We’ve been running Kubernetes for deep learning research for over two years. While our largest-scale workloads manage bare cloud VMs directly, Kubernetes provides a fast iteration cycle, reasonable scalability, and a lack of boilerplate which makes it ideal for most of our experiments.