[Linux-cluster] RHCS separate datacenter

Thu Aug 12 23:59:39 UTC 2010

On Thu, 2010-08-12 at 15:03 -0400, rhurst at bidmc.harvard.edu wrote:
> Fencing *can* be used to avoid those stated error conditions, all too true... I think that is a "spot on" assessment.
> But, my point was fencing is not required for RHCS IP takeover or even the resource group manager; it is required for GFS -- which again was not part of the question.  If it is, then Ana needs to state that clearly.  Red Hat will "support" RHCS even if you choose *not* to use fencing -- because GFS won't start without *something* configured for fencing, even if it is configured wrong -- so if your fencing is not configured properly and {bleeps} up your GFS-based data, you're always at fault.  And there are other means in place of fencing to avoid those disastrous conditions from an IP takeover (not GFS), i.e., an arbitrator such as Cisco GSS can be implemented.  Does that make sense, or was there some other point I am totally missing?

in the HA world (what RHCS is made for), when building a highly
available service, imagine a database service, 2 points are to focus:

1) data integrity
2) high availability

for the first, depending on the cluster stack you use, different methods
are possible. Service takeover in RHCS must not occur until the fencing
of the failing node suceeded (Sicilian code of honor) .
Even if the configuration allows you to run without defined fencing, it
is required, and mandatory.
Some other cluster stacks rely on suicide, the node itself commits
suicide if it comes to fail whatever part (IP, storage, etc....)
(Japanese code of honour).

> Yes, RHCS is NOT based on IPVS... I did not mean it that way as an absolute; IPVS is a part of that suite which may or may not be used.
> 
To come back to Ana's question about RHCS capability to be stretched
across 2 datacenters, if it is to build highly available services
(involving dual shared storage, etc...) fencing will be mandatory but
will be the blocking point unless the site disaster recovery procedure
is a defined manual exercise.

> -----Original Message-----
> From: linux-cluster-bounces at redhat.com [mailto:linux-cluster-bounces at redhat.com] On Behalf Of Brem Belguebli
> Sent: Thursday, August 12, 2010 2:35 PM
> To: linux clustering
> Subject: Re: [Linux-cluster] RHCS separate datacenter
> 
> 
> On Thu, 2010-08-12 at 13:19 -0400, rhurst at bidmc.harvard.edu wrote:
> > FWIIW, I thought the question was in regards to Cluster Suite (RHCS), not Global File System (GFS)?  In that regard, what does fencing have to do with this?
> 
> Fencing is mandatory (otherwise not supported by RH) even if not using GFS. How do you deal with split brain if not fencing or suiciding ?
> 
> > @Ana, are you concerned about pulse heartbeat?  That should not be an issue, but more so, there will (probably) be no kernel client-session tracking for fail-over, because your remote datacenter is not on the same physical LAN.  That means if you did fail-over, the virtual IP will takeover, but all current sessions would be dropped and need to re-connect.
> > 
> > Also, from my understanding of Linux IPVS on which RHCS is based, is that it can support a remote datacenter, without spanning tree, if you use the tunneling option (not direct or nat)... although we have never tried it.
> > 
> RHCS is NOT based on IPVS
> .
> > But your "support" question sounds like it is aimed squarely at Red Hat should something break down, and not about Linux IPVS being capable of working or not.  If your implementation works solely with the tools Red Hat supplies, then I see no reason why they would not support an issue should it present itself during your QA.
> > 
> 
> > -----Original Message-----
> > From: linux-cluster-bounces at redhat.com 
> > [mailto:linux-cluster-bounces at redhat.com] On Behalf Of brem belguebli
> > Sent: Thursday, August 12, 2010 9:41 AM
> > To: linux clustering
> > Subject: Re: [Linux-cluster] RHCS separate datacenter
> > 
> > Hi,
> > 
> > Fencing is the blocking point, in case you want site disaster to be automatically handled.
> > 
> > Brem
> > 
> > 2010/8/12 Laszlo Beres <laszlo at beres.me>:
> > > On Thu, Aug 12, 2010 at 2:46 PM, Marchezetti Macedo, Ana Cristina 
> > > <anacristina.macedo at gmail.com> wrote:
> > >
> > >> I have heard that RHCS has not been designed for use across 
> > >> separate locations but I didn't find any red hat statment about it? 
> > >> Is it supported by Red Hat?
> > >
> > > IMHO as long as you can ensure all RHCS requirements between 
> > > different geographical locations, it should not be a problem. But 
> > > that's true, there is no definitive support for SRDF or other 
> > > storage replication, as it exists in Sun Cluster for instance.
> > >
> > > --
> > > László Béres            Unix system engineer 
> > > http://www.google.com/profiles/beres.laszlo
> > >
> > > --
> > > Linux-cluster mailing list
> > > Linux-cluster at redhat.com
> > > https://www.redhat.com/mailman/listinfo/linux-cluster
> > 
> > --
> > Linux-cluster mailing list
> > Linux-cluster at redhat.com
> > https://www.redhat.com/mailman/listinfo/linux-cluster
> > 
> > --
> > Linux-cluster mailing list
> > Linux-cluster at redhat.com
> > https://www.redhat.com/mailman/listinfo/linux-cluster
> 
> 
> --
> Linux-cluster mailing list
> Linux-cluster at redhat.com
> https://www.redhat.com/mailman/listinfo/linux-cluster
> 
> --
> Linux-cluster mailing list
> Linux-cluster at redhat.com
> https://www.redhat.com/mailman/listinfo/linux-cluster