A combo of preboot and pipelines seems to help a bit, but Heroku's router is still to stupid not to send requests to shutting down dynos :/
