Comment 4 for bug 1735407

Some potential options (even bad ones):

1. We add the late-affinity check on the compute during evacuate. This could make the evacuate fail, but at least we wouldn't be breaking the server group policy. Since we don't have a reschedule loop for evacuate, the evacuate would just end there and the user would have to retry, but maybe that's OK in this case (it's maybe better than the alternative). I don't want to add a reschedule loop for the evacuate operation since reschedule loops are basically technical debt at this point, papering over bad scheduling decisions that we're trying to fix with Placement.

2. We could modify the evacuate API to allow you to evacuate an entire server group, which would get you the multi-instance evacuate in the same request functionality, which should allow the ServerGroupAntiAffinityFilter to work properly, but I doubt people would like that idea.

3. We wait for Placement to model affinity and hope that cures all problems...