Nick,
That seems to have done the trick - I reinstalled 2.63 and deleted the
old bayes tokens. I then removed the autolearn tags from my spam and ham
collections and retrained. It now seems to run in a sensible amount of time.
Thanks very much for the advice,
SA
From: Nick Leverton <[EMAIL PROTECTED]> To: [EMAIL PROTECTED] Subject: Re: SA 2.63 / 2.44 performance problems Date: Tue, 25 May 2004 10:46:00 +0100
On Tue, May 25, 2004 at 09:04:36AM +0000, not disclosed wrote:
> Dear SA,
>
> I recently changed from SA 2.44 to 2.63 and found that SA became really
> slow, ta
> king 15-30s or more on average per email while consuming 100%CPU on a fast
> linux
> PC with a fast internet connection.
> This is way slower than 2.44 and a considerable pain in the arse when you
> have a
> constant stream of spam to filter.
Did 2.44 have Bayes (I don't remember), and were you using it ? There is a known issue on upgrading to 2.6x, in that your Bayes database gets converted to the new format, and all the tokens get the same expiry time as a result. That means that when a Bayes expire is due, the code will go away and spend ages trying, and failing, to reduce the size of the database. It will fix itself when it's learnt enough new tokens to have be able to delete some old ones.
Check it by running a mail through "spamassassin -D" and look for any messages to do with bayes expiry.
If that is it, the solution is just to live with a large DB for a while. If you set "bayes_auto_expire 0" in local.cf, and run "sa-learn --force-expire" by hand (or by cron) occasionally, it will stop SA from trying to expire on-the-fly, and it should fix itself eventually (how quickly depends on how much mail you get).
ISTR you can tweak the bayes parameters to make it reach that point faster, but I can't remember (and I'm not sure it's worth worrying about :))
Nick
_________________________________________________________________ Express yourself with cool new emoticons http://www.msn.co.uk/specials/myemo
