Re: Can't figure out how to "aggregate" the spam training for aliased users

On 3/15/2025 07:48, Matus UHLAR - fantomas wrote:
> wOn 13.03.25 09:34, Karl Denninger wrote:
>> I'm sure this has come up before but I don't see a way to do it in 
>> the docs or Wiki.
>>
>> I have a number of "role" accounts that, in some cases, get email 
>> from more than one domain.  All are considered "local" by my postfix 
>> and dovecot config, thus all the mail from those domains gets dumped 
>> into the user account related to same, with the "role" accounts in 
>> the /etc/aliases file.
>>
>> For example, I have an "editor" user that is not really a user, but 
>> is in the aliases file to this account, and said "editor" address is 
>> valid not only for denninger.net, but also for a couple of other 
>> domains that I run a web property for on behalf of someone else.
>>
>> If someone spams that "editor" user Spamassassin will use its 
>> built-in rules -- but it does /not /honor the Bayesian classifier 
>> training that my account ("karl") has, and so if I put a spam that 
>> got through in "Junk" while the classifier will do its job if someone 
>> spams "[email protected]" later with that same classified content 
>> /it won't look at the classifier data all for a future email that 
>> comes to "[email protected]."/
>
> If you use spamd, it tries to use user (provided by spamc client) and 
> its home directory used for bayes database etc.
>
> ...unless you call spamd with "-u user", then it always uses that user.
>
> You can specify "-U username" to use as default user if provided does 
> not exist.

Aha.  That appears to be the magic for it.

I'm using spamd, yes, and database backend -- and since the "target" 
username for the email doesn't actually exist as a "real user" (its just 
an alias) and the classifier is trained by the "ham" and "spam" scripts 
which run against each individual user's junk folder, it had nothing to 
match against.

Now, for aliased things which go to a singular real role account, it does.

This would not work if you had aliases that went to multiple places 
along with real user accounts (then you need a way to resolve the alias 
first and pass it as the "user", which is where I thought I was missing 
something) but as long as the "role" accounts all point in the same 
place the "-U" switch aims the bayes classifier check there in the 
instance where the role account is not "real" yet all of them go to the 
same place.

Thanks.

-- 
Karl Denninger
[email protected]
/The Market Ticker/
/[S/MIME encrypted email preferred]/