Anyone? This isn't urgent or critical, but it's bugging me because it
changed from "working" to "not working" spontaneously.

Kris Deugau wrote:
> I have a pair of servers with global Bayes DBs that no longer remember
> which messages they've learned, for no apparent reason I can see. I
> can't think of any particular changes I might have made that might cause
> this.
> A third machine with per-user Bayes seems to learn fine for at least one
> user account (mine - I have a cron job that learns a newspam folder
> every night). I typically leave messages in a newspam folder until I've
> accumulated about 100 before deleting them; recently that's been taking
> several weeks. Learning via sa-learn is done manually on the
> misbehaving systems, usually once a day or every other day. (Aside from
> cron-based learning, this applies to all three systems.)
> Comparing debug output between the three machines shows no obvious
> processing differences aside from the different system names and Bayes
> pathnames.
> All three machines are running SA3.1.8, from the same package repository
> (RPMForge), although the system that's working is currently running
> CentOS3 where the other two are currently on CentOS 4. (They've all
> been moved from one physical box to another and had services migrated to
> new OS installs, versions, and distributions, all more than once. Other
> small ISP admins are no doubt familiar with that dance. )
> I thought it might be something sa-update had pulled in, but removing
> the update files on one misbehaving system didn't change anything.
> bayes_learn_to_journal is set on the two misbehaving systems, but it's
> been set since I originally upgraded from SA2.44 to 2.54 on both (IIRC).
> Expiry is disabled, and both systems have had a daily cron-based expiry
> run (again, set that way since SA-with-Bayes was originally set up on
> these machines, and working since originally configured). Autolearn is
> also enabled and tweaked (lower threshold set at -0.1); it's the only
> way I can really gather much ham on these systems. :/
> Odd not-remembering-what-we've-learned issue aside, learning (automatic
> and manual) seems to happen - single messages run through SA again after
> learning show different Bayes scores.
> Any suggestions on what to look for in the debug output?
> Anyone want to wade through pages and pages of debug output?
> -kgd