feat: support nested blob function signatures

fix: preserve namespace drop errors (#4099 )
## Summary - preserve typed namespace errors returned by `drop_table` - return `TableNotFound` when dropping an absent namespace table - cover repeated drop behavior in the namespace database test ## Validation - `cargo test --quiet --features remote -p lancedb database::namespace::tests::test_namespace_drop_table --lib` - `cargo clippy --quiet --features remote -p lancedb --lib --tests -- -D warnings` ## Context Sophon SQL implements `DROP TABLE IF EXISTS` by matching `lancedb::Error::TableNotFound`. The namespace database previously wrapped this error as `Runtime`, causing cleanup to fail and mask an earlier statement error.
2026-09-01 11:08:55 +00:00 · 2026-09-01 16:54:15 +08:00 · 2026-08-31 14:02:40 -07:00 · 2026-08-31 10:40:11 -07:00 · 2026-08-31 22:34:34 +08:00 · 2026-08-31 22:17:11 +08:00
54 changed files with 2331 additions and 11595 deletions
@@ -1,5 +1,5 @@
 [tool.bumpversion]
-current_version = "0.38.0-beta.12"
+current_version = "0.38.0"
 parse = """(?x)
    (?P<major>0|[1-9]\\d*)\\.
    (?P<minor>0|[1-9]\\d*)\\.
@@ -44,3 +44,27 @@ updates:
      python-deps:
        patterns:
          - "*"
+
+  # The npm ecosystem covers pnpm lockfiles. There are two separate installs:
+  # the bindings themselves and the examples, which have their own lockfile.
+  # As with cargo and pip above, only bump the lockfile — the version ranges
+  # in package.json are our consumers' constraints, not ours.
+  - package-ecosystem: npm
+    directory: /nodejs
+    schedule:
+      interval: weekly
+    versioning-strategy: lockfile-only
+    groups:
+      nodejs-deps:
+        patterns:
+          - "*"
+
+  - package-ecosystem: npm
+    directory: /nodejs/examples
+    schedule:
+      interval: weekly
+    versioning-strategy: lockfile-only
+    groups:
+      nodejs-examples-deps:
+        patterns:
+          - "*"
@@ -29,12 +29,14 @@ jobs:
    steps:
      - uses: actions/setup-node@v6
        with:
-          node-version: "18"
+          node-version: "24"
+      - uses: pnpm/action-setup@v6
+        with:
+          version: 11.1.1
      # These rules are disabled because Github will always ensure there
      # is a blank line between the title and the body and Github will
      # word wrap the description field to ensure a reasonable max line
      # length.
-      - run: npm install @commitlint/config-conventional
      - run: >
          echo 'module.exports = {
            "rules": {
@@ -43,7 +45,11 @@ jobs:
              "body-leading-blank": [0, "always"]
            }
          }' > .commitlintrc.js
-      - run: npx commitlint --extends @commitlint/config-conventional --verbose <<< $COMMIT_MSG
+      - run: >
+          pnpm dlx
+          --package @commitlint/cli@21.2.2
+          --package @commitlint/config-conventional@21.2.2
+          commitlint --extends @commitlint/config-conventional --verbose <<< $COMMIT_MSG
        env:
          COMMIT_MSG: >
            ${{ github.event.pull_request.title }}
@@ -54,7 +60,7 @@ jobs:
        with:
          script: |
            const message = `**ACTION NEEDED**
-              
+
              Lance follows the [Conventional Commits specification](https://www.conventionalcommits.org/en/v1.0.0/) for release automation.

              The PR title and description are used as the merge commit message.\
@@ -56,7 +56,7 @@ jobs:
        uses: lycheeverse/lychee-action@e7477775783ea5526144ba13e8db5eec57747ce8 # v2.9.0
        with:
          # Restricted to http(s) on purpose. Much of docs/src is generated
-          # API reference (the js/ tree comes from `npm run docs` in nodejs)
+          # API reference (the js/ tree comes from `pnpm run docs` in nodejs)
          # and the hand-written pages use mkdocstrings cross-references and
          # nav-relative paths that only resolve in the site mkdocs builds,
          # not in this checkout, so relative links would be reported as
@@ -55,9 +55,7 @@ jobs:
      - name: Set up node
        uses: actions/setup-node@v6
        with:
-          node-version: 20
-          cache: 'npm'
-          cache-dependency-path: docs/package-lock.json
+          node-version: 24
      - name: Install node dependencies
        working-directory: nodejs
        run: |
@@ -47,9 +47,8 @@ jobs:
        version: 11.1.1
    - uses: actions/setup-node@v6
      with:
-        # pnpm 11 requires Node >= 22.13; use 24 since 22 hits EOL
-        # in October. The library itself still supports Node >= 18
-        # (see test matrix below).
+        # Build on a supported LTS; the matrix job below covers every
+        # Node version the library claims to support.
        node-version: 24
        cache: 'pnpm'
        cache-dependency-path: nodejs/pnpm-lock.yaml
@@ -84,7 +83,7 @@ jobs:
    timeout-minutes: 30
    strategy:
      matrix:
-        node-version: [ "18", "20" ]
+        node-version: [ "22", "24", "26" ]
    runs-on: "ubuntu-22.04"
    defaults:
      run:
@@ -101,9 +100,9 @@ jobs:
    - uses: actions/setup-node@v6
      name: Setup Node.js 24 for build
      with:
-        # pnpm 11 requires Node >= 22.13; use 24 since 22 hits EOL
-        # in October. Build/install runs on Node 24; tests run on the
-        # matrix version below using direct jest invocation.
+        # Build and install once on a fixed version so the generated docs
+        # are identical across matrix legs; the tests below then run on each
+        # supported Node version.
        node-version: 24
        cache: 'pnpm'
        cache-dependency-path: nodejs/pnpm-lock.yaml
@@ -152,9 +151,9 @@ jobs:
        S3_TEST: "1"
        # Newer @smithy/core uses dynamic ESM imports.
        NODE_OPTIONS: "--experimental-vm-modules"
-      # Invoke jest directly because pnpm 11 itself requires Node 22+
-      # while the matrix tests on older Node versions.
-      run: npx jest --verbose
+      # Invoke the installed jest binary directly; the pnpm shim is set up
+      # against the build-phase Node, not the version selected above.
+      run: node_modules/.bin/jest --verbose
    - name: Test examples
      working-directory: ./
      env:
@@ -164,7 +163,7 @@ jobs:
      run: |
        python ci/mock_openai.py &
        cd nodejs/examples
-        npx jest --testEnvironment jest-environment-node-single-context --verbose
+        node_modules/.bin/jest --testEnvironment jest-environment-node-single-context --verbose
  macos:
    timeout-minutes: 30
    # macos-15 ships a newer linker; the older macos-14 linker fails to insert
@@ -185,8 +184,7 @@ jobs:
        version: 11.1.1
    - uses: actions/setup-node@v6
      with:
-        # pnpm 11 requires Node >= 22.13; use 24 since 22 hits EOL
-        # in October.
+        # pnpm 11 requires Node >= 22.13.
        node-version: 24
        cache: 'pnpm'
        cache-dependency-path: nodejs/pnpm-lock.yaml
@@ -168,8 +168,7 @@ jobs:
      - name: Setup node
        uses: actions/setup-node@v6
        with:
-          # pnpm 11 requires Node >= 22.13; use 24 since 22 hits EOL
-          # in October.
+          # pnpm 11 requires Node >= 22.13.
          node-version: 24
          cache: pnpm
          cache-dependency-path: nodejs/pnpm-lock.yaml
@@ -251,7 +250,7 @@ jobs:
          run: |
            set -e
            ${{ matrix.settings.pre_build }}
-            npx napi build --platform --release \
+            node_modules/.bin/napi build --platform --release \
              --features ${{ matrix.settings.features }} \
              --target ${{ matrix.settings.target }} \
              --dts ../lancedb/native.d.ts \
@@ -271,7 +270,7 @@ jobs:
      - name: Build
        run: |
          ${{ matrix.settings.pre_build }}
-          npx napi build --platform --release \
+          node_modules/.bin/napi build --platform --release \
              --features ${{ matrix.settings.features }} \
              --target ${{ matrix.settings.target }} \
              --dts ../lancedb/native.d.ts \
@@ -339,7 +338,7 @@ jobs:
          - target: aarch64-unknown-linux-gnu
            host: ubuntu-2404-8x-arm64
        node:
-          - '20'
+          - '22'
    runs-on: ${{ matrix.settings.host }}
    defaults:
      run:
@@ -385,9 +384,9 @@ jobs:
      - name: Move built files
        run: cp dist/native.d.ts dist/native.js dist/*.node lancedb/
      - name: Test bindings
-        # Invoke jest directly because pnpm 11 itself requires Node 22+
-        # while the matrix tests on older Node versions.
-        run: npx jest --verbose
+        # Invoke the installed jest binary directly; the pnpm shim is set up
+        # against the install-phase Node, not the version selected above.
+        run: node_modules/.bin/jest --verbose
  publish:
    name: Publish
    runs-on: ubuntu-latest
@@ -1,22 +0,0 @@
-name: Update package-lock.json
-
-on:
-  workflow_dispatch:
-
-permissions:
-  contents: read
-
-jobs:
-  publish:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v6
-        with:
-          ref: main
-          persist-credentials: false
-          fetch-depth: 0
-          lfs: true
-      - uses: ./.github/workflows/update_package_lock
-        with:
-          github_token: ${{ secrets.LANCEDB_RELEASE_TOKEN }}
@@ -1,22 +0,0 @@
-name: Update NodeJs package-lock.json
-
-on:
-  workflow_dispatch:
-
-permissions:
-  contents: read
-
-jobs:
-  publish:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v6
-        with:
-          ref: main
-          persist-credentials: false
-          fetch-depth: 0
-          lfs: true
-      - uses: ./.github/workflows/update_package_lock_nodejs
-        with:
-          github_token: ${{ secrets.LANCEDB_RELEASE_TOKEN }}
@@ -20,7 +20,10 @@ repos:
    hooks:
      - id: local-biome-check
        name: biome check
-        entry: npx @biomejs/biome@1.8.3 check --config-path nodejs/biome.json nodejs/
+        # Use the biome from nodejs/package.json rather than a separately
+        # pinned one: the two drifted apart and disagreed on formatting, so
+        # this hook rejected code that `pnpm lint` accepted.
+        entry: nodejs/node_modules/.bin/biome check --config-path nodejs/biome.json nodejs/
        language: system
        types: [text]
        files: "nodejs/.*"
@@ -38,7 +38,7 @@ Before committing changes, run formatting for every language you touched. At min
 * Rust changes: run `cargo fmt --all`.
 * Python changes: run `ruff format .` and `ruff check .` from the repository root,
  and run targeted tests through `cd python && uv run ...`.
-* TypeScript changes: run the relevant `npm`/`pnpm` lint, format, build, and docs commands in `nodejs`.
+* TypeScript changes: run the relevant `pnpm` lint, format, build, and docs commands in `nodejs`.

 Before creating a PR, the exact value passed to `gh pr create --title` must follow
 Conventional Commits, such as `fix: support nested field paths in native index creation`
@@ -101,12 +101,12 @@ Python bindings changes:
 TypeScript bindings changes:

 1. Add napi-rs method binding on `Table` in `nodejs/src/table.rs`.
-2. Run `npm run build` to generate TypeScript definitions.
+2. Run `pnpm build` to generate TypeScript definitions.
 3. Add typescript method on abstract class `Table` in `nodejs/src/table.ts`.
 4. Add concrete method on `LocalTable` class in `nodejs/src/native_table.ts`.
    * Note: despite the name, this class is also used for remote tables.
 5. Add test in `nodejs/__test__/table.test.ts`.
-6. Run `npm run docs` to generate TypeScript documentation.
+6. Run `pnpm run docs` to generate TypeScript documentation.

 ## Python API reference

@@ -535,9 +535,9 @@ dependencies = [

 [[package]]
 name = "async-trait"
-version = "0.1.91"
+version = "0.1.92"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ae36dc4177970ef04fde5178d3e2429882def40e57a451f919c098f72baa6cec"
+checksum = "82f6aeea286b8eb4dd3431a1be1b59d290ace00f5bfd8e2a159bc2a05e2c1667"
 dependencies = [
 "proc-macro2",
 "quote",
@@ -1443,9 +1443,9 @@ checksum = "175812e0be2bccb6abe50bb8d566126198344f707e304f45c648fd8f2cc0365e"

 [[package]]
 name = "bytemuck"
-version = "1.25.0"
+version = "1.25.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c8efb64bd706a16a1bdde310ae86b351e4d21550d98d056f22f8a7f7a2183fec"
+checksum = "95832e849adfb21180ccb6826a99da14e5d266ae5c2e668e1602cf234f153797"
 dependencies = [
 "bytemuck_derive",
 ]
@@ -3455,8 +3455,8 @@ checksum = "42703706b716c37f96a77aea830392ad231f44c9e9a67872fa5548707e11b11c"

 [[package]]
 name = "fsst"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow-array",
 "rand 0.9.5",
@@ -4815,8 +4815,8 @@ checksum = "e037a2e1d8d5fdbd49b16a4ea09d5d6401c1f29eca5ff29d03d3824dba16256a"

 [[package]]
 name = "lance"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arc-swap",
 "arrow",
@@ -4888,8 +4888,8 @@ dependencies = [

 [[package]]
 name = "lance-arrow"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow-array",
 "arrow-buffer",
@@ -4911,7 +4911,7 @@ dependencies = [
 [[package]]
 name = "lance-arrow-scalar"
 version = "58.0.0"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow-array",
 "arrow-buffer",
@@ -4925,7 +4925,7 @@ dependencies = [
 [[package]]
 name = "lance-arrow-stats"
 version = "58.0.0"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow-array",
 "arrow-schema",
@@ -4934,8 +4934,8 @@ dependencies = [

 [[package]]
 name = "lance-bitpacking"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrayref",
 "crunchy",
@@ -4945,8 +4945,8 @@ dependencies = [

 [[package]]
 name = "lance-core"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow-array",
 "arrow-buffer",
@@ -4983,8 +4983,8 @@ dependencies = [

 [[package]]
 name = "lance-datafusion"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow",
 "arrow-array",
@@ -5013,8 +5013,8 @@ dependencies = [

 [[package]]
 name = "lance-datagen"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow",
 "arrow-array",
@@ -5031,8 +5031,8 @@ dependencies = [

 [[package]]
 name = "lance-derive"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "proc-macro2",
 "quote",
@@ -5041,8 +5041,8 @@ dependencies = [

 [[package]]
 name = "lance-encoding"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow-arith",
 "arrow-array",
@@ -5075,8 +5075,8 @@ dependencies = [

 [[package]]
 name = "lance-file"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow-arith",
 "arrow-array",
@@ -5107,8 +5107,8 @@ dependencies = [

 [[package]]
 name = "lance-index"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arc-swap",
 "arrow",
@@ -5172,8 +5172,8 @@ dependencies = [

 [[package]]
 name = "lance-index-core"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow-array",
 "arrow-schema",
@@ -5195,8 +5195,8 @@ dependencies = [

 [[package]]
 name = "lance-io"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow",
 "arrow-array",
@@ -5236,8 +5236,8 @@ dependencies = [

 [[package]]
 name = "lance-linalg"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow-array",
 "arrow-schema",
@@ -5251,8 +5251,8 @@ dependencies = [

 [[package]]
 name = "lance-namespace"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow",
 "async-trait",
@@ -5264,8 +5264,8 @@ dependencies = [

 [[package]]
 name = "lance-namespace-impls"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow",
 "arrow-ipc",
@@ -5318,8 +5318,8 @@ dependencies = [

 [[package]]
 name = "lance-select"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow-array",
 "arrow-buffer",
@@ -5333,8 +5333,8 @@ dependencies = [

 [[package]]
 name = "lance-table"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow",
 "arrow-array",
@@ -5374,8 +5374,8 @@ dependencies = [

 [[package]]
 name = "lance-testing"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "arrow-array",
 "arrow-schema",
@@ -5388,8 +5388,8 @@ dependencies = [

 [[package]]
 name = "lance-tokenizer"
-version = "12.0.0-beta.2"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
+version = "12.0.0-beta.5"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
 dependencies = [
 "frostem",
 "icu_segmenter",
@@ -5402,7 +5402,7 @@ dependencies = [

 [[package]]
 name = "lancedb"
-version = "0.38.0-beta.12"
+version = "0.38.0"
 dependencies = [
 "ahash",
 "anyhow",
@@ -5490,7 +5490,7 @@ dependencies = [

 [[package]]
 name = "lancedb-nodejs"
-version = "0.38.0-beta.12"
+version = "0.38.0"
 dependencies = [
 "arrow-array",
 "arrow-buffer",
@@ -5515,7 +5515,7 @@ dependencies = [

 [[package]]
 name = "lancedb-python"
-version = "0.38.0-beta.12"
+version = "0.38.0"
 dependencies = [
 "arrow",
 "async-trait",
@@ -5748,9 +5748,9 @@ dependencies = [

 [[package]]
 name = "log"
-version = "0.4.33"
+version = "0.4.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0ceec5bc11778974d1bcb055b18002eba7f4b3518b6a0081b3af5f21666da9ad"
+checksum = "f9f8bd3e56ce4dfc153cf470fffbfa98c7620958b312ca5c3a4b8d5181fd13c6"

 [[package]]
 name = "loom"
@@ -6001,9 +6001,9 @@ dependencies = [

 [[package]]
 name = "moka"
-version = "0.12.15"
+version = "0.12.16"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "957228ad12042ee839f93c8f257b62b4c0ab5eaae1d4fa60de53b27c9d7c5046"
+checksum = "4293f18e7567a1caf3c584855554377025c65e0aa445344d04171f5ad63d19b9"
 dependencies = [
 "async-lock",
 "crossbeam-channel",
@@ -6097,14 +6097,15 @@ dependencies = [

 [[package]]
 name = "napi"
-version = "3.11.0"
+version = "3.12.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "de33522036981030a75c231829566bc63414e08101a6f5ff4ac6cef19c8e0941"
+checksum = "58c5f4d5375213fdb7be2655e152386e82f026f9a5ba36a75556e11359aafe09"
 dependencies = [
 "bitflags 2.11.1",
 "chrono",
 "ctor 1.0.12",
 "futures",
+ "libc",
 "napi-build",
 "napi-sys",
 "nohash-hasher",
@@ -6116,15 +6117,15 @@ dependencies = [

