Version 32.1.3
Released: November 20th 2026
This release adds direct-to-disk bulk import for real-time, sharded, and replicated tables, including concurrent shared loads; improves KNN, hybrid, and full-text search performance; expands auto-embedding and diagnostics capabilities; and fixes issues with replication, query correctness, data integrity, packaging, and shutdown. See the cluster upgrade notes below before upgrading from an older version.
❤️ We'd like to thank @tudorvasinca for PR #4858, PR #4924, and PR #4925; @vitlav for PR #5060; @RaphaelFakhri for PR #5031; and @drahy333 for Columnar PR #217.
Upgrading from 29.9.0
Not using replication? Skip these steps. No special migration is needed for standalone instances.
The replication protocol has changed, so rolling upgrades are not supported:
- Check that all nodes are synchronized, stop writes, and back up your data.
- Stop all nodes one at a time, waiting for each to shut down cleanly. Remember which one stopped last.
- Upgrade all nodes to 32.1.3 while they are stopped. Prevent automatic restarts during the upgrade.
- Start the last stopped node first with
manticore_new_cluster(systemd) orsearchd --new-cluster, then start the others normally. Wait for all nodes to synchronize before resuming writes.
If the cluster shut down unexpectedly, use the cluster recovery guide instead.
Breaking Changes
- ⚠️ v32.0.0 PR #5101 Added bulk import for replicated tables and changed how replicated table files are tracked. This is incompatible with older versions: upgrade every cluster node together, not one by one. See Upgrading from 29.9.0.
- ⚠️ v31.0.0 Issue #5016 In real-time mode, Manticore stores its table and cluster list in
manticore.jsonunder data_dir (usually/var/lib/manticore/manticore.json). A power loss could previously leave it empty. Manticore now saves it more safely and refuses to start with an empty file rather than silently losing the table list. Most users need no action. If startup reportsfile is empty, restore a valid backup; do not delete the file. - ⚠️ v30.0.0 PR #5015 Cluster joins now check that all nodes can reach the joining node before transferring data. The replication change is incompatible with older versions; see Upgrading from 29.9.0.
New Features and Improvements
- 🆕 v32.1.0 Commit Bulk import now passes the table's configured
indexerand common options to the bulk indexer and builds chunks with the table's external files, making imported data consistent with the table configuration. - 🆕 v32.0.0 PR #5101 Extended direct-to-disk bulk import to sharded tables over HTTP and SQL, including replicated shards. The coordinator routes documents to their shards, runs indexers on shard nodes, and publishes each shard's disk chunk on commit;
PURGE BULK_IMPORTcan clean up unfinished shared loads. On replicated tables, chunks are transferred to replicas and attached through replication, with SST recovery for nodes that missed a file change. - 🆕 v31.0.24 PR #5013 Issue #5110 Improved full-text search with dictionary-based
COUNT(*)for eligible single-term queries and rank-free counting for more complex queries. Added 4,096-row bitmap windows for eligibleOR/ANDcounts and unranked queries, SIMD bitmap operations, faster postings readers, redundant-filter elimination, andGROUP BYblob prefetch.SHOW METAidentifiesFastCount,RanklessCount,WindowOrCount, andWindowAndCountexecution paths. - 🆕 v31.0.24 PR #5013 Issue #5110 Changed the default
access_doclistsandaccess_hitliststommapinstead of buffered file reads and added a configurableskiplist_cache_sizeto reduce repeated skiplist decompression. - 🆕 v31.0.23 PR #5013 Issue #5110 Tagged an intermediate version of the full-text counting and postings optimizations on their development branch; the completed changes were integrated into
mainin v31.0.24. - 🆕 v31.0.20 PR #5092 Updated Executor to 1.4.8 with a newer LLM extension revision.
- 🆕 v31.0.17 PR #5074 Updated Executor to 1.4.7, including an updated LLM dependency.
- 🆕 v31.0.2 PR #5021 Updated Executor to 1.4.6 (including PHP 8.5.12 security fixes) and Buddy to 4.4.9, which avoids PHP 8.5 deprecation errors in fuzzy search and autocomplete when a character has no keyboard-layout mapping.
- 🆕 v29.18.6 PR #4987 Made crash dumps identify the current operation and table, including background optimizations, RT flushes, merges, rotation, replication-related work, and SQL/HTTP requests, rather than sometimes reporting a previous query.
