High Availability setup

[herberts]

Hi,

I am in the process of evaluating Zimbra CS for 10000 to 15000 users. We need performance, scalability and high availability.

I am not a big fan of active/standby setups, I prefer active/active and shared nothing architectures.

In order to achieve this goal I was wondering if I could set up the following:

N Message store nodes all mounting disks via NFS from our filers.

A MySQL Cluster to store the metadata.

Several replicas of the LDAP server.

With a HW load balancer, IMAP/POP/HTTP requests could be sent to any server.

With this kind of setup, all nodes have the same view of messages and metadata, we could therefore scale horizontally (until such time when NFS becomes the bottleneck).

Is there something I have missed? Are there any caveats to using NFS?

Any comment welcome.

Mathias.

[rsharpe]

I don't believe you can link all of your message stores back to an NFS mount, but I could wrong. For HA you should probably be looking more towards a cluster solution with a SAN.

[marcmac]

NFS is a bad, bad, bad idea. Lousy file locking, and it'll almost certainly fail when we try to create a hard link. Get a SAN, or build a box as an iScsi target.

[reza225]

use NFS3 or AFS perhaps to solve the locking issues. doesn't RHEL and possibly fedora have a nfs-locking daemon that runs to solve nfs's locking issues?

[herberts]

I agree on the lousiness of NFS, I was just trying to figure out a way to not have an active/standby setup.

Maybe having a layer that brodcasts messages using JGroups when storing them would be better. We could have a store nothing architecture.

We could therefore also have local Lucene indices which would correctly be updated.

One tough thing to handle would be to manage to catch up on data when a node goes down and up again.

Mathias.

[reza225]

If Linux supports CARP (Common address redundancy protocol (www.openbsd.org)) you might want to check that out for a HA solution.

[Klug]

Originally Posted by reza225

If Linux supports CARP (Common address redundancy protocol (www.openbsd.org)) you might want to check that out for a HA solution.

CARP rox but will not help on the syncing between the two servers.

There's a thread on DRBD in the forum : [SOLVED] Zimbra on DRBD.

Anyway DRBD + CARP (or VRRP or any other patented system) is nice but won't solve the initial problem (active/active with HA + load balancing).

[pbrunnen]

Has anyone made any headway on this? I am in the same train of thinking... but how would you do it with iSCSI. I am testing iSCSI now, but how does that help active/active?

I am really curious! :-)

-Cheers, Peter.