 [[package]]
 name = "napi-build"
-version = "2.4.0"
+version = "2.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5282704fbe8d49b0cf8b08e3f33233416a528658f205c7e5ace63b582de0b11c"
+checksum = "60fdf9b392c50e7c4170fa633bd909490ed7835cea4c046776d1a4dd8d2ae0ab"

 [[package]]
 name = "napi-derive"
-version = "3.6.1"
+version = "3.6.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4d5c9c02556ea6dc99dffd36c1ce60141411657438501a125b675776d011ce92"
+checksum = "0fa55ea69990c90b888e9e77044410e304ce7f35de599dc6d0b5c1923d2e59af"
 dependencies = [
 "convert_case",
 "ctor 1.0.12",
@@ -6136,9 +6137,9 @@ dependencies = [

 [[package]]
 name = "napi-derive-backend"
-version = "6.1.1"
+version = "6.1.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d60b5d773ad46c698c8cc2cd9fde0b283d39cbb7f71c04bee633c7bdba4423bd"
+checksum = "df4056ac7c18e4438ccf0edaed4340ca0d269278c8ec19284f7b23cb039fd0ae"
 dependencies = [
 "convert_case",
 "proc-macro2",
@@ -8601,9 +8602,9 @@ dependencies = [

 [[package]]
 name = "roaring"
-version = "0.11.4"
+version = "0.11.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1dedc5658c6ecb3bdb5ef5f3295bb9253f42dcf3fd1402c03f6b1f7659c3c4a9"
+checksum = "18bd8a37d17a58532776dcdf6041ce64929adca78e8489d5cacbafe99229d3e1"
 dependencies = [
 "bytemuck",
 "byteorder",
@@ -9063,9 +9064,9 @@ dependencies = [

 [[package]]
 name = "serde_with"
-version = "3.21.0"
+version = "3.22.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "76a5c54c7310e7b8b9577c286d7e399ddd876c3e12b3ed917a8aabc4b96e9e8c"
+checksum = "ee78f1fbe43ac4a0e47aadb3dbd357b69eb0d3793e948624cd03dd2750ab1c0a"
 dependencies = [
 "base64 0.22.1",
 "bs58",
@@ -9073,6 +9074,7 @@ dependencies = [
 "hex",
 "indexmap 1.9.3",
 "indexmap 2.14.0",
+ "jiff",
 "schemars 0.9.0",
 "schemars 1.2.1",
 "serde_core",
@@ -9083,9 +9085,9 @@ dependencies = [

 [[package]]
 name = "serde_with_macros"
-version = "3.21.0"
+version = "3.22.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "84d57bc0c8b9a17920c178daa6bb924850d54a9c97ab45194bb8c17ad66bb660"
+checksum = "8705578779c2b6bd90d84d66eb2e206b708b1a4d7b9f17641b293545bf1c7e46"
 dependencies = [
 "darling 0.23.0",
 "proc-macro2",
@@ -10452,9 +10454,9 @@ checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821"

 [[package]]
 name = "uuid"
-version = "1.24.0"
+version = "1.26.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bf3923a6f5c4c6382e0b653c4117f48d631ea17f38ed86e2a828e6f7412f5239"
+checksum = "b5772d71c9be8a8a6ac2117d949c5b224c1b72241bb611d9a3012edcf8af7812"
 dependencies = [
 "getrandom 0.4.2",
 "js-sys",
@@ -13,20 +13,20 @@ categories = ["database-implementations"]
 rust-version = "1.91.0"

 [workspace.dependencies]
-lance = { "version" = "=12.0.0-beta.2", default-features = false, "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
-lance-core = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
-lance-datagen = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
-lance-file = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
-lance-io = { "version" = "=12.0.0-beta.2", default-features = false, "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
-lance-index = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
-lance-linalg = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
-lance-namespace = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
-lance-namespace-impls = { "version" = "=12.0.0-beta.2", default-features = false, "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
-lance-table = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
-lance-testing = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
-lance-datafusion = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
-lance-encoding = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
-lance-arrow = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance = { "version" = "=12.0.0-beta.5", default-features = false, "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
+lance-core = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
+lance-datagen = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
+lance-file = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
+lance-io = { "version" = "=12.0.0-beta.5", default-features = false, "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
+lance-index = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
+lance-linalg = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
+lance-namespace = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
+lance-namespace-impls = { "version" = "=12.0.0-beta.5", default-features = false, "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
+lance-table = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
+lance-testing = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
+lance-datafusion = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
+lance-encoding = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
+lance-arrow = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
 lancedb = { path = "rust/lancedb", default-features = false }
 ahash = "0.8"
 # Note that this one does not include pyarrow
@@ -5,5 +5,5 @@ licenses:
 	cd python && cargo about generate ../about.hbs -o RUST_THIRD_PARTY_LICENSES.html -c ../about.toml
 	cd python && uv sync --all-extras && uv tool run pip-licenses --python .venv/bin/python --format=markdown --with-urls --output-file=PYTHON_THIRD_PARTY_LICENSES.md
 	cd nodejs && cargo about generate ../about.hbs -o RUST_THIRD_PARTY_LICENSES.html -c ../about.toml
-	cd nodejs && npx license-checker --markdown --out NODEJS_THIRD_PARTY_LICENSES.md
+	cd nodejs && pnpm dlx license-checker@25 --markdown --out NODEJS_THIRD_PARTY_LICENSES.md
 	cd java && ./mvnw license:aggregate-add-third-party -q
@@ -12,16 +12,12 @@ done
 # This updates the lockfile without building
 cargo metadata --quiet > /dev/null

-pushd nodejs || exit 1
-npm install --package-lock-only --silent
-popd
-
 if git diff --quiet --exit-code; then
  echo "No lockfile changes to commit; skipping amend."
 elif $AMEND; then
-  git add Cargo.lock nodejs/package-lock.json
+  git add Cargo.lock
  git commit --amend --no-edit
 else
-  git add Cargo.lock nodejs/package-lock.json
+  git add Cargo.lock
  git commit -m "Update lockfiles"
 fi
@@ -47,22 +47,24 @@ pytest -vv python/tests/docs

 ### Checking typescript examples

-The `@lancedb/lancedb` package must be built before running the tests:
+The examples depend on `@lancedb/lancedb` at `file:../dist`, so the package must be
+built before running the tests. This uses pnpm; see the
+[Typescript contributing guide](../nodejs/CONTRIBUTING.md) for the toolchain setup.

 ```shell
 pushd nodejs
-npm ci
-npm run build
+pnpm install
+pnpm build
 popd
 ```

-Then you can run the examples by going to the `nodejs/examples` directory and
-running the tests like a normal npm package:
+Then you can run the examples by going to the `nodejs/examples` directory, which is a
+separate pnpm package with its own lockfile:

 ```shell
 pushd nodejs/examples
-npm ci
-npm test
+pnpm install
+pnpm test
 popd
 ```

