April 2022 Summaries

Training machine learning models is a slow process that requires running many experiments with different options. Distributed machine learning addresses this problem by parallelizing training models using low-cost infrastructure in a clustered environment. This approach enables model-training time to improve from hours to minutes, and it's made possible by recent advances in distributed computing. Ray Train is a one-stop distributed training toolkit designed with ease of use, workstation friendliness, support for Jupyter Notebooks, fault-tolerance, and easy installation procedures in mind, promising to simplify the process of deploying machine learning models.

KubeRay is a popular toolkit for managing Ray clusters on Kubernetes, developed in collaboration with several major companies. It has become increasingly popular since its release in October 2021 and has recently introduced significant enhancements, including integration with Ray autoscaler, simplified installation using Kustomize, and a generic abstraction to improve cluster management. The tool aims to simplify the challenges of managing Ray clusters in native Kubernetes by providing a user-friendly interface and reducing the learning curve for users. Its community-driven development has led to bug fixes, stability and performance enhancements, and is now planning its next major release with new features such as job and serve CRD integration and improved workspace-centric development experience.

Ray 1.12 is here! The highlights in this release include the alpha release of Ray AI Runtime (AIR), a unified experience for Ray libraries, as well as various updates to Ray's data processing and machine learning capabilities such as the Deployment Graph API in Ray Serve and improvements to Ray Datasets. The team also introduced usage data collection to help discover and address pressing issues, while giving users full control over how their data is used and collected. With these features and more, users can now run pip install -U ray to access the latest version of Ray 1.12.

The Production RL Summit, hosted by Anyscale, concluded with a hands-on tutorial on Ray RLlib for building recommender systems. The tutorial was led by Sven Mika, the lead maintainer of RLlib, and covered topics such as reinforcement learning (RL), contextual bandits, and deep RL algorithms. Participants built their own Slate Recommender System using the RecSim environment and contextual bandit algorithms, including Thompson sampling and Linear UCB. The tutorial also explored offline RL algorithms, including behavioral cloning, CQL, MARWIL, and DQN. Additionally, participants learned how to deploy RL models using Ray Serve, a framework for serving trained machine learning models in production. The event concluded with a presentation at the ODSC-East conference in Boston, where attendees could engage with the Ray RLlib team on forums, Slack, and GitHub.