[Linux-cluster] heartbeat to rgmanager design question

[Brian Kroth]

Andrew Beekhof <beekhof at gmail.com> 2008-11-28 08:50:
> On Thu, Nov 27, 2008 at 15:57, Brian Kroth <bpkroth at gmail.com> wrote:
> > Hello all,
> >
> > I've been using Heartbeat in past to do resource failover with the
> > following scheme:
> >
> > 1) Each node in the cluster runs a dummy monitoring resource agent as a
> > clone.  This resource agent monitors the health of a service on the node
> > using whatever rules one wants to write into it.  For instance, make
> > sure the service is not in maintenance mode, mysql is running, queries
> > return timely, and replication is up to date.  If all the checks pass it
> > uses attrd_updater to set an attribute for that service on the node to
> > 1.  Else, it is set to 0.  Note that this resource agent in no way
> > affects the service it is monitoring.
> >
> > 2) The cluster configuration uses the attributes for each of the
> > monitored services to generate a score for the machine.  The machine
> > with the highest score gets to host the virtual ip for that service.
> >
> > This scheme allows one to, for instance, touch a file on a machine that
> > will signify that it's in maintenance mode.  The service ip would then
> > be moved to another node, leaving one to test out the service on the
> > machine's management ip without removing it from the cluster itself
> > which would cause a lack of gfs access.  It also provides for more
> > granular monitoring of each service.
> >
> > I want to know how I would configure rgmanager with something similar to
> > this - to have resource agents that continually monitor the status of a
> > service on each node and then move service IPs accordingly.
> 
> Just out of interest, where did the rgmanager requirement come from?
> 
> <blatant-advertisement>
> The Heartbeat resource manager also runs on OpenAIS now which, IIRC,
> is what rgmanager uses... so, in theory, it can manage anything
> rgmanager can.
> </blatant-advertisement>

I think that's still in development, but I may be wrong.  If it's the
case I'll look back into it.  rgmanager is not a requirement.  I'm just
curious if I can get by running one cluster stack instead of two.

It occurred to me last night that I might be able to cook something up
that used ccs_tool to adjust the priorities of a node in a failover
domain.  I'll report back if this works out.  I haven't done any further
research yet.

> > I see that one can write their own agents, but I don't see a scoring
> > scheme anywhere.  My concern is that if I simply write an agent to
> > monitor a service and have an ip depend upon the return code of that
> > monitoring agent the service would not ever be failed back to the
> > original host.
> >
> > Does this make sense?
> >
> > Thanks,
> > Brian
> >
> > --
> > Linux-cluster mailing list
> > Linux-cluster at redhat.com
> > https://www.redhat.com/mailman/listinfo/linux-cluster