- 🆕 v29.18.0 PR #4994 Extended
knn_rescore_prefetchto row-wise vector storage, reading candidate row headers and full-precision vectors in parallel when data is not already cached. - 🆕 v29.17.0 PR #5005 Added automatic document-ID generation to
bulk_import, so inputs need not supply every ID. - 🆕 v29.16.0 PR #5000 Added highlighting for auto-chunked embedding text: a KNN result can highlight the chunk that matched the query rather than only the original full text.
- 🆕 v29.15.0 PR #4999 Introduced shared HTTP bulk import: concurrent
/bulk?bulk_import=<table>&shared[=N]requests feed one table-wide indexer pool and publish up toNdisk chunks on commit. Added session-aware cleanup throughPURGE BULK_IMPORTand support for bare document lines. - 🆕 v29.15.0 PR #4996 Added
indextool --docstore-bad-idsto list live document IDs affected by damaged stored-field blocks so they can be reinserted;indextool --checknow detects missing or zero-filled docstore blocks. - 🆕 v29.14.6 manticore-load PR #29 Updated manticore-load with opt-in
--realistictext generation for benchmarking: frequency-weighted English vocabulary, natural document-length variation, repeated terms, and<text/common/...>,<text/medium/...>, and<text/rare/...>query terms. The previous synthetic generator remains the default. - 🆕 v29.14.2 Issue #4929 Made HNSW graph construction during
FLUSH RAMCHUNKparallel, splitting RAM chunks into balanced row ranges soknn_parallel_buildcan improve flush speed. - 🆕 v29.14.0 PR #4969 Optimized KNN search with a per-table rescoring budget, prefetch for columnar full-precision vectors, and adaptive HNSW
early_terminationon agents.knn_rescore_prefetchreduces cold-cache vector reads without changing matches. - 🆕 v29.13.1 Commit Improved
CREATE TABLEvalidation and errors for unknown, omitted, or unsupported field types, including document-ID and UUID columns. - 🆕 v29.13.0 PR #4932 Enabled facets and JSON aggregations on hybrid-search results after reciprocal-rank fusion, with bucket counts reflecting the final fused candidate set and post-fusion filters.
- 🆕 v29.12.0 Issue #4824 Added direct-to-disk bulk import for local real-time tables through SQL and HTTP, building disk chunks without staging the documents in a RAM chunk. Added commit/rollback handling, duplicate-ID processing, cleanup, and
PURGE BULK_IMPORT. - 🆕 v29.11.2 Buddy PR #701 Updated Buddy to 4.4.5, which validates conversational-search routing responses within the retry budget and recovers from invalid routes or malformed tool calls where possible.
- 🆕 v29.11.0 PR #4919 Extended
FREEZEsupport to plain tables. - 🆕 v29.10.0 MCL PR #208 Added auto-embedding support for ONNX models that have no token-type input, including
BAAI/bge-m3. - 🆕 v29.9.3 Buddy PR #699 Updated Buddy to 4.4.4 with coverage for
CHAT MODELoptionsapi_key,base_url, andtimeoutthrough the LLM PHP extension.
Bug Fixes
- 🪲 v32.1.3 PR #5121 Updated Executor to 1.4.9, bringing PHP 8.5.13 security fixes.
- 🪲 v32.1.2 Issue #5073 Updated Buddy to 4.4.11 to use the configured
node_addresswhen building sharding node IDs, fixing replicated sharded-table creation timing out in Docker bridge networks. - 🪲 v32.1.1 Issue #5107 Fixed Debian packages installing internal
.bundle_*dependency metadata files into the filesystem root; the files remain available at build time but are excluded from the package payload. - 🪲 v32.1.0 Commit Fixed
ATTACH TABLElosing the table's external exceptions-file metadata, which also affected subsequent bulk-indexed chunks. - 🪲 v32.0.5 PR #5109 Fixed percolate-query
INSERThanging when a concurrentFLUSH RAMCHUNKcrossed a coroutine thread boundary while holding a pthread lock. Also fixed concurrent PQ metadata saves and TID snapshot ordering that could corrupt metadata or lose saved rules after a crash. - 🪲 v32.0.4 Issue #5085 Fixed cached rankers using incorrect row-ID boundaries under pseudo-sharding, which could produce incorrect results when reusing the query cache.
