On 3/2/12 4:24 PM, Eric Shubert wrote:
On 03/02/2012 04:47 PM, Casey Price wrote:
Eric - do we have any plans to be using supervise and
multilog/daemontools log format for dovecot in the future? It sure is
nice when all (or at least many) of your components are using the same
logging mechanism/format, and storing them in a standardized location.
No. In fact, I expect QMT will be migrating to systemd and rsyslog at
some point in the future. I think it'll be best in the long run if we
use dovecot in the same way as it's packaged for the distro, in this
case using init scripts and syslog.
That makes total sense. I just wanted to figure out the gameplan.
Note, dovecot is very sensitive to time. This can be a little
problematic on VMs, as timekeeping is trickier there. A simple
solution is to run this script as a cron job every minute or so:
service dovecot status >/dev/null 2>&1 || \
service dovecot start >/dev/null 2>&1
I believe this is mentioned on the wiki as well, and I'm fairly certain
I'm using the exact same script if not something very close.
Not to get too far off track here, but while we are on the topic of
dovecot...has anyone had any experience setting up shared/public
namespaces in dovecot for global spam/ham folders?
I do that. Here's part of the dovecot config:
# This namespace can be used for people to dump their spam/ham into.
# A cron job can then use them as input to sa-learn.
# See qmailtoaster-plus for a sample script.
namespace {
location = maildir:/home/vpopmail/domains/mydomain.com/sa-learn
prefix = sa-learn.
separator = .
type = public
}
I'm trying to figure
out the best way to train spamassassin, and the issue at hand is that
the mailboxes are actually stored on a separate QMT machine, and my SA
boxes simply do the scanning/filtering before passing the mail to the
next hop using smtproutes. Just fishing for some ideas.
The sa-learn program updates the bayes database, using the shared
mailboxes as input. Exporting the shared mailboxes via nfs seems
suitable to me.
I think the trickier part is having multiple scanners. You'll need
some way of replicating the bayes databases between the two. I don't
know if a bayes database can be shared at all.
This shouldn't be an issue for me, as I'm not running spamd on the QMT
box which contains the users. Instead I'm currently using the following
hierarchy:
--> POP
GW1 --> SA1 --> Q2
--> vCluster1
(I'm currently building a new server that will eventually take the place
of GW1 & SA1 and consolidate them into one host). My real problem here
is that the GW & SA boxes don't actually know which users exist...they
simply have all of our domains listed in rcphosts & goodrptto, and then
smtproutes is setup accordingly to pass mail from GW1 to SA1, and then
SA1 passes it on to one of the 3 servers listed above. Obviously this
isn't the ideal configuration, and I'd like to change it...but that is a
major overhaul, and I'm trying to gradually improve my infrastructure.
Ideally I would like to follow Bill Schupp's Qmail ISP Array model that
Eric posted a link to awhile back...but it will be a little while before
that can happen. So in the meantime, I'm trying to at least find ways to
automate the learning of spam/ham from a remote mailbox to my SA boxes -
if that makes sense?
So the question is do I create a namespace for each domain on each of
the 3 servers, or can I share one for all domains on each server, or
finally - can I create a shared namespace somewhere that all 3 servers
could access?
Sorry to complicate things :-P
Casey Price
Smile Global Technical Support
Submit or check trouble tickets http://billing.smileglobal.com
www.smileglobal.com <http://www.smileglobal.com>
Follow us on Twitter <https://twitter.com/#%21/SmileInternet>
Find us on Facebook <https://www.facebook.com/smileglobal>
