We could also add in force step downs with no catchup period. This will use retryable writes since the primary is being killed, but provide maximum chaos and rollback scenarios.

This can be done with a single replica set and with a sharded cluster.