Requests, limits, and quality of service

The two numbers that decide your fate

The request is what the scheduler reserves; the limit is the ceiling the runtime enforces. Requests decide where a pod lands, limits decide how it misbehaves when it exceeds them.

CPU and memory fail differently. Exceeding a CPU limit throttles the process — it gets slow. Exceeding a memory limit kills it outright, with no application-level error and a very unhelpful exit.

Pods whose requests equal their limits get the highest quality-of-service class and are evicted last under pressure. Pods with no requests at all are the first to go.