I run my own mail server, but it's far from justifiable on a pure uptime basis. I don't have HA configured, so my mail (which serves <10 users on a box in colo) goes down for 2-3 minutes ~4-10 times per year when I reboot. Even worse, I don't really announce downtimes, but just randomly reboot it late at night if I've upgraded something important.
That's best-case. I have backup-mx, but no real failover for imap or normal smtp (other than reading mailspool directly on the backup mx). I used to do a crazy flood-fill thing with 3 servers smtp forwarding mail and marking it, but doing mail HA correctly is kind of hard (and then keeping mailstore synced from imap in sync, too).
How well would glusterfs work across the Internet? Doing local HA with some kind of SAN seems fine, but the thing I'm trying to solve is "entire datacenter goes down, make the user have no more pain than making a new imap connection". It seems like something that would really need to be handled inside IMAP (using some crazy backing store for mail, like a database) to be done well, but maybe filesystem level replication with maildir would be enough.
I wish there were some "best practices in systems administration" for different services, ranging from simple (just set up imap/dovecot) through complex (gmail-scale), per service.
There's been pretty good focus on how to do http, dns, and some database services, but not as much recently on smtp, imap, etc.
