feat: add mem_wal flag to merge insert for MemWAL write path

Add support for enabling MemWAL (Memory Write-Ahead Log) mode on merge insert
operations. This allows streaming writes to route through memory nodes for
high-performance buffered writes.

Changes:
- Add `mem_wal` field to MergeInsertBuilder with validation
- Add `x-lancedb-mem-wal-enabled` header for remote requests
- Add Python `mem_wal()` method to LanceMergeInsertBuilder
- Add validation to ensure only upsert pattern is supported:
  - when_matched_update_all() without filter
  - when_not_matched_insert_all()
- Throw NotSupported error for native tables
- Add mem_wal_enabled to ClientConfig for Python/Node bindings

Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>

.bumpversion.toml

@@ -1,5 +1,5 @@
 [tool.bumpversion]
-current_version = "0.26.2"
+current_version = "0.27.0-beta.3"
 parse = """(?x)
     (?P<major>0|[1-9]\\d*)\\.
     (?P<minor>0|[1-9]\\d*)\\.

.github/workflows/build_linux_wheel/action.yml

@@ -29,6 +29,7 @@ runs:
       if: ${{ inputs.arm-build == 'false' }}
       uses: PyO3/maturin-action@v1
       with:
+        maturin-version: "1.12.4"
         command: build
         working-directory: python
         docker-options: "-e PIP_EXTRA_INDEX_URL='https://pypi.fury.io/lance-format/ https://pypi.fury.io/lancedb/'"
@@ -44,6 +45,7 @@ runs:
       if: ${{ inputs.arm-build == 'true' }}
       uses: PyO3/maturin-action@v1
       with:
+        maturin-version: "1.12.4"
         command: build
         working-directory: python
         docker-options: "-e PIP_EXTRA_INDEX_URL='https://pypi.fury.io/lance-format/ https://pypi.fury.io/lancedb/'"

.github/workflows/build_mac_wheel/action.yml

@@ -20,6 +20,7 @@ runs:
       uses: PyO3/maturin-action@v1
       with:
         command: build
+        maturin-version: "1.12.4"
         # TODO: pass through interpreter
         args: ${{ inputs.args }}
         docker-options: "-e PIP_EXTRA_INDEX_URL='https://pypi.fury.io/lance-format/ https://pypi.fury.io/lancedb/'"

.github/workflows/build_windows_wheel/action.yml

@@ -25,6 +25,7 @@ runs:
       uses: PyO3/maturin-action@v1
       with:
         command: build
+        maturin-version: "1.12.4"
         args: ${{ inputs.args }}
         docker-options: "-e PIP_EXTRA_INDEX_URL='https://pypi.fury.io/lance-format/ https://pypi.fury.io/lancedb/'"
         working-directory: python

.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
@@ -355,7 +356,8 @@ jobs:
           if [[ $DRY_RUN == "true" ]]; then
             ARGS="$ARGS --dry-run"
           fi
-          if [[ $GITHUB_REF =~ refs/tags/v(.*)-beta.* ]]; then
+          VERSION=$(node -p "require('./package.json').version")
+          if [[ $VERSION == *-* ]]; then
             ARGS="$ARGS --tag preview"
           fi
           npm publish $ARGS

.github/workflows/python.yml

@@ -8,7 +8,12 @@ on:
     paths:
       - Cargo.toml
       - python/**
+      - rust/**
       - .github/workflows/python.yml
+      - .github/workflows/build_linux_wheel/**
+      - .github/workflows/build_mac_wheel/**
+      - .github/workflows/build_windows_wheel/**
+      - .github/workflows/run_tests/**
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}

.github/workflows/rust.yml

@@ -100,7 +100,9 @@ jobs:
           lfs: true
       - uses: Swatinem/rust-cache@v2
       - name: Install dependencies
-        run: sudo apt install -y protobuf-compiler libssl-dev
+        run: |
+          sudo apt update
+          sudo apt install -y protobuf-compiler libssl-dev
       - uses: rui314/setup-mold@v1
       - name: Make Swap
         run: |
@@ -183,7 +185,7 @@ 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.toml

@@ -5,30 +5,30 @@ exclude = ["python"]
 resolver = "2"
 
 [workspace.package]
-edition = "2021"
+edition = "2024"
 authors = ["LanceDB Devs <dev@lancedb.com>"]
 license = "Apache-2.0"
 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" = "=2.0.1", default-features = false }
-lance-core = "=2.0.1"
-lance-datagen = "=2.0.1"
-lance-file = "=2.0.1"
-lance-io = { "version" = "=2.0.1", default-features = false }
-lance-index = "=2.0.1"
-lance-linalg = "=2.0.1"
-lance-namespace = "=2.0.1"
-lance-namespace-impls = { "version" = "=2.0.1", default-features = false }
-lance-table = "=2.0.1"
-lance-testing = "=2.0.1"
-lance-datafusion = "=2.0.1"
-lance-encoding = "=2.0.1"
-lance-arrow = "=2.0.1"
+lance = { "version" = "=3.0.0-rc.3", default-features = false, "tag" = "v3.0.0-rc.3", "git" = "https://github.com/lance-format/lance.git" }
+lance-core = { "version" = "=3.0.0-rc.3", "tag" = "v3.0.0-rc.3", "git" = "https://github.com/lance-format/lance.git" }
+lance-datagen = { "version" = "=3.0.0-rc.3", "tag" = "v3.0.0-rc.3", "git" = "https://github.com/lance-format/lance.git" }
+lance-file = { "version" = "=3.0.0-rc.3", "tag" = "v3.0.0-rc.3", "git" = "https://github.com/lance-format/lance.git" }
+lance-io = { "version" = "=3.0.0-rc.3", default-features = false, "tag" = "v3.0.0-rc.3", "git" = "https://github.com/lance-format/lance.git" }
+lance-index = { "version" = "=3.0.0-rc.3", "tag" = "v3.0.0-rc.3", "git" = "https://github.com/lance-format/lance.git" }
+lance-linalg = { "version" = "=3.0.0-rc.3", "tag" = "v3.0.0-rc.3", "git" = "https://github.com/lance-format/lance.git" }
+lance-namespace = { "version" = "=3.0.0-rc.3", "tag" = "v3.0.0-rc.3", "git" = "https://github.com/lance-format/lance.git" }
+lance-namespace-impls = { "version" = "=3.0.0-rc.3", default-features = false, "tag" = "v3.0.0-rc.3", "git" = "https://github.com/lance-format/lance.git" }
+lance-table = { "version" = "=3.0.0-rc.3", "tag" = "v3.0.0-rc.3", "git" = "https://github.com/lance-format/lance.git" }
+lance-testing = { "version" = "=3.0.0-rc.3", "tag" = "v3.0.0-rc.3", "git" = "https://github.com/lance-format/lance.git" }
+lance-datafusion = { "version" = "=3.0.0-rc.3", "tag" = "v3.0.0-rc.3", "git" = "https://github.com/lance-format/lance.git" }
+lance-encoding = { "version" = "=3.0.0-rc.3", "tag" = "v3.0.0-rc.3", "git" = "https://github.com/lance-format/lance.git" }
+lance-arrow = { "version" = "=3.0.0-rc.3", "tag" = "v3.0.0-rc.3", "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 }
@@ -40,13 +40,15 @@ arrow-schema = "57.2"
 arrow-select = "57.2"
 arrow-cast = "57.2"
 async-trait = "0"
-datafusion = { version = "51.0", default-features = false }
-datafusion-catalog = "51.0"
-datafusion-common = { version = "51.0", default-features = false }
-datafusion-execution = "51.0"
-datafusion-expr = "51.0"
-datafusion-physical-plan = "51.0"
-datafusion-physical-expr = "51.0"
+datafusion = { version = "52.1", default-features = false }
+datafusion-catalog = "52.1"
+datafusion-common = { version = "52.1", default-features = false }
+datafusion-execution = "52.1"
+datafusion-expr = "52.1"
+datafusion-functions = "52.1"
+datafusion-physical-plan = "52.1"
+datafusion-physical-expr = "52.1"
+datafusion-sql = "52.1"
 env_logger = "0.11"
 half = { "version" = "2.7.1", default-features = false, features = [
     "num-traits",

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.3</version>
 </dependency>
 ```
 

docs/src/js/interfaces/DeleteResult.md

@@ -8,6 +8,14 @@
 
 ## Properties
 
+### numDeletedRows
+
+```ts
+numDeletedRows: number;
+```
+
+***
+
 ### version
 
 ```ts

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.3</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.3</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>2.0.1</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.3"
 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/__test__/table.test.ts

@@ -1697,6 +1697,65 @@ describe.each([arrow15, arrow16, arrow17, arrow18])(
       expect(results2[0].text).toBe(data[1].text);
     });
 
