Read replicas vs sharding: how far replicas got OpenAI, and when teams sharded
Published 2026-10-19 · Built only from first-party engineering write-ups in our database. Every number below comes from the linked source.
OpenAI is the strongest case for going far with replicas: ChatGPT runs on a single-primary Postgres with nearly 50 global read replicas, while write-heavy shardable workloads moved to Cosmos DB. Replicas scale reads, not writes, so the teams here whose writes outgrew one primary sharded instead: Figma, Notion, Slack and Shopify.
Read replicas vs Sharding: side by side
| Read replicas | Sharding | |
|---|---|---|
| Stories in our database | 6 | 6 |
| Companies | Figma, GitHub, GitLab, OpenAI and Shopify | Figma, Notion, PlanetScale, Shopify and Slack |
| Published | 2021–2026 | 2020–2026 |
| Stack in these stories | PostgreSQL (3), Azure (2), Azure Blob Storage (2), Git (2) | MySQL (3), PostgreSQL (3), Vitess (3), AWS RDS (2) |
| Scale numbers stated in sources | 99.99999% of repositories copied to Azure (GitHub); 473.3 billion Git events/month (GitHub); 800 million users (OpenAI) | 118.5M QPS sustained for 16 minutes (PlanetScale); 1 million QPS (PlanetScale); 2.3 million QPS (Slack) |
| Direct comparison story | OpenAI: Scaling PostgreSQL to power 800 million ChatGPT users | (same story) |
Verdict
Add replicas while reads are the bottleneck, and plan for replica consistency (Shopify) and replica stalls (GitLab). Shard, or move workloads out, when writes outgrow one primary.
Counts are stories in How They Scaled tagged with each side, see the read replicas and horizontal sharding pages. They describe what these teams wrote about, not market share or benchmark results.