If you’ve been with us a while you’ve probably noticed that our improvements tend to come in waves, and usually focus on specific areas of the service at a time. Most of the time our efforts are spent working on the outbound systems, as consistent delivery for the email you send is kind of a big deal to us.
Today we’re starting on a new round of work on our inbound spam filters. Inbound spam is a sensitive topic here. Some of our customers want every single email addressed to them, no matter where it comes from or how bad it is. Others look to us as the authority for halting spam before it reaches their inbox. To balance it, we focus on making a happy medium. The happy medium is this:
- We maintain a layer of filtering that rejects email outright, before any of the user’s own spam settings apply. Every email service does this, but we all have unique criteria for doing it. Ours has been fairly relaxed and in most cases favors receiving spam to avoid false positives.
- We allow the user to control their own settings for anything that passes the initial filtering layer. This is most of the inbound email.
To improve on our filtering efforts, and to reduce the amount of spam we receive (and, as a result, forward… this isn’t entirely without consideration for our outbound delivery quality) we are improving the top layer of the filtering. This is the layer that rejects email before the user’s settings apply. We understand that we have to be careful here because we’d like to avoid increasing false positives while decreasing inbound spam. To continue our tradition of transparency, and to allow customers to make contributions to our inbound filters, we’ve updated our Github repo with the latest configurations that we’ve added to our default rspamd installations on the inbound servers:
We will be rolling out this update slowly to all of our servers, observing the impact to ensure that we’re not introducing problems that we didn’t encounter in our staging environment. All servers will see a roll out of this by November 14, 2020.