[xorcist]

Yeah, I certainly wouldn't start at that end.

A useful metric for "load" is just as hard as doing the load balancing itself.

Requests can be vastly different, unless you have only one application they're also constantly changing and there are more load balancers involved (both horizontally and vertically as a single request can pass multiple). There are also numerous failure conditions under which responses are very fast.

In that situation it's easier to design for an even spread, then work to improve that metric as much as possible as more information becomes available.