Google Cloud PMLE Sample Questions with Explanations

These original sample questions are designed to help you check how the exam topics appear in decision-style prompts. They are not taken from the live exam.

Use these sample questions as a guided self-assessment for Google Cloud Professional Machine Learning Engineer (PMLE) topics such as data readiness, model selection, Vertex AI training, deployment, MLOps, monitoring, explainability, responsible AI, and generative AI evaluation.

Where these questions fit in the PMLE guide

The sample set below is part of the Google Cloud PMLE guide path:

• Google Cloud certification hub
• PMLE exam guide
• PMLE cheat sheet
• PMLE study plan
• PMLE FAQ
• PMLE resources
• PMLE glossary

PMLE ML-engineering sample questions

Work through each prompt before opening the explanation. Strong ML-engineering answers protect the full lifecycle: metric, data, training, serving, monitoring, and retraining.

---

Question 1

Topic: Preventing data leakage

A team trains a churn model and accidentally includes a feature that is only known after a customer cancels service. Offline accuracy looks excellent, but production predictions are poor. What is the most likely issue?

• A. The model should use a larger machine type for prediction.
• B. The model has data leakage because training used information unavailable at prediction time.
• C. The model should always be replaced with a generative AI model.
• D. The serving endpoint needs public unauthenticated access.

Best answer: B

Explanation: The feature leaks future information into training. PMLE scenarios often test whether offline metrics are trustworthy given the timing and availability of features.

Why the other choices are weaker:

• A may change capacity, but it does not fix invalid training data.
• C changes model class without addressing leakage.
• D creates a security problem and is unrelated to model quality.

What this tests: Feature availability, leakage, offline evaluation, and production validity.

Related topics: Data leakage; Features; Evaluation; Model quality

---

Question 2

Topic: Choosing batch prediction

A retailer scores 40 million product recommendations every night for use in the next day’s email campaign. Low latency is not required, but throughput and repeatability matter. Which serving pattern is strongest?

• A. Run batch prediction and write the scored outputs to storage or a warehouse for the campaign system.
• B. Deploy a low-latency online endpoint and call it once for each customer during the email send.
• C. Ask analysts to score customers manually in a spreadsheet.
• D. Skip scoring and send the same recommendation to every customer.

Best answer: A

Explanation: The requirement is large-scale scheduled scoring, not interactive prediction. Batch prediction fits high-throughput offline scoring and produces reusable outputs for downstream systems.

Why the other choices are weaker:

• B adds online serving complexity when latency is not needed.
• C is not scalable or repeatable.
• D ignores personalization and model use entirely.

What this tests: Online vs batch prediction, serving constraints, and throughput-oriented design.

Related topics: Batch prediction; Serving; Throughput; Vertex AI

---

Question 3

Topic: Monitoring drift after deployment

A fraud model was accurate at launch, but the business suspects transaction patterns have changed. The team wants to detect whether production feature distributions are moving away from training data and trigger investigation. What should be added?

• A. Delete all training metadata to reduce storage cost.
• B. Only monitor endpoint CPU utilization.
• C. Monitor feature skew or drift and alert when production distributions exceed defined thresholds.
• D. Retrain every minute regardless of data volume or validation results.

Best answer: C

Explanation: The scenario is about model-input behavior changing after deployment. Drift or skew monitoring compares production signals against expected distributions and gives the team an investigation trigger.

Why the other choices are weaker:

• A removes useful lineage and comparison data.
• B monitors infrastructure health, not model data quality.
• D is expensive and unsafe without validation gates.

What this tests: MLOps monitoring, drift detection, alerting, and post-deployment model governance.

Related topics: Model monitoring; Drift; Skew; MLOps

---

Question 4

Topic: Evaluating a grounded GenAI assistant

A team builds a support assistant that uses retrieval over approved product documentation. Before release, they need evidence that answers are grounded, relevant, safe, and not exposing sensitive content. Which step is most appropriate?

• A. Evaluate the assistant with representative prompts, source-grounding checks, safety checks, and human review for high-risk cases.
• B. Release it immediately because retrieval always prevents hallucinations.
• C. Disable logging so unsafe responses cannot be audited.
• D. Measure only average response length because longer answers are more complete.

Best answer: A

Explanation: Grounded GenAI systems still need evaluation. Representative prompts, grounding checks, safety review, and escalation for high-risk outputs provide release evidence beyond simple fluency.

Why the other choices are weaker:

• B overtrusts retrieval and skips validation.
• C weakens traceability and incident response.
• D measures a superficial proxy instead of quality and safety.

What this tests: GenAI evaluation, grounding, responsible AI, and release readiness.

Related topics: Generative AI; Grounding; Evaluation; Responsible AI

Independent study note

Tech Exam Lexicon and IT Mastery are independent study tools. They are not affiliated with, endorsed by, or sponsored by Google Cloud or any certification body.