PostgreSQL has been a vital component in supporting OpenAI’s products, including ChatGPT, especially as user numbers surged to 800 million. The database's load increased over tenfold in the past year alone. OpenAI’s strategy relied on a single primary Azure PostgreSQL server and nearly 50 read replicas across multiple regions. This setup, while effective, faced challenges—particularly during traffic spikes caused by issues like cache misses or complex database queries. These spikes led to increased query latency and request timeouts, creating a feedback loop that risked degrading service quality. Despite PostgreSQL's strengths in handling read-heavy workloads, it struggled with write-heavy scenarios due to its multiversion concurrency control (MVCC). Each update creates a new version of a row, leading to write amplification and increased resource demands. To manage this, OpenAI has started migrating shardable workloads to Azure Cosmos DB to reduce pressure on PostgreSQL. They’ve also stopped adding new tables to the PostgreSQL instance, opting instead to direct new workloads to sharded systems. Although sharding PostgreSQL remains a consideration, it’s not an immediate priority because the current architecture still supports ongoing growth. To further alleviate the load on the primary instance, OpenAI has implemented several optimizations. They offload as much read traffic as possible to replicas and have made changes to application logic to reduce unnecessary writes. For instance, they've addressed bugs causing redundant writes and introduced lazy writes to help smooth out traffic spikes. While some high-volume write operations remain on PostgreSQL, efforts are ongoing to migrate these to more suitable systems. These strategies aim to maintain performance and reliability as user demand continues to rise.