Original implementation was very slow (~70s) when the observer wasn't subscribed to communitities, because it had to search entire hive_posts table in order to determine there were no posts to return. This algorithm picks most recent _limit posts from each community that observer is subscribed to, then sorts then by most recent.