Also this doesn't require top-down management of overlapping/coordinated rate limits, since the sending rate is directly affected by congestion events.

This is based on IP network congestion control, but the interesting point I took away from this paper was that if you signal which stream of events has triggered the congestion, you can reduce the rate of events in the stream at source without affecting other low-throughput streams.

Once an external rate limit is saturated, receiving 'Too many requests' 429 responses can be used to back-off and retry, but this slows down all streams.
We don't want large buffers of pending requests, but do want to use full available capacity.