Using Ray On Large-scale Applications at Ant Group
summary
Ant Group replaces a glue of Flink/Spark/TF with Ray as a single engine, adding plugin runtimes for streaming, parallel batch and AI (TensorFlow-on-Ray), deploying via a Ray operator on Kubernetes with auto‑scaling and HPA. A custom MPP framework compiles Java code into Ray tasks and uses BSP-style barriers; it processes ~10 MB/s and more than 10 million dimensions end-to-end and returns results in 200 milliseconds during peak events with no data‑loss and a single SLA.
AI-written summary · about
chapters (9)
- 0:00Intro
- 4:04The problems of multiple computing systems
- 6:28Build a general propose distributed system to support E2E computing
- 7:28Ray is a fast and simple framework for building and running distributed applications
- 8:32Multiple computational patterns as plugin on top of Ray. • Based Ray on Yarn and Kubernetes.
- 10:15An online machine learning pipeline.
- 13:56E2E Online Learning architecture based on Ray
- 19:30Building MPP Framework based on Ray
- 24:44Fusion Computing, build complex computing systems without limits.
related talks
03mRay: A Distributed Execution Framework for Emerging AI Applications
26mData Processing on Ray
15mRay: A Distributed Compute Engine for AI
23mImproving Ray for Large-scale Applications
24mRay: A System for High-performance, Distributed ML Applications
19mScaling Frontier AI with Ray on Kubernetes
28mA Deep Dive into Ray’s Scheduling Policy
28mScaling Multimodal Data and Reinforcement Learning with Ray