Question 17
Domain 1: Databricks Machine LearningA feature transformation is applied during training but forgotten in serving. What risk does this create?
Correct answer: C
Explanation
Training-serving skew happens when the model sees different data processing in training and production. If a feature transformation is applied during training but forgotten in serving, the model receives inconsistent inputs, so its predictions can degrade because "training" and "serving" no longer match.
Why each option is right or wrong
A. Guaranteed model improvement
Missing a serving transformation causes mismatch, not guaranteed improvement in model quality.
B. Better governance automatically
Governance concerns lineage, access, and compliance; preprocessing mismatch does not automatically improve it.
C. Training-serving skew
The issue is a mismatch between the training pipeline and the inference pipeline: the model was fit on transformed features, but at serving time it receives raw, untransformed inputs. In ML systems terms, that is training-serving skew, because the feature values presented to the model at prediction time are not the same representation used during training, so performance can drop even if the model itself is unchanged.
D. Lower latency in every case
Skipping a transformation might reduce work, but latency is not always lower and accuracy can suffer.