Skip to content

Releases

pgvectorfast 0.2.1

Choose a tag to compare

@zxjcarrot zxjcarrot released this 16 Nov 06:26

Enable pgvector and the Calico buffer pool with one SQL command after preloading vectorfast and restarting PostgreSQL:

CREATE EXTENSION vectorfast;

The separate CREATE EXTENSION vector; CREATE EXTENSION calico; CALL calico.activate(); setup remains available.

  • Fix buffer-pin reservation during a Calico fallback in VACUUM.
  • Include all pgvector SQL upgrade scripts in Docker images.
  • Validate packaged installations through both setup paths and pgvector regressions.
  • Add reproducible, recall-calibrated HNSW benchmarks for pgvector, pgvectorfast, Chroma, and Qdrant on SIFT10M and DEEP10M, with in-memory and larger-than-memory limits.

Installation: Docker, APT, PGXN, and source.
Benchmark results and methodology: BENCHMARKS.md.

Validation: PostgreSQL 18/19 source builds, Docker, APT, PGXN.

pgvectorfast v0.2.0

Choose a tag to compare

@zxjcarrot zxjcarrot released this 15 Nov 22:37

Release v0.2.0 of pgvectorfast, featuring dynamic /dev/shm buffer pool allocation, HNSW/vector search optimizations, Docker/APT/PGXN distribution channels, and vectorfast.so shared library packaging.

Upstream Base & Bundled Extensions

  • Based on: pgvector v0.8.7
  • Bundled extensions: vector (0.8.7) and calico (0.2, including 0.1 -> 0.2 upgrade path)

Highlights

  • Dynamic /dev/shm Calico Buffer Pool (calico 0.2): Replaces the static 1 TiB /dev/zero arena and fixed prefix hash table with on-demand /dev/shm extents (16 MiB metadata extents and 512 MiB region extents, REGION_BITS = 26), partitioned 128-lock prefix buckets, and lazy backend attachment.
  • HNSW & Vector Search Optimizations: 4-way AVX+FMA L2/IP distance kernels, inline optimistic slot/tuple reads, single-pass optimistic slot lookup, generation-stamped visited tables, pipelined layer-0 traversal with streaming tuple/heap prefetches, coalesced async AIO batching, and parallel index build references to vectorfast.so.
  • Distribution Channels: Multi-channel installation via Docker (zxjcarrot/pgvectorfast:0.2.0-pg18 / :pg18), APT (Debian bookworm & trixie), PGXN (pgxn install pgvectorfast), and source (PGXS).

pgvectorfast v0.1.0

Choose a tag to compare

@zxjcarrot zxjcarrot released this 15 Nov 04:29

Initial release of pgvectorfast (v0.1.0), a high-performance drop-in enhancement for pgvector powered by the Calico shared buffer pool.

Upstream Base Version

  • Based on stable release: pgvector v0.8.7 (commit f37c13f68b57d2c3472b2214fbcff699d6d34876, rebased from initial fork base d823c44591798ba3b13c34e60657bc1a8999e44b)
  • Bundled extensions: vector (0.8.7) and calico (0.1)

Highlights

  • Calico Shared Buffer Pool (calico): Replaces PostgreSQL's shared buffer hash table with hardware-Direct virtual memory page translation and per-page latch/state co-location (calico.enable_fast_buf = on).
  • Optimistic Page Locking for Vector Search (vector): Lock-free optimistic page reads for HNSW graph traversal and IVFFlat scans (hnsw.optimistic_lock = on, ivfflat.optimistic_lock = on).
  • Zero-Patch PGXS Installation: Builds and installs directly against unmodified PostgreSQL 18 and 19 servers on Linux x86-64 via make && make install.