Monitoring AI solutions
Free Google Cloud Professional Machine Learning Engineer practice — 6 questions on Monitoring AI solutions, with explanations. No sign-up. Full 12-question mixed test →
A team deploys a tabular model to a Vertex AI endpoint with both prediction drift monitoring and feature attribution drift monitoring enabled. Feature attribution drift alerts begin firing for a feature that has historically had low predictive importance, even though overall prediction drift remains within threshold. What is the MOST likely explanation and correct action?
Feature attribution drift tracks how much a feature's contribution to model output changes over time using Vertex Explainable AI values, which can shift before the change is large enough to move overall prediction distributions — it's an early-warning signal distinct from prediction drift.
A team has an AutoML tabular model deployed to a Vertex AI endpoint. They want Model Monitoring to detect drift in feature attributions over time, not just drift in raw feature or prediction value distributions. What must they configure?
Feature attribution drift monitoring in Vertex AI requires Explainable AI to be enabled on the endpoint so attribution values can be computed for serving requests, plus configuring the attribution-score monitoring objective on the monitoring job — this is native functionality, not something built manually.
A Vertex AI endpoint serves roughly 50,000 predictions per hour. Model Monitoring's request-response logging to BigQuery is driving up storage and query costs, but the team still needs statistically meaningful training-serving skew and drift detection. What should they change in the monitoring job configuration?
The predictionSamplingRate parameter controls what fraction of serving requests are logged to BigQuery for monitoring; lowering it reduces logging volume and cost while still providing a statistically representative sample for skew and drift statistics.
An organization's Vertex AI resources sit inside a VPC Service Controls perimeter. After enabling Model Monitoring, monitoring jobs fail with a VPC Service Controls violation when attempting to write results to BigQuery. What is the correct fix?
VPC Service Controls enforce a service perimeter around resources regardless of IAM grants; to allow Model Monitoring's BigQuery writes, BigQuery and the target project must be inside the same perimeter or covered by an explicit ingress/egress access level.
Which statement correctly distinguishes training-serving skew detection from prediction drift detection in Vertex AI Model Monitoring?
Training-serving skew uses the training dataset as the fixed baseline for comparison, catching cases where serving data no longer resembles what the model was trained on, while prediction drift compares serving data against itself over successive time windows to detect gradual distributional change.
A newly rolled-out model version on a Vertex AI endpoint begins producing predictions with error rates far above SLA within an hour, triggering a Cloud Monitoring alert. The team must restore service immediately with minimal risk while preserving an audit trail for post-incident review. What should they do?
Vertex AI endpoints support traffic splitting across multiple deployed model versions; redirecting all traffic back to the known-good version is the fastest, lowest-risk rollback, and leaving the faulty version deployed (at 0% traffic) preserves it for root-cause analysis and audit purposes.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$129.99 $34.99 with code FREETEST33 — valid through September 7.
Get my $34.99 deal →