefinition, EmbeddingFunction, EmbeddingRegistry},
+    table::WriteOptions,
+    Error, Result, Table,
+};
+
+pub struct CreateTableBuilder {
+    parent: Arc<dyn Database>,
+    embeddings: Vec<(EmbeddingDefinition, Arc<dyn EmbeddingFunction>)>,
+    embedding_registry: Arc<dyn EmbeddingRegistry>,
+    request: CreateTableRequest,
+}
+
+impl CreateTableBuilder {
+    pub(super) fn new(
+        parent: Arc<dyn Database>,
+        embedding_registry: Arc<dyn EmbeddingRegistry>,
+        name: String,
+        data: Box<dyn Scannable>,
+    ) -> Self {
+        Self {
+            parent,
+            embeddings: Vec::new(),
+            embedding_registry,
+            request: CreateTableRequest::new(name, data),
+        }
+    }
+
+    /// Set the mode for creating the table
+    ///
+    /// This controls what happens if a table with the given name already exists
+    pub fn mode(mut self, mode: CreateTableMode) -> Self {
+        self.request.mode = mode;
+        self
+    }
+
+    /// Apply the given write options when writing the initial data
+    pub fn write_options(mut self, write_options: WriteOptions) -> Self {
+        self.request.write_options = write_options;
+        self
+    }
+
+    /// Set an option for the storage layer.
+    ///
+    /// Options already set on the connection will be inherited by the table,
+    /// but can be overridden here.
+    ///
+    /// See available options at <https://lancedb.com/docs/storage/>
+    pub fn storage_option(mut self, key: impl Into<String>, value: impl Into<String>) -> Self {
+        let store_params = self
+            .request
+            .write_options
+            .lance_write_params
+            .get_or_insert(Default::default())
+            .store_params
+            .get_or_insert(Default::default());
+        merge_storage_options(store_params, [(key.into(), value.into())]);
+        self
+    }
+
+    /// Set multiple options for the storage layer.
+    ///
+    /// Options already set on the connection will be inherited by the table,
+    /// but can be overridden here.
+    ///
+    /// See available options at <https://lancedb.com/docs/storage/>
+    pub fn storage_options(
+        mut self,
+        pairs: impl IntoIterator<Item = (impl Into<String>, impl Into<String>)>,
+    ) -> Self {
+        let store_params = self
+            .request
+            .write_options
+            .lance_write_params
+            .get_or_insert(Default::default())
+            .store_params
+            .get_or_insert(Default::default());
+        let updates = pairs
+            .into_iter()
+            .map(|(key, value)| (key.into(), value.into()));
+        merge_storage_options(store_params, updates);
+        self
+    }
+
+    /// Add an embedding definition to the table.
+    ///
+    /// The `embedding_name` must match the name of an embedding function that
+    /// was previously registered with the connection's [`EmbeddingRegistry`].
+    pub fn add_embedding(mut self, definition: EmbeddingDefinition) -> Result<Self> {
+        // Early verification of the embedding name
+        let embedding_func = self
+            .embedding_registry
+            .get(&definition.embedding_name)
+            .ok_or_else(|| Error::EmbeddingFunctionNotFound {
+                name: definition.embedding_name.clone(),
+                reason: "No embedding function found in the connection's embedding_registry"
+                    .to_string(),
+            })?;
+
+        self.embeddings.push((definition, embedding_func));
+        Ok(self)
+    }
+
+    /// Set the namespace for the table
+    pub fn namespace(mut self, namespace: Vec<String>) -> Self {
+        self.request.namespace = namespace;
+        self
+    }
+
+    /// Set a custom location for the table.
+    ///
+    /// If not set, the database will derive a location from its URI and the table name.
+    /// This is useful when integrating with namespace systems that manage table locations.
+    pub fn location(mut self, location: impl Into<String>) -> Self {
+        self.request.location = Some(location.into());
+        self
+    }
+
+    /// Set a storage options provider for automatic credential refresh.
+    ///
+    /// This allows tables to automatically refresh cloud storage credentials
+    /// when they expire, enabling long-running operations on remote storage.
+    pub fn storage_options_provider(mut self, provider: Arc<dyn StorageOptionsProvider>) -> Self {
+        let store_params = self
+            .request
+            .write_options
+            .lance_write_params
+            .get_or_insert(Default::default())
+            .store_params
+            .get_or_insert(Default::default());
+        set_storage_options_provider(store_params, provider);
+        self
+    }
+
+    /// Execute the create table operation
+    pub async fn execute(mut self) -> Result<Table> {
+        let embedding_registry = self.embedding_registry.clone();
+        let parent = self.parent.clone();
+
+        // If embeddings were configured via add_embedding(), wrap the data
+        if !self.embeddings.is_empty() {
+            let wrapped_data: Box<dyn Scannable> = Box::new(WithEmbeddingsScannable::try_new(
+                self.request.data,
+                self.embeddings,
+            )?);
+            self.request.data = wrapped_data;
+        }
+
+        Ok(Table::new_with_embedding_registry(
+            parent.create_table(self.request).await?,
+            parent,
+            embedding_registry,
+        ))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use arrow_array::{
+        record_batch, Array, FixedSizeListArray, Float32Array, RecordBatch, RecordBatchIterator,
+    };
+    use arrow_schema::{ArrowError, DataType, Field, Schema};
+    use futures::TryStreamExt;
+    use lance_file::version::LanceFileVersion;
+    use tempfile::tempdir;
+
+    use crate::{
+        arrow::{SendableRecordBatchStream, SimpleRecordBatchStream},
+        connect,
+        database::listing::{ListingDatabaseOptions, NewTableConfig},
+        embeddings::{EmbeddingDefinition, EmbeddingFunction, MemoryRegistry},
+        query::{ExecutableQuery, QueryBase, Select},
+        test_utils::embeddings::MockEmbed,
+    };
+    use std::borrow::Cow;
+
+    use super::*;
+
+    #[tokio::test]
+    async fn create_empty_table() {
+        let db = connect("memory://").execute().await.unwrap();
+        let schema = Arc::new(Schema::new(vec![
+            Field::new("id", DataType::Int64, false),
+            Field::new("value", DataType::Float64, false),
+        ]));
+        db.create_empty_table("name", schema.clone())
+            .execute()
+            .await
+            .unwrap();
+        let table = db.open_table("name").execute().await.unwrap();
+        assert_eq!(table.schema().await.unwrap(), schema);
+        assert_eq!(table.count_rows(None).await.unwrap(), 0);
+    }
+
+    async fn test_create_table_with_data<T>(data: T)
+    where
+        T: Scannable + 'static,
+    {
+        let db = connect("memory://").execute().await.unwrap();
+        let schema = data.schema();
+        db.create_table("data_table", data).execute().await.unwrap();
+        let table = db.open_table("data_table").execute().await.unwrap();
+        assert_eq!(table.count_rows(None).await.unwrap(), 3);
+        assert_eq!(table.schema().await.unwrap(), schema);
+    }
+
+    #[tokio::test]
+    async fn create_table_with_batch() {
+        let batch = record_batch!(("id", Int64, [1, 2, 3])).unwrap();
+        test_create_table_with_data(batch).await;
+    }
+
+    #[tokio::test]
+    async fn test_create_table_with_vec_batch() {
+        let data = vec![
+            record_batch!(("id", Int64, [1, 2])).unwrap(),
+            record_batch!(("id", Int64, [3])).unwrap(),
+        ];
+        test_create_table_with_data(data).await;
+    }
+
+    #[tokio::test]
+    async fn test_create_table_with_record_batch_reader() {
+        let data = vec![
+            record_batch!(("id", Int64, [1, 2])).unwrap(),
+            record_batch!(("id", Int64, [3])).unwrap(),
+        ];
+        let schema = data[0].schema();
+        let reader: Box<dyn arrow_array::RecordBatchReader + Send> = Box::new(
+            RecordBatchIterator::new(data.into_iter().map(Ok), schema.clone()),
+        );
+        test_create_table_with_data(reader).await;
+    }
+
+    #[tokio::test]
+    async fn test_create_table_with_stream() {
+        let data = vec![
+            record_batch!(("id", Int64, [1, 2])).unwrap(),
+            record_batch!(("id", Int64, [3])).unwrap(),
+        ];
+        let schema = data[0].schema();
+        let inner = futures::stream::iter(data.into_iter().map(Ok));
+        let stream: SendableRecordBatchStream = Box::pin(SimpleRecordBatchStream {
+            schema,
+            stream: inner,
+        });
+        test_create_table_with_data(stream).await;
+    }
+
+    #[derive(Debug)]
+    struct MyError;
+
+    impl std::fmt::Display for MyError {
+        fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+            write!(f, "MyError occurred")
+        }
+    }
+
+    impl std::error::Error for MyError {}
+
+    #[tokio::test]
+    async fn test_create_preserves_reader_error() {
+        let first_batch = record_batch!(("id", Int64, [1, 2])).unwrap();
+        let schema = first_batch.schema();
+        let iterator = vec![
+            Ok(first_batch),
+            Err(ArrowError::ExternalError(Box::new(MyError))),
+        ];
+        let reader: Box<dyn arrow_array::RecordBatchReader + Send> = Box::new(
+            RecordBatchIterator::new(iterator.into_iter(), schema.clone()),
+        );
+
+        let db = connect("memory://").execute().await.unwrap();
+        let result = db.create_table("failing_table", reader).execute().await;
+
+        assert!(result.is_err());
+        // TODO: when we upgrade to Lance 2.0.0, this should pass
+        // assert!(matches!(result, Err(Error::External { source})
+        //     if source.downcast_ref::<MyError>().is_some()
+        // ));
+    }
+
+    #[tokio::test]
+    async fn test_create_preserves_stream_error() {
+        let first_batch = record_batch!(("id", Int64, [1, 2])).unwrap();
+        let schema = first_batch.schema();
+        let iterator = vec![
+            Ok(first_batch),
+            Err(Error::External {
+                source: Box::new(MyError),
+            }),
+        ];
+        let stream = futures::stream::iter(iterator);
+        let stream: SendableRecordBatchStream = Box::pin(SimpleRecordBatchStream {
+            schema: schema.clone(),
+            stream,
+        });
+
+        let db = connect("memory://").execute().await.unwrap();
+        let result = db
+            .create_table("failing_stream_table", stream)
+            .execute()
+            .await;
+
+        assert!(result.is_err());
+        // TODO: when we upgrade to Lance 2.0.0, this should pass
+        // assert!(matches!(result, Err(Error::External { source})
+        //     if source.downcast_ref::<MyError>().is_some()
+        // ));
+    }
+
+    #[tokio::test]
+    #[allow(deprecated)]
+    async fn test_create_table_with_storage_options() {
+        let batch = record_batch!(("id", Int64, [1, 2, 3])).unwrap();
+        let db = connect("memory://").execute().await.unwrap();
+
+        let table = db
+            .create_table("options_table", batch)
+            .storage_option("timeout", "30s")
+            .storage_options([("retry_count", "3")])
+            .execute()
+            .await
+            .unwrap();
+
+        let final_options = table.storage_options().await.unwrap();
+        assert_eq!(final_options.get("timeout"), Some(&"30s".to_string()));
+        assert_eq!(final_options.get("retry_count"), Some(&"3".to_string()));
+    }
+
+    #[tokio::test]
+    async fn test_create_table_unregistered_embedding() {
+        let db = connect("memory://").execute().await.unwrap();
+        let batch = record_batch!(("text", Utf8, ["hello", "world"])).unwrap();
+
+        // Try to add an embedding that doesn't exist in the registry
+        let result = db
+            .create_table("embed_table", batch)
+            .add_embedding(EmbeddingDefinition::new(
+                "text",
+                "nonexistent_embedding_function",
+                None::<&str>,
+            ));
+
+        match result {
+            Err(Error::EmbeddingFunctionNotFound { name, .. }) => {
+                assert_eq!(name, "nonexistent_embedding_function");
+            }
+            Err(other) => panic!("Expected EmbeddingFunctionNotFound error, got: {:?}", other),
+            Ok(_) => panic!("Expected error, but got Ok"),
+        }
+    }
+
+    #[tokio::test]
+    async fn test_create_table_already_exists() {
+        let tmp_dir = tempdir().unwrap();
+        let uri = tmp_dir.path().to_str().unwrap();
+        let db = connect(uri).execute().await.unwrap();
+        let schema = Arc::new(Schema::new(vec![Field::new("x", DataType::Int32, false)]));
+        db.create_empty_table("test", schema.clone())
+            .execute()
+            .await
+            .unwrap();
+        db.create_empty_table("test", schema)
+            .mode(CreateTableMode::exist_ok(|mut req| {
+                req.index_cache_size = Some(16);
+                req
+            }))
+            .execute()
+            .await
+            .unwrap();
+        let other_schema = Arc::new(Schema::new(vec![Field::new("y", DataType::Int32, false)]));
+        assert!(db
+            .create_empty_table("test", other_schema.clone())
+            .execute()
+            .await
+            .is_err()); // TODO: assert what this error is
+        let overwritten = db
+            .create_empty_table("test", other_schema.clone())
+            .mode(CreateTableMode::Overwrite)
+            .execute()
+            .await
+            .unwrap();
+        assert_eq!(other_schema, overwritten.schema().await.unwrap());
+    }
+
+    #[tokio::test]
+    #[rstest::rstest]
+    #[case(LanceFileVersion::Legacy)]
+    #[case(LanceFileVersion::Stable)]
+    async fn test_create_table_with_storage_version(
+        #[case] data_storage_version: LanceFileVersion,
+    ) {
+        let db = connect("memory://")
+            .database_options(&ListingDatabaseOptions {
+                new_table_config: NewTableConfig {
+                    data_storage_version: Some(data_storage_version),
+                    ..Default::default()
+                },
+                ..Default::default()
+            })
+            .execute()
+            .await
+            .unwrap();
+
+        let batch = record_batch!(("id", Int64, [1, 2, 3])).unwrap();
+        let table = db
+            .create_table("legacy_table", batch)
+            .execute()
+            .await
+            .unwrap();
+
+        let native_table = table.as_native().unwrap();
+        let storage_format = native_table
+            .manifest()
+            .await
+            .unwrap()
+            .data_storage_format
+            .lance_file_version()
+            .unwrap();
+        // Compare resolved versions since Stable/Next are aliases that resolve at storage time
+        assert_eq!(storage_format.resolve(), data_storage_version.resolve());
+    }
+
+    #[tokio::test]
+    async fn test_create_table_with_embedding() {
+        // Register the mock embedding function
+        let registry = Arc::new(MemoryRegistry::new());
+        let mock_embedding: Arc<dyn EmbeddingFunction> = Arc::new(MockEmbed::new("mock", 4));
+        registry.register("mock", mock_embedding).unwrap();
+
+        // Connect with the custom registry
+        let conn = connect("memory://")
+            .embedding_registry(registry)
+            .execute()
+            .await
+            .unwrap();
+
+        // Create data without the embedding column
+        let batch = record_batch!(("text", Utf8, ["hello", "world", "test"])).unwrap();
+
+        // Create table with add_embedding - embeddings should be computed automatically
+        let table = conn
+            .create_table("embed_test", batch)
+            .add_embedding(EmbeddingDefinition::new(
+                "text",
+                "mock",
+                Some("text_embedding"),
+            ))
+            .unwrap()
+            .execute()
+            .await
+            .unwrap();
+
+        // Verify row count
+        assert_eq!(table.count_rows(None).await.unwrap(), 3);
+
+        // Verify the schema includes the embedding column
+        let result_schema = table.schema().await.unwrap();
+        assert_eq!(result_schema.fields().len(), 2);
+        assert_eq!(result_schema.field(0).name(), "text");
+        assert_eq!(result_schema.field(1).name(), "text_embedding");
+
+        // Verify the embedding column has the correct type
+        assert!(matches!(
+            result_schema.field(1).data_type(),
+            DataType::FixedSizeList(_, 4)
+        ));
+
+        // Query to verify the embeddings were computed
+        let results: Vec<RecordBatch> = table
+            .query()
+            .select(Select::columns(&["text", "text_embedding"]))
+            .execute()
+            .await
+            .unwrap()
+            .try_collect()
+            .await
+            .unwrap();
+
+        let total_rows: usize = results.iter().map(|b| b.num_rows()).sum();
+        assert_eq!(total_rows, 3);
+
+        // Check that all rows have embedding values (not null)
+        for batch in &results {
+            let embedding_col = batch.column(1);
+            assert_eq!(embedding_col.null_count(), 0);
+            assert_eq!(embedding_col.len(), batch.num_rows());
+        }
+
+        // Verify the schema metadata contains the column definitions
+        assert!(
+            result_schema
+                .metadata
+                .contains_key("lancedb::column_definitions"),
+            "Schema metadata should contain column definitions"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_create_empty_table_with_embeddings() {
+        #[derive(Debug, Clone)]
+        struct MockEmbedding {
+            dim: usize,
+        }
+
+        impl EmbeddingFunction for MockEmbedding {
+            fn name(&self) -> &str {
+                "test_embedding"
+            }
+
+            fn source_type(&self) -> Result<Cow<'_, DataType>> {
+                Ok(Cow::Owned(DataType::Utf8))
+            }
+
+            fn dest_type(&self) -> Result<Cow<'_, DataType>> {
+                Ok(Cow::Owned(DataType::new_fixed_size_list(
+                    DataType::Float32,
+                    self.dim as i32,
+                    true,
+                )))
+            }
+
+            fn compute_source_embeddings(&self, source: Arc<dyn Array>) -> Result<Arc<dyn Array>> {
+                let len = source.len();
+                let values = vec![1.0f32; len * self.dim];
+                let values = Arc::new(Float32Array::from(values));
+                let field = Arc::new(Field::new("item", DataType::Float32, true));
+                Ok(Arc::new(FixedSizeListArray::new(
+                    field,
+                    self.dim as i32,
+                    values,
+                    None,
+                )))
+            }
+
+            fn compute_query_embeddings(&self, _input: Arc<dyn Array>) -> Result<Arc<dyn Array>> {
+                unimplemented!()
+            }
+        }
+
+        let tmp_dir = tempdir().unwrap();
+        let uri = tmp_dir.path().to_str().unwrap();
+        let db = connect(uri).execute().await.unwrap();
+
+        let embed_func = Arc::new(MockEmbedding { dim: 128 });
+        db.embedding_registry()
+            .register("test_embedding", embed_func.clone())
+            .unwrap();
+
+        let schema = Arc::new(Schema::new(vec![Field::new("name", DataType::Utf8, true)]));
+        let ed = EmbeddingDefinition {
+            source_column: "name".to_owned(),
+            dest_column: Some("name_embedding".to_owned()),
+            embedding_name: "test_embedding".to_owned(),
+        };
+
+        let table = db
+            .create_empty_table("test", schema)
+            .mode(CreateTableMode::Overwrite)
+            .add_embedding(ed)
+            .unwrap()
+            .execute()
+            .await
+            .unwrap();
+
+        let table_schema = table.schema().await.unwrap();
+        assert!(table_schema.column_with_name("name").is_some());
+        assert!(table_schema.column_with_name("name_embedding").is_some());
+
+        let embedding_field = table_schema.field_with_name("name_embedding").unwrap();
+        assert_eq!(
+            embedding_field.data_type(),
+            &DataType::new_fixed_size_list(DataType::Float32, 128, true)
+        );
+
+        let input_batch = record_batch!(("name", Utf8, ["Alice", "Bob", "Charlie"])).unwrap();
+        table.add(input_batch).execute().await.unwrap();
+
+        let results = table
+            .query()
+            .execute()
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+
+        assert_eq!(results.len(), 1);
+        let batch = &results[0];
+        assert_eq!(batch.num_rows(), 3);
+        assert!(batch.column_by_name("name_embedding").is_some());
+
+        let embedding_col = batch
+            .column_by_name("name_embedding")
+            .unwrap()
+            .as_any()
+            .downcast_ref::<FixedSizeListArray>()
+            .unwrap();
+        assert_eq!(embedding_col.len(), 3);
+    }
+}

rust/lancedb/src/data.rs

@@ -5,3 +5,4 @@
 
 pub mod inspect;
 pub mod sanitize;
+pub mod scannable;

rust/lancedb/src/data/scannable.rs

@@ -0,0 +1,968 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+//! Data source abstraction for LanceDB.
+//!
+//! This module provides a [`Scannable`] trait that allows input data sources to express
+//! capabilities (row count, rescannability) so the insert pipeline can make
+//! better decisions about write parallelism and retry strategies.
+
+use std::sync::Arc;
+
+use crate::arrow::{
+    SendableRecordBatchStream, SendableRecordBatchStreamExt, SimpleRecordBatchStream,
+};
+use crate::embeddings::{
+    compute_embeddings_for_batch, compute_output_schema, EmbeddingDefinition, EmbeddingFunction,
+    EmbeddingRegistry,
+};
+use crate::table::{ColumnDefinition, ColumnKind, TableDefinition};
+use crate::{Error, Result};
+use arrow_array::{ArrayRef, RecordBatch, RecordBatchIterator, RecordBatchReader};
+use arrow_schema::{ArrowError, SchemaRef};
+use async_trait::async_trait;
+use futures::stream::once;
+use futures::StreamExt;
+use lance_datafusion::utils::StreamingWriteSource;
+
+pub trait Scannable: Send {
+    /// Returns the schema of the data.
+    fn schema(&self) -> SchemaRef;
+
+    /// Read data as a stream of record batches.
+    ///
+    /// For rescannable sources (in-memory data like RecordBatch, Vec<RecordBatch>),
+    /// this can be called multiple times and returns cloned data each time.
+    ///
+    /// For non-rescannable sources (streams, readers), this can only be called once.
+    /// Calling it a second time returns a stream whose first item is an error.
+    fn scan_as_stream(&mut self) -> SendableRecordBatchStream;
+
+    /// Optional hint about the number of rows.
+    ///
+    /// When available, this allows the pipeline to estimate total data size
+    /// and choose appropriate partitioning.
+    fn num_rows(&self) -> Option<usize> {
+        None
+    }
+
+    /// Whether the source can be re-read from the beginning.
+    ///
+    /// `true` for in-memory data (Tables, DataFrames) and disk-based sources (Datasets).
+    /// `false` for streaming sources (DuckDB results, network streams).
+    ///
+    /// When true, the pipeline can retry failed writes by rescanning.
+    fn rescannable(&self) -> bool {
+        false
+    }
+}
+
+impl std::fmt::Debug for dyn Scannable {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("Scannable")
+            .field("schema", &self.schema())
+            .field("num_rows", &self.num_rows())
+            .field("rescannable", &self.rescannable())
+            .finish()
+    }
+}
+
+impl Scannable for RecordBatch {
+    fn schema(&self) -> SchemaRef {
+        Self::schema(self)
+    }
+
+    fn scan_as_stream(&mut self) -> SendableRecordBatchStream {
+        let batch = self.clone();
+        let schema = batch.schema();
+        Box::pin(SimpleRecordBatchStream {
+            schema,
+            stream: once(async move { Ok(batch) }),
+        })
+    }
+
+    fn num_rows(&self) -> Option<usize> {
+        Some(Self::num_rows(self))
+    }
+
+    fn rescannable(&self) -> bool {
+        true
+    }
+}
+
+impl Scannable for Vec<RecordBatch> {
+    fn schema(&self) -> SchemaRef {
+        if self.is_empty() {
+            Arc::new(arrow_schema::Schema::empty())
+        } else {
+            self[0].schema()
+        }
+    }
+
+    fn scan_as_stream(&mut self) -> SendableRecordBatchStream {
+        if self.is_empty() {
+            let schema = Scannable::schema(self);
+            return Box::pin(SimpleRecordBatchStream {
+                schema,
+                stream: once(async {
+                    Err(Error::InvalidInput {
+                        message: "Cannot scan an empty Vec<RecordBatch>".to_string(),
+                    })
+                }),
+            });
+        }
+        let schema = Scannable::schema(self);
+        let batches = self.clone();
+        let stream = futures::stream::iter(batches.into_iter().map(Ok));
+        Box::pin(SimpleRecordBatchStream { schema, stream })
+    }
+
+    fn num_rows(&self) -> Option<usize> {
+        Some(self.iter().map(|b| b.num_rows()).sum())
+    }
+
+    fn rescannable(&self) -> bool {
+        true
+    }
+}
+
+impl Scannable for Box<dyn RecordBatchReader + Send> {
+    fn schema(&self) -> SchemaRef {
+        RecordBatchReader::schema(self.as_ref())
+    }
+
+    fn scan_as_stream(&mut self) -> SendableRecordBatchStream {
+        let schema = Scannable::schema(self);
+
+        // Swap self with a reader that errors on iteration, so a second call
+        // produces a clear error instead of silently returning empty data.
+        let err_reader: Box<dyn RecordBatchReader + Send> = Box::new(RecordBatchIterator::new(
+            vec![Err(ArrowError::InvalidArgumentError(
+                "Reader has already been consumed".into(),
+            ))],
+            schema.clone(),
+        ));
+        let reader = std::mem::replace(self, err_reader);
+
+        // Bridge the blocking RecordBatchReader to an async stream via a channel.
+        let (tx, rx) = tokio::sync::mpsc::channel::<crate::Result<RecordBatch>>(2);
+        tokio::task::spawn_blocking(move || {
+            for batch_result in reader {
+                let result = batch_result.map_err(Into::into);
+                if tx.blocking_send(result).is_err() {
+                    break;
+                }
+            }
+        });
+
+        let stream = futures::stream::unfold(rx, |mut rx| async move {
+            rx.recv().await.map(|batch| (batch, rx))
+        })
+        .fuse();
+
+        Box::pin(SimpleRecordBatchStream { schema, stream })
+    }
+}
+
+impl Scannable for SendableRecordBatchStream {
+    fn schema(&self) -> SchemaRef {
+        self.as_ref().schema()
+    }
+
+    fn scan_as_stream(&mut self) -> SendableRecordBatchStream {
+        let schema = Scannable::schema(self);
+
+        // Swap self with an error stream so a second call produces a clear error.
+        let error_stream = Box::pin(SimpleRecordBatchStream {
+            schema: schema.clone(),
+            stream: once(async {
+                Err(Error::InvalidInput {
+                    message: "Stream has already been consumed".to_string(),
+                })
+            }),
+        });
+        std::mem::replace(self, error_stream)
+    }
+}
+
+#[async_trait]
+impl StreamingWriteSource for Box<dyn Scannable> {
+    fn arrow_schema(&self) -> SchemaRef {
+        self.schema()
+    }
+
+    fn into_stream(mut self) -> datafusion_physical_plan::SendableRecordBatchStream {
+        self.scan_as_stream().into_df_stream()
+    }
+}
+
+/// A scannable that applies embeddings to the stream.
+pub struct WithEmbeddingsScannable {
+    inner: Box<dyn Scannable>,
+    embeddings: Vec<(EmbeddingDefinition, Arc<dyn EmbeddingFunction>)>,
+    output_schema: SchemaRef,
+}
+
+impl WithEmbeddingsScannable {
+    /// Create a new WithEmbeddingsScannable.
+    ///
+    /// The embeddings are applied to the inner scannable's data as new columns.
+    pub fn try_new(
+        inner: Box<dyn Scannable>,
+        embeddings: Vec<(EmbeddingDefinition, Arc<dyn EmbeddingFunction>)>,
+    ) -> Result<Self> {
+        let output_schema = compute_output_schema(&inner.schema(), &embeddings)?;
+
+        // Build column definitions: Physical for base columns, Embedding for new ones
+        let base_col_count = inner.schema().fields().len();
+        let column_definitions: Vec<ColumnDefinition> = (0..base_col_count)
+            .map(|_| ColumnDefinition {
+                kind: ColumnKind::Physical,
+            })
+            .chain(embeddings.iter().map(|(ed, _)| ColumnDefinition {
+                kind: ColumnKind::Embedding(ed.clone()),
+            }))
+            .collect();
+
+        let table_definition = TableDefinition::new(output_schema, column_definitions);
+        let output_schema = table_definition.into_rich_schema();
+
+        Self::with_schema(inner, embeddings, output_schema)
+    }
+
+    /// Create a WithEmbeddingsScannable with a specific output schema.
+    ///
+    /// Use this when the table schema is already known (e.g. during add) to
+    /// avoid nullability mismatches between the embedding function's declared
+    /// type and the table's stored type.
+    pub fn with_schema(
+        inner: Box<dyn Scannable>,
+        embeddings: Vec<(EmbeddingDefinition, Arc<dyn EmbeddingFunction>)>,
+        output_schema: SchemaRef,
+    ) -> Result<Self> {
+        Ok(Self {
+            inner,
+            embeddings,
+            output_schema,
+        })
+    }
+}
+
+impl Scannable for WithEmbeddingsScannable {
+    fn schema(&self) -> SchemaRef {
+        self.output_schema.clone()
+    }
+
+    fn scan_as_stream(&mut self) -> SendableRecordBatchStream {
+        let inner_stream = self.inner.scan_as_stream();
+        let embeddings = self.embeddings.clone();
+        let output_schema = self.output_schema.clone();
+        let stream_schema = output_schema.clone();
+
+        let mapped_stream = inner_stream.then(move |batch_result| {
+            let embeddings = embeddings.clone();
+            let output_schema = output_schema.clone();
+            async move {
+                let batch = batch_result?;
+                let result = tokio::task::spawn_blocking(move || {
+                    compute_embeddings_for_batch(batch, &embeddings)
+                })
+                .await
+                .map_err(|e| Error::Runtime {
+                    message: format!("Task panicked during embedding computation: {}", e),
+                })??;
+                // Cast columns to match the declared output schema. The data is
+                // identical but field metadata (e.g. nested nullability) may
+                // differ between the embedding function output and the table.
+                let columns: Vec<ArrayRef> = result
+                    .columns()
+                    .iter()
+                    .enumerate()
+                    .map(|(i, col)| {
+                        let target_type = output_schema.field(i).data_type();
+                        if col.data_type() == target_type {
+                            Ok(col.clone())
+                        } else {
+                            arrow_cast::cast(col, target_type).map_err(Error::from)
+                        }
+                    })
+                    .collect::<Result<_>>()?;
+                let result = RecordBatch::try_new(output_schema, columns)?;
+                Ok(result)
+            }
+        });
+
+        Box::pin(SimpleRecordBatchStream {
+            schema: stream_schema,
+            stream: mapped_stream,
+        })
+    }
+
+    fn num_rows(&self) -> Option<usize> {
+        self.inner.num_rows()
+    }
+
+    fn rescannable(&self) -> bool {
+        self.inner.rescannable()
+    }
+}
+
+pub fn scannable_with_embeddings(
+    inner: Box<dyn Scannable>,
+    table_definition: &TableDefinition,
+    registry: Option<&Arc<dyn EmbeddingRegistry>>,
+) -> Result<Box<dyn Scannable>> {
+    if let Some(registry) = registry {
+        let mut embeddings = Vec::with_capacity(table_definition.column_definitions.len());
+        for cd in table_definition.column_definitions.iter() {
+            if let ColumnKind::Embedding(embedding_def) = &cd.kind {
+                match registry.get(&embedding_def.embedding_name) {
+                    Some(func) => {
+                        embeddings.push((embedding_def.clone(), func));
+                    }
+                    None => {
+                        return Err(Error::EmbeddingFunctionNotFound {
+                            name: embedding_def.embedding_name.clone(),
+                            reason: format!(
+                                "Table was defined with an embedding column `{}` but no embedding function was found with that name within the registry.",
+                                embedding_def.embedding_name
+                            ),
+                        });
+                    }
+                }
+            }
+        }
+
+        if !embeddings.is_empty() {
+            // Use the table's schema so embedding column types (including nested
+            // nullability) match what's stored, avoiding mismatches with the
+            // embedding function's declared dest_type.
+            return Ok(Box::new(WithEmbeddingsScannable::with_schema(
+                inner,
+                embeddings,
+                table_definition.schema.clone(),
+            )?));
+        }
+    }
+
+    Ok(inner)
+}
+
+/// A wrapper that buffers the first RecordBatch from a Scannable so we can
+/// inspect it (e.g. to estimate data size) without losing it.
+pub(crate) struct PeekedScannable {
+    inner: Box<dyn Scannable>,
+    peeked: Option<RecordBatch>,
+    /// The first item from the stream, if it was an error. Stored so we can
+    /// re-emit it from `scan_as_stream` instead of silently dropping it.
+    first_error: Option<crate::Error>,
+    stream: Option<SendableRecordBatchStream>,
+}
+
+impl PeekedScannable {
+    pub fn new(inner: Box<dyn Scannable>) -> Self {
+        Self {
+            inner,
+            peeked: None,
+            first_error: None,
+            stream: None,
+        }
+    }
+
+    /// Reads and buffers the first batch from the inner scannable.
+    /// Returns a clone of it. Subsequent calls return the same batch.
+    ///
+    /// Returns `None` if the stream is empty or the first item is an error.
+    /// Errors are preserved and re-emitted by `scan_as_stream`.
+    pub async fn peek(&mut self) -> Option<RecordBatch> {
+        if self.peeked.is_some() {
+            return self.peeked.clone();
+        }
+        // Already peeked and got an error or empty stream.
+        if self.stream.is_some() || self.first_error.is_some() {
+            return None;
+        }
+        let mut stream = self.inner.scan_as_stream();
+        match stream.next().await {
+            Some(Ok(batch)) => {
+                self.peeked = Some(batch.clone());
+                self.stream = Some(stream);
+                Some(batch)
+            }
+            Some(Err(e)) => {
+                self.first_error = Some(e);
+                self.stream = Some(stream);
+                None
+            }
+            None => {
+                self.stream = Some(stream);
+                None
+            }
+        }
+    }
+}
+
+impl Scannable for PeekedScannable {
+    fn schema(&self) -> SchemaRef {
+        self.inner.schema()
+    }
+
+    fn num_rows(&self) -> Option<usize> {
+        self.inner.num_rows()
+    }
+
+    fn rescannable(&self) -> bool {
+        self.inner.rescannable()
+    }
+
+    fn scan_as_stream(&mut self) -> SendableRecordBatchStream {
+        let schema = self.inner.schema();
+
+        // If peek() hit an error, prepend it so downstream sees the error.
+        let error_item = self.first_error.take().map(Err);
+
+        match (self.peeked.take(), self.stream.take()) {
+            (Some(batch), Some(rest)) => {
+                let prepend = futures::stream::once(std::future::ready(Ok(batch)));
+                Box::pin(SimpleRecordBatchStream {
+                    schema,
+                    stream: prepend.chain(rest),
+                })
+            }
+            (Some(batch), None) => Box::pin(SimpleRecordBatchStream {
+                schema,
+                stream: futures::stream::once(std::future::ready(Ok(batch))),
+            }),
+            (None, Some(rest)) => {
+                if let Some(err) = error_item {
+                    let stream = futures::stream::once(std::future::ready(err));
+                    Box::pin(SimpleRecordBatchStream { schema, stream })
+                } else {
+                    rest
+                }
+            }
+            (None, None) => {
+                // peek() was never called — just delegate
+                self.inner.scan_as_stream()
+            }
+        }
+    }
+}
+
+/// Compute the number of write partitions based on data size estimates.
+///
+/// `sample_bytes` and `sample_rows` come from a representative batch and are
+/// used to estimate per-row size. `total_rows_hint` is the total row count
+/// when known; otherwise `sample_rows` row count is used as a lower bound
+/// estimate.
+///
+/// Targets roughly 1 million rows or 2 GB per partition, capped at
+/// `max_partitions` (typically the number of available CPU cores).
+pub(crate) fn estimate_write_partitions(
+    sample_bytes: usize,
+    sample_rows: usize,
+    total_rows_hint: Option<usize>,
+    max_partitions: usize,
+) -> usize {
+    if sample_rows == 0 {
+        return 1;
+    }
+    let bytes_per_row = sample_bytes / sample_rows;
+    let total_rows = total_rows_hint.unwrap_or(sample_rows);
+    let total_bytes = total_rows * bytes_per_row;
+    let by_rows = total_rows.div_ceil(1_000_000);
+    let by_bytes = total_bytes.div_ceil(2 * 1024 * 1024 * 1024);
+    by_rows.max(by_bytes).max(1).min(max_partitions)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use arrow_array::record_batch;
+    use futures::TryStreamExt;
+
+    #[tokio::test]
+    async fn test_record_batch_rescannable() {
+        let mut batch = record_batch!(("id", Int64, [0, 1, 2])).unwrap();
+
+        let stream1 = batch.scan_as_stream();
+        let batches1: Vec<RecordBatch> = stream1.try_collect().await.unwrap();
+        assert_eq!(batches1.len(), 1);
+        assert_eq!(batches1[0], batch);
+
+        assert!(batch.rescannable());
+        let stream2 = batch.scan_as_stream();
+        let batches2: Vec<RecordBatch> = stream2.try_collect().await.unwrap();
+        assert_eq!(batches2.len(), 1);
+        assert_eq!(batches2[0], batch);
+    }
+
+    #[tokio::test]
+    async fn test_vec_batch_rescannable() {
+        let mut batches = vec![
+            record_batch!(("id", Int64, [0, 1])).unwrap(),
+            record_batch!(("id", Int64, [2, 3, 4])).unwrap(),
+        ];
+
+        let stream1 = batches.scan_as_stream();
+        let result1: Vec<RecordBatch> = stream1.try_collect().await.unwrap();
+        assert_eq!(result1.len(), 2);
+        assert_eq!(result1[0], batches[0]);
+        assert_eq!(result1[1], batches[1]);
+
+        assert!(batches.rescannable());
+        let stream2 = batches.scan_as_stream();
+        let result2: Vec<RecordBatch> = stream2.try_collect().await.unwrap();
+        assert_eq!(result2.len(), 2);
+        assert_eq!(result2[0], batches[0]);
+        assert_eq!(result2[1], batches[1]);
+    }
+
+    #[tokio::test]
+    async fn test_vec_batch_empty_errors() {
+        let mut empty: Vec<RecordBatch> = vec![];
+        let mut stream = empty.scan_as_stream();
+        let result = stream.next().await;
+        assert!(result.is_some());
+        assert!(result.unwrap().is_err());
+    }
+
+    #[tokio::test]
+    async fn test_reader_not_rescannable() {
+        let batch = record_batch!(("id", Int64, [0, 1, 2])).unwrap();
+        let schema = batch.schema();
+        let mut reader: Box<dyn arrow_array::RecordBatchReader + Send> = Box::new(
+            RecordBatchIterator::new(vec![Ok(batch.clone())], schema.clone()),
+        );
+
+        let stream1 = reader.scan_as_stream();
+        let result1: Vec<RecordBatch> = stream1.try_collect().await.unwrap();
+        assert_eq!(result1.len(), 1);
+        assert_eq!(result1[0], batch);
+
+        assert!(!reader.rescannable());
+        // Second call returns a stream whose first item is an error
+        let mut stream2 = reader.scan_as_stream();
+        let result2 = stream2.next().await;
+        assert!(result2.is_some());
+        assert!(result2.unwrap().is_err());
+    }
+
+    #[tokio::test]
+    async fn test_stream_not_rescannable() {
+        let batch = record_batch!(("id", Int64, [0, 1, 2])).unwrap();
+        let schema = batch.schema();
+        let inner_stream = futures::stream::iter(vec![Ok(batch.clone())]);
+        let mut stream: SendableRecordBatchStream = Box::pin(SimpleRecordBatchStream {
+            schema: schema.clone(),
+            stream: inner_stream,
+        });
+
+        let stream1 = stream.scan_as_stream();
+        let result1: Vec<RecordBatch> = stream1.try_collect().await.unwrap();
+        assert_eq!(result1.len(), 1);
+        assert_eq!(result1[0], batch);
+
+        assert!(!stream.rescannable());
+        // Second call returns a stream whose first item is an error
+        let mut stream2 = stream.scan_as_stream();
+        let result2 = stream2.next().await;
+        assert!(result2.is_some());
+        assert!(result2.unwrap().is_err());
+    }
+
+    mod peeked_scannable_tests {
+        use crate::test_utils::TestCustomError;
+
+        use super::*;
+
+        #[tokio::test]
+        async fn test_peek_returns_first_batch() {
+            let batch = record_batch!(("id", Int64, [1, 2, 3])).unwrap();
+            let mut peeked = PeekedScannable::new(Box::new(batch.clone()));
+
+            let first = peeked.peek().await.unwrap();
+            assert_eq!(first, batch);
+        }
+
+        #[tokio::test]
+        async fn test_peek_is_idempotent() {
+            let batch = record_batch!(("id", Int64, [1, 2, 3])).unwrap();
+            let mut peeked = PeekedScannable::new(Box::new(batch.clone()));
+
+            let first = peeked.peek().await.unwrap();
+            let second = peeked.peek().await.unwrap();
+            assert_eq!(first, second);
+        }
+
+        #[tokio::test]
+        async fn test_scan_after_peek_returns_all_data() {
+            let batches = vec![
+                record_batch!(("id", Int64, [1, 2])).unwrap(),
+                record_batch!(("id", Int64, [3, 4, 5])).unwrap(),
+            ];
+            let mut peeked = PeekedScannable::new(Box::new(batches.clone()));
+
+            let first = peeked.peek().await.unwrap();
+            assert_eq!(first, batches[0]);
+
+            let result: Vec<RecordBatch> = peeked.scan_as_stream().try_collect().await.unwrap();
+            assert_eq!(result.len(), 2);
+            assert_eq!(result[0], batches[0]);
+            assert_eq!(result[1], batches[1]);
+        }
+
+        #[tokio::test]
+        async fn test_scan_without_peek_passes_through() {
+            let batch = record_batch!(("id", Int64, [1, 2, 3])).unwrap();
+            let mut peeked = PeekedScannable::new(Box::new(batch.clone()));
+
+            let result: Vec<RecordBatch> = peeked.scan_as_stream().try_collect().await.unwrap();
+            assert_eq!(result.len(), 1);
+            assert_eq!(result[0], batch);
+        }
+
+        #[tokio::test]
+        async fn test_delegates_num_rows() {
+            let batches = vec![
+                record_batch!(("id", Int64, [1, 2])).unwrap(),
+                record_batch!(("id", Int64, [3])).unwrap(),
+            ];
+            let peeked = PeekedScannable::new(Box::new(batches));
+            assert_eq!(peeked.num_rows(), Some(3));
+        }
+
+        #[tokio::test]
+        async fn test_non_rescannable_stream_data_preserved() {
+            let batches = vec![
+                record_batch!(("id", Int64, [1, 2])).unwrap(),
+                record_batch!(("id", Int64, [3])).unwrap(),
+            ];
+            let schema = batches[0].schema();
+            let inner = futures::stream::iter(batches.clone().into_iter().map(Ok));
+            let stream: SendableRecordBatchStream = Box::pin(SimpleRecordBatchStream {
+                schema,
+                stream: inner,
+            });
+
+            let mut peeked = PeekedScannable::new(Box::new(stream));
+            assert!(!peeked.rescannable());
+            assert_eq!(peeked.num_rows(), None);
+
+            let first = peeked.peek().await.unwrap();
+            assert_eq!(first, batches[0]);
+
+            // All data is still available via scan_as_stream
+            let result: Vec<RecordBatch> = peeked.scan_as_stream().try_collect().await.unwrap();
+            assert_eq!(result.len(), 2);
+            assert_eq!(result[0], batches[0]);
+            assert_eq!(result[1], batches[1]);
+        }
+
+        #[tokio::test]
+        async fn test_error_in_first_batch_propagates() {
+            let schema = Arc::new(arrow_schema::Schema::new(vec![arrow_schema::Field::new(
+                "id",
+                arrow_schema::DataType::Int64,
+                false,
+            )]));
+            let inner = futures::stream::iter(vec![Err(Error::External {
+                source: Box::new(TestCustomError),
+            })]);
+            let stream: SendableRecordBatchStream = Box::pin(SimpleRecordBatchStream {
+                schema,
+                stream: inner,
+            });
+
+            let mut peeked = PeekedScannable::new(Box::new(stream));
+
+            // peek returns None for errors
+            assert!(peeked.peek().await.is_none());
+
+            // But the error should come through when scanning
+            let mut stream = peeked.scan_as_stream();
+            let first = stream.next().await.unwrap();
+            assert!(first.is_err());
+            let err = first.unwrap_err();
+            assert!(
+                matches!(&err, Error::External { source } if source.downcast_ref::<TestCustomError>().is_some()),
+                "Expected TestCustomError to be preserved, got: {err}"
+            );
+        }
+
+        #[tokio::test]
+        async fn test_error_in_later_batch_propagates() {
+            let good_batch = record_batch!(("id", Int64, [1, 2])).unwrap();
+            let schema = good_batch.schema();
+            let inner = futures::stream::iter(vec![
+                Ok(good_batch.clone()),
+                Err(Error::External {
+                    source: Box::new(TestCustomError),
+                }),
+            ]);
+            let stream: SendableRecordBatchStream = Box::pin(SimpleRecordBatchStream {
+                schema,
+                stream: inner,
+            });
+
+            let mut peeked = PeekedScannable::new(Box::new(stream));
+
+            // peek succeeds with the first batch
+            let first = peeked.peek().await.unwrap();
+            assert_eq!(first, good_batch);
+
+            // scan_as_stream should yield the first batch, then the error
+            let mut stream = peeked.scan_as_stream();
+            let batch1 = stream.next().await.unwrap().unwrap();
+            assert_eq!(batch1, good_batch);
+
+            let batch2 = stream.next().await.unwrap();
+            assert!(batch2.is_err());
+            let err = batch2.unwrap_err();
+            assert!(
+                matches!(&err, Error::External { source } if source.downcast_ref::<TestCustomError>().is_some()),
+                "Expected TestCustomError to be preserved, got: {err}"
+            );
+        }
+
+        #[tokio::test]
+        async fn test_empty_stream_returns_none() {
+            let schema = Arc::new(arrow_schema::Schema::new(vec![arrow_schema::Field::new(
+                "id",
+                arrow_schema::DataType::Int64,
+                false,
+            )]));
+            let inner = futures::stream::empty();
+            let stream: SendableRecordBatchStream = Box::pin(SimpleRecordBatchStream {
+                schema,
+                stream: inner,
+            });
+
+            let mut peeked = PeekedScannable::new(Box::new(stream));
+            assert!(peeked.peek().await.is_none());
+
+            // Scanning an empty (post-peek) stream should yield nothing
+            let result: Vec<RecordBatch> = peeked.scan_as_stream().try_collect().await.unwrap();
+            assert!(result.is_empty());
+        }
+    }
+
+    mod estimate_write_partitions_tests {
+        use super::*;
+
+        #[test]
+        fn test_small_data_single_partition() {
+            // 100 rows * 24 bytes/row = 2400 bytes — well under both thresholds
+            assert_eq!(estimate_write_partitions(2400, 100, Some(100), 8), 1);
+        }
+
+        #[test]
+        fn test_scales_by_row_count() {
+            // 2.5M rows at 24 bytes/row — row threshold dominates
+            // ceil(2_500_000 / 1_000_000) = 3
+            assert_eq!(estimate_write_partitions(72, 3, Some(2_500_000), 8), 3);
+        }
+
+        #[test]
+        fn test_scales_by_byte_size() {
+            // 100k rows at 40KB/row = ~4GB total → ceil(4GB / 2GB) = 2
+            let sample_bytes = 40_000 * 10;
+            assert_eq!(
+                estimate_write_partitions(sample_bytes, 10, Some(100_000), 8),
+                2
+            );
+        }
+
+        #[test]
+        fn test_capped_at_max_partitions() {
+            // 10M rows would want 10 partitions, but capped at 4
+            assert_eq!(estimate_write_partitions(72, 3, Some(10_000_000), 4), 4);
+        }
+
+        #[test]
+        fn test_zero_sample_rows_returns_one() {
+            assert_eq!(estimate_write_partitions(0, 0, Some(1_000_000), 8), 1);
+        }
+
+        #[test]
+        fn test_no_row_hint_uses_sample_size() {
+            // Without a hint, uses sample_rows (3), which is small
+            assert_eq!(estimate_write_partitions(72, 3, None, 8), 1);
+        }
+
+        #[test]
+        fn test_always_at_least_one() {
+            assert_eq!(estimate_write_partitions(24, 1, Some(1), 8), 1);
+        }
+    }
+
+    mod embedding_tests {
+        use super::*;
+        use crate::embeddings::MemoryRegistry;
+        use crate::table::{ColumnDefinition, ColumnKind};
+        use crate::test_utils::embeddings::MockEmbed;
+        use arrow_array::Array as _;
+        use arrow_array::{ArrayRef, StringArray};
+        use arrow_schema::{DataType, Field, Schema};
+
+        #[tokio::test]
+        async fn test_with_embeddings_scannable() {
+            let schema = Arc::new(Schema::new(vec![Field::new("text", DataType::Utf8, false)]));
+            let text_array = StringArray::from(vec!["hello", "world", "test"]);
+            let batch =
+                RecordBatch::try_new(schema.clone(), vec![Arc::new(text_array) as ArrayRef])
+                    .unwrap();
+
+            let mock_embedding: Arc<dyn EmbeddingFunction> = Arc::new(MockEmbed::new("mock", 4));
+            let embedding_def = EmbeddingDefinition::new("text", "mock", Some("text_embedding"));
+
+            let mut scannable = WithEmbeddingsScannable::try_new(
+                Box::new(batch.clone()),
+                vec![(embedding_def, mock_embedding)],
+            )
+            .unwrap();
+
+            // Check that schema has the embedding column
+            let output_schema = scannable.schema();
+            assert_eq!(output_schema.fields().len(), 2);
+            assert_eq!(output_schema.field(0).name(), "text");
+            assert_eq!(output_schema.field(1).name(), "text_embedding");
+
+            // Check num_rows and rescannable are preserved
+            assert_eq!(scannable.num_rows(), Some(3));
+            assert!(scannable.rescannable());
+
+            // Read the data
+            let stream = scannable.scan_as_stream();
+            let results: Vec<RecordBatch> = stream.try_collect().await.unwrap();
+            assert_eq!(results.len(), 1);
+
+            let result_batch = &results[0];
+            assert_eq!(result_batch.num_rows(), 3);
+            assert_eq!(result_batch.num_columns(), 2);
+
+            // Verify the embedding column is present and has the right shape
+            let embedding_col = result_batch.column(1);
+            assert_eq!(embedding_col.len(), 3);
+        }
+
+        #[tokio::test]
+        async fn test_maybe_embedded_scannable_no_embeddings() {
+            let batch = record_batch!(("id", Int64, [1, 2, 3])).unwrap();
+
+            // Create a table definition with no embedding columns
+            let table_def = TableDefinition::new_from_schema(batch.schema());
+
+            // Even with a registry, if there are no embedding columns, it's a passthrough
+            let registry: Arc<dyn EmbeddingRegistry> = Arc::new(MemoryRegistry::new());
+            let mut scannable =
+                scannable_with_embeddings(Box::new(batch.clone()), &table_def, Some(&registry))
+                    .unwrap();
+
+            // Check that data passes through unchanged
+            let stream = scannable.scan_as_stream();
+            let results: Vec<RecordBatch> = stream.try_collect().await.unwrap();
+            assert_eq!(results.len(), 1);
+            assert_eq!(results[0], batch);
+        }
+
+        #[tokio::test]
+        async fn test_maybe_embedded_scannable_with_embeddings() {
+            let schema = Arc::new(Schema::new(vec![Field::new("text", DataType::Utf8, false)]));
+            let text_array = StringArray::from(vec!["hello", "world"]);
+            let batch =
+                RecordBatch::try_new(schema.clone(), vec![Arc::new(text_array) as ArrayRef])
+                    .unwrap();
+
+            // Create a table definition with an embedding column
+            let embedding_def = EmbeddingDefinition::new("text", "mock", Some("text_embedding"));
+            let embedding_schema = Arc::new(Schema::new(vec![
+                Field::new("text", DataType::Utf8, false),
+                Field::new(
+                    "text_embedding",
+                    DataType::FixedSizeList(
+                        Arc::new(Field::new("item", DataType::Float32, true)),
+                        4,
+                    ),
+                    false,
+                ),
+            ]));
+            let table_def = TableDefinition::new(
+                embedding_schema,
+                vec![
+                    ColumnDefinition {
+                        kind: ColumnKind::Physical,
+                    },
+                    ColumnDefinition {
+                        kind: ColumnKind::Embedding(embedding_def.clone()),
+                    },
+                ],
+            );
+
+            // Register the mock embedding function
+            let registry: Arc<dyn EmbeddingRegistry> = Arc::new(MemoryRegistry::new());
+            let mock_embedding: Arc<dyn EmbeddingFunction> = Arc::new(MockEmbed::new("mock", 4));
+            registry.register("mock", mock_embedding).unwrap();
+
+            let mut scannable =
+                scannable_with_embeddings(Box::new(batch), &table_def, Some(&registry)).unwrap();
+
+            // Read and verify the data has embeddings
+            let stream = scannable.scan_as_stream();
+            let results: Vec<RecordBatch> = stream.try_collect().await.unwrap();
+            assert_eq!(results.len(), 1);
+
+            let result_batch = &results[0];
+            assert_eq!(result_batch.num_columns(), 2);
+            assert_eq!(result_batch.schema().field(1).name(), "text_embedding");
+        }
+
+        #[tokio::test]
+        async fn test_maybe_embedded_scannable_missing_function() {
+            let schema = Arc::new(Schema::new(vec![Field::new("text", DataType::Utf8, false)]));
+            let text_array = StringArray::from(vec!["hello"]);
+            let batch =
+                RecordBatch::try_new(schema.clone(), vec![Arc::new(text_array) as ArrayRef])
+                    .unwrap();
+
+            // Create a table definition with an embedding column
+            let embedding_def =
+                EmbeddingDefinition::new("text", "nonexistent", Some("text_embedding"));
+            let embedding_schema = Arc::new(Schema::new(vec![
+                Field::new("text", DataType::Utf8, false),
+                Field::new(
+                    "text_embedding",
+                    DataType::FixedSizeList(
+                        Arc::new(Field::new("item", DataType::Float32, true)),
+                        4,
+                    ),
+                    false,
+                ),
+            ]));
+            let table_def = TableDefinition::new(
+                embedding_schema,
+                vec![
+                    ColumnDefinition {
+                        kind: ColumnKind::Physical,
+                    },
+                    ColumnDefinition {
+                        kind: ColumnKind::Embedding(embedding_def),
+                    },
+                ],
+            );
+
+            // Registry has no embedding functions registered
+            let registry: Arc<dyn EmbeddingRegistry> = Arc::new(MemoryRegistry::new());
+
+            let result = scannable_with_embeddings(Box::new(batch), &table_def, Some(&registry));
+
+            // Should fail because the embedding function is not found
+            assert!(result.is_err());
+            let err = result.err().unwrap();
+            assert!(
+                matches!(err, Error::EmbeddingFunctionNotFound { .. }),
+                "Expected EmbeddingFunctionNotFound"
+            );
+        }
+    }
+}

