← Back to the chapter · Previous detail · Next detail →
Technical detail · Production ML Monitoring
Distinguish drift from its possible causes
Data drift changes input distributions as sensors age, placement or calibration changes, firmware changes, or context shifts. Concept drift changes the relationship between features and outcomes. Label drift changes what operators call correct. Decision drift means the output no longer matches operational value because the process, user population, costs, or policy changed.
Runtime drift changes the deployment boundary through latency, memory, power, queueing, connectivity, telemetry, or update behavior. Release drift means control is lost because version identity, rollout scope, rollback rules, ownership, or audit evidence is missing. These changes need different evidence and owners.
Before retraining, check for a broken sensor, a unit or schema change, a new operating regime, a threshold that no longer matches alert cost, or dropped or delayed features. If confidence falls only during night shifts, inspect load patterns and label coverage. If latency rises after a quantized model update, inspect packaging, batching, and hardware scheduling. Neither symptom alone proves concept drift.