What manticoresoftware/manticoresearch shipped

Week's updates: streaming SQL columnar fixes, index_converter crash fixes, GCC 14 build fix, per-test CLT failure tracking, Grafana 13.3 coverage. Releases 31.0.14–31.0.17.

This week we fixed streaming SQL queries to materialize columnar attributes and stored fields even when not selected, avoiding mysqldump errors. We fixed index_converter crashes with --config and --path and added a clear error for --path --all. We fixed a GCC 14 build failure in sortergroup.cpp. CI now tracks per-test CLT failure rates over time with a daily or on-demand report. We added Grafana 13.3 to tested versions and updated supported range documentation. Releases 31.0.14 through 31.0.17 include these changes.

New releases 31.0.6 through 31.0.13 fix query correctness for columnar filters, JOIN with REGEX, JSON negation, and secondary index range boundaries. COUNT(DISTINCT) on strings is no longer 2x slower. Docs now cover bulk_import memory and disk usage.

This week we published releases 31.0.6 through 31.0.13, addressing several query correctness issues: filters through columnar scans no longer lose or repeat rows, JOIN with REGEX() on joined table attributes returns expected rows, JSON query_string with only negated terms returns the correct result, and!=, NOT IN, NOT BETWEEN through the secondary index no longer lose rows at range boundaries. We also fixed a performance regression where COUNT(DISTINCT) on rowwise strings was 2x slower, and made JSON and multi-value filters return consistent rows across storages and access paths. Documentation now includes memory and disk usage guidance for bulk_import.

31.0.0 is out: manticore.json writes are now crash-safe and replacement is atomic. Replication join now fails cleanly when nodes cannot reach each other; protocol moved to 0x110, so upgrade all nodes together. Duplicate-row fixes for secondary index unions, collation-aware COUNT(

This week's releases include 31.0.0, which makes manticore.json updates crash-safe by syncing the new file before atomically replacing the old one, so a power loss during DDL no longer leaves an empty registry. 30.0.0 is a breaking change for replication: JOIN CLUSTER now checks reachability before state transfer and fails with the failing node's error, and the cluster protocol is now 0x110, so all nodes must be upgraded together. Also fixed: duplicate rows from secondary index unions on MVA and JSON arrays, COUNT(DISTINCT) on strings ignoring session collation, a hung SST donor read loop on zero or short reads, out-of-bounds reads in the MySQL handshake parser, and string range comparisons with secondary indexes. mysqldump coverage now includes MySQL 26.10.

New releases this week: 29.18.2, 29.18.1, 29.18.0, 29.17.2, 29.17.1, 29.17.0 and 29.16.0. Hybrid search over distributed tables no longer crashes, replication file names are validated, vector rescoring gets prefetch, bulk_import handles auto-id, and KNN results can highlight the

Several Manticoresearch releases landed this week, from 29.16.0 through 29.18.2. A crash on hybrid search over a distributed table is fixed, replication file names are now validated as plain basenames, and MCL build testing is disabled on packaging. On the feature side, vector rescoring gains prefetch with rowwise storage, bulk_import handles auto-id, and KNN queries can highlight the chunk of an auto-chunked text that matched. indextool --docstore-bad-ids lists documents with damaged stored fields, and indextool --check now reports rows not covered by docstore blocks.

New releases: per-table knn rescore budget, parallel hnsw build, fixes for filtered knn regression, MVA merge, apostrophe escaping, and more. See changelog.

This week we shipped several fixes and improvements. Performance regressions in filtered knn and columnar queries were fixed. MVA group merging in multi-table queries now works without source attributes. We added a per-table knn rescore budget, vector prefetch for columnar rescoring, and early termination on agents. HNSW graph building now splits ramchunks into equal row ranges and runs in parallel on FLUSH RAMCHUNK. SST FILE_RESERVE hash metadata is now bounded. Fuzzy SQL apostrophe escaping was fixed via Buddy 4.4.7. manticore-load gained a --realistic mode. Full details in the changelog.

29.13.1 improves DDL error messages for unsupported column types. 29.13.0 adds facet/aggregation support in hybrid search with RRF fusion.

Two releases this week: 29.13.0 brings facet and JSON aggregation support to hybrid search, with counts properly reflecting post-fusion filtering. 29.13.1 improves DDL validation by reporting unsupported column types clearly and suggesting database types where applicable, making table creation less error-prone.

29.12.2 shipped: direct-to-disk bulk import for RT tables, fixed postlimit in distributed setups, resolved FREEZE/REPLACE races, and improved build compatibility.

Week 29 brought major improvements to Manticoresearch: direct-to-disk bulk import for real-time tables now enables faster large-volume data insertion, build stability improved across Debian and RHEL, and critical race conditions fixed in chunk merging and FREEZE operations. Documentation expanded for HTTP behavior, geo-distance models, and numerical precision.

29.10.0 ships support for BAAI bge-m3 ONNX models, fixes a deadlock in concurrent BERT embeddings, and resolves index format sync issues on ALTER operations.

29.10.0 is available with several important fixes. Concurrent BERT auto-embedding requests no longer deadlock. Docid lookup now syncs correctly with index format versions during ALTER operations. ONNX models without token types are now supported, including BAAI/bge-m3. KNN queries also get improved max_matches handling.