rust/lancedb/src/database.rs

@@ -18,12 +18,7 @@ use std::collections::HashMap;
 use std::sync::Arc;
 use std::time::Duration;
 
-use arrow_array::RecordBatchReader;
-use async_trait::async_trait;
-use datafusion_physical_plan::stream::RecordBatchStreamAdapter;
-use futures::stream;
 use lance::dataset::ReadParams;
-use lance_datafusion::utils::StreamingWriteSource;
 use lance_namespace::models::{
     CreateNamespaceRequest, CreateNamespaceResponse, DescribeNamespaceRequest,
     DescribeNamespaceResponse, DropNamespaceRequest, DropNamespaceResponse, ListNamespacesRequest,
@@ -31,9 +26,9 @@ use lance_namespace::models::{
 };
 use lance_namespace::LanceNamespace;
 
-use crate::arrow::{SendableRecordBatchStream, SendableRecordBatchStreamExt};
+use crate::data::scannable::Scannable;
 use crate::error::Result;
-use crate::table::{BaseTable, TableDefinition, WriteOptions};
+use crate::table::{BaseTable, WriteOptions};
 
 pub mod listing;
 pub mod namespace;
@@ -90,8 +85,10 @@ pub type TableBuilderCallback = Box<dyn FnOnce(OpenTableRequest) -> OpenTableReq
 
 /// Describes what happens when creating a table and a table with
 /// the same name already exists
+#[derive(Default)]
 pub enum CreateTableMode {
     /// If the table already exists, an error is returned
+    #[default]
     Create,
     /// If the table already exists, it is opened.  Any provided data is
     /// ignored.  The function will be passed an OpenTableBuilder to customize
@@ -109,57 +106,14 @@ impl CreateTableMode {
     }
 }
 
-impl Default for CreateTableMode {
-    fn default() -> Self {
-        Self::Create
-    }
-}
-
-/// The data to start a table or a schema to create an empty table
-pub enum CreateTableData {
-    /// Creates a table using an iterator of data, the schema will be obtained from the data
-    Data(Box<dyn RecordBatchReader + Send>),
-    /// Creates a table using a stream of data, the schema will be obtained from the data
-    StreamingData(SendableRecordBatchStream),
-    /// Creates an empty table, the definition / schema must be provided separately
-    Empty(TableDefinition),
-}
-
-impl CreateTableData {
-    pub fn schema(&self) -> Arc<arrow_schema::Schema> {
-        match self {
-            Self::Data(reader) => reader.schema(),
-            Self::StreamingData(stream) => stream.schema(),
-            Self::Empty(definition) => definition.schema.clone(),
-        }
-    }
-}
-
-#[async_trait]
-impl StreamingWriteSource for CreateTableData {
-    fn arrow_schema(&self) -> Arc<arrow_schema::Schema> {
-        self.schema()
-    }
-    fn into_stream(self) -> datafusion_physical_plan::SendableRecordBatchStream {
-        match self {
-            Self::Data(reader) => reader.into_stream(),
-            Self::StreamingData(stream) => stream.into_df_stream(),
-            Self::Empty(table_definition) => {
-                let schema = table_definition.schema.clone();
-                Box::pin(RecordBatchStreamAdapter::new(schema, stream::empty()))
-            }
-        }
-    }
-}
-
 /// A request to create a table
 pub struct CreateTableRequest {
     /// The name of the new table
     pub name: String,
     /// The namespace to create the table in. Empty list represents root namespace.
     pub namespace: Vec<String>,
-    /// Initial data to write to the table, can be None to create an empty table
-    pub data: CreateTableData,
+    /// Initial data to write to the table, can be empty.
+    pub data: Box<dyn Scannable>,
     /// The mode to use when creating the table
     pub mode: CreateTableMode,
     /// Options to use when writing data (only used if `data` is not None)
@@ -173,7 +127,7 @@ pub struct CreateTableRequest {
 }
 
 impl CreateTableRequest {
-    pub fn new(name: String, data: CreateTableData) -> Self {
+    pub fn new(name: String, data: Box<dyn Scannable>) -> Self {
         Self {
             name,
             namespace: vec![],

rust/lancedb/src/database/listing.rs

@@ -922,7 +922,7 @@ impl Database for ListingDatabase {
             .with_read_params(read_params.clone())
             .load()
             .await
-            .map_err(|e| Error::Lance { source: e })?;
+            .map_err(|e| -> Error { e.into() })?;
 
         let version_ref = match (request.source_version, request.source_tag) {
             (Some(v), None) => Ok(Ref::Version(None, Some(v))),
@@ -937,7 +937,7 @@ impl Database for ListingDatabase {
         source_dataset
             .shallow_clone(&target_uri, version_ref, Some(storage_params))
             .await
-            .map_err(|e| Error::Lance { source: e })?;
+            .map_err(|e| -> Error { e.into() })?;
 
         let cloned_table = NativeTable::open_with_params(
             &target_uri,
@@ -1098,8 +1098,10 @@ impl Database for ListingDatabase {
 mod tests {
     use super::*;
     use crate::connection::ConnectRequest;
-    use crate::database::{CreateTableData, CreateTableMode, CreateTableRequest, WriteOptions};
-    use crate::table::{Table, TableDefinition};
+    use crate::data::scannable::Scannable;
+    use crate::database::{CreateTableMode, CreateTableRequest};
+    use crate::table::WriteOptions;
+    use crate::Table;
     use arrow_array::{Int32Array, RecordBatch, StringArray};
     use arrow_schema::{DataType, Field, Schema};
     use std::path::PathBuf;
@@ -1139,7 +1141,7 @@ mod tests {
             .create_table(CreateTableRequest {
                 name: "source_table".to_string(),
                 namespace: vec![],
-                data: CreateTableData::Empty(TableDefinition::new_from_schema(schema.clone())),
+                data: Box::new(RecordBatch::new_empty(schema.clone())) as Box<dyn Scannable>,
                 mode: CreateTableMode::Create,
                 write_options: Default::default(),
                 location: None,
@@ -1196,16 +1198,11 @@ mod tests {
         )
         .unwrap();
 
-        let reader = Box::new(arrow_array::RecordBatchIterator::new(
-            vec![Ok(batch)],
-            schema.clone(),
-        ));
-
         let source_table = db
             .create_table(CreateTableRequest {
                 name: "source_with_data".to_string(),
                 namespace: vec![],
-                data: CreateTableData::Data(reader),
+                data: Box::new(batch) as Box<dyn Scannable>,
                 mode: CreateTableMode::Create,
                 write_options: Default::default(),
                 location: None,
@@ -1264,7 +1261,7 @@ mod tests {
         db.create_table(CreateTableRequest {
             name: "source".to_string(),
             namespace: vec![],
-            data: CreateTableData::Empty(TableDefinition::new_from_schema(schema)),
+            data: Box::new(RecordBatch::new_empty(schema)) as Box<dyn Scannable>,
             mode: CreateTableMode::Create,
             write_options: Default::default(),
             location: None,
@@ -1300,7 +1297,7 @@ mod tests {
         db.create_table(CreateTableRequest {
             name: "source".to_string(),
             namespace: vec![],
-            data: CreateTableData::Empty(TableDefinition::new_from_schema(schema)),
+            data: Box::new(RecordBatch::new_empty(schema)) as Box<dyn Scannable>,
             mode: CreateTableMode::Create,
             write_options: Default::default(),
             location: None,
@@ -1340,7 +1337,7 @@ mod tests {
         db.create_table(CreateTableRequest {
             name: "source".to_string(),
             namespace: vec![],
-            data: CreateTableData::Empty(TableDefinition::new_from_schema(schema)),
+            data: Box::new(RecordBatch::new_empty(schema)) as Box<dyn Scannable>,
             mode: CreateTableMode::Create,
             write_options: Default::default(),
             location: None,
@@ -1380,7 +1377,7 @@ mod tests {
         db.create_table(CreateTableRequest {
             name: "source".to_string(),
             namespace: vec![],
-            data: CreateTableData::Empty(TableDefinition::new_from_schema(schema)),
+            data: Box::new(RecordBatch::new_empty(schema)) as Box<dyn Scannable>,
             mode: CreateTableMode::Create,
             write_options: Default::default(),
             location: None,
@@ -1435,7 +1432,7 @@ mod tests {
         db.create_table(CreateTableRequest {
             name: "source".to_string(),
             namespace: vec![],
-            data: CreateTableData::Empty(TableDefinition::new_from_schema(schema)),
+            data: Box::new(RecordBatch::new_empty(schema)) as Box<dyn Scannable>,
             mode: CreateTableMode::Create,
             write_options: Default::default(),
             location: None,
@@ -1484,16 +1481,11 @@ mod tests {
         )
         .unwrap();
 
-        let reader = Box::new(arrow_array::RecordBatchIterator::new(
-            vec![Ok(batch1)],
-            schema.clone(),
-        ));
-
         let source_table = db
             .create_table(CreateTableRequest {
                 name: "versioned_source".to_string(),
                 namespace: vec![],
-                data: CreateTableData::Data(reader),
+                data: Box::new(batch1) as Box<dyn Scannable>,
                 mode: CreateTableMode::Create,
                 write_options: Default::default(),
                 location: None,
@@ -1517,14 +1509,7 @@ mod tests {
 
         let db = Arc::new(db);
         let source_table_obj = Table::new(source_table.clone(), db.clone());
-        source_table_obj
-            .add(Box::new(arrow_array::RecordBatchIterator::new(
-                vec![Ok(batch2)],
-                schema.clone(),
-            )))
-            .execute()
-            .await
-            .unwrap();
+        source_table_obj.add(batch2).execute().await.unwrap();
 
         // Verify source table now has 4 rows
         assert_eq!(source_table.count_rows(None).await.unwrap(), 4);
@@ -1570,16 +1555,11 @@ mod tests {
         )
         .unwrap();
 
-        let reader = Box::new(arrow_array::RecordBatchIterator::new(
-            vec![Ok(batch1)],
-            schema.clone(),
-        ));
-
         let source_table = db
             .create_table(CreateTableRequest {
                 name: "tagged_source".to_string(),
                 namespace: vec![],
-                data: CreateTableData::Data(reader),
+                data: Box::new(batch1),
                 mode: CreateTableMode::Create,
                 write_options: Default::default(),
                 location: None,
@@ -1607,14 +1587,7 @@ mod tests {
         .unwrap();
 
         let source_table_obj = Table::new(source_table.clone(), db.clone());
-        source_table_obj
-            .add(Box::new(arrow_array::RecordBatchIterator::new(
-                vec![Ok(batch2)],
-                schema.clone(),
-            )))
-            .execute()
-            .await
-            .unwrap();
+        source_table_obj.add(batch2).execute().await.unwrap();
 
         // Source table should have 4 rows
         assert_eq!(source_table.count_rows(None).await.unwrap(), 4);
@@ -1657,16 +1630,11 @@ mod tests {
         )
         .unwrap();
 
-        let reader = Box::new(arrow_array::RecordBatchIterator::new(
-            vec![Ok(batch1)],
-            schema.clone(),
-        ));
-
         let source_table = db
             .create_table(CreateTableRequest {
                 name: "independent_source".to_string(),
                 namespace: vec![],
-                data: CreateTableData::Data(reader),
+                data: Box::new(batch1),
                 mode: CreateTableMode::Create,
                 write_options: Default::default(),
                 location: None,
@@ -1706,14 +1674,7 @@ mod tests {
 
         let db = Arc::new(db);
         let cloned_table_obj = Table::new(cloned_table.clone(), db.clone());
-        cloned_table_obj
-            .add(Box::new(arrow_array::RecordBatchIterator::new(
-                vec![Ok(batch_clone)],
-                schema.clone(),
-            )))
-            .execute()
-            .await
-            .unwrap();
+        cloned_table_obj.add(batch_clone).execute().await.unwrap();
 
         // Add different data to the source table
         let batch_source = RecordBatch::try_new(
@@ -1726,14 +1687,7 @@ mod tests {
         .unwrap();
 
         let source_table_obj = Table::new(source_table.clone(), db);
-        source_table_obj
-            .add(Box::new(arrow_array::RecordBatchIterator::new(
-                vec![Ok(batch_source)],
-                schema.clone(),
-            )))
-            .execute()
-            .await
-            .unwrap();
+        source_table_obj.add(batch_source).execute().await.unwrap();
 
         // Verify they have evolved independently
         assert_eq!(source_table.count_rows(None).await.unwrap(), 4); // 2 + 2
@@ -1751,16 +1705,11 @@ mod tests {
             RecordBatch::try_new(schema.clone(), vec![Arc::new(Int32Array::from(vec![1, 2]))])
                 .unwrap();
 
-        let reader = Box::new(arrow_array::RecordBatchIterator::new(
-            vec![Ok(batch1)],
-            schema.clone(),
-        ));
-
         let source_table = db
             .create_table(CreateTableRequest {
                 name: "latest_version_source".to_string(),
                 namespace: vec![],
-                data: CreateTableData::Data(reader),
+                data: Box::new(batch1),
                 mode: CreateTableMode::Create,
                 write_options: Default::default(),
                 location: None,
@@ -1779,14 +1728,7 @@ mod tests {
             .unwrap();
 
             let source_table_obj = Table::new(source_table.clone(), db.clone());
-            source_table_obj
-                .add(Box::new(arrow_array::RecordBatchIterator::new(
-                    vec![Ok(batch)],
-                    schema.clone(),
-                )))
-                .execute()
-                .await
-                .unwrap();
+            source_table_obj.add(batch).execute().await.unwrap();
         }
 
         // Source should have 8 rows total (2 + 2 + 2 + 2)
@@ -1849,16 +1791,11 @@ mod tests {
         )
         .unwrap();
 
-        let reader = Box::new(arrow_array::RecordBatchIterator::new(
-            vec![Ok(batch)],
-            schema.clone(),
-        ));
-
         let table = db
             .create_table(CreateTableRequest {
                 name: "test_stable".to_string(),
                 namespace: vec![],
-                data: CreateTableData::Data(reader),
+                data: Box::new(batch),
                 mode: CreateTableMode::Create,
                 write_options: Default::default(),
                 location: None,
@@ -1887,11 +1824,6 @@ mod tests {
         )
         .unwrap();
 
-        let reader = Box::new(arrow_array::RecordBatchIterator::new(
-            vec![Ok(batch)],
-            schema.clone(),
-        ));
-
         let mut storage_options = HashMap::new();
         storage_options.insert(
             OPT_NEW_TABLE_ENABLE_STABLE_ROW_IDS.to_string(),
@@ -1914,7 +1846,7 @@ mod tests {
             .create_table(CreateTableRequest {
                 name: "test_stable_table_level".to_string(),
                 namespace: vec![],
-                data: CreateTableData::Data(reader),
+                data: Box::new(batch),
                 mode: CreateTableMode::Create,
                 write_options,
                 location: None,
@@ -1963,11 +1895,6 @@ mod tests {
         )
         .unwrap();
 
-        let reader = Box::new(arrow_array::RecordBatchIterator::new(
-            vec![Ok(batch)],
-            schema.clone(),
-        ));
-
         let mut storage_options = HashMap::new();
         storage_options.insert(
             OPT_NEW_TABLE_ENABLE_STABLE_ROW_IDS.to_string(),
@@ -1990,7 +1917,7 @@ mod tests {
             .create_table(CreateTableRequest {
                 name: "test_override".to_string(),
                 namespace: vec![],
-                data: CreateTableData::Data(reader),
+                data: Box::new(batch),
                 mode: CreateTableMode::Create,
                 write_options,
                 location: None,
@@ -2108,7 +2035,7 @@ mod tests {
         db.create_table(CreateTableRequest {
             name: "table1".to_string(),
             namespace: vec![],
-            data: CreateTableData::Empty(TableDefinition::new_from_schema(schema.clone())),
+            data: Box::new(RecordBatch::new_empty(schema.clone())) as Box<dyn Scannable>,
             mode: CreateTableMode::Create,
             write_options: Default::default(),
             location: None,
@@ -2120,7 +2047,7 @@ mod tests {
         db.create_table(CreateTableRequest {
             name: "table2".to_string(),
             namespace: vec![],
-            data: CreateTableData::Empty(TableDefinition::new_from_schema(schema)),
+            data: Box::new(RecordBatch::new_empty(schema)) as Box<dyn Scannable>,
             mode: CreateTableMode::Create,
             write_options: Default::default(),
             location: None,

rust/lancedb/src/database/namespace.rs

@@ -7,6 +7,7 @@ use std::collections::HashMap;
 use std::sync::Arc;
 
 use async_trait::async_trait;
+use lance_io::object_store::{ObjectStoreParams, StorageOptionsAccessor};
 use lance_namespace::{
     models::{
         CreateEmptyTableRequest, CreateNamespaceRequest, CreateNamespaceResponse,
@@ -212,45 +213,75 @@ impl Database for LanceNamespaceDatabase {
             ..Default::default()
         };
 
-        let location = match self.namespace.declare_table(declare_request).await {
-            Ok(response) => response.location.ok_or_else(|| Error::Runtime {
-                message: "Table location is missing from declare_table response".to_string(),
-            })?,
-            Err(e) => {
-                // Check if the error is "not supported" and try create_empty_table as fallback
-                let err_str = e.to_string().to_lowercase();
-                if err_str.contains("not supported") || err_str.contains("not implemented") {
-                    warn!(
-                        "declare_table is not supported by the namespace client, \
+        let (location, initial_storage_options) =
+            match self.namespace.declare_table(declare_request).await {
+                Ok(response) => {
+                    let loc = response.location.ok_or_else(|| Error::Runtime {
+                        message: "Table location is missing from declare_table response"
+                            .to_string(),
+                    })?;
+                    // Use storage options from response, fall back to self.storage_options
+                    let opts = response
+                        .storage_options
+                        .or_else(|| Some(self.storage_options.clone()))
+                        .filter(|o| !o.is_empty());
+                    (loc, opts)
+                }
+                Err(e) => {
+                    // Check if the error is "not supported" and try create_empty_table as fallback
+                    let err_str = e.to_string().to_lowercase();
+                    if err_str.contains("not supported") || err_str.contains("not implemented") {
+                        warn!(
+                            "declare_table is not supported by the namespace client, \
                         falling back to deprecated create_empty_table. \
                         create_empty_table is deprecated and will be removed in Lance 3.0.0. \
                         Please upgrade your namespace client to support declare_table."
-                    );
-                    #[allow(deprecated)]
-                    let create_empty_request = CreateEmptyTableRequest {
-                        id: Some(table_id.clone()),
-                        ..Default::default()
-                    };
+                        );
+                        #[allow(deprecated)]
+                        let create_empty_request = CreateEmptyTableRequest {
+                            id: Some(table_id.clone()),
+                            ..Default::default()
+                        };
 
-                    #[allow(deprecated)]
-                    let create_response = self
-                        .namespace
-                        .create_empty_table(create_empty_request)
-                        .await
-                        .map_err(|e| Error::Runtime {
-                            message: format!("Failed to create empty table: {}", e),
+                        #[allow(deprecated)]
+                        let create_response = self
+                            .namespace
+                            .create_empty_table(create_empty_request)
+                            .await
+                            .map_err(|e| Error::Runtime {
+                                message: format!("Failed to create empty table: {}", e),
+                            })?;
+
+                        let loc = create_response.location.ok_or_else(|| Error::Runtime {
+                            message: "Table location is missing from create_empty_table response"
+                                .to_string(),
                         })?;
-
-                    create_response.location.ok_or_else(|| Error::Runtime {
-                        message: "Table location is missing from create_empty_table response"
-                            .to_string(),
-                    })?
-                } else {
-                    return Err(Error::Runtime {
-                        message: format!("Failed to declare table: {}", e),
-                    });
+                        // For deprecated path, use self.storage_options
+                        let opts = if self.storage_options.is_empty() {
+                            None
+                        } else {
+                            Some(self.storage_options.clone())
+                        };
+                        (loc, opts)
+                    } else {
+                        return Err(Error::Runtime {
+                            message: format!("Failed to declare table: {}", e),
+                        });
+                    }
                 }
-            }
+            };
+
+        let write_params = if let Some(storage_opts) = initial_storage_options {
+            let mut params = request.write_options.lance_write_params.unwrap_or_default();
+            let store_params = params
+                .store_params
+                .get_or_insert_with(ObjectStoreParams::default);
+            store_params.storage_options_accessor = Some(Arc::new(
+                StorageOptionsAccessor::with_static_options(storage_opts),
+            ));
+            Some(params)
+        } else {
+            request.write_options.lance_write_params
         };
 
         let native_table = NativeTable::create_from_namespace(
@@ -260,7 +291,7 @@ impl Database for LanceNamespaceDatabase {
             request.namespace.clone(),
             request.data,
             None, // write_store_wrapper not used for namespace connections
-            request.write_options.lance_write_params,
+            write_params,
             self.read_consistency_interval,
             self.server_side_query_enabled,
             self.session.clone(),
@@ -354,15 +385,13 @@ mod tests {
     use super::*;
     use crate::connect_namespace;
     use crate::query::ExecutableQuery;
-    use arrow_array::{Int32Array, RecordBatch, RecordBatchIterator, StringArray};
+    use arrow_array::{Int32Array, RecordBatch, StringArray};
     use arrow_schema::{DataType, Field, Schema};
     use futures::TryStreamExt;
     use tempfile::tempdir;
 
