Kubernetes-Eager-OOM-Killer

Kubernetes containers can have memory limits set. These limits are only enforced when a node has memory pressure, so it is easy to have containers running well over their limits for a long time and only discover it when the system gets a load spike.

This tool watches a cluster and preemptively kills any pod that has containers over their limit, whether the node has pressure or not. It generates similar events to the real OOM killer.

Modes

Kubelet mode (default)

Runs as a single Deployment. Polls every node's kubelet /metrics/resource endpoint via the API server proxy to read container_memory_working_set_bytes, then compares against the pod spec's memory limit.

helm install eager-oom-killer ./helm
# or explicitly:
helm install eager-oom-killer ./helm --set mode=kubelet

Cgroup mode

Runs as a DaemonSet on every node. Reads cgroup v2 files (memory.current and memory.max) directly from the host filesystem. This avoids API server proxying and uses the same values the kernel uses for OOM decisions.

Requires:

• cgroup v2 (unified hierarchy) on the host
• containerd or CRI-O container runtime

helm install eager-oom-killer ./helm --set mode=cgroup

Configuration

Value	Default	Description
mode	kubelet	kubelet (Deployment) or cgroup (DaemonSet)
interval	1m	Polling interval (e.g. 30s, 2m)
cgroupRoot	/sys/fs/cgroup	Host cgroup v2 mount path (cgroup mode only)
image.repository	eager-oom-killer	Container image repository
image.tag	latest	Container image tag