We kept getting paged for latency.

The SRE team knew the drill. Shift load to replicas, scale the database, bounce connections. It worked, usually. Things settled down. The on-call engineer closed the incident and went back to sleep.

Then the same page fired three nights later.