Hi all
First off, a small introduction: I'm the Hosting Development Engineer for PIPEX Hosting in the UK. I have a small interest (depending how you look at size!) in NetApp kit, in that we're running them for a number of different applications.
On 10 December 2004 09:35, owner-toasters@mathworks.com wrote:
Conventional wisdom is (was) that you cannot run pop over NFS: the file/record locking is inadequate. Certainly my Qualcomm pop documentation says you mustn't use NFS.
Using file-based (UNIX spool, mbox format etc) mailboxes over NFS is bad. Using directory-based, such as Maildir format where every message is a separate file, over NFS is eminently possible - I know of several instances of very large installations using exactly this setup.
Similar issues with mail delivery agents (am I using the right term?) dropping mail into an NFS mounted directory (I use sendmail).
It depends on the format you're delivering - I've successfully used Exim 3.x and 4.x to achieve Maildir deliveries, with courier-imap (and POP) to allow users to read it.
Hope that helps
Graeme
On Fri, 10 Dec 2004, Graeme Fowler wrote:
On 10 December 2004 09:35, owner-toasters@mathworks.com wrote:
Conventional wisdom is (was) that you cannot run pop over NFS: the file/record locking is inadequate. Certainly my Qualcomm pop documentation says you mustn't use NFS.
Using file-based (UNIX spool, mbox format etc) mailboxes over NFS is bad. Using directory-based, such as Maildir format where every message is a separate file, over NFS is eminently possible - I know of several instances of very large installations using exactly this setup.
qpopper is still in use? And they still haven't put in file locking?
In 1995-96 my group implemented *and sent in* a set of patches for Qualcomm's POP server. At the time we had /var/mail on our mail server mounted from an F330 - serving over 25,000 mailboxes (and home dirs, too!). We used procmail for delivery, and made sure that our local builds of Pine, elm, and I think the Emacs mail reader all consistently used the same locking style. So /var/mail was mounted on our three user-accessible shell machines as well - and dialup users could use command line or curses-based mail readers.
This was done under SunOS 4.1.x. In 1995. We handled 450,000 emails a day on a dual-cpu Sparc 20 (in a FDDI ring with the filer). (How many dozen Exchange servers on zillion Ghz Pentiums does that translate to today?)
The only issue we ever had with the setup was clients dropping offline unexpectedly (ah, the days of banks and banks of analog 28.8 modems) and then logging back in before the lock had expired - maybe once or twice a week we'd manually remove a stale .lock file for someone. Considering the size of the userbase and the volume we handled, that was acceptable.
Not to pick on you, Graeme, but I just cannot believe that almost TEN YEARS LATER that this belief that "NFS is bad for mailboxes" persists. We solved it a decade ago, folks... true, NFS locking has never been easy or pretty, but with care it could work. Perhaps with NFS v4's locking support people will finally start to reevaluate those old NFS prejudices? (It took a *long* time for people to accept that a Netapp could be faster than local disk, too...)
And y'know, if Qualcomm never integrated our patches then, well, screw 'em. I switched to the UW POP/IMAP quite some time ago and never looked back. But now I'm just a cranky old guy who's behind the curve these days. All that Cyrus this, Exim that, blah blah blah - I'm too tired to care anymore. Besides, Real Men use 'cat' and flow control. :-)
Similar issues with mail delivery agents (am I using the right term?) dropping mail into an NFS mounted directory (I use sendmail).
It depends on the format you're delivering - I've successfully used Exim 3.x and 4.x to achieve Maildir deliveries, with courier-imap (and POP) to allow users to read it.
Just one fairly nasty side-effect of maildir format on filers: the 4K block size. Better double or quadruple the quota on your users mailboxes if you're moving them from a typical local 1K Unix FS to the filer, since all those small files will balloon way up... I've seen 700MB MH mail folders...<shudder>
Cheers,
-- Chris
-
Chris Lamb -
Graeme Fowler