@@ -84,6 +86,7 @@ The new files should be checked into the repository.

 ```shell
 pushd nodejs
-npm run docs
+# `pnpm docs` would invoke pnpm's built-in `docs` command, not the script.
+pnpm run docs
 popd
 ```
@@ -1,135 +0,0 @@
-{
-  "name": "lancedb-docs-test",
-  "version": "1.0.0",
-  "lockfileVersion": 3,
-  "requires": true,
-  "packages": {
-    "": {
-      "name": "lancedb-docs-test",
-      "version": "1.0.0",
-      "license": "Apache 2",
-      "dependencies": {
-        "apache-arrow": "file:../node/node_modules/apache-arrow",
-        "vectordb": "file:../node"
-      },
-      "devDependencies": {
-        "@types/node": "^20.11.8",
-        "typescript": "^5.3.3"
-      }
-    },
-    "../node": {
-      "name": "vectordb",
-      "version": "0.21.2-beta.0",
-      "cpu": [
-        "x64",
-        "arm64"
-      ],
-      "license": "Apache-2.0",
-      "os": [
-        "darwin",
-        "linux",
-        "win32"
-      ],
-      "dependencies": {
-        "@neon-rs/load": "^0.0.74",
-        "axios": "^1.4.0"
-      },
-      "devDependencies": {
-        "@neon-rs/cli": "^0.0.160",
-        "@types/chai": "^4.3.4",
-        "@types/chai-as-promised": "^7.1.5",
-        "@types/mocha": "^10.0.1",
-        "@types/node": "^18.16.2",
-        "@types/sinon": "^10.0.15",
-        "@types/temp": "^0.9.1",
-        "@types/uuid": "^9.0.3",
-        "@typescript-eslint/eslint-plugin": "^5.59.1",
-        "apache-arrow-old": "npm:apache-arrow@13.0.0",
-        "cargo-cp-artifact": "^0.1",
-        "chai": "^4.3.7",
-        "chai-as-promised": "^7.1.1",
-        "eslint": "^8.39.0",
-        "eslint-config-standard-with-typescript": "^34.0.1",
-        "eslint-plugin-import": "^2.26.0",
-        "eslint-plugin-n": "^15.7.0",
-        "eslint-plugin-promise": "^6.1.1",
-        "mocha": "^10.2.0",
-        "openai": "^4.24.1",
-        "sinon": "^15.1.0",
-        "temp": "^0.9.4",
-        "ts-node": "^10.9.1",
-        "ts-node-dev": "^2.0.0",
-        "typedoc": "^0.24.7",
-        "typedoc-plugin-markdown": "^3.15.3",
-        "typescript": "^5.1.0",
-        "uuid": "^9.0.0"
-      },
-      "optionalDependencies": {
-        "@lancedb/vectordb-darwin-arm64": "0.21.2-beta.0",
-        "@lancedb/vectordb-darwin-x64": "0.21.2-beta.0",
-        "@lancedb/vectordb-linux-arm64-gnu": "0.21.2-beta.0",
-        "@lancedb/vectordb-linux-x64-gnu": "0.21.2-beta.0",
-        "@lancedb/vectordb-win32-x64-msvc": "0.21.2-beta.0"
-      },
-      "peerDependencies": {
-        "@apache-arrow/ts": "^14.0.2",
-        "apache-arrow": "^14.0.2"
-      }
-    },
-    "../node/node_modules/apache-arrow": {
-      "version": "14.0.2",
-      "license": "Apache-2.0",
-      "dependencies": {
-        "@types/command-line-args": "5.2.0",
-        "@types/command-line-usage": "5.0.2",
-        "@types/node": "20.3.0",
-        "@types/pad-left": "2.1.1",
-        "command-line-args": "5.2.1",
-        "command-line-usage": "7.0.1",
-        "flatbuffers": "23.5.26",
-        "json-bignum": "^0.0.3",
-        "pad-left": "^2.1.0",
-        "tslib": "^2.5.3"
-      },
-      "bin": {
-        "arrow2csv": "bin/arrow2csv.js"
-      }
-    },
-    "node_modules/@types/node": {
-      "version": "20.11.8",
-      "resolved": "https://registry.npmjs.org/@types/node/-/node-20.11.8.tgz",
-      "integrity": "sha512-i7omyekpPTNdv4Jb/Rgqg0RU8YqLcNsI12quKSDkRXNfx7Wxdm6HhK1awT3xTgEkgxPn3bvnSpiEAc7a7Lpyow==",
-      "dev": true,
-      "dependencies": {
-        "undici-types": "~5.26.4"
-      }
-    },
-    "node_modules/apache-arrow": {
-      "resolved": "../node/node_modules/apache-arrow",
-      "link": true
-    },
-    "node_modules/typescript": {
-      "version": "5.3.3",
-      "resolved": "https://registry.npmjs.org/typescript/-/typescript-5.3.3.tgz",
-      "integrity": "sha512-pXWcraxM0uxAS+tN0AG/BF2TyqmHO014Z070UsJ+pFvYuRSq8KH8DmWpnbXe0pEPDHXZV3FcAbJkijJ5oNEnWw==",
-      "dev": true,
-      "bin": {
-        "tsc": "bin/tsc",
-        "tsserver": "bin/tsserver"
-      },
-      "engines": {
-        "node": ">=14.17"
-      }
-    },
-    "node_modules/undici-types": {
-      "version": "5.26.5",
-      "resolved": "https://registry.npmjs.org/undici-types/-/undici-types-5.26.5.tgz",
-      "integrity": "sha512-JlCMO+ehdEIKqlFxk6IfVoAUVmgz7cU7zD/h9XZ0qzeosSHmUJVOzSQvvYSYWXkFXC+IfLKSIffhv0sVZup6pA==",
-      "dev": true
-    },
-    "node_modules/vectordb": {
-      "resolved": "../node",
-      "link": true
-    }
-  }
-}
@@ -1,20 +0,0 @@
-{
-  "name": "lancedb-docs-test",
-  "version": "1.0.0",
-  "description": "auto-generated tests from doc",
-  "author": "dev@lancedb.com",
-  "license": "Apache 2",
-  "dependencies": {
-    "apache-arrow": "file:../node/node_modules/apache-arrow",
-    "vectordb": "file:../node"
-  },
-  "scripts": {
-    "build": "tsc -b && cd ../node && npm run build-release",
-    "example": "npm run build && node",
-    "test": "npm run build && ls dist/*.js | xargs -n 1 node"
-  },
-  "devDependencies": {
-    "@types/node": "^20.11.8",
-    "typescript": "^5.3.3"
-  }
-}
@@ -14,7 +14,7 @@ Add the following dependency to your `pom.xml`:
 <dependency>
    <groupId>com.lancedb</groupId>
    <artifactId>lancedb-core</artifactId>
-    <version>0.38.0-beta.12</version>
+    <version>0.38.0</version>
 </dependency>
 ```

@@ -1,17 +0,0 @@
-{
-  "include": [
-    "src/*.ts",
-  ],
-  "compilerOptions": {
-    "target": "es2022",
-    "module": "nodenext",
-    "declaration": true,
-    "outDir": "./dist",
-    "strict": true,
-    "allowJs": true,
-    "resolveJsonModule": true,
-  },
-  "exclude": [
-    "./dist/*",
-  ]
-}
@@ -8,7 +8,7 @@
    <parent>
      <groupId>com.lancedb</groupId>
      <artifactId>lancedb-parent</artifactId>
-      <version>0.38.0-beta.12</version>
+      <version>0.38.0-final.0</version>
      <relativePath>../pom.xml</relativePath>
    </parent>

@@ -6,7 +6,7 @@

    <groupId>com.lancedb</groupId>
    <artifactId>lancedb-parent</artifactId>
-    <version>0.38.0-beta.12</version>
+    <version>0.38.0-final.0</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>12.0.0-beta.2</lance-core.version>
+        <lance-core.version>12.0.0-beta.5</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>
@@ -1,7 +1,7 @@
 [package]
 name = "lancedb-nodejs"
 edition.workspace = true
-version = "0.38.0-beta.12"
+version = "0.38.0"
 publish = false
 license.workspace = true
 description.workspace = true
@@ -5,8 +5,8 @@ import packageJson = require("../package.json");

 describe("package metadata", () => {
  it("requires Node.js type declarations compatible with the runtime", () => {
-    expect(packageJson.engines.node).toBe(">= 18");
-    expect(packageJson.peerDependencies["@types/node"]).toBe(">=18");
+    expect(packageJson.engines.node).toBe(">= 22");
+    expect(packageJson.peerDependencies["@types/node"]).toBe(">=22");
    expect(packageJson.peerDependenciesMeta["@types/node"]).toEqual({
      optional: true,
    });
@@ -3,6 +3,7 @@

 import * as http from "http";
 import { RequestListener } from "http";
+import packageJson = require("../package.json");
 import {
  ClientConfig,
  Connection,
@@ -70,7 +71,13 @@ async function withMockDatabase(
  try {
    await callback(db);
  } finally {
-    server.close();
+    // `close()` alone leaves the port bound until keep-alive sockets drain, so
+    // a single failing test would cascade into EADDRINUSE for every test after
+    // it. Destroy the connections and wait for the port to actually be free.
+    await new Promise<void>((resolve) => {
+      server.closeAllConnections();
+      server.close(() => resolve());
+    });
  }
 }

@@ -131,7 +138,7 @@ describe("remote connection", () => {
      (req, res) => {
        expect(req.headers["x-api-key"]).toEqual("fake");
        expect(req.headers["user-agent"]).toEqual(
-          `LanceDB-Node-Client/${process.env.npm_package_version}`,
+          `LanceDB-Node-Client/${packageJson.version}`,
        );

        const body = JSON.stringify({ tables: [] });
@@ -8,7 +8,8 @@
    "//1": "--experimental-vm-modules is needed to run jest with sentence-transformers",
    "//2": "--testEnvironment is needed to run jest with sentence-transformers",
    "//3": "See: https://github.com/huggingface/transformers.js/issues/57",
-    "test": "node --experimental-vm-modules node_modules/.bin/jest --testEnvironment jest-environment-node-single-context --verbose",
+    "//4": "jest is invoked by its JS entry, not node_modules/.bin/jest: under pnpm that path is a shell shim, which `node` cannot execute",
+    "test": "node --experimental-vm-modules node_modules/jest/bin/jest.js --testEnvironment jest-environment-node-single-context --verbose",
    "lint": "biome check *.ts && biome format *.ts",
    "lint-ci": "biome ci .",
    "lint-fix": "biome check --write *.ts && pnpm format",
@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-darwin-arm64",
-	"version": "0.38.0-beta.12",
+	"version": "0.38.0",
 	"os": ["darwin"],
 	"cpu": ["arm64"],
 	"main": "lancedb.darwin-arm64.node",
@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-arm64-gnu",
-	"version": "0.38.0-beta.12",
+	"version": "0.38.0",
 	"os": ["linux"],
 	"cpu": ["arm64"],
 	"main": "lancedb.linux-arm64-gnu.node",
@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-arm64-musl",
-	"version": "0.38.0-beta.12",
+	"version": "0.38.0",
 	"os": ["linux"],
 	"cpu": ["arm64"],
 	"main": "lancedb.linux-arm64-musl.node",
@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-x64-gnu",
-	"version": "0.38.0-beta.12",
+	"version": "0.38.0",
 	"os": ["linux"],
 	"cpu": ["x64"],
 	"main": "lancedb.linux-x64-gnu.node",
@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-x64-musl",
-	"version": "0.38.0-beta.12",
+	"version": "0.38.0",
 	"os": ["linux"],
 	"cpu": ["x64"],
 	"main": "lancedb.linux-x64-musl.node",
@@ -1,6 +1,6 @@
 {
  "name": "@lancedb/lancedb-win32-arm64-msvc",
-  "version": "0.38.0-beta.12",
+  "version": "0.38.0",
  "os": [
    "win32"
  ],
@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-win32-x64-msvc",
-	"version": "0.38.0-beta.12",
+	"version": "0.38.0",
 	"os": ["win32"],
 	"cpu": ["x64"],
 	"main": "lancedb.win32-x64-msvc.node",
@@ -11,7 +11,7 @@
    "ann"
  ],
  "private": false,
-  "version": "0.38.0-beta.12",
+  "version": "0.38.0",
  "main": "dist/index.js",
  "exports": {
    ".": "./dist/index.js",
@@ -67,7 +67,7 @@
    "timeout": "3m"
  },
  "engines": {
-    "node": ">= 18"
+    "node": ">= 22"
  },
  "packageManager": "pnpm@11.1.1",
  "cpu": ["x64", "arm64"],
@@ -101,7 +101,7 @@
    "openai": "4.29.2"
  },
  "peerDependencies": {
-    "@types/node": ">=18",
+    "@types/node": ">=22",
    "apache-arrow": ">=15.0.0 <=18.1.0"
  },
  "peerDependenciesMeta": {
@@ -664,7 +664,11 @@ impl JsFullTextQuery {
 }

 fn parse_fts_query(query: Object) -> napi::Result<FullTextSearchQuery> {
-    if let Ok(Some(query)) = query.get::<&JsFullTextQuery>("query") {
+    // `&JsFullTextQuery` recovers a native class reference through napi's borrow-tracked
+    // path, which is only usable from generated `#[napi]` argument conversion. This is a
+    // manual lookup on a nested `Object` property instead, so use `ClassInstance`, which
+    // unwraps the class without requiring a borrow scope.
+    if let Ok(Some(query)) = query.get::<ClassInstance<JsFullTextQuery>>("query") {
        Ok(FullTextSearchQuery::new_query(query.inner.clone()))
    } else if let Ok(Some(query_text)) = query.get::<String>("query") {
        let mut query_text = query_text;
@@ -1,6 +1,6 @@
 [package]
 name = "lancedb-python"
-version = "0.38.0-beta.12"
+version = "0.38.0"
 publish = false
 edition.workspace = true
 description = "Python bindings for LanceDB"
@@ -150,6 +150,7 @@ class Connection(object):
    def job(self, job_id: str) -> Job: ...
    async def create_function_async(self, request_json: str) -> Job: ...
    async def get_function(self, name: str, version: str) -> str: ...
+    async def drop_function(self, name: str, version: str) -> bool: ...
    async def list_jobs(self) -> List[JobInfo]: ...
    async def get_job(self, job_id: str) -> Optional[JobDescription]: ...
    async def cancel_job(self, job_id: str) -> bool: ...
@@ -712,6 +712,16 @@ class DBConnection(EnforceOverrides):
            "Function catalog operations are not supported for this connection type"
        )

+    def drop_function(self, name: str, *, version: str) -> bool:
+        """Drop one exact immutable Function version from the remote catalog.
+
+        Returns True when the version changed to Dropped and False for an
+        idempotent replay. Local connections raise NotImplementedError.
+        """
+        raise NotImplementedError(
+            "Function catalog operations are not supported for this connection type"
+        )
+
    def job(self, job_id: str) -> Job:
        """A [Job][lancedb.job.Job] handle for a server-side job by id.

@@ -1413,6 +1423,10 @@ class LanceDBConnection(DBConnection):
    def get_function(self, name: str, *, version: str) -> FunctionVersion:
        return LOOP.run(self._conn.get_function(name, version=version))

+    @override
+    def drop_function(self, name: str, *, version: str) -> bool:
+        return LOOP.run(self._conn.drop_function(name, version=version))
+
    @override
    def list_jobs(self) -> List[JobInfo]:
        """List server-side jobs across the database's tables."""
@@ -2243,6 +2257,10 @@ class AsyncConnection(object):
        """Open one exact immutable Function version from the remote catalog."""
        return FunctionVersion.from_json(await self._inner.get_function(name, version))

+    async def drop_function(self, name: str, *, version: str) -> bool:
+        """Drop one exact immutable Function version from the remote catalog."""
+        return await self._inner.drop_function(name, version)
+
    async def list_jobs(self) -> List[JobInfo]:
        """List server-side jobs across the database's tables."""
        return await self._inner.list_jobs()
@@ -49,11 +49,25 @@ from pydantic import (
    model_validator,
 )

+from .schema import is_blob_v2_field as _is_blob_v2_field
+
 _Int32 = conint(strict=True, ge=-(2**31), le=2**31 - 1)
 _UInt32 = conint(strict=True, ge=0, le=2**32 - 1)
 _UInt64 = conint(strict=True, ge=0, le=2**64 - 1)


+def _validate_gpu_wire_marker(value: Any) -> bool:
+    if value is not True:
+        raise ValueError("runtime.gpu must be true")
+    return True
+
+
+def _normalize_gpu_marker(value: bool) -> Optional[bool]:
+    if not isinstance(value, bool):
+        raise ValueError("gpu must be a boolean")
+    return True if value else None
+
+
 class _FrozenDict(dict):
    def _immutable(self, *args, **kwargs):
        raise TypeError("remote canonical values are immutable")
@@ -239,6 +253,23 @@ class PythonRuntimeSpec(_RemoteValue):
    python_version: Optional[str] = None
    environment: Optional[PythonEnvironmentSpec] = None
    env: Optional[Mapping[str, str]] = None
+    gpu: Optional[bool] = None
+
+    @model_validator(mode="before")
+    @classmethod
+    def _discard_unknown_runtime_payload(cls, value):
+        if isinstance(value, Mapping):
+            kind = value.get("kind")
+            if isinstance(kind, str) and kind not in {"python", "python_v2"}:
+                return {"kind": kind}
+        return value
+
+    @field_validator("gpu", mode="before")
+    @classmethod
+    def _validate_gpu_marker(cls, value):
+        if value is None:
+            return None
+        return _validate_gpu_wire_marker(value)

    @model_validator(mode="after")
    def _validate_runtime_kind(self):
@@ -247,18 +278,28 @@ class PythonRuntimeSpec(_RemoteValue):
                raise ValueError("python runtime requires python_version")
            if self.environment is None:
                raise ValueError("python runtime requires environment")
+            if self.gpu is not None:
+                raise ValueError("python runtime with gpu requires kind='python_v2'")
+        elif self.kind == "python_v2":
+            if self.python_version is None:
+                raise ValueError("python_v2 runtime requires python_version")
+            if self.environment is None:
+                raise ValueError("python_v2 runtime requires environment")
+            if self.gpu is None:
+                raise ValueError("python_v2 runtime requires gpu")
        else:
            object.__setattr__(self, "python_version", None)
            object.__setattr__(self, "environment", None)
            object.__setattr__(self, "env", None)
+            object.__setattr__(self, "gpu", None)
        return self


 class FunctionVersion(_RemoteValue):
    """An exact immutable Function version returned by Enterprise.

-    Scheduling resources, priority, concurrency, and retry policy belong to
-    the submitting Job and are not part of this identity.
+    The GPU execution requirement is part of this identity. CPU and memory sizing,
+    priority, concurrency, and retry policy belong to the execution platform.
    """

    name: str
@@ -479,6 +520,13 @@ class RefreshColumnResult(_RemoteValue):


 _FUNCTION_NAME = re.compile(r"^[A-Za-z_][A-Za-z0-9_.-]*$")
+_FUNCTION_BLOB_V2_TYPE = "blob_v2"
+_ARROW_EXTENSION_NAME_KEY = "ARROW:extension:name"
+_BLOB_V2_EXTENSION_NAME = "lance.blob.v2"
+_NESTED_BLOB_COLLECTION_ERROR = (
+    "unsupported Arrow type for Function signature: Blob v2 fields nested under "
+    "collection types are not supported"
+)


 _GRAMMAR_PRIMITIVES = (
@@ -495,6 +543,7 @@ _GRAMMAR_PRIMITIVES = (
    (pa.float32(), "float32"),
    (pa.float64(), "float64"),
    (pa.string(), "utf8"),
+    (pa.large_string(), "large_utf8"),
    (pa.binary(), "binary"),
    (pa.date32(), "date32"),
    (pa.date64(), "date64"),
@@ -502,31 +551,258 @@ _GRAMMAR_PRIMITIVES = (


 def _canonical_arrow_type(data_type: pa.DataType) -> str:
-    """The server's V1 Function type grammar. Anything outside it is rejected
-    here rather than at registration."""
+    """The compact Function grammar, or canonical exact JSON for nested types."""
+    grammar = _grammar_arrow_type(data_type)
+    if grammar is not None:
+        return grammar
+    exact = _exact_arrow_type(data_type)
+    return json.dumps(exact, ensure_ascii=False, sort_keys=True, separators=(",", ":"))
+
+
+def _grammar_arrow_type(data_type: pa.DataType) -> Optional[str]:
    for candidate, name in _GRAMMAR_PRIMITIVES:
        if data_type == candidate:
            return name
    if pa.types.is_list(data_type) or pa.types.is_large_list(data_type):
+        item = _grammar_list_item(data_type)
+        if item is None:
+            return None
        prefix = "list" if pa.types.is_list(data_type) else "large_list"
-        return f"{prefix}<{_canonical_list_item(data_type)}>"
+        return f"{prefix}<{item}>"
    if pa.types.is_fixed_size_list(data_type) and data_type.list_size > 0:
-        return (
-            f"fixed_size_list<{_canonical_list_item(data_type)}, {data_type.list_size}>"
-        )
-    raise TypeError(f"unsupported Arrow type for Function signature: {data_type}")
+        item = _grammar_list_item(data_type)
+        if item is not None:
+            return f"fixed_size_list<{item}, {data_type.list_size}>"
+    return None


-def _canonical_list_item(data_type: pa.DataType) -> str:
+def _grammar_list_item(data_type: pa.DataType) -> Optional[str]:
    """The grammar names only the item type; it always means a non-nullable
-    child called `item`, so any other child metadata cannot be represented."""
+    child called `item`, so other child properties require exact JSON."""
    child = data_type.value_field
    if child.name != "item" or child.nullable or child.metadata:
+        return None
+    return _grammar_arrow_type(child.type)
+
+
+def _validate_exact_arrow_field(field: pa.Field) -> None:
+    if not field.name:
        raise TypeError(
-            "unsupported Arrow type for Function signature: list items must be a "
-            f"non-nullable field named 'item', got {child}"
+            "unsupported Arrow type for Function signature: field names "
+            "must not be empty"
        )
-    return _canonical_arrow_type(child.type)
+    if _is_blob_v2_field(field):
+        if not _has_supported_blob_v2_layout(field):
+            raise TypeError(
+                "unsupported Arrow type for Function signature: lance.blob.v2 "
+                f"requires a supported Blob storage layout, got {field}"
+            )
+        metadata = {
+            (key.decode() if isinstance(key, bytes) else key): (
+                value.decode() if isinstance(value, bytes) else value
