I have two mailservers running SA. On one (with a very small number of users), there is a shared bayes database (bayes_path /var/work/bayes/bayes in local.cf), and in the other one I use the default per-user databases.
In both machines I run sa-learn as a weekly cron job on spam in a "Spam" folder and ham on the users' inbox - on messages between 1 and 2 weeks old - on the assumption that users categorise their mail properly. The machine that uses the per-user database, there are much more hits on the BAYES_99 rule (it tops the list) compared to the other machine, where it is way down on the list. I therefore get the impression that this rule works better on the first box. My questions are: Should I use the shared database, if it doesn't seem to work as well? - or has something else gone wrong. Is it necessary to run sa-learn regularly, when SA auto-learns ham/spam anyway (except of course when I'm setting it up when I need to run it to initally set up the database)?
