Comment 5 for bug 1713783

It looks like the _destroy_evacuated_instances method in the compute manager has always filtered migrations on the 'accepted' status since originally this code was just meant to cleanup local resources once an evacuation from the source host has started, which is fine. The problem is with removing the source node allocations if the evacuation failed, but if it failed in the conductor, we can fix that here:

https://review.openstack.org/#/c/499237/

If it failed in the destination compute service, the migration status should be set to 'failed' and the migration filter in _destroy_evacuated_instances would filter it out.