Thomas Sobolik and Léopold Boudard highlight the challenges of maintaining the functionality of machine learning (ML) models in production, where subtle trends in the environment can radically alter their behavior. Unlike conventional applications, ML models require continuous evaluation to ensure accuracy and effectiveness. The authors discuss key metrics and strategies for monitoring model performance, including identifying evaluation metrics, monitoring proxy metrics like data and prediction drift, detecting data processing pipeline issues, directly evaluating prediction accuracy, and monitoring prediction and data drift. They also introduce the concept of drift, which refers to changes in the input data distributions or the model's predictions over time, and provide tools such as managed ML platforms and data analytics tools to detect these changes. The authors emphasize the importance of continuous evaluation and retraining of ML models to ensure they remain stable and performant over time.