## Summary
`@lancedb/lancedb` pins the optional dependency
`@huggingface/transformers` at 3.0.2, which requires `sharp ^0.33.5`.
Every npm install of LanceDB therefore pulled in a sharp version with
critical advisories (GHSA-wq5f-xc86-pv6w, GHSA-rgj7-g3m4-5g8c). No 3.x
release of transformers allows a patched sharp, so this upgrades to
4.3.2, which requires `sharp ^0.35.4`. A fresh install resolves that
range to 0.35.5, the first release that fixes both advisories.
- `nodejs/package.json`: `@huggingface/transformers` 3.0.2 → 4.3.2.
- `nodejs/pnpm-workspace.yaml`: remove the `sharp: ^0.35.4` override. It
only covered our own install, not consumers, and transformers now
requires that range itself.
- `nodejs/examples`: transformers 3.0.2 → 4.3.2, sharp 0.33.5 → 0.35.5.
- Both lockfiles resolve sharp to 0.35.5; `pnpm audit` reports no sharp
advisories in either.
`TransformersEmbeddingFunction` needed no code changes. The transformers
types are only used in private fields, so the published typings don't
change.
`docs/web-tests/ts` and `docs/web-tests/blob-v2/ts` are left alone: they
resolve transformers through a published `@lancedb/lancedb`, so they
pick this up after a release.
## Evidence
Embeddings from the built-in `huggingface` embedding function (default
`Xenova/all-MiniLM-L6-v2`) on 4.3.2, compared with the same mean pooling
on 3.0.2, for three sentences:
```
transformers 4.3.2: 3 x 384
text 0: max|diff|=6.62e-07 cosine=1.00000000
text 1: max|diff|=6.26e-07 cosine=1.00000000
text 2: max|diff|=7.33e-07 cosine=1.00000000
```
End to end through a table, `search("a feline sat down")` returns `["The
cat sat on the mat.", "A dog barked at the mailman."]`.
## Merge Risk
**Door:** Two-way. A revert restores the 3.0.2 pin.
**Blast Radius:** Node `huggingface` embedding function, and install
footprint of the optional dependency (onnxruntime-node 1.19.2 → 1.30.0,
which still downloads the CUDA provider on Linux x64 by default, as
1.19.2 did).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
## Summary
- inspect the assembled x64 musl addon for the unresolved AVX-512
distance-table symbol before npm publication
- fail the publish lifecycle with an actionable error if the broken
symbol is present
- add regression coverage using the dynamic-symbol output from the
affected 0.27.x artifact
## Root cause
LanceDB 0.27.x could enable the Rust AVX-512 distance-table callsite
when the separate C kernel build did not emit its object, leaving
`sum_4bit_dist_table_32bytes_batch_avx512` undefined in the musl addon.
The current Lance dependency includes the upstream split-kernel cfg fix,
and the published 0.33.0 musl artifact no longer has the undefined
symbol, but the LanceDB publish path did not verify that property before
releasing an addon.
This adds a release guard around the corrected implementation so the
same invalid binary cannot be published again.
## Validation
- `pnpm lint`
- `pnpm run build`
- `pnpm test --runInBand` (680 passed, 5 skipped)
- `pnpm run docs`
- checked published artifacts: 0.27.1 is rejected for the unresolved
symbol; 0.33.0 is accepted
Fixes#3200
<!-- lance-gatekeeper-fix:v1 agent=4094b9b6de4b43665a9ff50fa35322fb
generation=1 -->
Co-authored-by: Gatefixer <313497061+lancedb-gatefixer[bot]@users.noreply.github.com>
Thanks @vincentkoc for starting this! I was just looking into this
myself. I started with your PR and added:
- make arrow 21 the default
- widen the peer range to <22 so if they release 21.3 that'll be
included
- have to do skipLibCheck in tsconfig.json, because of flatbuffers 25
which needs TypeScript 5.7
- fix override in Symbol.hasInstance
- support new types in Arrow 20-21 (LargeList, Utf8View, BinaryView,
IntervalMonthDayNano)
- arrow-compat job in nodejs.yml that runs tests with arrow 19 and 20
(adds 2 sec)
- put the list of arrow types in one place
- update NODEJS_THIRD_PARTY_LICENSES.md
- regenerated typescript docs (sorry about reorderings, one time cost)
- reformat arrow.test.ts (sorry, almost all whitespace, thanks to biome.
git diff -w shows 110 lines edited)
- fixed Sliced Utf8View / BinaryView serialization error, by just
converting them to Utf8 and Binary (as our rust code would anyway)
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Lance Release
Will JonesandClaude Opus 5.5
lancedb-gatefixer[bot]