     /// Helper function to create test data
-    fn create_test_data() -> RecordBatchIterator<
-        std::vec::IntoIter<std::result::Result<RecordBatch, arrow_schema::ArrowError>>,
-    > {
+    fn create_test_data() -> RecordBatch {
         let schema = Arc::new(Schema::new(vec![
             Field::new("id", DataType::Int32, false),
             Field::new("name", DataType::Utf8, false),
@@ -371,12 +400,7 @@ mod tests {
         let id_array = Int32Array::from(vec![1, 2, 3, 4, 5]);
         let name_array = StringArray::from(vec!["Alice", "Bob", "Charlie", "David", "Eve"]);
 
-        let batch = RecordBatch::try_new(
-            schema.clone(),
-            vec![Arc::new(id_array), Arc::new(name_array)],
-        )
-        .unwrap();
-        RecordBatchIterator::new(vec![std::result::Result::Ok(batch)].into_iter(), schema)
+        RecordBatch::try_new(schema, vec![Arc::new(id_array), Arc::new(name_array)]).unwrap()
     }
 
     #[tokio::test]
@@ -618,13 +642,7 @@ mod tests {
 
         // Test: Overwrite the table
         let table2 = conn
-            .create_table(
-                "overwrite_test",
-                RecordBatchIterator::new(
-                    vec![std::result::Result::Ok(test_data2)].into_iter(),
-                    schema,
-                ),
-            )
+            .create_table("overwrite_test", test_data2)
             .namespace(vec!["test_ns".into()])
             .mode(CreateTableMode::Overwrite)
             .execute()

rust/lancedb/src/dataloader/permutation/builder.rs

@@ -13,7 +13,7 @@ use lance_datafusion::exec::SessionContextExt;
 use crate::{
     arrow::{SendableRecordBatchStream, SendableRecordBatchStreamExt, SimpleRecordBatchStream},
     connect,
-    database::{CreateTableData, CreateTableRequest, Database},
+    database::{CreateTableRequest, Database},
     dataloader::permutation::{
         shuffle::{Shuffler, ShufflerConfig},
         split::{SplitStrategy, Splitter, SPLIT_ID_COLUMN},
@@ -57,7 +57,7 @@ pub struct PermutationConfig {
 }
 
 /// Strategy for shuffling the data.
-#[derive(Debug, Clone)]
+#[derive(Debug, Clone, Default)]
 pub enum ShuffleStrategy {
     /// The data is randomly shuffled
     ///
@@ -78,15 +78,10 @@ pub enum ShuffleStrategy {
     /// The data is not shuffled
     ///
     /// This is useful for debugging and testing.
+    #[default]
     None,
 }
 
-impl Default for ShuffleStrategy {
-    fn default() -> Self {
-        Self::None
-    }
-}
-
 /// Builder for creating a permutation table.
 ///
 /// A permutation table is a table that stores split assignments and a shuffled order of rows.  This
@@ -313,10 +308,8 @@ impl PermutationBuilder {
             }
         };
 
-        let create_table_request = CreateTableRequest::new(
-            name.to_string(),
-            CreateTableData::StreamingData(streaming_data),
-        );
+        let create_table_request =
+            CreateTableRequest::new(name.to_string(), Box::new(streaming_data));
 
         let table = database.create_table(create_table_request).await?;
 
@@ -347,7 +340,7 @@ mod tests {
             .col("col_b", lance_datagen::array::step::<Int32Type>())
             .into_ldb_stream(RowCount::from(100), BatchCount::from(10));
         let data_table = db
-            .create_table_streaming("base_tbl", initial_data)
+            .create_table("base_tbl", initial_data)
             .execute()
             .await
             .unwrap();
@@ -387,7 +380,7 @@ mod tests {
             .col("some_value", lance_datagen::array::step::<Int32Type>())
             .into_ldb_stream(RowCount::from(100), BatchCount::from(10));
         let data_table = db
-            .create_table_streaming("mytbl", initial_data)
+            .create_table("mytbl", initial_data)
             .execute()
             .await
             .unwrap();

rust/lancedb/src/dataloader/permutation/reader.rs

@@ -426,6 +426,7 @@ impl PermutationReader {
             row_ids_query = row_ids_query.limit(limit as usize);
         }
         let mut row_ids = row_ids_query.execute().await?;
+        let mut idx_offset = 0;
         while let Some(batch) = row_ids.try_next().await? {
             let row_ids = batch
                 .column(0)
@@ -433,8 +434,9 @@ impl PermutationReader {
                 .values()
                 .to_vec();
             for (i, row_id) in row_ids.iter().enumerate() {
-                offset_map.insert(i as u64, *row_id);
+                offset_map.insert(i as u64 + idx_offset, *row_id);
             }
+            idx_offset += batch.num_rows() as u64;
         }
         let offset_map = Arc::new(offset_map);
         *offset_map_ref = Some(offset_map.clone());
@@ -845,4 +847,106 @@ mod tests {
             .to_vec();
         assert_eq!(idx_values, vec![row_ids[2] as i32]);
     }
+
+    #[tokio::test]
+    async fn test_filtered_permutation_full_iteration() {
+        use crate::dataloader::permutation::builder::PermutationBuilder;
+
+        // Create a base table with 10000 rows where idx goes 0..10000.
+        // Filter to even values only, giving 5000 rows in the permutation.
+        let base_table = lance_datagen::gen_batch()
+            .col("idx", lance_datagen::array::step::<Int32Type>())
+            .into_mem_table("tbl", RowCount::from(10000), BatchCount::from(1))
+            .await;
+
+        let permutation_table = PermutationBuilder::new(base_table.clone())
+            .with_filter("idx % 2 = 0".to_string())
+            .build()
+            .await
+            .unwrap();
+
+        assert_eq!(permutation_table.count_rows(None).await.unwrap(), 5000);
+
+        let reader = PermutationReader::try_from_tables(
+            base_table.base_table().clone(),
+            permutation_table.base_table().clone(),
+            0,
+        )
+        .await
+        .unwrap();
+
+        assert_eq!(reader.count_rows(), 5000);
+
+        // Iterate through all batches using a batch size that doesn't evenly divide
+        // the row count (5000 / 128 = 39 full batches + 1 batch of 8 rows).
+        let batch_size = 128;
+        let mut stream = reader
+            .read(
+                Select::All,
+                QueryExecutionOptions {
+                    max_batch_length: batch_size,
+                    ..Default::default()
+                },
+            )
+            .await
+            .unwrap();
+
+        let mut total_rows = 0u64;
+        let mut all_idx_values = Vec::new();
+        while let Some(batch) = stream.try_next().await.unwrap() {
+            assert!(batch.num_rows() <= batch_size as usize);
+            total_rows += batch.num_rows() as u64;
+            let idx_col = batch.column(0).as_primitive::<Int32Type>().values();
+            all_idx_values.extend(idx_col.iter().copied());
+        }
+
+        assert_eq!(total_rows, 5000);
+        assert_eq!(all_idx_values.len(), 5000);
+
+        // Every value should be even (from the filter)
+        assert!(all_idx_values.iter().all(|v| v % 2 == 0));
+
+        // Should have 5000 unique values
+        let unique: std::collections::HashSet<i32> = all_idx_values.iter().copied().collect();
+        assert_eq!(unique.len(), 5000);
+
+        // Use take_offsets to fetch rows from the beginning, middle, and end
+        // of the permutation. The values should match what we saw during iteration.
+
+        // Beginning
+        let batch = reader.take_offsets(&[0, 1, 2], Select::All).await.unwrap();
+        assert_eq!(batch.num_rows(), 3);
+        let idx_values = batch
+            .column(0)
+            .as_primitive::<Int32Type>()
+            .values()
+            .to_vec();
+        assert_eq!(idx_values, &all_idx_values[0..3]);
+
+        // Middle
+        let batch = reader
+            .take_offsets(&[2499, 2500, 2501], Select::All)
+            .await
+            .unwrap();
+        assert_eq!(batch.num_rows(), 3);
+        let idx_values = batch
+            .column(0)
+            .as_primitive::<Int32Type>()
+            .values()
+            .to_vec();
+        assert_eq!(idx_values, &all_idx_values[2499..2502]);
+
+        // End (last 3 rows)
+        let batch = reader
+            .take_offsets(&[4997, 4998, 4999], Select::All)
+            .await
+            .unwrap();
+        assert_eq!(batch.num_rows(), 3);
+        let idx_values = batch
+            .column(0)
+            .as_primitive::<Int32Type>()
+            .values()
+            .to_vec();
+        assert_eq!(idx_values, &all_idx_values[4997..5000]);
+    }
 }

rust/lancedb/src/dataloader/permutation/split.rs

@@ -27,9 +27,10 @@ use crate::{
 pub const SPLIT_ID_COLUMN: &str = "split_id";
 
 /// Strategy for assigning rows to splits
-#[derive(Debug, Clone)]
+#[derive(Debug, Clone, Default)]
 pub enum SplitStrategy {
     /// All rows will have split id 0
+    #[default]
     NoSplit,
     /// Rows will be randomly assigned to splits
     ///
@@ -73,15 +74,6 @@ pub enum SplitStrategy {
     Calculated { calculation: String },
 }
 
-// The default is not to split the data
-//
-// All data will be assigned to a single split.
-impl Default for SplitStrategy {
-    fn default() -> Self {
-        Self::NoSplit
-    }
-}
-
 impl SplitStrategy {
     pub fn validate(&self, num_rows: u64) -> Result<()> {
         match self {

rust/lancedb/src/embeddings.rs

@@ -18,7 +18,7 @@ use std::{
 };
 
 use arrow_array::{Array, RecordBatch, RecordBatchReader};
-use arrow_schema::{DataType, Field, SchemaBuilder};
+use arrow_schema::{DataType, Field, SchemaBuilder, SchemaRef};
 // use async_trait::async_trait;
 use serde::{Deserialize, Serialize};
 
@@ -190,6 +190,112 @@ impl<R: RecordBatchReader> WithEmbeddings<R> {
     }
 }
 
+/// Compute embedding arrays for a batch.
+///
+/// When multiple embedding functions are defined, they are computed in parallel using
+/// scoped threads. For a single embedding function, computation is done inline.
+fn compute_embedding_arrays(
+    batch: &RecordBatch,
+    embeddings: &[(EmbeddingDefinition, Arc<dyn EmbeddingFunction>)],
+) -> Result<Vec<Arc<dyn Array>>> {
+    if embeddings.len() == 1 {
+        let (fld, func) = &embeddings[0];
+        let src_column =
+            batch
+                .column_by_name(&fld.source_column)
+                .ok_or_else(|| Error::InvalidInput {
+                    message: format!("Source column '{}' not found", fld.source_column),
+                })?;
+        return Ok(vec![func.compute_source_embeddings(src_column.clone())?]);
+    }
+
+    // Parallel path: multiple embeddings
+    std::thread::scope(|s| {
+        let handles: Vec<_> = embeddings
+            .iter()
+            .map(|(fld, func)| {
+                let src_column = batch.column_by_name(&fld.source_column).ok_or_else(|| {
+                    Error::InvalidInput {
+                        message: format!("Source column '{}' not found", fld.source_column),
+                    }
+                })?;
+
+                let handle = s.spawn(move || func.compute_source_embeddings(src_column.clone()));
+
+                Ok(handle)
+            })
+            .collect::<Result<_>>()?;
+
+        handles
+            .into_iter()
+            .map(|h| {
+                h.join().map_err(|e| Error::Runtime {
+                    message: format!("Thread panicked during embedding computation: {:?}", e),
+                })?
+            })
+            .collect()
+    })
+}
+
+/// Compute the output schema when embeddings are applied to a base schema.
+///
+/// This returns the schema with embedding columns appended.
+pub fn compute_output_schema(
+    base_schema: &SchemaRef,
+    embeddings: &[(EmbeddingDefinition, Arc<dyn EmbeddingFunction>)],
+) -> Result<SchemaRef> {
+    let mut sb: SchemaBuilder = base_schema.as_ref().into();
+
+    for (ed, func) in embeddings {
+        let src_field = base_schema
+            .field_with_name(&ed.source_column)
+            .map_err(|_| Error::InvalidInput {
+                message: format!("Source column '{}' not found in schema", ed.source_column),
+            })?;
+
+        let field_name = ed
+            .dest_column
+            .clone()
+            .unwrap_or_else(|| format!("{}_embedding", &ed.source_column));
+
+        sb.push(Field::new(
+            field_name,
+            func.dest_type()?.into_owned(),
+            src_field.is_nullable(),
+        ));
+    }
+
+    Ok(Arc::new(sb.finish()))
+}
+
+/// Compute embeddings for a batch and append as new columns.
+///
+/// This function computes embeddings using the provided embedding functions and
+/// appends them as new columns to the batch.
+pub fn compute_embeddings_for_batch(
+    batch: RecordBatch,
+    embeddings: &[(EmbeddingDefinition, Arc<dyn EmbeddingFunction>)],
+) -> Result<RecordBatch> {
+    let embedding_arrays = compute_embedding_arrays(&batch, embeddings)?;
+
+    let mut result = batch;
+    for ((fld, _), embedding) in embeddings.iter().zip(embedding_arrays.iter()) {
+        let dst_field_name = fld
+            .dest_column
+            .clone()
+            .unwrap_or_else(|| format!("{}_embedding", &fld.source_column));
+
+        let dst_field = Field::new(
+            dst_field_name,
+            embedding.data_type().clone(),
+            embedding.nulls().is_some(),
+        );
+
+        result = result.try_with_column(dst_field, embedding.clone())?;
+    }
+    Ok(result)
+}
+
 impl<R: RecordBatchReader> WithEmbeddings<R> {
     fn dest_fields(&self) -> Result<Vec<Field>> {
         let schema = self.inner.schema();
@@ -240,48 +346,6 @@ impl<R: RecordBatchReader> WithEmbeddings<R> {
             column_definitions,
         })
     }
-
-    fn compute_embeddings_parallel(&self, batch: &RecordBatch) -> Result<Vec<Arc<dyn Array>>> {
-        if self.embeddings.len() == 1 {
-            let (fld, func) = &self.embeddings[0];
-            let src_column =
-                batch
-                    .column_by_name(&fld.source_column)
-                    .ok_or_else(|| Error::InvalidInput {
-                        message: format!("Source column '{}' not found", fld.source_column),
-                    })?;
-            return Ok(vec![func.compute_source_embeddings(src_column.clone())?]);
-        }
-
-        // Parallel path: multiple embeddings
-        std::thread::scope(|s| {
-            let handles: Vec<_> = self
-                .embeddings
-                .iter()
-                .map(|(fld, func)| {
-                    let src_column = batch.column_by_name(&fld.source_column).ok_or_else(|| {
-                        Error::InvalidInput {
-                            message: format!("Source column '{}' not found", fld.source_column),
-                        }
-                    })?;
-
-                    let handle =
-                        s.spawn(move || func.compute_source_embeddings(src_column.clone()));
-
-                    Ok(handle)
-                })
-                .collect::<Result<_>>()?;
-
-            handles
-                .into_iter()
-                .map(|h| {
-                    h.join().map_err(|e| Error::Runtime {
-                        message: format!("Thread panicked during embedding computation: {:?}", e),
-                    })?
-                })
-                .collect()
-        })
-    }
 }
 
 impl<R: RecordBatchReader> Iterator for MaybeEmbedded<R> {
@@ -309,37 +373,13 @@ impl<R: RecordBatchReader> Iterator for WithEmbeddings<R> {
     fn next(&mut self) -> Option<Self::Item> {
         let batch = self.inner.next()?;
         match batch {
-            Ok(batch) => {
-                let embeddings = match self.compute_embeddings_parallel(&batch) {
-                    Ok(emb) => emb,
-                    Err(e) => {
-                        return Some(Err(arrow_schema::ArrowError::ComputeError(format!(
-                            "Error computing embedding: {}",
-                            e
-                        ))))
-                    }
-                };
-
-                let mut batch = batch;
-                for ((fld, _), embedding) in self.embeddings.iter().zip(embeddings.iter()) {
-                    let dst_field_name = fld
-                        .dest_column
-                        .clone()
-                        .unwrap_or_else(|| format!("{}_embedding", &fld.source_column));
-
-                    let dst_field = Field::new(
-                        dst_field_name,
-                        embedding.data_type().clone(),
-                        embedding.nulls().is_some(),
-                    );
-
-                    match batch.try_with_column(dst_field.clone(), embedding.clone()) {
-                        Ok(b) => batch = b,
-                        Err(e) => return Some(Err(e)),
-                    };
-                }
-                Some(Ok(batch))
-            }
+            Ok(batch) => match compute_embeddings_for_batch(batch, &self.embeddings) {
+                Ok(batch_with_embeddings) => Some(Ok(batch_with_embeddings)),
+                Err(e) => Some(Err(arrow_schema::ArrowError::ComputeError(format!(
+                    "Error computing embedding: {}",
+                    e
+                )))),
+            },
             Err(e) => Some(Err(e)),
         }
     }

rust/lancedb/src/error.rs

@@ -4,9 +4,10 @@
 use std::sync::PoisonError;
 
 use arrow_schema::ArrowError;
+use datafusion_common::DataFusionError;
 use snafu::Snafu;
 
-type BoxError = Box<dyn std::error::Error + Send + Sync>;
+pub(crate) type BoxError = Box<dyn std::error::Error + Send + Sync>;
 
 #[derive(Debug, Snafu)]
 #[snafu(visibility(pub(crate)))]
@@ -80,6 +81,9 @@ pub enum Error {
     Arrow { source: ArrowError },
     #[snafu(display("LanceDBError: not supported: {message}"))]
     NotSupported { message: String },
+    /// External error pass through from user code.
+    #[snafu(transparent)]
+    External { source: BoxError },
     #[snafu(whatever, display("{message}"))]
     Other {
         message: String,
@@ -92,15 +96,72 @@ pub type Result<T> = std::result::Result<T, Error>;
 
 impl From<ArrowError> for Error {
     fn from(source: ArrowError) -> Self {
-        Self::Arrow { source }
+        match source {
+            ArrowError::ExternalError(source) => Self::from_box_error(source),
+            _ => Self::Arrow { source },
+        }
+    }
+}
+
+impl From<DataFusionError> for Error {
+    fn from(source: DataFusionError) -> Self {
+        match source {
+            DataFusionError::ArrowError(source, _) => (*source).into(),
+            DataFusionError::External(source) => Self::from_box_error(source),
+            other => Self::External {
+                source: Box::new(other),
+            },
+        }
     }
 }
 
 impl From<lance::Error> for Error {
     fn from(source: lance::Error) -> Self {
-        // TODO: Once Lance is changed to preserve ObjectStore, DataFusion, and Arrow errors, we can
-        // pass those variants through here as well.
-        Self::Lance { source }
+        // Try to unwrap external errors that were wrapped by lance
+        match source {
+            lance::Error::Wrapped { error, .. } => Self::from_box_error(error),
+            lance::Error::External { source } => Self::from_box_error(source),
+            _ => Self::Lance { source },
+        }
+    }
+}
+
+impl Error {
+    fn from_box_error(mut source: Box<dyn std::error::Error + Send + Sync>) -> Self {
+        source = match source.downcast::<Self>() {
+            Ok(e) => match *e {
+                Self::External { source } => return Self::from_box_error(source),
+                other => return other,
+            },
+            Err(source) => source,
+        };
+
+        source = match source.downcast::<lance::Error>() {
+            Ok(e) => match *e {
+                lance::Error::Wrapped { error, .. } => return Self::from_box_error(error),
+                other => return other.into(),
+            },
+            Err(source) => source,
+        };
+
+        source = match source.downcast::<ArrowError>() {
+            Ok(e) => match *e {
+                ArrowError::ExternalError(source) => return Self::from_box_error(source),
+                other => return other.into(),
+            },
+            Err(source) => source,
+        };
+
+        source = match source.downcast::<DataFusionError>() {
+            Ok(e) => match *e {
+                DataFusionError::ArrowError(source, _) => return (*source).into(),
+                DataFusionError::External(source) => return Self::from_box_error(source),
+                other => return other.into(),
+            },
+            Err(source) => source,
+        };
+
+        Self::External { source }
     }
 }
 

rust/lancedb/src/expr.rs

@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+//! Expression builder API for type-safe query construction
+//!
+//! This module provides a fluent API for building expressions that can be used
+//! in filters and projections. It wraps DataFusion's expression system.
+//!
+//! # Examples
+//!
+//! ```rust
+//! use std::ops::Mul;
+//! use lancedb::expr::{col, lit};
+//!
+//! let expr = col("age").gt(lit(18));
+//! let expr = col("age").gt(lit(18)).and(col("status").eq(lit("active")));
+//! let expr = col("price") * lit(1.1);
+//! ```
+
+mod sql;
+
+pub use sql::expr_to_sql_string;
+
+use std::sync::Arc;
+
+use arrow_schema::DataType;
+use datafusion_expr::{expr_fn::cast, Expr, ScalarUDF};
+use datafusion_functions::string::expr_fn as string_expr_fn;
+
+pub use datafusion_expr::{col, lit};
+
+pub use datafusion_expr::Expr as DfExpr;
+
+pub fn lower(expr: Expr) -> Expr {
+    string_expr_fn::lower(expr)
+}
+
+pub fn upper(expr: Expr) -> Expr {
+    string_expr_fn::upper(expr)
+}
+
+pub fn contains(expr: Expr, search: Expr) -> Expr {
+    string_expr_fn::contains(expr, search)
+}
+
+pub fn expr_cast(expr: Expr, data_type: DataType) -> Expr {
+    cast(expr, data_type)
+}
+
+lazy_static::lazy_static! {
+    static ref FUNC_REGISTRY: std::sync::RwLock<std::collections::HashMap<String, Arc<ScalarUDF>>> = {
+        let mut m = std::collections::HashMap::new();
+        m.insert("lower".to_string(), datafusion_functions::string::lower());
+        m.insert("upper".to_string(), datafusion_functions::string::upper());
+        m.insert("contains".to_string(), datafusion_functions::string::contains());
+        m.insert("btrim".to_string(), datafusion_functions::string::btrim());
+        m.insert("ltrim".to_string(), datafusion_functions::string::ltrim());
+        m.insert("rtrim".to_string(), datafusion_functions::string::rtrim());
+        m.insert("concat".to_string(), datafusion_functions::string::concat());
+        m.insert("octet_length".to_string(), datafusion_functions::string::octet_length());
+        std::sync::RwLock::new(m)
+    };
+}
+
+pub fn func(name: impl AsRef<str>, args: Vec<Expr>) -> crate::Result<Expr> {
+    let name = name.as_ref();
+    let registry = FUNC_REGISTRY
+        .read()
+        .map_err(|e| crate::Error::InvalidInput {
+            message: format!("lock poisoned: {}", e),
+        })?;
+    let udf = registry
+        .get(name)
+        .ok_or_else(|| crate::Error::InvalidInput {
+            message: format!("unknown function: {}", name),
+        })?;
+    Ok(Expr::ScalarFunction(
+        datafusion_expr::expr::ScalarFunction::new_udf(udf.clone(), args),
+    ))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_col_lit_comparisons() {
+        let expr = col("age").gt(lit(18));
+        let sql = expr_to_sql_string(&expr).unwrap();
+        assert!(sql.contains("age") && sql.contains("18"));
+
+        let expr = col("name").eq(lit("Alice"));
+        let sql = expr_to_sql_string(&expr).unwrap();
+        assert!(sql.contains("name") && sql.contains("Alice"));
+    }
+
+    #[test]
+    fn test_compound_expression() {
+        let expr = col("age").gt(lit(18)).and(col("status").eq(lit("active")));
+        let sql = expr_to_sql_string(&expr).unwrap();
+        assert!(sql.contains("age") && sql.contains("status"));
+    }
+
+    #[test]
+    fn test_string_functions() {
+        let expr = lower(col("name"));
+        let sql = expr_to_sql_string(&expr).unwrap();
+        assert!(sql.to_lowercase().contains("lower"));
+
+        let expr = contains(col("text"), lit("search"));
+        let sql = expr_to_sql_string(&expr).unwrap();
+        assert!(sql.to_lowercase().contains("contains"));
+    }
+
+    #[test]
+    fn test_func() {
+        let expr = func("lower", vec![col("x")]).unwrap();
+        let sql = expr_to_sql_string(&expr).unwrap();
+        assert!(sql.to_lowercase().contains("lower"));
+
+        let result = func("unknown_func", vec![col("x")]);
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_arithmetic() {
+        let expr = col("price") * lit(1.1);
+        let sql = expr_to_sql_string(&expr).unwrap();
+        assert!(sql.contains("price"));
+    }
+}

rust/lancedb/src/expr/sql.rs

@@ -0,0 +1,12 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+use datafusion_expr::Expr;
+use datafusion_sql::unparser;
+
+pub fn expr_to_sql_string(expr: &Expr) -> crate::Result<String> {
+    let ast = unparser::expr_to_sql(expr).map_err(|e| crate::Error::InvalidInput {
+        message: format!("failed to serialize expression to SQL: {}", e),
+    })?;
+    Ok(ast.to_string())
+}

rust/lancedb/src/io/object_store.rs

@@ -195,6 +195,11 @@ mod test {
         table::WriteOptions,
     };
 
+    // This test is ignored because lance 3.0 introduced LocalWriter optimization
+    // that bypasses the object store wrapper for local writes. The mirroring feature
+    // still works for remote/cloud storage, but can't be tested with local storage.
+    // See lance commit c878af433 "perf: create local writer for efficient local writes"
+    #[ignore]
     #[tokio::test]
     async fn test_e2e() {
         let dir1 = tempfile::tempdir().unwrap().keep().canonicalize().unwrap();
@@ -218,8 +223,9 @@ mod test {
         datagen = datagen.col(Box::<IncrementingInt32>::default());
         datagen = datagen.col(Box::new(RandomVector::default().named("vector".into())));
 
+        let data: Box<dyn arrow_array::RecordBatchReader + Send> = Box::new(datagen.batch(100));
         let res = db
-            .create_table("test", Box::new(datagen.batch(100)))
+            .create_table("test", data)
             .write_options(WriteOptions {
                 lance_write_params: Some(param),
             })
@@ -249,32 +255,38 @@ mod test {
         let primary_location = dir1.join("test.lance").canonicalize().unwrap();
         let secondary_location = dir2.join(primary_location.strip_prefix("/").unwrap());
 
-        let mut primary_iter = WalkDir::new(&primary_location).into_iter();
-        let mut secondary_iter = WalkDir::new(&secondary_location).into_iter();
+        // Skip lance internal directories (_versions, _transactions) and manifest files
+        let should_skip = |path: &std::path::Path| -> bool {
+            let path_str = path.to_str().unwrap();
+            path_str.contains("_latest.manifest")
+                || path_str.contains("_versions")
+                || path_str.contains("_transactions")
+        };
 
-        let mut primary_elem = primary_iter.next();
-        let mut secondary_elem = secondary_iter.next();
+        let primary_files: Vec<_> = WalkDir::new(&primary_location)
+            .into_iter()
+            .filter_entry(|e| !should_skip(e.path()))
+            .filter_map(|e| e.ok())
+            .map(|e| {
+                e.path()
+                    .strip_prefix(&primary_location)
+                    .unwrap()
+                    .to_path_buf()
+            })
+            .collect();
 
-        loop {
-            if primary_elem.is_none() && secondary_elem.is_none() {
-                break;
-            }
-            // primary has more data then secondary, should not run out before secondary
-            let primary_f = primary_elem.unwrap().unwrap();
-            // hit manifest, skip, _versions contains all the manifest and should not exist on secondary
-            let primary_raw_path = primary_f.file_name().to_str().unwrap();
-            if primary_raw_path.contains("_latest.manifest") {
-                primary_elem = primary_iter.next();
-                continue;
-            }
-            let secondary_f = secondary_elem.unwrap().unwrap();
-            assert_eq!(
-                primary_f.path().strip_prefix(&primary_location),
-                secondary_f.path().strip_prefix(&secondary_location)
-            );
+        let secondary_files: Vec<_> = WalkDir::new(&secondary_location)
+            .into_iter()
+            .filter_entry(|e| !should_skip(e.path()))
+            .filter_map(|e| e.ok())
+            .map(|e| {
+                e.path()
+                    .strip_prefix(&secondary_location)
+                    .unwrap()
+                    .to_path_buf()
+            })
+            .collect();
 
-            primary_elem = primary_iter.next();
-            secondary_elem = secondary_iter.next();
-        }
+        assert_eq!(primary_files, secondary_files, "File lists should match");
     }
 }

rust/lancedb/src/ipc.rs

@@ -12,10 +12,10 @@ use arrow_schema::Schema;
 use crate::{Error, Result};
 
 /// Convert a Arrow IPC file to a batch reader
-pub fn ipc_file_to_batches(buf: Vec<u8>) -> Result<impl RecordBatchReader> {
+pub fn ipc_file_to_batches(buf: Vec<u8>) -> Result<Box<dyn RecordBatchReader + Send>> {
     let buf_reader = Cursor::new(buf);
     let reader = FileReader::try_new(buf_reader, None)?;
-    Ok(reader)
+    Ok(Box::new(reader))
 }
 
 /// Convert record batches to Arrow IPC file

rust/lancedb/src/lib.rs

@@ -39,7 +39,6 @@
 //! #### Connect to a database.
 //!
 //! ```rust
-//! # use arrow_schema::{Field, Schema};
 //! # tokio::runtime::Runtime::new().unwrap().block_on(async {
 //! let db = lancedb::connect("data/sample-lancedb").execute().await.unwrap();
 //! # });
@@ -74,7 +73,10 @@
 //!
 //! #### Create a table
 //!
-//! To create a Table, you need to provide a [`arrow_schema::Schema`] and a [`arrow_array::RecordBatch`] stream.
+//! To create a Table, you need to provide an [`arrow_array::RecordBatch`]. The
+//! schema of the `RecordBatch` determines the schema of the table.
+//!
+//! Vector columns should be represented as `FixedSizeList<Float16/Float32>` data type.
 //!
 //! ```rust
 //! # use std::sync::Arc;
@@ -85,34 +87,29 @@
 //! # tokio::runtime::Runtime::new().unwrap().block_on(async {
 //! # let tmpdir = tempfile::tempdir().unwrap();
 //! # let db = lancedb::connect(tmpdir.path().to_str().unwrap()).execute().await.unwrap();
+//! let ndims = 128;
 //! let schema = Arc::new(Schema::new(vec![
 //!     Field::new("id", DataType::Int32, false),
 //!     Field::new(
 //!         "vector",
-//!         DataType::FixedSizeList(Arc::new(Field::new("item", DataType::Float32, true)), 128),
+//!         DataType::FixedSizeList(Arc::new(Field::new("item", DataType::Float32, true)), ndims),
 //!         true,
 //!     ),
 //! ]));
-//! // Create a RecordBatch stream.
-//! let batches = RecordBatchIterator::new(
-//!     vec![RecordBatch::try_new(
+//! let data = RecordBatch::try_new(
 //!         schema.clone(),
 //!         vec![
 //!             Arc::new(Int32Array::from_iter_values(0..256)),
 //!             Arc::new(
 //!                 FixedSizeListArray::from_iter_primitive::<Float32Type, _, _>(
-//!                     (0..256).map(|_| Some(vec![Some(1.0); 128])),
-//!                     128,
+//!                     (0..256).map(|_| Some(vec![Some(1.0); ndims as usize])),
+//!                     ndims,
 //!                 ),
 //!             ),
 //!         ],
 //!     )
-//!     .unwrap()]
-//!     .into_iter()
-//!     .map(Ok),
-//!     schema.clone(),
-//! );
-//! db.create_table("my_table", Box::new(batches))
+//!     .unwrap();
+//! db.create_table("my_table", data)
 //!     .execute()
 //!     .await
 //!     .unwrap();
@@ -151,42 +148,18 @@
 //! #### Open table and search
 //!
 //! ```rust
-//! # use std::sync::Arc;
 //! # use futures::TryStreamExt;
-//! # use arrow_schema::{DataType, Schema, Field};
-//! # use arrow_array::{RecordBatch, RecordBatchIterator};
-//! # use arrow_array::{FixedSizeListArray, Float32Array, Int32Array, types::Float32Type};
 //! # use lancedb::query::{ExecutableQuery, QueryBase};
-//! # tokio::runtime::Runtime::new().unwrap().block_on(async {
-//! # let tmpdir = tempfile::tempdir().unwrap();
-//! # let db = lancedb::connect(tmpdir.path().to_str().unwrap()).execute().await.unwrap();
-//! # let schema = Arc::new(Schema::new(vec![
-//! #  Field::new("id", DataType::Int32, false),
-//! #  Field::new("vector", DataType::FixedSizeList(
-//! #    Arc::new(Field::new("item", DataType::Float32, true)), 128), true),
-//! # ]));
-//! # let batches = RecordBatchIterator::new(vec![
-//! #    RecordBatch::try_new(schema.clone(),
-//! #       vec![
-//! #           Arc::new(Int32Array::from_iter_values(0..10)),
-//! #           Arc::new(FixedSizeListArray::from_iter_primitive::<Float32Type, _, _>(
-//! #               (0..10).map(|_| Some(vec![Some(1.0); 128])), 128)),
-//! #       ]).unwrap()
-//! #   ].into_iter().map(Ok),
-//! #    schema.clone());
-//! # db.create_table("my_table", Box::new(batches)).execute().await.unwrap();
-//! # let table = db.open_table("my_table").execute().await.unwrap();
+//! # async fn example(table: &lancedb::Table) -> lancedb::Result<()> {
 //! let results = table
 //!     .query()
-//!     .nearest_to(&[1.0; 128])
-//!     .unwrap()
+//!     .nearest_to(&[1.0; 128])?
 //!     .execute()
-//!     .await
-//!     .unwrap()
+//!     .await?
 //!     .try_collect::<Vec<_>>()
-//!     .await
-//!     .unwrap();
-//! # });
+//!     .await?;
+//! #   Ok(())
+//! # }
 //! ```
 
 pub mod arrow;
@@ -196,6 +169,7 @@ pub mod database;
 pub mod dataloader;
 pub mod embeddings;
 pub mod error;
+pub mod expr;
 pub mod index;
 pub mod io;
 pub mod ipc;
@@ -219,13 +193,14 @@ pub use error::{Error, Result};
 use lance_linalg::distance::DistanceType as LanceDistanceType;
 pub use table::Table;
 
-#[derive(Debug, Copy, Clone, PartialEq, Serialize, Deserialize)]
+#[derive(Debug, Copy, Clone, PartialEq, Serialize, Deserialize, Default)]
 #[non_exhaustive]
 #[serde(rename_all = "lowercase")]
 pub enum DistanceType {
     /// Euclidean distance. This is a very common distance metric that
     /// accounts for both magnitude and direction when determining the distance
     /// between vectors. l2 distance has a range of [0, ∞).
+    #[default]
     L2,
     /// Cosine distance.  Cosine distance is a distance metric
     /// calculated from the cosine similarity between two vectors. Cosine
@@ -247,12 +222,6 @@ pub enum DistanceType {
     Hamming,
 }
 
-impl Default for DistanceType {
-    fn default() -> Self {
-        Self::L2
-    }
-}
-
 impl From<DistanceType> for LanceDistanceType {
     fn from(value: DistanceType) -> Self {
         match value {

rust/lancedb/src/query.rs

@@ -359,6 +359,28 @@ pub trait QueryBase {
     /// on the filter column(s).
     fn only_if(self, filter: impl AsRef<str>) -> Self;
 
+    /// Only return rows which match the filter, using an expression builder.
+    ///
+    /// Use [`crate::expr`] for building type-safe expressions:
+    ///
+    /// ```
+    /// use lancedb::expr::{col, lit};
+    /// use lancedb::query::{QueryBase, ExecutableQuery};
+    ///
+    /// # use lancedb::Table;
+    /// # async fn query(table: &Table) -> Result<(), Box<dyn std::error::Error>> {
+    /// let results = table.query()
+    ///     .only_if_expr(col("age").gt(lit(18)).and(col("status").eq(lit("active"))))
+    ///     .execute()
+    ///     .await?;
+    /// # Ok(())
+    /// # }
+    /// ```
+    ///
+    /// Note: Expression filters are not supported for remote/server-side queries.
+    /// Use [`QueryBase::only_if`] with SQL strings for remote tables.
+    fn only_if_expr(self, filter: datafusion_expr::Expr) -> Self;
+
     /// Perform a full text search on the table.
     ///
     /// The results will be returned in order of BM25 scores.
@@ -468,6 +490,11 @@ impl<T: HasQuery> QueryBase for T {
         self
     }
 