+    test("full text search fast search", async () => {
+      const db = await connect(tmpDir.name);
+      const data = [{ text: "hello world", vector: [0.1, 0.2, 0.3], id: 1 }];
+      const table = await db.createTable("test", data);
+      await table.createIndex("text", {
+        config: Index.fts(),
+      });
+
+      // Insert unindexed data after creating the index.
+      await table.add([{ text: "xyz", vector: [0.4, 0.5, 0.6], id: 2 }]);
+
+      const withFlatSearch = await table
+        .search("xyz", "fts")
+        .limit(10)
+        .toArray();
+      expect(withFlatSearch.length).toBeGreaterThan(0);
+
+      const fastSearchResults = await table
+        .search("xyz", "fts")
+        .fastSearch()
+        .limit(10)
+        .toArray();
+      expect(fastSearchResults.length).toBe(0);
+
+      const nearestToTextFastSearch = await table
+        .query()
+        .nearestToText("xyz")
+        .fastSearch()
+        .limit(10)
+        .toArray();
+      expect(nearestToTextFastSearch.length).toBe(0);
+
+      // fastSearch should be chainable with other methods.
+      const chainedFastSearch = await table
+        .search("xyz", "fts")
+        .fastSearch()
+        .select(["text"])
+        .limit(5)
+        .toArray();
+      expect(chainedFastSearch.length).toBe(0);
+
+      await table.optimize();
+
+      const indexedFastSearch = await table
+        .search("xyz", "fts")
+        .fastSearch()
+        .limit(10)
+        .toArray();
+      expect(indexedFastSearch.length).toBeGreaterThan(0);
+
+      const indexedNearestToTextFastSearch = await table
+        .query()
+        .nearestToText("xyz")
+        .fastSearch()
+        .limit(10)
+        .toArray();
+      expect(indexedNearestToTextFastSearch.length).toBeGreaterThan(0);
+    });
+
     test("prewarm full text search index", async () => {
       const db = await connect(tmpDir.name);
       const data = [

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.3",
 	"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.3",
 	"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.3",
 	"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.3",
 	"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.3",
 	"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.3",
   "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.3",
 	"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.3",
   "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

@@ -8,10 +8,10 @@ use lancedb::database::{CreateTableMode, Database};
 use napi::bindgen_prelude::*;
 use napi_derive::*;
 
+use crate::ConnectionOptions;
 use crate::error::NapiErrorExt;
 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};

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/index.rs

@@ -3,12 +3,12 @@
 
 use std::sync::Mutex;
 
+use lancedb::index::Index as LanceDbIndex;
 use lancedb::index::scalar::{BTreeIndexBuilder, FtsIndexBuilder};
 use lancedb::index::vector::{
     IvfFlatIndexBuilder, IvfHnswPqIndexBuilder, IvfHnswSqIndexBuilder, IvfPqIndexBuilder,
     IvfRqIndexBuilder,
 };
-use lancedb::index::Index as LanceDbIndex;
 use napi_derive::napi;
 
 use crate::util::parse_distance_type;

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

@@ -17,11 +17,11 @@ use lancedb::query::VectorQuery as LanceDbVectorQuery;
 use napi::bindgen_prelude::*;
 use napi_derive::napi;
 
-use crate::error::convert_error;
 use crate::error::NapiErrorExt;
+use crate::error::convert_error;
 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('"');
@@ -549,15 +551,12 @@ fn parse_fts_query(query: napi::JsObject) -> napi::Result<FullTextSearchQuery> {
             }
         };
         let mut query = FullTextSearchQuery::new_query(query);
-        if let Some(cols) = columns {
-            if !cols.is_empty() {
-                query = query.with_columns(&cols).map_err(|e| {
-                    napi::Error::from_reason(format!(
-                        "Failed to set full text search columns: {}",
-                        e
-                    ))
-                })?;
-            }
+        if let Some(cols) = columns
+            && !cols.is_empty()
+        {
+            query = query.with_columns(&cols).map_err(|e| {
+                napi::Error::from_reason(format!("Failed to set full text search columns: {}", e))
+            })?;
         }
         Ok(query)
     } else {

nodejs/src/remote.rs

@@ -145,6 +145,7 @@ impl From<ClientConfig> for lancedb::remote::ClientConfig {
             id_delimiter: config.id_delimiter,
             tls_config: config.tls_config.map(Into::into),
             header_provider: None, // the header provider is set separately later
+            mem_wal_enabled: None, // mem_wal is set per-operation in merge_insert
         }
     }
 }

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

@@ -95,8 +95,7 @@ impl napi::bindgen_prelude::FromNapiValue for Session {
         napi_val: napi::sys::napi_value,
     ) -> 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)
+            unsafe { napi::bindgen_prelude::ClassInstance::from_napi_value(env, napi_val)? };
+        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" {
@@ -742,12 +753,14 @@ impl From<lancedb::table::AddResult> for AddResult {
 
 #[napi(object)]
 pub struct DeleteResult {
+    pub num_deleted_rows: i64,
     pub version: i64,
 }
 
 impl From<lancedb::table::DeleteResult> for DeleteResult {
     fn from(value: lancedb::table::DeleteResult) -> Self {
         Self {
+            num_deleted_rows: value.num_deleted_rows as i64,
             version: value.version as i64,
         }
     }

python/.bumpversion.toml

@@ -1,5 +1,5 @@
 [tool.bumpversion]
-current_version = "0.30.0-beta.0"
+current_version = "0.30.0-beta.3"
 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 = "0.30.0-beta.0"
+version = "0.30.0-beta.3"
 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

@@ -59,9 +59,9 @@ tests = [
     "polars>=0.19, <=1.3.0",
     "tantivy",
     "pyarrow-stubs",
-    "pylance>=1.0.0b14",
+    "pylance>=1.0.0b14,<3.0.0",
     "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/merge.py

@@ -34,6 +34,7 @@ class LanceMergeInsertBuilder(object):
         self._when_not_matched_by_source_condition = None
         self._timeout = None
         self._use_index = True
+        self._mem_wal = False
 
     def when_matched_update_all(
         self, *, where: Optional[str] = None
@@ -96,6 +97,47 @@ class LanceMergeInsertBuilder(object):
         self._use_index = use_index
         return self
 
+    def mem_wal(self, enabled: bool = True) -> LanceMergeInsertBuilder:
+        """
+        Enable MemWAL (Memory Write-Ahead Log) mode for this merge insert operation.
+
+        When enabled, the merge insert will route data through a memory node service
+        that buffers writes before flushing to storage. This is only supported for
+        remote (LanceDB Cloud) tables.
+
+        **Important:** MemWAL only supports the upsert pattern. You must use:
+        - `when_matched_update_all()` (without a filter condition)
+        - `when_not_matched_insert_all()`
+
+        MemWAL does NOT support:
+        - `when_matched_update_all(where=...)` with a filter condition
+        - `when_not_matched_by_source_delete()`
+
+        Parameters
+        ----------
+        enabled: bool
+            Whether to enable MemWAL mode. Defaults to `True`.
+
+        Raises
+        ------
+        NotImplementedError
+            If used on a native (local) table, as MemWAL is only supported for
+            remote tables.
+        ValueError
+            If the merge insert pattern is not supported by MemWAL.
+
+        Examples
+        --------
+        >>> # Correct usage with MemWAL
+        >>> table.merge_insert(["id"]) \\
+        ...     .when_matched_update_all() \\
+        ...     .when_not_matched_insert_all() \\
+        ...     .mem_wal() \\
+        ...     .execute(new_data)
+        """
+        self._mem_wal = enabled
+        return self
+
     def execute(
         self,
         new_data: DATA,

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

@@ -606,6 +606,7 @@ class LanceQueryBuilder(ABC):
                 query,
                 ordering_field_name=ordering_field_name,
                 fts_columns=fts_columns,
+                fast_search=fast_search,
             )
 
         if isinstance(query, list):
@@ -1456,12 +1457,14 @@ class LanceFtsQueryBuilder(LanceQueryBuilder):
         query: str | FullTextQuery,
         ordering_field_name: Optional[str] = None,
         fts_columns: Optional[Union[str, List[str]]] = None,
+        fast_search: bool = None,
     ):
         super().__init__(table)
         self._query = query
         self._phrase_query = False
         self.ordering_field_name = ordering_field_name
         self._reranker = None
+        self._fast_search = fast_search
         if isinstance(fts_columns, str):
             fts_columns = [fts_columns]
         self._fts_columns = fts_columns
@@ -1483,6 +1486,19 @@ class LanceFtsQueryBuilder(LanceQueryBuilder):
         self._phrase_query = phrase_query
         return self
 
+    def fast_search(self) -> LanceFtsQueryBuilder:
+        """
+        Skip a flat search of unindexed data. This will improve
+        search performance but search results will not include unindexed data.
+
+        Returns
+        -------
+        LanceFtsQueryBuilder
+            The LanceFtsQueryBuilder object.
+        """
+        self._fast_search = True
+        return self
+
     def to_query_object(self) -> Query:
         return Query(
             columns=self._columns,
@@ -1494,6 +1510,7 @@ class LanceFtsQueryBuilder(LanceQueryBuilder):
                 query=self._query, columns=self._fts_columns
             ),
             offset=self._offset,
+            fast_search=self._fast_search,
         )
 
     def output_schema(self) -> pa.Schema:
@@ -1782,6 +1799,26 @@ 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/remote/table.py

@@ -218,8 +218,6 @@ class RemoteTable(Table):
         train: bool = True,
     ):
         """Create an index on the table.
-        Currently, the only parameters that matter are
-        the metric and the vector column name.
 
