Hi all, I have a k3s cluster with 2 nodes(1GB RAM and 2 CPUs). When I run some compute intensive workload on the cluster the node goes to a not ready state(because of OOM). I want to avoid the node going to not ready state instead I want to kill the pod which is consuming more resources. I tried setting the

kube-reserved, system-reserved, kube-reserved-cgroup, system-reserved-cgroup

flags on kubelet. But If I don't have requests or limits in my pod spec it still crashes the node(NotReady state). I have tried combination of setting the above flags with eviction-hard flag as well. Is there any other solution to overcome this?

If I don’t have requests or limits in my pod spec it still crashes the node(NotReady state).

Yes, if you don’t put any limits on your pod, it is allowed to consume all the resources on your node - and with only 1GB of memory, that’s not going to take much. If you want the pod to not consume all the resources, you need to use limits. The kubelet reserved args only inform the scheduler how much to reserve for those functions. If you don’t place any requests or limits on the pod then setting the reservations is meaningless.