v1.38.2
📦 weaviateView on GitHub →
✨ 7 features🐛 26 fixes🔧 4 symbols
Summary
This release introduces the generative-deepseek module and enhances stability and performance across replication, indexing (HNSW, BM25), and backup operations. Several security fixes address SSRF vulnerabilities and dependency updates were performed.
Migration Steps
- Validate X-*-BaseURL request headers if you were relying on potentially insecure header forwarding.
✨ New Features
- Add generative-deepseek module.
- Add location configuration to text2vec-google module.
- Add support for endpoint setting in OpenAI client.
- Add support for location setting in generative-google module.
- Add support for dimensions setting in text2vec-aws module.
- Increase searchProbe default to 256 in hfresh.
- Add support for stop setting in module settings for generative-deepseek.
🐛 Bug Fixes
- Optimize async replication to use efficient property unmarshalling.
- Enforce tenant cap in RAFT.
- Make initTargetVector idempotent to prevent double-create.
- Serialize config updates and Iterate against the flat->hnsw upgrade.
- Read compression config under compressActionLock in insert validation for hnsw.
- Record max-impact pair on memtable BlockEntry for bm25.
- Tenant cap rejects ingest into existing tenants.
- Fix panic in db.WaitForLocalInflightWrites.
- Add missing forwarding of property names.
- Validate X-*-BaseURL request headers to close SSRF bypass in modules.
- Route multivector queue deletes through DeleteMulti.
- Defer queue registration until restore completes in hfresh.
- Remove flaky assertion for dynamic index.
- Return 5xx for operational failures in backup.
- Skip malformed runtime config values.
- Fix namespace test that is flaky due to EC.
- Block ShardReplicationFSM.AllPeersAtLeast on visible peers.
- Print weaviate testcontainer logs on namespace-test error.
- Fix reassign queue recovery hints in hfresh.
- Stop dropping nodes at snapshot block boundaries in hnsw.
- Fix backup RBAC checks.
- Snapshot in-place-mutated bbolt files during active-shard backup.
- Forbid certain schema mutations during replica movements.
- Don't abort the whole usage report in case of missing tenant.
- Reuse a single cursor across the hashbeat digest scan in async-rep.
- Parallel-cursor prefill for unbounded uncompressed vector cache in hnsw.