Uploaded September 2026 | Updated November 2026, 3 hours ago
A single ML pipeline has wildly different resource profiles at each stage: data preprocessing is CPU- and memory-bound, large-scale training is accelerator-bound. Forcing the whole pipeline onto one homogeneous GPU node means paying for idle accelerators during data-heavy stages and stalling on scarce GPU capacity.
At Ray Summit 2026, Robert Macy from Robinhood and Viraj Padte from AWS share chapter two of Robinhood's KubeRay journey: defining multiple worker groups in a single RayCluster, each backed by a different node type, and letting Ray's resource-aware scheduling map each pipeline stage to the hardware it actually needs. They walk through expressing per-task resource requirements, how the Ray autoscaler decides which worker groups to scale, and how Ray Data hands off to Ray Train across different hardware.
You'll leave knowing how to run heterogeneous Ray clusters on Amazon EKS inside one Ray job.
Liked this video? Check out other Ray Summit breakout session recordings: youtube.com/playlist?list=PLZNaJyYZRPdo
Subscribe to our YouTube channel to stay up-to-date on the future of AI: youtube.com/c/anyscale
đź”— Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com
A single ML pipeline has wildly different resource profiles at each stage: data preprocessing is CPU- and memory-bound, large-scale training is accelerator-bound. Forcing the whole pipeline onto one homogeneous GPU node means paying for idle accelerators during data-heavy stages and stalling on scarce GPU capacity.
At Ray Summit 2026, Robert Macy from Robinhood and Viraj Padte from AWS share chapter two of Robinhood's KubeRay journey: defining multiple worker groups in a single RayCluster, each backed by a different node type, and letting Ray's resource-aware scheduling map each pipeline stage to the hardware it actually needs. They walk through expressing per-task resource requirements, how the Ray autoscaler decides which worker groups to scale, and how Ray Data hands off to Ray Train across different hardware.
You'll leave knowing how to run heterogeneous Ray clusters on Amazon EKS inside one Ray job.
Liked this video? Check out other Ray Summit breakout session recordings: youtube.com/playlist?list=PLZNaJyYZRPdo
Subscribe to our YouTube channel to stay up-to-date on the future of AI: youtube.com/c/anyscale
đź”— Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com










