ParadeDB 0.26 Closes 8x Latency Gap With TIN on 150M Document BM25 Top-K Queries

ParadeDB engineers replaced sequential DocId-to-ctid mapping with direct page-aware postings and disabled dense_ratio search elision, matching TIN's reported 8x advantage on warm-cache read-only runs. The changes required 14 days of manual tuning on the same 8-client harness PlanetScale used. No architectural shift to native Postgres ctids occurred.

The rapid iteration mirrors how AI coding agents now compress weeks of index-tuning work into days. Similar tooling is already rewriting Git's object database layers ahead of the SHA-256 transition; early benchmarks show 9-12 % higher pack-file CPU cost under SHA-256 for repositories above 2 GB. ParadeDB's benchmark harness itself was generated with LLM assistance, revealing the same pattern.

Operationally, teams running mixed Postgres-Git workloads will face compounded lookup penalties once Git defaults to SHA-256 in 2027. Index authors must pre-allocate page-bitmap structures or accept visibility-check latency spikes. Future releases of both ParadeDB and libgit2 will need identical page-clustered identifier schemes to stay within current p99 budgets.

Next milestone is a 1 M document update workload; current code shows 3.2x regression versus TIN under concurrent writes.