OpenAI: Scaling PostgreSQL to the Next Level

Summary

At the PGConf.dev 2025 Global Developer Conference, Bohan Zhang from OpenAI shared OpenAI’s best practices with PostgreSQL, offering a glimpse into the database usage of one of the most prominent unicorn companies. At OpenAI, we utilize an unsharded architecture with one writer and multiple readers, demonstrating that PostgreSQL can scale gracefully under massive read loads.— PGConf.dev 2025, Bohan Zhang from OpenAI Bohan Zhang is a member of OpenAI’s Infrastructure team. He studied under Professor Andy Pavlo at Carnegie Mellon University and co-founded OtterTune with him. Background PostgreSQL serves as the core database supporting the majority of OpenAI’s critical systems. If PostgreSQL experiences downtime, many of OpenAI’s key services would be directly affected. There have been several instances in the past where issues related to PostgreSQL have led to outages of ChatGPT. OpenAI utilizes managed databases on Azure, employing a classic PostgreSQL primary-replica replication architecture without sharding. This setup consists of one primary database and over forty replicas. For a service like OpenAI, which boasts 500 million active users, scalability is a significant concern. Challenges In OpenAI’s primary-replica PostgreSQL architecture, read scalability is excellent. However, “write requests” have become a major bottleneck. OpenAI has implemented numerous optimizations in this area, such as offloading write loads wherever possible and avoiding the addition of new services to the primary database. PostgreSQL’s Multi-Version Concurrency Control (MVCC) design presents some known issues, including table and index bloat. Tuning automatic garbage collection (vacuuming) can be complex, as each write operation generates a complete new version, and index access may require additional visibility checks. These design aspects pose challenges when scaling read replicas: for example, increased Write-Ahead Logging (WAL) can lead to greater replication lag, and as the number o...