Reinforcement learning (RL) approaches have long interested the systems community as they promise to learn complex dynamic behaviour (e.g. scheduling) from raw feedback (e.g. latency, throughput). Recent successes in combining deep neural networks with RL have sparked significant new interest in this domain. While theoretically appealing, practical deployments remain elusive due to large training data requirements, algorithmic instability, and lack of standard tools. This talk will cover modern reinforcement learning approaches in computer systems research from a number of perspectives. I will first compare RL to widely used auto-tuning methods and provide intuition on standard algorithms. I will then discuss the gap between current research and practical deployments, describe potential solutions, and introduce a software stack for RL in systems research. Finally, I will discuss case studies in a number of application domains, and highlight future directions.