         Parameters
         ----------
@@ -250,11 +248,6 @@ class RemoteTable(Table):
         >>> table.create_index("l2", "vector") # doctest: +SKIP
         """
 
-        if num_sub_vectors is not None:
-            logging.warning(
-                "num_sub_vectors is not supported on LanceDB cloud."
-                "This parameter will be tuned automatically."
-            )
         if accelerator is not None:
             logging.warning(
                 "GPU accelerator is not yet supported on LanceDB cloud."

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
@@ -1329,7 +1331,7 @@ class Table(ABC):
         1  2  [3.0, 4.0]
         2  3  [5.0, 6.0]
         >>> table.delete("x = 2")
-        DeleteResult(version=2)
+        DeleteResult(num_deleted_rows=1, version=2)
         >>> table.to_pandas()
            x      vector
         0  1  [1.0, 2.0]
@@ -1343,7 +1345,7 @@ class Table(ABC):
         >>> to_remove
         '1, 5'
         >>> table.delete(f"x IN ({to_remove})")
-        DeleteResult(version=3)
+        DeleteResult(num_deleted_rows=1, version=3)
         >>> table.to_pandas()
            x      vector
         0  3  [5.0, 6.0]
@@ -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:
         """
@@ -4166,6 +4181,7 @@ class AsyncTable:
                 when_not_matched_by_source_condition=merge._when_not_matched_by_source_condition,
                 timeout=merge._timeout,
                 use_index=merge._use_index,
+                mem_wal=merge._mem_wal,
             ),
         )
 
@@ -4200,7 +4216,7 @@ class AsyncTable:
         1  2  [3.0, 4.0]
         2  3  [5.0, 6.0]
         >>> table.delete("x = 2")
-        DeleteResult(version=2)
+        DeleteResult(num_deleted_rows=1, version=2)
         >>> table.to_pandas()
            x      vector
         0  1  [1.0, 2.0]
@@ -4214,7 +4230,7 @@ class AsyncTable:
         >>> to_remove
         '1, 5'
         >>> table.delete(f"x IN ({to_remove})")
-        DeleteResult(version=3)
+        DeleteResult(num_deleted_rows=1, version=3)
         >>> table.to_pandas()
            x      vector
         0  3  [5.0, 6.0]

python/python/lancedb/util.py

@@ -324,6 +324,16 @@ def _(value: list):
     return "[" + ", ".join(map(value_to_sql, value)) + "]"
 
 
+@value_to_sql.register(dict)
+def _(value: dict):
+    # https://datafusion.apache.org/user-guide/sql/scalar_functions.html#named-struct
+    return (
+        "named_struct("
+        + ", ".join(f"'{k}', {value_to_sql(v)}" for k, v in value.items())
+        + ")"
+    )
+
+
 @value_to_sql.register(np.ndarray)
 def _(value: np.ndarray):
     return value_to_sql(value.tolist())

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_fts.py

@@ -27,6 +27,7 @@ from lancedb.query import (
     PhraseQuery,
     BooleanQuery,
     Occur,
+    LanceFtsQueryBuilder,
 )
 import numpy as np
 import pyarrow as pa
@@ -882,3 +883,109 @@ def test_fts_query_to_json():
         '"must_not":[]}}'
     )
     assert json_str == expected
+
+
+def test_fts_fast_search(table):
+    table.create_fts_index("text", use_tantivy=False)
+
+    # Insert some unindexed data
+    table.add(
+        [
+            {
+                "text": "xyz",
+                "vector": [0 for _ in range(128)],
+                "id": 101,
+                "text2": "xyz",
+                "nested": {"text": "xyz"},
+                "count": 10,
+            }
+        ]
+    )
+
+    # Without fast_search, the query object should not have fast_search set
+    builder = table.search("xyz", query_type="fts").limit(10)
+    query = builder.to_query_object()
+    assert query.fast_search is None
+
+    # With fast_search, the query object should have fast_search=True
+    builder = table.search("xyz", query_type="fts").fast_search().limit(10)
+    query = builder.to_query_object()
+    assert query.fast_search is True
+
+    # fast_search should be chainable with other methods
+    builder = (
+        table.search("xyz", query_type="fts").fast_search().select(["text"]).limit(5)
+    )
+    query = builder.to_query_object()
+    assert query.fast_search is True
+    assert query.limit == 5
+    assert query.columns == ["text"]
+
+    # fast_search should be enabled by keyword argument too
+    query = LanceFtsQueryBuilder(table, "xyz", fast_search=True).to_query_object()
+    assert query.fast_search is True
+
+    # Verify it executes without error and skips unindexed data
+    results = table.search("xyz", query_type="fts").fast_search().limit(5).to_list()
+    assert len(results) == 0
+
+    # Update index and verify it returns results
+    table.optimize()
+    results = table.search("xyz", query_type="fts").fast_search().limit(5).to_list()
+    assert len(results) > 0
+
+
+@pytest.mark.asyncio
+async def test_fts_fast_search_async(async_table):
+    await async_table.create_index("text", config=FTS())
+
+    # Insert some unindexed data
+    await async_table.add(
+        [
+            {
+                "text": "xyz",
+                "vector": [0 for _ in range(128)],
+                "id": 101,
+                "text2": "xyz",
+                "nested": {"text": "xyz"},
+                "count": 10,
+            }
+        ]
+    )
+
+    # Without fast_search, should return results
+    results = await async_table.query().nearest_to_text("xyz").limit(5).to_list()
+    assert len(results) > 0
+
+    # With fast_search, should return no results data unindexed
+    fast_results = (
+        await async_table.query()
+        .nearest_to_text("xyz")
+        .fast_search()
+        .limit(5)
+        .to_list()
+    )
+    assert len(fast_results) == 0
+
+    # Update index and verify it returns results
+    await async_table.optimize()
+
+    fast_results = (
+        await async_table.query()
+        .nearest_to_text("xyz")
+        .fast_search()
+        .limit(5)
+        .to_list()
+    )
+    assert len(fast_results) > 0
+
+    # fast_search should be chainable with other methods
+    results = (
+        await async_table.query()
+        .nearest_to_text("xyz")
+        .fast_search()
+        .select(["text"])
+        .limit(5)
+        .to_list()
+    )
+    assert len(results) > 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

@@ -121,6 +121,32 @@ def test_value_to_sql_string(tmp_path):
         assert table.to_pandas().query("search == @value")["replace"].item() == value
 
 
+def test_value_to_sql_dict():
+    # Simple flat struct
+    assert value_to_sql({"a": 1, "b": "hello"}) == "named_struct('a', 1, 'b', 'hello')"
+
+    # Nested struct
+    assert (
+        value_to_sql({"outer": {"inner": 1}})
+        == "named_struct('outer', named_struct('inner', 1))"
+    )
+
+    # List inside struct
+    assert value_to_sql({"a": [1, 2]}) == "named_struct('a', [1, 2])"
+
+    # Mixed types
+    assert (
+        value_to_sql({"name": "test", "count": 42, "rate": 3.14, "active": True})
+        == "named_struct('name', 'test', 'count', 42, 'rate', 3.14, 'active', TRUE)"
+    )
+
+    # Null value inside struct
+    assert value_to_sql({"a": None}) == "named_struct('a', NULL)"
+
+    # Empty dict
+    assert value_to_sql({}) == "named_struct()"
+
+
 def test_append_vector_columns():
     registry = EmbeddingFunctionRegistry.get_instance()
     registry.register("test")(MockTextEmbeddingFunction)
@@ -292,18 +318,14 @@ 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/arrow.rs

@@ -10,7 +10,7 @@ use arrow::{
 use futures::stream::StreamExt;
 use lancedb::arrow::SendableRecordBatchStream;
 use pyo3::{
-    exceptions::PyStopAsyncIteration, pyclass, pymethods, Bound, Py, PyAny, PyRef, PyResult, Python,
+    Bound, Py, PyAny, PyRef, PyResult, Python, exceptions::PyStopAsyncIteration, pyclass, pymethods,
 };
 use pyo3_async_runtimes::tokio::future_into_py;
 

python/src/connection.rs

@@ -9,10 +9,10 @@ use lancedb::{
     database::{CreateTableMode, Database, ReadConsistency},
 };
 use pyo3::{
+    Bound, FromPyObject, Py, PyAny, PyRef, PyResult, Python,
     exceptions::{PyRuntimeError, PyValueError},
     pyclass, pyfunction, pymethods,
     types::{PyDict, PyDictMethods},
-    Bound, FromPyObject, Py, PyAny, PyRef, PyResult, Python,
 };
 use pyo3_async_runtimes::tokio::future_into_py;
 
@@ -506,6 +506,7 @@ pub struct PyClientConfig {
     id_delimiter: Option<String>,
     tls_config: Option<PyClientTlsConfig>,
     header_provider: Option<Py<PyAny>>,
+    mem_wal_enabled: Option<bool>,
 }
 
