A flexible, high-performance serving system for machine learning models