2 min read | by Jordi Prats
The cluster autoscaler takes into consideration several factors when it chooses a node to remove (evicting it's Pods)
When choosing a node to remove it considers:
It also takes into account Pods that cannot be moved elsewhere due to any set of constraints like nodeSelectors, affinity or simply due to lack of resources
But we can also set an annotation to tell it not to evict a particular Pod:
$ kubectl describe pod ampa-5948679894-vcmnv -n ampa Name: ampa-5948679894-vcmnv Namespace: ampa Priority: 0 Node: scopuli.pet2cattle.com/10.12.16.41 Start Time: Mon, 18 Oct 2021 19:56:53 +0200 Labels: app.kubernetes.io/instance=ampa app.kubernetes.io/name=website pod-template-hash=5948679894 Annotations: cluster-autoscaler.kubernetes.io/safe-to-evict: false (...) This also works the other way around, we can tell the cluster autoscaler that is is safe to evict a particular Pod that meets any of the other considerations using the same annotation set to true:
"cluster-autoscaler.kubernetes.io/safe-to-evict": "true" Posted on 28/10/2021