[Rspamd-Users] Bayes and Redis on multiple mx servers

[Vsevolod Stakhov]

On 17/02/2021 13:28, Sandy Drobic wrote:
> Hello,
> 
> I have rspamd with redis as backend on several servers in our company. During
> testing I discovered that the trained bayes db was only used on mails on the
> primary server, although the redis replication does replicate the bayes db to
> the other servers.
> 
> local.d/classifier.conf:
> 
> backend = "redis";
> 
> read_servers = "localhost:6376, mx-srv1.example.com:6376";
> write_servers = "mx-srv1.example.com:6376";
> 
> new_schema = true;
> expire = 86400000;  # 1000 days
> autolearn = [-1, 12]
> min_learns = 200;
> 
> The "rspamc stat" shows that bayes was properly replicated.
> 
> [root at mx-srv1 tmp]# rspamc stat
> Results for command: stat (2.000 seconds)
> Messages scanned: 443
> Messages with action reject: 10, 2.25%
> Messages with action soft reject: 0, 0.00%
> Messages with action rewrite subject: 37, 8.35%
> Messages with action add header: 4, 0.90%
> Messages with action greylist: 7, 1.58%
> Messages with action no action: 385, 86.90%
> Messages treated as spam: 51, 11.51%
> Messages treated as ham: 392, 88.48%
> Messages learned: 419
> Connections count: 0
> Control connections count: 95
> Pools allocated: 170
> Pools freed: 143
> Bytes allocated: 26.72MiB
> Memory chunks allocated: 140
> Shared chunks allocated: 16
> Chunks freed: 0
> Oversized chunks: 1
> Statfile: BAYES_SPAM type: redis; length: 0; free blocks: 0; total blocks: 0;
> free: 0.00%; learned: 216; users: 1; languages: 0
> Statfile: BAYES_HAM type: redis; length: 0; free blocks: 0; total blocks: 0;
> free: 0.00%; learned: 202; users: 1; languages: 0
> Total learns: 418
> 
> [root at mx-ext ~]# rspamc stat
> Results for command: stat (2.001 seconds)
> Messages scanned: 16
> Messages with action reject: 0, 0.00%
> Messages with action soft reject: 0, 0.00%
> Messages with action rewrite subject: 7, 43.75%
> Messages with action add header: 0, 0.00%
> Messages with action greylist: 2, 12.50%
> Messages with action no action: 7, 43.75%
> Messages treated as spam: 7, 43.75%
> Messages treated as ham: 9, 56.25%
> Messages learned: 0
> Connections count: 0
> Control connections count: 46
> Pools allocated: 93
> Pools freed: 67
> Bytes allocated: 26.72MiB
> Memory chunks allocated: 142
> Shared chunks allocated: 16
> Chunks freed: 0
> Oversized chunks: 1
> Statfile: BAYES_SPAM type: redis; length: 0; free blocks: 0; total blocks: 0;
> free: 0.00%; learned: 216; users: 1; languages: 0
> Statfile: BAYES_HAM type: redis; length: 0; free blocks: 0; total blocks: 0;
> free: 0.00%; learned: 202; users: 1; languages: 0
> Total learns: 418
> 
> Or is Bayes only used when the number of scanned mails exceeds the min_learns
> on the local host regardless of the number of learned in BAYES_SPAM and BAYES_HAM?
> Is there a workaround to tell the other servers to use bayes?
> 
> Greetings
> Sandy Drobic
> 

It should be fixed by
https://github.com/rspamd/rspamd/commit/dbc9ff655dfb459eb8af328a82a5b8c848cda480

As a workaround, you can specify explicit weights for all upstreams, e.g.:

read_servers = "localhost:6376:1, mx-srv1.example.com:6376:1";