v1.37.12

🐛 Bug Fixes

• Fixed HNSW maintenance flag leak.
• Implemented cgroup v2 support by using automaxprocs to set GOMAXPROCS.
• Fixed HNSW search entrypoint prerequisites (though this was later reverted).
• Returned HTTP 500 instead of nil Responder on getGroups controller error.
• Fixed HNSW repair global entrypoint (though this was later reverted).
• Retried Go modules downloads in CI.
• Increased hfresh searchProbe default to 256.
• Fixed metric name typo.
• Improved performance by sharing one zstd encoder in REST client instead of using sync.Pool.
• Improved performance by allocating propagation buffers per-hashBeat, not per-range in async replication.
• Optimized cycle manager for large MT collections.
• Made halt-for-transfer inactivity monitor race-free.
• Added additional validation on PQ query vector length.
• Fixed collection issue in incremental backups.
• Fixed quantizer initialization race in hfresh.
• Fixed RBAC class-getter memo key mismatch in gRPC service.
• Resolved class lookup to happen only once on the gRPC layer.
• Narrowed traverser single-class reads to use a class getter.
• Implemented recovery from corrupt chunk headers in the queue instead of failing to load.
• Added percentage of eager shards loaded to raft waiting for database to restore log.
• Added "weaviate_" prefix to a missed metric.
• Split go-client-group-2 CI tests into group-2 (usage) and group-3.
• Deduplicated duplicate properties in BM25 (last boost wins).
• Propagated combineResults error in BM25 (previously resulted in silent empty results).
• Replaced per-shard jitter with bounded concurrent shard readers for performance.
• Fixed passing k1/b in the correct order when flushing LSMKV block-max impacts.
• Improved performance in async replication by using batched hashtree-root pre-filter for many-tenant clusters.
• Hoisted hashtree-init wait out of RLock in async replication merge paths.
• Prevented TimeBased repair from clobbering a newer local write in async replication.
• Avoided nil-pointer dereference in read-repair when a live winner's refetch fails.
• Salvaged and quarantined torn sealed chunks in the queue.
• Reused consistent view for fetches of multiple objects.
• Validated queue chunk record count and length against file size.
• Implemented back off on recoverable task errors and contained task/dispatch panics in the queue.
• Improved BM25 performance by approximating object count for IDF and reusing query stats on block-max fallback.
• Serialized hashtree rebuild against shard shutdown in async replication.
• Deregistered under the write lock and kept replication on while overrides are active in async replication.
• Closed hashtree-init channel on early goroutine exit to unblock parked merges in async replication.
• Applied propertyBoost to WAND pivot and block-impact bounds in BM25.
• Repaired REST root-prefilter JSON and bounded request size in async replication.
• Bounded halt-for-transfer backup preparation.
• Fixed geo queue backup preparation timeout.
• Hardened async-replication and replica read/write paths.
• Guarded nil logger in GetConsistentView slow-lock warning for LSMKV.