Trying to get per-user statistics working, what am I doing wrong?

[Leonardo Boiko]

My /etc/rspamd/override.d/statistic.conf:

# global classifier with per-language tokens
classifier "bayes" {
    tokenizer {
        name = "osb";
    }
    cache {
        path = "${DBDIR}/learn_cache.global.sqlite";
    }
    name = "global";
    min_tokens = 11;
    min_learns = 200;
    backend = "sqlite3";
    languages_enabled = true;

    statfile {
        symbol = "BAYES_HAM";
        path = "${DBDIR}/bayes.global.ham.sqlite";
        spam = false;
    }
    statfile {
        symbol = "BAYES_SPAM";
        path = "${DBDIR}/bayes.global.spam.sqlite";
        spam = true;
    }
}

# per-user classifier
classifier "bayes" {
    tokenizer {
        name = "osb";
    }
    cache {
        path = "${DBDIR}/learn_cache.users.sqlite";
    }
    name = "users";
    min_tokens = 11;
    min_learns = 200;
    backend = "sqlite3";
    # languages_enabled = true;
    users_enabled = true;

    statfile {
        symbol = "BAYES_HAM";
        path = "${DBDIR}/bayes.users.ham.sqlite";
        spam = false;
    }
    statfile {
        symbol = "BAYES_SPAM";
        path = "${DBDIR}/bayes.users.spam.sqlite";
        spam = true;
    }
}

It does create all six files.  Then:

$ echo 'select count(*) from tokens;' | sqlite3 bayes.users.spam.sqlite
0
$ echo 'select * from users;' | sqlite3 bayes.users.spam.sqlite
0|default|0

$ rspamc --classifier=users --deliver=example@example.com learn_spam < ~/sample_spam.eml
Results for file: stdin (0.031 seconds)
success = true;

It does seem to learn something:

$ echo 'select count(*) from tokens;' | sqlite3 bayes.users.spam.sqlite
707

However, the username column seems to be empty?:

$ echo 'select * from users;' | sqlite3 bayes.users.spam.sqlite
0|default|0
1||0

# trying another user
$ rspamc --classifier=users --deliver=example2@example.com learn_spam < ~/sample_spam.eml
Results for file: stdin (0.007 seconds)
HTTP error: 404, <AU20160712140011.00003-70961.00016-7159319@168.90.190.39> has been already learned as spam, ignore it

I'm using rspamd 1.2.8-1~jessie on debian jessie.

[Leonardo Boiko]

Update:

$ rspamc --classifier=users -r example@example.com learn_spam < ~/sample_spam.eml
Results for file: stdin (0.035 seconds)
success = true;

$ rspamc --classifier=users -r exam...@example.com learn_spam < ~/sample_spam.eml
Results for file: stdin (0.035 seconds)
success = true;

$ echo 'select * from users;' | sqlite3 bayes.users.spam.sqlite
0|default|0
1||0

2|example@example.com|0
3|example2@example.com|0

It seems that the option affecting per-user Bayesian training is '-r' and not '-d' (contrary to the documentation in the manpage and https://rspamd.com/doc/configuration/statistic.html )?