Bayes stats expiration and multi-level checks

[kobuki]

Hi, new to rspamd, switching my first system from SA + Dspam over to rspamd only. So far, it's running in testing mode (no mail gets moved/rejected yet, based on its scores) and has pretty good accuracy. Now I want to implement something similar to what Dspam provides: use a global stats db, so I can have a base for all users, and have a per-user db that users can individually train (via Dovecot antispam plugin). I can create the 2 databases per the docs, but I have a few questions.

1. How should I name the global and the per-user stats? Are they consulted sequentially, producing scores individually that get added to the final score?
2. What is the expiration on the Bayes databases? Fuzzy hash stores have an expiration, I'd like to employ one on the Bayes databases as well. Keeping those tokens forever makes little sense to me.

Do you have any suggestions?

[Felix Schwarz]

Am 06.12.2017 um 18:43 schrieb kobuki:
> 2. What is the expiration on the Bayes databases? Fuzzy hash stores have an
> expiration, I'd like to employ one on the Bayes databases as well. Keeping
> those tokens forever makes little sense to me.

IIRC there is no expiration for bayes tokens right now.

Felix

[kobuki]

That's too bad. Something must be done with it on the long term.

Anyone else has ideas for the other question?

[kobuki]

Could anyone please help with the question related to my original post, without answer?

How are the scores calculated when 2 classifier classes exist?

I've created a per-user and a global bayes stats db. They both get trained without (global) and with (per-user) user name. How is the score calculated in this scenario? Do I need to add some Lua function, do they get added, averaged, or the first that matches is applied?

Thanks in advance.