hj lee wrote: > Thank very much for the reply. > > I tested it both stonith-enabled and no-quorum-policy. As Dejan pointed, > this is related to stonith-enabled. With stonith-enabled true (which is > default), > if I kill the master node, the slave stays as a slave, it seems expecting > something from stonith. With stonith-enabled false, the slave was promoted > to master.
Right. Because the surviving node did not care there might be something running on the dead node. That's usually not a good idea. > The no-quorum-policy controls the behavior when the quorum is not enough. > With no-quorum-policy stop, when one of two nodes dies, the resources in the > other node get stopped. With no-quorum-poolicy ignore, when one of two nodes > dies, nothing happens. If master die, the slave stays as a slave. If slave > dies, the master stays as a master. > > By the way during the test, no-quorum-policy works only in openais stack. > With heartbeat stack, there is no difference between no-quorum-policy stop > and ignore. It seems the heartbeat stack always behavior as no-quorum-policy > ignore. > > Here are my new questions: > How does stonith detect the node was actually down? stonithd just uses the plugin you configured and tells it to kill the node. If that is not successful, the node's state is consequently still considered "unclean". As Dejan already pointed out: Since your stonith plugin is ssh and the other node is _powered off_, ssh cannot connect and therefore is not successful. ssh is just a (bad) test plugin for stonith. It does only work if the node is reachable and is not for production use. "real" stonith devices need to be usable even if the node is already powered off (like in your case). Ipmi usually is just a little better, since - afaik - most ipmi devices won't respond if the node has no power. The devices are not OS-dependant like ssh, but still node-hardware-dependant. However, there are some drawbacks on all the stonith devices I have personally come across. Regards Dominik > Doesn't it get this info > from the cluster stack? Does it have its own mechanism like pinging the > other nodes? Can stonith:external/ipmi detect the node is down? Which type > of stonith has this ability? _______________________________________________ Pacemaker mailing list Pacemaker@oss.clusterlabs.org http://oss.clusterlabs.org/mailman/listinfo/pacemaker
