What happened:

We were testing our internal fork which was recently rebased on top of the 1.16 release. As part of testing we did some cluster failovers with MultiplePodTemplatesScheduling feature set to true, but noticed that failover does not work as expected. Once workloads are evicted from their cluster, they do not get rescheduled.

The root cause seems to be due to the existing implementation of IsBindingReplicasChanged, which only takes spec.replicas into account. We need to also account for components.

What you expected to happen:

Scheduler should correctly detect when replicas have been scaled up or down when using multiple component scheduling.

How to reproduce it (as minimally and precisely as possible):

Attempt a cluster failover for a workload that has multiple components with the MultiplePodTemplatesScheduling feature set to true. This will be reproducible every time.