Global File system and Cyrus

[Janne Peltonen]

On Fri, Nov 24, 2006 at 11:20:04PM -0300, Diego Woitasen wrote:
> On Fri, Nov 24, 2006 at 11:10:05PM +0100, Miros?aw Jaworski wrote:
> > On Fri, 2006-11-24 at 18:17 -0300, Diego Woitasen wrote:
> > > I 'm trying to implement a cyrus server using GFS (with a shared storage
> > > connected with fiber channel) but can't find information or experiencies
> > > about this. In the FAQ somebody says that is not recommended but his
> > > entry is two years old. Searching in google and mailing list I don't see
> > > a precise information about this combination.
> > 
> > It's general advice not to use networked filesystems due to inevitable
> > problems with locking.
> 
> GFS is not a networked filesystem, it's a shared filesystems that
> implement the full unix semantics, including locking. What kind of locking
> problem are you talking about? If the DBD is the problem, can it use
> other database?

There are success reports with skiplist. I had to compile without BDB
support to get rid of all error messages - setting all databases to be
something else than BDB wasn't enough. There might be trouble with a
non-Berkeley deliverdb and receiving a lot of mail.

Using skiplist, it all seems to just work. But I haven't used this
combination in a production environment yet, just testing, with no real
load. We should have this in production next summer, so I'll be wiser
then.

People have been using such a cluster system with a Veritas cluster and
a Tru64 cluster in production environments, too.

> > > The idea is install two or more servers using a shared storage for all
> > > mailboxes, to obtain High Availability and Load Balance.
> > You may want to check Cyrus' murder for HA.
> mmm, cyrus murder is not a fully HA system, it can suffer of partial outages.

On the other hand, you can set up a replicating Murder the way they have
in Cambridge: set each backend server to be the replication master for
some mailboxen and the replica for others. So, if you lose one of the
backends, you can tell the murder that the mailboxen reside on the
replica backend(s) - ok, requires manual intervention.


--Janne Peltonen
Univ. of Helsinki
email admin (one of them)