Aller au contenu

KubeCPUOvercommit#

Sum of CPU request for all pods on all nodes is greater than the CPU capacity available on all the nodes minus one. The CPU request is defined in container manifests.

This alerts fires when the CPU requested by all pods would not fit on the cluster if it loosed a node.

Note

This alert only triggers for clusters where the autoscaler is not deployed.

Troubleshooting#

  • To have a look if nodes are overcommited you can use this :
kubectl describe nodes | grep -E "Name:|"cpu"" | grep -v "cpu:"

  • Verify with kubectl describe nodes to find pods that require a lot of CPU. Note that when the request is 0% in this output, the pod will have a default request of 10%. Thus the CPU requests total seen there is false.

  • For these pods there, check that the request CPU is expected.

Solutions#

  • for a customer cluster: The support team, inform the client of this alert.

  • some objects can be scaled down

  • the cluster can be resized (flavor, number of nodes). Prefer to add a node than changing the flavor.