Right-sizing Kubernetes requests and limits

Requests decide scheduling. Limits decide throttling and eviction. Confusing the two is the most common reason a cluster looks full while nodes sit at 20% utilization.

Measure before you guess

Start from actual usage over a period that includes your worst traffic day.

kubectl top pods --containers --sum

Use the p95 of CPU and the max of memory. CPU is compressible, memory is not, so the two need different safety margins.

A practical starting point

Resource	Request	Limit
CPU	p95 usage	2-4x request or unset
Memory	max usage + 20%	same as request

Setting memory limit equal to request gives the pod a Guaranteed QoS class, which makes eviction behavior predictable.

Keep it honest over time

Re-check sizing whenever the workload changes shape: a new caching layer, a different serialization format, or a runtime upgrade can move both numbers significantly.