When the Default Postgres Pool Died at 3 AM

Veltrix replaced a Postgres-backed JSONB update path with a Rust scoring pipeline over sharded Parquet files.

The old design caused recurring connection pressure, row-level lock contention, and a Redis cache stampede under heavy traffic.

By moving relevance scoring out of Postgres and tightening connection limits, p99 latency dropped from 1.4 seconds to 42 ms.

Database load stabilized, autovacuum recovered, and memory/allocation behavior improved significantly.