Egregoros

@SilverDeth @Caudill @graf Oh, yeah, the "we have a million users" was, while they were present on fedi, like, they had, over the course of three months, 27,000 users that were actually active and made at least one post, and only 9,000 that made more than one post. This includes bots, some of which posted several times per minute. I had to implement k-means clustering and a bunch of heuristics to stop the botflood; the one I eventually settled on was brutally simple but effective: segmenting by activity type ("Create"/"Announce", at the protocol level, which are posts/reposts, versus other types, likes/follows/etc.; the idea was spambots are mostly interested in doing things that cause you to see the spam) and then count activities by hour of the day (the idea being that bots don't sleep). So, at least 99% visible activity as the filter and then sort by variance across time of day and manual review (no automated blocking, and bots that said "I'm a bot" or whatever were fine).