Skip to content

Release 1.2.1

Latest
Latest

Choose a tag to compare

View all tags
@KarthikSubbarao KarthikSubbarao released this 07 Jul 23:14
1.2.1
788e8c9
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

A patch release on top of 1.2.0 focused on stability, correctness, and efficiency. It hardens command and load paths against crashes found through fuzzing, fixes concurrency and shutdown issues, tightens cluster and query correctness, and lands vector, tag, numeric, and text index optimizations.
Like the 1.2 release, this version requires Valkey core 9.0.1 release or higher.

Bug fixes

  • Hardened load and command paths against crashes found through fuzzing: added validation for HNSW index data loaded from an external file and for query vector size against index dimensions at parse time (#1191, #1192)
  • Fixed FT.INTERNAL_UPDATE crashes during AOF and RDB load that could leave a server in a restart loop (#1146)
  • Fixed a FLAT index ingestion livelock when BLOCK_SIZE is 0, and bounded BLOCK_SIZE and INITIAL_CAP so a single command cannot OOM-kill the server (#1122)
  • Fixed several crashes on malformed or edge-case commands: FT.SEARCH with an empty right-hand side of an OR, empty FT.AGGREGATE expressions, REDUCE COUNT_DISTINCT, loading text values from JSON in FT.AGGREGATE, and FT.DROPINDEX while queries are in flight (#1115, #966, #1032, #999, #956)
  • Fixed a race condition in concurrent HSET operations against an HNSW vector index (#1079)
  • Fixed a replica crash in RemoveSubscription when an index schema fails to load partway through RDB load (#1187)
  • Fixed HNSW replacement of deleted elements, including allow_replace_deleted labeling and ensuring it is set before index load (#926, #1019)
  • Enabled thread-safety checking and fixed the violations it surfaced (#1082)
  • Reworked index schema destruction to run asynchronously and restored background result cleanup (#1202, #1203)
  • Fixed shutdown leaks and timing holes: ThreadPool shutdown timing, RunByMain draining pending one-shots, in-flight async queries drained on cancel and timeout, and a heap-use-after-free in PausePoint during multi-database shutdown (#1062, #1029, #1087, #958)
  • Fixed a race in untracked-keys ingestion access versus negative queries, and guarded against a disengaged optional in tracked-mutation handling (#1060, #1120)
  • In cluster mode, treat an empty CLUSTER SLOTS response as incomplete, and correctly resolve slots in the middle or end of a range during fanout (#1181, #1117)
  • Fixed the query vector not being normalized on the prefiltering execution path for cosine-distance vector indexes (#1222)
  • Tag queries now handle an escaped closing brace correctly (#1148)
  • JSON string values containing backslash escapes are now decoded before indexing rather than stored raw (#1174)
  • Numeric range searches now accept values in scientific notation (#937)
  • Field serialization now uses a fixed ordering, which matters for text indexes (#1003)
  • Corrected memory and INFO reporting: accurate memory reporting when reader/writer thread counts change at runtime, a reclaimable_memory underflow, and coordinator bytes-in counting before response mutation (#1119, #1131, #1046)
  • search_number_of_indexes is now updated correctly during restore, and aggregate recursion depth is handled correctly (#974, #973)
  • Included a set of thread-sanitizer (TSAN) fixes (#860)
  • Improvements to tests and test infrastructure with no corresponding module code changes.

Optimizations

  • Vector search performance improvement (#1163)
  • Space savings for tag and numeric indexes (#1026)
  • Reduced reference counting in text iterators with short-circuited field-containment checks, backed by borrowed interned-string pointers (#1044, #1043)
  • Optimized content fetch to eliminate a double key open and use HashGet for specific field access (#1045)
  • Removed a reader-thread scaling bottleneck caused by the PausePoint mutex (#1037)
  • Used a protobuf arena for fanout response deserialization (#1040)
  • Optimized tag search to avoid relying on the root iterator in common cases (#994)
  • Replaced a deque with a flat hash map for tracked_vectors (#1015)

Contributors

@Aksha1812, @allenss-amazon, @KarthikSubbarao, @boda26, @baswanth09, @bandalgomsu, @Frank-Gu-81, @hjkwon0, @smkher, @BCathcart, @Nivesh-01, @daddaman-amz, @yairgott, @linbran5123

Contributors

smkher, yairgott, and 12 other contributors
Assets 2
Loading