- Sources: primary, discussion
- Summary: TIN uses the Postgres ctid directly as the document identifier, encoded as page-level and offset-level bitmaps sized to AVX2 and AVX-512 registers, which removes the identifier-to-ctid mapping that ParadeDB and pg_textsearch maintain and returns matches in heap order. The published benchmarks are vendor-run on a single 85 GB corpus and report 8 to 541 times the throughput of those alternatives. The release is GA only for PlanetScale Postgres and Neki databases rather than an extension to install elsewhere.
- Why it matters: The design removes a mapping layer that every comparable Postgres full-text extension has to keep consistent with the heap.
- Follow-up: Track whether TIN becomes installable outside PlanetScale and whether any third party reproduces the throughput figures.
send feedback on this story