Was on it until late last year, switched off recently. The cost for the performance and reliability of their Redis and Postgres instances was pretty bad, and it didn't seem like the dynos were well co-located with the db. A few of the issues we had:

  1. Redis was wildly unreliable, would go down for 2-5 minutes every few days with no explanation from support
  2. We had to pay for a much beefier instance of Redis just to increase our connection count. We weren't using more than 2-3% of the available memory, we just needed the connections, but you can't pick and choose.
  3. Their limits on websocket connections were also a killer, I think each dyno could have a max of like 1,000 sockets so we had to add more dynos just for more sockets.
  4. Switching over to AWS reduced the RTT (just networking, not actual query runnig time) on Postgres queries from ~8ms to ~1ms. Seems like the dynos and db weren't closely co-located.

However, the amount of time spent and code we have to maintain our infrastructure is waaay more than it used to be on Heroku. No free lunch for sure.