+    fn only_if_expr(mut self, filter: datafusion_expr::Expr) -> Self {
+        self.mut_query().filter = Some(QueryFilter::Datafusion(filter));
+        self
+    }
+
     fn full_text_search(mut self, query: FullTextSearchQuery) -> Self {
         if self.mut_query().limit.is_none() {
             self.mut_query().limit = Some(DEFAULT_TOP_K);
@@ -1381,7 +1408,7 @@ mod tests {
     use arrow::{array::downcast_array, compute::concat_batches, datatypes::Int32Type};
     use arrow_array::{
         cast::AsArray, types::Float32Type, FixedSizeListArray, Float32Array, Int32Array,
-        RecordBatch, RecordBatchIterator, RecordBatchReader, StringArray,
+        RecordBatch, StringArray,
     };
     use arrow_schema::{DataType, Field as ArrowField, Schema as ArrowSchema};
     use futures::{StreamExt, TryStreamExt};
@@ -1402,7 +1429,7 @@ mod tests {
         let batches = make_test_batches();
         let conn = connect(uri).execute().await.unwrap();
         let table = conn
-            .create_table("my_table", Box::new(batches))
+            .create_table("my_table", batches)
             .execute()
             .await
             .unwrap();
@@ -1463,7 +1490,7 @@ mod tests {
         let batches = make_non_empty_batches();
         let conn = connect(uri).execute().await.unwrap();
         let table = conn
-            .create_table("my_table", Box::new(batches))
+            .create_table("my_table", batches)
             .execute()
             .await
             .unwrap();
@@ -1525,7 +1552,7 @@ mod tests {
         let batches = make_non_empty_batches();
         let conn = connect(uri).execute().await.unwrap();
         let table = conn
-            .create_table("my_table", Box::new(batches))
+            .create_table("my_table", batches)
             .execute()
             .await
             .unwrap();
@@ -1578,7 +1605,7 @@ mod tests {
         let batches = make_non_empty_batches();
         let conn = connect(uri).execute().await.unwrap();
         let table = conn
-            .create_table("my_table", Box::new(batches))
+            .create_table("my_table", batches)
             .execute()
             .await
             .unwrap();
@@ -1599,13 +1626,13 @@ mod tests {
         assert!(result.is_err());
     }
 
-    fn make_non_empty_batches() -> impl RecordBatchReader + Send + 'static {
+    fn make_non_empty_batches() -> Box<dyn arrow_array::RecordBatchReader + Send> {
         let vec = Box::new(RandomVector::new().named("vector".to_string()));
         let id = Box::new(IncrementingInt32::new().named("id".to_string()));
-        BatchGenerator::new().col(vec).col(id).batch(512)
+        Box::new(BatchGenerator::new().col(vec).col(id).batch(512))
     }
 
-    fn make_test_batches() -> impl RecordBatchReader + Send + 'static {
+    fn make_test_batches() -> RecordBatch {
         let dim: usize = 128;
         let schema = Arc::new(ArrowSchema::new(vec![
             ArrowField::new("key", DataType::Int32, false),
@@ -1619,12 +1646,7 @@ mod tests {
             ),
             ArrowField::new("uri", DataType::Utf8, true),
         ]));
-        RecordBatchIterator::new(
-            vec![RecordBatch::new_empty(schema.clone())]
-                .into_iter()
-                .map(Ok),
-            schema,
-        )
+        RecordBatch::new_empty(schema)
     }
 
     async fn make_test_table(tmp_dir: &tempfile::TempDir) -> Table {
@@ -1633,7 +1655,7 @@ mod tests {
 
         let batches = make_non_empty_batches();
         let conn = connect(uri).execute().await.unwrap();
-        conn.create_table("my_table", Box::new(batches))
+        conn.create_table("my_table", batches)
             .execute()
             .await
             .unwrap()
@@ -1862,10 +1884,8 @@ mod tests {
 
         let record_batch =
             RecordBatch::try_new(schema.clone(), vec![Arc::new(text), Arc::new(vector)]).unwrap();
-        let record_batch_iter =
-            RecordBatchIterator::new(vec![record_batch].into_iter().map(Ok), schema.clone());
         let table = conn
-            .create_table("my_table", record_batch_iter)
+            .create_table("my_table", record_batch)
             .execute()
             .await
             .unwrap();
@@ -1949,10 +1969,8 @@ mod tests {
             ],
         )
         .unwrap();
-        let record_batch_iter =
-            RecordBatchIterator::new(vec![record_batch].into_iter().map(Ok), schema.clone());
         let table = conn
-            .create_table("my_table", record_batch_iter)
+            .create_table("my_table", record_batch)
             .mode(CreateTableMode::Overwrite)
             .execute()
             .await
@@ -2062,8 +2080,6 @@ mod tests {
     async fn test_pagination_with_fts() {
         let db = connect("memory://test").execute().await.unwrap();
         let data = fts_test_data(400);
-        let schema = data.schema();
-        let data = RecordBatchIterator::new(vec![Ok(data)], schema);
         let table = db.create_table("test_table", data).execute().await.unwrap();
 
         table

rust/lancedb/src/remote/client.rs

@@ -491,7 +491,7 @@ impl<S: HttpSend> RestfulLanceDbClient<S> {
     }
 
     /// Apply dynamic headers from the header provider if configured
-    async fn apply_dynamic_headers(&self, mut request: Request) -> Result<Request> {
+    pub(crate) async fn apply_dynamic_headers(&self, mut request: Request) -> Result<Request> {
         if let Some(ref provider) = self.header_provider {
             let headers = provider.get_headers().await?;
             let request_headers = request.headers_mut();
@@ -555,7 +555,9 @@ impl<S: HttpSend> RestfulLanceDbClient<S> {
             message: "Attempted to retry a request that cannot be cloned".to_string(),
         })?;
         let (_, r) = tmp_req.build_split();
-        let mut r = r.unwrap();
+        let mut r = r.map_err(|e| Error::Runtime {
+            message: format!("Failed to build request: {}", e),
+        })?;
         let request_id = self.extract_request_id(&mut r);
         let mut retry_counter = RetryCounter::new(retry_config, request_id.clone());
 
@@ -571,7 +573,9 @@ impl<S: HttpSend> RestfulLanceDbClient<S> {
             }
 
             let (c, request) = req_builder.build_split();
-            let mut request = request.unwrap();
+            let mut request = request.map_err(|e| Error::Runtime {
+                message: format!("Failed to build request: {}", e),
+            })?;
             self.set_request_id(&mut request, &request_id.clone());
 
             // Apply dynamic headers before each retry attempt
@@ -621,7 +625,7 @@ impl<S: HttpSend> RestfulLanceDbClient<S> {
         }
     }
 
-    fn log_request(&self, request: &Request, request_id: &String) {
+    pub(crate) fn log_request(&self, request: &Request, request_id: &String) {
         if log::log_enabled!(log::Level::Debug) {
             let content_type = request
                 .headers()
@@ -720,12 +724,58 @@ pub mod test_utils {
         }
     }
 
+    /// Consume a reqwest body into bytes, returning an error if the body
+    /// stream fails. This is used by MockSender to materialize streaming
+    /// bodies so that data pipeline errors (e.g. NaN rejection) are triggered
+    /// during mock sends just as they would be during a real HTTP upload.
+    pub async fn try_collect_body(body: reqwest::Body) -> std::result::Result<Vec<u8>, String> {
+        use http_body::Body;
+        use std::pin::Pin;
+
+        let mut body = body;
+        let mut data = Vec::new();
+        let mut body_pin = Pin::new(&mut body);
+        while let Some(frame) = futures::StreamExt::next(&mut futures::stream::poll_fn(|cx| {
+            body_pin.as_mut().poll_frame(cx)
+        }))
+        .await
+        {
+            match frame {
+                Ok(frame) => {
+                    if let Some(bytes) = frame.data_ref() {
+                        data.extend_from_slice(bytes);
+                    }
+                }
+                Err(e) => return Err(e.to_string()),
+            }
+        }
+        Ok(data)
+    }
+
     impl HttpSend for MockSender {
         async fn send(
             &self,
             _client: &reqwest::Client,
-            request: reqwest::Request,
+            mut request: reqwest::Request,
         ) -> reqwest::Result<reqwest::Response> {
+            // Consume any streaming body to materialize it into bytes.
+            // This triggers data pipeline errors (e.g. NaN rejection) that
+            // would otherwise only fire when a real HTTP client reads the body.
+            if let Some(body) = request.body_mut().take() {
+                match try_collect_body(body).await {
+                    Ok(bytes) => {
+                        *request.body_mut() = Some(reqwest::Body::from(bytes));
+                    }
+                    Err(msg) => {
+                        // Simulate a failed request by returning a 500 response.
+                        return Ok(http::Response::builder()
+                            .status(500)
+                            .body(msg)
+                            .unwrap()
+                            .into());
+                    }
+                }
+            }
             let response = (self.f)(request);
             Ok(response)
         }

rust/lancedb/src/remote/db.rs

@@ -4,13 +4,11 @@
 use std::collections::HashMap;
 use std::sync::Arc;
 
-use arrow_array::RecordBatchIterator;
 use async_trait::async_trait;
 use http::StatusCode;
 use lance_io::object_store::StorageOptions;
 use moka::future::Cache;
 use reqwest::header::CONTENT_TYPE;
-use tokio::task::spawn_blocking;
 
 use lance_namespace::models::{
     CreateNamespaceRequest, CreateNamespaceResponse, DescribeNamespaceRequest,
@@ -19,16 +17,17 @@ use lance_namespace::models::{
 };
 
 use crate::database::{
-    CloneTableRequest, CreateTableData, CreateTableMode, CreateTableRequest, Database,
-    DatabaseOptions, OpenTableRequest, ReadConsistency, TableNamesRequest,
+    CloneTableRequest, CreateTableMode, CreateTableRequest, Database, DatabaseOptions,
+    OpenTableRequest, ReadConsistency, TableNamesRequest,
 };
 use crate::error::Result;
+use crate::remote::util::stream_as_body;
 use crate::table::BaseTable;
 use crate::Error;
 
 use super::client::{ClientConfig, HttpSend, RequestResultExt, RestfulLanceDbClient, Sender};
 use super::table::RemoteTable;
-use super::util::{batches_to_ipc_bytes, parse_server_version};
+use super::util::parse_server_version;
 use super::ARROW_STREAM_CONTENT_TYPE;
 
 // Request structure for the remote clone table API
@@ -436,26 +435,8 @@ impl<S: HttpSend> Database for RemoteDatabase<S> {
         Ok(response)
     }
 
-    async fn create_table(&self, request: CreateTableRequest) -> Result<Arc<dyn BaseTable>> {
-        let data = match request.data {
-            CreateTableData::Data(data) => data,
-            CreateTableData::StreamingData(_) => {
-                return Err(Error::NotSupported {
-                    message: "Creating a remote table from a streaming source".to_string(),
-                })
-            }
-            CreateTableData::Empty(table_definition) => {
-                let schema = table_definition.schema.clone();
-                Box::new(RecordBatchIterator::new(vec![], schema))
-            }
-        };
-
-        // TODO: https://github.com/lancedb/lancedb/issues/1026
-        // We should accept data from an async source.  In the meantime, spawn this as blocking
-        // to make sure we don't block the tokio runtime if the source is slow.
-        let data_buffer = spawn_blocking(move || batches_to_ipc_bytes(data))
-            .await
-            .unwrap()?;
+    async fn create_table(&self, mut request: CreateTableRequest) -> Result<Arc<dyn BaseTable>> {
+        let body = stream_as_body(request.data.scan_as_stream())?;
 
         let identifier =
             build_table_identifier(&request.name, &request.namespace, &self.client.id_delimiter);
@@ -463,7 +444,7 @@ impl<S: HttpSend> Database for RemoteDatabase<S> {
             .client
             .post(&format!("/v1/table/{}/create/", identifier))
             .query(&[("mode", Into::<&str>::into(&request.mode))])
-            .body(data_buffer)
+            .body(body)
             .header(CONTENT_TYPE, ARROW_STREAM_CONTENT_TYPE);
 
         let (request_id, rsp) = self.client.send(req).await?;
@@ -813,7 +794,7 @@ mod tests {
     use std::collections::HashMap;
     use std::sync::{Arc, OnceLock};
 
-    use arrow_array::{Int32Array, RecordBatch, RecordBatchIterator};
+    use arrow_array::{Int32Array, RecordBatch};
     use arrow_schema::{DataType, Field, Schema};
 
     use crate::connection::ConnectBuilder;
@@ -993,8 +974,7 @@ mod tests {
             vec![Arc::new(Int32Array::from(vec![1, 2, 3]))],
         )
         .unwrap();
-        let reader = RecordBatchIterator::new([Ok(data.clone())], data.schema());
-        let table = conn.create_table("table1", reader).execute().await.unwrap();
+        let table = conn.create_table("table1", data).execute().await.unwrap();
         assert_eq!(table.name(), "table1");
     }
 
@@ -1011,8 +991,7 @@ mod tests {
             vec![Arc::new(Int32Array::from(vec![1, 2, 3]))],
         )
         .unwrap();
-        let reader = RecordBatchIterator::new([Ok(data.clone())], data.schema());
-        let result = conn.create_table("table1", reader).execute().await;
+        let result = conn.create_table("table1", data).execute().await;
         assert!(result.is_err());
         assert!(
             matches!(result, Err(crate::Error::TableAlreadyExists { name }) if name == "table1")
@@ -1045,8 +1024,7 @@ mod tests {
                 vec![Arc::new(Int32Array::from(vec![1, 2, 3]))],
             )
             .unwrap();
-            let reader = RecordBatchIterator::new([Ok(data.clone())], data.schema());
-            let mut builder = conn.create_table("table1", reader);
+            let mut builder = conn.create_table("table1", data.clone());
             if let Some(mode) = mode {
                 builder = builder.mode(mode);
             }
@@ -1071,9 +1049,8 @@ mod tests {
         .unwrap();
 
         let called: Arc<OnceLock<bool>> = Arc::new(OnceLock::new());
-        let reader = RecordBatchIterator::new([Ok(data.clone())], data.schema());
         let called_in_cb = called.clone();
-        conn.create_table("table1", reader)
+        conn.create_table("table1", data)
             .mode(CreateTableMode::ExistOk(Box::new(move |b| {
                 called_in_cb.clone().set(true).unwrap();
                 b
@@ -1262,9 +1239,8 @@ mod tests {
             vec![Arc::new(Int32Array::from(vec![1, 2, 3]))],
         )
         .unwrap();
-        let reader = RecordBatchIterator::new([Ok(data.clone())], data.schema());
         let table = conn
-            .create_table("table1", reader)
+            .create_table("table1", data)
             .namespace(vec!["ns1".to_string()])
             .execute()
             .await
@@ -1730,10 +1706,8 @@ mod tests {
                 vec![Arc::new(Int32Array::from(vec![1, 2, 3]))],
             )
             .unwrap();
-            let reader = RecordBatchIterator::new([Ok(data.clone())], schema.clone());
-
             let table = conn
-                .create_table("test_table", reader)
+                .create_table("test_table", data)
                 .namespace(namespace.clone())
                 .execute()
                 .await;
@@ -1806,9 +1780,7 @@ mod tests {
                 let data =
                     RecordBatch::try_new(schema.clone(), vec![Arc::new(Int32Array::from(vec![i]))])
                         .unwrap();
-                let reader = RecordBatchIterator::new([Ok(data.clone())], schema.clone());
-
-                conn.create_table(format!("table{}", i), reader)
+                conn.create_table(format!("table{}", i), data)
                     .namespace(namespace.clone())
                     .execute()
                     .await

rust/lancedb/src/remote/retry.rs

@@ -60,6 +60,34 @@ impl<'a> RetryCounter<'a> {
         self.check_out_of_retries(Box::new(source), status_code)
     }
 
+    /// Increment the appropriate failure counter based on the error type.
+    ///
+    /// For `Error::Http` whose source is a connect error, increments
+    /// `connect_failures`. For read errors (`is_body` or `is_decode`),
+    /// increments `read_failures`. For all other errors, increments
+    /// `request_failures`. Calls `check_out_of_retries` to enforce global limits.
+    pub fn increment_from_error(&mut self, source: crate::Error) -> crate::Result<()> {
+        let reqwest_err = match &source {
+            crate::Error::Http { source, .. } => source.downcast_ref::<reqwest::Error>(),
+            _ => None,
+        };
+
+        if reqwest_err.is_some_and(|e| e.is_connect()) {
+            self.connect_failures += 1;
+        } else if reqwest_err.is_some_and(|e| e.is_body() || e.is_decode()) {
+            self.read_failures += 1;
+        } else {
+            self.request_failures += 1;
+        }
+
+        let status_code = if let crate::Error::Http { status_code, .. } = &source {
+            *status_code
+        } else {
+            None
+        };
+        self.check_out_of_retries(Box::new(source), status_code)
+    }
+
     pub fn increment_connect_failures(&mut self, source: reqwest::Error) -> crate::Result<()> {
         self.connect_failures += 1;
         let status_code = source.status();
@@ -77,7 +105,7 @@ impl<'a> RetryCounter<'a> {
         let jitter = rand::random::<f32>() * self.config.backoff_jitter;
         let sleep_time = Duration::from_secs_f32(backoff + jitter);
         debug!(
-            "Retrying request {:?} ({}/{} connect, {}/{} read, {}/{} read) in {:?}",
+            "Retrying request {:?} ({}/{} connect, {}/{} request, {}/{} read) in {:?}",
             self.request_id,
             self.connect_failures,
             self.config.connect_retries,
@@ -91,6 +119,115 @@ impl<'a> RetryCounter<'a> {
     }
 }
 
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn test_config() -> ResolvedRetryConfig {
+        ResolvedRetryConfig {
+            retries: 3,
+            connect_retries: 2,
+            read_retries: 3,
+            backoff_factor: 0.0,
+            backoff_jitter: 0.0,
+            statuses: vec![reqwest::StatusCode::BAD_GATEWAY],
+        }
+    }
+
+    /// Get a real reqwest connect error by trying to connect to a refused port.
+    async fn make_connect_error() -> reqwest::Error {
+        // Port 1 is almost always refused/unavailable.
+        reqwest::Client::new()
+            .get("http://127.0.0.1:1")
+            .send()
+            .await
+            .unwrap_err()
+    }
+
+    #[tokio::test]
+    async fn test_increment_from_error_connect() {
+        let config = test_config();
+        let mut counter = RetryCounter::new(&config, "test".to_string());
+
+        let connect_err = make_connect_error().await;
+        assert!(connect_err.is_connect());
+
+        let http_err = crate::Error::Http {
+            source: Box::new(connect_err),
+            request_id: "test".to_string(),
+            status_code: None,
+        };
+
+        // First connect failure: should be ok (1 < 2)
+        counter.increment_from_error(http_err).unwrap();
+        assert_eq!(counter.connect_failures, 1);
+        assert_eq!(counter.request_failures, 0);
+
+        // Second connect failure: should hit the limit (2 >= 2)
+        let connect_err2 = make_connect_error().await;
+        let http_err2 = crate::Error::Http {
+            source: Box::new(connect_err2),
+            request_id: "test".to_string(),
+            status_code: None,
+        };
+        let result = counter.increment_from_error(http_err2);
+        assert!(result.is_err());
+        assert!(matches!(
+            result.unwrap_err(),
+            crate::Error::Retry {
+                connect_failures: 2,
+                max_connect_failures: 2,
+                ..
+            }
+        ));
+    }
+
+    #[test]
+    fn test_increment_from_error_request() {
+        let config = test_config();
+        let mut counter = RetryCounter::new(&config, "test".to_string());
+
+        let http_err = crate::Error::Http {
+            source: "bad gateway".into(),
+            request_id: "test".to_string(),
+            status_code: Some(reqwest::StatusCode::BAD_GATEWAY),
+        };
+
+        counter.increment_from_error(http_err).unwrap();
+        assert_eq!(counter.request_failures, 1);
+        assert_eq!(counter.connect_failures, 0);
+    }
+
+    #[tokio::test]
+    async fn test_increment_from_error_respects_global_limits() {
+        // If request_failures is already at max, a connect error should still
+        // trigger the global limit check.
+        let config = test_config();
+        let mut counter = RetryCounter::new(&config, "test".to_string());
+        counter.request_failures = 3; // at max
+
+        let connect_err = make_connect_error().await;
+        let http_err = crate::Error::Http {
+            source: Box::new(connect_err),
+            request_id: "test".to_string(),
+            status_code: None,
+        };
+
+        // Even though connect_failures would be 1 (under limit of 2),
+        // request_failures is already at 3 (>= limit of 3), so this should fail.
+        let result = counter.increment_from_error(http_err);
+        assert!(result.is_err());
+        assert!(matches!(
+            result.unwrap_err(),
+            crate::Error::Retry {
+                request_failures: 3,
+                connect_failures: 1,
+                ..
+            }
+        ));
+    }
+}
+
 #[derive(Debug, Clone)]
 pub struct ResolvedRetryConfig {
     pub retries: u8,

rust/lancedb/src/remote/table/insert.rs

@@ -8,7 +8,6 @@ use std::sync::{Arc, Mutex};
 
 use arrow_array::{ArrayRef, RecordBatch, UInt64Array};
 use arrow_ipc::CompressionType;
-use arrow_schema::ArrowError;
 use datafusion_common::{DataFusionError, Result as DataFusionResult};
 use datafusion_execution::{SendableRecordBatchStream, TaskContext};
 use datafusion_physical_expr::EquivalenceProperties;
@@ -76,7 +75,15 @@ impl<S: HttpSend + 'static> RemoteInsertExec<S> {
         self.add_result.lock().unwrap().clone()
     }
 
-    fn stream_as_body(data: SendableRecordBatchStream) -> DataFusionResult<reqwest::Body> {
+    /// Stream the input into an HTTP body as an Arrow IPC stream, capturing any
+    /// stream errors into the provided channel. Errors from the input plan
+    /// (e.g. NaN rejection) would otherwise be swallowed inside the HTTP body
+    /// upload; by stashing them in the channel we can surface them with their
+    /// original message after the request completes.
+    fn stream_as_http_body(
+        data: SendableRecordBatchStream,
+        error_tx: tokio::sync::oneshot::Sender<DataFusionError>,
+    ) -> DataFusionResult<reqwest::Body> {
         let options = arrow_ipc::writer::IpcWriteOptions::default()
             .try_with_compression(Some(CompressionType::LZ4_FRAME))?;
         let writer = arrow_ipc::writer::StreamWriter::try_new_with_options(
@@ -85,26 +92,44 @@ impl<S: HttpSend + 'static> RemoteInsertExec<S> {
             options,
         )?;
 
-        let stream = futures::stream::try_unfold((data, writer), move |(mut data, mut writer)| {
-            async move {
+        let stream = futures::stream::try_unfold(
+            (data, writer, Some(error_tx), false),
+            move |(mut data, mut writer, error_tx, finished)| async move {
+                if finished {
+                    return Ok(None);
+                }
                 match data.next().await {
                     Some(Ok(batch)) => {
-                        writer.write(&batch)?;
+                        writer
+                            .write(&batch)
+                            .map_err(|e| std::io::Error::other(e.to_string()))?;
                         let buffer = std::mem::take(writer.get_mut());
-                        Ok(Some((buffer, (data, writer))))
+                        Ok(Some((buffer, (data, writer, error_tx, false))))
+                    }
+                    Some(Err(e)) => {
+                        // Send the original error through the channel before
+                        // returning a generic error to reqwest.
+                        if let Some(tx) = error_tx {
+                            let _ = tx.send(e);
+                        }
+                        Err(std::io::Error::other(
+                            "input stream error (see error channel)",
+                        ))
                     }
-                    Some(Err(e)) => Err(e),
                     None => {
-                        if let Err(ArrowError::IpcError(_msg)) = writer.finish() {
-                            // Will error if already closed.
-                            return Ok(None);
-                        };
+                        writer
+                            .finish()
+                            .map_err(|e| std::io::Error::other(e.to_string()))?;
                         let buffer = std::mem::take(writer.get_mut());
-                        Ok(Some((buffer, (data, writer))))
+                        if buffer.is_empty() {
+                            Ok(None)
+                        } else {
+                            Ok(Some((buffer, (data, writer, None, true))))
+                        }
                     }
                 }
-            }
-        });
+            },
+        );
 
         Ok(reqwest::Body::wrap_stream(stream))
     }
@@ -202,24 +227,41 @@ impl<S: HttpSend + 'static> ExecutionPlan for RemoteInsertExec<S> {
                 request = request.query(&[("mode", "overwrite")]);
             }
 
-            let body = Self::stream_as_body(input_stream)?;
+            let (error_tx, mut error_rx) = tokio::sync::oneshot::channel();
+            let body = Self::stream_as_http_body(input_stream, error_tx)?;
             let request = request.body(body);
 
-            let (request_id, response) = client
-                .send(request)
-                .await
-                .map_err(|e| DataFusionError::External(Box::new(e)))?;
-
-            let response =
-                RemoteTable::<Sender>::handle_table_not_found(&table_name, response, &request_id)
+            let result: DataFusionResult<(String, _)> = async {
+                let (request_id, response) = client
+                    .send(request)
                     .await
                     .map_err(|e| DataFusionError::External(Box::new(e)))?;
 
-            let response = client
-                .check_response(&request_id, response)
+                let response = RemoteTable::<Sender>::handle_table_not_found(
+                    &table_name,
+                    response,
+                    &request_id,
+                )
                 .await
                 .map_err(|e| DataFusionError::External(Box::new(e)))?;
 
+                let response = client
+                    .check_response(&request_id, response)
+                    .await
+                    .map_err(|e| DataFusionError::External(Box::new(e)))?;
+
+                Ok((request_id, response))
+            }
+            .await;
+
+            // If the request failed due to an input stream error, surface the
+            // original error (e.g. NaN rejection) instead of the HTTP error.
+            if let Ok(stream_err) = error_rx.try_recv() {
+                return Err(stream_err);
+            }
+
+            let (request_id, response) = result?;
+
             let body_text = response.text().await.map_err(|e| {
                 DataFusionError::External(Box::new(Error::Http {
                     source: Box::new(e),

rust/lancedb/src/remote/util.rs

@@ -1,29 +1,50 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
-use std::io::Cursor;
-
-use arrow_array::RecordBatchReader;
+use arrow_ipc::CompressionType;
+use futures::{Stream, StreamExt};
 use reqwest::Response;
 
-use crate::Result;
+use crate::{arrow::SendableRecordBatchStream, Result};
 
 use super::db::ServerVersion;
 
-pub fn batches_to_ipc_bytes(batches: impl RecordBatchReader) -> Result<Vec<u8>> {
+pub fn stream_as_ipc(
+    data: SendableRecordBatchStream,
+) -> Result<impl Stream<Item = Result<bytes::Bytes>>> {
+    let options = arrow_ipc::writer::IpcWriteOptions::default()
+        .try_with_compression(Some(CompressionType::LZ4_FRAME))?;
     const WRITE_BUF_SIZE: usize = 4096;
     let buf = Vec::with_capacity(WRITE_BUF_SIZE);
-    let mut buf = Cursor::new(buf);
-    {
-        let mut writer = arrow_ipc::writer::StreamWriter::try_new(&mut buf, &batches.schema())?;
+    let writer =
+        arrow_ipc::writer::StreamWriter::try_new_with_options(buf, &data.schema(), options)?;
+    let stream = futures::stream::try_unfold(
+        (data, writer, false),
+        move |(mut data, mut writer, finished)| async move {
+            if finished {
+                return Ok(None);
+            }
+            match data.next().await {
+                Some(Ok(batch)) => {
+                    writer.write(&batch)?;
+                    let buffer = std::mem::take(writer.get_mut());
+                    Ok(Some((bytes::Bytes::from(buffer), (data, writer, false))))
+                }
+                Some(Err(e)) => Err(e),
+                None => {
+                    writer.finish()?;
+                    let buffer = std::mem::take(writer.get_mut());
+                    Ok(Some((bytes::Bytes::from(buffer), (data, writer, true))))
+                }
+            }
+        },
+    );
+    Ok(stream)
+}
 
-        for batch in batches {
-            let batch = batch?;
-            writer.write(&batch)?;
-        }
-        writer.finish()?;
-    }
-    Ok(buf.into_inner())
+pub fn stream_as_body(data: SendableRecordBatchStream) -> Result<reqwest::Body> {
+    let stream = stream_as_ipc(data)?;
+    Ok(reqwest::Body::wrap_stream(stream))
 }
 
 pub fn parse_server_version(req_id: &str, rsp: &Response) -> Result<ServerVersion> {