+            )
+            for key, value in (field.metadata or {}).items()
+        }
+        if metadata and metadata != {
+            _ARROW_EXTENSION_NAME_KEY: _BLOB_V2_EXTENSION_NAME
+        }:
+            raise TypeError(
+                "unsupported Arrow type for Function signature: lance.blob.v2 "
+                "field metadata must contain only its canonical extension marker"
+            )
+    elif field.metadata:
+        raise TypeError(
+            "unsupported Arrow type for Function signature: field metadata "
+            f"is not supported, got {field}"
+        )
+
+
+def _has_supported_blob_v2_layout(field: pa.Field) -> bool:
+    data_type = field.type
+    if isinstance(data_type, pa.ExtensionType):
+        data_type = data_type.storage_type
+    if not pa.types.is_struct(data_type):
+        return False
+
+    fields = tuple(data_type)
+
+    def matches(spec, compare_nullable) -> bool:
+        return len(fields) == len(spec) and all(
+            actual.name == name
+            and actual.type == expected_type
+            and (not check_nullable or actual.nullable == nullable)
+            for actual, (name, expected_type, nullable), check_nullable in zip(
+                fields, spec, compare_nullable
+            )
+        )
+
+    logical_minimal = (
+        ("data", pa.large_binary(), True),
+        ("uri", pa.utf8(), True),
+    )
+    logical_full = logical_minimal + (
+        ("position", pa.uint64(), True),
+        ("size", pa.uint64(), True),
+    )
+    prepared = (
+        ("kind", pa.uint8(), True),
+        ("data", pa.large_binary(), True),
+        ("uri", pa.utf8(), True),
+        ("blob_id", pa.uint32(), True),
+        ("blob_size", pa.uint64(), True),
+        ("position", pa.uint64(), True),
+    )
+    descriptor = (
+        ("kind", pa.uint8(), False),
+        ("position", pa.uint64(), False),
+        ("size", pa.uint64(), False),
+        ("blob_id", pa.uint32(), False),
+        ("blob_uri", pa.utf8(), False),
+    )
+    return (
+        matches(logical_minimal, (True, True))
+        or matches(logical_full, (True, True, False, False))
+        or matches(prepared, (True,) * len(prepared))
+        or matches(descriptor, (False,) * len(descriptor))
+    )
+
+
+def _canonical_arrow_field(field: pa.Field) -> str:
+    _validate_exact_arrow_field(field)
+    if _is_blob_v2_field(field):
+        return _FUNCTION_BLOB_V2_TYPE
+    return _canonical_arrow_type(field.type)
+
+
+def _blob_storage_type(field: pa.Field) -> pa.DataType:
+    data_type = field.type
+    if isinstance(data_type, pa.ExtensionType):
+        return data_type.storage_type
+    return data_type
+
+
+def _exact_blob_storage_type(field: pa.Field) -> dict[str, Any]:
+    storage = _blob_storage_type(field)
+    if not pa.types.is_struct(storage):
+        raise TypeError(
+            "unsupported Arrow type for Function signature: lance.blob.v2 "
+            "requires struct storage"
+        )
+    return {
+        "type": "struct",
+        "fields": [
+            {
+                "name": child.name,
+                "nullable": child.nullable,
+                "type": (
+                    {"type": "large_binary"}
+                    if pa.types.is_large_binary(child.type)
+                    else _exact_arrow_type(child.type)
+                ),
+            }
+            for child in storage
+        ],
+    }
+
+
+def _data_type_has_blob_v2(data_type: pa.DataType) -> bool:
+    if pa.types.is_struct(data_type):
+        return any(
+            _is_blob_v2_field(field) or _data_type_has_blob_v2(field.type)
+            for field in data_type
+        )
+    if (
+        pa.types.is_list(data_type)
+        or pa.types.is_large_list(data_type)
+        or pa.types.is_fixed_size_list(data_type)
+    ):
+        field = data_type.value_field
+        return _is_blob_v2_field(field) or _data_type_has_blob_v2(field.type)
+    if pa.types.is_map(data_type):
+        return any(
+            _is_blob_v2_field(field) or _data_type_has_blob_v2(field.type)
+            for field in (data_type.key_field, data_type.item_field)
+        )
+    return False
+
+
+def _exact_arrow_field(
+    field: pa.Field, *, inside_collection: bool = False
+) -> dict[str, Any]:
+    _validate_exact_arrow_field(field)
+    if _is_blob_v2_field(field):
+        if inside_collection:
+            raise TypeError(_NESTED_BLOB_COLLECTION_ERROR)
+        return {
+            "name": field.name,
+            "nullable": field.nullable,
+            "type": _exact_blob_storage_type(field),
+            "metadata": {
+                _ARROW_EXTENSION_NAME_KEY: _BLOB_V2_EXTENSION_NAME,
+            },
+        }
+    value = {
+        "name": field.name,
+        "nullable": field.nullable,
+        "type": _exact_arrow_type(field.type, inside_collection=inside_collection),
+    }
+    return value
+
+
+def _exact_arrow_type(
+    data_type: pa.DataType, *, inside_collection: bool = False
+) -> dict[str, Any]:
+    for candidate, name in _GRAMMAR_PRIMITIVES:
+        if data_type == candidate:
+            return {"type": name}
+    if pa.types.is_struct(data_type):
+        fields = list(data_type)
+        names = [field.name for field in fields]
+        if not fields or len(set(names)) != len(names):
+            raise TypeError(
+                "unsupported Arrow type for Function signature: structs must have "
+                "non-empty, uniquely named fields"
+            )
+        return {
+            "type": "struct",
+            "fields": [
+                _exact_arrow_field(field, inside_collection=inside_collection)
+                for field in fields
+            ],
+        }
+    if (
+        pa.types.is_list(data_type)
+        or pa.types.is_large_list(data_type)
+        or pa.types.is_fixed_size_list(data_type)
+    ):
+        if pa.types.is_fixed_size_list(data_type):
+            if data_type.value_field.name != "item":
+                raise TypeError(
+                    "unsupported Arrow type for Function signature: fixed-size list "
+                    "items must be named 'item'"
+                )
+            if data_type.list_size <= 0:
+                raise TypeError(
+                    f"unsupported Arrow type for Function signature: {data_type}"
+                )
+        value: dict[str, Any] = {
+            "type": (
+                "list"
+                if pa.types.is_list(data_type)
+                else "large_list"
+                if pa.types.is_large_list(data_type)
+                else "fixed_size_list"
+            ),
+            "fields": [
+                _exact_arrow_field(data_type.value_field, inside_collection=True)
+            ],
+        }
+        if pa.types.is_fixed_size_list(data_type):
+            value["length"] = data_type.list_size
+        return value
+    if pa.types.is_map(data_type) and _data_type_has_blob_v2(data_type):
+        raise TypeError(_NESTED_BLOB_COLLECTION_ERROR)
+    raise TypeError(f"unsupported Arrow type for Function signature: {data_type}")


 def _list_of(item: pa.DataType) -> pa.DataType:
@@ -600,8 +876,15 @@ def _callable_parameters(function: Callable[..., Any]) -> tuple[inspect.Paramete

 def _function_output(output: pa.DataType | pa.Field | pa.Schema) -> FunctionOutput:
    if isinstance(output, pa.Schema):
+        if output.metadata:
+            raise TypeError("Function output schema metadata is not supported")
        fields = tuple(output)
-    elif isinstance(output, pa.Field) and pa.types.is_struct(output.type):
+    elif (
+        isinstance(output, pa.Field)
+        and not _is_blob_v2_field(output)
+        and pa.types.is_struct(output.type)
+    ):
+        _validate_exact_arrow_field(output)
        if output.nullable:
            raise ValueError("Function output must be non-nullable")
        fields = tuple(output.type)
@@ -617,11 +900,12 @@ def _function_output(output: pa.DataType | pa.Field | pa.Schema) -> FunctionOutp
            raise TypeError(
                "output_schema must be a PyArrow DataType, Field, or Schema"
            )
+        _validate_exact_arrow_field(field)
        if field.nullable:
            raise ValueError("Function output must be non-nullable")
        return FunctionOutput(
            kind="scalar",
-            arrow_type=_canonical_arrow_type(field.type),
+            arrow_type=_canonical_arrow_field(field),
            nullable=False,
        )

@@ -629,6 +913,8 @@ def _function_output(output: pa.DataType | pa.Field | pa.Schema) -> FunctionOutp
        raise ValueError("named-struct Function output must contain at least one field")
    if any(field.nullable for field in fields):
        raise ValueError("Function output fields must be non-nullable")
+    for field in fields:
+        _validate_exact_arrow_field(field)
    names = [field.name for field in fields]
    if len(set(names)) != len(names):
        raise ValueError("Function output field names must be unique")
@@ -637,7 +923,7 @@ def _function_output(output: pa.DataType | pa.Field | pa.Schema) -> FunctionOutp
        fields=tuple(
            FunctionResultField(
                name=field.name,
-                arrow_type=_canonical_arrow_type(field.type),
+                arrow_type=_canonical_arrow_field(field),
                nullable=False,
            )
            for field in fields
@@ -657,6 +943,10 @@ def _infer_signature(
    if input_schema is not None:
        if not isinstance(input_schema, pa.Schema):
            raise TypeError("input_schema must be a PyArrow Schema")
+        if input_schema.metadata:
+            raise TypeError("Function input schema metadata is not supported")
+        for field in input_schema:
+            _validate_exact_arrow_field(field)
        expected = tuple(parameter.name for parameter in parameters)
        actual = tuple(input_schema.names)
        if actual != expected:
@@ -667,7 +957,7 @@ def _infer_signature(
        inputs = tuple(
            FunctionParameter(
                name=field.name,
-                arrow_type=_canonical_arrow_type(field.type),
+                arrow_type=_canonical_arrow_field(field),
                nullable=field.nullable,
            )
            for field in input_schema
@@ -690,7 +980,9 @@ def _infer_signature(
        inputs.append(
            FunctionParameter(
                name=parameter.name,
-                arrow_type=_canonical_arrow_type(data_type),
+                arrow_type=_canonical_arrow_field(
+                    pa.field(parameter.name, data_type, nullable=nullable)
+                ),
                nullable=nullable,
            )
        )
@@ -910,6 +1202,7 @@ class UdfDefinition:
        pip: tuple[str, ...],
        env: Mapping[str, str],
        python_version: Optional[str],
+        gpu: bool = False,
        conda: tuple[str, ...] = (),
        conda_channels: tuple[str, ...] = (),
    ):
@@ -938,12 +1231,14 @@ class UdfDefinition:
        signature = _infer_signature(function, input_schema, output_schema)
        source = _package_source(function)
        digest = f"sha256:{hashlib.sha256(source).hexdigest()}"
+        gpu_marker = _normalize_gpu_marker(gpu)
        runtime = PythonRuntimeSpec(
-            kind="python",
+            kind="python_v2" if gpu_marker is not None else "python",
            python_version=python_version
            or f"{sys.version_info.major}.{sys.version_info.minor}",
            environment=environment_spec,
            env=environment,
+            gpu=gpu_marker,
        )
        self._function = function
        self._request = FunctionRegistrationRequest(
@@ -989,6 +1284,7 @@ def udf(
    pip: tuple[str, ...] | list[str] = (),
    env: Optional[Mapping[str, str]] = None,
    python_version: Optional[str] = None,
+    gpu: bool = False,
    conda: tuple[str, ...] | list[str] = (),
    conda_channels: tuple[str, ...] | list[str] = (),
 ) -> Callable[[Callable[..., Any]], UdfDefinition]: ...
@@ -1003,6 +1299,7 @@ def udf(
    pip: tuple[str, ...] | list[str] = (),
    env: Optional[Mapping[str, str]] = None,
    python_version: Optional[str] = None,
+    gpu: bool = False,
    conda: tuple[str, ...] | list[str] = (),
    conda_channels: tuple[str, ...] | list[str] = (),
 ):
@@ -1035,6 +1332,10 @@ def udf(
        Environment variables included in the Function definition.
    python_version : str, optional
        Remote Python major/minor version. Defaults to the client version.
+    gpu : bool, default False
+        Whether every remote execution requires a GPU. The execution platform
+        selects one compatible GPU for each worker. The requirement is part of
+        the immutable Function version.

    The packaged artifact is a snapshot: the function source plus exactly
    the module-level names it references (modules as imports, importable
@@ -1059,6 +1360,11 @@ def udf(
    ...     return value * 2
    >>> score(1.5)
    3.0
+    >>> @udf(pip=["cupy-cuda12x"], gpu=True)
+    ... def gpu_score(value: int) -> int:
+    ...     return value * 2
+    >>> gpu_score.registration_request.runtime.gpu
+    True
    """

    def decorate(target: Callable[..., Any]) -> UdfDefinition:
@@ -1070,6 +1376,7 @@ def udf(
            pip=tuple(pip),
            env={} if env is None else env,
            python_version=python_version,
+            gpu=gpu,
            conda=tuple(conda),
            conda_channels=tuple(conda_channels),
        )
@@ -749,6 +749,10 @@ class RemoteDBConnection(DBConnection):
    def get_function(self, name: str, *, version: str) -> FunctionVersion:
        return LOOP.run(self._conn.get_function(name, version=version))

+    @override
+    def drop_function(self, name: str, *, version: str) -> bool:
+        return LOOP.run(self._conn.drop_function(name, version=version))
+
    @override
    def list_jobs(self) -> List["JobInfo"]:
        """List server-side jobs across the database's tables."""
@@ -12,6 +12,8 @@ from datetime import date
 import http.server
 import json
 from pathlib import Path
+import subprocess
+import sys
 import threading
 from typing import Optional

@@ -19,7 +21,13 @@ import pyarrow as pa
 import pytest

 import lancedb
-from lancedb.functions import UdfDefinition, udf
+from lancedb.functions import (
+    PythonRuntimeSpec,
+    UdfDefinition,
+    _canonical_arrow_type,
+    _GRAMMAR_PRIMITIVES,
+    udf,
+)

 THRESHOLD = 20
 _CACHE = None
@@ -61,6 +69,80 @@ def test_scalar_udf_matches_shared_registration_golden_and_remains_callable():
    }


+def _main_udf_source(
+    *, threshold: int = 20, input_annotation: str = "int", comparison: str = ">="
+) -> str:
+    return (
+        "from __future__ import annotations\n"
+        "from lancedb.functions import udf\n"
+        f"THRESHOLD = {threshold}\n"
+        "\n"
+        "@udf\n"
+        f"def label(value: {input_annotation}) -> str:\n"
+        f"    return 'big' if value {comparison} THRESHOLD else 'small'\n"
+        "\n"
+        "assert label.__module__ == '__main__'\n"
+        "print(label.registration_request.to_canonical_json())\n"
+    )
+
+
+def _run_main_udf(path: Path, source: str) -> dict:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(source)
+    result = subprocess.run(
+        [sys.executable, str(path)],
+        check=True,
+        capture_output=True,
+        text=True,
+    )
+    return json.loads(result.stdout)
+
+
+def test_main_udf_registration_identity_is_stable_across_processes_and_paths(
+    tmp_path,
+):
+    source = _main_udf_source()
+    original_path = tmp_path / "original" / "job.py"
+    moved_path = tmp_path / "moved" / "renamed_job.py"
+
+    original_runs = [_run_main_udf(original_path, source) for _ in range(2)]
+    moved_run = _run_main_udf(moved_path, source)
+
+    assert len({run["artifact"]["digest"] for run in [*original_runs, moved_run]}) == 1
+    assert all(
+        run["signature"] == original_runs[0]["signature"]
+        for run in [original_runs[1], moved_run]
+    )
+    assert original_runs[0] == original_runs[1] == moved_run
+
+    body_change = _run_main_udf(
+        tmp_path / "changes" / "body.py", _main_udf_source(comparison=">")
+    )
+    global_change = _run_main_udf(
+        tmp_path / "changes" / "global.py", _main_udf_source(threshold=21)
+    )
+    annotation_change = _run_main_udf(
+        tmp_path / "changes" / "annotation.py",
+        _main_udf_source(input_annotation="float"),
+    )
+
+    baseline = original_runs[0]
+    assert baseline["signature"] == body_change["signature"]
+    assert baseline["signature"] == global_change["signature"]
+    assert baseline["signature"] != annotation_change["signature"]
+    assert (
+        len(
+            {
+                baseline["artifact"]["digest"],
+                body_change["artifact"]["digest"],
+                global_change["artifact"]["digest"],
+                annotation_change["artifact"]["digest"],
+            }
+        )
+        == 4
+    )
+
+
 def _run_packaged(definition, *args):
    """Execute the shipped artifact in a fresh namespace, as a worker would."""
    source = base64.b64decode(definition.registration_request.artifact.content.data)
@@ -89,6 +171,58 @@ def test_udf_conda_environment():
        udf(name="channels", conda_channels=["conda-forge"])(lambda value: value)


+def test_udf_gpu_marker_uses_gpu_runtime():
+    @udf(pip=["cupy-cuda12x"], gpu=True)
+    def double_on_gpu(value: int) -> int:
+        return value * 2
+
+    request = json.loads(double_on_gpu.registration_request.to_canonical_json())
+    assert request["runtime"]["kind"] == "python_v2"
+    assert request["runtime"]["gpu"] is True
+
+    @udf(pip=["pyarrow"])
+    def cpu_function(value: int) -> int:
+        return value
+
+    cpu_runtime = json.loads(cpu_function.registration_request.to_canonical_json())[
+        "runtime"
+    ]
+    assert cpu_runtime["kind"] == "python"
+    assert "gpu" not in cpu_runtime
+
+    def identity(value: int) -> int:
+        return value
+
+    for invalid in [None, 0, 1, -1, 1.5, "", "true", "1", "H100"]:
+        with pytest.raises(ValueError, match="gpu must be a boolean"):
+            udf(name="invalid_gpu", gpu=invalid)(identity)
+
+    base_runtime = {
+        "kind": "python_v2",
+        "python_version": "3.12",
+        "environment": {"kind": "pip"},
+    }
+    runtime = PythonRuntimeSpec.model_validate({**base_runtime, "gpu": True})
+    assert runtime.gpu is True
+    for invalid in [False, 1, 0, "", "true", "1", "H100"]:
+        with pytest.raises(ValueError, match="runtime.gpu must be true"):
+            PythonRuntimeSpec.model_validate({**base_runtime, "gpu": invalid})
+
+
+def test_unknown_runtime_discards_payload_before_known_field_validation():
+    for payload in [
+        {"kind": "python_v3", "gpu": {"model": "H100"}},
+        {"kind": "python_v3", "resources": []},
+        {
+            "kind": "python_v3",
+            "environment": {"kind": []},
+            "python_version": 3.15,
+        },
+    ]:
+        runtime = PythonRuntimeSpec.model_validate(payload)
+        assert runtime.to_canonical_json() == '{"kind":"python_v3"}'
+
+
 def test_udf_packages_attribute_access_and_body_imports():
    @udf
    def word_norm(body: str) -> float:
@@ -168,9 +302,7 @@ def test_udf_resolves_module_globals_before_builtins(tmp_path):
        udf(module.uses_callable_shadow)


-def test_canonical_arrow_type_is_exactly_the_grammar():
-    from lancedb.functions import _GRAMMAR_PRIMITIVES, _canonical_arrow_type
-
+def test_canonical_arrow_type_prefers_the_compact_grammar():
    golden = json.loads(
        (
            Path(__file__).parents[3]
@@ -181,14 +313,19 @@ def test_canonical_arrow_type_is_exactly_the_grammar():
        case["arrow_type"] for case in golden["valid"] if "<" not in case["arrow_type"]
    ]
    assert [name for _, name in _GRAMMAR_PRIMITIVES] == primitives
+    assert _canonical_arrow_type(pa.list_(pa.field("item", pa.float32(), False))) == (
+        "list<float32>"
+    )
+    assert (
+        _canonical_arrow_type(pa.large_list(pa.field("item", pa.float32(), False)))
+        == "large_list<float32>"
+    )
    for outside in [
        pa.timestamp("us"),
        pa.decimal128(10, 2),
-        pa.large_string(),
        pa.large_binary(),
        pa.binary(4),
        pa.duration("s"),
-        pa.struct([pa.field("a", pa.int32())]),
        pa.list_(pa.float32(), 0),
        pa.list_(pa.timestamp("us")),
    ]:
@@ -378,14 +515,27 @@ def test_udf_recursion_versus_a_rebound_module_name(tmp_path):
        udf(raw_fact)


-def test_canonical_arrow_type_rejects_unrepresentable_list_children():
