RBL checks and milter

[sp...@sasknow.com]

Hi all,

I've been using spamass-milter in production for quite some time now, on
a single FreeBSD/sendmail server that sustains 2,500 messages per hour
during peak times. I'm currently using Mail::SpamAssassin 2.61.

With RBL checks enabled, spamc takes several seconds (7.8, on average)
to process each message. I expect it to. However, the problem is that
when run as a milter, the sending MTA is forced to wait, and, in some
cases (i.e., when spamc needs somewhat longer than 7.8 seconds), this is
causing timeouts with delivery, and consuming needless resources
(sockets and RAM) to filter spam in "real time", when I'll never have
any real-time status to report to the sender. With RBL enabled, the
system becomes extremely sluggish (load > 10.00) and starts rejecting
connections (451 4.7.1 Try again later), timing out, or just ignoring
the filter (thanks to "F="). Without RBL, it's very snappy, and load
averages hover < 1.00.

I'm also using clamav-milter, so here's the relevant M4 (mind the line
breaks). Nothing special:

  INPUT_MAIL_FILTER(`clamav',`S=local:/var/run/clmilter.sock, F=,
  T=S:4m;R:4m')dnl
  INPUT_MAIL_FILTER(`spamassassin', `S=local:/var/run/spamass-milter.sock,
  F=,T=C:15m;S:4m;R:4m;E:10m')

For now, I've disabled RBL checks with spamd -L, and the combination of
clamav and spamassassin takes much less than a second for the same
messages tested with RBL enabled, which is definitely acceptable. I'd
really like to continue to reap the benefits of RBL, though!

Where should I look next?

Thanks,
- Ryan