If you can't already do it with the rate limit module I wrote, open an issue with your detailed requirements: https://github.com/mholt/caddy-ratelimit -- should be pretty straightforward for the most part.

> QoS for a shared-multitenant system, in the presence of customers with really badly tuned and spiky request workloads, whose traffic you must nevertheless mostly accept.

Yah, we see that sometimes. Caddy usually handles it fine, sometimes with a bit of massaging the config.