It doesn't have to be HA and replication. Sharding by customer would mean a region outage takes out only 1/N customers. That wouldn't have a higher cost, other than whatever dev effort is needed to work that way.