Weaviate patches a bug that silently dropped vectors from search

Weaviate's HNSW compaction bug could isolate index nodes without throwing an error, silently dropping vectors from search, per its 1.39.4 release notes. The notes list "New Features: none" for the September 11 release. An isolated vector stays on disk but never surfaces in a query result, and nothing in the system flags the gap. That's the worst failure mode a vector database can have. Data disappears without an error, and the gap shows up only when a search that should return a result comes back empty.

The same release recovers orphaned postings through a reassign-all pass and bounds-checks read repair on replicas so it can't run past the data it's correcting. It also adds a barrier that makes a newly elected Raft leader sync its state machine before it judges new proposals. A separate fix creates queue chunk files with the O_EXCL flag, closing a race that let two same-microsecond chunks overwrite each other.

The one item Weaviate labels a feature is node-level query admission control. It lets operators cap how many concurrent queries a single node accepts before it rejects or queues new ones. Paired with the reliability fixes, the release reads as built for production clusters under steady load, not new capability.

The notes don't say how long the compaction bug existed or which versions are affected. Anyone running HNSW indexes under Weaviate has no way to know whether they already lost data, only that new losses should stop with this patch.