I would use tocken approach: each X seconds token is being granted, calling a service consumes n tokens, then add a cap so you won't have too many tokens accumulated overnight for example and you're good to go!