How to detect model drift with on-prem data on Vertex AI?

PMLE By islever Editorial Last updated

Short answer

D Create a Vertex AI Model Monitoring job. Enable feature attribution skew and drift detection for your model.

Question

You built a deep learning-based image classification model by using on-premises data. You want to use Vertex AI to deploy the model to production. Due to security concerns, you cannot move your data to the cloud. You are aware that the input data distribution might change over time. You need to detect model performance changes in production. What should you do?
  1. Use Vertex Explainable AI for model explainability. Configure feature-based explanations.
  2. Use Vertex Explainable AI for model explainability. Configure example-based explanations.
  3. Create a Vertex AI Model Monitoring job. Enable training-serving skew detection for your model.
  4. Create a Vertex AI Model Monitoring job. Enable feature attribution skew and drift detection for your model. Correct Answer

Key Takeaways

This question tests your understanding of Vertex AI Model Monitoring prerequisites, specifically that training-serving skew requires access to training data, while feature attribution drift can monitor production changes without it.

To detect model performance changes on Vertex AI when on-prem training data cannot be moved to the cloud, you must enable feature attribution skew and drift detection. The community agrees that standard training-serving skew detection is infeasible without access to the original training dataset.

Common mistake: Choosing Option C (training-serving skew) is the most common mistake because candidates overlook the constraint that on-prem data cannot be moved to the cloud, making the baseline comparison impossible.

Expert Analysis

Why the Answer Is Correct

Option D is correct because it addresses the security constraint that training data cannot leave the on-premises environment. Feature attribution skew and drift detection monitors the distribution of feature attributions (explanations) in production traffic to detect changes in model behavior without needing the original training dataset stored in the cloud.

Why the Other Options Are Wrong

Option C is incorrect because training-serving skew detection explicitly requires comparing the statistical distribution of production data against the training data, which is impossible if the training data is not available in Vertex AI. Options A and B are incorrect because Vertex Explainable AI is designed for interpreting individual predictions (debugging and transparency), not for monitoring ongoing model performance or detecting data drift over time.

Community Comment Notes

The community consensus heavily favors D, primarily due to the data availability constraint. Commenters [2] and [7] emphasize that Option C is invalid because the training data is not available in production. Comment [5] supports this by noting that Vertex AI Model Monitoring can analyze inputs and predictions without uploading the raw training data, whereas Comment [6] correctly identifies that training-serving skew is infeasible under these conditions.

Exam Strategy

When a scenario states that data cannot move to the cloud, immediately eliminate options that require cloud-side access to the training dataset, such as training-serving skew. Instead, look for monitoring strategies that rely solely on production inputs and predictions, such as feature attribution drift.

Community Votes

D
77%
C
23%

77% of anonymous learners picked answer D. Votes are pick records left by other test-takers — they are not the verified answer.

Discussion 8 contributions

Compiled from public community threads and reader submissions, lightly edited for clarity. Handles and vote tallies are shown as recorded, and comments reflect the views of individual contributors.

From public community threads 8

  1. b1a8fae 👍 8 Answered D
    D. You want to control how much the distribution of the data changes over time -> that's drift.
  2. sonicclasps 👍 5 Answered D
    the answer cannot be C, cause your training data is not available in production. So D is the only viable answer
  3. rajshiv 👍 3 Answered C
    Option D is incorrect in my opinion. "Feature attribution skew and drift detection" focus on knowing how feature values are contributing to the model’s predictions, but "training-serving skew" is more direct and efficient to detect distribution changes that could lead to performance issues
  4. gscharly 👍 2 Answered D
    D, as the training data is not available
  5. fitri001 👍 1 Answered D
    Security: Vertex AI Model Monitoring doesn't require uploading your training data to the cloud. It analyzes model predictions and input features on your on-premises server. Data Distribution Shifts: Feature attribution techniques like LIME or SHAP within Vertex AI Model Monitoring can identify how different features contribute to model predictions. Detecting drifts in these feature attributions can indicate changes in the underlying data distribution compared to the training data.
Show 3 more comments
  1. pikachu007 👍 1 Answered C
    Option A and B: Vertex Explainable AI provides insights into model behavior but doesn't directly detect performance changes or concept drift. It's more suitable for understanding model decisions, not monitoring production performance. Option D: Feature attribution skew and drift detection requires feature attributions calculated during training, which might not be feasible without cloud access to the data.
  2. omermahgoub 👍 1 Answered C
    Feature Attribution Skew and Drift Detection, this type of monitoring is useful in some cases, it requires access to the training and serving data for analysis. Since data cannot move to the cloud, Option D wouldn't be feasible. I vote for C. Create a Vertex AI Model Monitoring job. Enable training-serving skew detection for your model.
  3. pinimichele01 👍 1 Answered D
    the answer cannot be C, cause your training data is not available in production.

Reader comments & corrections

No reader corrections yet — spotted an error or have a note? Share it below.

Log in to comment, report an error, or add a note about this question.

How to Minimize Development Time for Weekly TensorFlow Retraining in Vertex AI?

S-grade

How to Handle Categorical Features in BigQuery ML for Fast Deployment?

S-grade

How to preprocess BigQuery data for TensorFlow?

S-grade

How to Track Model and Prediction Lineage in Vertex AI?

S-grade

Practice All PMLE Questions

Access 65 questions with complete answers and detailed explanations.

View Full PMLE Practice Test →

← Back to PMLE Study Guide