Uploaded November 2025 | Updated September 2026, 1 week ago
At Ray Summit 2025, Avin Regmi and Matan Appelbaum from Netflix share how the company is building Mako, a next-generation ML training platform engineered to meet the demands of modern, large-scale AI workloads.
They begin by outlining how Netflix evolved its legacy training platform to support increasingly complex models, larger datasets, and rapidly growing GPU requirements. Central to this evolution was the development of a custom GPU scheduler designed to significantly improve utilization, reduce fragmentation, and ensure that large multi-node training jobs run efficiently and reliably.
The speakers then walk through the architectural changes that enabled Netflix to scale training across diverse workloads—highlighting critical decisions around resource orchestration, distributed execution, and system resilience. They also detail where Ray fits into this journey, including how its flexible distributed runtime supports key components of the platform and what the team learned from integrating Ray into high-performance training pipelines.
Attendees will gain insights into designing modern ML training platforms, optimizing GPU usage at scale, and leveraging Ray to support the next wave of AI infrastructure at Netflix.
Liked this video? Check out other Ray Summit breakout session recordings youtube.com/playlist?list=PLzTswPQNepXllnU0C36WtkC0dqkAoDulh
Subscribe to our YouTube channel to stay up-to-date on the future of AI! youtube.com/c/anyscale
🔗 Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com
At Ray Summit 2025, Avin Regmi and Matan Appelbaum from Netflix share how the company is building Mako, a next-generation ML training platform engineered to meet the demands of modern, large-scale AI workloads.
They begin by outlining how Netflix evolved its legacy training platform to support increasingly complex models, larger datasets, and rapidly growing GPU requirements. Central to this evolution was the development of a custom GPU scheduler designed to significantly improve utilization, reduce fragmentation, and ensure that large multi-node training jobs run efficiently and reliably.
The speakers then walk through the architectural changes that enabled Netflix to scale training across diverse workloads—highlighting critical decisions around resource orchestration, distributed execution, and system resilience. They also detail where Ray fits into this journey, including how its flexible distributed runtime supports key components of the platform and what the team learned from integrating Ray into high-performance training pipelines.
Attendees will gain insights into designing modern ML training platforms, optimizing GPU usage at scale, and leveraging Ray to support the next wave of AI infrastructure at Netflix.
Liked this video? Check out other Ray Summit breakout session recordings youtube.com/playlist?list=PLzTswPQNepXllnU0C36WtkC0dqkAoDulh
Subscribe to our YouTube channel to stay up-to-date on the future of AI! youtube.com/c/anyscale
🔗 Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com
