Check the database state.
See Getting the Status of HADB
If the database state is Operational or better:
The machines needing unplanned maintenance do not include mirror nodes. Follow the single machine procedure for each failed machine, one DRU at a time. HADB service is not interrupted.
If the database state is Non-Operational:
The machines needing unplanned maintenance include mirror nodes. One such case is when the entire HADB is on a single failed machine. Get all the machines up and running first. Then clear HADB and recreate the session store. See Clearing a database. This interrupts HADB service.