GCP Professional ML Engineer Practice Question 9: Training Serving Skew Detection

Model Monitoring distinguishes the two objectives by their baseline. Training-serving skew detection compares the feature distributions in live inference requests against the original training data, so it can only be enabled if you supply that training dataset, which is why the configuration asks for a training data source plus the target column, the latter being excluded from the analysis. Prediction drift detection compares recent serving data against earlier serving data and needs no training set. The stem states the training table is available and names the training data as the comparison baseline, so skew detection is the objective, and the v1 endpoint-based monitoring it uses is the generally available, production-supported offering.

Option A is the other objective. Drift would tell you that today's traffic differs from last month's traffic, which is a genuinely useful signal but not the question risk management asked.

Option B could be made to produce a similar number, but it is a bespoke pipeline: you own the logging schema, the binning, the divergence computation, the thresholds, and the alerting, all to reimplement a managed feature.

Option C measures something different again. Feature attribution monitoring tracks how much each feature contributes to inferences over time, it catches a key feature losing importance, but it does not compare input feature distributions against the training data.

Reference: Monitor feature skew and drift