[Mimedefang] sa-learn load, blocking?

[Kevin A. McGrail]

I've had the same problem and Nice didn't help.  I think it's just grinding
and doing so much disk i/o or something.  I did find that using SQL for the
storage rather than DB3 helped a lot at least with the disk churn and if you
are using the SQL storage, I did find that the atime in the bayes_token
table did not have an index which looked like it would help a lot.

Have you looked if an sa-learn is perhaps stalling things?  When we have a
table lock on because of a token expire, we see a high load because of all
the queued up procmails/sendmails/spamd/mimedefangs/whatever waiting to run.

On systems using site-wide Bayesian, I had excellent luck using SQL, turning
off auto-expire and running one sa-learn --force-expire nightly in cron at
an off-peak time.

However, day-by-day,  I find the benefit of bayesian rules in comparison to
the "cost" in CPU timejust hasn't been there for me overall.

Regards,
KAM

> > when feeding sa-learn hundreds of messages at a time, the load average
> > skyrockets
>
> You can try to run sa-learn with nice (or renice the process) and it will
> probably be better.