Data-Centric AI and MLOPS: Drift Detection and Model Monitoring

Abstract

Machine learning systems in production fail in ways that training metrics rarely predict. Once a model serves live traffic, the data it receives drifts away from the distribution it learned, and accuracy degrades silently. This paper examines the shift from a model-centric view, where the dataset is fixed and architectures are tuned, toward a data-centric practice in which data quality, validation, and continuous monitoring become first-class engineering concerns. The discussion is organized around the machine learning operations (MLOps) lifecycle, covering continuous integration, delivery, and training, together with feature stores and model registries that make pipelines reproducible. A taxonomy of distribution shift is presented, separating covariate shift, prior or label shift, and concept drift, each with distinct detection needs. Detection methods are reviewed across two settings: batch statistics such as the population stability index, Kullback-Leibler and Jensen-Shannon divergence, the Kolmogorov-Smirnov test, and maximum mean discrepancy, and streaming detectors such as ADWIN and the drift detection method. Illustrative metrics show detection delay and false-alarm trade-offs across these methods, alongside a monitored stability index crossing an alert threshold that triggers retraining. Retraining strategies, governance, and current tooling are discussed. The aim is a practical reference for teams that must keep deployed models reliable as their input data changes.