Pod Priority and Autoscaling
This component can enable autoscaling behaviour for a cluster. If a cluster uses PriorityClass objects to affect pod scheduling and eviction, it is important to know how these two systems interact.
Preventing Autoscaling
By default, the ClusterAutoscaler defines a spec.podPriorityThreshold of -10.
Pods with a priority less than this threshold will not trigger scale-ups (if the cluster has no capacity, they will simply not be scheduleable), and they also don’t prevent scale-downs (the pods will get evicted if that allows the cluster to scale down).
Pods with a priority equal or greater than the threshold will be treated normally - the cluster will expand to ensure they can be scheduled, and they will not be preempted to shrink the cluster. The same also holds for pods without priority.
For further information, see the Kubernetes autoscaler FAQ.
|
Changing the default |