Pinterest's ML Evolution: Distributed Training with Ray

Uploaded October 2024 | Updated September 2026, 2 weeks ago

Pinterest's success hinges on its advanced recommender systems, which power key products like Homefeed, Related Pins, and Ads. In this session, Saurabh Vishwas Joshi and Jiun-Yu Lee from Pinterest's ML Platform team reveal how they've harnessed Ray to optimize data loading for their data-intensive recommender model training.

The presenters detail Pinterest's journey in leveraging Ray and the Ray Data ecosystem to decompose and orchestrate massive workloads across thousands of training jobs. They discuss the challenges faced, including memory pinning and multi-threaded collate, and the innovative solutions implemented to scale data-loading beyond trainer nodes. Learn how these optimizations have significantly boosted training throughput and how Pinterest is addressing new challenges through internal abstractions and open-source contributions.

--

Interested in more?
- Watch the full Day 1 Keynote: youtu.be/jwZHJthQvXo
- Watch the full Day 2 Keynote youtu.be/Lury2ad6KG8

--

🔗 Connect with us:
- Subscribe to our YouTube channel: youtube.com/@anyscale
- Twitter: https://x.com/anyscalecompute
- LinkedIn: linkedin.com/company/joinanyscale
- Website: anyscale.com

Pinterests ML Evolution: Distributed Training with Ray | Ray Summit 2024

How xAI Scales Image & Video Processing with Ray | Ray Summit 2025

Ion Stoica on Agentic Systems and AI Reliability | Ray on the Road – NYC 2025

Anyscale on Azure: Build and deploy AI at scale in your own tenant

An Overview of CloudKitchenss Ray-Powered ML Platform | Ray Summit 2024

$Prompt Learning: A Reinforcement Learning-Inspired Approach to AI Optimization | Ray Summit 2025 At Ray Summit 2025, Jason Lopatecki from Arize AI shares a new paradigm for iterative model improvement—Prompt Learning (PL)—a technique inspired by reinforcement learning but performed entirely through prompts, without weight updates, large datasets, or gradient-based training. He begins by framing the central question: What if we could achieve RL-like gains in model performance simply by refining prompts using natural-language feedback? Building on ideas from NVIDIA’s Voyager and broader community discussions, Arize explores how structured English feedback—evaluations, explanations, annotations—can act as an “error term” that guides models toward better behavior. Jason walks through the conceptual foundation of Prompt Learning and why encoding corrections directly in language can dramatically reduce data requirements—sometimes down to only a handful of examples. Key highlights include: The Prompt Learning framework: How PL adapts RL-style improvement loops to the prompt layer Why natural-language feedback is powerful: Turning textual “error signals” into prompt refinements without extra training PL vs RL efficiency: A head-to-head comparison showing where PL outperforms traditional RL-driven refinement Real-world experiment: Guiding JSON generation with latent constraints using only a few examples TextPRO system: Arize’s PL engine that performs end-to-end optimization through single LLM calls—at a fraction of the supervision cost Jason shares experimental results showing that Prompt Learning can achieve up to 70% test accuracy using just five feedback rules, outperforming standard prompting techniques and demonstrating that lightweight, language-based optimization can yield surprisingly strong results. Attendees will leave with a new lens on iterative LLM improvement—one that opens the door to powerful, low-cost prompt optimization without full RL pipelines or large training sets. Liked this video? Check out other Ray Summit breakout session recordings https://www.youtube.com/playlist?list=PLzTswPQNepXllnU0C36WtkC0dqkAoDulh Subscribe to our YouTube channel to stay up-to-date on the future of AI! https://www.youtube.com/c/anyscale 🔗 Connect with us: LinkedIn: https://www.linkedin.com/company/joinanyscale/$

Introducing Terminal-Bench: Evaluating LLM Agents in Realistic Terminal Settings | Ray Summit 2025