OpenAI runs ChatGPT for 800 million users on one Postgres primary

OpenAI detailed how it serves 800 million ChatGPT users from a single PostgreSQL primary without sharding. Database load grew more than 10x in a year: reads moved to nearly 50 replicas, shardable writes to Azure Cosmos DB, and PgBouncer cut connection time from 50 ms to 5 ms.

• Reads run on nearly 50 replicas while one primary handles writes
• PgBouncer cut connection time from 50 ms to 5 ms
• One SEV-0 in 12 months, during the image-generation launch
• Low double-digit millisecond p99 at millions of queries per second