----- Original Message -----
From: "Matt Kettler" <[EMAIL PROTECTED]>
Also, with that much mail coming in at the same time, there could be
contention for bayes locks. You might try adding
"bayes_learn_to_journal 1" to your local.cf, and see if that helps. This
will cause learning to be done into a "journal" file which periodically
gets merged into the main bayes DB. This causes the live bayes to be
delayed in update until the next sync (once a day or every 100k of bayes
data by default), but you can force-sync any manual training runs by
running sa-learn --sync afterwards.
I don't have "bayes_learn_to_journal 1" in my local.cf. But I see
bayes_journal file in the bayes directory. So it must be default
behaviour. As I had written , I do sa-learn --sync every hour .
By default SA does not learn to the journal. AFAIK, it only updates
atimes in the journal, by default.
Looking at BayeStore/DBM.pm, tok_touch always uses defer_update, even if
learn_to_journal is disabled.
I noticed another thing which might give a clue. Although my smptd
connections were maxed 100/100, the cpu load was still low ~0.4. But when I
saw the load when the hourly sync was going on, the load was high ~5-7 even
though the smtpd connections were not maxed. I have set Auto_expire to 1.
Could that be the problem ? Should I set auto_expire to 0 and
do --force-expire thro a cronjob and check ?
Anand
