I wonder if it's better in a non sharded use of something like Cockroach to push the replication and HA outside the database, let the control plane(e.g. kubernetes) handle failure and restarts on different nodes, and let something like Ceph with it's efficient write path and replication handle data durability.