Resource Configuration

Personal best practice:

• Always set memory request
• Always set memory limit to same amount as memory request
• Always set CPU request
• Don’t set CPU limits

Why memory limit = memory request?¶

• Memory can’t be reclaimed as easily as CPU shares
• Pods will be scheduled by requests, limits are not taken into account
• Thusly, having a couple of pods with memory limit > memory request on one node may starve some of them
• Therefore, setting limit = request eases capacity planning

Exception: Shortly lived pods like Cronjob Pods have their memory reclaimed after they are stopped and could therefore have memory limits > memory requests. But the issue of memory starvation in case several such pods are scheduled on one node still persists.

Why no CPU limits?¶

• Rarely (if ever) useful, as additionally claimed CPU shares can easily be reclaimed
• The myth, that Kubernetes system processes are starved without is wrong

Further reads¶

Kubernetes OOM and CPU Throttling by Sysdig