problems with bayes

[Bill Omer <bi...@spis.net>]

Hello list

I am having an issue with bayes.  I'm using MailScanner which invokes
Mail:SpamAssassin (a mix of 3.0.3 and 3.0.2).  This is being ran off an LVS
cluster consisting of 3 real servers.  I have been able to get bayes to use
MySQL so all the nodes can share the same bayes db.  This has been verified
by running MailScanner and spamassassin in debug mode.

The thing is, when I have bayes disabled I get a lot of false negatives.
But when I enable it, I get even more.  For some reason, nearly every single
piece of spam that gets through gets a -2.60 from BAYES_00     I have run
sa-learn across thousands of known spam messages but to no avail.  I'm not
sure what else the problem could be.  I have cleared the bayes db (sa-learn
-clear as well as re-creating the mysql tables) and that seems to help, but
then it just keeps getting worse and worse no matter how many messages
sa-learn learns from.   

I assume this is caused by some form of bayes poisoning.   Are there any
suggestions on how to get around this?

-Bill Omer