Toward a system-wide and interdisciplinary perspective on ML system performance