Pecan: Cost-Efficient ML Data Preprocessing with Automatic Transformation Ordering and Hybrid Placement

Explore a conference talk from USENIX ATC '24 that introduces Pecan, an innovative ML data preprocessing service designed to optimize costs and efficiency in machine learning workflows. Delve into the challenges of input data preprocessing as a common bottleneck in ML jobs, and discover how Pecan addresses these issues through two key approaches. Learn about the dynamic scheduling of data preprocessing workers on ML accelerator host resources and the automatic reordering of transformations to increase worker throughput. Examine how Pecan's techniques can significantly reduce preprocessing costs by an average of 87% and total training costs by up to 60% compared to existing methods. Gain insights into the importance of balancing commutativity and throughput in ML data pipelines while maintaining high model accuracy. This 21-minute presentation by researchers from ETH Zurich and Google offers valuable knowledge for ML practitioners and researchers looking to enhance the cost-efficiency of their data preprocessing workflows.