SeattleServer.com wrote: > On Saturday 14 October 2006 09:14, W B Hacker wrote: > >>We don't let SA run SPAM_BAYES at all. >> >>Too little gain for the pain. > > > Agreed. There are better bayesian engines than SA, but after a lot of months > of pain, we decided to eliminate bayesian filtering altogether. > > Not only is in incredibly resource-intensive and demands you buy better disks > and such for your servers to handle the load, but even the best bayesian > engines can only offer 98-99% accuracy with proper training (in the real > world, with lazy users, 80% was about tops), which isn't enough, and leaves > the users having to dig through a pile of spam all the time anyways, which in > my book, defeats the whole purpose of filtering it in the first place. > > Bayesian - neat theory, bad reality. Now, what am I supposed to do with all > these 15kRPM SCSI disks? ;-) > > Cheers,
We are not *totally* down on it - Bayes works quite well when 'boxed' to the traffic of a single user - as in an MUA. But there, one has the entire local resources for just one account, so the resource load is not a big deal. As Ferdinand Feghoot said at the cannibal BBQ: "One man's meat is another man's poi, son" Segregating as narrowly as that on a multi-user MTA would probably improve the accuracy, but make the resource load even worse. Note also that several folks here have mentioned ways of managing the learning process & DB rebuild that exacerbate the resource load issue. *However* - at the end of the day, what most turned us off is that every once in a while, when the moon is full - or something far less predictable, really - Bayes 'learns' a contrarian pattern, either blessing certain spam types or condemning good traffic. Catching this before it has done harm requires intense dedication to perusing parts of traffic - individual message content, and in 'near-real time' at that - than we cannot reasonably allocate. If you think your Bayes never does this, just archive all the rejects and peruse them once in a while. The OSF will eventually surface. Bill -- ## List details at http://www.exim.org/mailman/listinfo/exim-users ## Exim details at http://www.exim.org/ ## Please use the Wiki with this list - http://www.exim.org/eximwiki/
