OpenAI’s engineering team expanded a single‑primary PostgreSQL cluster with nearly 50 read‑only replicas, migrated write‑heavy workloads to Azure Cosmos DB, and applied extensive optimizations to reliably serve the global traffic of ChatGPT and the OpenAI API for 800 million users at multi‑million queries per second.
The article debunks OpenAI's sensational claim of handling 800 million ChatGPT users with a single PostgreSQL instance, revealing a pragmatic hybrid architecture that combines many read replicas, Azure CosmosDB for write‑heavy workloads, and top‑tier hardware, while highlighting cost and complexity considerations.
The article describes a high‑traffic core application with a MySQL master‑slave setup facing high resource usage; it details analysis of hardware, disk, table space, QPS, slow SQL, and presents governance actions including data archiving, moving reads to replicas, a MyBatis interceptor, and monitoring scripts, achieving significant performance improvements.
