Opinions required?

[John Waller]

This is not an actual problem, but I would like to hear some other opinions.

A few years ago (1997) we installed a 2 node ServiceGuard cluster on a customers site. This cluster ran 2 applications one being a testing / training area and one a live working area of the same application. Due to the customers financial budget at the time the test/train application VG was not mirrored and just lived on a single disk. The cluster was configured in such a way that this single disk was also configured as a secondary lock disk.
Unfortunatly this situation was forgot about and has been running fine until last Friday when this disk failed. Of course the worse thing occureded and when we shutdown the system which was running this test applicaton the alternate node which was running the live app TOC as it had problems obtaining the secondary lock disk. The bad thing is that the engineer has told the cluster that this configuration was not vailid.

OK , this setup was not ideal, but I cannot find anything that says you have to use a mirrored VG for a ServiceGuard package and believe that given the circumstances this setup is valid.

Any comments ??

[Justo Exposito]

Hi John,

I think that this is a good configuration for any hardware problem except for a disk failure as you can see. I suppose that the mirror software solve this problem and this must be a recomendation for a cluster, but if you know the problem and accept the risk of a disk failure your configuration is valid.

Regards,
Justo.

[David Navarro]

I think, you must eliminate all spof's for be supported by HP.
Yes, if you don't do so, SG will be work fine, but HP engineers, can't do anything in unsupported plataforms.
Are you sure that unmirrored VG is supported?

[melvyn burnard]

Couple of points here:

1) dual (2) cluster lock discs in a cluster are not recommended except for campus cluster scenarios, or where the cluster lock disc is powered from the same source as a node.

2) ServiceGuard is designed to provide resiliency for a package (application) but relies on disc technologies to protect the physiacl data, i.e. using RAID arrays, or MirrorDisk/UX

3) depending on the configuration, this may or may not have contributed to the problem you have experienced.

If a cluster is configured and you do not protect the data by either a RAID array or mirroring, this is not "unsupported", but means you are open to failures or issues caused by having a SPOF (Single Point of Failure). It is certainly not considered a good HA environment, and as such is seen to be "unsupported".

Hope that helps

[Steven Gillard_2]

Of course you don't technically "have" to mirror disks in a ServiceGuard environment, but by not doing so you are creating a single point of failure which is goes against what you are trying to achieve by installing SG in the first place.

What behaviour were you expecting when this disk failed? The situation you described seems perfectly reasonable under the circumstances. I'm with the HP engineer, that configuration is not valid in a SG environment.

Just my $0.02

Regards,
Steve

[John Bolene]

Melvin probably said it best.

To be resilient means having duplicate systems in case of a cpu or memory or lan failure.

To be redundant means having duplicate data that is mirrored somewhere.

The best scenario is resilient and redundant.

Disk technology has greatly improved in the last few years, but it is still a mechanical device.

[fg_1]

John

Worst thing that could happen to a cluster but as a fix it to the problem, your client may want to consider just using a small desktop size array (sc10) which is not very expensive and would allow him to have mirroring and redundancy for not only the O/S but the applications as well.

As everyone has said already, eliminating the SPOF's is key to any cluster solution.

Did you have mk_recovery tapes of the O/S and backups of the application? if so this will help with your restore process.

[John Waller]

Many thanks for your responses. You all seem to be agreeing with my initial thoughts, even though it is not ideal and leaves a SPOF, the engineer was wrong to tell the customer that it is an invalid setup.
A couple of interesting points, though, any ideas why having dual cluster lock disks in not a good idea except in a campus scenario.
Also Steven, the question is not really what did we expect to happen. I already knew we were in for a rough time when I realised we had a disk failure, but I was suprised the other node TOC as it could not obtain the secondary lock disk.

As for recovery, the O/S was unaffected and we just had to restore the data from the previous nights backup tape once the VG was recreated on new disks (yes we have learnt a lesson and now have a mirrored the VG