Scaling Pinterest ML Infrastructure with Ray: From Training to End-to-End ML Pipelines

Pinterest expanded Ray's capabilities beyond training to power their entire ML infrastructure stack, including feature development, sampling, and labeling workflows. They built a Ray Data native pipeline API, implemented Iceberg bucket joins for efficient data joining, added data persistence mechanisms, and optimized Ray Data for large workloads. These improvements reduced ML iteration times by 10x while cutting infrastructure costs, enabling faster feature experimentation and end-to-end ML pipeline development.