Skip to content

CHANGELOG.md

Latest commit

 

History

History
116 lines (78 loc) · 3.27 KB

CHANGELOG.md

File metadata and controls

116 lines (78 loc) · 3.27 KB

Changelog

v0.4.0

New Features

  • Segment block compression: LZ4 compression for posting list blocks, reducing index size

Improvements

  • Coverity static analysis integration for detecting code quality issues
  • Implement AMPROP_DISTANCE_ORDERABLE in amproperty for better planner support

Bug Fixes

  • Fixed 'too many LWLocks taken' error on partitioned tables
  • Fixed Coverity static analysis issues

Breaking Changes

  • Segment format v3: Indexes created with v0.3.0 or earlier must be recreated after upgrading to v0.4.0. Use REINDEX INDEX index_name or drop and recreate the index.

v0.3.0

New Features

  • Block-Max WAND optimization: 4x faster top-k queries using block-level upper bounds to skip non-contributing blocks

Improvements

  • Added code coverage workflow with Codecov integration
  • Added segment integrity tests
  • Added nightly stress tests with memory leak detection
  • Added competitive benchmarks comparing pg_textsearch to other search engines
  • Code consolidation and refactoring for better maintainability

Bug Fixes

  • Fixed memory leaks by using private DSA for index builds
  • Fixed bm25query varlena detoasting for binary I/O and scoring

v0.2.0

New Features

  • V2 segment format: Block-based posting storage (128 docs/block) with skip index metadata for future Block-Max WAND optimization

  • Unlimited indexes: Replaced fixed-size registry with dshash for unlimited concurrent BM25 indexes

  • Benchmark suite: MS MARCO and Wikipedia benchmarks with historical performance tracking on GitHub Pages

Improvements

  • Major code refactoring: organized source into am/, memtable/, segment/, types/, state/, planner/, and debug/ directories
  • Page reclamation after segment compaction
  • Better cost estimation for query planning

Bug Fixes

  • Fixed excessive memory allocation in document scoring
  • Fixed buildempty() to write init fork correctly

v0.1.0

First open-source release!

New Features

  • Implicit index resolution: Queries can now use simpler syntax with automatic index lookup:

    SELECT * FROM docs ORDER BY content <@> 'search terms' LIMIT 10;

    Instead of the explicit form:

    SELECT * FROM docs ORDER BY content <@> to_bm25query('search terms', 'docs_idx') LIMIT 10;

  • Partitioned table support: BM25 indexes now work on partitioned tables. Each partition maintains its own index, with parent index OID automatically mapping to partition indexes.

Improvements

  • Added bm25_summarize_index() function for fast index statistics without content dump
  • Added COST 1000 to scoring functions to help the query planner prefer index scans over sequential scans
  • Improved page versioning with distinct magic numbers for different page types
  • PostgreSQL 18 support verified in CI

Bug Fixes

  • Fixed extension upgrade paths from 0.0.4 and 0.0.5
  • Fixed compiler warnings and standardized header guards

Breaking Changes

  • Removed to_bm25vector() function (use to_bm25query() instead)
  • Renamed scoring functions to bm25_ prefix (bm25_text_bm25query_score, bm25_text_text_score)
  • Existing BM25 indexes must be rebuilt after upgrading

Housekeeping

  • Added PostgreSQL License and NOTICE file
  • Added license headers to source files