- 🪲 v32.0.3 Commit Fixed cluster bootstrap losing its saved node list and cluster user when
--new-clusterwas used or there were no reachable peers, preventing subsequent joins with inconsistent-cluster-user errors. - 🪲 v32.0.2 Issue #5095 Restored
--new-clusterand--new-cluster-forceafter a full cluster shutdown: explicit bootstrap no longer requires contacting stopped peers for their version or authentication details, while Galera still enforces safe-bootstrap checks. - 🪲 v32.0.1 Issue #4726 Fixed concurrent first MySQL connections racing during lazy SSL initialization and advertising TLS without usable keys. SSL availability is now decided once at startup, and invalid configured keys are reported there.
- 🪲 v31.0.22 PR #5093 Fixed
searchd --stopwaithanging during a cluster join or SST donation when shutdown had stopped the agent poller. Replication code now stops sending selected cluster requests once shutdown is in progress. - 🪲 v31.0.21 PR #5094 Fixed
FREEZEcrashing when its table set contained a distributed table; unavailable local indexes are skipped safely. - 🪲 v31.0.19 Issue #5066 Fixed deferred RT-table deletes and replacements ordering document IDs as signed values instead of unsigned values, which could leave documents with high-bit IDs undeleted.
- 🪲 v31.0.18 PR #5086 Updated Buddy to 4.4.10 for compatibility with PHP 8.5.
- 🪲 v31.0.16 Issue #5062 Fixed streaming
SELECT(includingLIMIT -1andSQL_NO_CACHE) crashing when it accessed columnar attributes not present in the selected row-wise blob data. - 🪲 v31.0.15 PR #5060, PR #5069 Fixed
index_convertercrashes and corrected--allwith--configto process the first configured table as well as subsequent ones; clarified the distinction between--pathand--all. ❤️ Thank you, @vitlav for the PR. - 🪲 v31.0.14 PR #5031 Fixed a
sortergroup.cppaccess declaration that caused compilation to fail with GCC 14. ❤️ Thank you, @RaphaelFakhri for the PR. - 🪲 v31.0.13 PR #5059 Fixed
ORfilters containing an alias derived fromWEIGHT()returning incorrect rows, including on the full-scan path. - 🪲 v31.0.12 PR #5058 Fixed columnar filters being skipped or losing rows when one filter used
ColumnarScanand another used a different access path, especially when columnar scanning was disabled for one attribute. - 🪲 v31.0.11 Issue #5047 Updated MCL to 14.5.7 to fix columnar scans dropping or repeating rows at the end of disk chunks.
- 🪲 v31.0.10 Issue #5008 Fixed
JOINwithREGEX()on joined-table attributes returning no rows and corrected filter routing when literals or JSON keys resemble table names. - 🪲 v31.0.9 Issue #5041 Fixed JSON
query_stringrequests containing only negative terms matching the opposite documents because an extra query-tree normalization removed the full-scan node. - 🪲 v31.0.8 Issue #5033 Columnar PR #217 Updated MCL to 14.5.6 to fix
!=,NOT IN, andNOT BETWEENsecondary-index filters silently dropping the final 64 rows of a chunk at a 64-row boundary. ❤️ Thank you, @drahy333 for the PR. - 🪲 v31.0.7 Commit Removed a regression that made row-wise string
COUNT(DISTINCT)roughly twice as slow by replacing repeated per-bytetolower()calls with a lookup table; string grouping also benefits. - 🪲 v31.0.6 PR #5034 Corrected JSON and MVA filtering inconsistencies across full scans, secondary indexes, row-wise/columnar storage, and SQL/HTTP. Fixes include array comparisons,
ANY()/ALL()conditions, mixed JSON types, exclusions, andCOUNT(*)/COUNT(DISTINCT)on multivalued data. Existing JSON secondary indexes containing repeated values should be rebuilt to remove duplicate entries. - 🪲 v31.0.5 PR #5011 Fixed secondary-index unions over MVAs and JSON arrays returning a document more than once when multiple values matched; row IDs are deduplicated across iterator and result-block boundaries.
- 🪲 v31.0.4 PR #5026 Fixed Debian bundle dependency resolution when an exact pinned development build was published only in the
release_candidaterepository; cache entries now remain repository- and architecture-specific. - 🪲 v31.0.3 PR #5025 Updated Test Kit package resolution to find the published Executor 1.4.6 across release, development, and release-candidate repositories rather than failing on a missing build artifact.
- 🪲 v31.0.1 Issue #4980 Fixed SST donor file reads hanging indefinitely on an unexpected EOF or repeating bytes after a short read; replication now advances the file offset and reports incomplete transfers explicitly.
