To apply machine learning to a broader range of applications, companies need better tools for model monitoring and debugging. In this post, Shreya Shankar discusses the limitations of current approaches to model monitoring (e.g. that calculate metrics based on a rolling window), which may not represent model performance on streams of data. Moreover, she highlights the need for better processes to define service-level objectives that relate a set of business objectives to a collection of metrics, window sizes, and alert procedures for a given task.