On Tue, Feb 21, 2012 at 3:12 PM, Vladislav Bogdanov <bub...@hoster-ok.com> wrote: > 21.02.2012 02:40, Andrew Beekhof wrote: >> On Mon, Feb 20, 2012 at 11:49 PM, Vladislav Bogdanov >> <bub...@hoster-ok.com> wrote: >>> 20.02.2012 14:36, Andrew Beekhof wrote: >>>> On Mon, Feb 20, 2012 at 10:26 PM, Adrian Fita <adrian.f...@gmail.com> >>>> wrote: >>>>> Thanks, I figured it out by now. But the real problem I'm facing is >>>>> explained in >>>>> http://oss.clusterlabs.org/pipermail/pacemaker/2012-February/013124.html >>>>> . Please also take a look there. This advisory thing was me trying to >>>>> solve the problem described in that message. >>>> >>>> Ok, I see that thread. I'll get to that one soon. >>> >>> I now look at the similar case (advisory ordering doesn't work). >>> >>> I hope I can add some pointers on what exactly happens. >>> >>> Imagine you have two advisory-ordered resources each also depend >>> (colocated and mandatory ordered) on some other resource (f.e. >>> filesystem on drbd device Master state). >>> I now play with lustre filesystem. It consists of several parts: >>> * MGS (one per cluster, cluster can consist of several lustre filesystems) >>> * MDT (one per filesystem) >>> * OST (many per filesystem) >>> It is very important for lustre to operate properly that MDT is started >>> after MGS and OSTs started after MDT. >> >> The you must not use advisory constraints for this sequence. > > Heh, but then I will be unable to move filesystem parts between drbd > replicas for failover... Advisory ordering allows me to move MGS or MDT > to another replica without restarting all the OSTs. > > Need to think again about making them migratable... > >> >>> >>> In my case I have MDT configured to start after MGS (advisory) and each >>> OST to start after MDT. Each of lustre parts is resided on a drbd device >>> (stacked one, so there is one more drbd device under it). >>> >>> What I see after transition from stop-all-resources="true" to "false" is >>> that some OSTs (which are located on nodes different from where MDT is) >>> are started before MDT. I see that advisory ordering is not honored when >>> dependencies of "then" resource are satisfied before ones of "first" >>> resource. >> >> So /both/ sides of the constraint are starting or stopping but the PE >> isn't enforcing the ordering? > > At one moment LogActions prints that they all are about to start. > Then deps for two "then" resources are satisfied and they are actually > started while "first" resource is not.
Ok, can you file a bug for that please? I'll get David to have a look. _______________________________________________ Pacemaker mailing list: Pacemaker@oss.clusterlabs.org http://oss.clusterlabs.org/mailman/listinfo/pacemaker Project Home: http://www.clusterlabs.org Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf Bugs: http://bugs.clusterlabs.org