 #[derive(FromPyObject)]
@@ -590,6 +591,7 @@ impl From<PyClientConfig> for lancedb::remote::ClientConfig {
             id_delimiter: value.id_delimiter,
             tls_config: value.tls_config.map(Into::into),
             header_provider,
+            mem_wal_enabled: value.mem_wal_enabled,
         }
     }
 }

python/src/error.rs

@@ -2,10 +2,10 @@
 // SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
 use pyo3::{
+    PyErr, PyResult, Python,
     exceptions::{PyIOError, PyNotImplementedError, PyOSError, PyRuntimeError, PyValueError},
     intern,
     types::{PyAnyMethods, PyNone},
-    PyErr, PyResult, Python,
 };
 
 use lancedb::error::Error as LanceError;

python/src/index.rs

@@ -3,17 +3,17 @@
 
 use lancedb::index::vector::{IvfFlatIndexBuilder, IvfRqIndexBuilder, IvfSqIndexBuilder};
 use lancedb::index::{
+    Index as LanceDbIndex,
     scalar::{BTreeIndexBuilder, FtsIndexBuilder},
     vector::{IvfHnswPqIndexBuilder, IvfHnswSqIndexBuilder, IvfPqIndexBuilder},
-    Index as LanceDbIndex,
 };
-use pyo3::types::PyStringMethods;
 use pyo3::IntoPyObject;
+use pyo3::types::PyStringMethods;
 use pyo3::{
+    Bound, FromPyObject, PyAny, PyResult, Python,
     exceptions::{PyKeyError, PyValueError},
     intern, pyclass, pymethods,
     types::PyAnyMethods,
-    Bound, FromPyObject, PyAny, PyResult, Python,
 };
 
 use crate::util::parse_distance_type;
@@ -41,7 +41,12 @@ pub fn extract_index_params(source: &Option<Bound<'_, PyAny>>) -> PyResult<Lance
                 let inner_opts = FtsIndexBuilder::default()
                     .base_tokenizer(params.base_tokenizer)
                     .language(&params.language)
-                    .map_err(|_| PyValueError::new_err(format!("LanceDB does not support the requested language: '{}'", params.language)))?
+                    .map_err(|_| {
+                        PyValueError::new_err(format!(
+                            "LanceDB does not support the requested language: '{}'",
+                            params.language
+                        ))
+                    })?
                     .with_position(params.with_position)
                     .lower_case(params.lower_case)
                     .max_token_length(params.max_token_length)
@@ -52,7 +57,7 @@ pub fn extract_index_params(source: &Option<Bound<'_, PyAny>>) -> PyResult<Lance
                     .ngram_max_length(params.ngram_max_length)
                     .ngram_prefix_only(params.prefix_only);
                 Ok(LanceDbIndex::FTS(inner_opts))
-            },
+            }
             "IvfFlat" => {
                 let params = source.extract::<IvfFlatParams>()?;
                 let distance_type = parse_distance_type(params.distance_type)?;
@@ -64,10 +69,11 @@ pub fn extract_index_params(source: &Option<Bound<'_, PyAny>>) -> PyResult<Lance
                     ivf_flat_builder = ivf_flat_builder.num_partitions(num_partitions);
                 }
                 if let Some(target_partition_size) = params.target_partition_size {
-                    ivf_flat_builder = ivf_flat_builder.target_partition_size(target_partition_size);
+                    ivf_flat_builder =
+                        ivf_flat_builder.target_partition_size(target_partition_size);
                 }
                 Ok(LanceDbIndex::IvfFlat(ivf_flat_builder))
-            },
+            }
             "IvfPq" => {
                 let params = source.extract::<IvfPqParams>()?;
                 let distance_type = parse_distance_type(params.distance_type)?;
@@ -86,7 +92,7 @@ pub fn extract_index_params(source: &Option<Bound<'_, PyAny>>) -> PyResult<Lance
                     ivf_pq_builder = ivf_pq_builder.num_sub_vectors(num_sub_vectors);
                 }
                 Ok(LanceDbIndex::IvfPq(ivf_pq_builder))
-            },
+            }
             "IvfSq" => {
                 let params = source.extract::<IvfSqParams>()?;
                 let distance_type = parse_distance_type(params.distance_type)?;
@@ -101,7 +107,7 @@ pub fn extract_index_params(source: &Option<Bound<'_, PyAny>>) -> PyResult<Lance
                     ivf_sq_builder = ivf_sq_builder.target_partition_size(target_partition_size);
                 }
                 Ok(LanceDbIndex::IvfSq(ivf_sq_builder))
-            },
+            }
             "IvfRq" => {
                 let params = source.extract::<IvfRqParams>()?;
                 let distance_type = parse_distance_type(params.distance_type)?;
@@ -117,7 +123,7 @@ pub fn extract_index_params(source: &Option<Bound<'_, PyAny>>) -> PyResult<Lance
                     ivf_rq_builder = ivf_rq_builder.target_partition_size(target_partition_size);
                 }
                 Ok(LanceDbIndex::IvfRq(ivf_rq_builder))
-            },
+            }
             "HnswPq" => {
                 let params = source.extract::<IvfHnswPqParams>()?;
                 let distance_type = parse_distance_type(params.distance_type)?;
@@ -138,7 +144,7 @@ pub fn extract_index_params(source: &Option<Bound<'_, PyAny>>) -> PyResult<Lance
                     hnsw_pq_builder = hnsw_pq_builder.num_sub_vectors(num_sub_vectors);
                 }
                 Ok(LanceDbIndex::IvfHnswPq(hnsw_pq_builder))
-            },
+            }
             "HnswSq" => {
                 let params = source.extract::<IvfHnswSqParams>()?;
                 let distance_type = parse_distance_type(params.distance_type)?;
@@ -155,7 +161,7 @@ pub fn extract_index_params(source: &Option<Bound<'_, PyAny>>) -> PyResult<Lance
                     hnsw_sq_builder = hnsw_sq_builder.target_partition_size(target_partition_size);
                 }
                 Ok(LanceDbIndex::IvfHnswSq(hnsw_sq_builder))
-            },
+            }
             not_supported => Err(PyValueError::new_err(format!(
                 "Invalid index type '{}'.  Must be one of BTree, Bitmap, LabelList, FTS, IvfPq, IvfSq, IvfHnswPq, or IvfHnswSq",
                 not_supported

python/src/lib.rs

@@ -2,14 +2,14 @@
 // SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
 use arrow::RecordBatchStream;
-use connection::{connect, Connection};
+use connection::{Connection, connect};
 use env_logger::Env;
 use index::IndexConfig;
 use permutation::{PyAsyncPermutationBuilder, PyPermutationReader};
 use pyo3::{
-    pymodule,
+    Bound, PyResult, Python, pymodule,
     types::{PyModule, PyModuleMethods},
-    wrap_pyfunction, Bound, PyResult, Python,
+    wrap_pyfunction,
 };
 use query::{FTSQuery, HybridQuery, Query, VectorQuery};
 use session::Session;

python/src/permutation.rs

@@ -16,17 +16,32 @@ use lancedb::{
     query::Select,
 };
 use pyo3::{
+    Bound, PyAny, PyRef, PyRefMut, PyResult, Python,
     exceptions::PyRuntimeError,
     pyclass, pymethods,
     types::{PyAnyMethods, PyDict, PyDictMethods, PyType},
-    Bound, PyAny, PyRef, PyRefMut, PyResult, Python,
 };
 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/query.rs

@@ -4,9 +4,9 @@
 use std::sync::Arc;
 use std::time::Duration;
 
-use arrow::array::make_array;
 use arrow::array::Array;
 use arrow::array::ArrayData;
+use arrow::array::make_array;
 use arrow::pyarrow::FromPyArrow;
 use arrow::pyarrow::IntoPyArrow;
 use arrow::pyarrow::ToPyArrow;
@@ -22,23 +22,23 @@ use lancedb::query::{
     VectorQuery as LanceDbVectorQuery,
 };
 use lancedb::table::AnyQuery;
-use pyo3::prelude::{PyAnyMethods, PyDictMethods};
-use pyo3::pyfunction;
-use pyo3::pymethods;
-use pyo3::types::PyList;
-use pyo3::types::{PyDict, PyString};
 use pyo3::Bound;
 use pyo3::IntoPyObject;
 use pyo3::PyAny;
 use pyo3::PyRef;
 use pyo3::PyResult;
 use pyo3::Python;
-use pyo3::{exceptions::PyRuntimeError, FromPyObject};
+use pyo3::prelude::{PyAnyMethods, PyDictMethods};
+use pyo3::pyfunction;
+use pyo3::pymethods;
+use pyo3::types::PyList;
+use pyo3::types::{PyDict, PyString};
+use pyo3::{FromPyObject, exceptions::PyRuntimeError};
+use pyo3::{PyErr, pyclass};
 use pyo3::{
     exceptions::{PyNotImplementedError, PyValueError},
     intern,
 };
-use pyo3::{pyclass, PyErr};
 use pyo3_async_runtimes::tokio::future_into_py;
 
 use crate::util::parse_distance_type;

python/src/session.rs

@@ -4,7 +4,7 @@
 use std::sync::Arc;
 
 use lancedb::{ObjectStoreRegistry, Session as LanceSession};
-use pyo3::{pyclass, pymethods, PyResult};
+use pyo3::{PyResult, pyclass, pymethods};
 
