Why does the autoscaler remove Nodes that are still running my workload?

The Kubernetes Cluster Autoscaler decides that a Node is unneeded using its own utilization-based logic. It does not consult CoreWeave’s CWActive Node condition or the Node Pool scaleDownStrategy. Those govern scale-down driven by the NodePool Operator when you lower targetNodes directly, not the autoscaler. As a result, if a Node becomes underutilized for the configured period while it is still active, the Cluster Autoscaler can determine that the Node is unneeded and mark it for removal. To prevent workload disruption, use one of the following annotations:

To prevent the autoscaler from evicting a Pod, annotate the Pod with cluster-autoscaler.kubernetes.io/safe-to-evict: "false".
To prevent the autoscaler from marking a Node for removal, annotate the Node with cluster-autoscaler.kubernetes.io/scale-down-disabled: "true".

For full details, see Node utilization in the autoscaling documentation.

Nodes Workload Scheduling