[spambayes-dev] Spontaneous training in Outlook addin?

[Tim Peters]

[Mark Hammond]
> I am yet to see an assertion error.

Me too.

> I *suspect* that it was the result of the database only being saved
> at shutdown time.

Well, I've been running three distinct saved-only-at-shutdown pickled-dict
classifiers on 3 different boxes "forever", endure relatively high email
volume, and never saw an assertion error on those either.

> Since incremental training I am yet to hear a report on this.

Good!

> Note that with incremental training/bsddb, the database is *never*
> saved at shutdown - we check, but it is never dirty!

Right, and that didn't even surprise me for an instant <wink>.

> This makes it far less likely the machine will shutdown during
> our db write.

That's an interesting theory!  I should bug Barry about this -- a modern
Berkeley "should be" saving enough info on its own to recover from premature
shutdown.  Remind me if this pops up again.

> Note, however, that the Outlook addin specifically traps this
> AssertionError [manager.py/score()] and reports the error to the
> user.  It tells them the database is corrupt and needs to be
> retrained (whereas before, SpamBayes just appeared to stop working
> for no reason)
>
> So it is unclear if the error has truly stopped, or people are
> silently re-training from scratch.
>
> Maybe I should change the text to indicate we would appreciate a
> quick mail to spambayes-dev at python.org indicating they saw the error?
> I'm inclined to believe that it is rarely seen though.

I vote to let sleeping bugs die.