The scariest line in a Halloween log:
FATAL: sorry, too many clients already
Postgres allows 100 connections by default, 97 for your app. 40 pods × 10 each = 400.
Seen it? What fixed it?
#postgresql#halloween
Wednesday: instant retries turned a slow minute into an outage.
Six rules:
1. Back off exponentially
2. Add jitter
3. Cap retries
4. Respect Retry-After
5. Don't retry 400/401/403
6. Give each client a retry budget
Which one does your client skip?
#sre#backend