Your strict rate-limit fallback lets one user through at five times the limit

transcript

A Redis node dies. Your rate limiter falls back to strict local limits. One user still gets five times their limit.

Your counters live in a 3-node Redis cluster. Every request checks one. Node 3 dies. A third of your counters just became unreachable. Failover takes 10 to 30 seconds. User B's key lived on node 3.

Fail open, and those users have no limit at all. Fail closed, and a third of your users are blocked by your outage. So the lesson picks a local fallback. A conservative limit on each server.

Each server allows 10 percent of the global limit. Sounds strict. But each server only knows its own share. Not the others'. You have 50 servers. Spread across all of them, one user gets 500 percent.

For a 30-second window, the lesson accepts that. No user is fully blocked. The system stays up. And you alert on it. In the interview, say the worst case out loud. Five times, for 30 seconds.

A limit split across servers multiplies by the number of servers. You didn't hope the failure away. You priced it.

Full breakdown in the lecture.