 /// A session for managing caches and object stores across LanceDB operations.
 ///

python/src/storage_options.rs

@@ -66,13 +66,10 @@ impl StorageOptionsProvider for PyStorageOptionsProviderWrapper {
                     .inner
                     .bind(py)
                     .call_method0("fetch_storage_options")
-                    .map_err(|e| lance_core::Error::IO {
-                        source: Box::new(std::io::Error::other(format!(
-                            "Failed to call fetch_storage_options: {}",
-                            e
-                        ))),
-                        location: snafu::location!(),
-                    })?;
+                    .map_err(|e| lance_core::Error::io_source(Box::new(std::io::Error::other(format!(
+                        "Failed to call fetch_storage_options: {}",
+                        e
+                    )))))?;
 
                 // If result is None, return None
                 if result.is_none() {
@@ -81,26 +78,19 @@ impl StorageOptionsProvider for PyStorageOptionsProviderWrapper {
 
                 // Extract the result dict - should be a flat Map<String, String>
                 let result_dict = result.downcast::<PyDict>().map_err(|_| {
-                    lance_core::Error::InvalidInput {
-                        source: "fetch_storage_options() must return None or a dict of string key-value pairs".into(),
-                        location: snafu::location!(),
-                    }
+                    lance_core::Error::invalid_input(
+                        "fetch_storage_options() must return a dict of string key-value pairs or None",
+                    )
                 })?;
 
                 // Convert all entries to HashMap<String, String>
                 let mut storage_options = HashMap::new();
                 for (key, value) in result_dict.iter() {
                     let key_str: String = key.extract().map_err(|e| {
-                        lance_core::Error::InvalidInput {
-                            source: format!("Storage option key must be a string: {}", e).into(),
-                            location: snafu::location!(),
-                        }
+                        lance_core::Error::invalid_input(format!("Storage option key must be a string: {}", e))
                     })?;
                     let value_str: String = value.extract().map_err(|e| {
-                        lance_core::Error::InvalidInput {
-                            source: format!("Storage option value must be a string: {}", e).into(),
-                            location: snafu::location!(),
-                        }
+                        lance_core::Error::invalid_input(format!("Storage option value must be a string: {}", e))
                     })?;
                     storage_options.insert(key_str, value_str);
                 }
@@ -109,13 +99,10 @@ impl StorageOptionsProvider for PyStorageOptionsProviderWrapper {
             })
         })
         .await
-        .map_err(|e| lance_core::Error::IO {
-            source: Box::new(std::io::Error::other(format!(
-                "Task join error: {}",
-                e
-            ))),
-            location: snafu::location!(),
-        })?
+        .map_err(|e| lance_core::Error::io_source(Box::new(std::io::Error::other(format!(
+            "Task join error: {}",
+            e
+        )))))?
     }
 
     fn provider_id(&self) -> String {

python/src/table.rs

@@ -5,8 +5,9 @@ use std::{collections::HashMap, sync::Arc};
 use crate::{
     connection::Connection,
     error::PythonErrorExt,
-    index::{extract_index_params, IndexConfig},
+    index::{IndexConfig, extract_index_params},
     query::{Query, TakeQuery},
+    table::scannable::PyScannable,
 };
 use arrow::{
     datatypes::{DataType, Schema},
@@ -18,13 +19,15 @@ use lancedb::table::{
     Table as LanceDbTable,
 };
 use pyo3::{
+    Bound, FromPyObject, PyAny, PyRef, PyResult, Python,
     exceptions::{PyKeyError, PyRuntimeError, PyValueError},
     pyclass, pymethods,
     types::{IntoPyDict, PyAnyMethods, PyDict, PyDictMethods},
-    Bound, FromPyObject, PyAny, PyRef, PyResult, Python,
 };
 use pyo3_async_runtimes::tokio::future_into_py;
 
+mod scannable;
+
 /// Statistics about a compaction operation.
 #[pyclass(get_all)]
 #[derive(Clone, Debug)]
@@ -109,19 +112,24 @@ impl From<lancedb::table::AddResult> for AddResult {
 #[pyclass(get_all)]
 #[derive(Clone, Debug)]
 pub struct DeleteResult {
+    pub num_deleted_rows: u64,
     pub version: u64,
 }
 
 #[pymethods]
 impl DeleteResult {
     pub fn __repr__(&self) -> String {
-        format!("DeleteResult(version={})", self.version)
+        format!(
+            "DeleteResult(num_deleted_rows={}, version={})",
+            self.num_deleted_rows, self.version
+        )
     }
 }
 
 impl From<lancedb::table::DeleteResult> for DeleteResult {
     fn from(result: lancedb::table::DeleteResult) -> Self {
         Self {
+            num_deleted_rows: result.num_deleted_rows,
             version: result.version,
         }
     }
@@ -293,12 +301,10 @@ impl Table {
 
     pub fn add<'a>(
         self_: PyRef<'a, Self>,
-        data: Bound<'_, PyAny>,
+        data: PyScannable,
         mode: String,
     ) -> PyResult<Bound<'a, PyAny>> {
-        let batches: Box<dyn arrow::array::RecordBatchReader + Send> =
-            Box::new(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" {
@@ -536,7 +542,7 @@ impl Table {
         let inner = self_.inner_ref()?.clone();
         future_into_py(self_.py(), async move {
             let versions = inner.list_versions().await.infer_error()?;
-            let versions_as_dict = Python::attach(|py| {
+            Python::attach(|py| {
                 versions
                     .iter()
                     .map(|v| {
@@ -553,9 +559,7 @@ impl Table {
                         Ok(dict.unbind())
                     })
                     .collect::<PyResult<Vec<_>>>()
-            });
-
-            versions_as_dict
+            })
         })
     }
 
@@ -706,6 +710,9 @@ impl Table {
         if let Some(use_index) = parameters.use_index {
             builder.use_index(use_index);
         }
+        if let Some(mem_wal) = parameters.mem_wal {
+            builder.mem_wal(mem_wal);
+        }
 
         future_into_py(self_.py(), async move {
             let res = builder.execute(Box::new(batches)).await.infer_error()?;
@@ -866,6 +873,7 @@ pub struct MergeInsertParams {
     when_not_matched_by_source_condition: Option<String>,
     timeout: Option<std::time::Duration>,
     use_index: Option<bool>,
+    mem_wal: Option<bool>,
 }
 
 #[pyclass]

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::{
+    Error,
+    arrow::{SendableRecordBatchStream, SimpleRecordBatchStream},
+    data::scannable::Scannable,
+};
+use pyo3::{FromPyObject, Py, PyAny, Python, types::PyAnyMethods};
+
+/// 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
+                            && 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,
+        })
+    }
+}

python/src/util.rs

@@ -5,8 +5,9 @@ use std::sync::Mutex;
 
 use lancedb::DistanceType;
 use pyo3::{
+    PyResult,
     exceptions::{PyRuntimeError, PyValueError},
-    pyfunction, PyResult,
+    pyfunction,
 };
 
 /// A wrapper around a rust builder

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.3"
 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.workspace = true
 datafusion-physical-expr.workspace = true
+datafusion-sql.workspace = true
 datafusion-physical-plan.workspace = true
 datafusion.workspace = true
 object_store = { workspace = true }

rust/lancedb/examples/bedrock.rs

@@ -9,10 +9,9 @@ use aws_config::Region;
 use aws_sdk_bedrockruntime::Client;
 use futures::StreamExt;
 use lancedb::{
-    connect,
-    embeddings::{bedrock::BedrockEmbeddingFunction, EmbeddingDefinition, EmbeddingFunction},
+    Result, connect,
+    embeddings::{EmbeddingDefinition, EmbeddingFunction, bedrock::BedrockEmbeddingFunction},
     query::{ExecutableQuery, QueryBase},
-    Result,
 };
 
 #[tokio::main]

rust/lancedb/examples/full_text_search.rs

@@ -10,10 +10,10 @@ use futures::TryStreamExt;
 use lance_index::scalar::FullTextSearchQuery;
 use lancedb::connection::Connection;
 
-use lancedb::index::scalar::FtsIndexBuilder;
 use lancedb::index::Index;
+use lancedb::index::scalar::FtsIndexBuilder;
 use lancedb::query::{ExecutableQuery, QueryBase};
-use lancedb::{connect, Result, Table};
+use lancedb::{Result, Table, connect};
 use rand::random;
 
