Restarts in the case of repeated dyno crashes now have exponential backoff. In the normal case of a long-running web or worker process getting an occasional crash, the dyno will be restarted without any intervention on your part. If a dyno crashes multiple times in a row, it will be restarted with exponentially increasing cool-off periods. If you push a new release to your app, the cool-off period will be reset.
Previously, the backoff policy was to restart crashed dynos every 10 minutes.