Description of problem:

• The Cluster Autoscaler CAPI provider incorrectly calculates safety constraints during MachineDeployment rolling updates, leading to massive over-deletion of machines.

Version-Release number of selected component (if applicable):

    4.18 (but we believe this on all the versions)

How reproducible:

  Often  

Steps to Reproduce:

1. Create MachineDeployment with  minSize=1 maxSize = n , assume current spec.replicas=20
2. Set maxUnavailable=0 for rolling update
3. Trigger rolling update (e.g., change machine template)
4. use PDB or K8s-Shredder to prevent old machine from old machineset deleting
5. For example 20 new machines  be created, current 40 total machines
6. The workloads has migration from old machines to new version machinese (making old machines empty)
7. Observe Cluster Autoscaler expect to remove 19 empty nodes which stucking rolling upgrade
8. Result: MachineDeployment scales down to 1 machine, deleting 39 machinese total

Actual results:

1. Cluster Autoscaler calls SetSize(20-19) → reduces spec.replicas to 1
2. MachineDeployment controller sees spec.replicas=1 vs 40 actual machines
3. As old machine in old machineset will not be deleting, all new machine in new machineset were deleted expect 1.
4. MachineDeployment controller actually deletes 39 machines (not just the 19 empty ones)
5. Machinedeployment scales from 40 machines to 1 machine

Expected results:

Safety calculations should consider actual running machine count during rolling updates.

If CAS find the running empty nodes are from the old machines, they should considering it's pending delete, so no scale down actions to reduce the machinedeployment replicas

Additional info: