ML systems fail quietly when data changes.

A renamed column will break a job, and you find out fast. The dangerous issues are the ones that break nothing: a field quietly switches units, a client stops sending an event, a timestamp shifts time zones, a categorical column explodes in cardinality. The pipeline still runs, the model still serves, and the predictions just get worse.

Data validation is the set of gates that catch bad data before it reaches features, training datasets, or serving.