Definition
Data drift is a change over time in the distribution of inputs or features seen by a system. Examples include a shift in customer geography, document length, device type, language, or the frequency of missing values.
Drift can indicate a real population change, a broken pipeline, a new product, seasonality, or an attack. Statistical distance can flag that something moved, but it does not say whether the change matters. Pair drift monitoring with data-quality checks and performance evidence on affected slices.
Choose a reference period deliberately. A stable annual pattern may look alarming against last month's data, while slow deterioration can disappear if the baseline continually rolls forward.
Distinguish it from nearby terms
Concept drift changes the relationship between inputs and desired outputs. Data drift can occur while that relationship remains valid. Data-quality failure corrupts inputs and may imitate drift without representing a real population change.
Check your understanding
Document length doubles after a new upload feature launches, but extraction accuracy stays stable. The input distribution drifted; the evidence does not yet show concept or performance drift.