-    from lancedb.functions import _canonical_arrow_type
-
+def test_canonical_arrow_type_uses_exact_json_for_list_child_properties():
+    nullable = pa.list_(pa.float32())
+    assert json.loads(_canonical_arrow_type(nullable)) == {
+        "type": "list",
+        "fields": [
+            {
+                "name": "item",
+                "nullable": True,
+                "type": {"type": "float32"},
+            }
+        ],
+    }
+    named = pa.list_(pa.field("custom", pa.float32(), nullable=False))
+    assert json.loads(_canonical_arrow_type(named))["fields"][0]["name"] == "custom"
    for outside in [
-        pa.list_(pa.float32()),  # pyarrow default: nullable child
-        pa.list_(pa.field("custom", pa.float32(), nullable=False)),
        pa.list_(pa.field("item", pa.float32(), nullable=False, metadata={"k": "v"})),
        pa.list_(pa.field("item", pa.float32(), nullable=False), 0),
+        pa.list_(
+            pa.field("item", pa.float32(), nullable=False, metadata={"k": "v"}), 3
+        ),
+        pa.list_(pa.field("custom", pa.float32(), nullable=False), 3),
    ]:
        with pytest.raises(TypeError, match="unsupported Arrow type"):
            _canonical_arrow_type(outside)
@@ -395,6 +545,29 @@ def test_canonical_arrow_type_rejects_unrepresentable_list_children():
        )
        == "fixed_size_list<float32, 3>"
    )
+    fixed = json.loads(_canonical_arrow_type(pa.list_(pa.float32(), 3)))
+    assert fixed == {
+        "type": "fixed_size_list",
+        "fields": [
+            {
+                "name": "item",
+                "nullable": True,
+                "type": {"type": "float32"},
+            }
+        ],
+        "length": 3,
+    }
+    large = json.loads(_canonical_arrow_type(pa.large_list(pa.float32())))
+    assert large["type"] == "large_list"
+    assert large["fields"][0]["nullable"] is True
+
+    for invalid_struct in [
+        pa.struct([]),
+        pa.struct([pa.field("a", pa.int32()), pa.field("a", pa.int64())]),
+        pa.struct([pa.field("", pa.int32())]),
+    ]:
+        with pytest.raises(TypeError, match="unsupported Arrow type"):
+            _canonical_arrow_type(invalid_struct)


 def _calls_missing(value: int) -> int:
@@ -432,6 +605,7 @@ def _arrow_type_from_golden(spec: dict) -> pa.DataType:
        "null": pa.null(),
        "bool": pa.bool_(),
        "utf8": pa.string(),
+        "large_utf8": pa.large_string(),
        "binary": pa.binary(),
        "float16": pa.float16(),
        "float32": pa.float32(),
@@ -448,8 +622,6 @@ def test_arrow_type_grammar_matches_the_shared_golden():
            / "rust/lancedb/tests/fixtures/first_class_functions/v1/arrow_types.json"
        ).read_text()
    )
-    from lancedb.functions import _canonical_arrow_type
-
    emitted = {
        case["arrow_type"]: _canonical_arrow_type(_arrow_type_from_golden(case["json"]))
        for case in golden["valid"]
@@ -482,6 +654,395 @@ def test_explicit_arrow_schema_is_deterministic():
    assert signature.output.nullable is False


+def test_blob_fields_use_the_scalar_function_semantic_type():
+    @udf(
+        input_schema=pa.schema([lancedb.blob("image", nullable=False)]),
+        output_schema=lancedb.blob("result", nullable=False),
+    )
+    def copy_blob(image):
+        return image
+
+    signature = copy_blob.registration_request.signature
+    assert signature.inputs[0].arrow_type == "blob_v2"
+    assert signature.output.kind == "scalar"
+    assert signature.output.arrow_type == "blob_v2"
+
+
+def test_whole_named_struct_function_can_include_a_blob_result_field():
+    @udf(
+        input_schema=pa.schema([lancedb.blob("image", nullable=False)]),
+        output_schema=pa.field(
+            "payload",
+            pa.struct(
+                [
+                    pa.field("mime_type", pa.string(), nullable=False),
+                    lancedb.blob("image", nullable=False),
+                ]
+            ),
+            nullable=False,
+        ),
+    )
+    def inspect_blob(image):
+        return {"mime_type": "image/png", "image": image}
+
+    output = inspect_blob.registration_request.signature.output
+    assert output.kind == "named_struct"
+    assert [(field.name, field.arrow_type) for field in output.fields] == [
+        ("mime_type", "utf8"),
+        ("image", "blob_v2"),
+    ]
+
+
+def test_struct_blob_signature_fields_preserve_exact_metadata_and_nullability():
+    nested_input = pa.field(
+        "payload",
+        pa.struct(
+            [
+                pa.field("mime_type", pa.string(), nullable=False),
+                pa.field(
+                    "nested",
+                    pa.struct([lancedb.blob("image", nullable=True)]),
+                    nullable=True,
+                ),
+            ]
+        ),
+        nullable=True,
+    )
+    nested_output = pa.field(
+        "result",
+        pa.struct(
+            [
+                pa.field("mime_type", pa.string(), nullable=False),
+                pa.field(
+                    "nested",
+                    pa.struct([lancedb.blob("image", nullable=True)]),
+                    nullable=False,
+                ),
+            ]
+        ),
+        nullable=False,
+    )
+
+    @udf(input_schema=pa.schema([nested_input]), output_schema=nested_output)
+    def copy_payload(payload):
+        return payload
+
+    signature = copy_payload.registration_request.signature
+    input_type = json.loads(signature.inputs[0].arrow_type)
+    assert input_type["fields"][1]["nullable"] is True
+    input_blob = input_type["fields"][1]["type"]["fields"][0]
+    assert input_blob["nullable"] is True
+    assert input_blob["metadata"] == {"ARROW:extension:name": "lance.blob.v2"}
+
+    assert signature.output.kind == "named_struct"
+    nested_result = next(
+        field for field in signature.output.fields if field.name == "nested"
+    )
+    output_type = json.loads(nested_result.arrow_type)
+    output_blob = output_type["fields"][0]
+    assert output_blob["nullable"] is True
+    assert output_blob["metadata"] == {"ARROW:extension:name": "lance.blob.v2"}
+
+
+def test_struct_blob_signature_supports_multiple_struct_levels():
+    recursive = pa.field(
+        "value",
+        pa.struct(
+            [
+                pa.field(
+                    "level_1",
+                    pa.struct(
+                        [
+                            pa.field(
+                                "level_2",
+                                pa.struct([lancedb.blob("image", nullable=False)]),
+                                nullable=False,
+                            )
+                        ]
+                    ),
+                    nullable=False,
+                )
+            ]
+        ),
+        nullable=False,
+    )
+
+    @udf(
+        input_schema=pa.schema([recursive]),
+        output_schema=pa.field("size", pa.int64(), nullable=False),
+    )
+    def blob_size(value):
+        return len(value["level_1"]["level_2"]["image"])
+
+    encoded = json.loads(blob_size.registration_request.signature.inputs[0].arrow_type)
+    blob = encoded["fields"][0]["type"]["fields"][0]["type"]["fields"][0]
+    assert blob["metadata"]["ARROW:extension:name"] == "lance.blob.v2"
+
+
+@pytest.mark.parametrize(
+    "data_type",
+    [
+        pa.list_(lancedb.blob("item", nullable=False)),
+        pa.large_list(lancedb.blob("item", nullable=False)),
+        pa.list_(lancedb.blob("item", nullable=False), 2),
+        pa.map_(pa.string(), lancedb.blob("value", nullable=False).type),
+    ],
+)
+def test_blob_signature_rejects_collection_ancestors(data_type):
+    with pytest.raises(
+        TypeError,
+        match="Blob v2 fields nested under collection types are not supported",
+    ):
+
+        @udf(
+            input_schema=pa.schema([pa.field("value", data_type, nullable=False)]),
+            output_schema=pa.field("size", pa.int64(), nullable=False),
+        )
+        def blob_size(value):
+            return len(value)
+
+
+def test_blob_signature_rejects_collection_below_a_struct():
+    nested = pa.field(
+        "value",
+        pa.struct(
+            [
+                pa.field(
+                    "images",
+                    pa.list_(lancedb.blob("item", nullable=False)),
+                    nullable=False,
+                )
+            ]
+        ),
+        nullable=False,
+    )
+    with pytest.raises(
+        TypeError,
+        match="Blob v2 fields nested under collection types are not supported",
+    ):
+
+        @udf(
+            input_schema=pa.schema([nested]),
+            output_schema=pa.field("size", pa.int64(), nullable=False),
+        )
+        def blob_size(value):
+            return len(value["images"])
+
+
+def test_named_struct_function_can_include_a_blob_result_field():
+    @udf(
+        input_schema=pa.schema([lancedb.blob("image", nullable=False)]),
+        output_schema=pa.schema(
+            [
+                lancedb.blob("thumbnail", nullable=False),
+                pa.field("width", pa.int32(), nullable=False),
+            ]
+        ),
+    )
+    def inspect_blob(image):
+        return {"thumbnail": image, "width": 1}
+
+    output = inspect_blob.registration_request.signature.output
+    assert output.kind == "named_struct"
+    assert [(field.name, field.arrow_type) for field in output.fields] == [
+        ("thumbnail", "blob_v2"),
+        ("width", "int32"),
+    ]
+
+
+def test_metadata_marked_blob_field_uses_the_semantic_type():
+    extension = lancedb.blob("image", nullable=False).type
+    storage = (
+        extension.storage_type if isinstance(extension, pa.ExtensionType) else extension
+    )
+    metadata_blob = pa.field(
+        "image",
+        storage,
+        nullable=False,
+        metadata={"ARROW:extension:name": "lance.blob.v2"},
+    )
+
+    @udf(
+        input_schema=pa.schema([metadata_blob]),
+        output_schema=pa.field("size", pa.int64(), nullable=False),
+    )
+    def blob_size(image):
+        return len(image)
+
+    assert blob_size.registration_request.signature.inputs[0].arrow_type == "blob_v2"
+
+
+def test_blob_marker_rejects_invalid_storage_layout():
+    malformed = pa.field(
+        "image",
+        pa.int64(),
+        nullable=False,
+        metadata={"ARROW:extension:name": "lance.blob.v2"},
+    )
+
+    with pytest.raises(TypeError, match="requires a supported Blob storage layout"):
+
+        @udf(
+            input_schema=pa.schema([malformed]),
+            output_schema=pa.field("size", pa.int64(), nullable=False),
+        )
+        def blob_size(image):
+            return len(image)
+
+
+def test_nested_non_blob_extension_is_not_silently_unwrapped():
+    class TestExtension(pa.ExtensionType):
+        def __init__(self):
+            super().__init__(pa.int64(), "test.function.extension")
+
+        def __arrow_ext_serialize__(self):
+            return b""
+
+        @classmethod
+        def __arrow_ext_deserialize__(cls, storage_type, serialized):
+            return cls()
+
+    nested = pa.field(
+        "value",
+        pa.struct([pa.field("extended", TestExtension(), nullable=False)]),
+        nullable=False,
+    )
+    with pytest.raises(TypeError, match="unsupported Arrow type"):
+
+        @udf(
+            input_schema=pa.schema([nested]),
+            output_schema=pa.field("result", pa.int64(), nullable=False),
+        )
+        def extension_value(value):
+            return value["extended"]
+
+
+def test_explicit_large_utf8_schemas_use_the_canonical_function_name():
+    input_schema = pa.schema([pa.field("text", pa.large_string(), nullable=True)])
+    output_schema = pa.field("result", pa.large_string(), nullable=False)
+
+    @udf(input_schema=input_schema, output_schema=output_schema)
+    def preserve(text):
+        return text
+
+    signature = preserve.registration_request.signature
+    assert signature.inputs[0].arrow_type == "large_utf8"
+    assert signature.inputs[0].nullable is True
+    assert signature.output.arrow_type == "large_utf8"
+    assert signature.output.nullable is False
+
+    nested = pa.struct([pa.field("text", pa.large_string(), nullable=True)])
+    assert json.loads(_canonical_arrow_type(nested)) == {
+        "type": "struct",
+        "fields": [
+            {
+                "name": "text",
+                "nullable": True,
+                "type": {"type": "large_utf8"},
+            }
+        ],
+    }
+
+
+def test_nested_struct_output_uses_canonical_exact_json():
+    token = pa.struct(
+        [
+            pa.field("position", pa.int32(), nullable=False),
+            pa.field("value", pa.string(), nullable=False),
+            pa.field("length", pa.int32(), nullable=False),
+        ]
+    )
+    analysis = pa.struct(
+        [
+            pa.field("normalized_text", pa.string(), nullable=False),
+            pa.field("has_content", pa.bool_(), nullable=False),
+            pa.field(
+                "metrics",
+                pa.struct(
+                    [
+                        pa.field("character_count", pa.int64(), nullable=False),
+                        pa.field("word_count", pa.int32(), nullable=False),
+                        pa.field("average_word_length", pa.float64(), nullable=False),
+                    ]
+                ),
+                nullable=False,
+            ),
+            pa.field(
+                "diagnostics",
+                pa.struct(
+                    [
+                        pa.field("status", pa.string(), nullable=False),
+                        pa.field(
+                            "normalization",
+                            pa.struct(
+                                [
+                                    pa.field("changed", pa.bool_(), nullable=False),
+                                    pa.field(
+                                        "original_length", pa.int64(), nullable=False
+                                    ),
+                                ]
+                            ),
+                            nullable=False,
+                        ),
+                    ]
+                ),
+                nullable=False,
+            ),
+            pa.field(
+                "token_preview",
+                pa.list_(pa.field("item", token, nullable=False)),
+                nullable=False,
+            ),
+        ]
+    )
+
+    @udf(
+        input_schema=pa.schema([pa.field("text", pa.string(), nullable=False)]),
+        output_schema=pa.field("analysis", analysis, nullable=False),
+    )
+    def analyze(text):
+        return {"normalized_text": text}
+
+    output = analyze.registration_request.signature.output
+    assert output.kind == "named_struct"
+    assert [field.name for field in output.fields] == [
+        "normalized_text",
+        "has_content",
+        "metrics",
+        "diagnostics",
+        "token_preview",
+    ]
+    metrics = json.loads(output.fields[2].arrow_type)
+    assert metrics == {
+        "type": "struct",
+        "fields": [
+            {
+                "name": "character_count",
+                "nullable": False,
+                "type": {"type": "int64"},
+            },
+            {
+                "name": "word_count",
+                "nullable": False,
+                "type": {"type": "int32"},
+            },
+            {
+                "name": "average_word_length",
+                "nullable": False,
+                "type": {"type": "float64"},
+            },
+        ],
+    }
+    preview = json.loads(output.fields[4].arrow_type)
+    assert preview["type"] == "list"
+    assert preview["fields"][0]["type"]["type"] == "struct"
+    assert [field["name"] for field in preview["fields"][0]["type"]["fields"]] == [
+        "position",
+        "value",
+        "length",
+    ]
+
+
 def test_annotation_and_explicit_schema_validation_fail_closed():
    with pytest.raises(TypeError, match="missing Function annotations"):

@@ -525,6 +1086,72 @@ def test_annotation_and_explicit_schema_validation_fail_closed():
        def nullable_explicit(value):
            return value

+    for invalid_field in [
+        pa.field("", pa.int32(), nullable=False),
+        pa.field("result", pa.int32(), nullable=False, metadata={"k": "v"}),
+    ]:
+        with pytest.raises(TypeError, match="unsupported Arrow type"):
+
+            @udf(
+                input_schema=pa.schema([pa.field("value", pa.int64())]),
+                output_schema=pa.schema([invalid_field]),
+            )
+            def invalid_explicit_field(value):
+                return value
+
+    with pytest.raises(TypeError, match="unsupported Arrow type"):
+
+        @udf(
+            input_schema=pa.schema(
+                [pa.field("value", pa.int64(), metadata={"k": "v"})]
+            ),
+            output_schema=pa.int64(),
+        )
+        def input_field_metadata(value):
+            return value
+
+    with pytest.raises(TypeError, match="unsupported Arrow type"):
+
+        @udf(
+            input_schema=pa.schema([pa.field("value", pa.int64())]),
+            output_schema=pa.field(
+                "result", pa.int64(), nullable=False, metadata={"k": "v"}
+            ),
+        )
+        def scalar_output_field_metadata(value):
+            return value
+
+    struct_type = pa.struct([pa.field("value", pa.int64(), nullable=False)])
+    with pytest.raises(TypeError, match="unsupported Arrow type"):
+
+        @udf(
+            input_schema=pa.schema([pa.field("value", pa.int64())]),
+            output_schema=pa.field(
+                "result", struct_type, nullable=False, metadata={"k": "v"}
+            ),
+        )
+        def struct_output_field_metadata(value):
+            return {"value": value}
+
+    for input_schema, output_schema in [
+        (
+            pa.schema([pa.field("value", pa.int64())], metadata={"k": "v"}),
+            pa.int64(),
+        ),
+        (
+            pa.schema([pa.field("value", pa.int64())]),
+            pa.schema(
+                [pa.field("result", pa.int64(), nullable=False)],
+                metadata={"k": "v"},
+            ),
+        ),
+    ]:
+        with pytest.raises(TypeError, match="schema metadata"):
+
+            @udf(input_schema=input_schema, output_schema=output_schema)
+            def schema_metadata(value):
+                return value
+

 def test_local_function_catalog_operations_are_not_supported(tmp_path):
    db = lancedb.connect(tmp_path)
@@ -535,6 +1162,8 @@ def test_local_function_catalog_operations_are_not_supported(tmp_path):
        db.create_function_async(normalize_score)
    with pytest.raises(NotImplementedError, match=message):
        db.get_function("normalize_score", version="fv_exact")
+    with pytest.raises(NotImplementedError, match=message):
+        db.drop_function("normalize_score", version="fv_exact")


@contextlib.contextmanager
@@ -580,6 +1209,12 @@ def _mock_remote_function_catalog():
                    "version": "fv_exact",
                }
                response = state["version"]
+            elif self.path == "/v1/functions/drop":
+                assert body == {
+                    "name": "normalize_score",
+                    "version": "fv_exact",
+                }
+                response = {"dropped": True}
            else:
                status = 404
                response = {"error": "not found"}
@@ -638,3 +1273,40 @@ def test_blocking_remote_registration_returns_function_version():
        "/v1/functions/create",
        "/v1/jobs/describe",
    ]
+
+
+def test_remote_drop_function_sends_exact_version():
+    with _mock_remote_function_catalog() as (host, state):
+        db = lancedb.connect(
+            "db://dev",
+            api_key="fake",
+            host_override=host,
+            client_config={"retry_config": {"retries": 0}},
+        )
+        assert db.drop_function("normalize_score", version="fv_exact") is True
+
+    assert state["requests"] == [
+        (
+            "/v1/functions/drop",
+            {"name": "normalize_score", "version": "fv_exact"},
+        )
+    ]
+
+
+@pytest.mark.asyncio
+async def test_async_remote_drop_function_sends_exact_version():
+    with _mock_remote_function_catalog() as (host, state):
+        db = await lancedb.connect_async(
+            "db://dev",
+            api_key="fake",
+            host_override=host,
+            client_config={"retry_config": {"retries": 0}},
+        )
+        assert await db.drop_function("normalize_score", version="fv_exact") is True
+
+    assert state["requests"] == [