 #[tokio::main]
@@ -46,19 +46,21 @@ fn create_some_records() -> Result<Box<dyn arrow_array::RecordBatchReader + Send
         .collect::<Vec<_>>();
     let n_terms = 3;
     let batches = RecordBatchIterator::new(
-        vec![RecordBatch::try_new(
-            schema.clone(),
-            vec![
-                Arc::new(Int32Array::from_iter_values(0..TOTAL as i32)),
-                Arc::new(StringArray::from_iter_values((0..TOTAL).map(|_| {
-                    (0..n_terms)
-                        .map(|_| words[random::<u32>() as usize % words.len()])
-                        .collect::<Vec<_>>()
-                        .join(" ")
-                }))),
-            ],
-        )
-        .unwrap()]
+        vec![
+            RecordBatch::try_new(
+                schema.clone(),
+                vec![
+                    Arc::new(Int32Array::from_iter_values(0..TOTAL as i32)),
+                    Arc::new(StringArray::from_iter_values((0..TOTAL).map(|_| {
+                        (0..n_terms)
+                            .map(|_| words[random::<u32>() as usize % words.len()])
+                            .collect::<Vec<_>>()
+                            .join(" ")
+                    }))),
+                ],
+            )
+            .unwrap(),
+        ]
         .into_iter()
         .map(Ok),
         schema.clone(),

rust/lancedb/examples/hybrid_search.rs

@@ -5,16 +5,15 @@ 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::index::scalar::FtsIndexBuilder;
 use lancedb::{
-    connect,
+    Result, Table, connect,
     embeddings::{
-        sentence_transformers::SentenceTransformersEmbeddings, EmbeddingDefinition,
-        EmbeddingFunction,
+        EmbeddingDefinition, EmbeddingFunction,
+        sentence_transformers::SentenceTransformersEmbeddings,
     },
     query::{QueryBase, QueryExecutionOptions},
-    Result, Table,
 };
 use std::{iter::once, sync::Arc};
 

rust/lancedb/examples/ivf_pq.rs

@@ -14,10 +14,10 @@ use arrow_schema::{DataType, Field, Schema};
 use futures::TryStreamExt;
 use lancedb::connection::Connection;
 
-use lancedb::index::vector::IvfPqIndexBuilder;
 use lancedb::index::Index;
+use lancedb::index::vector::IvfPqIndexBuilder;
 use lancedb::query::{ExecutableQuery, QueryBase};
-use lancedb::{connect, DistanceType, Result, Table};
+use lancedb::{DistanceType, Result, Table, connect};
 
 #[tokio::main]
 async fn main() -> Result<()> {
@@ -51,19 +51,21 @@ fn create_some_records() -> Result<Box<dyn arrow_array::RecordBatchReader + Send
 
     // 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,
+        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()]
+                ],
+            )
+            .unwrap(),
+        ]
         .into_iter()
         .map(Ok),
         schema.clone(),

rust/lancedb/examples/openai.rs

@@ -8,10 +8,9 @@ use std::{iter::once, sync::Arc};
 use arrow_array::{RecordBatch, StringArray};
 use futures::StreamExt;
 use lancedb::{
-    connect,
-    embeddings::{openai::OpenAIEmbeddingFunction, EmbeddingDefinition, EmbeddingFunction},
+    Result, connect,
+    embeddings::{EmbeddingDefinition, EmbeddingFunction, openai::OpenAIEmbeddingFunction},
     query::{ExecutableQuery, QueryBase},
-    Result,
 };
 
 // --8<-- [end:imports]

rust/lancedb/examples/sentence_transformers.rs

@@ -7,13 +7,12 @@ use arrow_array::{RecordBatch, StringArray};
 use arrow_schema::{DataType, Field, Schema};
 use futures::StreamExt;
 use lancedb::{
-    connect,
+    Result, connect,
     embeddings::{
-        sentence_transformers::SentenceTransformersEmbeddings, EmbeddingDefinition,
-        EmbeddingFunction,
+        EmbeddingDefinition, EmbeddingFunction,
+        sentence_transformers::SentenceTransformersEmbeddings,
     },
     query::{ExecutableQuery, QueryBase},
-    Result,
 };
 
 #[tokio::main]

rust/lancedb/examples/simple.rs

@@ -14,7 +14,7 @@ use futures::TryStreamExt;
 use lancedb::connection::Connection;
 use lancedb::index::Index;
 use lancedb::query::{ExecutableQuery, QueryBase};
-use lancedb::{connect, Result, Table as LanceDbTable};
+use lancedb::{Result, Table as LanceDbTable, connect};
 
 #[tokio::main]
 async fn main() -> Result<()> {

rust/lancedb/src/arrow.rs

@@ -12,7 +12,7 @@ use lance_datagen::{BatchCount, BatchGeneratorBuilder, RowCount};
 #[cfg(feature = "polars")]
 use {crate::polars_arrow_convertors, polars::frame::ArrowChunk, polars::prelude::DataFrame};
 
-use crate::{error::Result, Error};
+use crate::{Error, error::Result};
 
 /// An iterator of batches that also has a schema
 pub trait RecordBatchReader: Iterator<Item = Result<arrow_array::RecordBatch>> {
@@ -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

@@ -17,6 +17,7 @@ use lance_namespace::models::{
 #[cfg(feature = "aws")]
 use object_store::aws::AwsCredential;
 
+use crate::Table;
 use crate::connection::create_table::CreateTableBuilder;
 use crate::data::scannable::Scannable;
 use crate::database::listing::ListingDatabase;
@@ -31,7 +32,6 @@ use crate::remote::{
     client::ClientConfig,
     db::{OPT_REMOTE_API_KEY, OPT_REMOTE_HOST_OVERRIDE, OPT_REMOTE_REGION},
 };
-use crate::Table;
 use lance::io::ObjectStoreParams;
 pub use lance_encoding::version::LanceFileVersion;
 #[cfg(feature = "remote")]
@@ -566,8 +566,11 @@ pub struct ConnectBuilder {
 }
 
 #[cfg(feature = "remote")]
-const ENV_VARS_TO_STORAGE_OPTS: [(&str, &str); 1] =
-    [("AZURE_STORAGE_ACCOUNT_NAME", "azure_storage_account_name")];
+const ENV_VARS_TO_STORAGE_OPTS: [(&str, &str); 3] = [
+    ("AZURE_STORAGE_ACCOUNT_NAME", "azure_storage_account_name"),
+    ("AZURE_CLIENT_ID", "azure_client_id"),
+    ("AZURE_TENANT_ID", "azure_tenant_id"),
+];
 
 impl ConnectBuilder {
     /// Create a new [`ConnectOptions`] with the given database URI.
@@ -758,10 +761,10 @@ impl ConnectBuilder {
         options: &mut HashMap<String, String>,
     ) {
         for (env_key, opt_key) in env_var_to_remote_storage_option {
-            if let Ok(env_value) = std::env::var(env_key) {
-                if !options.contains_key(*opt_key) {
-                    options.insert((*opt_key).to_string(), env_value);
-                }
+            if let Ok(env_value) = std::env::var(env_key)
+                && !options.contains_key(*opt_key)
+            {
+                options.insert((*opt_key).to_string(), env_value);
             }
         }
     }
@@ -781,13 +784,19 @@ impl ConnectBuilder {
             message: "An api_key is required when connecting to LanceDb Cloud".to_string(),
         })?;
 
+        // Propagate mem_wal_enabled from options to client_config
+        let mut client_config = self.request.client_config;
+        if options.mem_wal_enabled.is_some() {
+            client_config.mem_wal_enabled = options.mem_wal_enabled;
+        }
+
         let storage_options = StorageOptions(options.storage_options.clone());
         let internal = Arc::new(crate::remote::db::RemoteDatabase::try_new(
             &self.request.uri,
             &api_key,
             &region,
             options.host_override,
-            self.request.client_config,
+            client_config,
             storage_options.into(),
         )?);
         Ok(Connection {
@@ -1011,14 +1020,13 @@ mod tests {
     #[cfg(feature = "remote")]
     #[test]
     fn test_apply_env_defaults() {
-        let env_key = "TEST_APPLY_ENV_DEFAULTS_ENVIRONMENT_VARIABLE_ENV_KEY";
-        let env_val = "TEST_APPLY_ENV_DEFAULTS_ENVIRONMENT_VARIABLE_ENV_VAL";
+        let env_key = "PATH";
+        let env_val = std::env::var(env_key).expect("PATH should be set in test environment");
         let opts_key = "test_apply_env_defaults_environment_variable_opts_key";
-        std::env::set_var(env_key, env_val);
 
         let mut options = HashMap::new();
         ConnectBuilder::apply_env_defaults(&[(env_key, opts_key)], &mut options);
-        assert_eq!(Some(&env_val.to_string()), options.get(opts_key));
+        assert_eq!(Some(&env_val), options.get(opts_key));
 
         options.insert(opts_key.to_string(), "EXPLICIT-VALUE".to_string());
         ConnectBuilder::apply_env_defaults(&[(env_key, opts_key)], &mut options);

rust/lancedb/src/connection/create_table.rs

@@ -6,12 +6,12 @@ use std::sync::Arc;
 use lance_io::object_store::StorageOptionsProvider;
 
 use crate::{
+    Error, Result, Table,
     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 {
@@ -167,7 +167,7 @@ impl CreateTableBuilder {
 #[cfg(test)]
 mod tests {
     use arrow_array::{
-        record_batch, Array, FixedSizeListArray, Float32Array, RecordBatch, RecordBatchIterator,
+        Array, FixedSizeListArray, Float32Array, RecordBatch, RecordBatchIterator, record_batch,
     };
     use arrow_schema::{ArrowError, DataType, Field, Schema};
     use futures::TryStreamExt;
@@ -380,11 +380,12 @@ mod tests {
             .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
+        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)

