Yes we found that two years ago and i studied it. It looks like promising but when you find some statistics comparative with spamassassin, results are not so different and this is logical as there is (as far i found) no rules bases, but a compatibility with spamassassin rules !
Considering than Spamassassin rules are far from being kept updated regularly (are they ? they mainly decrease score automatically using scripts as far as i can notice. They have almost no people supporting it and no recent db of messages. They were asking for help about that some years ago and when i wrote to them last year, i got no answer at all. Can't say it is dead but this is an open source project needing help I think).
Finally i found two interesting things in RSpamd:
1) Was written in C, so it is faster than SA
2) Fuzzy Hashes (I wrote a module for MailCleaner, never published yet because of performance hit to do exactly that, would be better to integrate Rspamd)
3) Neural network: I like people launching words like this. So the bayesian learn from mass message to find recurrent patterns ? I would like to see difference with the bayesian from Spamc. But could be efficient if algorythms are good.
Ok so there was three points...
Hmmm.... Lots of things to do after the migration but i would like to implement it anyway now