The growing complexity and heterogeneity of networked systems have spurred a plethora of machine learning (ML) solutions, each promising a tantalizing improvement in performance. However, their path to real-world adoption is fraught with obstacles due to concerns from system operators about ML’s generalization, transparency, robustness, and efficiency.

My research takes a holistic approach to enabling practical ML for networked systems: 1) building open research platforms to lay the foundation for ML-based algorithms; 2) complementing ML with classical techniques (e.g., time-tested heuristics, control algorithms, or optimization methods) for enhanced deployability; and 3) validating ML-augmented methods through extensive empirical evidence gathered from real users or production systems. In this talk, I will demonstrate this research approach using three studies: Puffer/Fugu learns to adapt video bitrate in situ on a live streaming service we developed (with over 280,000 users to date), Autothrottle learns to assist resource management for cloud microservices, and Teal learns to accelerate traffic engineering on wide-area networks. Finally, I will conclude by outlining my research agenda for further pushing the boundaries of practical ML in networked systems.