rust/lancedb/src/table/add_data.rs

@@ -0,0 +1,714 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+use std::sync::Arc;
+
+use arrow_schema::{DataType, Fields, Schema};
+use lance::dataset::WriteMode;
+use serde::{Deserialize, Serialize};
+
+use crate::data::scannable::scannable_with_embeddings;
+use crate::data::scannable::Scannable;
+use crate::embeddings::EmbeddingRegistry;
+use crate::table::datafusion::cast::cast_to_table_schema;
+use crate::table::datafusion::reject_nan::reject_nan_vectors;
+use crate::table::datafusion::scannable_exec::ScannableExec;
+use crate::{Error, Result};
+
+use super::{BaseTable, TableDefinition, WriteOptions};
+
+#[derive(Debug, Clone, Default)]
+pub enum AddDataMode {
+    /// Rows will be appended to the table (the default)
+    #[default]
+    Append,
+    /// The existing table will be overwritten with the new data
+    Overwrite,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, Default)]
+pub struct AddResult {
+    // The commit version associated with the operation.
+    // A version of `0` indicates compatibility with legacy servers that do not return
+    /// a commit version.
+    #[serde(default)]
+    pub version: u64,
+}
+
+#[derive(Debug, Default, Clone, Copy)]
+pub enum NaNVectorBehavior {
+    /// Reject any vectors containing NaN values (the default)
+    #[default]
+    Error,
+    /// Allow NaN values to be added, but they will not be indexed for search
+    Keep,
+}
+
+/// A builder for configuring a [`crate::table::Table::add`] operation
+pub struct AddDataBuilder {
+    pub(crate) parent: Arc<dyn BaseTable>,
+    pub(crate) data: Box<dyn Scannable>,
+    pub(crate) mode: AddDataMode,
+    pub(crate) write_options: WriteOptions,
+    pub(crate) on_nan_vectors: NaNVectorBehavior,
+    pub(crate) embedding_registry: Option<Arc<dyn EmbeddingRegistry>>,
+}
+
+impl std::fmt::Debug for AddDataBuilder {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("AddDataBuilder")
+            .field("parent", &self.parent)
+            .field("mode", &self.mode)
+            .field("write_options", &self.write_options)
+            .finish()
+    }
+}
+
+impl AddDataBuilder {
+    pub(crate) fn new(
+        parent: Arc<dyn BaseTable>,
+        data: Box<dyn Scannable>,
+        embedding_registry: Option<Arc<dyn EmbeddingRegistry>>,
+    ) -> Self {
+        Self {
+            parent,
+            data,
+            mode: AddDataMode::Append,
+            write_options: WriteOptions::default(),
+            on_nan_vectors: NaNVectorBehavior::default(),
+            embedding_registry,
+        }
+    }
+
+    pub fn mode(mut self, mode: AddDataMode) -> Self {
+        self.mode = mode;
+        self
+    }
+
+    pub fn write_options(mut self, options: WriteOptions) -> Self {
+        self.write_options = options;
+        self
+    }
+
+    /// Configure how to handle NaN values in vector columns.
+    ///
+    /// By default, any vectors containing NaN values will be rejected with an
+    /// error, since NaNs cannot be indexed for search. Setting this to `Keep`
+    /// will allow NaN values to be added to the table, but they will not be
+    /// indexed and will not be searchable.
+    pub fn on_nan_vectors(mut self, behavior: NaNVectorBehavior) -> Self {
+        self.on_nan_vectors = behavior;
+        self
+    }
+
+    pub async fn execute(self) -> Result<AddResult> {
+        self.parent.clone().add(self).await
+    }
+
+    /// Build a DataFusion execution plan that applies embeddings, casts data to
+    /// the table schema, and optionally rejects NaN vectors.
+    ///
+    /// Returns the plan along with whether the input is rescannable (for retry
+    /// decisions) and whether this is an overwrite operation.
+    pub(crate) fn into_plan(
+        mut self,
+        table_schema: &Schema,
+        table_def: &TableDefinition,
+    ) -> Result<PreprocessingOutput> {
+        let overwrite = self
+            .write_options
+            .lance_write_params
+            .as_ref()
+            .is_some_and(|p| matches!(p.mode, WriteMode::Overwrite))
+            || matches!(self.mode, AddDataMode::Overwrite);
+
+        if !overwrite {
+            validate_schema(&self.data.schema(), table_schema)?;
+        }
+
+        self.data =
+            scannable_with_embeddings(self.data, table_def, self.embedding_registry.as_ref())?;
+
+        let rescannable = self.data.rescannable();
+        let plan: Arc<dyn datafusion_physical_plan::ExecutionPlan> =
+            Arc::new(ScannableExec::new(self.data));
+        // Skip casting when overwriting — the input schema replaces the table schema.
+        let plan = if overwrite {
+            plan
+        } else {
+            cast_to_table_schema(plan, table_schema)?
+        };
+        let plan = match self.on_nan_vectors {
+            NaNVectorBehavior::Error => reject_nan_vectors(plan)?,
+            NaNVectorBehavior::Keep => plan,
+        };
+
+        Ok(PreprocessingOutput {
+            plan,
+            overwrite,
+            rescannable,
+            write_options: self.write_options,
+            mode: self.mode,
+        })
+    }
+}
+
+pub struct PreprocessingOutput {
+    pub plan: Arc<dyn datafusion_physical_plan::ExecutionPlan>,
+    pub overwrite: bool,
+    pub rescannable: bool,
+    pub write_options: WriteOptions,
+    pub mode: AddDataMode,
+}
+
+/// Check that the input schema is valid for insert.
+///
+/// Fields can be in different orders, so match by name.
+///
+/// If a column exists in input but not in table, error (no extra columns allowed).
+///
+/// If a column exists in table but not in input, that is okay - it may be filled with nulls.
+///
+/// If the types are not exactly the same, we will attempt to cast later - so that is also okay at this stage.
+///
+/// If the nullability is different, that is also okay - we can relax nullability when casting.
+fn validate_schema(input: &Schema, table: &Schema) -> Result<()> {
+    validate_fields(input.fields(), table.fields())
+}
+
+fn validate_fields(input: &Fields, table: &Fields) -> Result<()> {
+    for field in input {
+        match table.iter().find(|f| f.name() == field.name()) {
+            None => {
+                return Err(Error::InvalidInput {
+                    message: format!("field '{}' does not exist in table schema", field.name()),
+                });
+            }
+            Some(table_field) => {
+                if let (DataType::Struct(in_children), DataType::Struct(tbl_children)) =
+                    (field.data_type(), table_field.data_type())
+                {
+                    validate_fields(in_children, tbl_children)?;
+                }
+            }
+        }
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use arrow::datatypes::Float64Type;
+    use arrow_array::{
+        record_batch, FixedSizeListArray, Float32Array, Int32Array, LargeStringArray, ListArray,
+        RecordBatch, RecordBatchIterator,
+    };
+    use arrow_schema::{ArrowError, DataType, Field, Schema};
+    use futures::TryStreamExt;
+    use lance::dataset::{WriteMode, WriteParams};
+
+    use crate::arrow::{SendableRecordBatchStream, SimpleRecordBatchStream};
+    use crate::connect;
+    use crate::data::scannable::Scannable;
+    use crate::embeddings::{
+        EmbeddingDefinition, EmbeddingFunction, EmbeddingRegistry, MemoryRegistry,
+    };
+    use crate::query::{ExecutableQuery, QueryBase, Select};
+    use crate::table::add_data::NaNVectorBehavior;
+    use crate::table::{ColumnDefinition, ColumnKind, Table, TableDefinition, WriteOptions};
+    use crate::test_utils::embeddings::MockEmbed;
+    use crate::test_utils::TestCustomError;
+    use crate::Error;
+
+    use super::AddDataMode;
+
+    async fn create_test_table() -> Table {
+        let conn = connect("memory://").execute().await.unwrap();
+        let batch = record_batch!(("id", Int64, [1, 2, 3])).unwrap();
+        conn.create_table("test", batch).execute().await.unwrap()
+    }
+
+    async fn test_add_with_data<T>(data: T)
+    where
+        T: Scannable + 'static,
+    {
+        let table = create_test_table().await;
+        let schema = data.schema();
+        table.add(data).execute().await.unwrap();
+        assert_eq!(table.count_rows(None).await.unwrap(), 5); // 3 initial + 2 added
+        assert_eq!(table.schema().await.unwrap(), schema);
+    }
+
+    #[tokio::test]
+    async fn test_add_with_batch() {
+        let batch = record_batch!(("id", Int64, [4, 5])).unwrap();
+        test_add_with_data(batch).await;
+    }
+
+    #[tokio::test]
+    async fn test_add_with_vec_batch() {
+        let data = vec![
+            record_batch!(("id", Int64, [4])).unwrap(),
+            record_batch!(("id", Int64, [5])).unwrap(),
+        ];
+        test_add_with_data(data).await;
+    }
+
+    #[tokio::test]
+    async fn test_add_with_record_batch_reader() {
+        let data = vec![
+            record_batch!(("id", Int64, [4])).unwrap(),
+            record_batch!(("id", Int64, [5])).unwrap(),
+        ];
+        let schema = data[0].schema();
+        let reader: Box<dyn arrow_array::RecordBatchReader + Send> = Box::new(
+            RecordBatchIterator::new(data.into_iter().map(Ok), schema.clone()),
+        );
+        test_add_with_data(reader).await;
+    }
+
+    #[tokio::test]
+    async fn test_add_with_stream() {
+        let data = vec![
+            record_batch!(("id", Int64, [4])).unwrap(),
+            record_batch!(("id", Int64, [5])).unwrap(),
+        ];
+        let schema = data[0].schema();
+        let inner = futures::stream::iter(data.into_iter().map(Ok));
+        let stream: SendableRecordBatchStream = Box::pin(SimpleRecordBatchStream {
+            schema,
+            stream: inner,
+        });
+        test_add_with_data(stream).await;
+    }
+
+    fn assert_preserves_external_error(err: &Error) {
+        assert!(
+            matches!(err, Error::External { source } if source.downcast_ref::<TestCustomError>().is_some()),
+            "Expected Error::External, got: {err:?}"
+        );
+        // The original TestCustomError message should be preserved through the
+        // error chain, even if the error gets wrapped multiple times by
+        // lance's insert pipeline.
+        assert!(
+            err.to_string().contains("TestCustomError occurred"),
+            "Expected original error message to be preserved, got: {err}"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_add_preserves_reader_error() {
+        let table = create_test_table().await;
+        let first_batch = record_batch!(("id", Int64, [4])).unwrap();
+        let schema = first_batch.schema();
+        let iterator = vec![
+            Ok(first_batch),
+            Err(ArrowError::ExternalError(Box::new(TestCustomError))),
+        ];
+        let reader: Box<dyn arrow_array::RecordBatchReader + Send> = Box::new(
+            RecordBatchIterator::new(iterator.into_iter(), schema.clone()),
+        );
+
+        let result = table.add(reader).execute().await;
+
+        assert_preserves_external_error(&result.unwrap_err());
+    }
+
+    #[tokio::test]
+    async fn test_add_preserves_stream_error() {
+        let table = create_test_table().await;
+        let first_batch = record_batch!(("id", Int64, [4])).unwrap();
+        let schema = first_batch.schema();
+        let iterator = vec![
+            Ok(first_batch),
+            Err(Error::External {
+                source: Box::new(TestCustomError),
+            }),
+        ];
+        let stream = futures::stream::iter(iterator);
+        let stream: SendableRecordBatchStream = Box::pin(SimpleRecordBatchStream {
+            schema: schema.clone(),
+            stream,
+        });
+
+        let result = table.add(stream).execute().await;
+
+        assert_preserves_external_error(&result.unwrap_err());
+    }
+
+    #[tokio::test]
+    async fn test_add() {
+        let conn = connect("memory://").execute().await.unwrap();
+
+        let batch = record_batch!(("i", Int32, [0, 1, 2])).unwrap();
+        let table = conn
+            .create_table("test", batch.clone())
+            .execute()
+            .await
+            .unwrap();
+        assert_eq!(table.count_rows(None).await.unwrap(), 3);
+
+        let new_batch = record_batch!(("i", Int32, [3])).unwrap();
+        table.add(new_batch).execute().await.unwrap();
+
+        assert_eq!(table.count_rows(None).await.unwrap(), 4);
+        assert_eq!(table.schema().await.unwrap(), batch.schema());
+    }
+
+    #[tokio::test]
+    async fn test_add_overwrite() {
+        let conn = connect("memory://").execute().await.unwrap();
+
+        let batch = record_batch!(("i", Int32, [0, 1, 2])).unwrap();
+        let table = conn
+            .create_table("test", batch.clone())
+            .execute()
+            .await
+            .unwrap();
+        assert_eq!(table.count_rows(None).await.unwrap(), batch.num_rows());
+
+        let new_batch = record_batch!(("x", Float32, [0.0, 1.0])).unwrap();
+        let res = table
+            .add(new_batch.clone())
+            .mode(AddDataMode::Overwrite)
+            .execute()
+            .await
+            .unwrap();
+        assert_eq!(res.version, table.version().await.unwrap());
+        assert_eq!(table.count_rows(None).await.unwrap(), new_batch.num_rows());
+        assert_eq!(table.schema().await.unwrap(), new_batch.schema());
+
+        // Can overwrite using underlying WriteParams (which
+        // take precedence over AddDataMode)
+        let param: WriteParams = WriteParams {
+            mode: WriteMode::Overwrite,
+            ..Default::default()
+        };
+
+        table
+            .add(new_batch.clone())
+            .write_options(WriteOptions {
+                lance_write_params: Some(param),
+            })
+            .mode(AddDataMode::Append)
+            .execute()
+            .await
+            .unwrap();
+        assert_eq!(table.count_rows(None).await.unwrap(), new_batch.num_rows());
+    }
+
+    #[tokio::test]
+    async fn test_add_with_embeddings() {
+        let registry = Arc::new(MemoryRegistry::new());
+        let mock_embedding: Arc<dyn EmbeddingFunction> = Arc::new(MockEmbed::new("mock", 4));
+        registry.register("mock", mock_embedding).unwrap();
+
+        let conn = connect("memory://")
+            .embedding_registry(registry)
+            .execute()
+            .await
+            .unwrap();
+
+        let schema = Arc::new(Schema::new(vec![
+            Field::new("text", DataType::Utf8, false),
+            Field::new(
+                "text_embedding",
+                DataType::FixedSizeList(Arc::new(Field::new("item", DataType::Float32, true)), 4),
+                false,
+            ),
+        ]));
+
+        // Add embedding metadata to the schema
+        let embedding_def = EmbeddingDefinition::new("text", "mock", Some("text_embedding"));
+        let table_def = TableDefinition::new(
+            schema.clone(),
+            vec![
+                ColumnDefinition {
+                    kind: ColumnKind::Physical,
+                },
+                ColumnDefinition {
+                    kind: ColumnKind::Embedding(embedding_def),
+                },
+            ],
+        );
+        let rich_schema = table_def.into_rich_schema();
+
+        let table = conn
+            .create_empty_table("embed_test", rich_schema)
+            .execute()
+            .await
+            .unwrap();
+
+        // Now add new data WITHOUT the embedding column - it should be computed automatically
+        let new_batch = record_batch!(("text", Utf8, ["hello", "world"])).unwrap();
+        table.add(new_batch).execute().await.unwrap();
+
+        assert_eq!(table.count_rows(None).await.unwrap(), 2);
+
+        // Query to verify the embeddings were computed for the new rows
+        let results: Vec<RecordBatch> = table
+            .query()
+            .select(Select::columns(&["text", "text_embedding"]))
+            .execute()
+            .await
+            .unwrap()
+            .try_collect()
+            .await
+            .unwrap();
+
+        let total_rows: usize = results.iter().map(|b| b.num_rows()).sum();
+        assert_eq!(total_rows, 2);
+
+        // Check that all rows have embedding values (not null)
+        for batch in &results {
+            let embedding_col = batch.column(1);
+            assert_eq!(embedding_col.null_count(), 0);
+        }
+    }
+
+    #[tokio::test]
+    async fn test_add_casts_to_table_schema() {
+        let table_schema = Arc::new(Schema::new(vec![
+            Field::new("id", DataType::Int64, false),
+            Field::new("text", DataType::Utf8, false),
+            Field::new(
+                "embedding",
+                DataType::FixedSizeList(Arc::new(Field::new("item", DataType::Float32, true)), 4),
+                false,
+            ),
+        ]));
+
+        let input_schema = Arc::new(Schema::new(vec![
+            Field::new("id", DataType::Int32, false), // Upcast integer
+            Field::new("text", DataType::LargeUtf8, false), // Re-encode string
+            // Cast list of float64 to fixed-size list of float32
+            // (This will only work if list size is correct. See next test.
+            Field::new(
+                "embedding",
+                DataType::List(Arc::new(Field::new("item", DataType::Float64, true))),
+                false,
+            ),
+        ]));
+
+        let db = connect("memory://").execute().await.unwrap();
+        let table = db
+            .create_empty_table("cast_test", table_schema.clone())
+            .execute()
+            .await
+            .unwrap();
+
+        let batch = RecordBatch::try_new(
+            input_schema,
+            vec![
+                Arc::new(Int32Array::from(vec![1, 2])),
+                Arc::new(LargeStringArray::from(vec!["hello", "world"])),
+                Arc::new(ListArray::from_iter_primitive::<Float64Type, _, _>(vec![
+                    Some(vec![0.1, 0.2, 0.3, 0.4].into_iter().map(Some)),
+                    Some(vec![0.5, 0.6, 0.7, 0.8].into_iter().map(Some)),
+                ])),
+            ],
+        )
+        .unwrap();
+        table.add(batch).execute().await.unwrap();
+
+        let row_count = table.count_rows(None).await.unwrap();
+        assert_eq!(row_count, 2);
+    }
+
+    #[tokio::test]
+    async fn test_add_rejects_bad_vector_dimensions() {
+        let table_schema = Arc::new(Schema::new(vec![Field::new(
+            "embedding",
+            DataType::FixedSizeList(Arc::new(Field::new("item", DataType::Float32, true)), 4),
+            false,
+        )]));
+
+        let input_schema = Arc::new(Schema::new(vec![Field::new(
+            "embedding",
+            DataType::List(Arc::new(Field::new("item", DataType::Float64, true))),
+            false,
+        )]));
+
+        let db = connect("memory://").execute().await.unwrap();
+        let table = db
+            .create_empty_table("cast_test", table_schema.clone())
+            .execute()
+            .await
+            .unwrap();
+
+        let batch = RecordBatch::try_new(
+            input_schema,
+            vec![Arc::new(
+                ListArray::from_iter_primitive::<Float64Type, _, _>(vec![
+                    Some(vec![0.1, 0.2, 0.3, 0.4].into_iter().map(Some)),
+                    Some(vec![0.5, 0.6, 0.8].into_iter().map(Some)),
+                ]),
+            )],
+        )
+        .unwrap();
+        let res = table.add(batch).execute().await;
+
+        // TODO: to recover the error, we will need fix upstream in Lance.
+        // assert!(
+        //     matches!(res, Err(Error::Arrow { source: ArrowError::CastError(_) })),
+        //     "Expected schema mismatch error due to wrong vector dimensions, but got: {res:?}"
+        // );
+        assert!(
+            res.is_err(),
+            "Expected error due to wrong vector dimensions, but got success"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_add_rejects_nan_vectors() {
+        let schema = Arc::new(Schema::new(vec![Field::new(
+            "embedding",
+            DataType::FixedSizeList(Arc::new(Field::new("item", DataType::Float32, true)), 4),
+            false,
+        )]));
+
+        let db = connect("memory://").execute().await.unwrap();
+        let table = db
+            .create_empty_table("nan_test", schema.clone())
+            .execute()
+            .await
+            .unwrap();
+
+        let batch = RecordBatch::try_new(
+            schema,
+            vec![Arc::new(
+                FixedSizeListArray::try_new(
+                    Arc::new(Field::new("item", DataType::Float32, true)),
+                    4,
+                    Arc::new(Float32Array::from(vec![0.1, 0.2, f32::NAN, 0.4])),
+                    None,
+                )
+                .unwrap(),
+            )],
+        )
+        .unwrap();
+        let res = table.add(batch.clone()).execute().await;
+        let err = res.unwrap_err();
+        assert!(
+            err.to_string().contains("NaN"),
+            "Expected error mentioning NaN values, but got: {err:?}"
+        );
+
+        table
+            .add(batch)
+            .on_nan_vectors(NaNVectorBehavior::Keep)
+            .execute()
+            .await
+            .unwrap();
+
+        let row_count = table.count_rows(None).await.unwrap();
+        assert_eq!(row_count, 1);
+    }
+
+    #[tokio::test]
+    async fn test_add_subschema() {
+        let data = record_batch!(("id", Int64, [4, 5]), ("text", Utf8, ["foo", "bar"])).unwrap();
+        let db = connect("memory://").execute().await.unwrap();
+        let table = db
+            .create_table("test", data.clone())
+            .execute()
+            .await
+            .unwrap();
+
+        let new_data = record_batch!(("id", Int64, [6, 7])).unwrap();
+        table.add(new_data).execute().await.unwrap();
+
+        assert_eq!(table.count_rows(None).await.unwrap(), 4);
+        assert_eq!(
+            table
+                .count_rows(Some("id IS NOT NULL".to_string()))
+                .await
+                .unwrap(),
+            4
+        );
+        assert_eq!(
+            table
+                .count_rows(Some("text IS NOT NULL".to_string()))
+                .await
+                .unwrap(),
+            2
+        );
+
+        // We can still cast
+        let new_data = record_batch!(("text", LargeUtf8, ["baz", "qux"])).unwrap();
+        table.add(new_data).execute().await.unwrap();
+
+        assert_eq!(table.count_rows(None).await.unwrap(), 6);
+        assert_eq!(
+            table
+                .count_rows(Some("id IS NOT NULL".to_string()))
+                .await
+                .unwrap(),
+            4
+        );
+        assert_eq!(
+            table
+                .count_rows(Some("text IS NOT NULL".to_string()))
+                .await
+                .unwrap(),
+            4
+        );
+
+        // Extra columns mean an error
+        let new_data =
+            record_batch!(("id", Int64, [8, 9]), ("extra", Utf8, ["extra1", "extra2"])).unwrap();
+        let res = table.add(new_data).execute().await;
+        assert!(
+            res.is_err(),
+            "Expected error due to extra column, but got: {res:?}"
+        );
+
+        // Insert with a subset of struct sub-fields
+        let struct_schema = Arc::new(Schema::new(vec![
+            Field::new("id", DataType::Int64, false),
+            Field::new(
+                "metadata",
+                DataType::Struct(
+                    vec![
+                        Field::new("a", DataType::Int64, true),
+                        Field::new("b", DataType::Utf8, true),
+                    ]
+                    .into(),
+                ),
+                true,
+            ),
+        ]));
+        let db2 = connect("memory://").execute().await.unwrap();
+        let table2 = db2
+            .create_empty_table("struct_test", struct_schema)
+            .execute()
+            .await
+            .unwrap();
+
+        // Insert with only the "a" sub-field of the struct
+        let sub_struct_schema = Arc::new(Schema::new(vec![
+            Field::new("id", DataType::Int64, false),
+            Field::new(
+                "metadata",
+                DataType::Struct(vec![Field::new("a", DataType::Int64, true)].into()),
+                true,
+            ),
+        ]));
+        let struct_batch = RecordBatch::try_new(
+            sub_struct_schema,
+            vec![
+                Arc::new(arrow_array::Int64Array::from(vec![1, 2])),
+                Arc::new(arrow_array::StructArray::from(vec![(
+                    Arc::new(Field::new("a", DataType::Int64, true)),
+                    Arc::new(arrow_array::Int64Array::from(vec![10, 20]))
+                        as Arc<dyn arrow_array::Array>,
+                )])),
+            ],
+        )
+        .unwrap();
+        table2.add(struct_batch).execute().await.unwrap();
+        assert_eq!(table2.count_rows(None).await.unwrap(), 2);
+    }
+}

rust/lancedb/src/table/datafusion.rs

@@ -3,7 +3,10 @@
 
 //! This module contains adapters to allow LanceDB tables to be used as DataFusion table providers.
 
+pub mod cast;
 pub mod insert;
+pub mod reject_nan;
+pub mod scannable_exec;
 pub mod udtf;
 
 use std::{collections::HashMap, sync::Arc};
@@ -287,8 +290,7 @@ pub mod tests {
 
     use arrow::array::AsArray;
     use arrow_array::{
-        BinaryArray, Float64Array, Int32Array, Int64Array, RecordBatch, RecordBatchIterator,
-        RecordBatchReader, StringArray, UInt32Array,
+        BinaryArray, Float64Array, Int32Array, Int64Array, RecordBatch, StringArray, UInt32Array,
     };
     use arrow_schema::{DataType, Field, Schema};
     use datafusion::{
@@ -308,7 +310,7 @@ pub mod tests {
         table::datafusion::BaseTableAdapter,
     };
 
-    fn make_test_batches() -> impl RecordBatchReader + Send + Sync + 'static {
+    fn make_test_batches() -> RecordBatch {
         let metadata = HashMap::from_iter(vec![("foo".to_string(), "bar".to_string())]);
         let schema = Arc::new(
             Schema::new(vec![
@@ -317,19 +319,17 @@ pub mod tests {
             ])
             .with_metadata(metadata),
         );
-        RecordBatchIterator::new(
-            vec![RecordBatch::try_new(
-                schema.clone(),
-                vec![
-                    Arc::new(Int32Array::from_iter_values(0..10)),
-                    Arc::new(UInt32Array::from_iter_values(0..10)),
-                ],
-            )],
+        RecordBatch::try_new(
             schema,
+            vec![
+                Arc::new(Int32Array::from_iter_values(0..10)),
+                Arc::new(UInt32Array::from_iter_values(0..10)),
+            ],
         )
+        .unwrap()
     }
 
-    fn make_tbl_two_test_batches() -> impl RecordBatchReader + Send + Sync + 'static {
+    fn make_tbl_two_test_batches() -> RecordBatch {
         let metadata = HashMap::from_iter(vec![("foo".to_string(), "bar".to_string())]);
         let schema = Arc::new(
             Schema::new(vec![
@@ -342,28 +342,26 @@ pub mod tests {
             ])
             .with_metadata(metadata),
         );
-        RecordBatchIterator::new(
-            vec![RecordBatch::try_new(
-                schema.clone(),
-                vec![
-                    Arc::new(Int64Array::from_iter_values(0..1000)),
-                    Arc::new(StringArray::from_iter_values(
-                        (0..1000).map(|i| i.to_string()),
-                    )),
-                    Arc::new(Float64Array::from_iter_values((0..1000).map(|i| i as f64))),
-                    Arc::new(StringArray::from_iter_values(
-                        (0..1000).map(|i| format!("{{\"i\":{}}}", i)),
-                    )),
-                    Arc::new(BinaryArray::from_iter_values(
-                        (0..1000).map(|i| (i as u32).to_be_bytes().to_vec()),
-                    )),
-                    Arc::new(StringArray::from_iter_values(
-                        (0..1000).map(|i| i.to_string()),
-                    )),
-                ],
-            )],
+        RecordBatch::try_new(
             schema,
+            vec![
+                Arc::new(Int64Array::from_iter_values(0..1000)),
+                Arc::new(StringArray::from_iter_values(
+                    (0..1000).map(|i| i.to_string()),
+                )),
+                Arc::new(Float64Array::from_iter_values((0..1000).map(|i| i as f64))),
+                Arc::new(StringArray::from_iter_values(
+                    (0..1000).map(|i| format!("{{\"i\":{}}}", i)),
+                )),
+                Arc::new(BinaryArray::from_iter_values(
+                    (0..1000).map(|i| (i as u32).to_be_bytes().to_vec()),
+                )),
+                Arc::new(StringArray::from_iter_values(
+                    (0..1000).map(|i| i.to_string()),
+                )),
+            ],
         )
+        .unwrap()
     }
 
     struct TestFixture {

rust/lancedb/src/table/datafusion/cast.rs

@@ -0,0 +1,498 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+use std::sync::Arc;
+
+use arrow_schema::{DataType, Field, FieldRef, Fields, Schema};
+use datafusion::functions::core::{get_field, named_struct};
+use datafusion_common::config::ConfigOptions;
+use datafusion_common::ScalarValue;
+use datafusion_physical_expr::expressions::{cast, Literal};
+use datafusion_physical_expr::ScalarFunctionExpr;
+use datafusion_physical_plan::expressions::Column;
+use datafusion_physical_plan::projection::ProjectionExec;
+use datafusion_physical_plan::{ExecutionPlan, PhysicalExpr};
+
+use crate::{Error, Result};
+
+pub fn cast_to_table_schema(
+    input: Arc<dyn ExecutionPlan>,
+    table_schema: &Schema,
+) -> Result<Arc<dyn ExecutionPlan>> {
+    let input_schema = input.schema();
+
+    if input_schema.fields() == table_schema.fields() {
+        return Ok(input);
+    }
+
+    let exprs = build_field_exprs(
+        input_schema.fields(),
+        table_schema.fields(),
+        &|idx| Arc::new(Column::new(input_schema.field(idx).name(), idx)) as Arc<dyn PhysicalExpr>,
+        &input_schema,
+    )?;
+
+    let exprs: Vec<(Arc<dyn PhysicalExpr>, String)> = exprs
+        .into_iter()
+        .map(|(expr, field)| (expr, field.name().clone()))
+        .collect();
+
+    let projection = ProjectionExec::try_new(exprs, input).map_err(crate::Error::from)?;
+
+    Ok(Arc::new(projection))
+}
+
+/// Build expressions to project input fields to match the table schema.
+///
+/// For each table field that exists in the input, produce an expression that
+/// reads from the input and casts if needed. Fields in the table but not in the
+/// input are omitted (the storage layer handles missing columns).
+fn build_field_exprs(
+    input_fields: &Fields,
+    table_fields: &Fields,
+    get_input_expr: &dyn Fn(usize) -> Arc<dyn PhysicalExpr>,
+    input_schema: &Schema,
+) -> Result<Vec<(Arc<dyn PhysicalExpr>, FieldRef)>> {
+    let config = Arc::new(ConfigOptions::default());
+    let mut result = Vec::new();
+
+    for table_field in table_fields {
+        let Some(input_idx) = input_fields
+            .iter()
+            .position(|f| f.name() == table_field.name())
+        else {
+            continue;
+        };
+
+        let input_field = &input_fields[input_idx];
+        let input_expr = get_input_expr(input_idx);
+
+        let expr = match (input_field.data_type(), table_field.data_type()) {
+            // Both are structs: recurse into sub-fields to handle subschemas and casts.
+            (DataType::Struct(in_children), DataType::Struct(tbl_children))
+                if in_children != tbl_children =>
+            {
+                let sub_exprs = build_field_exprs(
+                    in_children,
+                    tbl_children,
+                    &|child_idx| {
+                        let child_name = in_children[child_idx].name();
+                        Arc::new(ScalarFunctionExpr::new(
+                            &format!("get_field({child_name})"),
+                            get_field(),
+                            vec![
+                                input_expr.clone(),
+                                Arc::new(Literal::new(ScalarValue::from(child_name.as_str()))),
+                            ],
+                            Arc::new(in_children[child_idx].as_ref().clone()),
+                            config.clone(),
+                        )) as Arc<dyn PhysicalExpr>
+                    },
+                    input_schema,
+                )?;
+
+                let output_struct_fields: Fields = sub_exprs
+                    .iter()
+                    .map(|(_, f)| f.clone())
+                    .collect::<Vec<_>>()
+                    .into();
+                let output_field: FieldRef = Arc::new(Field::new(
+                    table_field.name(),
+                    DataType::Struct(output_struct_fields),
+                    table_field.is_nullable(),
+                ));
+
+                // Build named_struct(lit("a"), expr_a, lit("b"), expr_b, ...)
+                let mut ns_args: Vec<Arc<dyn PhysicalExpr>> = Vec::new();
+                for (sub_expr, sub_field) in &sub_exprs {
+                    ns_args.push(Arc::new(Literal::new(ScalarValue::from(
+                        sub_field.name().as_str(),
+                    ))));
+                    ns_args.push(sub_expr.clone());
+                }
+
+                let ns_expr: Arc<dyn PhysicalExpr> = Arc::new(ScalarFunctionExpr::new(
+                    &format!("named_struct({})", table_field.name()),
+                    named_struct(),
+                    ns_args,
+                    output_field.clone(),
+                    config.clone(),
+                ));
+
+                result.push((ns_expr, output_field));
+                continue;
+            }
+            // Types match: pass through.
+            (inp, tbl) if inp == tbl => input_expr,
+            // Types differ: cast.
+            _ => cast(input_expr, input_schema, table_field.data_type().clone()).map_err(|e| {
+                Error::InvalidInput {
+                    message: format!(
+                        "cannot cast field '{}' from {} to {}: {}",
+                        table_field.name(),
+                        input_field.data_type(),
+                        table_field.data_type(),
+                        e
+                    ),
+                }
+            })?,
+        };
+
+        let output_field = Arc::new(Field::new(
+            table_field.name(),
+            table_field.data_type().clone(),
+            table_field.is_nullable(),
+        ));
+        result.push((expr, output_field));
+    }
+
+    Ok(result)
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use arrow_array::{
+        Float32Array, Float64Array, Int32Array, Int64Array, RecordBatch, StringArray, StructArray,
+    };
+    use arrow_schema::{DataType, Field, Schema};
+    use datafusion::prelude::SessionContext;
+    use datafusion_catalog::MemTable;
+    use futures::TryStreamExt;
+
+    use super::cast_to_table_schema;
+
+    async fn plan_from_batch(
+        batch: RecordBatch,
+    ) -> Arc<dyn datafusion_physical_plan::ExecutionPlan> {
+        let schema = batch.schema();
+        let table = MemTable::try_new(schema, vec![vec![batch]]).unwrap();
+        let ctx = SessionContext::new();
+        ctx.register_table("t", Arc::new(table)).unwrap();
+        let df = ctx.table("t").await.unwrap();
+        df.create_physical_plan().await.unwrap()
+    }
+
+    async fn collect(plan: Arc<dyn datafusion_physical_plan::ExecutionPlan>) -> RecordBatch {
+        let ctx = SessionContext::new();
+        let stream = plan.execute(0, ctx.task_ctx()).unwrap();
+        let batches: Vec<RecordBatch> = stream.try_collect().await.unwrap();
+        arrow_select::concat::concat_batches(&plan.schema(), &batches).unwrap()
+    }
+
+    #[tokio::test]
+    async fn test_noop_when_schemas_match() {
+        let schema = Arc::new(Schema::new(vec![
+            Field::new("a", DataType::Int32, false),
+            Field::new("b", DataType::Utf8, false),
+        ]));
+        let batch = RecordBatch::try_new(
+            schema.clone(),
+            vec![
+                Arc::new(Int32Array::from(vec![1, 2])),
+                Arc::new(StringArray::from(vec!["x", "y"])),
+            ],
+        )
+        .unwrap();
+
+        let input = plan_from_batch(batch).await;
+        let input_ptr = Arc::as_ptr(&input);
+        let result = cast_to_table_schema(input, &schema).unwrap();
+        assert_eq!(Arc::as_ptr(&result), input_ptr);
+    }
+
+    #[tokio::test]
+    async fn test_simple_type_cast() {
+        let input_batch = RecordBatch::try_new(
+            Arc::new(Schema::new(vec![
+                Field::new("id", DataType::Int32, false),
+                Field::new("val", DataType::Float32, false),
+            ])),
+            vec![
+                Arc::new(Int32Array::from(vec![1, 2, 3])),
+                Arc::new(Float32Array::from(vec![1.5, 2.5, 3.5])),
+            ],
+        )
+        .unwrap();
+
+        let table_schema = Schema::new(vec![
+            Field::new("id", DataType::Int64, false),
+            Field::new("val", DataType::Float64, false),
+        ]);
+
+        let plan = plan_from_batch(input_batch).await;
+        let casted = cast_to_table_schema(plan, &table_schema).unwrap();
+        let result = collect(casted).await;
+
+        assert_eq!(result.schema().field(0).data_type(), &DataType::Int64);
+        assert_eq!(result.schema().field(1).data_type(), &DataType::Float64);
+
+        let ids: &Int64Array = result.column(0).as_any().downcast_ref().unwrap();
+        assert_eq!(ids.values(), &[1, 2, 3]);
+
+        let vals: &Float64Array = result.column(1).as_any().downcast_ref().unwrap();
+        assert!((vals.value(0) - 1.5).abs() < 1e-6);
+        assert!((vals.value(1) - 2.5).abs() < 1e-6);
+        assert!((vals.value(2) - 3.5).abs() < 1e-6);
+    }
+
+    #[tokio::test]
+    async fn test_missing_table_field_skipped() {
+        // Input has "a", table expects "a" and "b". "b" is omitted from the
+        // projection since the storage layer fills in missing columns.
+        let input_batch = RecordBatch::try_new(
+            Arc::new(Schema::new(vec![Field::new("a", DataType::Int32, false)])),
+            vec![Arc::new(Int32Array::from(vec![10, 20]))],
+        )
+        .unwrap();
+
+        let table_schema = Schema::new(vec![
+            Field::new("a", DataType::Int32, false),
+            Field::new("b", DataType::Utf8, true),
+        ]);
+
+        let plan = plan_from_batch(input_batch).await;
+        let casted = cast_to_table_schema(plan, &table_schema).unwrap();
+        let result = collect(casted).await;
+
+        assert_eq!(result.num_columns(), 1);
+        assert_eq!(result.schema().field(0).name(), "a");
+    }
+
+    #[tokio::test]
+    async fn test_extra_input_fields_dropped() {
+        // Input has "a" and "extra"; table only expects "a".
+        let input_batch = RecordBatch::try_new(
+            Arc::new(Schema::new(vec![
+                Field::new("a", DataType::Int32, false),
+                Field::new("extra", DataType::Utf8, false),
+            ])),
+            vec![
+                Arc::new(Int32Array::from(vec![1, 2])),
+                Arc::new(StringArray::from(vec!["x", "y"])),
+            ],
+        )
+        .unwrap();
+
+        let table_schema = Schema::new(vec![Field::new("a", DataType::Int64, false)]);
+
+        let plan = plan_from_batch(input_batch).await;
+        let casted = cast_to_table_schema(plan, &table_schema).unwrap();
+        let result = collect(casted).await;
+
+        assert_eq!(result.num_columns(), 1);
+        assert_eq!(result.schema().field(0).name(), "a");
+        assert_eq!(result.schema().field(0).data_type(), &DataType::Int64);
+    }
+
+    #[tokio::test]
+    async fn test_reorders_to_table_schema() {
+        let input_batch = RecordBatch::try_new(
+            Arc::new(Schema::new(vec![
+                Field::new("b", DataType::Utf8, false),
+                Field::new("a", DataType::Int32, false),
+            ])),
+            vec![
+                Arc::new(StringArray::from(vec!["x", "y"])),
+                Arc::new(Int32Array::from(vec![1, 2])),
+            ],
+        )
+        .unwrap();
+
+        let table_schema = Schema::new(vec![
+            Field::new("a", DataType::Int32, false),
+            Field::new("b", DataType::Utf8, false),
+        ]);
+
+        let plan = plan_from_batch(input_batch).await;
+        let casted = cast_to_table_schema(plan, &table_schema).unwrap();
+        let result = collect(casted).await;
+
+        assert_eq!(result.schema().field(0).name(), "a");
+        assert_eq!(result.schema().field(1).name(), "b");
+
+        let a: &Int32Array = result.column(0).as_any().downcast_ref().unwrap();
+        assert_eq!(a.values(), &[1, 2]);
+        let b: &StringArray = result.column(1).as_any().downcast_ref().unwrap();
+        assert_eq!(b.value(0), "x");
+    }
+
+    #[tokio::test]
+    async fn test_struct_subfield_cast() {
+        // Input struct has {x: Int32, y: Int32}, table expects {x: Int64, y: Int64}.
+        let inner_fields = vec![
+            Field::new("x", DataType::Int32, false),
+            Field::new("y", DataType::Int32, false),
+        ];
+        let struct_array = StructArray::from(vec![
+            (
+                Arc::new(inner_fields[0].clone()),
+                Arc::new(Int32Array::from(vec![1, 2])) as _,
+            ),
+            (
+                Arc::new(inner_fields[1].clone()),
+                Arc::new(Int32Array::from(vec![3, 4])) as _,
+            ),
+        ]);
+        let input_batch = RecordBatch::try_new(
+            Arc::new(Schema::new(vec![Field::new(
+                "s",
+                DataType::Struct(inner_fields.into()),
+                false,
+            )])),
+            vec![Arc::new(struct_array)],
+        )
+        .unwrap();
+
+        let table_inner = vec![
+            Field::new("x", DataType::Int64, false),
+            Field::new("y", DataType::Int64, false),
+        ];
+        let table_schema = Schema::new(vec![Field::new(
+            "s",
+            DataType::Struct(table_inner.into()),
+            false,
+        )]);
+
+        let plan = plan_from_batch(input_batch).await;
+        let casted = cast_to_table_schema(plan, &table_schema).unwrap();
+        let result = collect(casted).await;
+
+        let struct_col = result
+            .column(0)
+            .as_any()
+            .downcast_ref::<StructArray>()
+            .unwrap();
+        assert_eq!(struct_col.column(0).data_type(), &DataType::Int64);
+        assert_eq!(struct_col.column(1).data_type(), &DataType::Int64);
+
+        let x: &Int64Array = struct_col.column(0).as_any().downcast_ref().unwrap();
+        assert_eq!(x.values(), &[1, 2]);
+        let y: &Int64Array = struct_col.column(1).as_any().downcast_ref().unwrap();
+        assert_eq!(y.values(), &[3, 4]);
+    }
+
+    #[tokio::test]
+    async fn test_struct_subschema() {
+        // Input struct has {x, y, z}, table only expects {x, z}.
+        let inner_fields = vec![
+            Field::new("x", DataType::Int32, false),
+            Field::new("y", DataType::Int32, false),
+            Field::new("z", DataType::Int32, false),
+        ];
+        let struct_array = StructArray::from(vec![
+            (
+                Arc::new(inner_fields[0].clone()),
+                Arc::new(Int32Array::from(vec![1, 2])) as _,
+            ),
+            (
+                Arc::new(inner_fields[1].clone()),
+                Arc::new(Int32Array::from(vec![10, 20])) as _,
+            ),
+            (
+                Arc::new(inner_fields[2].clone()),
+                Arc::new(Int32Array::from(vec![100, 200])) as _,
+            ),
+        ]);
+        let input_batch = RecordBatch::try_new(
+            Arc::new(Schema::new(vec![Field::new(
+                "s",
+                DataType::Struct(inner_fields.into()),
+                false,
+            )])),
+            vec![Arc::new(struct_array)],
+        )
+        .unwrap();
+
+        let table_inner = vec![
+            Field::new("x", DataType::Int32, false),
+            Field::new("z", DataType::Int32, false),
+        ];
+        let table_schema = Schema::new(vec![Field::new(
+            "s",
+            DataType::Struct(table_inner.into()),
+            false,
+        )]);
+
+        let plan = plan_from_batch(input_batch).await;
+        let casted = cast_to_table_schema(plan, &table_schema).unwrap();
+        let result = collect(casted).await;
+
+        let struct_col = result
+            .column(0)
+            .as_any()
+            .downcast_ref::<StructArray>()
+            .unwrap();
+        assert_eq!(struct_col.num_columns(), 2);
+
+        let x: &Int32Array = struct_col
+            .column_by_name("x")
+            .unwrap()
+            .as_any()
+            .downcast_ref()
+            .unwrap();
+        assert_eq!(x.values(), &[1, 2]);
+        let z: &Int32Array = struct_col
+            .column_by_name("z")
+            .unwrap()
+            .as_any()
+            .downcast_ref()
+            .unwrap();
+        assert_eq!(z.values(), &[100, 200]);
+    }
+
+    #[tokio::test]
+    async fn test_incompatible_cast_errors() {
+        let input_batch = RecordBatch::try_new(
+            Arc::new(Schema::new(vec![Field::new("a", DataType::Binary, false)])),
+            vec![Arc::new(arrow_array::BinaryArray::from_vec(vec![b"hi"]))],
+        )
+        .unwrap();
+
+        let table_schema = Schema::new(vec![Field::new("a", DataType::Int32, false)]);
+
+        let plan = plan_from_batch(input_batch).await;
+        let result = cast_to_table_schema(plan, &table_schema);
+        assert!(result.is_err());
+        let err_msg = result.unwrap_err().to_string();
+        assert!(
+            err_msg.contains("cannot cast field 'a'"),
+            "unexpected error: {err_msg}"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_mixed_cast_and_passthrough() {
+        // "a" needs cast (Int32→Int64), "b" passes through unchanged.
+        let input_batch = RecordBatch::try_new(
+            Arc::new(Schema::new(vec![
+                Field::new("a", DataType::Int32, false),
+                Field::new("b", DataType::Utf8, false),
+            ])),
+            vec![
+                Arc::new(Int32Array::from(vec![7, 8])),
+                Arc::new(StringArray::from(vec!["hello", "world"])),
+            ],
+        )
+        .unwrap();
+
+        let table_schema = Schema::new(vec![
+            Field::new("a", DataType::Int64, false),
+            Field::new("b", DataType::Utf8, false),
+        ]);
+
+        let plan = plan_from_batch(input_batch).await;
+        let casted = cast_to_table_schema(plan, &table_schema).unwrap();
+        let result = collect(casted).await;
+
+        assert_eq!(result.schema().field(0).data_type(), &DataType::Int64);
+        assert_eq!(result.schema().field(1).data_type(), &DataType::Utf8);
+
+        let a: &Int64Array = result.column(0).as_any().downcast_ref().unwrap();
+        assert_eq!(a.values(), &[7, 8]);
+        let b: &StringArray = result.column(1).as_any().downcast_ref().unwrap();
+        assert_eq!(b.value(0), "hello");
+        assert_eq!(b.value(1), "world");
+    }
+}

rust/lancedb/src/table/datafusion/insert.rs

@@ -200,7 +200,7 @@ impl ExecutionPlan for InsertExec {
                     let new_dataset = CommitBuilder::new(dataset.clone())
                         .execute(merged_txn)
                         .await?;
-                    ds_wrapper.set_latest(new_dataset).await;
+                    ds_wrapper.update(new_dataset);
                 }
             }
 
@@ -222,7 +222,7 @@ mod tests {
     use std::vec;
 
     use super::*;
-    use arrow_array::{record_batch, Int32Array, RecordBatchIterator};
+    use arrow_array::{record_batch, RecordBatchIterator};
     use datafusion::prelude::SessionContext;
     use datafusion_catalog::MemTable;
     use tempfile::tempdir;
@@ -238,11 +238,8 @@ mod tests {
 
         // Create initial table
         let batch = record_batch!(("id", Int32, [1, 2, 3])).unwrap();
-        let schema = batch.schema();
-        let reader = RecordBatchIterator::new(vec![Ok(batch)], schema);
-
         let table = db
-            .create_table("test_insert", Box::new(reader))
+            .create_table("test_insert", batch)
             .execute()
             .await
             .unwrap();
@@ -279,11 +276,8 @@ mod tests {
 
         // Create initial table with 3 rows
         let batch = record_batch!(("id", Int32, [1, 2, 3])).unwrap();
-        let schema = batch.schema();
-        let reader = RecordBatchIterator::new(vec![Ok(batch)], schema);
-
         let table = db
-            .create_table("test_overwrite", Box::new(reader))
+            .create_table("test_overwrite", batch)
             .execute()
             .await
             .unwrap();
@@ -318,20 +312,9 @@ mod tests {
         let db = connect(uri).execute().await.unwrap();
 
