need to test realistic CRUSH hierarchies
A user reported on the mailing list that their clients couldn't connect to their cluster. It turned out to be because their cluster was incorrectly requiring the CRUSH_V4 feature bit (https://github.com/ceph/ceph/pull/4322/files), which is going to be the case in any user cluster. Our upgrade test clusters didn't hit it because they have a custom flat CRUSH hierarchy.
Change things so that the upgrade tests are more representative of customer needs in this way, so we detect such failures in future.