Definition
Drift detection is the practice of spotting when a deployed model's world has changed enough to hurt its accuracy. Models are trained on a snapshot of data, but production data keeps moving — user behavior changes, markets shift, hardware mixes evolve — and that mismatch quietly degrades performance unless something flags it.
Data drift versus concept drift
There are two distinct failure modes. Data drift is a change in the distribution of the input data — the production inputs no longer look like the training set, even if the underlying task is the same. Concept drift is deeper: the relationship between inputs and the target itself changes, so the same input should now produce a different correct answer. Both degrade quality and often coincide; in monitoring, a shift in data distribution can be an early symptom of concept drift.
How drift is caught and handled
Detection typically compares live production data against a baseline (usually the training distribution) using statistical tests on feature distributions and, where ground truth is available, on prediction accuracy. When drift is flagged, the response is a root-cause analysis followed by action — most often retraining the model on fresh data and promoting the new version through the model registry.
Drift detection is a core part of model monitoring and the trigger for continuous retraining in MLOps. For an operator running their own models, it is the early-warning system that keeps a self-hosted deployment honest over time.
In Simple Terms
Drift detection is the practice of spotting when a deployed model’s world has changed enough to hurt its accuracy. Models are trained on a snapshot…