Requests are queued in a bucket and processed at a constant "leak rate". If the bucket is full, clients must back off.