Rate limiting looks outdated. Suppose you can spawn new servers (and turn down unused ones) fast enough to react to demand, and have a good LB that sends the traffic where you have capacity, then you do not need to limit anything until you reach full capacity, in which case you probably want to degrade (by query class and/or client class) rather than limit anyway.

Rate limiting is an inefficient way to distribute a service, that makes sense only if you preallocate your resources and have queries with predictable cost. Let's use this technique only as a bug-prevention tool not for resource economy, as organised scarcity is likely as inefficient for the data center as it is for the distribution of goods :)