> I am looking into switching from a global bayes/awl/setting environment
> to a per-user environment with MySQL as a backend.
>
> <puts on asbestos suit>
> Would anyone care to offer an opinion as to whether and/or to what
> degree this might make in overall effectiveness? Anyone back up that
> opinion with cold hard facts?
>
> Will I be able to migrate small sets of users from global to per-user or
> will I have to make the jump for all my end-users/domains at once?
I have a suggestion to spare for the "setting" environment which works pretty
well to me and would avoid the per-user/global question.
My amavis settings are in a postgres db in which each organization (more or
less = domain) has a schema. Each organization has a table with user-defined
settings. My public schema (i.e.: the default one) has a table too with
organizational settings. Also, the public schema has a view which is the mean
by which I get the amavis settings for a given user. It attempts fetching the
per-user settings table in the organizational schema and, if they are missing,
it attempts fetching the per-organization settings. If this too are missing, it
uses some default which may be tought as "local" (ie.: server-wide) settings.
I did find this three-layers way of handling settings pretty useful: a user
wants to get .exe attachments without having them wrapped into a warning
message? Put a record in the per-user table. An org wants to have treats
reported to a specific user? Tell it to the per-organization table.
However, I don't have bayes data on the db (it's in the global bdb). This is
because most of my customers use pop3 to download messages, so I have quite no
way to train bayes efficently. Do I?
Also, awl settings are in a database only to ease their adjustment (and, in
future, replication), but are global as well: I'm serving small communities in
my town, so often a source ip/e-mail scores may be reasonably used for all my
potential destinators.
Cheers,
giampaolo
> I'd like to preload the bayes db for each user so that's it's 'primed'
> and ready to do. Obviously, it would be preferable to preload with
> their specific mail but is it possible to feed bayes for each user with
> a generic set of spam/ham?
>
