ALTER TABLE lock queue

A routine migration to add a column ran during a low-traffic window and 'hung'. Within a minute, the entire users table became effectively unavailable, every read and write to it is timing out, app error rate spiked to 100% on those endpoints, and the migration's ALTER TABLE is still active. pg_locks shows the ALTER waiting on an AccessExclusiveLock, and behind it a long queue of normally-fast SELECTs and UPDATEs all waiting. There's one old transaction at the front holding an AccessShareLock on users that's been open for 40 minutes. Triage and mitigate.