Le mar 18 juillet 2006 09:34, Lionel Elie Mamane a écrit : > This will still include legitimate mail. something like 50 over 300k is less than 0.016%. which is also really less than the usual number of false positives of your bayesian mail filter. see end of mail. > > and if you never actually realized, there *IS* such a slowdown on > > debian mail lists, it's called crossassassin, it kills master on a > > regular basis, and is *REALLY* less effective than greylisting. > > I don't remember the "master cannot cope under mail load, we need > desperate measures" point being brought up before. I may have missed > it. these days master has a high load on a regular basis: load average: 239.68, 299.68, 326.84 from IRC a couple of days ago, What I experience as a debian developer is that: * 80% of the overall spam that eventually comes into my inbox went through my debian.org account, that renders the read of such a mailbox really hard, and I'm pretty sure that I miss more than 0.016% of legitimate mail in my readings. * my @debian.org address has considerable slowdowns due to our MXs beeing overloaded from time to time. 80% of the time, it's because of crossassassin becoming mad, or some spam attack. Just take some factual numbers: I receive sth like 300 mails a day (top, I think the mean value is more around 150). that makes 109.500 mails a year. I know for a fact that my bayesian filter makes sth like 4 to 5 errors per year. And yes I know how to train one. So my bayesian mail filter has at least a 0.05% false positive rate, and I'm really convinced in fact it's more like 0.1% (maybe even more). SA is used extensively on debian hosts, I'm also quite sure it also has worse rates than a 0.1%. So you are claiming that greylisting is a really bad method ? come on ! currently, I receive so many spams from debian, that I just CANT sort them. it's sth like 90spams a *day* sometimes. How do you find the time to look at the good mails in there ? I can't. So by not delaying 0.016% of the legitimate mails, you make a lot of people *LOOSE* for real way more than that. please, your point is only made of impressions, now you have numbers. -- ·O· Pierre Habouzit ··O madcoder@debian.org OOO http://www.madism.org
Attachment:
pgpJtu6ajfaYI.pgp
Description: PGP signature