Hatchet
A guide to preventing Postgres from toppling over.
Engineering Case Study
Summary
OpenAI explains how PostgreSQL powers ChatGPT for 800 million users via a single primary Azure PostgreSQL instance with nearly 50 global read replicas, covering optimizations, challenges with write-heavy workloads, and lessons learned at scale.
Highlights
- A single primary Azure PostgreSQL flexible server with nearly 50 read replicas across multiple global regions supports millions of queries per second
- PostgreSQL load at OpenAI grew more than 10x over the past year
- Read-heavy scaling relies on caching layers, multi-layer rate limiting, query routing, and application-level optimizations
- Write-heavy workloads are shifted to sharded systems such as Azure CosmosDB because PostgreSQL's MVCC is less efficient for writes
- The architecture achieves low double-digit millisecond latency and five-nines availability with only one major incident in the past year
auto-generated
Context
Audience
Database engineers, infrastructure engineers, SREs, and architects working with PostgreSQL or relational databases at scale
DomainDatabase Engineering
FormatTechnical blog post
Accessfree online
Source Overview
OpenAI scaled PostgreSQL to handle millions of queries per second for 800 million ChatGPT users by deploying a single primary database instance with nearly 50 read replicas across multiple regions, offloading read-heavy traffic while minimizing writes.
The system handles massive global traffic through optimizations including caching layers, rate limiting across multiple layers, query routing, and migration of write-heavy workloads to sharded systems like Azure CosmosDB. Despite PostgreSQL's inherent challenges with write-heavy workloads due to its multiversion concurrency control implementation, this architecture maintains low double-digit millisecond latency and five-nines availability with only one major incident in the past year.
Topics
Related
Azure PostgreSQL Flexible ServerAzure CosmosDBPostgreSQL MVCCRead Replica ArchitecturesDatabase ShardingHigh-Availability Systems