+        (
+            "/v1/functions/drop",
+            {"name": "normalize_score", "version": "fv_exact"},
+        )
+    ]
@@ -629,6 +629,17 @@ impl Connection {
        })
    }

+    pub fn drop_function(
+        self_: PyRef<'_, Self>,
+        name: String,
+        version: String,
+    ) -> PyResult<Bound<'_, PyAny>> {
+        let inner = self_.get_inner()?.clone();
+        future_into_py(self_.py(), async move {
+            inner.drop_function(name, version).await.infer_error()
+        })
+    }
+
    pub fn list_jobs(self_: PyRef<'_, Self>) -> PyResult<Bound<'_, PyAny>> {
        let inner = self_.get_inner()?.clone();
        future_into_py(self_.py(), async move {
@@ -1,6 +1,6 @@
 [package]
 name = "lancedb"
-version = "0.38.0-beta.12"
+version = "0.38.0"
 edition.workspace = true
 description = "LanceDB: A serverless, low-latency vector database for AI applications"
 license.workspace = true
@@ -523,6 +523,21 @@ impl Connection {
            .await
    }

+    /// Drop one exact immutable Function version from the remote catalog.
+    ///
+    /// Returns `true` when the server appended a Dropped transition and
+    /// `false` for an idempotent replay. Local databases return
+    /// [`Error::NotSupported`].
+    pub async fn drop_function(
+        &self,
+        name: impl AsRef<str>,
+        version: impl AsRef<str>,
+    ) -> Result<bool> {
+        self.internal
+            .drop_function(name.as_ref(), version.as_ref())
+            .await
+    }
+
    /// Rename a table in the database.
    ///
    /// This is only supported in LanceDB Cloud.
@@ -307,6 +307,10 @@ pub trait Database:
    ) -> Result<crate::function::FunctionVersion> {
        function_catalog_not_supported()
    }
+    /// Drop one exact immutable Function version from the remote catalog.
+    async fn drop_function(&self, _name: &str, _version: &str) -> Result<bool> {
+        function_catalog_not_supported()
+    }
    /// A [`crate::job::Job`] handle for a server-side job by id, suitable for
    /// waiting on or cancelling the job. The handle is constructed without a
    /// server round trip; an unknown id surfaces when the handle is used.
@@ -539,9 +539,7 @@ impl Database for LanceNamespaceDatabase {
        self.namespace
            .drop_table(drop_request)
            .await
-            .map_err(|e| Error::Runtime {
-                message: format!("Failed to drop table: {}", e),
-            })?;
+            .map_err(|e| map_namespace_lance_error(e, name))?;

        Ok(())
    }
@@ -1495,6 +1493,15 @@ mod tests {
            .expect("Failed to list tables");
        assert!(!table_names_after.contains(&"drop_test".to_string()));

+        let error = conn
+            .drop_table("drop_test", &["test_ns".into()])
+            .await
+            .expect_err("dropping a missing table should fail");
+        assert!(
+            matches!(error, Error::TableNotFound { ref name, .. } if name == "drop_test"),
+            "expected TableNotFound, got: {error:?}"
+        );
+
        // Verify: Cannot open dropped table
        let open_result = conn.open_table("drop_test").execute().await;
        assert!(open_result.is_err());
@@ -15,6 +15,9 @@ use serde_json::Value;

 use crate::{Error, Result};

+/// Semantic Function type for a Blob v2 value.
+pub const FUNCTION_BLOB_V2_TYPE: &str = "blob_v2";
+
 fn invalid_json(error: impl std::fmt::Display) -> Error {
    Error::InvalidInput {
        message: format!("invalid remote Function JSON: {error}"),
@@ -207,6 +210,33 @@ pub enum PythonRuntimeSpec {
        environment: PythonEnvironmentSpec,
        env: BTreeMap<String, String>,
    },
+    /// The GPU-enabled Sophon-managed Python runtime.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use std::collections::BTreeMap;
+    /// use lancedb::function::{PythonEnvironmentSpec, PythonRuntimeSpec};
+    ///
+    /// let runtime = PythonRuntimeSpec::PythonV2 {
+    ///     python_version: "3.12".to_string(),
+    ///     environment: PythonEnvironmentSpec {
+    ///         kind: "pip".to_string(),
+    ///         packages: vec!["cupy-cuda12x".to_string()],
+    ///         channels: Vec::new(),
+    ///         path: None,
+    ///         modules: Vec::new(),
+    ///         image: None,
+    ///     },
+    ///     env: BTreeMap::new(),
+    /// };
+    /// assert!(runtime.requires_gpu());
+    /// ```
+    PythonV2 {
+        python_version: String,
+        environment: PythonEnvironmentSpec,
+        env: BTreeMap<String, String>,
+    },
    /// A runtime kind introduced by a newer server.
    ///
    /// Unknown payload fields are intentionally not retained because the
@@ -219,22 +249,27 @@ impl PythonRuntimeSpec {
    pub fn kind(&self) -> &str {
        match self {
            Self::Python { .. } => "python",
+            Self::PythonV2 { .. } => "python_v2",
            Self::Unrecognized { kind } => kind,
        }
    }

-    /// The Python version for the V1 runtime, or `None` for an unknown kind.
+    /// The Python version for a known Python runtime, or `None` for an unknown kind.
    pub fn python_version(&self) -> Option<&str> {
        match self {
-            Self::Python { python_version, .. } => Some(python_version),
+            Self::Python { python_version, .. } | Self::PythonV2 { python_version, .. } => {
+                Some(python_version)
+            }
            Self::Unrecognized { .. } => None,
        }
    }

-    /// The Python environment for the V1 runtime, or `None` for an unknown kind.
+    /// The Python environment for a known Python runtime, or `None` for an unknown kind.
    pub fn environment(&self) -> Option<&PythonEnvironmentSpec> {
        match self {
-            Self::Python { environment, .. } => Some(environment),
+            Self::Python { environment, .. } | Self::PythonV2 { environment, .. } => {
+                Some(environment)
+            }
            Self::Unrecognized { .. } => None,
        }
    }
@@ -242,38 +277,73 @@ impl PythonRuntimeSpec {
    /// Environment variables, or `None` for an unknown kind.
    pub fn env(&self) -> Option<&BTreeMap<String, String>> {
        match self {
-            Self::Python { env, .. } => Some(env),
+            Self::Python { env, .. } | Self::PythonV2 { env, .. } => Some(env),
            Self::Unrecognized { .. } => None,
        }
    }
+
+    /// Whether the runtime requires a GPU selected by the execution platform.
+    pub fn requires_gpu(&self) -> bool {
+        matches!(self, Self::PythonV2 { .. })
+    }
 }

 #[derive(Deserialize)]
-struct PythonRuntimeWire {
-    kind: String,
-    #[serde(default)]
-    python_version: Option<String>,
-    #[serde(default)]
-    environment: Option<PythonEnvironmentSpec>,
+struct PythonRuntimeV1Wire {
+    python_version: String,
+    environment: PythonEnvironmentSpec,
    #[serde(default)]
    env: BTreeMap<String, String>,
+    #[serde(default)]
+    gpu: Option<Value>,
+}
+
+#[derive(Deserialize)]
+struct PythonRuntimeV2Wire {
+    python_version: String,
+    environment: PythonEnvironmentSpec,
+    #[serde(default)]
+    env: BTreeMap<String, String>,
+    gpu: bool,
 }

 impl<'de> Deserialize<'de> for PythonRuntimeSpec {
    fn deserialize<D: Deserializer<'de>>(deserializer: D) -> std::result::Result<Self, D::Error> {
-        let wire = PythonRuntimeWire::deserialize(deserializer)?;
-        if wire.kind == "python" {
-            Ok(Self::Python {
-                python_version: wire
-                    .python_version
-                    .ok_or_else(|| de::Error::missing_field("python_version"))?,
-                environment: wire
-                    .environment
-                    .ok_or_else(|| de::Error::missing_field("environment"))?,
-                env: wire.env,
-            })
-        } else {
-            Ok(Self::Unrecognized { kind: wire.kind })
+        let value = Value::deserialize(deserializer)?;
+        let kind = value
+            .get("kind")
+            .ok_or_else(|| de::Error::missing_field("kind"))?
+            .as_str()
+            .ok_or_else(|| de::Error::custom("runtime.kind must be a string"))?
+            .to_string();
+        match kind.as_str() {
+            "python" => {
+                let wire: PythonRuntimeV1Wire =
+                    serde_json::from_value(value).map_err(de::Error::custom)?;
+                if wire.gpu.is_some() {
+                    return Err(de::Error::custom(
+                        "python runtime with gpu requires kind='python_v2'",
+                    ));
+                }
+                Ok(Self::Python {
+                    python_version: wire.python_version,
+                    environment: wire.environment,
+                    env: wire.env,
+                })
+            }
+            "python_v2" => {
+                let wire: PythonRuntimeV2Wire =
+                    serde_json::from_value(value).map_err(de::Error::custom)?;
+                if !wire.gpu {
+                    return Err(de::Error::custom("runtime.gpu must be true"));
+                }
+                Ok(Self::PythonV2 {
+                    python_version: wire.python_version,
+                    environment: wire.environment,
+                    env: wire.env,
+                })
+            }
+            _ => Ok(Self::Unrecognized { kind }),
        }
    }
 }
@@ -287,6 +357,8 @@ impl Serialize for PythonRuntimeSpec {
            environment: &'a PythonEnvironmentSpec,
            #[serde(skip_serializing_if = "BTreeMap::is_empty")]
            env: &'a BTreeMap<String, String>,
+            #[serde(skip_serializing_if = "Option::is_none")]
+            gpu: Option<bool>,
        }

        #[derive(Serialize)]
@@ -304,6 +376,19 @@ impl Serialize for PythonRuntimeSpec {
                python_version,
                environment,
                env,
+                gpu: None,
+            }
+            .serialize(serializer),
+            Self::PythonV2 {
+                python_version,
+                environment,
+                env,
+            } => PythonRuntimeRef {
+                kind: "python_v2",
+                python_version,
+                environment,
+                env,
+                gpu: Some(true),
            }
            .serialize(serializer),
            Self::Unrecognized { kind } => UnrecognizedRuntimeRef { kind }.serialize(serializer),
@@ -313,8 +398,8 @@ impl Serialize for PythonRuntimeSpec {

 /// Immutable Function version returned by the Enterprise catalog.
 ///
-/// Scheduling resources, priority, concurrency, and retry policy belong to
-/// the submitting Job and are not part of this identity.
+/// The GPU execution requirement is part of this identity. CPU and memory sizing,
+/// priority, concurrency, and retry policy belong to the execution platform.
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
 pub struct FunctionVersion {
    name: String,
@@ -589,7 +674,7 @@ impl_json!(RefreshColumnResult);

 #[cfg(test)]
 mod conda_environment_tests {
-    use super::PythonEnvironmentSpec;
+    use super::{PythonEnvironmentSpec, PythonRuntimeSpec};

    #[test]
    fn conda_channels_round_trip_and_pip_stays_bare() {
@@ -608,4 +693,45 @@ mod conda_environment_tests {
            serde_json::from_str(r#"{"kind":"pip","packages":["numpy"]}"#).unwrap();
        assert!(!serde_json::to_string(&pip).unwrap().contains("channels"));
    }
+
+    #[test]
+    fn gpu_python_runtime_marker_round_trips_and_validates() {
+        let runtime: PythonRuntimeSpec = serde_json::from_str(
+            r#"{"kind":"python_v2","python_version":"3.12","environment":{"kind":"pip"},"gpu":true}"#,
+        )
+        .unwrap();
+        assert_eq!(runtime.kind(), "python_v2");
+        assert!(runtime.requires_gpu());
+        assert_eq!(
+            super::canonical_json(&runtime).unwrap(),
+            r#"{"environment":{"kind":"pip"},"gpu":true,"kind":"python_v2","python_version":"3.12"}"#
+        );
+
+        for invalid in [
+            r#"{"kind":"python","python_version":"3.12","environment":{"kind":"pip"},"gpu":true}"#,
+            r#"{"kind":"python_v2","python_version":"3.12","environment":{"kind":"pip"}}"#,
+            r#"{"kind":"python_v2","python_version":"3.12","environment":{"kind":"pip"},"gpu":1}"#,
+            r#"{"kind":"python_v2","python_version":"3.12","environment":{"kind":"pip"},"gpu":false}"#,
+            r#"{"kind":"python_v2","python_version":"3.12","environment":{"kind":"pip"},"gpu":"true"}"#,
+            r#"{"kind":"python_v2","python_version":"3.12","environment":{"kind":"pip"},"gpu":"H100"}"#,
+        ] {
+            assert!(serde_json::from_str::<PythonRuntimeSpec>(invalid).is_err());
+        }
+    }
+
+    #[test]
+    fn unknown_runtime_discards_payload_before_known_field_validation() {
+        for encoded in [
+            r#"{"kind":"python_v3","gpu":{"model":"H100"}}"#,
+            r#"{"kind":"python_v3","resources":[]}"#,
+            r#"{"kind":"python_v3","python_version":3.15,"environment":{"kind":[]}}"#,
+        ] {
+            let runtime: PythonRuntimeSpec = serde_json::from_str(encoded).unwrap();
+            assert_eq!(runtime.kind(), "python_v3");
+            assert_eq!(
+                super::canonical_json(&runtime).unwrap(),
+                r#"{"kind":"python_v3"}"#
+            );
+        }
+    }
 }
@@ -533,6 +533,11 @@ struct RemoteListJobsResponse {
    page_token: Option<String>,
 }

+#[derive(serde::Deserialize)]
+struct RemoteDropFunctionResponse {
+    dropped: bool,
+}
+
 /// Bound on `list_jobs` page walking; a warning is logged when the listing
 /// is truncated at this many pages.
 const MAX_LIST_JOBS_PAGES: usize = 100;
@@ -583,6 +588,20 @@ impl<S: HttpSend> Database for RemoteDatabase<S> {
        response.json().await.err_to_http(request_id)
    }

+    async fn drop_function(&self, name: &str, version: &str) -> Result<bool> {
+        let req = self
+            .client
+            .post("/v1/functions/drop")
+            .json(&serde_json::json!({
+                "name": name,
+                "version": version,
+            }));
+        let (request_id, response) = self.client.send(req).await?;
+        let response = self.client.check_response(&request_id, response).await?;
+        let response: RemoteDropFunctionResponse = response.json().await.err_to_http(request_id)?;
+        Ok(response.dropped)
+    }
+
    fn job(&self, job_id: &str) -> Result<crate::job::Job> {
        Ok(crate::job::Job::new(Box::new(super::job::RemoteJob::new(
            self.client.clone(),
@@ -2689,6 +2708,25 @@ mod tests {
        assert_eq!(version.version(), "fv_01K3EXACT");
    }

+    #[tokio::test]
+    async fn test_drop_function_sends_exact_version_and_decodes_replay() {
+        let conn = Connection::new_with_handler(|request| {
+            assert_eq!(request.method(), &reqwest::Method::POST);
+            assert_eq!(request.url().path(), "/v1/functions/drop");
+            let body: serde_json::Value =
+                serde_json::from_slice(request.body().unwrap().as_bytes().unwrap()).unwrap();
+            assert_eq!(
+                body,
+                serde_json::json!({"name": "embed", "version": "fv_01K3EXACT"})
+            );
+            http::Response::builder()
+                .status(200)
+                .body(r#"{"dropped":false}"#)
+                .unwrap()
+        });
+        assert!(!conn.drop_function("embed", "fv_01K3EXACT").await.unwrap());
+    }
+
    #[tokio::test]
    async fn test_conn_job_waits_to_done() {
        let polls = Arc::new(AtomicUsize::new(0));
@@ -5734,9 +5734,8 @@ mod tests {
            ))
            .execute()
            .await;
-        let err = match result {
-            Ok(_) => panic!("legacy remote query unexpectedly succeeded"),
-            Err(err) => err,
+        let Err(err) = result else {
+            panic!("legacy remote query unexpectedly succeeded")
        };

        assert!(
@@ -7465,6 +7464,93 @@ mod tests {
        assert_eq!(result.version, 8);
    }

+    #[tokio::test]
+    async fn test_add_function_column_allows_an_existing_binding() {
+        let binding = crate::function::FunctionBinding::from_json(include_str!(
+            "../../tests/fixtures/first_class_functions/v1/remote_function_binding.json"
+        ))
+        .unwrap();
+        let binding_metadata = crate::table::computed_columns::function_bindings_metadata(
+            std::slice::from_ref(&binding),
+        )
+        .unwrap();
+        let mut fields = vec![
+            Field::new("title", DataType::Utf8, true),
+            Field::new("body", DataType::Utf8, true),
+        ];
+        fields.extend(binding.outputs().iter().map(|output| {
+            let data_type = match output.arrow_type.as_str() {
+                "utf8" => DataType::Utf8,
+                "int64" => DataType::Int64,
+                other => panic!("unexpected fixture output type {other}"),
+            };
+            Field::new(&output.output_name, data_type, true).with_metadata(
+                crate::table::computed_columns::function_computed_column_metadata(
+                    binding.binding_id(),
+                    output.output_ordinal,
+                    &["title".into(), "body".into()],
+                ),
+            )
+        }));
+        let schema = Schema::new_with_metadata(
+            fields,
+            HashMap::from([(
+                crate::table::computed_columns::FUNCTION_BINDINGS_META_KEY.to_string(),
+                binding_metadata,
+            )]),
+        );
+        let table =
+            Table::new_with_handler("my_table", move |request| match request.url().path() {
+                "/v1/table/my_table/describe/" => http::Response::builder()
+                    .status(200)
+                    .body(describe_response(&schema))
+                    .unwrap(),
+                "/v1/table/my_table/add_columns/" => {
+                    let actual: serde_json::Value =
+                        serde_json::from_slice(request.body().unwrap().as_bytes().unwrap())
+                            .unwrap();
+                    assert_eq!(
+                        actual["new_columns"],
+                        serde_json::json!([
+                            {"name":"secondary_text","all_null":true},
+                            {"name":"secondary_token_count","all_null":true}
+                        ])
+                    );
+                    http::Response::builder()
+                        .status(200)
+                        .body(r#"{"version":10}"#.to_string())
+                        .unwrap()
+                }
+                path => panic!("Unexpected path: {path}"),
+            });
+        let application = crate::function::FunctionApplication::from_json(
+            r#"{
+                "function":{"name":"text_features","version":"fv_01K3TEXT"},
+                "inputs":[
+                    {"parameter":"title","kind":"column","value":{"path":"title"}},
+                    {"parameter":"body","kind":"column","value":{"path":"body"}}
+                ],
+                "output":{"kind":"named_struct","fields":[
+                    {"name":"normalized_text","arrow_type":"utf8","nullable":false},
+                    {"name":"token_count","arrow_type":"int64","nullable":false}
+                ]},
+                "columns":{
+                    "normalized_text":"secondary_text",
+                    "token_count":"secondary_token_count"
+                }
+            }"#,
+        )
+        .unwrap();
+
+        let result = table
+            .add_columns()
+            .function(application)
+            .execute()
+            .await
+            .unwrap();
+        assert_eq!(result.version, 10);
+    }
+
    #[tokio::test]
    async fn test_add_fixed_size_list_function_column_declares_the_vector_type() {
        let table = Table::new_with_handler("my_table", |request| {
@@ -5763,6 +5763,13 @@ mod tests {
        assert!(index_bytes > 0);
        assert_eq!(with_index, data_only + index_bytes);

+        // Release builds reject unstable overlay datasets unless explicitly opted in.
+        if !lance_table::feature_flags::can_read_dataset(
+            lance_table::feature_flags::FLAG_UNSTABLE_DATA_OVERLAY_FILES,
+        ) {
+            return;
+        }
+
        // Commit an overlay file supplying new `foo` values for the first three
        // rows of fragment 0. There is no high-level API that writes overlays
        // yet, so write the overlay's data file and commit the `DataOverlay`
