Question

Over the last hour, requests to a couple of API endpoints started timing out waiting for a database connection, then spread. PgBouncer in transaction mode reports its server pool fully checked out, but `pg_stat_activity` shows most of those server connections sitting in `idle in transaction`, some for many minutes, all from one service deployed at the start of the hour. DB CPU is low. The deploy added a new code path that calls an external HTTP API in the middle of a DB transaction. Triage and mitigate.