I think the thing people - and especially the OpenStack guys - don't appreciate is how terrible it is to (a) lose a workload or (b) require someone visit the datacenter (which may actually be a colo in another state or hours drive away). Having to fall back to some sort of terrible insecure management like IPMI or a dedicated mandatory management network (of questionable security), etc. is just not viable.

Systems and infrastructure architectures need to address a few things that really matter - error handling, continuous self monitoring, state compression and linear, systematic self-recovery - and while some OpenStack components handle this (ceph is pretty great except for a few things around access control and security) the whole doesn't handle them well at all. It's not enough to log a message (or worse, just log an exception stack).