The current approach requires catching exceptions to handle these errors
and callers are already handling model validations so it seems more
appropriate.  Also it seemed to convoluted to add this logic directly to
the model since the model needs to check too many possible associations
to determine whether or not there are more than one cluster since the
model doesn't know what it's being created on. Additionally we only
wanted to validate during create to avoid the risk of existing models
becoming invalid by many different edge cases.