A Pod can sit Pending beside a node whose CPU graph looks quiet. The scheduler counts requested resources against node capacity. A low utilization graph shows recent usage, while a request reserves space in the scheduler’s accounting.
Start with the symptom
kubectl describe pod -n staging api-7fc86
kubectl top pod -n staging --containers
kubectl get pod -n staging api-7fc86 \
-o jsonpath='{.status.containerStatuses[*].lastState.terminated.reason}'
- Pending with Insufficient cpu or memory: inspect the requests, node capacity, and placement constraints.
- Running with slow responses: compare the latency trend with CPU throttling and request volume. A CPU limit can cap the time a container gets to run.
- OOMKilled after a burst: look for memory peaks, concurrency changes, or retained allocations. The latest snapshot may miss the spike that caused the kill.
Watch a full traffic cycle
Collect per-container usage during startup, a normal peak, and an unusually expensive request. Track the main process and sidecars separately. A single daily average hides allocator growth, request bursts, and batch jobs that overlap with the peak.
Check the knock-on effects before lowering a request. If an autoscaler uses CPU utilization as a percentage of the request, a smaller request can make the same workload look busier. Change one setting, observe the result over a comparable window, and keep the previous values with the deployment history.