On 10/30/06, <b class="gmail_sendername">Joshua Goodall</b> <<a href="mailto:joshua_goodall@pacific.net.au">joshua_goodall@pacific.net.au</a>> wrote:<div><span class="gmail_quote"></span><blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;">
... <br>
</blockquote><blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;">I recommend a pair of NetApp filers (e.g. 2 x 3020c) exporting NFS from<br>shared disk shelves in their high-availability mode. This is a hybrid
<br>of the two options, except a) it's a fully supported and stable NetApp<br>configuration, b) they're sharing disk storage, c) you're not hog-tied<br>to GFS, d) it's the best NFS implementation in the world (you can even
<br>trust the fcntl locking), e) it's really fast, f) NetApp's onboard<br>snapshots and other funky software bits will make you wonder how you<br>lived without them, g) the HA actually works.<br><br>I have a six-digit userbase with mailboxes in this configuration.
<br><br>JG<br><br></blockquote></div>We
are in the process of rolling out / migrating mail from various "other"
imap stores to a cluster of two servers running dovecot-1.0.beta9
attached to a 3050c cluster. Hosts are running RHEL4u3 on x86_64
on the latest official kernel, 2.6.9-42.0.3.ELsmp. NFS mount
options on both machines are
(rw,nosuid,nodev,tcp,nfsvers=3,rsize=8192,wsize=8192,soft,intr,noac,actimeo=0).
We are using default_mail_env with Maildir/indices stored over
NFS. Postfix with procmail delivery. Currently we are using
DNS round robin as there is no SLB in the mix just yet. It is
extremely fast.<br>
<br>
The only thing I am still mildly concerned about are some errors
regarding index corruption. These are seen on both
machines. Hopefully a newer version fixes these. See below:<br>
<br>
dovecot: Oct 31 11:56:52 Error: IMAP(user): Corrupted index cache file
mailstore/user/Maildir/.My-Folder.blah/dovecot.index.cache: invalid
record size<br>
<br>
These errors don't appear to have any impact on the client side, performance or otherwise.<br>