tensorchord/VectorChord 0.4.0

on GitHub

VectorChord 0.4 Release Notes

Major Improvements

• Streaming I/O & Page Prefetch

  • Complete rewrite of page layout to enable pipelined computation with streaming I/O.
  • On PostgreSQL 17, uses fadvise to prefetch buffers into the OS page cache, eliminating per-buffer read waits and fully leveraging disk throughput.
  • In upcoming PostgreSQL 18, direct support for io_uring will further streamline asynchronous I/O.
  • Benchmarks: 2–3× lower latency on cold queries (no buffer or page cache), translating to significantly improved tail latency in production.

• Prefilter Acceleration

  • Introduces true prefilter support for vector + filter queries.
  • Previous postfilter approach ranked full result sets and then applied filters—inefficient when selectivity is low (e.g., 1% filter rate).
  • Applies SQL filters before full precision vector distance computations, avoiding unnecessary work.
  • Benchmarks: Up to 3× faster end-to-end search on highly selective filters without any additional tuning.

Other Improvements

• Optimized Residual Quantization

  • Collaboration with RaBitQ author Jianyang: refactored distance term $|⟨o, q–c⟩|$ into $⟨o, q⟩ – ⟨o, c⟩$, so the query vector is quantized only once.
  • Result: ~20% QPS improvement over 0.3.
  • Recommendation: Enable residual quantization for L2 workloads.

• Fast Walsh-Hadamard Transform for Rotation

  • Collaboration with RaBitQ author Jianyang: replaced manual vchordrq.prewarm_dim GUC with an on-the-fly Fast Walsh-Hadamard Transform.
  • Removes the need to configure a prewarmed dimension list and yields marginal speed gains during setup.

Thank you for using VectorChord! As always, we welcome feedback and contributions on GitHub.

Full Changelog: 0.3.0...0.4.0