On Sun, Jul 09, 2023 at 07:06:10PM +0200, Robert Senger wrote:
> I've set up a testing environment that also uses master-master
> replication of the mysql bayes database, with priority in dns set to
> equal for both mx to get incoming mail distributed evenly to both
> systems. So far, this seems to work, but this is a low load
> environment.

it boils down on how much you trust mysql master-master replication
stability and performance, which is heavily dependent on your
experiences and exact versions used (are we talking about Oracle
Mysql, or MariaDB or Percona forks? which versions? What replication
setup? etc.)

I've had problems under high concurrent load (not performance, but
replication setup breaking) in the past, so I prefer to avoid
master-master replication if possible, especially if I anticipate
high concurrent load.

But if you are confident in it, sure, go ahead.

> Any suggestions?

Well, how are you training your bayes DB? If it is via cron and
manually curated ham/spam corpuses (the recommended way), I'd rather
suggest keeping databases separate and simply running training on
both servers (you can duplicate or share ham/spam corpuses as you wish,
from rsync to SMB/NFS).

If you are using auto-learn (which was not recommended last time I
looked), well, you'd probably better off NOT syncing bayes at all
IMHO, as it should be prefered that risk of bayes poisoning is
reduced to one server instead of replicating that (and there is not
much benefit, as auto-learn will quickly learn on each server
separately anyway, and if one set of domains is not getting some type
of spam, it is not beneficial to learn it anyway)

-- 
Opinions above are GNU-copylefted.

Reply via email to