From Idle to Savings: Building a Global Scheduler for Cost‑Efficient Data P... Rainie Li & Ang Zhang

This talk focuses on the development of a global scheduling service at Pinterest that enhances cost efficiency for data processing on Kubernetes. The speakers, Rainie Li and Ang Zhang, discuss how this service significantly reduces CPU and GPU compute costs for big data and AI/ML jobs by optimizing resource usage across clusters using a combination of temporary capacity and fixed pools. They explain the algorithm that evaluates job placement based on runtime, urgency, and cost while maintaining service level objectives. The session covers the implementation of K8s primitives, including PriorityClass and node affinity, and presents insights into service design aspects like cost-aware routing and utilization dashboards.