         // Create initial table
-        let schema = Arc::new(ArrowSchema::new(vec![Field::new(
-            "id",
-            DataType::Int32,
-            false,
-        )]));
-        let batches = vec![RecordBatch::try_new(
-            schema.clone(),
-            vec![Arc::new(Int32Array::from(vec![1, 2, 3]))],
-        )
-        .unwrap()];
-        let reader = RecordBatchIterator::new(batches.into_iter().map(Ok), schema.clone());
-
+        let batch = record_batch!(("id", Int32, [1, 2, 3])).unwrap();
         let table = db
-            .create_table("test_empty", Box::new(reader))
+            .create_table("test_empty", batch)
             .execute()
             .await
             .unwrap();
@@ -352,12 +335,13 @@ mod tests {
             false,
         )]));
         // Empty batches
-        let source_reader = RecordBatchIterator::new(
-            std::iter::empty::<Result<RecordBatch, arrow_schema::ArrowError>>(),
-            source_schema,
-        );
+        let source_reader: Box<dyn arrow_array::RecordBatchReader + Send> =
+            Box::new(RecordBatchIterator::new(
+                std::iter::empty::<Result<RecordBatch, arrow_schema::ArrowError>>(),
+                source_schema,
+            ));
         let source_table = db
-            .create_table("empty_source", Box::new(source_reader))
+            .create_table("empty_source", source_reader)
             .execute()
             .await
             .unwrap();
@@ -389,20 +373,10 @@ mod tests {
         let db = connect(uri).execute().await.unwrap();
 
         // Create initial table
-        let schema = Arc::new(ArrowSchema::new(vec![Field::new(
-            "id",
-            DataType::Int32,
-            true,
-        )]));
-        let batches =
-            vec![
-                RecordBatch::try_new(schema.clone(), vec![Arc::new(Int32Array::from(vec![1]))])
-                    .unwrap(),
-            ];
-        let reader = RecordBatchIterator::new(batches.into_iter().map(Ok), schema.clone());
-
+        let batch = record_batch!(("id", Int32, [1])).unwrap();
+        let schema = batch.schema();
         let table = db
-            .create_table("test_multi_batch", Box::new(reader))
+            .create_table("test_multi_batch", batch)
             .execute()
             .await
             .unwrap();

rust/lancedb/src/table/datafusion/reject_nan.rs

@@ -0,0 +1,269 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+//! A DataFusion projection that rejects vectors containing NaN values.
+
+use std::any::Any;
+use std::sync::{Arc, LazyLock};
+
+use arrow_array::{Array, FixedSizeListArray};
+use arrow_schema::{DataType, Field, FieldRef};
+use datafusion_common::config::ConfigOptions;
+use datafusion_expr::{ColumnarValue, ScalarFunctionArgs, ScalarUDFImpl, Signature, Volatility};
+use datafusion_physical_expr::ScalarFunctionExpr;
+use datafusion_physical_plan::expressions::Column;
+use datafusion_physical_plan::projection::ProjectionExec;
+use datafusion_physical_plan::{ExecutionPlan, PhysicalExpr};
+
+use crate::{Error, Result};
+
+static REJECT_NAN_UDF: LazyLock<Arc<datafusion_expr::ScalarUDF>> =
+    LazyLock::new(|| Arc::new(datafusion_expr::ScalarUDF::from(RejectNanUdf::new())));
+
+/// Returns true if the field is a vector column: FixedSizeList<Float16/32/64>.
+fn is_vector_field(field: &Field) -> bool {
+    if let DataType::FixedSizeList(child, _) = field.data_type() {
+        matches!(
+            child.data_type(),
+            DataType::Float16 | DataType::Float32 | DataType::Float64
+        )
+    } else {
+        false
+    }
+}
+
+/// Wraps the input plan with a projection that checks vector columns for NaN values.
+///
+/// Non-vector columns pass through unchanged. Vector columns are wrapped with a
+/// UDF that returns the column as-is if no NaNs are present, or errors otherwise.
+pub fn reject_nan_vectors(input: Arc<dyn ExecutionPlan>) -> Result<Arc<dyn ExecutionPlan>> {
+    let schema = input.schema();
+    let config = Arc::new(ConfigOptions::default());
+    let udf = REJECT_NAN_UDF.clone();
+
+    let mut has_vector_cols = false;
+    let mut exprs: Vec<(Arc<dyn PhysicalExpr>, String)> = Vec::new();
+
+    for (idx, field) in schema.fields().iter().enumerate() {
+        let col_expr: Arc<dyn PhysicalExpr> = Arc::new(Column::new(field.name(), idx));
+
+        if is_vector_field(field) {
+            has_vector_cols = true;
+            let wrapped: Arc<dyn PhysicalExpr> = Arc::new(ScalarFunctionExpr::new(
+                &format!("reject_nan({})", field.name()),
+                udf.clone(),
+                vec![col_expr],
+                Arc::clone(field) as FieldRef,
+                config.clone(),
+            ));
+            exprs.push((wrapped, field.name().clone()));
+        } else {
+            exprs.push((col_expr, field.name().clone()));
+        }
+    }
+
+    if !has_vector_cols {
+        return Ok(input);
+    }
+
+    let projection = ProjectionExec::try_new(exprs, input).map_err(Error::from)?;
+    Ok(Arc::new(projection))
+}
+
+/// A scalar UDF that passes through FixedSizeList arrays unchanged, but errors
+/// if any float values in the list are NaN.
+#[derive(Debug, Hash, PartialEq, Eq)]
+struct RejectNanUdf {
+    signature: Signature,
+}
+
+impl RejectNanUdf {
+    fn new() -> Self {
+        Self {
+            signature: Signature::any(1, Volatility::Immutable),
+        }
+    }
+}
+
+impl ScalarUDFImpl for RejectNanUdf {
+    fn as_any(&self) -> &dyn Any {
+        self
+    }
+
+    fn name(&self) -> &str {
+        "reject_nan"
+    }
+
+    fn signature(&self) -> &Signature {
+        &self.signature
+    }
+
+    fn return_type(&self, arg_types: &[DataType]) -> datafusion_common::Result<DataType> {
+        Ok(arg_types[0].clone())
+    }
+
+    fn invoke_with_args(
+        &self,
+        args: ScalarFunctionArgs,
+    ) -> datafusion_common::Result<ColumnarValue> {
+        let arg = &args.args[0];
+        match arg {
+            ColumnarValue::Array(array) => {
+                check_no_nans(array)?;
+                Ok(ColumnarValue::Array(array.clone()))
+            }
+            ColumnarValue::Scalar(_) => Ok(arg.clone()),
+        }
+    }
+}
+
+fn check_no_nans(array: &dyn Array) -> datafusion_common::Result<()> {
+    let fsl = array
+        .as_any()
+        .downcast_ref::<FixedSizeListArray>()
+        .ok_or_else(|| {
+            datafusion_common::DataFusionError::Internal(
+                "reject_nan expected FixedSizeList".to_string(),
+            )
+        })?;
+
+    // Only inspect elements that are both in a valid parent row and non-null
+    // themselves. Values backing null parent rows or null child elements may
+    // contain garbage (including NaN) per the Arrow spec.
+    let has_nan = (0..fsl.len()).filter(|i| fsl.is_valid(*i)).any(|i| {
+        let row = fsl.value(i);
+        match row.data_type() {
+            DataType::Float16 => row
+                .as_any()
+                .downcast_ref::<arrow_array::Float16Array>()
+                .unwrap()
+                .iter()
+                .any(|v| v.is_some_and(|v| v.is_nan())),
+            DataType::Float32 => row
+                .as_any()
+                .downcast_ref::<arrow_array::Float32Array>()
+                .unwrap()
+                .iter()
+                .any(|v| v.is_some_and(|v| v.is_nan())),
+            DataType::Float64 => row
+                .as_any()
+                .downcast_ref::<arrow_array::Float64Array>()
+                .unwrap()
+                .iter()
+                .any(|v| v.is_some_and(|v| v.is_nan())),
+            _ => false,
+        }
+    });
+
+    if has_nan {
+        return Err(datafusion_common::DataFusionError::ArrowError(
+            Box::new(arrow_schema::ArrowError::ComputeError(
+                "Vector column contains NaN values".to_string(),
+            )),
+            None,
+        ));
+    }
+
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use arrow_array::Float32Array;
+
+    #[test]
+    fn test_passes_clean_vectors() {
+        let fsl = FixedSizeListArray::try_new(
+            Arc::new(Field::new("item", DataType::Float32, true)),
+            2,
+            Arc::new(Float32Array::from(vec![1.0, 2.0, 3.0, 4.0])),
+            None,
+        )
+        .unwrap();
+        assert!(check_no_nans(&fsl).is_ok());
+    }
+
+    #[test]
+    fn test_rejects_nan_vectors() {
+        let fsl = FixedSizeListArray::try_new(
+            Arc::new(Field::new("item", DataType::Float32, true)),
+            2,
+            Arc::new(Float32Array::from(vec![1.0, f32::NAN, 3.0, 4.0])),
+            None,
+        )
+        .unwrap();
+        assert!(check_no_nans(&fsl).is_err());
+    }
+
+    #[test]
+    fn test_skips_null_rows() {
+        // Values backing null rows may contain NaN per the Arrow spec.
+        // We should not reject a batch just because of garbage in null slots.
+        let values = Float32Array::from(vec![1.0, 2.0, f32::NAN, f32::NAN]);
+        let fsl = FixedSizeListArray::try_new(
+            Arc::new(Field::new("item", DataType::Float32, true)),
+            2,
+            Arc::new(values),
+            // Row 0 is valid [1.0, 2.0], row 1 is null [NAN, NAN]
+            Some(vec![true, false].into()),
+        )
+        .unwrap();
+        assert!(fsl.is_null(1));
+        assert!(check_no_nans(&fsl).is_ok());
+    }
+
+    #[test]
+    fn test_skips_null_elements_within_valid_row() {
+        // A valid row with null child elements: the underlying buffer may hold
+        // NaN but the null bitmap says they're absent — should not reject.
+        let values = Float32Array::from(vec![
+            Some(1.0),
+            None, // null element — buffer may contain NaN
+            Some(3.0),
+            None, // null element
+        ]);
+        let fsl = FixedSizeListArray::try_new(
+            Arc::new(Field::new("item", DataType::Float32, true)),
+            2,
+            Arc::new(values),
+            None, // both rows are valid
+        )
+        .unwrap();
+        assert!(check_no_nans(&fsl).is_ok());
+    }
+
+    #[test]
+    fn test_rejects_nan_in_valid_row_with_nulls_present() {
+        // Row 0 is null, row 1 is valid but contains NaN — should reject.
+        let values = Float32Array::from(vec![0.0, 0.0, 1.0, f32::NAN]);
+        let fsl = FixedSizeListArray::try_new(
+            Arc::new(Field::new("item", DataType::Float32, true)),
+            2,
+            Arc::new(values),
+            Some(vec![false, true].into()),
+        )
+        .unwrap();
+        assert!(check_no_nans(&fsl).is_err());
+    }
+
+    #[test]
+    fn test_is_vector_field() {
+        assert!(is_vector_field(&Field::new(
+            "v",
+            DataType::FixedSizeList(Arc::new(Field::new("item", DataType::Float32, true)), 4),
+            false,
+        )));
+        assert!(is_vector_field(&Field::new(
+            "v",
+            DataType::FixedSizeList(Arc::new(Field::new("item", DataType::Float64, true)), 4),
+            false,
+        )));
+        assert!(!is_vector_field(&Field::new("id", DataType::Int32, false)));
+        assert!(!is_vector_field(&Field::new(
+            "v",
+            DataType::FixedSizeList(Arc::new(Field::new("item", DataType::Int32, true)), 4),
+            false,
+        )));
+    }
+}

rust/lancedb/src/table/datafusion/scannable_exec.rs

@@ -0,0 +1,118 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+use core::fmt;
+use std::sync::{Arc, Mutex};
+
+use datafusion_common::{stats::Precision, DataFusionError, Result as DFResult, Statistics};
+use datafusion_execution::{SendableRecordBatchStream, TaskContext};
+use datafusion_physical_expr::{EquivalenceProperties, Partitioning};
+use datafusion_physical_plan::{
+    execution_plan::EmissionType, DisplayAs, DisplayFormatType, ExecutionPlan, PlanProperties,
+};
+
+use crate::{arrow::SendableRecordBatchStreamExt, data::scannable::Scannable};
+
+pub struct ScannableExec {
+    // We don't require Scannable to by Sync, so we wrap it in a Mutex to allow safe concurrent access.
+    source: Mutex<Box<dyn Scannable>>,
+    num_rows: Option<usize>,
+    properties: PlanProperties,
+}
+
+impl std::fmt::Debug for ScannableExec {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("ScannableExec")
+            .field("schema", &self.schema())
+            .field("num_rows", &self.num_rows)
+            .finish()
+    }
+}
+
+impl ScannableExec {
+    pub fn new(source: Box<dyn Scannable>) -> Self {
+        let schema = source.schema();
+        let eq_properties = EquivalenceProperties::new(schema);
+        let properties = PlanProperties::new(
+            eq_properties,
+            Partitioning::UnknownPartitioning(1),
+            EmissionType::Incremental,
+            datafusion_physical_plan::execution_plan::Boundedness::Bounded,
+        );
+
+        let num_rows = source.num_rows();
+        let source = Mutex::new(source);
+        Self {
+            source,
+            num_rows,
+            properties,
+        }
+    }
+}
+
+impl DisplayAs for ScannableExec {
+    fn fmt_as(&self, _t: DisplayFormatType, f: &mut std::fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "ScannableExec: num_rows={:?}", self.num_rows)
+    }
+}
+
+impl ExecutionPlan for ScannableExec {
+    fn name(&self) -> &str {
+        "ScannableExec"
+    }
+
+    fn as_any(&self) -> &dyn std::any::Any {
+        self
+    }
+
+    fn properties(&self) -> &PlanProperties {
+        &self.properties
+    }
+
+    fn children(&self) -> Vec<&Arc<dyn ExecutionPlan>> {
+        vec![]
+    }
+
+    fn with_new_children(
+        self: Arc<Self>,
+        children: Vec<Arc<dyn ExecutionPlan>>,
+    ) -> DFResult<Arc<dyn ExecutionPlan>> {
+        if !children.is_empty() {
+            return Err(DataFusionError::Internal(
+                "ScannableExec does not have children".to_string(),
+            ));
+        }
+        Ok(self)
+    }
+
+    fn execute(
+        &self,
+        partition: usize,
+        _context: Arc<TaskContext>,
+    ) -> DFResult<SendableRecordBatchStream> {
+        if partition != 0 {
+            return Err(DataFusionError::Internal(format!(
+                "ScannableExec only supports partition 0, got {}",
+                partition
+            )));
+        }
+
+        let stream = match self.source.lock() {
+            Ok(mut guard) => guard.scan_as_stream(),
+            Err(poison) => poison.into_inner().scan_as_stream(),
+        };
+
+        Ok(stream.into_df_stream())
+    }
+
+    fn partition_statistics(&self, _partition: Option<usize>) -> DFResult<Statistics> {
+        Ok(Statistics {
+            num_rows: self
+                .num_rows
+                .map(Precision::Exact)
+                .unwrap_or(Precision::Absent),
+            total_byte_size: Precision::Absent,
+            column_statistics: vec![],
+        })
+    }
+}

rust/lancedb/src/table/datafusion/udtf/fts.rs

@@ -97,7 +97,7 @@ mod tests {
         table::datafusion::BaseTableAdapter,
         Connection, Table,
     };
-    use arrow_array::{Int32Array, RecordBatch, RecordBatchIterator, StringArray};
+    use arrow_array::{Int32Array, RecordBatch, StringArray};
     use arrow_schema::{DataType, Field, Schema as ArrowSchema};
     use datafusion::prelude::SessionContext;
 
@@ -173,14 +173,7 @@ mod tests {
 
         // Create LanceDB database and table
         let db = crate::connect("memory://test").execute().await.unwrap();
-        let table = db
-            .create_table(
-                "foo",
-                RecordBatchIterator::new(vec![Ok(batch)].into_iter(), schema),
-            )
-            .execute()
-            .await
-            .unwrap();
+        let table = db.create_table("foo", batch).execute().await.unwrap();
 
         // Create FTS index
         table
@@ -323,13 +316,7 @@ mod tests {
             RecordBatch::try_new(metadata_schema.clone(), vec![metadata_col, extra_col]).unwrap();
 
         let _metadata_table = db
-            .create_table(
-                "metadata",
-                RecordBatchIterator::new(
-                    vec![Ok(metadata_batch.clone())].into_iter(),
-                    metadata_schema.clone(),
-                ),
-            )
+            .create_table("metadata", metadata_batch.clone())
             .execute()
             .await
             .unwrap();
@@ -393,14 +380,7 @@ mod tests {
         let batch =
             RecordBatch::try_new(schema.clone(), vec![id_col, text_col, category_col]).unwrap();
 
-        let table = db
-            .create_table(
-                table_name,
-                RecordBatchIterator::new(vec![Ok(batch)].into_iter(), schema),
-            )
-            .execute()
-            .await
-            .unwrap();
+        let table = db.create_table(table_name, batch).execute().await.unwrap();
 
         // Create FTS index
         table
@@ -546,14 +526,7 @@ mod tests {
         ]));
         let batch = RecordBatch::try_new(schema.clone(), vec![id_col, text_col]).unwrap();
 
-        let table = db
-            .create_table(
-                "docs",
-                RecordBatchIterator::new(vec![Ok(batch)].into_iter(), schema),
-            )
-            .execute()
-            .await
-            .unwrap();
+        let table = db.create_table("docs", batch).execute().await.unwrap();
 
         // Create FTS index with position information for phrase queries
         table
@@ -691,14 +664,7 @@ mod tests {
         let batch =
             RecordBatch::try_new(schema.clone(), vec![id_col, title_col, content_col]).unwrap();
 
-        let table = db
-            .create_table(
-                "multi_col",
-                RecordBatchIterator::new(vec![Ok(batch)].into_iter(), schema),
-            )
-            .execute()
-            .await
-            .unwrap();
+        let table = db.create_table("multi_col", batch).execute().await.unwrap();
 
         // Create FTS indices on both columns
         table
@@ -963,13 +929,7 @@ mod tests {
         let metadata_batch =
             RecordBatch::try_new(metadata_schema.clone(), vec![metadata_id, extra_info]).unwrap();
         let _metadata_table = db
-            .create_table(
-                "metadata",
-                RecordBatchIterator::new(
-                    vec![Ok(metadata_batch.clone())].into_iter(),
-                    metadata_schema,
-                ),
-            )
+            .create_table("metadata", metadata_batch.clone())
             .execute()
             .await
             .unwrap();
@@ -1358,14 +1318,7 @@ mod tests {
         ]));
         let batch = RecordBatch::try_new(schema.clone(), vec![id_col, text_col]).unwrap();
 
-        let table = db
-            .create_table(
-                "docs",
-                RecordBatchIterator::new(vec![Ok(batch)].into_iter(), schema),
-            )
-            .execute()
-            .await
-            .unwrap();
+        let table = db.create_table("docs", batch).execute().await.unwrap();
 
         // Create FTS index with position information
         table
@@ -1510,14 +1463,7 @@ mod tests {
         let batch =
             RecordBatch::try_new(schema.clone(), vec![id_col, title_col, content_col]).unwrap();
 
-        let table = db
-            .create_table(
-                "docs",
-                RecordBatchIterator::new(vec![Ok(batch)].into_iter(), schema),
-            )
-            .execute()
-            .await
-            .unwrap();
+        let table = db.create_table("docs", batch).execute().await.unwrap();
 
         // Create FTS indices on both columns
         table
@@ -1591,14 +1537,7 @@ mod tests {
         let batch =
             RecordBatch::try_new(schema.clone(), vec![id_col, title_col, content_col]).unwrap();
 
-        let table = db
-            .create_table(
-                "docs",
-                RecordBatchIterator::new(vec![Ok(batch)].into_iter(), schema),
-            )
-            .execute()
-            .await
-            .unwrap();
+        let table = db.create_table("docs", batch).execute().await.unwrap();
 
         // Create FTS indices
         table
@@ -1724,36 +1663,23 @@ mod tests {
             .unwrap();
 
         // Create table with simple text for n-gram testing
-        let data = RecordBatchIterator::new(
-            vec![RecordBatch::try_new(
-                Arc::new(ArrowSchema::new(vec![
-                    Field::new("id", DataType::Int32, false),
-                    Field::new("text", DataType::Utf8, false),
-                ])),
-                vec![
-                    Arc::new(Int32Array::from(vec![1, 2, 3])),
-                    Arc::new(StringArray::from(vec![
-                        "hello world",
-                        "lance database",
-                        "lance is cool",
-                    ])),
-                ],
-            )
-            .unwrap()]
-            .into_iter()
-            .map(Ok),
+        let data = RecordBatch::try_new(
             Arc::new(ArrowSchema::new(vec![
                 Field::new("id", DataType::Int32, false),
                 Field::new("text", DataType::Utf8, false),
             ])),
-        );
+            vec![
+                Arc::new(Int32Array::from(vec![1, 2, 3])),
+                Arc::new(StringArray::from(vec![
+                    "hello world",
+                    "lance database",
+                    "lance is cool",
+                ])),
+            ],
+        )
+        .unwrap();
 
-        let table = Arc::new(
-            db.create_table("docs", Box::new(data))
-                .execute()
-                .await
-                .unwrap(),
-        );
+        let table = Arc::new(db.create_table("docs", data).execute().await.unwrap());
 
         // Create FTS index with n-gram tokenizer (default min_ngram_length=3)
         table
@@ -1876,43 +1802,29 @@ mod tests {
             .unwrap();
 
         // Create table with two text columns
-        let data = RecordBatchIterator::new(
-            vec![RecordBatch::try_new(
-                Arc::new(ArrowSchema::new(vec![
-                    Field::new("id", DataType::Int32, false),
-                    Field::new("title", DataType::Utf8, false),
-                    Field::new("content", DataType::Utf8, false),
-                ])),
-                vec![
-                    Arc::new(Int32Array::from(vec![1, 2, 3])),
-                    Arc::new(StringArray::from(vec![
-                        "Important Document",
-                        "Another Document",
-                        "Random Text",
-                    ])),
-                    Arc::new(StringArray::from(vec![
-                        "This is important information",
-                        "This has details",
-                        "Nothing special here",
-                    ])),
-                ],
-            )
-            .unwrap()]
-            .into_iter()
-            .map(Ok),
+        let data = RecordBatch::try_new(
             Arc::new(ArrowSchema::new(vec![
                 Field::new("id", DataType::Int32, false),
                 Field::new("title", DataType::Utf8, false),
                 Field::new("content", DataType::Utf8, false),
             ])),
-        );
+            vec![
+                Arc::new(Int32Array::from(vec![1, 2, 3])),
+                Arc::new(StringArray::from(vec![
+                    "Important Document",
+                    "Another Document",
+                    "Random Text",
+                ])),
+                Arc::new(StringArray::from(vec![
+                    "This is important information",
+                    "This has details",
+                    "Nothing special here",
+                ])),
+            ],
+        )
+        .unwrap();
 
-        let table = Arc::new(
-            db.create_table("docs", Box::new(data))
-                .execute()
-                .await
-                .unwrap(),
-        );
+        let table = Arc::new(db.create_table("docs", data).execute().await.unwrap());
 
         // Create FTS indices on both columns
         table

rust/lancedb/src/table/dataset.rs

@@ -2,321 +2,501 @@
 // SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
 use std::{
-    ops::{Deref, DerefMut},
-    sync::Arc,
-    time::{self, Duration, Instant},
+    sync::{Arc, Mutex},
+    time::Duration,
 };
 
 use lance::{dataset::refs, Dataset};
-use tokio::sync::{RwLock, RwLockReadGuard, RwLockWriteGuard};
 
-use crate::error::Result;
-
-/// A wrapper around a [Dataset] that provides lazy-loading and consistency checks.
-///
-/// This can be cloned cheaply. It supports concurrent reads or exclusive writes.
-#[derive(Debug, Clone)]
-pub struct DatasetConsistencyWrapper(Arc<RwLock<DatasetRef>>);
+use crate::{error::Result, utils::background_cache::BackgroundCache, Error};
 
 /// A wrapper around a [Dataset] that provides consistency checks.
 ///
-/// The dataset is lazily loaded, and starts off as None. On the first access,
-/// the dataset is loaded.
+/// This can be cloned cheaply. Callers get an [`Arc<Dataset>`] from [`get()`](Self::get)
+/// and call [`update()`](Self::update) after writes to store the new version.
 #[derive(Debug, Clone)]
-enum DatasetRef {
-    /// In this mode, the dataset is always the latest version.
-    Latest {
-        dataset: Dataset,
-        read_consistency_interval: Option<Duration>,
-        last_consistency_check: Option<time::Instant>,
-    },
-    /// In this mode, the dataset is a specific version. It cannot be mutated.
-    TimeTravel { dataset: Dataset, version: u64 },
+pub struct DatasetConsistencyWrapper {
+    state: Arc<Mutex<DatasetState>>,
+    consistency: ConsistencyMode,
 }
 
-impl DatasetRef {
-    /// Reload the dataset to the appropriate version.
-    async fn reload(&mut self) -> Result<()> {
-        match self {
-            Self::Latest {
-                dataset,
-                last_consistency_check,
-                ..
-            } => {
-                dataset.checkout_latest().await?;
-                last_consistency_check.replace(Instant::now());
-            }
-            Self::TimeTravel { dataset, version } => {
-                dataset.checkout_version(*version).await?;
-            }
-        }
-        Ok(())
-    }
+/// The current dataset and whether it is pinned to a specific version.
+#[derive(Debug, Clone)]
+struct DatasetState {
+    dataset: Arc<Dataset>,
+    /// `Some(version)` = pinned to a specific version (time travel),
+    /// `None` = tracking latest.
+    pinned_version: Option<u64>,
+}
 
-    fn is_latest(&self) -> bool {
-        matches!(self, Self::Latest { .. })
-    }
-
-    async fn need_reload(&self) -> Result<bool> {
-        Ok(match self {
-            Self::Latest { dataset, .. } => {
-                dataset.latest_version_id().await? != dataset.version().version
-            }
-            Self::TimeTravel { dataset, version } => dataset.version().version != *version,
-        })
-    }
-
-    async fn as_latest(&mut self, read_consistency_interval: Option<Duration>) -> Result<()> {
-        match self {
-            Self::Latest { .. } => Ok(()),
-            Self::TimeTravel { dataset, .. } => {
-                dataset
-                    .checkout_version(dataset.latest_version_id().await?)
-                    .await?;
-                *self = Self::Latest {
-                    dataset: dataset.clone(),
-                    read_consistency_interval,
-                    last_consistency_check: Some(Instant::now()),
-                };
-                Ok(())
-            }
-        }
-    }
-
-    async fn as_time_travel(&mut self, target_version: impl Into<refs::Ref>) -> Result<()> {
-        let target_ref = target_version.into();
-
-        match self {
-            Self::Latest { dataset, .. } => {
-                let new_dataset = dataset.checkout_version(target_ref.clone()).await?;
-                let version_value = new_dataset.version().version;
-
-                *self = Self::TimeTravel {
-                    dataset: new_dataset,
-                    version: version_value,
-                };
-            }
-            Self::TimeTravel { dataset, version } => {
-                let should_checkout = match &target_ref {
-                    refs::Ref::Version(_, Some(target_ver)) => version != target_ver,
-                    refs::Ref::Version(_, None) => true, // No specific version, always checkout
-                    refs::Ref::VersionNumber(target_ver) => version != target_ver,
-                    refs::Ref::Tag(_) => true, // Always checkout for tags
-                };
-
-                if should_checkout {
-                    let new_dataset = dataset.checkout_version(target_ref).await?;
-                    let version_value = new_dataset.version().version;
-
-                    *self = Self::TimeTravel {
-                        dataset: new_dataset,
-                        version: version_value,
-                    };
-                }
-            }
-        }
-        Ok(())
-    }
-
-    fn time_travel_version(&self) -> Option<u64> {
-        match self {
-            Self::Latest { .. } => None,
-            Self::TimeTravel { version, .. } => Some(*version),
-        }
-    }
-
-    fn set_latest(&mut self, dataset: Dataset) {
-        match self {
-            Self::Latest {
-                dataset: ref mut ds,
-                ..
-            } => {
-                if dataset.manifest().version > ds.manifest().version {
-                    *ds = dataset;
-                }
-            }
-            _ => unreachable!("Dataset should be in latest mode at this point"),
-        }
-    }
+#[derive(Debug, Clone)]
+enum ConsistencyMode {
+    /// Only update table state when explicitly asked.
+    Lazy,
+    /// Always check for a new version on every read.
+    Strong,
+    /// Periodically check for new version in the background. If the table is being
+    /// regularly accessed, refresh will happen in the background. If the table is idle for a while,
+    /// the next access will trigger a refresh before returning the dataset.
+    ///
+    /// read_consistency_interval = TTL
+    /// refresh_window = min(3s, TTL/4)
+    ///
+    /// | t < TTL - refresh_window | t < TTL                           | t >= TTL            |
+    /// |  Return value            | Background refresh & return value |  syncronous refresh |
+    Eventual(BackgroundCache<Arc<Dataset>, Error>),
 }
 
 impl DatasetConsistencyWrapper {
     /// Create a new wrapper in the latest version mode.
     pub fn new_latest(dataset: Dataset, read_consistency_interval: Option<Duration>) -> Self {
-        Self(Arc::new(RwLock::new(DatasetRef::Latest {
-            dataset,
-            read_consistency_interval,
-            last_consistency_check: Some(Instant::now()),
-        })))
+        let dataset = Arc::new(dataset);
+        let consistency = match read_consistency_interval {
+            Some(d) if d == Duration::ZERO => ConsistencyMode::Strong,
+            Some(d) => {
+                let refresh_window = std::cmp::min(std::time::Duration::from_secs(3), d / 4);
+                let cache = BackgroundCache::new(d, refresh_window);
+                cache.seed(dataset.clone());
+                ConsistencyMode::Eventual(cache)
+            }
+            None => ConsistencyMode::Lazy,
+        };
+        Self {
+            state: Arc::new(Mutex::new(DatasetState {
+                dataset,
+                pinned_version: None,
+            })),
+            consistency,
+        }
     }
 
-    /// Get an immutable reference to the dataset.
-    pub async fn get(&self) -> Result<DatasetReadGuard<'_>> {
-        self.ensure_up_to_date().await?;
-        Ok(DatasetReadGuard {
-            guard: self.0.read().await,
-        })
-    }
-
-    /// Get a mutable reference to the dataset.
+    /// Get the current dataset.
     ///
-    /// If the dataset is in time travel mode this will fail
-    pub async fn get_mut(&self) -> Result<DatasetWriteGuard<'_>> {
-        self.ensure_mutable().await?;
-        self.ensure_up_to_date().await?;
-        Ok(DatasetWriteGuard {
-            guard: self.0.write().await,
-        })
-    }
-
-    /// Get a mutable reference to the dataset without requiring the
-    /// dataset to be in a Latest mode.
-    pub async fn get_mut_unchecked(&self) -> Result<DatasetWriteGuard<'_>> {
-        self.ensure_up_to_date().await?;
-        Ok(DatasetWriteGuard {
-            guard: self.0.write().await,
-        })
-    }
-
-    /// Convert into a wrapper in latest version mode
-    pub async fn as_latest(&self, read_consistency_interval: Option<Duration>) -> Result<()> {
-        if self.0.read().await.is_latest() {
-            return Ok(());
-        }
-
-        let mut write_guard = self.0.write().await;
-        if write_guard.is_latest() {
-            return Ok(());
-        }
-
-        write_guard.as_latest(read_consistency_interval).await
-    }
-
-    pub async fn as_time_travel(&self, target_version: impl Into<refs::Ref>) -> Result<()> {
-        self.0.write().await.as_time_travel(target_version).await
-    }
-
-    /// Provide a known latest version of the dataset.
+    /// Behavior depends on the consistency mode:
+    /// - **Lazy** (`None`): returns the cached dataset immediately.
+    /// - **Strong** (`Some(ZERO)`): checks for a new version before returning.
+    /// - **Eventual** (`Some(d)` where `d > 0`): returns a cached value immediately
+    ///   while refreshing in the background when the TTL expires.
     ///
-    /// This is usually done after some write operation, which inherently will
-    /// have the latest version.
-    pub async fn set_latest(&self, dataset: Dataset) {
-        self.0.write().await.set_latest(dataset);
-    }
-
-    pub async fn reload(&self) -> Result<()> {
-        if !self.0.read().await.need_reload().await? {
-            return Ok(());
+    /// If pinned to a specific version (time travel), always returns the
+    /// pinned dataset regardless of consistency mode.
+    pub async fn get(&self) -> Result<Arc<Dataset>> {
+        {
+            let state = self.state.lock().unwrap();
+            if state.pinned_version.is_some() {
+                return Ok(state.dataset.clone());
+            }
         }
 
-        let mut write_guard = self.0.write().await;
-        // on lock escalation -- check if someone else has already reloaded
-        if !write_guard.need_reload().await? {
-            return Ok(());
-        }
-
-        // actually need reloading
-        write_guard.reload().await
-    }
-
-    /// Returns the version, if in time travel mode, or None otherwise
-    pub async fn time_travel_version(&self) -> Option<u64> {
-        self.0.read().await.time_travel_version()
-    }
-
-    pub async fn ensure_mutable(&self) -> Result<()> {
-        let dataset_ref = self.0.read().await;
-        match &*dataset_ref {
-            DatasetRef::Latest { .. } => Ok(()),
-            DatasetRef::TimeTravel { .. } => Err(crate::Error::InvalidInput {
-                message: "table cannot be modified when a specific version is checked out"
-                    .to_string(),
-            }),
-        }
-    }
-
-    async fn is_up_to_date(&self) -> Result<bool> {
-        let dataset_ref = self.0.read().await;
-        match &*dataset_ref {
-            DatasetRef::Latest {
-                read_consistency_interval,
-                last_consistency_check,
-                ..
-            } => match (read_consistency_interval, last_consistency_check) {
-                (None, _) => Ok(true),
-                (Some(_), None) => Ok(false),
-                (Some(read_consistency_interval), Some(last_consistency_check)) => {
-                    if &last_consistency_check.elapsed() < read_consistency_interval {
-                        Ok(true)
-                    } else {
-                        Ok(false)
-                    }
+        match &self.consistency {
+            ConsistencyMode::Eventual(bg_cache) => {
+                if let Some(dataset) = bg_cache.try_get() {
+                    return Ok(dataset);
                 }
-            },
-            DatasetRef::TimeTravel { dataset, version } => {
-                Ok(dataset.version().version == *version)
+                let state = self.state.clone();
+                bg_cache
+                    .get(move || refresh_latest(state))
+                    .await
+                    .map_err(unwrap_shared_error)
+            }
+            ConsistencyMode::Strong => refresh_latest(self.state.clone()).await,
+            ConsistencyMode::Lazy => {
+                let state = self.state.lock().unwrap();
+                Ok(state.dataset.clone())
             }
         }
     }
 
-    /// Ensures that the dataset is loaded and up-to-date with consistency and
-    /// version parameters.
-    async fn ensure_up_to_date(&self) -> Result<()> {
-        if !self.is_up_to_date().await? {
-            self.reload().await?;
+    /// Store a new dataset version after a write operation.
+    ///
+    /// Only stores the dataset if its version is at least as new as the current one.
+    /// Same-version updates are accepted for operations like manifest path migration
+    /// that modify the dataset without creating a new version.
+    /// If the wrapper has since transitioned to time-travel mode (e.g. via a
+    /// concurrent [`as_time_travel`](Self::as_time_travel) call), the update
+    /// is silently ignored — the write already committed to storage.
+    pub fn update(&self, dataset: Dataset) {
+        let mut state = self.state.lock().unwrap();
+        if state.pinned_version.is_some() {
+            // A concurrent as_time_travel() beat us here. The write succeeded
+            // in storage, but since we're now pinned we don't advance the
+            // cached pointer.
+            return;
         }
+        if dataset.manifest().version >= state.dataset.manifest().version {
+            state.dataset = Arc::new(dataset);
+        }
+        drop(state);
+        if let ConsistencyMode::Eventual(bg_cache) = &self.consistency {
+            bg_cache.invalidate();
+        }
+    }
+
+    /// Checkout a branch and track its HEAD for new versions.
+    pub async fn as_branch(&self, _branch: impl Into<String>) -> Result<()> {
+        todo!("Branch support not yet implemented")
+    }
+
+    /// Check that the dataset is in a mutable mode (Latest).
+    pub fn ensure_mutable(&self) -> Result<()> {
+        let state = self.state.lock().unwrap();
+        if state.pinned_version.is_some() {
+            Err(crate::Error::InvalidInput {
+                message: "table cannot be modified when a specific version is checked out"
+                    .to_string(),
+            })
+        } else {
+            Ok(())
+        }
+    }
+
+    /// Returns the version, if in time travel mode, or None otherwise.
+    pub fn time_travel_version(&self) -> Option<u64> {
+        self.state.lock().unwrap().pinned_version
+    }
+
+    /// Convert into a wrapper in latest version mode.
+    pub async fn as_latest(&self) -> Result<()> {
+        let dataset = {
+            let state = self.state.lock().unwrap();
+            if state.pinned_version.is_none() {
+                return Ok(());
+            }
+            state.dataset.clone()
+        };
+
+        let latest_version = dataset.latest_version_id().await?;
+        let new_dataset = dataset.checkout_version(latest_version).await?;
+
+        let mut state = self.state.lock().unwrap();
+        if state.pinned_version.is_some() {
+            state.dataset = Arc::new(new_dataset);
+            state.pinned_version = None;
+        }
+        drop(state);
+        if let ConsistencyMode::Eventual(bg_cache) = &self.consistency {
+            bg_cache.invalidate();
+        }
+        Ok(())
+    }
+
+    pub async fn as_time_travel(&self, target_version: impl Into<refs::Ref>) -> Result<()> {
+        let target_ref = target_version.into();
+
+        let (should_checkout, dataset) = {
+            let state = self.state.lock().unwrap();
+            let should = match state.pinned_version {
+                None => true,
+                Some(version) => match &target_ref {
+                    refs::Ref::Version(_, Some(target_ver)) => version != *target_ver,
+                    refs::Ref::Version(_, None) => true,
+                    refs::Ref::VersionNumber(target_ver) => version != *target_ver,
+                    refs::Ref::Tag(_) => true,
+                },
+            };
+            (should, state.dataset.clone())
+        };
+
+        if !should_checkout {
+            return Ok(());
+        }
+
+        let new_dataset = dataset.checkout_version(target_ref).await?;
+        let version_value = new_dataset.version().version;
+
+        let mut state = self.state.lock().unwrap();
+        state.dataset = Arc::new(new_dataset);
+        state.pinned_version = Some(version_value);
+        Ok(())
+    }
+
+    pub async fn reload(&self) -> Result<()> {
+        let (dataset, pinned_version) = {
+            let state = self.state.lock().unwrap();
+            (state.dataset.clone(), state.pinned_version)
+        };
+
+        match pinned_version {
+            None => {
+                refresh_latest(self.state.clone()).await?;
+                if let ConsistencyMode::Eventual(bg_cache) = &self.consistency {
+                    bg_cache.invalidate();
+                }
+            }
+            Some(version) => {
+                if dataset.version().version == version {
+                    return Ok(());
+                }
+
+                let new_dataset = dataset.checkout_version(version).await?;
+
+                let mut state = self.state.lock().unwrap();
+                if state.pinned_version == Some(version) {
+                    state.dataset = Arc::new(new_dataset);
+                }
+            }
+        }
+
         Ok(())
     }
 }
 
