Uploaded November 2025 | Updated September 2026, 2 weeks ago
At Ray Summit 2025, Avin Regmi and Matan Appelbaum from Netflix share architectural patterns for processing petabyte-scale, multi-modal data for Generative AI—spanning text, video, audio, and more—using Ray as the backbone for large, heterogeneous distributed pipelines.
They begin by outlining why multi-modal data processing at this scale is uniquely challenging. These pipelines must coordinate diverse workloads, manage stateful operations such as deduplication, and efficiently utilize GPU acceleration—all while maintaining reliability across massive distributed clusters.
Drawing on their experience building NVIDIA NeMo Curator, the speakers demonstrate how Ray’s core primitives make this possible. They show how:
Ray Actors support stateful, long-running operations needed for tasks like deduplication, metadata tracking, and multi-step transformations
Ray Tasks provide highly parallel, stateless processing for large-scale batch transformations
Heterogeneous CPU/GPU resource management maximizes throughput and pipeline robustness across multi-modal workloads
Avin and Matan break down the patterns and best practices that enabled their pipelines to operate efficiently across petabytes of data and thousands of distributed workers.
Attendees will walk away with a clear understanding of how to use Ray to architect scalable, resilient, GPU-accelerated data pipelines for next-generation Generative AI applications.
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 architectural patterns for processing petabyte-scale, multi-modal data for Generative AI—spanning text, video, audio, and more—using Ray as the backbone for large, heterogeneous distributed pipelines.
They begin by outlining why multi-modal data processing at this scale is uniquely challenging. These pipelines must coordinate diverse workloads, manage stateful operations such as deduplication, and efficiently utilize GPU acceleration—all while maintaining reliability across massive distributed clusters.
Drawing on their experience building NVIDIA NeMo Curator, the speakers demonstrate how Ray’s core primitives make this possible. They show how:
Ray Actors support stateful, long-running operations needed for tasks like deduplication, metadata tracking, and multi-step transformations
Ray Tasks provide highly parallel, stateless processing for large-scale batch transformations
Heterogeneous CPU/GPU resource management maximizes throughput and pipeline robustness across multi-modal workloads
Avin and Matan break down the patterns and best practices that enabled their pipelines to operate efficiently across petabytes of data and thousands of distributed workers.
Attendees will walk away with a clear understanding of how to use Ray to architect scalable, resilient, GPU-accelerated data pipelines for next-generation Generative AI applications.
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
![[Ray Meetup] Ray + vLLM in Action: Lessons from Pinterest and Large Scale Distributed Inference](https://i.ytimg.com/vi/HDSy09hrm2I/mqdefault.jpg "[Ray Meetup] Ray + vLLM in Action: Lessons from Pinterest and Large Scale Distributed Inference
Listen in to our Ray Meetup where we explored batch inference at scale with Ray and vLLM! Learn how Pinterest scales batch inference using Ray, and get a first look at Anyscale’s latest tools—Ray Serve and Data LLM—for orchestrating large-scale LLM inference. We’ll cover topics like batch inference, prefill-decode disaggregation, DP/EP parallelism, and custom request routing.
Speakers:
Chia-Wei Chen, Software Engineer, ML Training Infra, Pinterest
Kourosh Hakhamaneshi, AI Lead, Anyscale")
