PlanetScale launches TIN Postgres search extension

PlanetScale's choice to build directly on PostgreSQL's internal physical addressing rather than adapting generic Lucene-style search paradigms delivers remarkable performance gains, though locking the production engine behind PlanetScale's managed platform leaves self-hosters with only the unoptimized "Lead" fallback.

* By adopting 48-bit `ctid`s as document IDs, TIN avoids the expensive translation tables and cache misses required by external indexing architectures.

* Two-level page and offset bitmaps map cleanly into AVX2 and AVX-512 registers, turning conjunctions, disjunctions, and count queries into hardware-vectorized bitwise operations.

* Segment merges eliminate document renumbering, allowing direct ownership transfers of immutable bitmaps and drastically reducing disk write amplification.

* Pushing MVCC checks into native intersections with PostgreSQL visibility map bitmaps allows rapid index-only query execution without touching heap blocks.

* Restricting the optimized TIN engine to PlanetScale's cloud while open-sourcing only the slow, scan-based "Lead" companion for local testing creates lock-in and friction for teams wanting vendor-agnostic infrastructure.