Harris, Jason (DIS) wrote:
I'm wanting to know many times per year do SA admins have to retrain ?
In a well-maintained install, you should NEVER have to retrain unless you have a catastrophic failure that crashes your live system *and* your backups.
Regular manual training of missed spam and mistagged ham is also critical to keep Bayes healthy.
SA Version 2.64
Have you patched it with the SURBL addon? http://sourceforge.net/projects/spamcopuri/
I've been using this on three servers, and it's been a MAJOR help in keeping SA2.64 effective. 3.x versions are just too resource-intensive for these systems.
Regular feedback in the form of manual training of missed spam and the occasional (~1 every few months) mistagged ham has kept Bayes pretty accurate, too. Customer feedback gives me ~15-25 missed spams per week on this system.
This setup has been running with minor tweaks since ~SA2.44, with sitewide Bayes introduced along about 2.54 or 2.55 (I never installed earlier 2.5x versions due to the series of bugs that popped up). It's survived hardware upgrades and a cross-distro move from RH7.3 to Debian woody - but I've never had to completely wipe and retrain the Bayes database. (I had some fun getting Debian woody to recognize the RH7.3 Bayes db; I had to build a custom DB_File and force it to install over top of the stock Debian version.)
The same pretty much applies to my personal server's SA install (which I run per-user Bayes instead of sitewide). I get ~5-6 messages slipping through each week, which I collect in a "newspam" folder that I manually feed to sa-learn on an irregular "Oh look, three missed spams today" basis.
-kgd
