Michael -
I'd be happy to. I've copied the list since someone else may find the additional detail useful.

Michael Bellears wrote:
Hi Nick,

What I do is run a local read-only copy of the 
mysql database on each of my qmail servers, which are 
replicated off of 
a master read-write copy.
Delivering roughly 500K messages per day to this cluster I'm running 
three servers for qmail, vpopmail, courier-imap/pop3, 

Sounds nice - We are about to migrate to a multi-server environment, and
was wondering if you could answer a couple of questions for me.

Are your servers sitting behind a local director or equiv.?

Are you running a common nfs share for your clients mailbox's (i.e.

What happens if your master mysql box fails?

Thanks in advance,
We're using a redundant pair of Foundry ServerIron load balancers, though I strongly discourage you from buying such devices. My experience is that they are very unbalanced in their load balancing, and their interfaces (web and cli) are hideous for every day usage. Other experiences of mine say to buy a pair of cheap linux boxes and use Linux Virtual Servers (LVS) as a load balancing interface. Its far more configurable to balance based on load, and is far, far cheaper to deploy.
We're mounting /var/vpopmail/domains as an nfs mount for all of our users maildirs from a pair of Sun E250s (which also run the MySQL read/write DB). These are hooked up to a mirrored pair of Raid5 arrays, with the mirroring being done by Veritas Volume Manager, and the boxes are clustered with Veritas Cluster Services for high availability. Again I'd not actually recommend using this setup unless you've got these or similar pieces already paid for and lying around. Either  real filers (such as Network Appliance, Procomm or others) or linux boxes using tools from the Linux-HA project are what I'd recommend. I'm looking to replace our E250s with a pair of NetApp F825c filer heads with clustering and the whole ball of wax as high availability is an absolute necessity in my network.
As I mentioned previously the MySQL boxes are clustered, and the MySQL database itself is on shared SCSI storage (part of the volume manager setup).
When I replace the E250s I'm thinking about instead setting up a circular MySQL replication scheme, since MySQL tags replication data with the originating server, so once it makes its way through the circle it won't be applied again. My only problem so far is that this would end up like a token ring network, in that one dead link and the whole thing is unusable.  Otherwise I'll end up using linux-ha and some perl scripts to dynamically reconfigure a pair of servers as Master/Slave (when the master crashes, the slave becomes the master, and when the previous master comes back up it automagically knows to become the new slave).
Let me know if there's any thing else I can answer.
Hope this helps,
Nick Harring
Webley Systems

Reply via email to