Scaling ML Pipelines with Feast, Ray and Kubeflow

Learn to transform slow, file-based ML pipelines into distributed, production-ready architectures in this conference talk from DevConf.IN 2026. Discover how to overcome common bottlenecks like feature engineering delays and data loading issues that plague ML training workflows. Explore the integration of three powerful open-source tools: Feast for feature management, Ray for distributed data processing, and Kubeflow Training Operator for orchestrating distributed training on Kubernetes. Follow along with a comprehensive demonstration featuring an end-to-end pipeline that powers a Temporal Fusion Transformer trained on 421K rows of Walmart sales data, achieving 10.5% MAPE compared to the typical 15-20% industry baseline. Witness how PyTorch DDP across multiple GPUs significantly reduces training time while maintaining model performance. Examine faster feature loading techniques using Ray and Feast, observe raw data flowing through a fully managed feature platform, and understand how distributed PyTorch jobs are launched and scaled with Kubeflow Training Operator. Gain insights into production inference paths powered by Feast's hybrid storage and compute capabilities, learn how Ray transforms feature engineering performance at scale, and discover how Feast standardizes feature computation across both training and inference phases. Walk away with a repeatable blueprint for building ML pipelines that scale effectively as your models, data, and teams grow, plus the confidence to implement these tools in your own production environment.