Scaling Postgres to the next level at OpenAI | POSETTE: An Event for Postgres 2025

Video of a conference talk discussing how PostgreSQL serves as a cornerstone of OpenAI's backend infrastructure, powering many of our most critical features. In this talk, I will share how we've taken PostgreSQL to the next level to handle OpenAI's demanding workloads on Azure. I'll cover the challenges we faced, the lessons learned from outages, and the strategies we implemented to scale Azure Database for PostgreSQL to handle millions of queries per second (QPS). Key topics include: • Connection Pooling: How we effectively manage connection pooling with PgBouncer. • Read Replication: Strategies for scaling read replicas to meet increasing demand. • Schema Migrations and Long-Running Queries: Approaches to seamless schema migrations and resolving performance bottlenecks caused by long-running queries. • Observability: Enhancing monitoring and observability to ensure reliability and performance. Bohan Zhang is a member of the technical staff at OpenAI. Prior to joining OpenAI, he co-founded OtterTune, a startup specializing in automatic database tuning services. He has also been a speaker at various PostgreSQL conferences, including PostgresConf 2024 and PostgresConf Silicon Valley 2023. He also co-authored with Prof. Andy Pavlo the article "The Part of PostgreSQL We Hate the Most," which critically examines PostgreSQL's implementation of MVCC, and is cited in the PostgreSQL's Wikipedia page. ► Video bookmarks: ⏩ 00:00 Introduction ⏩ 01:35 Background – how Postgres was scaled to meet demand at OpenAI ⏩ 04:02 Challenges in write-heavy workloads ⏩ 05:49 How read-heavy workloads are served by unsharded Postgres ⏩ 06:19 Reduce load on primary ⏩ 09:09 Query optimization ⏩ 11:58 Single point of failure ⏩ 15:03 Rate limit ⏩ 17:46 Connection pooling ⏩ 19:53 Schema management ⏩ 23:40 Wrap-up 📕 Everything you need to know about POSETTE: An Event for Postgres 2025 can be found at: https://posetteconf.com ✅ Lear