[Linux-cluster] ways used for auto-eviction clarifications

[brem belguebli]

With multisites clusters, STONITH as only fencing method won't work in
case of network partition (intersite network global failure).

Tools to build multisites clusters already exist (lvm mirror, drbd) or
are going to be available (dm-replicator) but still the cluster stack
forbids the setup due to the only STONITH support.

Brem



2009/11/11 Lon H. Hohberger <lhh at redhat.com>:
> On Thu, 2009-11-05 at 02:17 +0100, Gianluca Cecchi wrote:
>> Hello,
>> can anyone summarize the possible events generating a self-eviction of
>> a node for an rhcs cluster?
>> Are these only executed via halt/reboot commands inside OS or also
>> through connection to the self-fence-device?
>
> linux-cluster does not generally have a notion of "self-fencing".
>
> Unless there's a confirmed "dead" by another cluster member, the node is
> considered alive.
>
> self_fence in the 'fs.sh' script is an exception, but not the way you
> might think...
>
> The node calls 'reboot -fn' if the umount command fails.
>
> However, this does /not/ obviate the requirement that another node
> successfully fence the newly-rebooted node prior to allowing recovery.
>
> In effect, the node is rebooted twice: once by itself, and once by
> another cluster member via iLO or whatever other power device you are
> using.
>
> -- Lon
>
>
> --
> Linux-cluster mailing list
> Linux-cluster at redhat.com
> https://www.redhat.com/mailman/listinfo/linux-cluster
>