How do you know it's safe to lower Kubernetes resource requests?

November 16, 2026 4 answers
Asked by _p58!tramnotch.4

If a workload usually uses only a fraction of its requests, what should I check before lowering them? CPU utilization autoscaling and past OOM kills seem like they could make current usage misleading.

Answers

2w+drainpocket_591

Check the HPA calculation before changing CPU requests. With a CPU utilization target, utilization is measured relative to the request. Lower the request and the reported percentage rises, even if the actual CPU usage stays the same.

For example, a container using 250 millicores is at 25 percent of a 1 CPU request, but 50 percent of a 500 millicore request. Compare that with the HPA target and consider whether the workload could add replicas after the change. Watch replica count and usage as it settles.

Reply to this answer
d.527~ferryspool

I'm much more cautious with memory, especially if the limit is changing too. A container that was OOM killed and restarted can look perfectly comfortable now, while the spike that killed it is no longer visible in a current-usage chart.

Look at historical peaks and OOM events over a useful window, not just the average. A lower request and a lower limit are different changes, so don't treat evidence for one as automatic evidence for the other.

5m!socketcaddy_827:

Would you wait for a longer history if there isn't enough data to see occasional peaks?

Reply to this answer
6k~ferrulecrank_583

Don't overlook replica count. If there are lots of lightly used pods, reducing the minimum replicas may be worth considering before cutting every pod's request. Let the HPA settle and check the metrics after the change, rather than changing both at once.

Reply to this answer
4f!mapshuttle_927

The VPA recommender can be a useful starting point for sizing, even if you don't want it applying changes automatically. I'd still compare its recommendations with usage history and the HPA configuration. A recommendation is not the same thing as evidence that a lower memory limit will survive a peak.

Reply to this answer

Add your answer

Answers are reviewed before publication.

Related Discussions