Skip to content

v1.19.2

Choose a tag to compare

@timvisee timvisee released this 05 Oct 14:16
· 308 commits to master since this release
v1.19.2
016542a

Change log

Improvements

  • #10682 - Don't rewrite full sparse posting list on every upsert, making sparse vector upserts significantly faster
  • #10349, #10445, #10661, #10700 - Make proxy segments persist their state, prevents WAL pinning and long recovery times on restart
  • #10686 - Fix HNSW builds not always saturating all assigned threads
  • #10624 - Resolve filter IDs only once per search request, improving performance
  • #10853 - Allow to bind internal gRPC to separate network interface
  • #10852 - Add startup warning if no API keys are configured
  • #10891 - Don't hold collections lock during Raft snapshot, which previously blocked access to all collections
  • #10570 - Double time for consensus synchronization await to allow time for re-election
  • #10617 - Encode queue proxy batches in blocking thread, prevent blocking Tokio runtimes
  • #10480 - Prevent blocking updates while synchronizing segment manifest
  • #10541 - Prevent blocking updates while planning optimizations
  • #10610 - Move BM25 sparse embedding to blocking thread to prevent blocking runtimes
  • #10599 - Move queue proxy WAL reads and batch building off async runtime
  • #10794 - Bound cardinality estimates by number of points
  • #10618 - Expose resharding state in telemetry
  • #10458 - Make reporting memory statistics more efficient with parallel workers
  • #10813 - Report transferred bytes and rate in snapshot transfers comment
  • #10781 - Report Acorn search statistics in telemetry
  • #7714 - Add number of failed transfers in metrics (collection_shard_transfer_failed)
  • #10542 - Add iteration count in 'Trying to read-lock a segment' log line
  • #10714 - Support GCS and Azure Blob Storage for collection snapshots
  • #10715 - Add configurable prefix for object storage snapshots

Bug Fixes

  • #10903 - Fix zero search results when HNSW entry point does not satisfy filter
  • #10870 - Fix adding a named multivector to indexed collection breaking search
  • #10577 - Don't claim unfinished operations when flushing, fixing data inconsistencies
  • #10507, #10531 - Fix proxy segments not propagating some named vector or payload index changes
  • #10837 - Prevent write tearing in all Gridstore storages, which could previously corrupt storage
  • #10812 - Fix mutable payload index not able to load on torn write, causing crash loop
  • #10643 - Fix stale snapshot transfer on recovering peer breaking other peers
  • #10897 - Fix group_by on paths with an array index
  • #10567 - Fix MMR pagination with offsets
  • #10878 - Fix MMR candidates not ordered by vector distance before merge, improving accuracy
  • #10593 - Fix TextAny match on unindexed payload fields
  • #10777 - Fix 2-hop Acorn not being capped
  • #10855 - Fix alt_api_key not being used in internal gRPC API calls
  • #10862 - Reject full-text filters with min_token_len > max_token_len
  • #10277 - Don't re-apply committed entries already applied during consensus start

Security

  • #10758 - Accept only read-write API keys on the internal gRPC API
  • #10556 - Don't expose path in logs when uploading snapshot without collection configuration