oh, i didn't see it that way :]
to me, it seems that skipping autolearn emails is a bug (unless it was designed that way), thus i suggested two possible fixes. Though you seem to think this problem is not really a problem.
Well, it most certainly was designed that way. Most programs upon lock contention wait for the lock to clear and don't proceed until it does.
The SA devels very much intentionally made it skip the learning. No question about it. That behavior doesn't happen by accident.
As for how I think of it, I think of it as a very minor issue, one of so little consequence that it's not worth much effort to "fix" it.
(As has already been suggested, you can use the learn to journal option, but this will delay the entry of the data into your live bayes DB until the next journal synch occurs)
i guess it all depends if autolearn has any *real* effect or not.
It does have a real effect, but in the grand scheme of things, missing a few of them is hardly critical. I think Theo put it best when he said:
"IMHO, if you get enough mail that you have autolearning not happening due to lock contention, you have enough learning going on already."
That's really a quote worth it's weight in platinum. (ok, so hot air doesn't weigh much, but what do you expect on spamassassin-users?)
By having some autolearning you've already reaped most of the benefit. The skipping of learning on a handful of emails is a pretty trivial matter. The goal is to have fresh dynamic input into the bayes DB. If you've got contention, you're already at the point of having a good flow of fresh autolearning.
Would it be "better" if you could learn every learnable message? yes.
Would you be able to measure the difference? yes, with a sufficiently large test sample.
Would you be able to notice the difference without detailed measurement? probably not.
There's lots of other areas of work which can reap greater benefits to SA's accuracy with the same amount or less effort involved. (i.e.: getting 3.0 out the door with SPF support).
