PMLE sample questions

Practice making decisions in Google Cloud ML scenarios. No account is needed for these public samples or the diagnostic.

What does the PMLE test?

The Professional Machine Learning Engineer exam covers low-code AI, data and model collaboration, scaling prototypes, serving models, ML pipelines and monitoring AI solutions. Use the official exam guide to check the full scope.

How are these sample questions written?

Our public sample format uses original, scenario-based questions with four options. Each published question must be checked against Google documentation and explain every option, not just the correct answer. These are not Google exam questions. They are separate from Testero's diagnostic and practice bank.

Architecting low-code AI solutions

1. Your company recently deployed an Agent Platform AutoML classifier to predict retail order cancellations for inventory planning. Its random holdout results looked strong, but prediction quality fell after a pricing change. Historical data now contains labeled orders with dates from before and after the change. You want to estimate how a retrained model handles future orders while minimizing changes to the existing training workflow. What should you do?A. Replace AutoML with an Agent Platform custom classifier, use shuffled cross-validation to average cancellation performance across several folds of historical orders, and select the strongest candidate based on those results.B. Move training to a BigQuery ML boosted tree classifier, separate orders chronologically for training and evaluation, and assess cancellation predictions on the newest labeled orders before deploying the replacement.C. Retrain the Agent Platform AutoML classifier using manually assigned partitions: earliest orders for training, subsequent orders for validation, and newest orders for the final evaluation test set.D. Repeat Agent Platform AutoML training with the default random split, increasing the training budget and using the resulting test metrics to assess future cancellation predictions.

Architecting low-code AI solutions

2. You work for an insurer whose Gemini assistant drafts explanations of claim decisions through BigQuery ML. Monitoring shows errors on newly encountered claims. Analysts have collected corrected input-output pairs in BigQuery, with several revisions associated with each claim, and your team is preparing a tuning run. You want to estimate how the tuned model handles unseen claims while minimizing changes to the SQL workflow. Which approach should you use?A. Group revisions by claim in SQL and assign each claim a Boolean evaluation flag. Fine-tune Gemini through BigQuery ML using that custom split, then evaluate the tuned model on the reserved claims.B. Fine-tune Gemini through BigQuery ML using all corrected draft pairs as training data. Evaluate the tuned model on those same pairs to estimate how accurately it handles unseen claim explanations.C. Configure BigQuery ML to randomly split the corrected draft rows into training and evaluation data. Fine-tune Gemini with those groups, then use the tuned model's evaluation results to estimate performance on unseen claims.D. Evaluate the current Gemini model on separate unseen claims through BigQuery ML. Then fine-tune it on the corrected drafts and use the earlier evaluation to estimate the tuned model's performance.

Automating and orchestrating ML pipelines

3. You manage a greenhouse heating-demand predictor that helps plan energy purchases. Its running BigQuery ML linear regression model automatically standardizes numeric sensor readings during training and prediction. As prediction requests become more frequent, you are migrating serving to Agent Platform. You want to preserve training-time preprocessing while minimizing prediction latency. What should you do?A. Standardize incoming sensor readings in a Dataflow streaming pipeline using statistics from recent readings. Send the transformed features to the exported BigQuery ML model deployed on an Agent Platform endpoint.B. Build a custom prediction service on Agent Platform that submits each request's raw sensor readings to the existing BigQuery ML model. Return the query's heating-demand prediction to the calling application.C. Train a replacement AutoML tabular regression model in Agent Platform using historical raw sensor readings. Deploy the replacement to an endpoint and let AutoML preprocess incoming readings for heating-demand predictions.D. Export the existing BigQuery ML model to Cloud Storage and deploy it to an Agent Platform endpoint. Send raw sensor readings so the model applies its retained preprocessing.

Automating and orchestrating ML pipelines

4. Your company predicts wind-turbine failures using a nightly workflow that trains, evaluates, and deploys models. During a migration from virtual-machine scripts, your team has converted these steps into an Agent Platform Pipelines pipeline. Workflow releases and recurring production runs still need automation. You want to test and release workflow changes when code changes while minimizing repeated build work during nightly retraining. What should you do?A. Schedule a Cloud Run job with Cloud Scheduler to retrieve current repository code, compile the pipeline, and submit each nightly run. Treat successful production execution as acceptance of workflow changes for subsequent training, evaluation, and deployment.B. Configure Cloud Build triggers to test, compile, and publish pipeline versions to Artifact Registry after code changes. Update the Agent Platform Pipelines schedule to use each released version for nightly training, evaluation, and deployment.C. Reimplement the workflow in Managed Service for Apache Airflow. Configure its nightly execution to invoke Cloud Build for component tests and builds, then run the training, evaluation, and deployment tasks with those newly built components.D. Trigger Cloud Build nightly to retrieve repository code, run tests, compile the pipeline, and submit a production run. Use this recurring build to deliver workflow changes and execute training, evaluation, and deployment.

Collaborating within and across teams to manage data and models

5. Your team develops a PyTorch model that groups furniture photographs for a resale marketplace in Colab Enterprise. As the catalog grows, training epochs now take hours on the existing CPU runtime. Profiling shows most time is spent in tensor operations, and the dataset still fits in memory. You want to shorten interactive experiments while minimizing changes to the notebook workflow. What should you do first?A. Create a GPU-backed Colab Enterprise runtime, connect the existing notebook, and move the PyTorch model and training tensors onto the GPU.B. Package the PyTorch training code for GPU-backed Agent Platform custom training jobs, and submit a separate job for each notebook experiment.C. Increase the Colab Enterprise runtime's memory allocation, retain CPU execution, and rerun the existing PyTorch training cells interactively.D. Migrate the notebook to a GPU-enabled Agent Platform Workbench instance, recreate its environment, and move the PyTorch model and tensors onto the GPU.

