Large, unsharded services are generally going to run one instance per machine alone as that's the most efficient way to run most of them (reduce per-instance ov

Large, unsharded services are generally going to run one instance per machine alone as that's the most efficient way to run most of them (reduce per-instance overhead as much as possible). Small (true "micro") services or other things which have instances per use case (like the inference platform example given in the paper) may use a small fraction of a machine.