[Linux-cluster] fencing question, self stonith/suicide as last resort fencing

brem belguebli brem.belguebli at gmail.com
Tue Oct 13 12:45:59 UTC 2009


Hi,

This discussion may have occured many times here, but I'd like to
reintroduce it.

I know a lot of you Guru guys won't agree with the principle but
wouldn't that be something that could be implemented and reserved for
specific cases (multisite clusters for instance).

The idea being, in theses specific multisite cluster cases, if a
network partition occurs, no legacy fencing method will success,
preventing the services to failover to the sane cluster partition.

One solution (it seems to be implemented with different known cluster
stacks) is to trigger a hard reboot or a panic (this would be
prefered) after legacy fencing times out on the nodes who
transitionned from quorate  to inquorate.

This would allow the sane partition to recover the services from the
lost partition after the legacy fencing methods have timed out + the
(configurable) self fencing timer (a few minutes).

Regards




More information about the Linux-cluster mailing list