[Dovecot] 1.0rc8 status report

[David Lee]

A quick status report on how 1.0rc8 behaved in service for a few hours
with several hundred simultaneous users, at a site very new to dovecot.

Oh, and a question at the end.

Summary: Reasonable for a first shot but one significant problem,
requiring backing off.


Background:

We have a long-established UW-IMAP service for a user population of about
20,000 based on a few Linux (Redhat) machines running IMAP/POP and inbound
delivery.  We try to ensure, but cannot guarantee, that all activity for a
given user takes place within one machine.  Each machine mounts the INBOX
area ("/var/spool/mail"; traditional UNIX mbox) via NFS with tight NFS
arguments ("noac,actimeo=0", etc.) and similarly mounts the users' folder
areas which are subdirs of their home directories.  (We know that Mark
Crispin recommends against NFS for UW-IMAP, but we seem to have been OK.)

There is also some processing:  .forward->"| procmail"->folder-or-inbox

Each machine typically has several hundred simultaneous IMAP connections.

This has basically worked well, but the UW-IMAP loading has been heavy.



The plan:

In an ideal world, I would like to restructure the above.  But our world
is not ideal, so we have to stay with the structure.  But we are looking
for a transparent (user perspective) migration to dovecot.



The dovecot experience:

Yesterday, I quietly adjusted a DNS entry to redirect one of the live
email hostnames at an additional machine in the "farm", running dovecot
1.0rc8, including deliver/LDA (and taking into account some post-rc7
dovecot changes in this area).

On an earlier, smaller-scale test, one problem had been some periods of
"temporary authentication failures".  Increasing "login_processes_count"
and "login_max_processes_count" (each by a factor of 8) seems to have
fixed this, and I'm not aware of any problems in that area yesterday.

It basically went well.  But just over two hours hours later I had to back
off, because of a significant dovecot problem, namely that dovecot
crashed, almost silently.  The only traces of this event in the log file
seem to be:
   Oct 10 16:26:12 [...] dovecot: child 24525 (login) returned error 89
   Oct 10 16:26:14 [...] dovecot: Login process died too early - shutting down

Any thoughts?  Any fixes?  If the problem needs debugging (or additional
data/log collection) how might that be attempted in this environment?



-- 

:  David Lee                                I.T. Service          :
:  Senior Systems Programmer                Computer Centre       :
:                                           Durham University     :
:  http://www.dur.ac.uk/t.d.lee/            South Road            :
:                                           Durham DH1 3LE        :
:  Phone: +44 191 334 2752                  U.K.                  :