Sharding our core Postgres database (without any downtime)

Gadget successfully sharded their core Postgres database to handle 400% traffic increases during Black Friday/Cyber Monday without any downtime. They split their control plane and data plane, using a custom maintenance mode primitive that pauses traffic for milliseconds during atomic cutovers. The process involved app-by-app migration using Postgres logical replication, achieving p95 pause windows of just 250ms while maintaining data consistency throughout the transition.