While testing Kubernetes redundancy and testing the Cluster’s reaction to a pod becoming unavailable – I found that the cluster took over 5 minutes to recreate pods after stopping the Kubelet service on one of the nodes.
This is due to the default pod-eviction-timeout level in Kubernetes being set to 5 minutes. To save delving into the default Kubelet configuration this can be bypassed simply by adding the below spec to Your Kubernetes deployment configuration:
template: tolerations: - key: "node.kubernetes.io/unreachable" operator: "Exists" effect: "NoExecute" tolerationSeconds: 5 - key: "node.kubernetes.io/not-ready" operator: "Exists" effect: "NoExecute" tolerationSeconds: 5
Applying this configuration will cause Your pods to be recreated on a healthy node after 5 seconds of the Node being in status “NotReady”