Cyrus, clusters, GFS - HA yet again

[Scott Adkins]

I have to agree here, the reason we like the cluster filesystem method
is its shear simplicity.  It is nothing to add extra nodes to split the
load across more servers.  We don't have to partition users, we don't
have to worry about losing part of the user community when a server goes
down, only one place to hit for backups (any one server will do), one
address for the users to hit (without complicated means of achieving it),
etc etc.

In essence, if you have a good CFS system running, then it is as simple
as "if you can run Cyrus on one server, you can run it on all of them".

As somebody (Sebastion maybe?) pointed out, you do have to take care that
certain directories are *NOT* shared between the cluster nodes... the
/var/imap/socket directory is one, and I would highly recommend not
sharing the /var/imap/proc directory.  On our system, we also do not
share the /var/imap/log directory, though I do not recall why... it is
empty, and probably is empty because we abandoned BerkeleyDB...  For
SASL, we run SASL on all the nodes using UNIX domain sockets, and thus
the socket directory SASL uses also needs not be shared...

After that, it all just works...

Scott

--On Monday, October 30, 2006 11:09 AM +0200 Janne Peltonen <janne.peltonen at helsinki.fi> wrote:

> Hi, and thanks for the answer.
>
> On Sun, Oct 29, 2006 at 09:11:53PM -0500, Wesley Craig wrote:
>> On 27 Oct 2006, at 03:35, Janne Peltonen wrote:
>> > Or should I just give up and start considering Murder?
>> Before you decide on whether to give up on clustering, you should
>> thoroughly consider Murder.  At a minimum, it provides the location
>> independence that you'd like.  With replication, you have a
>> realistic, immediate recovery path for machine failure.  Both of
>> those are much better than what you have, while not as good as a real
>> HA system.  But murder + replication will be a lot less work than
>> developing an HA architecture from scratch.
>
> It is very true that building an HA architecture from scratch is a lot
> of work. But I'm not really trying to build it from scratch: I'm trying
> to recreate with Linux and RedHat GFS an architecture that's been in
> production use with Tru64 and its cluster FS (in Ohio), or with some
> other *nix and Veritas (in Pittsburgh).
>
> There were a couple of reasons I wanted to try this solution before
> Murder, replication and friends. The main reason was, if this could be
> get to work, it would have a lot less moving parts than Murder: no need
> for differently configured front and back end cyrus servers; no need for
> a MUPDATE master server. Each server could be more or less identical,
> and if one of them went down, there would be no loss of service (ok, the
> load-balancing box would have to notice that one of the ip addresses it
> shares went away, and the active imap connections to the dead server
> would die, but that's it, really). And adding a new server to the pool
> could be almost as easy as putting a new HP Blade on the rack,
> configuring its MAC in our DHCP server, installing an HD image on its
> local HD, booting it, and configuring its IP on the IP-sharing box.
>
> Of course, if it won't work, it won't work, and Murder is the supported
> solution. When my predecessor made his decision on what solution to
> pursue, he considered Murder much too complicated, and wasn't sure
> whether 2.3 was mature enough (that is, replication and friends), but
> this all might have changed - and he might have been wrong even then. ;)
>
> But then, other people seem to have got this to work - or something very
> similar. The simplicity of the setup is so appealing that I'd really
> like to give it a try.
>
>
> Regards,
> --Janne
> ----
> Cyrus Home Page: http://cyrusimap.web.cmu.edu/
> Cyrus Wiki/FAQ: http://cyrusimap.web.cmu.edu/twiki
> List Archives/Info: http://asg.web.cmu.edu/cyrus/mailing-list.html



-- 
 +-----------------------------------------------------------------------+
      Scott W. Adkins                http://www.cns.ohiou.edu/~sadkins/
   UNIX Systems Engineer                  mailto:adkinss at ohio.edu
        ICQ 7626282                 Work (740)593-9478 Fax (740)593-1944
 +-----------------------------------------------------------------------+
     PGP Public Key available at http://www.cns.ohiou.edu/~sadkins/pgp/
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 229 bytes
Desc: not available
Url : http://lists.andrew.cmu.edu/pipermail/info-cyrus/attachments/20061101/6a659168/attachment.bin