Uploaded November 2025 | Updated November 2026, 21 minutes ago
At Ray Summit 2025, Chen Zhu, Tao Huang, Yang Pei from LinkedIn share how the company adopted Ray to power a scalable online training system for real-time recommendation models—transforming the “Jobs You Might Be Interested In” ranking pipeline from slow, offline training to continuous, high-throughput updates.
They begin by outlining the limitations of LinkedIn’s traditional offline recommender workflows, where model freshness lagged by days or even weeks. This delay led to outdated recommendations, inconsistent features, and missed opportunities to adapt to rapidly changing user behavior and market dynamics.
By moving to online training with Ray, LinkedIn created a tight feedback loop that keeps models continuously refreshed—improving recommendation quality, boosting click-through rates, and cutting compute costs through incremental retraining. The result: up to 8.9× cost savings while delivering more relevant recommendations in real time.
The speakers then walk through the design of LinkedIn’s production online training system, including:
High-throughput streaming data generation
Scalable Kafka ingestion pipelines
A nearline training loop powered by Ray for fast, iterative updates
Key optimizations for throughput, latency, and system stability
They also share lessons learned from deploying Ray in a mission-critical production environment and the architectural choices that proved essential for reliability and scalability.
Attendees will leave with actionable insights into building Ray-powered online training systems that maintain model freshness, reduce costs, and drive measurable business impact.
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, Chen Zhu, Tao Huang, Yang Pei from LinkedIn share how the company adopted Ray to power a scalable online training system for real-time recommendation models—transforming the “Jobs You Might Be Interested In” ranking pipeline from slow, offline training to continuous, high-throughput updates.
They begin by outlining the limitations of LinkedIn’s traditional offline recommender workflows, where model freshness lagged by days or even weeks. This delay led to outdated recommendations, inconsistent features, and missed opportunities to adapt to rapidly changing user behavior and market dynamics.
By moving to online training with Ray, LinkedIn created a tight feedback loop that keeps models continuously refreshed—improving recommendation quality, boosting click-through rates, and cutting compute costs through incremental retraining. The result: up to 8.9× cost savings while delivering more relevant recommendations in real time.
The speakers then walk through the design of LinkedIn’s production online training system, including:
High-throughput streaming data generation
Scalable Kafka ingestion pipelines
A nearline training loop powered by Ray for fast, iterative updates
Key optimizations for throughput, latency, and system stability
They also share lessons learned from deploying Ray in a mission-critical production environment and the architectural choices that proved essential for reliability and scalability.
Attendees will leave with actionable insights into building Ray-powered online training systems that maintain model freshness, reduce costs, and drive measurable business impact.
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










