TIL OpenAI uses (used?) one primary write instance for their PostgreSQL cluster with dozens of read replicas. This powers the core ChatGPT service which has hundreds of millions of users and, needless to say, is a critical backbone to it. The talk implies they shard now, but the whole video emphasises all the optimizations they did in order to support their workload through a single primary. It isn’t mentioned at what time they switched to sharding, but it’s heavily implied that the single primary took them a very long way. It’s an incredibly testament to the power of modern hardware alongside witty optimizations. Comments