Is Bayes learning for individual or for entire domain?

[Ken Morley <ke...@jmtg.com>]

I've got Sendmail, MIMEDefang, Clam AV, SpamAssassin (with Bayes support)
and Razor setup on a RH Linux 8.0 box that front-ends our domains Exchange
mail server. This seems to be working well so far, but I'm trying to "learn"
the Bayes component and have several questions.

In our environment, there are only a couple of "users" defined on the Linux
box and none of those correlate to actual e-mail users.  For example, while
there is a ken@jmtg.com email address in our domain, there is not a "ken"
user on the Linux box.  Likewise, their are root, mailadmin and clamav users
on the Linux system and none of these are valid email addresses in our
domain.  

>From reading many of the posts concerning Bayes, I'm understanding that when
sa-learn is used, it is only updating the tokens for the user currently
logged in. 

In our environment, when SpamAssassin runs, which Bayes database would it
access?  The database associated with the user that SA is running as
(mailadmin)?  If that's the case, should I log into the Linux box as
mailadmin to use sa-learn?  Lastly, am I accomplishing anything other than
wearing out my keyboard by using sa-learn?  By this I mean is Bayes creating
signatures that are unique to the specific email an individual would receive
and not effective when applied against more varied  email the domain would
receive?

Thanks very much for your help!

Ken Morley