Ray Summit 2020

As tools trend toward scaling to larger and larger data, their requirements are growing as well, which presents a non-trivial human cost. To bring the focus of data science tools back to the data scientist, we built Modin, a platform that scales data scientist capabilities without requiring them to learn about distributed computing concepts. In this talk, we discuss the design of Modin and how Ray enables our modular design to scale. We also discuss our experience with enabling the data scientist to connect to multiple Ray clusters from the same notebook.

Today's Industrial-AI faces 3 key challenges: “Can't Start”, “Can't Profit”, and “Can't Deploy”. Christopher will share how his team has successfully overcome these challenges using “Human-First AI”, an open-source, collaborative data-science framework that leverages Ray at scale.

In this talk, we will present our effort to run Ray on YARN, and the integration of Ray on LinkedIn’s open-sourced offline infrastructure: Azkaban (workflow scheduling service) and TonY (Tensorflow on Yarn). We will provide a demo of running a Ray job end-to-end, discuss the architectural decisions and talk about our cooperation with the Ray team on this effort.
In the later half of the talk, we will share how we used Ray Tune on Kubernetes in a real world use case. Tune helped us identify promising model configurations, using state of the art Bayesian optimization algorithms like TPE & PBT, with minimal supervision. The data pipeline is also optimized using Tune to extract the maximum throughput and we were able to train 2x faster by reducing the GPU idle times.

Developing distributed applications is difficult, and Ray tries to make it easier. However, it’s a well-known fact that most of our time is spent debugging, not actually writing the initial code. This talk will describe the vision for making debugging applications during development and monitoring them in production simple when using Ray. First, we’ll describe the recent improvements that we’ve made to metrics monitoring and the Ray dashboard. Then, we’ll present a few possible future directions including dashboard improvements, structured events, and distributed tracing. This will include a short demo of how to debug an application using these tools.

Many of the issues faced when applying ML techniques to problems in finance greatly benefit from a robust framework for distributed applications like Ray. In this talk we will go over some of the interesting challenges that arise, from model selection and backtesting to gaining user confidence, and how Ray may be used overcome them.

As a FinTech company, Ant Group emphasizes the core value of digital payments platform as it underpins e-commerce and financial services such as loans and insurance. Behind the digital payments platform lies the big data system, which serves as a pillar for effective data analysis and decision-making. Therefore, Ant Group particularly concerns about the latency and reliability of distributed systems. Last year Ant Group successfully applied Ray in financial service, marketing-promotion, risk control and other scenarios in Double 11 Shopping Festival 2019. In this talk, Jiaying will share their experiences in performance assurance and stability control in this process.

Ray was originally created as a framework for building distributed Python applications, mainly targeting at the AI area. At Ant Group, besides using Python, we also developed the Java API on top of Ray and use it to build large-scale big-data systems. In addition, we developed the cross-language invocation functionality, which allows you to seamlessly build a system that contains components written in different languages.

In this talk, we’ll introduce Ray’s Java API and show you how to write multi-language distributed applications with Ray.