Towards Observability for Machine Learning Pipelines

Explore the challenges and solutions for achieving end-to-end observability in machine learning pipelines in this insightful talk by Shreya Shankar, a Ph.D. student at the University of Berkeley. Delve into the complexities of managing ML workflows in heterogeneous tool stacks and learn about innovative approaches to address post-deployment issues. Discover mltrace, a platform-agnostic system designed to provide comprehensive observability for ML practitioners. Gain valuable insights into executing predefined tests, monitoring ML-specific metrics at runtime, tracking end-to-end data flow, and enabling post-hoc pipeline health inquiries. Understand the importance of observability in addressing unexpected output values and lower-quality predictions in production ML applications. This talk offers a deep dive into the cutting-edge research aimed at improving the operationalization and maintenance of machine learning systems in complex software environments.