-pub struct DatasetReadGuard<'a> {
-    guard: RwLockReadGuard<'a, DatasetRef>,
-}
+async fn refresh_latest(state: Arc<Mutex<DatasetState>>) -> Result<Arc<Dataset>> {
+    let dataset = { state.lock().unwrap().dataset.clone() };
 
-impl Deref for DatasetReadGuard<'_> {
-    type Target = Dataset;
+    let mut ds = (*dataset).clone();
+    ds.checkout_latest().await?;
+    let new_arc = Arc::new(ds);
 
-    fn deref(&self) -> &Self::Target {
-        match &*self.guard {
-            DatasetRef::Latest { dataset, .. } => dataset,
-            DatasetRef::TimeTravel { dataset, .. } => dataset,
+    {
+        let mut state = state.lock().unwrap();
+        if state.pinned_version.is_none()
+            && new_arc.manifest().version >= state.dataset.manifest().version
+        {
+            state.dataset = new_arc.clone();
         }
     }
+
+    Ok(new_arc)
 }
 
-pub struct DatasetWriteGuard<'a> {
-    guard: RwLockWriteGuard<'a, DatasetRef>,
-}
-
-impl Deref for DatasetWriteGuard<'_> {
-    type Target = Dataset;
-
-    fn deref(&self) -> &Self::Target {
-        match &*self.guard {
-            DatasetRef::Latest { dataset, .. } => dataset,
-            DatasetRef::TimeTravel { dataset, .. } => dataset,
-        }
-    }
-}
-
-impl DerefMut for DatasetWriteGuard<'_> {
-    fn deref_mut(&mut self) -> &mut Self::Target {
-        match &mut *self.guard {
-            DatasetRef::Latest { dataset, .. } => dataset,
-            DatasetRef::TimeTravel { dataset, .. } => dataset,
-        }
+fn unwrap_shared_error(arc: Arc<Error>) -> Error {
+    match Arc::try_unwrap(arc) {
+        Ok(err) => err,
+        Err(arc) => Error::Runtime {
+            message: arc.to_string(),
+        },
     }
 }
 
 #[cfg(test)]
 mod tests {
+    use std::time::Instant;
+
+    use arrow_array::{Int32Array, RecordBatch, RecordBatchIterator};
     use arrow_schema::{DataType, Field, Schema};
-    use lance::{dataset::WriteParams, io::ObjectStoreParams};
+    use lance::{
+        dataset::{WriteMode, WriteParams},
+        io::ObjectStoreParams,
+    };
 
     use super::*;
 
     use crate::{connect, io::object_store::io_tracking::IoStatsHolder, table::WriteOptions};
 
+    async fn create_test_dataset(uri: &str) -> Dataset {
+        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));
+        let batch = RecordBatch::try_new(
+            schema.clone(),
+            vec![Arc::new(Int32Array::from(vec![1, 2, 3]))],
+        )
+        .unwrap();
+        Dataset::write(
+            RecordBatchIterator::new(vec![Ok(batch)], schema),
+            uri,
+            Some(WriteParams::default()),
+        )
+        .await
+        .unwrap()
+    }
+
+    async fn append_to_dataset(uri: &str) -> Dataset {
+        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));
+        let batch = RecordBatch::try_new(
+            schema.clone(),
+            vec![Arc::new(Int32Array::from(vec![4, 5, 6]))],
+        )
+        .unwrap();
+        Dataset::write(
+            RecordBatchIterator::new(vec![Ok(batch)], schema),
+            uri,
+            Some(WriteParams {
+                mode: WriteMode::Append,
+                ..Default::default()
+            }),
+        )
+        .await
+        .unwrap()
+    }
+
+    #[tokio::test]
+    async fn test_get_returns_dataset() {
+        let dir = tempfile::tempdir().unwrap();
+        let uri = dir.path().to_str().unwrap();
+        let ds = create_test_dataset(uri).await;
+        let version = ds.version().version;
+
+        let wrapper = DatasetConsistencyWrapper::new_latest(ds, None);
+        let ds1 = wrapper.get().await.unwrap();
+        let ds2 = wrapper.get().await.unwrap();
+
+        assert_eq!(ds1.version().version, version);
+        assert_eq!(ds2.version().version, version);
+
+        // Arc<Dataset> is independent — not borrowing from wrapper
+        drop(wrapper);
+        assert_eq!(ds1.version().version, version);
+    }
+
+    #[tokio::test]
+    async fn test_update_stores_newer_version() {
+        let dir = tempfile::tempdir().unwrap();
+        let uri = dir.path().to_str().unwrap();
+        let ds_v1 = create_test_dataset(uri).await;
+        assert_eq!(ds_v1.version().version, 1);
+
+        let wrapper = DatasetConsistencyWrapper::new_latest(ds_v1, None);
+
+        let ds_v2 = append_to_dataset(uri).await;
+        assert_eq!(ds_v2.version().version, 2);
+
+        wrapper.update(ds_v2);
+
+        let ds = wrapper.get().await.unwrap();
+        assert_eq!(ds.version().version, 2);
+    }
+
+    #[tokio::test]
+    async fn test_update_ignores_older_version() {
+        let dir = tempfile::tempdir().unwrap();
+        let uri = dir.path().to_str().unwrap();
+        let ds_v1 = create_test_dataset(uri).await;
+        let ds_v2 = append_to_dataset(uri).await;
+
+        let wrapper = DatasetConsistencyWrapper::new_latest(ds_v2, None);
+        wrapper.update(ds_v1);
+
+        let ds = wrapper.get().await.unwrap();
+        assert_eq!(ds.version().version, 2);
+    }
+
+    #[tokio::test]
+    async fn test_ensure_mutable_allows_latest() {
+        let dir = tempfile::tempdir().unwrap();
+        let uri = dir.path().to_str().unwrap();
+        let ds = create_test_dataset(uri).await;
+
+        let wrapper = DatasetConsistencyWrapper::new_latest(ds, None);
+        assert!(wrapper.ensure_mutable().is_ok());
+    }
+
+    #[tokio::test]
+    async fn test_ensure_mutable_rejects_time_travel() {
+        let dir = tempfile::tempdir().unwrap();
+        let uri = dir.path().to_str().unwrap();
+        let ds = create_test_dataset(uri).await;
+
+        let wrapper = DatasetConsistencyWrapper::new_latest(ds, None);
+        wrapper.as_time_travel(1u64).await.unwrap();
+
+        assert!(wrapper.ensure_mutable().is_err());
+    }
+
+    #[tokio::test]
+    async fn test_time_travel_version() {
+        let dir = tempfile::tempdir().unwrap();
+        let uri = dir.path().to_str().unwrap();
+        let ds = create_test_dataset(uri).await;
+
+        let wrapper = DatasetConsistencyWrapper::new_latest(ds, None);
+        assert_eq!(wrapper.time_travel_version(), None);
+
+        wrapper.as_time_travel(1u64).await.unwrap();
+        assert_eq!(wrapper.time_travel_version(), Some(1));
+    }
+
+    #[tokio::test]
+    async fn test_as_latest_from_time_travel() {
+        let dir = tempfile::tempdir().unwrap();
+        let uri = dir.path().to_str().unwrap();
+        let ds = create_test_dataset(uri).await;
+
+        let wrapper = DatasetConsistencyWrapper::new_latest(ds, None);
+        wrapper.as_time_travel(1u64).await.unwrap();
+        assert!(wrapper.ensure_mutable().is_err());
+
+        wrapper.as_latest().await.unwrap();
+        assert!(wrapper.ensure_mutable().is_ok());
+        assert_eq!(wrapper.time_travel_version(), None);
+    }
+
+    #[tokio::test]
+    async fn test_lazy_consistency_never_refreshes() {
+        let dir = tempfile::tempdir().unwrap();
+        let uri = dir.path().to_str().unwrap();
+        let ds = create_test_dataset(uri).await;
+
+        let wrapper = DatasetConsistencyWrapper::new_latest(ds, None);
+        let v1 = wrapper.get().await.unwrap().version().version;
+
+        // External write
+        append_to_dataset(uri).await;
+
+        // Lazy consistency should not pick up external write
+        let v_after = wrapper.get().await.unwrap().version().version;
+        assert_eq!(v1, v_after);
+    }
+
+    #[tokio::test]
+    async fn test_strong_consistency_always_refreshes() {
+        let dir = tempfile::tempdir().unwrap();
+        let uri = dir.path().to_str().unwrap();
+        let ds = create_test_dataset(uri).await;
+
+        let wrapper = DatasetConsistencyWrapper::new_latest(ds, Some(Duration::ZERO));
+        let v1 = wrapper.get().await.unwrap().version().version;
+
+        // External write
+        append_to_dataset(uri).await;
+
+        // Strong consistency should pick up external write
+        let v_after = wrapper.get().await.unwrap().version().version;
+        assert_eq!(v_after, v1 + 1);
+    }
+
+    #[tokio::test]
+    async fn test_eventual_consistency_background_refresh() {
+        let dir = tempfile::tempdir().unwrap();
+        let uri = dir.path().to_str().unwrap();
+        let ds = create_test_dataset(uri).await;
+
+        let wrapper = DatasetConsistencyWrapper::new_latest(ds, Some(Duration::from_millis(200)));
+
+        // Populate the cache
+        let v1 = wrapper.get().await.unwrap().version().version;
+        assert_eq!(v1, 1);
+
+        // External write
+        append_to_dataset(uri).await;
+
+        // Should return cached value immediately (within TTL)
+        let v_cached = wrapper.get().await.unwrap().version().version;
+        assert_eq!(v_cached, 1);
+
+        // Wait for TTL to expire, then get() should trigger a refresh
+        tokio::time::sleep(Duration::from_millis(300)).await;
+        let v_after = wrapper.get().await.unwrap().version().version;
+        assert_eq!(v_after, 2);
+    }
+
+    #[tokio::test]
+    async fn test_eventual_consistency_update_invalidates_cache() {
+        let dir = tempfile::tempdir().unwrap();
+        let uri = dir.path().to_str().unwrap();
+        let ds_v1 = create_test_dataset(uri).await;
+
+        let wrapper = DatasetConsistencyWrapper::new_latest(ds_v1, Some(Duration::from_secs(60)));
+
+        // Simulate a write that produces v2
+        let ds_v2 = append_to_dataset(uri).await;
+        wrapper.update(ds_v2);
+
+        // get() should return v2 immediately (update invalidated the bg_cache,
+        // and the mutex state was updated)
+        let v = wrapper.get().await.unwrap().version().version;
+        assert_eq!(v, 2);
+    }
+
     #[tokio::test]
     async fn test_iops_open_strong_consistency() {
         let db = connect("memory://")
@@ -332,7 +512,7 @@ mod tests {
             .create_empty_table("test", schema)
             .write_options(WriteOptions {
                 lance_write_params: Some(WriteParams {
-                    store_params: Some(ObjectStoreParams {
+                    store_params: Some(lance::io::ObjectStoreParams {
                         object_store_wrapper: Some(Arc::new(io_stats.clone())),
                         ..Default::default()
                     }),
@@ -351,4 +531,85 @@ mod tests {
         let stats = io_stats.incremental_stats();
         assert_eq!(stats.read_iops, 1);
     }
+
+    /// Regression test: a write that races with as_time_travel() must not panic.
+    ///
+    /// Sequence: ensure_mutable() passes → as_time_travel() completes → write
+    /// calls update().  Previously the assert!() in update() would fire.
+    #[tokio::test]
+    async fn test_update_after_concurrent_time_travel_does_not_panic() {
+        let dir = tempfile::tempdir().unwrap();
+        let uri = dir.path().to_str().unwrap();
+        let ds_v1 = create_test_dataset(uri).await;
+
+        let wrapper = DatasetConsistencyWrapper::new_latest(ds_v1, None);
+
+        // Simulate: as_time_travel() completes just before the write's update().
+        wrapper.as_time_travel(1u64).await.unwrap();
+        assert_eq!(wrapper.time_travel_version(), Some(1));
+
+        // The write already committed to storage; now it calls update().
+        // This must not panic, and the wrapper must stay pinned.
+        let ds_v2 = append_to_dataset(uri).await;
+        wrapper.update(ds_v2);
+
+        let ds = wrapper.get().await.unwrap();
+        assert_eq!(ds.version().version, 1);
+    }
+
+    /// Regression test: before the fix, the reload fast-path (no version change)
+    /// did not reset `last_consistency_check`, causing a list call on every
+    /// subsequent query once the interval expired.
+    #[tokio::test]
+    async fn test_reload_resets_consistency_timer() {
+        let db = connect("memory://")
+            .read_consistency_interval(Duration::from_secs(1))
+            .execute()
+            .await
+            .unwrap();
+        let io_stats = IoStatsHolder::default();
+        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));
+        let table = db
+            .create_empty_table("test", schema)
+            .write_options(WriteOptions {
+                lance_write_params: Some(WriteParams {
+                    store_params: Some(ObjectStoreParams {
+                        object_store_wrapper: Some(Arc::new(io_stats.clone())),
+                        ..Default::default()
+                    }),
+                    ..Default::default()
+                }),
+            })
+            .execute()
+            .await
+            .unwrap();
+
+        let start = Instant::now();
+        io_stats.incremental_stats(); // reset
+
+        // Step 1: within interval — no list
+        table.schema().await.unwrap();
+        let s = io_stats.incremental_stats();
+        assert_eq!(s.read_iops, 0, "step 1, elapsed={:?}", start.elapsed());
+
+        // Step 2: still within interval — no list
+        table.schema().await.unwrap();
+        let s = io_stats.incremental_stats();
+        assert_eq!(s.read_iops, 0, "step 2, elapsed={:?}", start.elapsed());
+
+        // Step 3: sleep past the 1s boundary
+        tokio::time::sleep(Duration::from_secs(1)).await;
+
+        // Step 4: interval expired — exactly 1 list, timer resets
+        table.schema().await.unwrap();
+        let s = io_stats.incremental_stats();
+        assert_eq!(s.read_iops, 1, "step 4, elapsed={:?}", start.elapsed());
+
+        // Step 5: 10 more calls — timer just reset, no lists (THIS is the regression test).
+        for _ in 0..10 {
+            table.schema().await.unwrap();
+        }
+        let s = io_stats.incremental_stats();
+        assert_eq!(s.read_iops, 0, "step 5, elapsed={:?}", start.elapsed());
+    }
 }

rust/lancedb/src/table/delete.rs

@@ -18,23 +18,18 @@ pub struct DeleteResult {
 ///
 /// This logic was moved from NativeTable::delete to keep table.rs clean.
 pub(crate) async fn execute_delete(table: &NativeTable, predicate: &str) -> Result<DeleteResult> {
-    // We access the dataset from the table. Since this is in the same module hierarchy (super),
-    // and 'dataset' is pub(crate), we can access it.
-    let mut dataset = table.dataset.get_mut().await?;
-
-    // Perform the actual delete on the Lance dataset
+    table.dataset.ensure_mutable()?;
+    let mut dataset = (*table.dataset.get().await?).clone();
     dataset.delete(predicate).await?;
-
-    // Return the result with the new version
-    Ok(DeleteResult {
-        version: dataset.version().version,
-    })
+    let version = dataset.version().version;
+    table.dataset.update(dataset);
+    Ok(DeleteResult { version })
 }
 
 #[cfg(test)]
 mod tests {
     use crate::connect;
-    use arrow_array::{record_batch, Int32Array, RecordBatch, RecordBatchIterator};
+    use arrow_array::{record_batch, Int32Array, RecordBatch};
     use arrow_schema::{DataType, Field, Schema};
     use std::sync::Arc;
 
@@ -53,10 +48,7 @@ mod tests {
         .unwrap();
 
         let table = conn
-            .create_table(
-                "test_delete",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_delete", batch)
             .execute()
             .await
             .unwrap();
@@ -102,10 +94,7 @@ mod tests {
         let original_schema = batch.schema();
 
         let table = conn
-            .create_table(
-                "test_delete_all",
-                RecordBatchIterator::new(vec![Ok(batch)], original_schema.clone()),
-            )
+            .create_table("test_delete_all", batch)
             .execute()
             .await
             .unwrap();
@@ -126,13 +115,8 @@ mod tests {
         // Create a table with 5 rows
         let batch = record_batch!(("id", Int32, [1, 2, 3, 4, 5])).unwrap();
 
-        let schema = batch.schema();
-
         let table = conn
-            .create_table(
-                "test_delete_noop",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_delete_noop", batch)
             .execute()
             .await
             .unwrap();

rust/lancedb/src/table/merge.rs

@@ -1,13 +1,45 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
-use std::{sync::Arc, time::Duration};
+use std::sync::Arc;
+use std::time::Duration;
 
 use arrow_array::RecordBatchReader;
+use futures::future::Either;
+use futures::{FutureExt, TryFutureExt};
+use lance::dataset::{
+    MergeInsertBuilder as LanceMergeInsertBuilder, WhenMatched, WhenNotMatchedBySource,
+};
+use serde::{Deserialize, Serialize};
 
-use crate::Result;
+use crate::error::{Error, Result};
 
-use super::{BaseTable, MergeResult};
+use super::{BaseTable, NativeTable};
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, Default)]
+pub struct MergeResult {
+    // The commit version associated with the operation.
+    // A version of `0` indicates compatibility with legacy servers that do not return
+    /// a commit version.
+    #[serde(default)]
+    pub version: u64,
+    /// Number of inserted rows (for user statistics)
+    #[serde(default)]
+    pub num_inserted_rows: u64,
+    /// Number of updated rows (for user statistics)
+    #[serde(default)]
+    pub num_updated_rows: u64,
+    /// Number of deleted rows (for user statistics)
+    /// Note: This is different from internal references to 'deleted_rows', since we technically "delete" updated rows during processing.
+    /// However those rows are not shared with the user.
+    #[serde(default)]
+    pub num_deleted_rows: u64,
+    /// Number of attempts performed during the merge operation.
+    /// This includes the initial attempt plus any retries due to transaction conflicts.
+    /// A value of 1 means the operation succeeded on the first try.
+    #[serde(default)]
+    pub num_attempts: u32,
+}
 
 /// A builder used to create and run a merge insert operation
 ///
@@ -124,3 +156,172 @@ impl MergeInsertBuilder {
         self.table.clone().merge_insert(self, new_data).await
     }
 }
+
+/// Internal implementation of the merge insert logic
+///
+/// This logic was moved from NativeTable::merge_insert to keep table.rs clean.
+pub(crate) async fn execute_merge_insert(
+    table: &NativeTable,
+    params: MergeInsertBuilder,
+    new_data: Box<dyn RecordBatchReader + Send>,
+) -> Result<MergeResult> {
+    let dataset = table.dataset.get().await?;
+    let mut builder = LanceMergeInsertBuilder::try_new(dataset.clone(), params.on)?;
+    match (
+        params.when_matched_update_all,
+        params.when_matched_update_all_filt,
+    ) {
+        (false, _) => builder.when_matched(WhenMatched::DoNothing),
+        (true, None) => builder.when_matched(WhenMatched::UpdateAll),
+        (true, Some(filt)) => builder.when_matched(WhenMatched::update_if(&dataset, &filt)?),
+    };
+    if params.when_not_matched_insert_all {
+        builder.when_not_matched(lance::dataset::WhenNotMatched::InsertAll);
+    } else {
+        builder.when_not_matched(lance::dataset::WhenNotMatched::DoNothing);
+    }
+    if params.when_not_matched_by_source_delete {
+        let behavior = if let Some(filter) = params.when_not_matched_by_source_delete_filt {
+            WhenNotMatchedBySource::delete_if(dataset.as_ref(), &filter)?
+        } else {
+            WhenNotMatchedBySource::Delete
+        };
+        builder.when_not_matched_by_source(behavior);
+    } else {
+        builder.when_not_matched_by_source(WhenNotMatchedBySource::Keep);
+    }
+    builder.use_index(params.use_index);
+
+    let future = if let Some(timeout) = params.timeout {
+        let future = builder
+            .retry_timeout(timeout)
+            .try_build()?
+            .execute_reader(new_data);
+        Either::Left(tokio::time::timeout(timeout, future).map(|res| match res {
+            Ok(Ok((new_dataset, stats))) => Ok((new_dataset, stats)),
+            Ok(Err(e)) => Err(e.into()),
+            Err(_) => Err(Error::Runtime {
+                message: "merge insert timed out".to_string(),
+            }),
+        }))
+    } else {
+        let job = builder.try_build()?;
+        Either::Right(job.execute_reader(new_data).map_err(|e| e.into()))
+    };
+    let (new_dataset, stats) = future.await?;
+    let version = new_dataset.manifest().version;
+    table.dataset.update(new_dataset.as_ref().clone());
+    Ok(MergeResult {
+        version,
+        num_updated_rows: stats.num_updated_rows,
+        num_inserted_rows: stats.num_inserted_rows,
+        num_deleted_rows: stats.num_deleted_rows,
+        num_attempts: stats.num_attempts,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use arrow_array::{Int32Array, RecordBatch, RecordBatchIterator, RecordBatchReader};
+    use arrow_schema::{DataType, Field, Schema};
+    use std::sync::Arc;
+
+    use crate::connect;
+
+    fn merge_insert_test_batches(offset: i32, age: i32) -> Box<dyn RecordBatchReader + Send> {
+        let schema = Arc::new(Schema::new(vec![
+            Field::new("i", DataType::Int32, false),
+            Field::new("age", DataType::Int32, false),
+        ]));
+        let batch = RecordBatch::try_new(
+            schema.clone(),
+            vec![
+                Arc::new(Int32Array::from_iter_values(offset..(offset + 10))),
+                Arc::new(Int32Array::from_iter_values(std::iter::repeat_n(age, 10))),
+            ],
+        )
+        .unwrap();
+        Box::new(RecordBatchIterator::new(vec![Ok(batch)], schema))
+    }
+
+    #[tokio::test]
+    async fn test_merge_insert() {
+        let conn = connect("memory://").execute().await.unwrap();
+
+        // Create a dataset with i=0..10
+        let batches = merge_insert_test_batches(0, 0);
+        let table = conn
+            .create_table("my_table", batches)
+            .execute()
+            .await
+            .unwrap();
+        assert_eq!(table.count_rows(None).await.unwrap(), 10);
+
+        // Create new data with i=5..15
+        let new_batches = merge_insert_test_batches(5, 1);
+
+        // Perform a "insert if not exists"
+        let mut merge_insert_builder = table.merge_insert(&["i"]);
+        merge_insert_builder.when_not_matched_insert_all();
+        let result = merge_insert_builder.execute(new_batches).await.unwrap();
+        // Only 5 rows should actually be inserted
+        assert_eq!(table.count_rows(None).await.unwrap(), 15);
+        assert_eq!(result.num_inserted_rows, 5);
+        assert_eq!(result.num_updated_rows, 0);
+        assert_eq!(result.num_deleted_rows, 0);
+        assert_eq!(result.num_attempts, 1);
+
+        // Create new data with i=15..25 (no id matches)
+        let new_batches = merge_insert_test_batches(15, 2);
+        // Perform a "bulk update" (should not affect anything)
+        let mut merge_insert_builder = table.merge_insert(&["i"]);
+        merge_insert_builder.when_matched_update_all(None);
+        merge_insert_builder.execute(new_batches).await.unwrap();
+        // No new rows should have been inserted
+        assert_eq!(table.count_rows(None).await.unwrap(), 15);
+        assert_eq!(
+            table.count_rows(Some("age = 2".to_string())).await.unwrap(),
+            0
+        );
+
+        // Conditional update that only replaces the age=0 data
+        let new_batches = merge_insert_test_batches(5, 3);
+        let mut merge_insert_builder = table.merge_insert(&["i"]);
+        merge_insert_builder.when_matched_update_all(Some("target.age = 0".to_string()));
+        merge_insert_builder.execute(new_batches).await.unwrap();
+        assert_eq!(
+            table.count_rows(Some("age = 3".to_string())).await.unwrap(),
+            5
+        );
+    }
+
+    #[tokio::test]
+    async fn test_merge_insert_use_index() {
+        let conn = connect("memory://").execute().await.unwrap();
+
+        // Create a dataset with i=0..10
+        let batches = merge_insert_test_batches(0, 0);
+        let table = conn
+            .create_table("my_table", batches)
+            .execute()
+            .await
+            .unwrap();
+        assert_eq!(table.count_rows(None).await.unwrap(), 10);
+
+        // Test use_index=true (default behavior)
+        let new_batches = merge_insert_test_batches(5, 1);
+        let mut merge_insert_builder = table.merge_insert(&["i"]);
+        merge_insert_builder.when_not_matched_insert_all();
+        merge_insert_builder.use_index(true);
+        merge_insert_builder.execute(new_batches).await.unwrap();
+        assert_eq!(table.count_rows(None).await.unwrap(), 15);
+
+        // Test use_index=false (force table scan)
+        let new_batches = merge_insert_test_batches(15, 2);
+        let mut merge_insert_builder = table.merge_insert(&["i"]);
+        merge_insert_builder.when_not_matched_insert_all();
+        merge_insert_builder.use_index(false);
+        merge_insert_builder.execute(new_batches).await.unwrap();
+        assert_eq!(table.count_rows(None).await.unwrap(), 25);
+    }
+}

rust/lancedb/src/table/optimize.rs

@@ -26,8 +26,10 @@ use crate::error::Result;
 /// optimize different parts of the table on disk.
 ///
 /// By default, it optimizes everything, as [`OptimizeAction::All`].
+#[derive(Default)]
 pub enum OptimizeAction {
     /// Run all optimizations with default values
+    #[default]
     All,
     /// Compacts files in the dataset
     ///
@@ -84,12 +86,6 @@ pub enum OptimizeAction {
     Index(OptimizeOptions),
 }
 
-impl Default for OptimizeAction {
-    fn default() -> Self {
-        Self::All
-    }
-}
-
 /// Statistics about the optimization.
 #[derive(Debug, Default)]
 pub struct OptimizeStats {
@@ -105,12 +101,10 @@ pub struct OptimizeStats {
 /// This logic was moved from NativeTable to keep table.rs clean.
 pub(crate) async fn optimize_indices(table: &NativeTable, options: &OptimizeOptions) -> Result<()> {
     info!("LanceDB: optimizing indices: {:?}", options);
-    table
-        .dataset
-        .get_mut()
-        .await?
-        .optimize_indices(options)
-        .await?;
+    table.dataset.ensure_mutable()?;
+    let mut dataset = (*table.dataset.get().await?).clone();
+    dataset.optimize_indices(options).await?;
+    table.dataset.update(dataset);
     Ok(())
 }
 
@@ -131,10 +125,9 @@ pub(crate) async fn cleanup_old_versions(
     delete_unverified: Option<bool>,
     error_if_tagged_old_versions: Option<bool>,
 ) -> Result<RemovalStats> {
-    Ok(table
-        .dataset
-        .get_mut()
-        .await?
+    table.dataset.ensure_mutable()?;
+    let dataset = table.dataset.get().await?;
+    Ok(dataset
         .cleanup_old_versions(older_than, delete_unverified, error_if_tagged_old_versions)
         .await?)
 }
@@ -150,8 +143,10 @@ pub(crate) async fn compact_files_impl(
     options: CompactionOptions,
     remap_options: Option<Arc<dyn IndexRemapperOptions>>,
 ) -> Result<CompactionMetrics> {
-    let mut dataset_mut = table.dataset.get_mut().await?;
-    let metrics = compact_files(&mut dataset_mut, options, remap_options).await?;
+    table.dataset.ensure_mutable()?;
+    let mut dataset = (*table.dataset.get().await?).clone();
+    let metrics = compact_files(&mut dataset, options, remap_options).await?;
+    table.dataset.update(dataset);
     Ok(metrics)
 }
 
@@ -212,7 +207,7 @@ pub(crate) async fn execute_optimize(
 
 #[cfg(test)]
 mod tests {
-    use arrow_array::{Int32Array, RecordBatch, RecordBatchIterator, StringArray};
+    use arrow_array::{Int32Array, RecordBatch, StringArray};
     use arrow_schema::{DataType, Field, Schema};
     use rstest::rstest;
     use std::sync::Arc;
@@ -236,10 +231,7 @@ mod tests {
         .unwrap();
 
         let table = conn
-            .create_table(
-                "test_compact",
-                RecordBatchIterator::new(vec![Ok(batch)], schema.clone()),
-            )
+            .create_table("test_compact", batch)
             .execute()
             .await
             .unwrap();
@@ -253,11 +245,7 @@ mod tests {
                 ))],
             )
             .unwrap();
-            table
-                .add(RecordBatchIterator::new(vec![Ok(batch)], schema.clone()))
-                .execute()
-                .await
-                .unwrap();
+            table.add(batch).execute().await.unwrap();
         }
 
         // Verify we have multiple fragments before compaction
@@ -322,10 +310,7 @@ mod tests {
         .unwrap();
 
         let table = conn
-            .create_table(
-                "test_prune",
-                RecordBatchIterator::new(vec![Ok(batch)], schema.clone()),
-            )
+            .create_table("test_prune", batch)
             .execute()
             .await
             .unwrap();
@@ -339,11 +324,7 @@ mod tests {
                 ))],
             )
             .unwrap();
-            table
-                .add(RecordBatchIterator::new(vec![Ok(batch)], schema.clone()))
-                .execute()
-                .await
-                .unwrap();
+            table.add(batch).execute().await.unwrap();
         }
 
         // Verify multiple versions exist
@@ -405,10 +386,7 @@ mod tests {
         .unwrap();
 
         let table = conn
-            .create_table(
-                "test_index_optimize",
-                RecordBatchIterator::new(vec![Ok(batch)], schema.clone()),
-            )
+            .create_table("test_index_optimize", batch)
             .execute()
             .await
             .unwrap();
@@ -426,11 +404,7 @@ mod tests {
             vec![Arc::new(Int32Array::from_iter_values(100..200))],
         )
         .unwrap();
-        table
-            .add(RecordBatchIterator::new(vec![Ok(batch)], schema.clone()))
-            .execute()
-            .await
-            .unwrap();
+        table.add(batch).execute().await.unwrap();
 
         // Verify index stats before optimization
         let indices = table.list_indices().await.unwrap();
@@ -474,10 +448,7 @@ mod tests {
         .unwrap();
 
         let table = conn
-            .create_table(
-                "test_optimize_all",
-                RecordBatchIterator::new(vec![Ok(batch)], schema.clone()),
-            )
+            .create_table("test_optimize_all", batch)
             .execute()
             .await
             .unwrap();
@@ -491,11 +462,7 @@ mod tests {
                 ))],
             )
             .unwrap();
-            table
-                .add(RecordBatchIterator::new(vec![Ok(batch)], schema.clone()))
-                .execute()
-                .await
-                .unwrap();
+            table.add(batch).execute().await.unwrap();
         }
 
         // Run all optimizations
@@ -559,20 +526,13 @@ mod tests {
         .unwrap();
 
         let table = conn
-            .create_table(
-                "test_deferred_remap",
-                RecordBatchIterator::new(vec![Ok(batch.clone())], schema.clone()),
-            )
+            .create_table("test_deferred_remap", batch.clone())
             .execute()
             .await
             .unwrap();
 
         // Add more data
-        table
-            .add(RecordBatchIterator::new(vec![Ok(batch)], schema.clone()))
-            .execute()
-            .await
-            .unwrap();
+        table.add(batch).execute().await.unwrap();
 
         // Create an index
         table
@@ -648,20 +608,13 @@ mod tests {
         let original_schema = batch.schema();
 
         let table = conn
-            .create_table(
-                "test_schema_preserved",
-                RecordBatchIterator::new(vec![Ok(batch.clone())], schema.clone()),
-            )
+            .create_table("test_schema_preserved", batch.clone())
             .execute()
             .await
             .unwrap();
 
         // Add more data
-        table
-            .add(RecordBatchIterator::new(vec![Ok(batch)], schema.clone()))
-            .execute()
-            .await
-            .unwrap();
+        table.add(batch).execute().await.unwrap();
 
         // Run compaction
         table
@@ -703,10 +656,7 @@ mod tests {
         .unwrap();
 
         let table = conn
-            .create_table(
-                "test_empty_optimize",
-                RecordBatchIterator::new(vec![Ok(batch)], schema.clone()),
-            )
+            .create_table("test_empty_optimize", batch)
             .execute()
             .await
             .unwrap();
@@ -752,19 +702,12 @@ mod tests {
         .unwrap();
 
         let table = conn
-            .create_table(
-                "test_checkout_optimize",
-                RecordBatchIterator::new(vec![Ok(batch.clone())], schema.clone()),
-            )
+            .create_table("test_checkout_optimize", batch.clone())
             .execute()
             .await
             .unwrap();
 
