As bad as the latency-creating bottlenecks with random-routing and single-threaded apps can get, it is always the case that "twisting the magic knob" of adding

As bad as the latency-creating bottlenecks with random-routing and single-threaded apps can get, it is always the case that "twisting the magic knob" of adding more dynos would help. It just wouldn't help as much as people expected it to. Further, the amount of paid-for capacity left idle (while other dynos were backed-up) offends many sensibilities about how a system "ought to" work. But at least at the margin, the "magic knob" still worked a little. I think that's why the problem festered: there was always a default recommendation -- "more dynos!" -- that helped, which also just happened to fit the mental model Heroku promotes (while also boosting its revenue). In one of the older support logs RapGenius posted, it was evident that while RapGenius was suspicious about so many dynos being necessary, they let the ticket close once that expensive-but-temporarily-suitable 'fix' was in place. They (and Heroku) turned the magic knob and then let their attention move on to other things.