Bays DB connection and sharing

[DerDanilo]

I think what PMG is missing is an external bays DB connection if the mail server do the bays training since all mails are stored there.
This is quiet simple to setup via CLI but has no official support in the UI and possible the pmg logic and system itself.
The PMG quarantine features are not always needed or wanted. So this doesn't always work. Training the bayes needs to happen somewhere else.

@proxmox staff
Are the Bayes shared between all nodes and pushed to new nodes if the cluster is extended for some reason?

There could also be a simple DB cluster shared between all pmgs that. On some projects we are using an already existing mariadb cluster for that.

 

[Stoiko Ivanov]

Are the Bayes shared between all nodes and pushed to new nodes if the cluster is extended for some reason?

The bayes database are local files in /root/.spamassassin, which are not shared across a cluster.

From our experience most deployments to not benefit too much from bayes training (more often than not it leads to many false positives) - so it's not on our roadmap to change this for now ( we would need at least a few example setups where bayes filtering indeed did improve the spam detection with solid numbers and the learning strategy)

 

[CRCinAU]

I noticed a lot of spam getting through when setting up my PMG install for the first time... I've only been using PMG for ~3 days - and its mostly been positive - but my bayes DB from my old server has many years of training and hardly let any spam through...

I'm hoping that my guide here helps learning in some setups: https://forum.proxmox.com/threads/how-to-train-for-spam.41122/#post-310567

 

[heutger]

The bayes database are local files in /root/.spamassassin, which are not shared across a cluster.

From our experience most deployments to not benefit too much from bayes training (more often than not it leads to many false positives) - so it's not on our roadmap to change this for now ( we would need at least a few example setups where bayes filtering indeed did improve the spam detection with solid numbers and the learning strategy)

@Stoiko Ivanov Would be great to sync Bayes DB as well. For sure, it's currently hard to train via PMG. Quarantine does not provide to learn the still one PMG placed spam as spam or ham as ham to PMG and once the mail is delivered to the destination server it needs coding to get the message "back" to PMG for training (e.g. rspamd has here input fields to learn spam and ham, but that's also not nice). The script provided by @CRCinAU should help or a spam and ham mailbox been read and learned or finally a sieve script be invoked by moving mails to spam or out of spam. However, that's why people asked here about integration with archiving or a mail server solution as then all mail will be handled on one system and improving bayes DB would be much easier and bayes is really great ("Das Zünglein an der Waage" after blacklists and improved SpamAssassin).