High level:

High level: 1. Picked tenant_id as the shard key (every query already had it) 2. Built a routing layer in the app: tenant → shard lookup cached in memory 3. Logical replication to copy each tenant group to its new shard 4. Dual-read verification for a week (compare results, alert on mismatch) 5. Cut over tenant groups in batches during low traffic, with a 1-click rollback Total: 4 months, 2 engineers. The dual-read week found 3 bugs that would've been incidents.