IMO this is the problem with all storage clusters that you run yourself, not just Ceph. Ultimately, keeping data alive through instance failures is just a lot of maintenance that needs to happen (even with automation).