Work out pod resource requests instantly with clear inputs, formula shown and shareable results.
Requests drive scheduling and are what the cluster reserves; limits cap actual consumption. Setting requests near the p95 with modest headroom avoids both waste and eviction. Memory limits should equal requests because exceeding a memory limit kills the pod, whereas CPU limits merely throttle — which is why a CPU limit above the request is safe and a memory one is not.
Requests and limits
request = p95 usage x (1 + headroom); CPU limit = request x multiplier; Guaranteed QoS requires requests equal to limits for both CPU and memory
It is contested. Limits cause throttling that hurts tail latency even below full utilisation, so many operators set CPU requests only and rely on requests for fair sharing.
Because that gives Guaranteed QoS, which makes the pod the last to be evicted under node memory pressure.