On Wed, 11 Dec 2019 08:59:53 -0500 Bill Cole wrote:
> It's infeasible and would be unwise to put a uucode detector/decoder > into SA. There's no limit to the corner cases that can arise with a > sloppy format that has an unknowable number of mostly unmaintained > implementations which uses lore and local habits in place of > specification. I think that's overstating it. It could probably be implemented so it works or falls back to the status quo. But it's the status quo that's most wrong. I think it could have been a lot worst if the backend had been faster and/or the the attachment smaller so that it completed. In which case training could have swamped the Bayes database, expiring the useful tokens. If Redis does its internal LRU expiry promptly, it's the most vulnerable to this. With manual training a separate limit can used in sa-learn or spamc. With autotraining I think anything that's small enough to scan can be trained.