-        table
-            .add(RecordBatchIterator::new(vec![Ok(batch)], schema.clone()))
-            .execute()
-            .await
-            .unwrap();
+        table.add(batch).execute().await.unwrap();
 
         table.checkout(1).await.unwrap();
 

rust/lancedb/src/table/query.rs

@@ -0,0 +1,738 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+use std::sync::Arc;
+
+use super::NativeTable;
+use crate::error::{Error, Result};
+use crate::expr::expr_to_sql_string;
+use crate::query::{
+    QueryExecutionOptions, QueryFilter, QueryRequest, Select, VectorQueryRequest, DEFAULT_TOP_K,
+};
+use crate::utils::{default_vector_column, TimeoutStream};
+use arrow::array::{AsArray, FixedSizeListBuilder, Float32Builder};
+use arrow::datatypes::{Float32Type, UInt8Type};
+use arrow_array::Array;
+use arrow_schema::{DataType, Schema};
+use datafusion_physical_plan::projection::ProjectionExec;
+use datafusion_physical_plan::repartition::RepartitionExec;
+use datafusion_physical_plan::union::UnionExec;
+use datafusion_physical_plan::ExecutionPlan;
+use futures::future::try_join_all;
+use lance::dataset::scanner::DatasetRecordBatchStream;
+use lance::dataset::scanner::Scanner;
+use lance_datafusion::exec::{analyze_plan as lance_analyze_plan, execute_plan};
+use lance_namespace::models::{
+    QueryTableRequest as NsQueryTableRequest, QueryTableRequestColumns,
+    QueryTableRequestFullTextQuery, QueryTableRequestVector, StringFtsQuery,
+};
+use lance_namespace::LanceNamespace;
+
+#[derive(Debug, Clone)]
+pub enum AnyQuery {
+    Query(QueryRequest),
+    VectorQuery(VectorQueryRequest),
+}
+
+//Decide between namespace or local
+pub async fn execute_query(
+    table: &NativeTable,
+    query: &AnyQuery,
+    options: QueryExecutionOptions,
+) -> Result<DatasetRecordBatchStream> {
+    // If namespace client is configured, use server-side query execution
+    if let Some(ref namespace_client) = table.namespace_client {
+        return execute_namespace_query(table, namespace_client.clone(), query, options).await;
+    }
+    execute_generic_query(table, query, options).await
+}
+
+pub async fn analyze_query_plan(
+    table: &NativeTable,
+    query: &AnyQuery,
+    options: QueryExecutionOptions,
+) -> Result<String> {
+    let plan = create_plan(table, query, options).await?;
+    Ok(lance_analyze_plan(plan, Default::default()).await?)
+}
+
+/// Local Execution Path (DataFusion)
+async fn execute_generic_query(
+    table: &NativeTable,
+    query: &AnyQuery,
+    options: QueryExecutionOptions,
+) -> Result<DatasetRecordBatchStream> {
+    let plan = create_plan(table, query, options.clone()).await?;
+    let inner = execute_plan(plan, Default::default())?;
+    let inner = if let Some(timeout) = options.timeout {
+        TimeoutStream::new_boxed(inner, timeout)
+    } else {
+        inner
+    };
+    Ok(DatasetRecordBatchStream::new(inner))
+}
+
+pub async fn create_plan(
+    table: &NativeTable,
+    query: &AnyQuery,
+    options: QueryExecutionOptions,
+) -> Result<Arc<dyn ExecutionPlan>> {
+    let query = match query {
+        AnyQuery::VectorQuery(query) => query.clone(),
+        AnyQuery::Query(query) => VectorQueryRequest::from_plain_query(query.clone()),
+    };
+
+    let ds_ref = table.dataset.get().await?;
+    let schema = ds_ref.schema();
+    let mut column = query.column.clone();
+
+    let mut query_vector = query.query_vector.first().cloned();
+    if query.query_vector.len() > 1 {
+        if column.is_none() {
+            // Infer a vector column with the same dimension of the query vector.
+            let arrow_schema = Schema::from(ds_ref.schema());
+            column = Some(default_vector_column(
+                &arrow_schema,
+                Some(query.query_vector[0].len() as i32),
+            )?);
+        }
+        let vector_field = schema.field(column.as_ref().unwrap()).unwrap();
+        if let DataType::List(_) = vector_field.data_type() {
+            // Multivector handling: concatenate into FixedSizeList<FixedSizeList<_>>
+            let vectors = query
+                .query_vector
+                .iter()
+                .map(|arr| arr.as_ref())
+                .collect::<Vec<_>>();
+            let dim = vectors[0].len();
+            let mut fsl_builder = FixedSizeListBuilder::with_capacity(
+                Float32Builder::with_capacity(dim),
+                dim as i32,
+                vectors.len(),
+            );
+            for vec in vectors {
+                fsl_builder
+                    .values()
+                    .append_slice(vec.as_primitive::<Float32Type>().values());
+                fsl_builder.append(true);
+            }
+            query_vector = Some(Arc::new(fsl_builder.finish()));
+        } else {
+            // Multiple query vectors: create a plan for each and union them
+            let query_vecs = query.query_vector.clone();
+            let plan_futures = query_vecs
+                .into_iter()
+                .map(|query_vector| {
+                    let mut sub_query = query.clone();
+                    sub_query.query_vector = vec![query_vector];
+                    let options_ref = options.clone();
+                    async move {
+                        create_plan(table, &AnyQuery::VectorQuery(sub_query), options_ref).await
+                    }
+                })
+                .collect::<Vec<_>>();
+            let plans = try_join_all(plan_futures).await?;
+            return create_multi_vector_plan(plans);
+        }
+    }
+
+    let mut scanner: Scanner = ds_ref.scan();
+
+    if let Some(query_vector) = query_vector {
+        let column = if let Some(col) = column {
+            col
+        } else {
+            let arrow_schema = Schema::from(ds_ref.schema());
+            default_vector_column(&arrow_schema, Some(query_vector.len() as i32))?
+        };
+
+        let (_, element_type) = lance::index::vector::utils::get_vector_type(schema, &column)?;
+        let is_binary = matches!(element_type, DataType::UInt8);
+        let top_k = query.base.limit.unwrap_or(DEFAULT_TOP_K) + query.base.offset.unwrap_or(0);
+
+        if is_binary {
+            let query_vector = arrow::compute::cast(&query_vector, &DataType::UInt8)?;
+            let query_vector = query_vector.as_primitive::<UInt8Type>();
+            scanner.nearest(&column, query_vector, top_k)?;
+        } else {
+            scanner.nearest(&column, query_vector.as_ref(), top_k)?;
+        }
+
+        scanner.minimum_nprobes(query.minimum_nprobes);
+        if let Some(maximum_nprobes) = query.maximum_nprobes {
+            scanner.maximum_nprobes(maximum_nprobes);
+        }
+    }
+
+    scanner.limit(
+        query.base.limit.map(|limit| limit as i64),
+        query.base.offset.map(|offset| offset as i64),
+    )?;
+
+    if let Some(ef) = query.ef {
+        scanner.ef(ef);
+    }
+
+    scanner.distance_range(query.lower_bound, query.upper_bound);
+    scanner.use_index(query.use_index);
+    scanner.prefilter(query.base.prefilter);
+
+    match query.base.select {
+        Select::Columns(ref columns) => {
+            scanner.project(columns.as_slice())?;
+        }
+        Select::Dynamic(ref select_with_transform) => {
+            scanner.project_with_transform(select_with_transform.as_slice())?;
+        }
+        Select::All => {}
+    }
+
+    if query.base.with_row_id {
+        scanner.with_row_id();
+    }
+
+    scanner.batch_size(options.max_batch_length as usize);
+
+    if query.base.fast_search {
+        scanner.fast_search();
+    }
+
+    if let Some(filter) = &query.base.filter {
+        match filter {
+            QueryFilter::Sql(sql) => {
+                scanner.filter(sql)?;
+            }
+            QueryFilter::Substrait(substrait) => {
+                scanner.filter_substrait(substrait)?;
+            }
+            QueryFilter::Datafusion(expr) => {
+                scanner.filter_expr(expr.clone());
+            }
+        }
+    }
+
+    if let Some(fts) = &query.base.full_text_search {
+        scanner.full_text_search(fts.clone())?;
+    }
+
+    if let Some(refine_factor) = query.refine_factor {
+        scanner.refine(refine_factor);
+    }
+
+    if let Some(distance_type) = query.distance_type {
+        scanner.distance_metric(distance_type.into());
+    }
+
+    if query.base.disable_scoring_autoprojection {
+        scanner.disable_scoring_autoprojection();
+    }
+
+    Ok(scanner.create_plan().await?)
+}
+
+//Helper functions below
+
+// Take many execution plans and map them into a single plan that adds
+// a query_index column and unions them.
+pub(crate) fn create_multi_vector_plan(
+    plans: Vec<Arc<dyn ExecutionPlan>>,
+) -> Result<Arc<dyn ExecutionPlan>> {
+    if plans.is_empty() {
+        return Err(Error::InvalidInput {
+            message: "No plans provided".to_string(),
+        });
+    }
+    // Projection to keeping all existing columns
+    let first_plan = plans[0].clone();
+    let project_all_columns = first_plan
+        .schema()
+        .fields()
+        .iter()
+        .enumerate()
+        .map(|(i, field)| {
+            let expr = datafusion_physical_plan::expressions::Column::new(field.name().as_str(), i);
+            let expr = Arc::new(expr) as Arc<dyn datafusion_physical_plan::PhysicalExpr>;
+            (expr, field.name().clone())
+        })
+        .collect::<Vec<_>>();
+
+    let projected_plans = plans
+        .into_iter()
+        .enumerate()
+        .map(|(plan_i, plan)| {
+            let query_index = datafusion_common::ScalarValue::Int32(Some(plan_i as i32));
+            let query_index_expr = datafusion_physical_plan::expressions::Literal::new(query_index);
+            let query_index_expr =
+                Arc::new(query_index_expr) as Arc<dyn datafusion_physical_plan::PhysicalExpr>;
+            let mut projections = vec![(query_index_expr, "query_index".to_string())];
+            projections.extend_from_slice(&project_all_columns);
+            let projection = ProjectionExec::try_new(projections, plan).unwrap();
+            Arc::new(projection) as Arc<dyn datafusion_physical_plan::ExecutionPlan>
+        })
+        .collect::<Vec<_>>();
+
+    let unioned = UnionExec::try_new(projected_plans).map_err(|err| Error::Runtime {
+        message: err.to_string(),
+    })?;
+    // We require 1 partition in the final output
+    let repartitioned = RepartitionExec::try_new(
+        unioned,
+        datafusion_physical_plan::Partitioning::RoundRobinBatch(1),
+    )
+    .unwrap();
+    Ok(Arc::new(repartitioned))
+}
+
+/// Execute a query on the namespace server instead of locally.
+async fn execute_namespace_query(
+    table: &NativeTable,
+    namespace_client: Arc<dyn LanceNamespace>,
+    query: &AnyQuery,
+    _options: QueryExecutionOptions,
+) -> Result<DatasetRecordBatchStream> {
+    // Build table_id from namespace + table name
+    let mut table_id = table.namespace.clone();
+    table_id.push(table.name.clone());
+
+    // Convert AnyQuery to namespace QueryTableRequest
+    let mut ns_request = convert_to_namespace_query(query)?;
+    // Set the table ID on the request
+    ns_request.id = Some(table_id);
+
+    // Call the namespace query_table API
+    let response_bytes = namespace_client
+        .query_table(ns_request)
+        .await
+        .map_err(|e| Error::Runtime {
+            message: format!("Failed to execute server-side query: {}", e),
+        })?;
+
+    // Parse the Arrow IPC response into a RecordBatchStream
+    parse_arrow_ipc_response(response_bytes).await
+}
+
+/// Convert an AnyQuery to the namespace QueryTableRequest format.
+fn convert_to_namespace_query(query: &AnyQuery) -> Result<NsQueryTableRequest> {
+    match query {
+        AnyQuery::VectorQuery(vq) => {
+            // Extract the query vector(s)
+            let vector = extract_query_vector(&vq.query_vector)?;
+
+            // Convert filter to SQL string
+            let filter = match &vq.base.filter {
+                Some(f) => Some(filter_to_sql(f)?),
+                None => None,
+            };
+
+            // Convert select to columns list
+            let columns = match &vq.base.select {
+                Select::All => None,
+                Select::Columns(cols) => Some(Box::new(QueryTableRequestColumns {
+                    column_names: Some(cols.clone()),
+                    column_aliases: None,
+                })),
+                Select::Dynamic(_) => {
+                    return Err(Error::NotSupported {
+                        message:
+                            "Dynamic column selection is not supported for server-side queries"
+                                .to_string(),
+                    });
+                }
+            };
+
+            // Check for unsupported features
+            if vq.base.reranker.is_some() {
+                return Err(Error::NotSupported {
+                    message: "Reranker is not supported for server-side queries".to_string(),
+                });
+            }
+
+            // Convert FTS query if present
+            let full_text_query = vq.base.full_text_search.as_ref().map(|fts| {
+                let columns = fts.columns();
+                let columns_vec = if columns.is_empty() {
+                    None
+                } else {
+                    Some(columns.into_iter().collect())
+                };
+                Box::new(QueryTableRequestFullTextQuery {
+                    string_query: Some(Box::new(StringFtsQuery {
+                        query: fts.query.to_string(),
+                        columns: columns_vec,
+                    })),
+                    structured_query: None,
+                })
+            });
+
+            Ok(NsQueryTableRequest {
+                id: None, // Will be set in namespace_query
+                k: vq.base.limit.unwrap_or(10) as i32,
+                vector: Box::new(vector),
+                vector_column: vq.column.clone(),
+                filter,
+                columns,
+                offset: vq.base.offset.map(|o| o as i32),
+                distance_type: vq.distance_type.map(|dt| dt.to_string()),
+                nprobes: Some(vq.minimum_nprobes as i32),
+                ef: vq.ef.map(|e| e as i32),
+                refine_factor: vq.refine_factor.map(|r| r as i32),
+                lower_bound: vq.lower_bound,
+                upper_bound: vq.upper_bound,
+                prefilter: Some(vq.base.prefilter),
+                fast_search: Some(vq.base.fast_search),
+                with_row_id: Some(vq.base.with_row_id),
+                bypass_vector_index: Some(!vq.use_index),
+                full_text_query,
+                ..Default::default()
+            })
+        }
+        AnyQuery::Query(q) => {
+            // For non-vector queries, pass an empty vector (similar to remote table implementation)
+            if q.reranker.is_some() {
+                return Err(Error::NotSupported {
+                    message: "Reranker is not supported for server-side query execution"
+                        .to_string(),
+                });
+            }
+
+            let filter = q.filter.as_ref().map(filter_to_sql).transpose()?;
+
+            let columns = match &q.select {
+                Select::All => None,
+                Select::Columns(cols) => Some(Box::new(QueryTableRequestColumns {
+                    column_names: Some(cols.clone()),
+                    column_aliases: None,
+                })),
+                Select::Dynamic(_) => {
+                    return Err(Error::NotSupported {
+                        message: "Dynamic columns are not supported for server-side query"
+                            .to_string(),
+                    });
+                }
+            };
+
+            // Handle full text search if present
+            let full_text_query = q.full_text_search.as_ref().map(|fts| {
+                let columns_vec = if fts.columns().is_empty() {
+                    None
+                } else {
+                    Some(fts.columns().iter().cloned().collect())
+                };
+                Box::new(QueryTableRequestFullTextQuery {
+                    string_query: Some(Box::new(StringFtsQuery {
+                        query: fts.query.to_string(),
+                        columns: columns_vec,
+                    })),
+                    structured_query: None,
+                })
+            });
+
+            // Empty vector for non-vector queries
+            let vector = Box::new(QueryTableRequestVector {
+                single_vector: Some(vec![]),
+                multi_vector: None,
+            });
+
+            Ok(NsQueryTableRequest {
+                id: None, // Will be set by caller
+                vector,
+                k: q.limit.unwrap_or(10) as i32,
+                filter,
+                columns,
+                prefilter: Some(q.prefilter),
+                offset: q.offset.map(|o| o as i32),
+                vector_column: None, // No vector column for plain queries
+                with_row_id: Some(q.with_row_id),
+                bypass_vector_index: Some(true), // No vector index for plain queries
+                full_text_query,
+                ..Default::default()
+            })
+        }
+    }
+}
+
+fn filter_to_sql(filter: &QueryFilter) -> Result<String> {
+    match filter {
+        QueryFilter::Sql(sql) => Ok(sql.clone()),
+        QueryFilter::Substrait(_) => Err(Error::NotSupported {
+            message: "Substrait filters are not supported for server-side queries".to_string(),
+        }),
+        QueryFilter::Datafusion(expr) => expr_to_sql_string(expr),
+    }
+}
+
+/// Extract query vector(s) from Arrow arrays into the namespace format.
+fn extract_query_vector(
+    query_vectors: &[Arc<dyn arrow_array::Array>],
+) -> Result<QueryTableRequestVector> {
+    if query_vectors.is_empty() {
+        return Err(Error::InvalidInput {
+            message: "Query vector is required for vector search".to_string(),
+        });
+    }
+
+    // Handle single vector case
+    if query_vectors.len() == 1 {
+        let arr = &query_vectors[0];
+        let single_vector = array_to_f32_vec(arr)?;
+        Ok(QueryTableRequestVector {
+            single_vector: Some(single_vector),
+            multi_vector: None,
+        })
+    } else {
+        // Handle multi-vector case
+        let multi_vector: Result<Vec<Vec<f32>>> =
+            query_vectors.iter().map(array_to_f32_vec).collect();
+        Ok(QueryTableRequestVector {
+            single_vector: None,
+            multi_vector: Some(multi_vector?),
+        })
+    }
+}
+
+/// Convert an Arrow array to a Vec<f32>.
+fn array_to_f32_vec(arr: &Arc<dyn arrow_array::Array>) -> Result<Vec<f32>> {
+    // Handle FixedSizeList (common for vectors)
+    if let Some(fsl) = arr
+        .as_any()
+        .downcast_ref::<arrow_array::FixedSizeListArray>()
+    {
+        let values = fsl.values();
+        if let Some(f32_arr) = values.as_any().downcast_ref::<arrow_array::Float32Array>() {
+            return Ok(f32_arr.values().to_vec());
+        }
+    }
+
+    // Handle direct Float32Array
+    if let Some(f32_arr) = arr.as_any().downcast_ref::<arrow_array::Float32Array>() {
+        return Ok(f32_arr.values().to_vec());
+    }
+
+    Err(Error::InvalidInput {
+        message: "Query vector must be Float32 type".to_string(),
+    })
+}
+
+/// Parse Arrow IPC response from the namespace server.
+async fn parse_arrow_ipc_response(bytes: bytes::Bytes) -> Result<DatasetRecordBatchStream> {
+    use arrow_ipc::reader::StreamReader;
+    use std::io::Cursor;
+
+    let cursor = Cursor::new(bytes);
+    let reader = StreamReader::try_new(cursor, None).map_err(|e| Error::Runtime {
+        message: format!("Failed to parse Arrow IPC response: {}", e),
+    })?;
+
+    // Collect all record batches
+    let schema = reader.schema();
+    let batches: Vec<_> = reader
+        .into_iter()
+        .collect::<std::result::Result<Vec<_>, _>>()
+        .map_err(|e| Error::Runtime {
+            message: format!("Failed to read Arrow IPC batches: {}", e),
+        })?;
+
+    // Create a stream from the batches
+    let stream = futures::stream::iter(batches.into_iter().map(Ok));
+    let record_batch_stream =
+        Box::pin(datafusion_physical_plan::stream::RecordBatchStreamAdapter::new(schema, stream));
+
+    Ok(DatasetRecordBatchStream::new(record_batch_stream))
+}
+
+#[cfg(test)]
+#[allow(deprecated)]
+mod tests {
+    use arrow_array::Float32Array;
+    use futures::TryStreamExt;
+    use std::sync::Arc;
+
+    use super::*;
+    use crate::query::QueryExecutionOptions;
+
+    #[test]
+    fn test_convert_to_namespace_query_vector() {
+        let query_vector = Arc::new(Float32Array::from(vec![1.0, 2.0, 3.0, 4.0]));
+
+        let vq = VectorQueryRequest {
+            base: QueryRequest {
+                limit: Some(10),
+                offset: Some(5),
+                filter: Some(QueryFilter::Sql("id > 0".to_string())),
+                select: Select::Columns(vec!["id".to_string()]),
+                ..Default::default()
+            },
+            column: Some("vector".to_string()),
+            // We cast here to satisfy the struct definition
+            query_vector: vec![query_vector as Arc<dyn Array>],
+            minimum_nprobes: 20,
+            distance_type: Some(crate::DistanceType::L2),
+            ..Default::default()
+        };
+
+        let any_query = AnyQuery::VectorQuery(vq);
+
+        let ns_request = convert_to_namespace_query(&any_query).unwrap();
+
+        assert_eq!(ns_request.k, 10);
+        assert_eq!(ns_request.offset, Some(5));
+        assert_eq!(ns_request.filter, Some("id > 0".to_string()));
+        assert_eq!(
+            ns_request
+                .columns
+                .as_ref()
+                .and_then(|c| c.column_names.as_ref()),
+            Some(&vec!["id".to_string()])
+        );
+        assert_eq!(ns_request.vector_column, Some("vector".to_string()));
+        assert_eq!(ns_request.distance_type, Some("l2".to_string()));
+
+        // Verify the vector data was extracted correctly
+        assert!(ns_request.vector.single_vector.is_some());
+        assert_eq!(
+            ns_request.vector.single_vector.as_ref().unwrap(),
+            &vec![1.0, 2.0, 3.0, 4.0]
+        );
+    }
+
+    #[test]
+    fn test_convert_to_namespace_query_plain_query() {
+        let q = QueryRequest {
+            limit: Some(20),
+            offset: Some(5),
+            filter: Some(QueryFilter::Sql("id > 5".to_string())),
+            select: Select::Columns(vec!["id".to_string()]),
+            with_row_id: true,
+            ..Default::default()
+        };
+
+        let any_query = AnyQuery::Query(q);
+
+        let ns_request = convert_to_namespace_query(&any_query).unwrap();
+
+        assert_eq!(ns_request.k, 20);
+        assert_eq!(ns_request.offset, Some(5));
+        assert_eq!(ns_request.filter, Some("id > 5".to_string()));
+        assert_eq!(
+            ns_request
+                .columns
+                .as_ref()
+                .and_then(|c| c.column_names.as_ref()),
+            Some(&vec!["id".to_string()])
+        );
+        assert_eq!(ns_request.with_row_id, Some(true));
+        assert_eq!(ns_request.bypass_vector_index, Some(true));
+        assert!(ns_request.vector_column.is_none());
+
+        assert!(ns_request.vector.single_vector.as_ref().unwrap().is_empty());
+    }
+
+    #[tokio::test]
+    async fn test_execute_query_local_routing() {
+        use crate::connect;
+        use crate::table::query::execute_query;
+        use arrow_array::{Int32Array, RecordBatch};
+        use arrow_schema::{DataType, Field, Schema};
+
+        let conn = connect("memory://").execute().await.unwrap();
+
+        let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int32, false)]));
+        let batch = RecordBatch::try_new(
+            schema.clone(),
+            vec![Arc::new(Int32Array::from(vec![1, 2, 3, 4, 5]))],
+        )
+        .unwrap();
+
+        let table = conn
+            .create_table("test_routing", vec![batch])
+            .execute()
+            .await
+            .unwrap();
+
+        let native_table = table.as_native().unwrap();
+
+        // Setup a request
+        let req = QueryRequest {
+            filter: Some(QueryFilter::Sql("id > 3".to_string())),
+            ..Default::default()
+        };
+        let query = AnyQuery::Query(req);
+
+        // Action: Call execute_query directly
+        // This validates that execute_query correctly routes to the local DataFusion engine
+        // when table.namespace_client is None.
+        let stream = execute_query(native_table, &query, QueryExecutionOptions::default())
+            .await
+            .unwrap();
+
+        // Verify results
+        let batches = stream.try_collect::<Vec<_>>().await.unwrap();
+        let count: usize = batches.iter().map(|b| b.num_rows()).sum();
+        assert_eq!(count, 2); // 4 and 5
+    }
+
+    #[tokio::test]
+    async fn test_create_plan_multivector_structure() {
+        use arrow_array::{Float32Array, RecordBatch};
+        use arrow_schema::{DataType, Field, Schema};
+        use datafusion_physical_plan::display::DisplayableExecutionPlan;
+
+        use crate::table::query::create_plan;
+
+        use crate::connect;
+
+        let conn = connect("memory://").execute().await.unwrap();
+        let schema = Arc::new(Schema::new(vec![
+            Field::new("id", DataType::Int32, false),
+            Field::new(
+                "vector",
+                DataType::FixedSizeList(Arc::new(Field::new("item", DataType::Float32, true)), 2),
+                false,
+            ),
+        ]));
+
+        let batch = RecordBatch::new_empty(schema.clone());
+        let table = conn
+            .create_table("test_plan", vec![batch])
+            .execute()
+            .await
+            .unwrap();
+        let native_table = table.as_native().unwrap();
+
+        // This triggers the "create_multi_vector_plan" logic branch
+        let q1 = Arc::new(Float32Array::from(vec![1.0, 2.0]));
+        let q2 = Arc::new(Float32Array::from(vec![3.0, 4.0]));
+
+        let req = VectorQueryRequest {
+            column: Some("vector".to_string()),
+            query_vector: vec![q1, q2],
+            ..Default::default()
+        };
+        let query = AnyQuery::VectorQuery(req);
+
+        // Create the Plan
+        let plan = create_plan(native_table, &query, QueryExecutionOptions::default())
+            .await
+            .unwrap();
+
+        // formatting it allows us to see the hierarchy
+        let display = DisplayableExecutionPlan::new(plan.as_ref())
+            .indent(true)
+            .to_string();
+
+        // We expect a RepartitionExec wrapping a UnionExec
+        assert!(
+            display.contains("RepartitionExec"),
+            "Plan should include Repartitioning"
+        );
+        assert!(
+            display.contains("UnionExec"),
+            "Plan should include a Union of multiple searches"
+        );
+        // We expect the projection to add the 'query_index' column (logic inside multi_vector_plan)
+        assert!(
+            display.contains("query_index"),
+            "Plan should add query_index column"
+        );
+    }
+}

rust/lancedb/src/table/schema_evolution.rs

@@ -52,11 +52,12 @@ pub(crate) async fn execute_add_columns(
     transforms: NewColumnTransform,
     read_columns: Option<Vec<String>>,
 ) -> Result<AddColumnsResult> {
-    let mut dataset = table.dataset.get_mut().await?;
+    table.dataset.ensure_mutable()?;
+    let mut dataset = (*table.dataset.get().await?).clone();
     dataset.add_columns(transforms, read_columns, None).await?;
-    Ok(AddColumnsResult {
-        version: dataset.version().version,
-    })
+    let version = dataset.version().version;
+    table.dataset.update(dataset);
+    Ok(AddColumnsResult { version })
 }
 
 /// Internal implementation of the alter columns logic.
@@ -66,11 +67,12 @@ pub(crate) async fn execute_alter_columns(
     table: &NativeTable,
     alterations: &[ColumnAlteration],
 ) -> Result<AlterColumnsResult> {
-    let mut dataset = table.dataset.get_mut().await?;
+    table.dataset.ensure_mutable()?;
+    let mut dataset = (*table.dataset.get().await?).clone();
     dataset.alter_columns(alterations).await?;
-    Ok(AlterColumnsResult {
-        version: dataset.version().version,
-    })
+    let version = dataset.version().version;
+    table.dataset.update(dataset);
+    Ok(AlterColumnsResult { version })
 }
 
 /// Internal implementation of the drop columns logic.
@@ -80,16 +82,17 @@ pub(crate) async fn execute_drop_columns(
     table: &NativeTable,
     columns: &[&str],
 ) -> Result<DropColumnsResult> {
-    let mut dataset = table.dataset.get_mut().await?;
+    table.dataset.ensure_mutable()?;
+    let mut dataset = (*table.dataset.get().await?).clone();
     dataset.drop_columns(columns).await?;
-    Ok(DropColumnsResult {
-        version: dataset.version().version,
-    })
+    let version = dataset.version().version;
+    table.dataset.update(dataset);
+    Ok(DropColumnsResult { version })
 }
 
 #[cfg(test)]
 mod tests {
-    use arrow_array::{record_batch, Int32Array, RecordBatchIterator, StringArray};
+    use arrow_array::{record_batch, Int32Array, StringArray};
     use arrow_schema::DataType;
     use futures::TryStreamExt;
     use lance::dataset::ColumnAlteration;
@@ -105,13 +108,9 @@ mod tests {
         let conn = connect("memory://").execute().await.unwrap();
 
         let batch = record_batch!(("id", Int32, [1, 2, 3, 4, 5])).unwrap();
-        let schema = batch.schema();
 
         let table = conn
-            .create_table(
-                "test_add_columns",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_add_columns", batch)
             .execute()
             .await
             .unwrap();
@@ -169,13 +168,9 @@ mod tests {
         let conn = connect("memory://").execute().await.unwrap();
 
         let batch = record_batch!(("x", Int32, [10, 20, 30])).unwrap();
-        let schema = batch.schema();
 
         let table = conn
-            .create_table(
-                "test_add_multi_columns",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_add_multi_columns", batch)
             .execute()
             .await
             .unwrap();
@@ -205,13 +200,9 @@ mod tests {
         let conn = connect("memory://").execute().await.unwrap();
 
         let batch = record_batch!(("id", Int32, [1, 2, 3])).unwrap();
-        let schema = batch.schema();
 
         let table = conn
-            .create_table(
-                "test_add_const_column",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_add_const_column", batch)
             .execute()
             .await
             .unwrap();
@@ -255,13 +246,9 @@ mod tests {
         let conn = connect("memory://").execute().await.unwrap();
 
         let batch = record_batch!(("old_name", Int32, [1, 2, 3])).unwrap();
-        let schema = batch.schema();
 
         let table = conn
-            .create_table(
-                "test_alter_rename",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_alter_rename", batch)
             .execute()
             .await
             .unwrap();
@@ -304,10 +291,7 @@ mod tests {
         .unwrap();
 
         let table = conn
-            .create_table(
-                "test_alter_nullable",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_alter_nullable", batch)
             .execute()
             .await
             .unwrap();
@@ -332,13 +316,9 @@ mod tests {
         let conn = connect("memory://").execute().await.unwrap();
 
         let batch = record_batch!(("num", Int32, [1, 2, 3])).unwrap();
-        let schema = batch.schema();
 
         let table = conn
-            .create_table(
-                "test_cast_type",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_cast_type", batch)
             .execute()
             .await
             .unwrap();
@@ -379,13 +359,9 @@ mod tests {
         let conn = connect("memory://").execute().await.unwrap();
 
         let batch = record_batch!(("num", Int32, [1, 2, 3])).unwrap();
-        let schema = batch.schema();
 
         let table = conn
-            .create_table(
-                "test_invalid_cast",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_invalid_cast", batch)
             .execute()
             .await
             .unwrap();
@@ -407,13 +383,9 @@ mod tests {
         let conn = connect("memory://").execute().await.unwrap();
 
         let batch = record_batch!(("a", Int32, [1, 2, 3]), ("b", Int32, [4, 5, 6])).unwrap();
-        let schema = batch.schema();
 
         let table = conn
-            .create_table(
-                "test_alter_multi",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_alter_multi", batch)
             .execute()
             .await
             .unwrap();
@@ -441,13 +413,9 @@ mod tests {
 
         let batch =
             record_batch!(("keep", Int32, [1, 2, 3]), ("remove", Int32, [4, 5, 6])).unwrap();
-        let schema = batch.schema();
 
         let table = conn
-            .create_table(
-                "test_drop_single",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_drop_single", batch)
             .execute()
             .await
             .unwrap();
@@ -478,13 +446,9 @@ mod tests {
             ("d", Int32, [7, 8])
         )
         .unwrap();
-        let schema = batch.schema();
 
         let table = conn
-            .create_table(
-                "test_drop_multi",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_drop_multi", batch)
             .execute()
             .await
             .unwrap();
@@ -511,13 +475,9 @@ mod tests {
             ("extra", Int32, [10, 20, 30])
         )
         .unwrap();
-        let schema = batch.schema();
 
         let table = conn
-            .create_table(
-                "test_drop_preserves",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_drop_preserves", batch)
             .execute()
             .await
             .unwrap();
@@ -567,13 +527,9 @@ mod tests {
         let conn = connect("memory://").execute().await.unwrap();
 
         let batch = record_batch!(("existing", Int32, [1, 2, 3])).unwrap();
-        let schema = batch.schema();
 
         let table = conn
-            .create_table(
-                "test_drop_nonexistent",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_drop_nonexistent", batch)
             .execute()
             .await
             .unwrap();
@@ -593,13 +549,9 @@ mod tests {
         let conn = connect("memory://").execute().await.unwrap();
 
         let batch = record_batch!(("existing", Int32, [1, 2, 3])).unwrap();
-        let schema = batch.schema();
 
         let table = conn
-            .create_table(
-                "test_alter_nonexistent",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_alter_nonexistent", batch)
             .execute()
             .await
             .unwrap();
@@ -623,13 +575,8 @@ mod tests {
         let conn = connect("memory://").execute().await.unwrap();
 
         let batch = record_batch!(("a", Int32, [1, 2, 3]), ("b", Int32, [4, 5, 6])).unwrap();
-        let schema = batch.schema();
-
         let table = conn
-            .create_table(
-                "test_version_increment",
-                RecordBatchIterator::new(vec![Ok(batch)], schema),
-            )
+            .create_table("test_version_increment", batch)
             .execute()
             .await
             .unwrap();

rust/lancedb/src/table/update.rs

@@ -78,11 +78,13 @@ pub(crate) async fn execute_update(
     table: &NativeTable,
     update: UpdateBuilder,
 ) -> Result<UpdateResult> {
+    table.dataset.ensure_mutable()?;
+
     // 1. Snapshot the current dataset
-    let dataset = table.dataset.get().await?.clone();
+    let dataset = table.dataset.get().await?;
 
     // 2. Initialize the Lance Core builder
-    let mut builder = LanceUpdateBuilder::new(Arc::new(dataset));
+    let mut builder = LanceUpdateBuilder::new(dataset);
 
     // 3. Apply the filter (WHERE clause)
     if let Some(predicate) = update.filter {
@@ -99,10 +101,7 @@ pub(crate) async fn execute_update(
     let res = operation.execute().await?;
 
     // 6. Update the table's view of the latest version
-    table
-        .dataset
-        .set_latest(res.new_dataset.as_ref().clone())
-        .await;
+    table.dataset.update(res.new_dataset.as_ref().clone());
 
     Ok(UpdateResult {
         rows_updated: res.rows_updated,
@@ -117,9 +116,8 @@ mod tests {
     use crate::query::{ExecutableQuery, Select};
     use arrow_array::{
         record_batch, Array, BooleanArray, Date32Array, FixedSizeListArray, Float32Array,
-        Float64Array, Int32Array, Int64Array, LargeStringArray, RecordBatch, RecordBatchIterator,
-        RecordBatchReader, StringArray, TimestampMillisecondArray, TimestampNanosecondArray,
-        UInt32Array,
+        Float64Array, Int32Array, Int64Array, LargeStringArray, RecordBatch, StringArray,
+        TimestampMillisecondArray, TimestampNanosecondArray, UInt32Array,
     };
     use arrow_data::ArrayDataBuilder;
     use arrow_schema::{ArrowError, DataType, Field, Schema, TimeUnit};
@@ -167,51 +165,46 @@ mod tests {
             ),
         ]));
 
-        let record_batch_iter = RecordBatchIterator::new(
-            vec![RecordBatch::try_new(
-                schema.clone(),
-                vec![
-                    Arc::new(Int32Array::from_iter_values(0..10)),
-                    Arc::new(Int64Array::from_iter_values(0..10)),
-                    Arc::new(UInt32Array::from_iter_values(0..10)),
-                    Arc::new(StringArray::from_iter_values(vec![
-                        "a", "b", "c", "d", "e", "f", "g", "h", "i", "j",
-                    ])),
-                    Arc::new(LargeStringArray::from_iter_values(vec![
-                        "a", "b", "c", "d", "e", "f", "g", "h", "i", "j",
-                    ])),
-                    Arc::new(Float32Array::from_iter_values((0..10).map(|i| i as f32))),
-                    Arc::new(Float64Array::from_iter_values((0..10).map(|i| i as f64))),
-                    Arc::new(Into::<BooleanArray>::into(vec![
-                        true, false, true, false, true, false, true, false, true, false,
-                    ])),
-                    Arc::new(Date32Array::from_iter_values(0..10)),
-                    Arc::new(TimestampNanosecondArray::from_iter_values(0..10)),
-                    Arc::new(TimestampMillisecondArray::from_iter_values(0..10)),
-                    Arc::new(
-                        create_fixed_size_list(
-                            Float32Array::from_iter_values((0..20).map(|i| i as f32)),
-                            2,
-                        )
-                        .unwrap(),
-                    ),
-                    Arc::new(
-                        create_fixed_size_list(
-                            Float64Array::from_iter_values((0..20).map(|i| i as f64)),
-                            2,
-                        )
-                        .unwrap(),
-                    ),
-                ],
-            )
-            .unwrap()]
-            .into_iter()
-            .map(Ok),
+        let batch = RecordBatch::try_new(
             schema.clone(),
-        );
+            vec![
+                Arc::new(Int32Array::from_iter_values(0..10)),
+                Arc::new(Int64Array::from_iter_values(0..10)),
+                Arc::new(UInt32Array::from_iter_values(0..10)),
+                Arc::new(StringArray::from_iter_values(vec![
+                    "a", "b", "c", "d", "e", "f", "g", "h", "i", "j",
+                ])),
+                Arc::new(LargeStringArray::from_iter_values(vec![
+                    "a", "b", "c", "d", "e", "f", "g", "h", "i", "j",
+                ])),
+                Arc::new(Float32Array::from_iter_values((0..10).map(|i| i as f32))),
+                Arc::new(Float64Array::from_iter_values((0..10).map(|i| i as f64))),
+                Arc::new(Into::<BooleanArray>::into(vec![
+                    true, false, true, false, true, false, true, false, true, false,
+                ])),
+                Arc::new(Date32Array::from_iter_values(0..10)),
+                Arc::new(TimestampNanosecondArray::from_iter_values(0..10)),
+                Arc::new(TimestampMillisecondArray::from_iter_values(0..10)),
+                Arc::new(
+                    create_fixed_size_list(
+                        Float32Array::from_iter_values((0..20).map(|i| i as f32)),
+                        2,
+                    )
+                    .unwrap(),
+                ),
+                Arc::new(
+                    create_fixed_size_list(
+                        Float64Array::from_iter_values((0..20).map(|i| i as f64)),
+                        2,
+                    )
+                    .unwrap(),
+                ),
+            ],
+        )
+        .unwrap();
 
         let table = conn
-            .create_table("my_table", record_batch_iter)
+            .create_table("my_table", batch)
             .execute()
             .await
             .unwrap();
@@ -338,15 +331,13 @@ mod tests {
         Ok(FixedSizeListArray::from(data))
     }
 
-    fn make_test_batches() -> impl RecordBatchReader + Send + Sync + 'static {
+    fn make_test_batch() -> RecordBatch {
         let schema = Arc::new(Schema::new(vec![Field::new("i", DataType::Int32, false)]));
-        RecordBatchIterator::new(
-            vec![RecordBatch::try_new(
-                schema.clone(),
-                vec![Arc::new(Int32Array::from_iter_values(0..10))],
-            )],
-            schema,
+        RecordBatch::try_new(
+            schema.clone(),
+            vec![Arc::new(Int32Array::from_iter_values(0..10))],
         )
+        .unwrap()
     }
 
     #[tokio::test]
@@ -367,12 +358,8 @@ mod tests {
         )
         .unwrap();
 
-        let schema = batch.schema();
-        // need the iterator for create table
-        let record_batch_iter = RecordBatchIterator::new(vec![Ok(batch)], schema);
-
         let table = conn
-            .create_table("my_table", record_batch_iter)
+            .create_table("my_table", batch)
             .execute()
             .await
             .unwrap();
@@ -430,7 +417,7 @@ mod tests {
             .await
             .unwrap();
         let tbl = conn
-            .create_table("my_table", make_test_batches())
+            .create_table("my_table", make_test_batch())
             .execute()
             .await
             .unwrap();

rust/lancedb/src/test_utils.rs

@@ -3,3 +3,15 @@
 
 pub mod connection;
 pub mod datagen;
+pub mod embeddings;
+
+#[derive(Debug)]
+pub struct TestCustomError;
+
+impl std::fmt::Display for TestCustomError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "TestCustomError occurred")
+    }
+}
+
+impl std::error::Error for TestCustomError {}