rust/lancedb/src/data/inspect.rs

@@ -5,9 +5,9 @@ use std::collections::HashMap;
 
 use arrow::compute::kernels::{aggregate::bool_and, length::length};
 use arrow_array::{
+    Array, GenericListArray, OffsetSizeTrait, PrimitiveArray, RecordBatchReader,
     cast::AsArray,
     types::{ArrowPrimitiveType, Int32Type, Int64Type},
-    Array, GenericListArray, OffsetSizeTrait, PrimitiveArray, RecordBatchReader,
 };
 use arrow_ord::cmp::eq;
 use arrow_schema::DataType;
@@ -78,7 +78,7 @@ pub fn infer_vector_columns(
                 _ => {
                     return Err(Error::Schema {
                         message: format!("Column {} is not a list", col_name),
-                    })
+                    });
                 }
             } {
                 if let Some(Some(prev_dim)) = columns_to_infer.get(&col_name) {
@@ -102,8 +102,8 @@ mod tests {
     use super::*;
 
     use arrow_array::{
-        types::{Float32Type, Float64Type},
         FixedSizeListArray, Float32Array, ListArray, RecordBatch, RecordBatchIterator, StringArray,
+        types::{Float32Type, Float64Type},
     };
     use arrow_schema::{DataType, Field, Schema};
     use std::{sync::Arc, vec};

rust/lancedb/src/data/sanitize.rs

@@ -4,10 +4,10 @@
 use std::{iter::repeat_with, sync::Arc};
 
 use arrow_array::{
-    cast::AsArray,
-    types::{Float16Type, Float32Type, Float64Type, Int32Type, Int64Type},
     Array, ArrowNumericType, FixedSizeListArray, PrimitiveArray, RecordBatch, RecordBatchIterator,
     RecordBatchReader,
+    cast::AsArray,
+    types::{Float16Type, Float32Type, Float64Type, Int32Type, Int64Type},
 };
 use arrow_cast::{can_cast_types, cast};
 use arrow_schema::{ArrowError, DataType, Field, Schema};
@@ -184,7 +184,7 @@ mod tests {
     use std::sync::Arc;
 
     use arrow_array::{
-        FixedSizeListArray, Float16Array, Float32Array, Float64Array, Int32Array, Int8Array,
+        FixedSizeListArray, Float16Array, Float32Array, Float64Array, Int8Array, Int32Array,
         RecordBatch, RecordBatchIterator, StringArray,
     };
     use arrow_schema::Field;

rust/lancedb/src/data/scannable.rs

@@ -9,22 +9,21 @@
 
 use std::sync::Arc;
 
-use arrow_array::{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;
-
 use crate::arrow::{
     SendableRecordBatchStream, SendableRecordBatchStreamExt, SimpleRecordBatchStream,
 };
 use crate::embeddings::{
-    compute_embeddings_for_batch, compute_output_schema, EmbeddingDefinition, EmbeddingFunction,
-    EmbeddingRegistry,
+    EmbeddingDefinition, EmbeddingFunction, EmbeddingRegistry, compute_embeddings_for_batch,
+    compute_output_schema,
 };
 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::StreamExt;
+use futures::stream::once;
+use lance_datafusion::utils::StreamingWriteSource;
 
 pub trait Scannable: Send {
     /// Returns the schema of the data.
@@ -228,6 +227,19 @@ impl WithEmbeddingsScannable {
         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,
@@ -245,9 +257,11 @@ impl Scannable for WithEmbeddingsScannable {
         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 || {
@@ -257,12 +271,29 @@ impl Scannable for WithEmbeddingsScannable {
                 .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: output_schema,
+            schema: stream_schema,
             stream: mapped_stream,
         })
     }
@@ -303,8 +334,13 @@ pub fn scannable_with_embeddings(
         }
 
         if !embeddings.is_empty() {
-            return Ok(Box::new(WithEmbeddingsScannable::try_new(
-                inner, embeddings,
+            // 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(),
             )?));
         }
     }
@@ -312,6 +348,133 @@ pub fn scannable_with_embeddings(
     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::*;
@@ -408,6 +571,231 @@ mod tests {
         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;

rust/lancedb/src/database.rs

@@ -19,12 +19,12 @@ use std::sync::Arc;
 use std::time::Duration;
 
 use lance::dataset::ReadParams;
+use lance_namespace::LanceNamespace;
 use lance_namespace::models::{
     CreateNamespaceRequest, CreateNamespaceResponse, DescribeNamespaceRequest,
     DescribeNamespaceResponse, DropNamespaceRequest, DropNamespaceResponse, ListNamespacesRequest,
     ListNamespacesResponse, ListTablesRequest, ListTablesResponse,
 };
-use lance_namespace::LanceNamespace;
 
 use crate::data::scannable::Scannable;
 use crate::error::Result;
@@ -85,8 +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
@@ -104,12 +106,6 @@ impl CreateTableMode {
     }
 }
 
-impl Default for CreateTableMode {
-    fn default() -> Self {
-        Self::Create
-    }
-}
-
 /// A request to create a table
 pub struct CreateTableRequest {
     /// The name of the new table

rust/lancedb/src/database/listing.rs

@@ -8,7 +8,7 @@ use std::path::Path;
 use std::{collections::HashMap, sync::Arc};
 
 use lance::dataset::refs::Ref;
-use lance::dataset::{builder::DatasetBuilder, ReadParams, WriteMode};
+use lance::dataset::{ReadParams, WriteMode, builder::DatasetBuilder};
 use lance::io::{ObjectStore, ObjectStoreParams, WrappingObjectStore};
 use lance_datafusion::utils::StreamingWriteSource;
 use lance_encoding::version::LanceFileVersion;
@@ -1097,11 +1097,11 @@ impl Database for ListingDatabase {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::Table;
     use crate::connection::ConnectRequest;
     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;

rust/lancedb/src/database/namespace.rs

@@ -7,17 +7,17 @@ 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,
-        DeclareTableRequest, DescribeNamespaceRequest, DescribeNamespaceResponse,
-        DescribeTableRequest, DropNamespaceRequest, DropNamespaceResponse, DropTableRequest,
-        ListNamespacesRequest, ListNamespacesResponse, ListTablesRequest, ListTablesResponse,
-    },
     LanceNamespace,
+    models::{
+        CreateNamespaceRequest, CreateNamespaceResponse, DeclareTableRequest,
+        DescribeNamespaceRequest, DescribeNamespaceResponse, DescribeTableRequest,
+        DropNamespaceRequest, DropNamespaceResponse, DropTableRequest, ListNamespacesRequest,
+        ListNamespacesResponse, ListTablesRequest, ListTablesResponse,
+    },
 };
 use lance_namespace_impls::ConnectBuilder;
-use log::warn;
 
 use crate::database::ReadConsistency;
 use crate::error::{Error, Result};
@@ -212,45 +212,30 @@ 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 {
+        let (location, initial_storage_options) = {
+            let response = self.namespace.declare_table(declare_request).await?;
+            let loc = 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, \
-                        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()
-                    };
+            })?;
+            // 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)
+        };
 
-                    #[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),
-                        })?;
-
-                    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),
-                    });
-                }
-            }
+        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 +245,7 @@ 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(),

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

@@ -11,16 +11,16 @@ use lance_core::ROW_ID;
 use lance_datafusion::exec::SessionContextExt;
 
 use crate::{
+    Error, Result, Table,
     arrow::{SendableRecordBatchStream, SendableRecordBatchStreamExt, SimpleRecordBatchStream},
     connect,
     database::{CreateTableRequest, Database},
     dataloader::permutation::{
         shuffle::{Shuffler, ShufflerConfig},
-        split::{SplitStrategy, Splitter, SPLIT_ID_COLUMN},
-        util::{rename_column, TemporaryDirectory},
+        split::{SPLIT_ID_COLUMN, SplitStrategy, Splitter},
+        util::{TemporaryDirectory, rename_column},
     },
     query::{ExecutableQuery, QueryBase, Select},
-    Error, Result, Table,
 };
 
 pub const SRC_ROW_ID_COL: &str = "row_id";
@@ -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

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

@@ -25,8 +25,8 @@ use futures::{StreamExt, TryStreamExt};
 use lance::dataset::scanner::DatasetRecordBatchStream;
 use lance::io::RecordBatchStream;
 use lance_arrow::RecordBatchExt;
-use lance_core::error::LanceOptionExt;
 use lance_core::ROW_ID;
+use lance_core::error::LanceOptionExt;
 use std::collections::HashMap;
 use std::sync::Arc;
 
@@ -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());
@@ -498,10 +500,10 @@ mod tests {
     use rand::seq::SliceRandom;
 
     use crate::{
+        Table,
         arrow::SendableRecordBatchStream,
         query::{ExecutableQuery, QueryBase},
-        test_utils::datagen::{virtual_table, LanceDbDatagenExt},
-        Table,
+        test_utils::datagen::{LanceDbDatagenExt, virtual_table},
     };
 
     use super::*;
@@ -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/shuffle.rs

