In this case, mail is stored using operations to an NFS server, which is dumb as far as mail is concerned: it doesn't know anything about mailboxes, just files and directories.

This approach works very well in practice: it can be more efficient because only one database lookup needs to be made for each incoming piece of mail and each incoming POP3/IMAP connection, and it's easy to scale, by adding more MX frontends, more POP3/IMAP frontends, and more NFS server backends. (You just mount these on different directories, say /mail1, /mail2, /mail3 etc, and the database includes the full path to the user's Maildir, so you can specify which users are on which NFS servers)

The only hard part is making an NFS server which is fast and resilient. You can buy your way out of this problem, by getting something like a Network Appliance NFS server (www.netapp.com). Using their clustered machines, like the FAS270C, you have two independent NFS server boxes, each of which can take over the disks belonging to the other front-end if it fails. The performance is astounding and they are also easy to manage, e.g. you can grow volumes on a live system just by adding more disks.