@@ -29,14 +29,16 @@ use datafusion_common::{ScalarValue, tree_node::TreeNode};
 use datafusion_expr::Expr;
 use datafusion_physical_plan::PhysicalExpr;
 use lance::dataset::NewColumnTransform;
-use lance_arrow::FieldExt;
-use lance_core::datatypes::{BLOB_V2_DESC_FIELD, format_field_path_minimal, parse_field_path};
+use lance_arrow::{ARROW_EXT_NAME_KEY, BLOB_V2_EXT_NAME, FieldExt};
+use lance_core::datatypes::{
+    BLOB_V2_DESC_FIELD, BlobV2Layout, format_field_path_minimal, parse_field_path,
+};
 use lance_datafusion::planner::Planner;
 use lance_namespace::models::{JsonArrowDataType, JsonArrowField, JsonArrowSchema};
 use serde::{Deserialize, Serialize};
 use serde_json::Value;

-use crate::function::{FunctionApplication, FunctionBinding};
+use crate::function::{FUNCTION_BLOB_V2_TYPE, FunctionApplication, FunctionBinding};
 use crate::utils::resolve_arrow_field_path;
 use crate::{Error, Result};

@@ -547,7 +549,12 @@ fn ensure_known_binding_shape(value: &Value) -> Result<()> {
    Ok(())
 }

-fn resolve_field_path<'a>(schema: &'a ArrowSchema, path: &str) -> Result<&'a ArrowField> {
+struct ResolvedFieldPath<'a> {
+    root: &'a ArrowField,
+    leaf: &'a ArrowField,
+}
+
+fn resolve_field_path<'a>(schema: &'a ArrowSchema, path: &str) -> Result<ResolvedFieldPath<'a>> {
    let parts = lance_core::datatypes::parse_field_path(path).map_err(|e| {
        invalid_function(format!("invalid Function input field path '{path}': {e}"))
    })?;
@@ -556,25 +563,40 @@ fn resolve_field_path<'a>(schema: &'a ArrowSchema, path: &str) -> Result<&'a Arr
            "Function input field path cannot be empty",
        ));
    };
-    let mut field = schema
+    let root = schema
        .field_with_name(root)
        .map_err(|_| invalid_function(format!("unknown Function input column '{path}'")))?;
+    let mut leaf = root;
    for child in children {
-        let DataType::Struct(fields) = field.data_type() else {
+        let DataType::Struct(fields) = leaf.data_type() else {
            return Err(invalid_function(format!(
                "Function input field path '{path}' traverses a non-struct field"
            )));
        };
-        field = fields
+        leaf = fields
            .iter()
            .find(|field| field.name() == child)
            .map(AsRef::as_ref)
            .ok_or_else(|| invalid_function(format!("unknown Function input column '{path}'")))?;
    }
-    Ok(field)
+    Ok(ResolvedFieldPath { root, leaf })
 }

 fn canonical_input_arrow_type(field: &JsonArrowField) -> Result<String> {
+    let is_blob_v2 = field
+        .metadata
+        .as_ref()
+        .and_then(|metadata| metadata.get(ARROW_EXT_NAME_KEY))
+        .map(String::as_str)
+        == Some(BLOB_V2_EXT_NAME);
+    if is_blob_v2 || field.r#type.fields.is_some() {
+        let arrow_field = lance_namespace::schema::convert_json_arrow_field(field)
+            .map_err(|e| invalid_function(format!("invalid Function input field: {e}")))?;
+        validate_function_blob_nesting(&arrow_field, false)?;
+        if is_blob_v2 {
+            return Ok(FUNCTION_BLOB_V2_TYPE.to_string());
+        }
+    }
    if field.r#type.fields.is_none() && field.r#type.length.is_none() {
        Ok(field.r#type.r#type.clone())
    } else {
@@ -584,6 +606,42 @@ fn canonical_input_arrow_type(field: &JsonArrowField) -> Result<String> {
    }
 }

+fn has_supported_blob_v2_layout(field: &ArrowField) -> bool {
+    field.is_blob_v2()
+        && matches!(
+            field.data_type(),
+            DataType::Struct(fields) if BlobV2Layout::classify(fields).is_some()
+        )
+}
+
+fn validate_function_blob_nesting(field: &ArrowField, inside_collection: bool) -> Result<()> {
+    if field.is_blob_v2() {
+        if inside_collection {
+            return Err(invalid_function(format!(
+                "Function field '{}' nests Blob v2 under a collection, which Function signatures do not support",
+                field.name()
+            )));
+        }
+        if !has_supported_blob_v2_layout(field) {
+            return Err(invalid_function(format!(
+                "Function field '{}' has an invalid Blob v2 storage layout",
+                field.name()
+            )));
+        }
+        return Ok(());
+    }
+    match field.data_type() {
+        DataType::Struct(fields) => fields
+            .iter()
+            .try_for_each(|field| validate_function_blob_nesting(field, inside_collection)),
+        DataType::List(field)
+        | DataType::LargeList(field)
+        | DataType::FixedSizeList(field, _)
+        | DataType::Map(field, _) => validate_function_blob_nesting(field, true),
+        _ => Ok(()),
+    }
+}
+
 /// `fixed_size_list<item, size>` -> (`item`, `size`); the comma must sit outside
 /// any nested `<...>`.
 fn split_fixed_size_list(raw: &str) -> Option<(&str, i32)> {
@@ -663,10 +721,82 @@ fn parse_output_arrow_type(raw: &str) -> Result<JsonArrowDataType> {
    Ok(data_type)
 }

+fn function_output_field(name: &str, nullable: bool, raw: &str) -> Result<JsonArrowField> {
+    let field = if raw == FUNCTION_BLOB_V2_TYPE {
+        lance_namespace::schema::arrow_schema_to_json(&ArrowSchema::new(vec![crate::blob(
+            name, nullable,
+        )]))
+        .map_err(|e| invalid_function(format!("could not encode Blob v2 output field: {e}")))?
+        .fields
+        .into_iter()
+        .next()
+        .ok_or_else(|| invalid_function("Blob v2 output field is missing"))?
+    } else {
+        JsonArrowField::new(name.to_string(), nullable, parse_output_arrow_type(raw)?)
+    };
+    let arrow_field = lance_namespace::schema::convert_json_arrow_field(&field)
+        .map_err(|e| invalid_function(format!("invalid Function output field: {e}")))?;
+    validate_function_blob_nesting(&arrow_field, false)?;
+    Ok(field)
+}
+
+fn function_output_field_matches(expected: &ArrowField, actual: &ArrowField) -> bool {
+    expected.name() == actual.name()
+        && expected.is_nullable() == actual.is_nullable()
+        && if expected.is_blob_v2() {
+            has_supported_blob_v2_layout(expected) && has_supported_blob_v2_layout(actual)
+        } else {
+            function_output_type_matches(expected.data_type(), actual.data_type())
+        }
+}
+
+fn function_output_type_matches(expected: &DataType, actual: &DataType) -> bool {
+    if expected == actual {
+        return true;
+    }
+    match (expected, actual) {
+        (DataType::Struct(expected), DataType::Struct(actual)) => {
+            expected.len() == actual.len()
+                && expected
+                    .iter()
+                    .zip(actual)
+                    .all(|(expected, actual)| function_output_field_matches(expected, actual))
+        }
+        (DataType::List(expected), DataType::List(actual))
+        | (DataType::LargeList(expected), DataType::LargeList(actual)) => {
+            function_output_field_matches(expected, actual)
+        }
+        (
+            DataType::FixedSizeList(expected, expected_size),
+            DataType::FixedSizeList(actual, actual_size),
+        ) => expected_size == actual_size && function_output_field_matches(expected, actual),
+        (DataType::Map(expected, expected_sorted), DataType::Map(actual, actual_sorted)) => {
+            expected_sorted == actual_sorted && function_output_field_matches(expected, actual)
+        }
+        _ => false,
+    }
+}
+
+fn function_output_type_has_blob(data_type: &DataType) -> bool {
+    match data_type {
+        DataType::Struct(fields) => fields
+            .iter()
+            .any(|field| field.is_blob_v2() || function_output_type_has_blob(field.data_type())),
+        DataType::List(field)
+        | DataType::LargeList(field)
+        | DataType::FixedSizeList(field, _)
+        | DataType::Map(field, _) => {
+            field.is_blob_v2() || function_output_type_has_blob(field.data_type())
+        }
+        _ => false,
+    }
+}
+
 fn ensure_binding_matches_schema(schema: &ArrowSchema, binding: &FunctionBinding) -> Result<()> {
    let mut input_fields = Vec::with_capacity(binding.inputs().len());
    for input in binding.inputs() {
-        let field = resolve_field_path(schema, &input.field_path)?;
+        let resolved = resolve_field_path(schema, &input.field_path)?;
+        let field = resolved.leaf;
        if field
            .metadata()
            .get(COMPUTED_COLUMN_META_KEY)
@@ -721,6 +851,11 @@ fn ensure_binding_matches_schema(schema: &ArrowSchema, binding: &FunctionBinding
        )));
    }

+    let expected_inputs = binding
+        .inputs()
+        .iter()
+        .map(|input| input.field_path.clone())
+        .collect::<Vec<_>>();
    let mut output_fields = Vec::with_capacity(binding.outputs().len());
    for output in binding.outputs() {
        let field = schema.field_with_name(&output.output_name).map_err(|_| {
@@ -737,25 +872,61 @@ fn ensure_binding_matches_schema(schema: &ArrowSchema, binding: &FunctionBinding
                binding.binding_id()
            )));
        }
-        let expected_type = parse_output_arrow_type(&output.arrow_type)?;
-        let expected_type = lance_namespace::schema::convert_json_arrow_type(&expected_type)
-            .map_err(|e| invalid_function(format!("invalid Function output type: {e}")))?;
-        if field.data_type() != &expected_type {
+        let (type_matches, has_semantic_blob) = if output.arrow_type == FUNCTION_BLOB_V2_TYPE {
+            (has_supported_blob_v2_layout(field), true)
+        } else {
+            let expected_type = parse_output_arrow_type(&output.arrow_type)?;
+            let expected_type = lance_namespace::schema::convert_json_arrow_type(&expected_type)
+                .map_err(|e| invalid_function(format!("invalid Function output type: {e}")))?;
+            (
+                function_output_type_matches(&expected_type, field.data_type()),
+                function_output_type_has_blob(&expected_type),
+            )
+        };
+        if !type_matches {
            return Err(invalid_function(format!(
                "Function output '{}' type no longer matches binding '{}'",
                output.output_name,
                binding.binding_id()
            )));
        }
-        output_fields.push(ArrowField::new(
-            field.name().clone(),
-            field.data_type().clone(),
-            true,
-        ));
-    }
-    let output_schema =
-        lance_namespace::schema::arrow_schema_to_json(&ArrowSchema::new(output_fields))
+        let metadata = field.metadata();
+        let declared_inputs = metadata
+            .get(INPUTS_META_KEY)
+            .and_then(|raw| serde_json::from_str::<Vec<String>>(raw).ok());
+        if metadata.get(COMPUTED_COLUMN_META_KEY).map(String::as_str) != Some("true")
+            || metadata.get(KIND_META_KEY).map(String::as_str) != Some(FUNCTION_KIND)
+            || metadata
+                .get(FUNCTION_BINDING_ID_META_KEY)
+                .map(String::as_str)
+                != Some(binding.binding_id())
+            || metadata
+                .get(FUNCTION_OUTPUT_ORDINAL_META_KEY)
+                .and_then(|value| value.parse::<u32>().ok())
+                != Some(output.output_ordinal)
+            || declared_inputs.as_deref() != Some(expected_inputs.as_slice())
+        {
+            return Err(invalid_function(format!(
+                "Function output '{}' declaration metadata does not match binding '{}'",
+                output.output_name,
+                binding.binding_id()
+            )));
+        }
+        if has_semantic_blob {
+            output_fields.push(function_output_field(
+                field.name(),
+                true,
+                &output.arrow_type,
+            )?);
+        } else {
+            let json = lance_namespace::schema::arrow_schema_to_json(&ArrowSchema::new(vec![
+                ArrowField::new(field.name().clone(), field.data_type().clone(), true),
+            ]))
            .map_err(|e| invalid_function(format!("invalid Function output schema: {e}")))?;
+            output_fields.push(json.fields.into_iter().next().unwrap());
+        }
+    }
+    let output_schema = JsonArrowSchema::new(output_fields);
    let output_schema = serde_json::to_value(output_schema).map_err(|e| {
        invalid_function(format!(
            "could not encode exact Function output schema: {e}"
@@ -778,7 +949,7 @@ pub(crate) fn plan_function_application(
    application: &FunctionApplication,
    output_name: Option<&str>,
 ) -> Result<FunctionDeclarationPlan> {
-    ensure_no_function_bindings_for_mutation(schema, "Function binding declaration")?;
+    ensure_supported_function_metadata(schema)?;
    if application.has_unknown_fields() {
        return Err(Error::NotSupported {
            message: "Function application contains fields from a newer contract".into(),
@@ -828,8 +999,9 @@ pub(crate) fn plan_function_application(
                input.parameter
            ))
        })?;
-        let field = resolve_field_path(schema, path)?;
-        if field
+        let resolved = resolve_field_path(schema, path)?;
+        if resolved
+            .root
            .metadata()
            .get(COMPUTED_COLUMN_META_KEY)
            .map(String::as_str)
@@ -839,6 +1011,7 @@ pub(crate) fn plan_function_application(
                "Function input '{path}' is computed; computed-on-computed bindings are not supported"
            )));
        }
+        let field = resolved.leaf;
        let parameter_field = ArrowField::new(
            input.parameter.clone(),
            field.data_type().clone(),
@@ -882,16 +1055,15 @@ pub(crate) fn plan_function_application(
                    "Function logical outputs must be non-nullable during NULL assignment",
                ));
            }
-            let data_type =
-                parse_output_arrow_type(output.arrow_type.as_deref().ok_or_else(|| {
-                    invalid_function("scalar Function output is missing its Arrow type")
-                })?)?;
+            let arrow_type = output.arrow_type.as_deref().ok_or_else(|| {
+                invalid_function("scalar Function output is missing its Arrow type")
+            })?;
            outputs.push(FunctionOutputTarget {
                result_field: WHOLE_RESULT_FIELD.to_string(),
                output_name: name.to_string(),
                output_ordinal: 0,
            });
-            output_fields.push(JsonArrowField::new(name.to_string(), true, data_type));
+            output_fields.push(function_output_field(name, true, arrow_type)?);
        }
        "named_struct" => {
            if output.fields.is_empty() {
@@ -935,13 +1107,7 @@ pub(crate) fn plan_function_application(
                let fields = output
                    .fields
                    .iter()
-                    .map(|field| {
-                        Ok(JsonArrowField::new(
-                            field.name.clone(),
-                            false,
-                            parse_output_arrow_type(&field.arrow_type)?,
-                        ))
-                    })
+                    .map(|field| function_output_field(&field.name, false, &field.arrow_type))
                    .collect::<Result<Vec<_>>>()?;
                let mut data_type = JsonArrowDataType::new("struct".to_string());
                data_type.fields = Some(fields);
@@ -968,11 +1134,7 @@ pub(crate) fn plan_function_application(
                        output_name: name.clone(),
                        output_ordinal: ordinal as u32,
                    });
-                    output_fields.push(JsonArrowField::new(
-                        name.clone(),
-                        true,
-                        parse_output_arrow_type(&field.arrow_type)?,
-                    ));
+                    output_fields.push(function_output_field(name, true, &field.arrow_type)?);
                }
            }
        }
@@ -1462,9 +1624,7 @@ fn plan_declarations(schema: SchemaRef, columns: &[(String, String)]) -> Result<

    for (name, expression) in columns {
        if schema.field_with_name(name).is_ok() {
-            return Err(Error::ColumnAlreadyExists {
-                name: name.to_string(),
-            });
+            return Err(Error::ColumnAlreadyExists { name: name.clone() });
        }

        let bound = bind(schema.clone(), name, expression)?;
@@ -1611,7 +1771,7 @@ mod tests {
    }

    use arrow_array::record_batch;
-    use arrow_schema::DataType;
+    use arrow_schema::{DataType, TimeUnit};
    use futures::TryStreamExt;
    use lance::dataset::ColumnAlteration;

@@ -2572,6 +2732,95 @@ mod tests {
        .unwrap()
    }

+    fn blob_application(output: &str) -> FunctionApplication {
+        FunctionApplication::from_json(&format!(
+            r#"{{
+                "function":{{"name":"blob_features","version":"fv_blob"}},
+                "inputs":[
+                    {{"parameter":"image","kind":"column","value":{{"path":"image"}}}}
+                ],
+                "output":{output}
+            }}"#
+        ))
+        .unwrap()
+    }
+
+    fn exact_arrow_type(field: ArrowField) -> String {
+        let json =
+            lance_namespace::schema::arrow_schema_to_json(&ArrowSchema::new(vec![field])).unwrap();
+        serde_json::to_string(json.fields[0].r#type.as_ref()).unwrap()
+    }
+
+    fn single_input_application(path: &str) -> FunctionApplication {
+        FunctionApplication::from_json(
+            &serde_json::json!({
+                "function": {"name": "inspect", "version": "fv_nested_blob"},
+                "inputs": [{
+                    "parameter": "value",
+                    "kind": "column",
+                    "value": {"path": path}
+                }],
+                "output": {"kind": "scalar", "arrow_type": "int64", "nullable": false}
+            })
+            .to_string(),
+        )
+        .unwrap()
+    }
+
+    fn binding_from_plan(plan: &FunctionDeclarationPlan) -> FunctionBinding {
+        let inputs = plan
+            .input_bindings
+            .iter()
+            .enumerate()
+            .map(|(index, input)| {
+                serde_json::json!({
+                    "parameter": input.parameter,
+                    "field_id": index,
+                    "field_path": input.field_path,
+                    "arrow_type": input.arrow_type,
+                    "nullable": input.nullable,
+                })
+            })
+            .collect::<Vec<_>>();
+        let outputs = plan
+            .outputs
+            .iter()
+            .zip(&plan.output_schema.fields)
+            .enumerate()
+            .map(|(index, (output, field))| {
+                serde_json::json!({
+                    "result_field": output.result_field,
+                    "output_name": output.output_name,
+                    "output_field_id": 100 + index,
+                    "output_ordinal": output.output_ordinal,
+                    "arrow_type": canonical_input_arrow_type(field).unwrap(),
+                    "nullable": false,
+                })
+            })
+            .collect::<Vec<_>>();
+        FunctionBinding::from_json(
+            &serde_json::json!({
+                "binding_id": "fb_blob",
+                "function": plan.application.function(),
+                "inputs": inputs,
+                "outputs": outputs,
+                "input_schema": plan.input_schema,
+                "output_schema": plan.output_schema,
+            })
+            .to_string(),
+        )
+        .unwrap()
+    }
+
+    fn full_blob_field(name: &str, nullable: bool) -> ArrowField {
+        ArrowField::new(
+            name,
+            DataType::Struct(lance_core::datatypes::BLOB_V2_LOGICAL_FIELDS.clone()),
+            nullable,
+        )
+        .with_metadata(crate::blob(name, nullable).metadata().clone())
+    }
+
    fn function_binding_schema(title_nullable: bool, body_nullable: bool) -> ArrowSchema {
        ArrowSchema::new(vec![
            ArrowField::new("title", DataType::Utf8, title_nullable),
@@ -2581,6 +2830,37 @@ mod tests {
        ])
    }

+    fn valid_function_binding_schema(
+        title_nullable: bool,
+        body_nullable: bool,
+        binding: &FunctionBinding,
+    ) -> ArrowSchema {
+        let mut fields = function_binding_schema(title_nullable, body_nullable)
+            .fields()
+            .iter()
+            .map(|field| field.as_ref().clone())
+            .collect::<Vec<_>>();
+        let inputs = binding
+            .inputs()
+            .iter()
+            .map(|input| input.field_path.clone())
+            .collect::<Vec<_>>();
