[Rspamd-Users] Migrate bayes from sqlite to redis, learns are lost

[Tim Harman]

On 31/07/2019 8:17 am, Aki Kyo wrote:
<snip>
> If I must reset all bayes to zero, it's not desired but I could live
> (how would I purge all stats from redis?) Previous sqlite learns were
> from autolearn[0,9] but as advised on mailing list I change now to
> autolearn[-20,20] so probably learns will be infrequent. What is best
> alternative to autolearn... how about feeding messages coming to not
> existing accounts on my domain (I get a fair amount of those
> attempts)?

Well, I can't answer why the migrate didn't work.  And while I have 
removed bayes information before using redis-cli I can't remember 
exactly my command(s) and I don't want to suggest things in case I'm 
wrong and nuke your redis database.

As for your autolearn: why -20,20 though?  That seems very unlikely to 
learn much, especially on a small mailserver.
I have a small mailserver (150-200 emails a day) and I have set 
"autolearn = [-2, 15];" which works very well *for me*

I would advise you to look at your mail coming in using the rspamd GUI 
and figure out what numbers you need to know "this is definitely ham" 
and "this is definitely spam"

I suspect -20,20 is great on a mailserver processing thousands of emails 
a minute, but if you're only getting 100 emails a day like me, tuning it 
a bit closer to your limits seems to make sense.  You can check every 
day or two and if mail is being incorrectly learnt, you can manually 
step in and change it using the webgui learn feature, or rspamc.  That's 
assuming you still have the entire email and haven't forwarded it on out 
of your control.

You can also use rspamc to learn your existing ham and spam corpus 
quickly (I used this when coming from spamassassin.  I had all my old 
ham/spam so I just fed it to rspamc)

Feeding in emails to domains you don't host is _probably_ a good idea, 
if you know *for sure* all those emails are really spam.

Tim