Dan Bongert wrote: > I recently did an email server change/upgrade from Sendmail on > FreeBSD (w/ Spamassassin 2.6.4) to Postfix on RHEL 3 (w/Spamassassin > 3.1). On both systems, Spamassassin is called from user's .procmailrc > files--not every user wants to be running SA (I'm not quite sure why). > > I wasn't able to convert people's Bayes databases from one system to > the other--the Linux system didn't recognize them at all as valid DB > files, so everyone had to start Bayes over from scratch. > > Here's my problem: the new SA doesn't work nearly as well as the old > one. Some of my users are reporting 50% false negatives in their > inbox in the morning, even after their Bayes autolearning has kicked > in. We run a nightly learning script for them, and have been telling > everyone to put any and all false negatives in the appropriate > mailbox so that sa-learn can snag them. > > For my own experiences, I'm seeing a lot more spam that's being > autolearned as ham--scores of 0.0 and even negative ones for things > that to my eyes are very obviously spam. > > It's a pretty vanilla set up so far--are there any recommended > optional rules sets or tweaks I haven't discovered for 3.1 yet? > Unfortunately, I don't have any hard numbers comparing the set ups, > just lots of complaints that the new version isn't as good.
Sounds like you've got some configuration issues. Take a look at your local.cf, init.pre, and v310.pre files and see if you see anything obvious. run "spamassassin --lint" and make sure you don't see any errors. Take a look at the headers of some of the emails. If you see hits for ALL_TRUSTED on any emails from outside your network, you need to fix your trust path (trusted_networks). http://wiki.apache.org/spamassassin/TrustPath As for recommended rule sets, I run most of the SARE rule sets. http://rulesemporium.com/rules.htm Make sure your network tests are working. Razor2, DCC, and Pyzor can also make a big difference. Once you get this sorted out, nuke your Bayes databases to get rid of all of the bad learning and start over learning ham and spam. -- Bowie
