OpenAI pushed PostgreSQL to handle millions of QPS across 800M users. How? Nearly 50 read replicas, heavy read offloading, and serious trimming on write pressure.
Writes? Sent elsewhere. Sharded systems like CosmosDB, lazy writes, and app-level tweaks helped sidestep PostgreSQL’s MVCC write amplification mess.
Cache misses don’t get a free pass either—a custom cache locking setup rate-limits bursty traffic before it hits the primary.
Still not enough? They’re testing WAL relay replication. Relay nodes forward the write-ahead log, offloading replicas and buying time beyond normal scaling ceilings.
