A variety of important email enhancements to announce!
Training your spam filter just got a whole lot easier, simply drag email out of your “Spam” folder to learn as non-spam and drag an email into “Spam” to learn as spam. The “AutoSpam” folder is no longer in use on Atlas and Delta platforms.
SRS is enabled by default for forwarded email. Before when mail came in with an outbound forward as its next destination, email headers were preserved resulting in potential SPF/DKIM mismatches. This improves deliverability of mail forwarded for example to your GMail account that uses a bevy of heuristics to block/greylist spam.
Speaking of greylisting, forwarding has its challenges. Without inspection of mail that passes through it’s possible for a server to relay spam. SpamAssassin uses a very imprecise learning algorithm and relies on a concoction of expensive regular expressions to qualify the spamminess of an email. Worse yet it operates synchronously: in high volume scenarios it’s quite easy to clog email delivery. rspamd works asynchronously with a limited set of optimized regexes but derives its power from an adaptive Bayes classifier (Hidden Markov Model) that determines spam/ham not on the absolute content of the message but rather the sequence of content. As spam techniques adapt so too does this model.
I compared SA with rspamd over previously tagged data. Top, left to right: rspamd ham/spam without training, ham after training, SA ham/spam previously recorded, rspamd sample running over the ham trained data set again, my 2017 ham inbox filtered by rspamd after feeding it the Enron corpus vs what SA scored. rspamd needs a corpus to ramp up accuracy, so it’s bootstrapping off SA.
Both Atlas and Delta will switch over to rspamd exclusively once precision is dialed in in a couple months with an expectation to deliver mail faster and with fewer false positives
– Matt