+        for output in binding.outputs() {
+            let index = fields
+                .iter()
+                .position(|field| field.name() == &output.output_name)
+                .unwrap();
+            fields[index] = fields[index]
+                .clone()
+                .with_metadata(function_computed_column_metadata(
+                    binding.binding_id(),
+                    output.output_ordinal,
+                    &inputs,
+                ));
+        }
+        ArrowSchema::new(fields)
+    }
+
    #[test]
    fn test_non_nullable_function_inputs_can_bind_to_nullable_parameters() {
        let binding = FunctionBinding::from_json(include_str!(
@@ -2588,7 +2868,11 @@ mod tests {
        ))
        .unwrap();

-        ensure_binding_matches_schema(&function_binding_schema(false, false), &binding).unwrap();
+        ensure_binding_matches_schema(
+            &valid_function_binding_schema(false, false, &binding),
+            &binding,
+        )
+        .unwrap();
    }

    #[test]
@@ -2601,8 +2885,11 @@ mod tests {
        raw_binding["input_schema"]["fields"][0]["nullable"] = Value::Bool(false);
        let binding: FunctionBinding = serde_json::from_value(raw_binding).unwrap();

-        let err = ensure_binding_matches_schema(&function_binding_schema(true, false), &binding)
-            .unwrap_err();
+        let err = ensure_binding_matches_schema(
+            &valid_function_binding_schema(true, false, &binding),
+            &binding,
+        )
+        .unwrap_err();
        assert!(
            matches!(&err, Error::InvalidInput { message }
                if message.contains("input column 'title' is nullable")
@@ -2613,6 +2900,73 @@ mod tests {
        );
    }

+    #[test]
+    fn test_second_binding_rejects_outputs_without_reciprocal_metadata() {
+        let binding = FunctionBinding::from_json(include_str!(
+            "../../tests/fixtures/first_class_functions/v1/remote_function_binding.json"
+        ))
+        .unwrap();
+        let schema = ArrowSchema::new_with_metadata(
+            function_binding_schema(true, true).fields().to_vec(),
+            HashMap::from([(
+                FUNCTION_BINDINGS_META_KEY.to_string(),
+                function_bindings_metadata(std::slice::from_ref(&binding)).unwrap(),
+            )]),
+        );
+
+        let err = plan_function_application(
+            &schema,
+            &named_struct_application(
+                r#"{"normalized_text":"secondary_text","token_count":"secondary_token_count"}"#,
+            ),
+            None,
+        )
+        .unwrap_err();
+        assert!(
+            matches!(&err, Error::InvalidInput { message }
+                if message.contains("declaration metadata")
+                    && message.contains("fb_01K3TEXT")),
+            "{err:?}"
+        );
+    }
+
+    #[test]
+    fn test_persisted_nested_input_keeps_leaf_level_validation() {
+        let mut raw_binding: Value = serde_json::from_str(include_str!(
+            "../../tests/fixtures/first_class_functions/v1/remote_function_binding.json"
+        ))
+        .unwrap();
+        raw_binding["inputs"][0]["field_path"] = Value::String("title.value".to_string());
+        let binding: FunctionBinding = serde_json::from_value(raw_binding).unwrap();
+        let title = ArrowField::new(
+            "title",
+            DataType::Struct(vec![ArrowField::new("value", DataType::Utf8, true)].into()),
+            true,
+        )
+        .with_metadata(HashMap::from([
+            (COMPUTED_COLUMN_META_KEY.to_string(), "true".to_string()),
+            (KIND_META_KEY.to_string(), SQL_KIND.to_string()),
+            (EXPRESSION_META_KEY.to_string(), "title".to_string()),
+        ]));
+        let mut fields = vec![title, ArrowField::new("body", DataType::Utf8, true)];
+        fields.extend(binding.outputs().iter().map(|output| {
+            let data_type = match output.arrow_type.as_str() {
+                "utf8" => DataType::Utf8,
+                "int64" => DataType::Int64,
+                other => panic!("unexpected fixture output type {other}"),
+            };
+            ArrowField::new(&output.output_name, data_type, true).with_metadata(
+                function_computed_column_metadata(
+                    binding.binding_id(),
+                    output.output_ordinal,
+                    &["title.value".into(), "body".into()],
+                ),
+            )
+        }));
+
+        ensure_binding_matches_schema(&ArrowSchema::new(fields), &binding).unwrap();
+    }
+
    #[test]
    fn test_function_binding_metadata_survives_schema_round_trip() {
        let binding = FunctionBinding::from_json(include_str!(
@@ -2661,9 +3015,36 @@ mod tests {
                output_ordinal: 1,
            } if binding_id == "fb_01K3TEXT"
        ));
-        let err = plan_function_application(&reopened, &named_struct_application("{}"), None)
+        let dependent_application = FunctionApplication::from_json(
+            r#"{
+                "function":{"name":"dependent","version":"fv_dependent"},
+                "inputs":[
+                    {"parameter":"text","kind":"column","value":{"path":"search_text"}}
+                ],
+                "output":{"kind":"scalar","arrow_type":"int64","nullable":false}
+            }"#,
+        )
+        .unwrap();
+        let err = plan_function_application(&reopened, &dependent_application, Some("dependent"))
            .unwrap_err();
-        assert!(matches!(err, Error::NotSupported { .. }));
+        assert!(
+            matches!(&err, Error::InvalidInput { message } if message.contains("computed-on-computed"))
+        );
+        let plan = plan_function_application(
+            &reopened,
+            &named_struct_application(
+                r#"{"normalized_text":"secondary_text","token_count":"secondary_token_count"}"#,
+            ),
+            None,
+        )
+        .unwrap();
+        assert_eq!(
+            plan.outputs
+                .iter()
+                .map(|output| output.output_name.as_str())
+                .collect::<Vec<_>>(),
+            ["secondary_text", "secondary_token_count"]
+        );
    }

    #[test]
@@ -2713,6 +3094,273 @@ mod tests {
        );
    }

+    #[test]
+    fn test_blob_function_plans_semantic_input_and_scalar_output() {
+        let schema = ArrowSchema::new(vec![crate::blob("image", false)]);
+        let application =
+            blob_application(r#"{"kind":"scalar","arrow_type":"blob_v2","nullable":false}"#);
+        let plan = plan_function_application(&schema, &application, Some("thumbnail")).unwrap();
+
+        assert_eq!(plan.input_bindings[0].arrow_type, FUNCTION_BLOB_V2_TYPE);
+        let input_schema =
+            lance_namespace::schema::convert_json_arrow_schema(&plan.input_schema).unwrap();
+        assert!(input_schema.field(0).is_blob_v2());
+        let output_schema =
+            lance_namespace::schema::convert_json_arrow_schema(&plan.output_schema).unwrap();
+        assert!(output_schema.field(0).is_blob_v2());
+    }
+
+    #[test]
+    fn test_blob_scalar_binding_accepts_full_logical_layout() {
+        let input = crate::blob("image", false);
+        let application =
+            blob_application(r#"{"kind":"scalar","arrow_type":"blob_v2","nullable":false}"#);
+        let plan = plan_function_application(
+            &ArrowSchema::new(vec![input.clone()]),
+            &application,
+            Some("thumbnail"),
+        )
+        .unwrap();
+        let binding = binding_from_plan(&plan);
+        let mut metadata = full_blob_field("thumbnail", true).metadata().clone();
+        metadata.extend(function_computed_column_metadata(
+            binding.binding_id(),
+            0,
+            &["image".into()],
+        ));
+        let output = full_blob_field("thumbnail", true).with_metadata(metadata);
+
+        ensure_binding_matches_schema(&ArrowSchema::new(vec![input, output]), &binding).unwrap();
+    }
+
+    #[test]
+    fn test_blob_binding_rejects_marker_on_invalid_storage_layout() {
+        let input = crate::blob("image", false);
+        let application =
+            blob_application(r#"{"kind":"scalar","arrow_type":"blob_v2","nullable":false}"#);
+        let plan = plan_function_application(
+            &ArrowSchema::new(vec![input.clone()]),
+            &application,
+            Some("thumbnail"),
+        )
+        .unwrap();
+        let binding = binding_from_plan(&plan);
+        let malformed = ArrowField::new("thumbnail", DataType::Int64, true)
+            .with_metadata(crate::blob("thumbnail", true).metadata().clone());
+
+        ensure_binding_matches_schema(&ArrowSchema::new(vec![input, malformed]), &binding)
+            .unwrap_err();
+    }
+
+    #[test]
+    fn test_blob_input_rejects_marker_on_invalid_storage_layout() {
+        let malformed = ArrowField::new("image", DataType::Int64, false)
+            .with_metadata(crate::blob("image", false).metadata().clone());
+        let application =
+            blob_application(r#"{"kind":"scalar","arrow_type":"blob_v2","nullable":false}"#);
+
+        plan_function_application(
+            &ArrowSchema::new(vec![malformed]),
+            &application,
+            Some("thumbnail"),
+        )
+        .unwrap_err();
+    }
+
+    #[test]
+    fn test_non_blob_input_does_not_require_json_round_trip() {
+        let json = lance_namespace::schema::arrow_schema_to_json(&ArrowSchema::new(vec![
+            ArrowField::new("event_time", DataType::Time64(TimeUnit::Microsecond), false),
+        ]))
+        .unwrap();
+
+        assert_eq!(
+            canonical_input_arrow_type(&json.fields[0]).unwrap(),
+            "time64"
+        );
+    }
+
+    #[test]
+    fn test_blob_named_struct_plans_expanded_and_whole_outputs() {
+        let schema = ArrowSchema::new(vec![crate::blob("image", false)]);
+        let application = blob_application(
+            r#"{"kind":"named_struct","fields":[
+                {"name":"thumbnail","arrow_type":"blob_v2","nullable":false},
+                {"name":"width","arrow_type":"int32","nullable":false}
+            ]}"#,
+        );
+
+        let expanded = plan_function_application(&schema, &application, None).unwrap();
+        let expanded_schema =
+            lance_namespace::schema::convert_json_arrow_schema(&expanded.output_schema).unwrap();
+        assert!(expanded_schema.field(0).is_blob_v2());
+        assert_eq!(expanded_schema.field(1).data_type(), &DataType::Int32);
+
+        let whole = plan_function_application(&schema, &application, Some("analysis")).unwrap();
+        let whole_schema =
+            lance_namespace::schema::convert_json_arrow_schema(&whole.output_schema).unwrap();
+        let DataType::Struct(fields) = whole_schema.field(0).data_type() else {
+            panic!("whole Function output should be a struct");
+        };
+        assert!(fields[0].is_blob_v2());
+        assert_eq!(fields[1].data_type(), &DataType::Int32);
+    }
+
+    #[test]
+    fn test_struct_blob_input_preserves_exact_schema_and_nullability() {
+        let payload = ArrowField::new(
+            "payload",
+            DataType::Struct(Fields::from(vec![
+                ArrowField::new("mime_type", DataType::Utf8, false),
+                ArrowField::new(
+                    "nested",
+                    DataType::Struct(Fields::from(vec![crate::blob("image", true)])),
+                    true,
+                ),
+            ])),
+            true,
+        );
+        let plan = plan_function_application(
+            &ArrowSchema::new(vec![payload]),
+            &single_input_application("payload"),
+            Some("size"),
+        )
+        .unwrap();
+
+        let declared: JsonArrowDataType =
+            serde_json::from_str(&plan.input_bindings[0].arrow_type).unwrap();
+        let DataType::Struct(fields) =
+            lance_namespace::schema::convert_json_arrow_type(&declared).unwrap()
+        else {
+            panic!("expected a struct Function input")
+        };
+        assert!(fields[1].is_nullable());
+        let DataType::Struct(nested) = fields[1].data_type() else {
+            panic!("expected a recursive struct Function input")
+        };
+        assert!(nested[0].is_blob_v2());
+        assert!(nested[0].is_nullable());
+
+        let exact = lance_namespace::schema::convert_json_arrow_schema(&plan.input_schema).unwrap();
+        let DataType::Struct(fields) = exact.field(0).data_type() else {
+            panic!("expected exact input schema to retain the struct")
+        };
+        let DataType::Struct(nested) = fields[1].data_type() else {
+            panic!("expected exact input schema to retain the nested struct")
+        };
+        assert!(nested[0].is_blob_v2());
+    }
+
+    #[test]
+    fn test_recursive_blob_result_plans_one_whole_named_struct_column() {
+        let details_type = exact_arrow_type(ArrowField::new(
+            "details",
+            DataType::Struct(Fields::from(vec![crate::blob("image", true)])),
+            false,
+        ));
+        let application = FunctionApplication::from_json(
+            &serde_json::json!({
+                "function": {"name": "inspect", "version": "fv_nested_blob"},
+                "inputs": [],
+                "output": {
+                    "kind": "named_struct",
+                    "fields": [
+                        {"name": "mime_type", "arrow_type": "utf8", "nullable": false},
+                        {"name": "details", "arrow_type": details_type, "nullable": false}
+                    ]
+                }
+            })
+            .to_string(),
+        )
+        .unwrap();
+        let plan = plan_function_application(&ArrowSchema::empty(), &application, Some("payload"))
+            .unwrap();
+
+        assert_eq!(plan.outputs.len(), 1);
+        assert_eq!(plan.outputs[0].result_field, WHOLE_RESULT_FIELD);
+        let schema =
+            lance_namespace::schema::convert_json_arrow_schema(&plan.output_schema).unwrap();
+        assert_eq!(schema.field(0).name(), "payload");
+        let DataType::Struct(fields) = schema.field(0).data_type() else {
+            panic!("whole named result must be one struct column")
+        };
+        assert_eq!(
+            fields.iter().map(|field| field.name()).collect::<Vec<_>>(),
+            ["mime_type", "details"]
+        );
+        let DataType::Struct(details) = fields[1].data_type() else {
+            panic!("expected recursive result struct")
+        };
+        assert!(details[0].is_blob_v2());
+        assert!(!fields.iter().any(|field| field.name() == "payload"));
+    }
+
+    #[test]
+    fn test_blob_children_under_collections_are_rejected() {
+        let collections = vec![
+            DataType::List(Arc::new(crate::blob("item", false))),
+            DataType::LargeList(Arc::new(crate::blob("item", false))),
+            DataType::FixedSizeList(Arc::new(crate::blob("item", false)), 2),
+            DataType::Map(
+                Arc::new(ArrowField::new(
+                    "entries",
+                    DataType::Struct(Fields::from(vec![
+                        ArrowField::new("key", DataType::Utf8, false),
+                        crate::blob("value", false),
+                    ])),
+                    false,
+                )),
+                false,
+            ),
+        ];
+        for data_type in collections {
+            let schema = ArrowSchema::new(vec![ArrowField::new("value", data_type, false)]);
+            let error = plan_function_application(
+                &schema,
+                &single_input_application("value"),
+                Some("size"),
+            )
+            .unwrap_err();
+            assert!(
+                error.to_string().contains("under a collection"),
+                "got: {error}"
+            );
+        }
+    }
+
+    #[test]
+    fn test_blob_whole_struct_binding_accepts_full_logical_layout() {
+        let input = crate::blob("image", false);
+        let application = blob_application(
+            r#"{"kind":"named_struct","fields":[
+                {"name":"thumbnail","arrow_type":"blob_v2","nullable":false},
+                {"name":"width","arrow_type":"int32","nullable":false}
+            ]}"#,
+        );
+        let plan = plan_function_application(
+            &ArrowSchema::new(vec![input.clone()]),
+            &application,
+            Some("analysis"),
+        )
+        .unwrap();
+        let binding = binding_from_plan(&plan);
+
+        let output = ArrowField::new(
+            "analysis",
+            DataType::Struct(Fields::from(vec![
+                full_blob_field("thumbnail", false),
+                ArrowField::new("width", DataType::Int32, false),
+            ])),
+            true,
+        )
+        .with_metadata(function_computed_column_metadata(
+            binding.binding_id(),
+            0,
+            &["image".into()],
+        ));
+        ensure_binding_matches_schema(&ArrowSchema::new(vec![input, output]), &binding).unwrap();
+    }
+
    #[test]
    fn test_function_mapping_and_sibling_collisions_fail_before_request() {
        let unknown = named_struct_application(r#"{"missing":"renamed"}"#);
@@ -2819,5 +3467,38 @@ mod tests {
        assert!(
            matches!(&err, Error::InvalidInput { message } if message.contains("computed-on-computed"))
        );
+
+        let nested_title = ArrowField::new(
+            "title",
+            DataType::Struct(vec![ArrowField::new("value", DataType::Utf8, true)].into()),
+            true,
+        )
+        .with_metadata(HashMap::from([
+            (COMPUTED_COLUMN_META_KEY.to_string(), "true".to_string()),
+            (KIND_META_KEY.to_string(), SQL_KIND.to_string()),
+            (
+                EXPRESSION_META_KEY.to_string(),
+                "struct('value')".to_string(),
+            ),
+        ]));
+        let nested_schema = ArrowSchema::new(vec![nested_title, schema.field(1).as_ref().clone()]);
+        let nested_application = FunctionApplication::from_json(
+            r#"{
+                "function":{"name":"text_features","version":"fv_exact"},
+                "inputs":[
+                    {"parameter":"title","kind":"column","value":{"path":"title.value"}},
+                    {"parameter":"body","kind":"column","value":{"path":"body"}}
+                ],
+                "output":{"kind":"named_struct","fields":[
+                    {"name":"normalized_text","arrow_type":"utf8","nullable":false},
+                    {"name":"token_count","arrow_type":"int64","nullable":false}
+                ]}
+            }"#,
+        )
+        .unwrap();
+        let err = plan_function_application(&nested_schema, &nested_application, None).unwrap_err();
+        assert!(
+            matches!(&err, Error::InvalidInput { message } if message.contains("computed-on-computed"))
+        );
    }
 }
@@ -45,7 +45,11 @@ async fn local_function_catalog_operations_return_stable_not_supported() {
        .get_function("normalize_score", "fv_exact")
        .await
        .unwrap_err();
-    for error in [create_error, lookup_error] {
+    let drop_error = connection
+        .drop_function("normalize_score", "fv_exact")
+        .await
+        .unwrap_err();
+    for error in [create_error, lookup_error, drop_error] {
        assert!(matches!(
            error,
            Error::NotSupported { message }
@@ -78,6 +78,12 @@
        "type": "utf8"
      }
    },
+    {
+      "arrow_type": "large_utf8",
+      "json": {
+        "type": "large_utf8"
+      }
+    },
    {
      "arrow_type": "binary",
      "json": {
@@ -171,6 +177,21 @@
        ]
      }
    },
+    {
+      "arrow_type": "list<large_utf8>",
+      "json": {
+        "type": "list",
+        "fields": [
+          {
+            "name": "item",
+            "nullable": false,
+            "type": {
+              "type": "large_utf8"
+            }
+          }
+        ]
+      }
+    },
    {
      "arrow_type": "large_list<utf8>",
      "json": {
@@ -186,6 +207,21 @@
        ]
      }
    },
+    {
+      "arrow_type": "large_list<large_utf8>",
+      "json": {
+        "type": "large_list",
+        "fields": [
+          {
+            "name": "item",
+            "nullable": false,
+            "type": {
+              "type": "large_utf8"
+            }
+          }
+        ]
+      }
+    },
    {
      "arrow_type": "fixed_size_list<float32, 384>",
      "json": {
@@ -330,4 +366,4 @@
    "timestamp[us]",
    "struct<a: int32>"
  ]
-}
+}