On Wed, 2010-03-17 at 12:15 -0700, Mark Moseley wrote:
- Our #1 main motivation for looking Dovecot is relief for our currently overtaxed NFS servers, mostly in the form of the index files. Benchmarking dovecot looks great, even with the index files in the maildir.
Have you read the thread starting from http://dovecot.org/list/dovecot/2010-January/046106.html and spanning a month or so? That provides a good view of potential problems with NFS.
I'd read through that thread and subsequently forgot the implications. Presumably the same pitfalls apply to 2.0?
Yes.
From the wiki and from the thread, it sounds like this just affects index files. One thing I didn't see in the thread (though it'd be easy to miss in a thread that long) is whether performance suffered horribly with 'noac' on a *separate* index mount (the mentions I can find where they tried 'noac', it was on the entire mail store, not a separate index mount) -- though doing 'noac' on anything seems like a recipe for disaster.
noac increases the NFS traffic for indexes by something like 10x, regardless of how it's mounted.
Another extreme option might be to just keep indexes on local disk only. Obviously users hitting another server wouldn't benefit from the other server and it'd incur updates to the indexes but they seem somewhat minimal -- in sane cases at least where the number of newer-than-the-index messages is reasonable. With aggressive load balancer stickiness at maybe a /24 netblock level, I'd be able to keep people localized for a little while (with no way around people accessing the same mailbox from home and work). Am I underestimating the costs of incremental index updates on local disks? In testing, tt seemed like dovecot was pretty conservative in disk accesses/reads when updating. I'm hoping the 95%-of-the-time worst case will be someone hitting the same mailbox from two locations, so they'd be hitting at most two servers. With local index caches, I'm mildly sure I could live with that (esp if I can talk my way into an SSD on each one). Am I being too naive?
I don't know if anyone has run local indexes in larger setups, so I can't really give any good answers. The worst case of rebuilding the whole index isn't anyway any worse than what Courier has to do every time.
I'll play around with that more. So it should be possible to have more than one private 'hidden=no; list=yes' namespace that'd show up in a mail client's subscribable folders list?
Yes. And probably hidden=yes is better, since it doesn't really need to show up as a separate namespace for clients (even though most would ignore it anyway, but some might show it a bit weirdly).
Offhand, before I start sifting through source code, is Dovecot on a 32-bit system limited to a 32-bit -- i.e. 2 gig -- limit on quota size like Courier is?
Quota is always 64bit. And Dovecot usually tries to enable 64 bit file offsets for 32 bit systems too, and then there aren't any 32 bit file size limits anywhere.
If we're going to have to live with users complaining about a one-time redownload of just post-conversion mail, I'll need to get started convincing the higher-ups that that's life.
Check what the new Courier POP3 UIDLs look like. If all of them are maildir base filename, setting pop3_uidl_format=%f should make this conversion easy. Just run it through once to get old UIDLs converted and then you don't need to worry about it, because both Courier and Dovecot give the same UIDLs. But I never really understood how Courier assigns new UIDLs, sometimes it seems to use the filename and sometimes not..
So do you mean keep pop3_uidl_format=%f basically forever? That'd work for me, but are there any big implications, performance-wise or other, of not using the default "%08Xu%08Xv"?
%f is at least more reliable, because it's not fully guaranteed that UIDVALIDITY+UID never changes (although in practice they don't). If they do, then POP3 UIDL changes and clients redownload mail. The only disadvantage is the somewhat larger UIDLs causing a bit more and network bandwidth usage.
It's also possible to set pop3_save_uidl=yes, and that's probably a good idea also, so that whenever UIDL is sent to client, it's also saved to dovecot-uidlist. That allows you to later change pop3_uidl_format without changing existing UIDLs.
Does that require a corresponding change to the conversion script?
No.
So I'd have Dovecot use the Courier-style UIDLs, run the conversion to get the old UIDLs (i.e. anything not using the Courier "%f", which is a bit) into dovecot-uidlist, and then not worry about any new mail, since clients will see the same UIDLs (and dovecot will just update dovecot-uidlist with those UIDLs the first time they log into Dovecot's POP3)?
Yeah.