[Pacemaker] Resource-level fencing without stonith?

[Lajos Pajtek]

Using IPMI was also my first idea, but the LOM interfaces are (for security reasons) on a different LAN and there is no route to them.

Thanks for pointing me to the existing sg_persist RA. Also it seems that the author intended to use prout types 1 or 3. I thought of using 5 (write exclusive, registrants only). I saw your comments as well. Definitely a good starting point, and far more generic than what I have right now.


You wrote that I "likely still want STONITH even if using this RA." This is an OCF RA. How does STONITH come into the picture here?Or do you mean that I should use a real stonith device together with this RA?


Thanks and best regards,
Lajos



----- Original Message -----
From: Florian Haas <florian at hastexo.com>
To: Lajos Pajtek <lajospajtek at yahoo.com>; The Pacemaker cluster resource manager <pacemaker at oss.clusterlabs.org>
Cc: 
Sent: Friday, March 23, 2012 6:27 PM
Subject: Re: [Pacemaker] Resource-level fencing without stonith?

On Fri, Mar 23, 2012 at 6:07 PM, Lajos Pajtek <lajospajtek at yahoo.com> wrote:
>
>
> Hi,
>
> I am building a two-node, active-standby cluster with shared storage. I think I got the basic primitives right, but fencing, implemented using SCSI persistent reservations, gives me some headache.First, I am unable to get stonith:fence_scsi work on RH/CentOS 6. (Using the sg_persist utility I am able to register keys, etc so that's not the problem.)

Any specific reason for not using IPMI? That's practically ubiquitous,
and pretty much always works.

> This made me think about the fact that conceptually SCSI fencing should be resource-level fencing, not node-level fencing. The other node is not powered down or rebooted so perhaps I shouldn't be using stonith at all. Currently I think about having stonith-enabled="false" and I am writing a master-slave resource agent script to manage the SCSI persistent reservations in case of fail-over.

That idea of such a resource agent is fine, but please don't write one
from scratch. Instead, expand on this one:

https://github.com/nif/ClusterLabs__resource-agents/blob/master/heartbeat/sg_persist
That one's off to a good start, but the original author never had time
to finish it.

Mind you; you'll likely still want STONITH, even if you use the sg_persist RA.

Hope this helps.
Cheers,
Florian

-- 
Need help with High Availability?
http://www.hastexo.com/now