On Tue, 20 Mar 2007, Erik Slooff wrote:
all,
I have an interesting observation on my mail gateway (policyd for
greylisting, postfix, amavisd-new and spamassassin); after implementing
greylisting and other measures such as RBLs there aren't enough spam
messages coming through to keep bayes trained. While bayes has worked fine
over the past two years, weekly expiry of the database now leaves me with
approx. 80 spam messages (checked using sa-learn --dump magic) which is
not enough for bayes to kick in anymore.
These are the spam stats for the past 12 months:
http://www.slooff.net/cgi-bin/mailgraph.cgi?3-e
Is this something others have also observed or is something wrong with my
setup?
We are using greylisting here as well. Our server is set up to auto learn.
It has been working out nicely for a while now (over six months). Just
recently I've altered things so bayes learning is done on a per user
basis. I have noticed it has taken time for accounts to reach the 200
minimum learning limit for ham/spam. Currently there are 17,833 rows in my
bayes_vars table and over 75,000,000 rows in my bayes_token table. There
are only six accounts that are over the 200 minimum learning. This has
been over the course of two weeks.
We are in the process of setting up a front end so the customers can do
their own processing of ham and spam. I'm sure that will speed the
learning process up for the accounts.