@@ -18,12 +18,12 @@ use lance_io::{
     scheduler::{ScanScheduler, SchedulerConfig},
     utils::CachedFileSize,
 };
-use rand::{seq::SliceRandom, Rng, RngCore};
+use rand::{Rng, RngCore, seq::SliceRandom};
 
 use crate::{
-    arrow::{SendableRecordBatchStream, SimpleRecordBatchStream},
-    dataloader::permutation::util::{non_crypto_rng, TemporaryDirectory},
     Error, Result,
+    arrow::{SendableRecordBatchStream, SimpleRecordBatchStream},
+    dataloader::permutation::util::{TemporaryDirectory, non_crypto_rng},
 };
 
 #[derive(Debug, Clone)]
@@ -281,7 +281,7 @@ mod tests {
     use datafusion_expr::col;
     use futures::TryStreamExt;
     use lance_datagen::{BatchCount, BatchGeneratorBuilder, ByteCount, RowCount, Seed};
-    use rand::{rngs::SmallRng, SeedableRng};
+    use rand::{SeedableRng, rngs::SmallRng};
 
     fn test_gen() -> BatchGeneratorBuilder {
         lance_datagen::gen_batch()

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

@@ -2,8 +2,8 @@
 // SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
 use std::sync::{
-    atomic::{AtomicBool, AtomicU64, AtomicUsize, Ordering},
     Arc,
+    atomic::{AtomicBool, AtomicU64, AtomicUsize, Ordering},
 };
 
 use arrow_array::{Array, BooleanArray, RecordBatch, UInt64Array};
@@ -15,21 +15,22 @@ use lance_arrow::SchemaExt;
 use lance_core::ROW_ID;
 
 use crate::{
+    Error, Result,
     arrow::{SendableRecordBatchStream, SimpleRecordBatchStream},
     dataloader::{
         permutation::shuffle::{Shuffler, ShufflerConfig},
         permutation::util::TemporaryDirectory,
     },
     query::{Query, QueryBase, Select},
-    Error, Result,
 };
 
 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/dataloader/permutation/util.rs

@@ -7,12 +7,12 @@ use arrow_array::RecordBatch;
 use arrow_schema::{Fields, Schema};
 use datafusion_execution::disk_manager::DiskManagerMode;
 use futures::TryStreamExt;
-use rand::{rngs::SmallRng, RngCore, SeedableRng};
+use rand::{RngCore, SeedableRng, rngs::SmallRng};
 use tempfile::TempDir;
 
 use crate::{
-    arrow::{SendableRecordBatchStream, SimpleRecordBatchStream},
     Error, Result,
+    arrow::{SendableRecordBatchStream, SimpleRecordBatchStream},
 };
 
 /// Directory to use for temporary files

rust/lancedb/src/embeddings.rs

@@ -23,9 +23,9 @@ use arrow_schema::{DataType, Field, SchemaBuilder, SchemaRef};
 use serde::{Deserialize, Serialize};
 
 use crate::{
+    Error,
     error::Result,
     table::{ColumnDefinition, ColumnKind, TableDefinition},
-    Error,
 };
 
 /// Trait for embedding functions

rust/lancedb/src/embeddings/bedrock.rs

@@ -8,7 +8,7 @@ use arrow::array::{AsArray, Float32Builder};
 use arrow_array::{Array, ArrayRef, FixedSizeListArray, Float32Array};
 use arrow_data::ArrayData;
 use arrow_schema::DataType;
-use serde_json::{json, Value};
+use serde_json::{Value, json};
 
 use super::EmbeddingFunction;
 use crate::{Error, Result};

rust/lancedb/src/embeddings/openai.rs

@@ -8,9 +8,9 @@ use arrow_array::{Array, ArrayRef, FixedSizeListArray, Float32Array};
 use arrow_data::ArrayData;
 use arrow_schema::DataType;
 use async_openai::{
+    Client,
     config::OpenAIConfig,
     types::{CreateEmbeddingRequest, Embedding, EmbeddingInput, EncodingFormat},
-    Client,
 };
 use tokio::{runtime::Handle, task};
 

rust/lancedb/src/embeddings/sentence_transformers.rs

@@ -7,7 +7,7 @@ use super::EmbeddingFunction;
 use arrow::{
     array::{AsArray, PrimitiveBuilder},
     datatypes::{
-        ArrowPrimitiveType, Float16Type, Float32Type, Float64Type, Int64Type, UInt32Type, UInt8Type,
+        ArrowPrimitiveType, Float16Type, Float32Type, Float64Type, Int64Type, UInt8Type, UInt32Type,
     },
 };
 use arrow_array::{Array, FixedSizeListArray, PrimitiveArray};
@@ -16,8 +16,8 @@ use arrow_schema::DataType;
 use candle_core::{CpuStorage, Device, Layout, Storage, Tensor};
 use candle_nn::VarBuilder;
 use candle_transformers::models::bert::{BertModel, DTYPE};
-use hf_hub::{api::sync::Api, Repo, RepoType};
-use tokenizers::{tokenizer::Tokenizer, PaddingParams};
+use hf_hub::{Repo, RepoType, api::sync::Api};
+use tokenizers::{PaddingParams, tokenizer::Tokenizer};
 
 /// Compute embeddings using huggingface sentence-transformers.
 pub struct SentenceTransformersEmbeddingsBuilder {
@@ -230,7 +230,7 @@ impl SentenceTransformersEmbeddings {
             Storage::Cpu(CpuStorage::BF16(_)) => {
                 return Err(crate::Error::Runtime {
                     message: "unsupported data type".to_string(),
-                })
+                });
             }
             _ => unreachable!("we already moved the tensor to the CPU device"),
         };
@@ -298,12 +298,12 @@ impl SentenceTransformersEmbeddings {
             DataType::Utf8View => {
                 return Err(crate::Error::Runtime {
                     message: "Utf8View not yet implemented".to_string(),
-                })
+                });
             }
             _ => {
                 return Err(crate::Error::Runtime {
                     message: "invalid type".to_string(),
-                })
+                });
             }
         };
 

rust/lancedb/src/error.rs

@@ -4,6 +4,7 @@
 use std::sync::PoisonError;
 
 use arrow_schema::ArrowError;
+use datafusion_common::DataFusionError;
 use snafu::Snafu;
 
 pub(crate) type BoxError = Box<dyn std::error::Error + Send + Sync>;
@@ -96,28 +97,74 @@ pub type Result<T> = std::result::Result<T, Error>;
 impl From<ArrowError> for Error {
     fn from(source: ArrowError) -> Self {
         match source {
-            ArrowError::ExternalError(source) => match source.downcast::<Self>() {
-                Ok(e) => *e,
-                Err(source) => Self::External { 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 {
         // Try to unwrap external errors that were wrapped by lance
         match source {
-            lance::Error::Wrapped { error, .. } => match error.downcast::<Self>() {
-                Ok(e) => *e,
-                Err(source) => Self::External { 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 }
+    }
+}
+
 impl From<object_store::Error> for Error {
     fn from(source: object_store::Error) -> Self {
         Self::ObjectStore { 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, ScalarUDF, expr_fn::cast};
+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/index.rs

@@ -9,7 +9,7 @@ use std::time::Duration;
 use vector::IvfFlatIndexBuilder;
 
 use crate::index::vector::IvfRqIndexBuilder;
-use crate::{table::BaseTable, DistanceType, Error, Result};
+use crate::{DistanceType, Error, Result, table::BaseTable};
 
 use self::{
     scalar::{BTreeIndexBuilder, BitmapIndexBuilder, LabelListIndexBuilder},

rust/lancedb/src/index/scalar.rs

@@ -27,7 +27,7 @@
 ///
 /// The btree index does not currently have any parameters though parameters such as the
 /// block size may be added in the future.
-#[derive(Default, Debug, Clone)]
+#[derive(Default, Debug, Clone, serde::Serialize)]
 pub struct BTreeIndexBuilder {}
 
 impl BTreeIndexBuilder {}
@@ -39,7 +39,7 @@ impl BTreeIndexBuilder {}
 /// This index works best for low-cardinality (i.e., less than 1000 unique values) columns,
 /// where the number of unique values is small.
 /// The bitmap stores a list of row ids where the value is present.
-#[derive(Debug, Clone, Default)]
+#[derive(Debug, Clone, Default, serde::Serialize)]
 pub struct BitmapIndexBuilder {}
 
 /// Builder for LabelList index.
@@ -48,10 +48,10 @@ pub struct BitmapIndexBuilder {}
 /// support queries with `array_contains_all` and `array_contains_any`
 /// using an underlying bitmap index.
 ///
-#[derive(Debug, Clone, Default)]
+#[derive(Debug, Clone, Default, serde::Serialize)]
 pub struct LabelListIndexBuilder {}
 
-pub use lance_index::scalar::inverted::query::*;
 pub use lance_index::scalar::FullTextSearchQuery;
 pub use lance_index::scalar::InvertedIndexParams as FtsIndexBuilder;
 pub use lance_index::scalar::InvertedIndexParams;
+pub use lance_index::scalar::inverted::query::*;