Re: [Evms-cluster] Question about MD and clustering

[Steve D. <st...@us...>]

Joshua M. Thompson wrote:
> On Mon, 2004-02-23 at 15:33 -0600, Steve Dobbelstein wrote:
>
> > As you can see, MD is safe in a private disk group because a private
disk
> > group is only accessible by one node the cluster.  There is no risk of
> > another node corrupting the metadata.
>
> Oh cool, that's even better than I expected. I was actually trying to
> reason out WHY it would be unsafe in a private disk group since I
> couldn't see how one node dying mid-update was any different than a
> single machine dying mid-update. The few posts I found on Google weren't
> too clear on the topic which is why I asked.
>
> Since it does indeed work I think I found a glitch: if the MD RAID-5
> region is currently rebuilding/syncing and you try to failover to the
> other node, both nodes end up syncing the array at the same time.

STONITH makes sure the failed node is dead before executing the failover.
STONITH is part of a robust HA cluster implementation.  Without STONITH you
run the risk of the situation you describe.

> I would imagine that running evms_activate after the region has been
> reassigned should shut down the resync process, shouldn't it?

Probably not.  If the disk group (CSM container) has been failed over to
node B, then node A will no longer see what's in the disk group.  A run of
evms_activate on node A won't find the resyncing array and will not know to
shut it down.

Steve D.