I’d like to mention it since Uber has a really cool load shedder [1], also implemented similarly by Netflix [2] and failsafe-go [3]. It basically looks for points where more requests per second suddenly cause a significant increase in latency and calls that the concurrency limit.

1: https://www.uber.com/us/en/blog/cinnamon-using-century-old-t...

2: https://github.com/Netflix/concurrency-limits

3: https://failsafe-go.dev/adaptive-limiter/

429 (and sometimes 503) errors returned by servers might well be a symptom of intentional load shedding. Perhaps it's just not explicitly called out as a server behavior that induces client retries.