- 🪲 v29.18.7 Buddy PR #703 Updated Buddy to 4.4.8, making KNN substitution use the parsed query owned by each request payload instead of shared parser state.
- 🪲 v29.18.5 PR #5010 Fixed
COUNT(DISTINCT)on strings ignoring the active session collation; the string-hash and secondary-index fast paths now use stored hashes only when the collation permits it. - 🪲 v29.18.4 PR #5007 Fixed string range comparisons through secondary indexes being treated as equality and corrected JSON/expression string ordering, negative filters, non-default collations, cache-key reuse, and evaluated-string leaks.
- 🪲 v29.18.3 PR #4997 Added bounds checks to MySQL-protocol handshakes and prepared-statement parameter handling to reject malformed packets safely.
- 🪲 v29.18.2 Commit Updated MCL's secondary-index builder to bound sort-bin buffers by bytes, write complete bins, and fail cleanly on short writes.
- 🪲 v29.18.1 Issue #5003 Fixed a crash in hybrid searches against distributed tables.
- 🪲 v29.17.2 Commit Fixed packaging builds that unnecessarily attempted to run MCL's library-build tests.
- 🪲 v29.17.1 PR #4998 Added validation for file names accepted during replication transfers to reject unsafe or invalid paths.
- 🪲 v29.14.7 PR #4993 Fixed a race between timed coroutine waiter registration and timer dispatch, with regression coverage.
- 🪲 v29.14.5 Issue #4989 Fixed merging grouped MVA results from multiple tables when source attributes are unavailable, avoiding incorrect or missing groups.
- 🪲 v29.14.4 Issue #4971 Fixed a performance regression in filtered KNN queries by deciding when to bypass HNSW before estimating secondary-index filter costs.
- 🪲 v29.14.3 Buddy PR #705 Updated Buddy to 4.4.7, fixing SQL fuzzy-search requests containing escaped apostrophes inside
MATCH(). - 🪲 v29.14.1 Issue #4977 Fixed replication SST transfers failing when file-hash metadata exceeded a network packet by adjusting the transfer chunk size to fit its metadata.
- 🪲 v29.13.2 Issue #4804 Updated MCL to address a columnar-query performance regression observed in benchmarks.
- 🪲 v29.12.2 Issue #4769 Fixed expressions evaluated after
LIMITreturning incorrect results on distributed tables with several local tables using different settings. - 🪲 v29.12.1 PR #4940 Fixed build failures in the new bulk-indexing code by using Boost filesystem instead of
std::filesystemon affected build targets. - 🪲 v29.11.6 PR #4925 Fixed
REPLACEandDELETEoperations paused byFREEZEmissing an in-flight disk chunk save and leaving duplicate or undeleted documents afterUNFREEZE. ❤️ Thank you, @tudorvasinca for the PR. - 🪲 v29.11.5 PR #4924 Fixed a parallel RT chunk-merge race where two workers assigned the same chunk ID, potentially overwriting files, losing documents, and later crashing. Chunk IDs now use an atomic counter. ❤️ Thank you, @tudorvasinca for the PR.
- 🪲 v29.11.4 Issue #4891 Fixed the type of a floating-point
EXISTexpression being lost in geospatial polygon containment. - 🪲 v29.11.3 Issue #4920 Fixed group-by handling retaining an unnecessary source attribute for MVA groups.
- 🪲 v29.11.1 Issue #4906 Fixed JSON highlight queries with apostrophes being escaped incorrectly and rejected or parsed incorrectly.
- 🪲 v29.9.6 Issue #4915 Fixed concurrent BERT auto-embedding requests deadlocking in the shared Rayon worker pool by serializing model conversions before entering parallel inference.
- 🪲 v29.9.5 Issue #4787 Fixed a t-digest memory leak in the implicit group sorter during parallel percentile aggregation.
- 🪲 v29.9.4 PR #4858 Fixed
ALTER TABLEleaving a document-ID lookup (.spt) in an older format than its disk-chunk header, which could corrupt ID lookups after an upgrade. Added checks and an offline repair path inindextool. ❤️ Thank you, @tudorvasinca for the PR. - 🪲 v29.9.2 PR #4900 Fixed non-rescored KNN searches dropping requested nearest neighbors when implicit
max_matcheswas smaller thank. - 🪲 v29.9.1 PR #4899 Improved batched vector rescoring performance and disabled batching for small
kvalues where the overhead outweighed the benefit.