I guess cluster nodes should be scheduled to be taken down for random amounts of time so that they fail in sequence more gracefully.

Have an “off on weekends” node. And 24/7 nodes.