[DRBD-user] primary-primary with split brain resulting in read-only operation?

Graham Wood drbd at spam.dragonhold.org
Mon Apr 20 13:45:28 CEST 2009

Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.


On Sat, Apr 18, 2009 at 11:44:24AM -0700, Scott Bronson wrote:
> I'd like to set up a primary-primary DRBD setup under GFS or OCFS2
> (haven't decided which one yet).  The catch is, I have somewhat odd
> split brain needs.
This is normally not a good idea...

> Very quickly, if the two nodes split, I'd like one node to continue
> operating as usual and the other one to go read-only for the duration
> of the split brain.  Is this a good idea?  Is it reasonable to do with
> DRBD?
No, and not really...


> Would DRBD be a good tool to use?  Thanks
I think it would be as good as any, but that the idea itself seems flawed as described...

You seem to be trying to get HA on 2 nodes, without actually allowing for what happens in 
the failure case.

Split brain often cannot be distuinguished from node failure, with the result that if you 
set the cluster (which requires something on top of drbd, of course) as described then 
you're losing half the benefit.

Should PrimaryA die, then PrimaryB will go into read only mode - which is probably not what 
you want.

In addition to this, if you DO get a split brain (and it's an infrequent enough issue most 
of the time to not be worth this effort), then you have the problems associated with resync 
- which you seem to have glossed over.

If you're happy for PrimaryB to reboot as soon as PrimaryA gomes back online, then you could 
possibly mess with the cluster enough to do it, but you're still having a mess.

Can you explain what PrimaryA/PrimaryB are actually doing?  There may be another way of 
implementing this.

Graham



More information about the drbd-user mailing list