Add annotation to skip rescheduling simulation for pods that don't need to survive node deletion #9110
Description
/area cluster-autoscaler
When scaling down nodes with unmanaged pods (no Deployment, ReplicaSet, or other controller), the autoscaler still runs rescheduling
simulation to find destination nodes for these pods. This causes issues:
- Unmanaged pods with safe-to-evict: true are evictable, but simulation still runs
- The simulation cascade effect (pods virtually scheduled on remaining nodes) causes later nodes to fail simulation
- Scale-down becomes slow or blocked, even though these pods don't need to be rescheduled - they can simply be terminated
Current behavior:
- safe-to-evict: true → Pod can be evicted, but still needs destination in simulation
- DaemonSet pods → Excluded from PodsToReschedule, no simulation needed
- Mirror pods → SkipDrain status, no simulation needed
Proposed solution:
Add a new annotation:
cluster-autoscaler.kubernetes.io/skip-reschedule: "true"
Pods with this annotation would:
- Be evictable (same as safe-to-evict: true)
- Receive SkipDrain status (like mirror pods)
- Not be added to PodsToReschedule
- Not require destination simulation
Use case:
Unmanaged pods that are intentionally ephemeral - they don't need to survive node deletion. The pod will simply be terminated when the node is removed, and that's acceptable. There's no controller to reschedule them anyway.
Implementation:
In simulator/drainability/rules/safetoevict/rule.go, add check:
if pod.Annotations["cluster-autoscaler.kubernetes.io/skip-reschedule"] == "true" {
return drainability.NewSkipStatus()
}