Collaborating within and across teams to manage data and models

6. You recently deployed a repair-cost predictor that helps an insurer estimate claim expenses. Your team tests candidate scikit-learn regressors through independent Agent Platform custom training jobs. Each script trains and evaluates one candidate, and engineers copy settings and results into a shared spreadsheet. You want a shared environment for candidate-model experimentation while minimizing changes to this working execution flow. What should you do?A. Move candidate training and evaluation into Agent Platform Workbench notebooks. Execute the scikit-learn code interactively and capture each candidate's parameters and evaluation metrics in notebook cells that engineers share for review.B. Adopt Agent Platform AutoML for tabular regression experiments. Submit the repair-cost dataset for automated model training and hyperparameter search, then review the trained model's evaluation metrics through the service's evaluation interface.C. Use Agent Platform Experiments with the existing custom training jobs. Record each candidate's parameters and evaluation metrics as an experiment run through the Agent Platform SDK for Python, retaining the scikit-learn training scripts.D. Build an Agent Platform Pipelines workflow around the training and evaluation scripts using Kubeflow Pipelines components. Submit each candidate configuration as a pipeline run and associate the runs with an experiment for comparison.

Monitoring AI solutions

7. You are monitoring a courier's delivery-delay predictor on Agent Platform Inference. Dispatchers use its scores to prioritize shipments for intervention. The deployed tabular model already returns feature attributions, but your team reviews only accuracy and input distributions as the route mix changes. You want to detect shifts in which features contribute to production predictions while minimizing manual review. What should you do first?A. Monitor the predictor's delivery-delay score distributions with Agent Platform Model Monitoring. Schedule comparisons against an earlier production baseline, and alert dispatch analysts when predicted scores shift, using those alerts to assess changes in feature influence.B. Configure feature attribution drift monitoring in Agent Platform Model Monitoring for the deployed predictor. Schedule comparisons of production attribution distributions against an earlier production baseline, and alert dispatch analysts when feature contributions shift beyond the configured thresholds.C. Enable input feature drift monitoring in Agent Platform Model Monitoring for the deployed predictor. Compare shipment input distributions against an earlier production baseline, and alert dispatch analysts when shipment characteristics change beyond the configured thresholds.D. Generate a global feature-importance report by aggregating Agent Platform explanations over the predictor's training dataset. Share the resulting feature rankings with dispatch analysts, and use this historical summary to assess feature contributions as production delivery routes change.

Scaling prototypes into ML models

8. Your company is building an irrigation planner that predicts each greenhouse's next-day water demand. Historical sensor summaries, crop types, and recorded water use are available as labeled tabular data in BigQuery. Your team has prepared the training examples but has no existing training code or preferred model family. You want to automate the search across model architectures and their tuning while minimizing the modeling code your team maintains. What should you do?A. Develop a neural network with Agent Platform custom training on the greenhouse examples, using managed hyperparameter tuning to explore its architecture and training parameters for water-demand prediction.B. Create a boosted-tree regression model in BigQuery ML from the prepared greenhouse examples, using hyperparameter tuning to optimize the selected algorithm for water-demand prediction.C. Build an Agent Platform Pipelines workflow to train and compare regression algorithms on the greenhouse examples, implementing candidate models and search logic in custom components for water-demand prediction.D. Train an Agent Platform AutoML tabular regression model on the prepared greenhouse examples, using its managed model search and hyperparameter tuning to build the water-demand predictor.

Serving and scaling models

9. You are running a retail demand forecasting service that helps stores plan inventory. Its Python inference code loads an XGBoost model and uses a company Python package to adjust forecasts for local promotions. As more stores adopt the service, your team is moving model hosting to Agent Platform. You want to preserve current forecast behavior while minimizing prediction-server code your team maintains. What should you do?A. Use Agent Platform's prebuilt XGBoost prediction container to serve the existing model. Upload the Python inference code and company package alongside the model artifacts in Cloud Storage so the serving runtime can load them.B. Use Agent Platform custom prediction routines to wrap the existing XGBoost inference code in a Predictor and install the company package. Build the custom serving container with the supplied prediction server, then deploy it to an Agent Platform endpoint.C. Write a dedicated HTTP prediction server that wraps the existing XGBoost inference code and calls the company package. Install the dependencies in a custom serving container, then upload and deploy the container and model to an Agent Platform endpoint.D. Retrain the forecasting model in TensorFlow and incorporate the company's promotion adjustments into the exported model. Upload the TensorFlow model artifacts to Cloud Storage, then deploy with Agent Platform's prebuilt TensorFlow prediction container.

Serving and scaling models

10. You are building an assistant that summarizes customer support conversations. Your team has selected a Llama model from Model Garden that is available through a managed API and for self-deployment. Support agents will request summaries as they review cases, and early usage will be intermittent. You want serving costs to track actual usage during the initial rollout while minimizing serving infrastructure maintenance. What should you do?A. Build a model-serving container for Cloud Run, and configure GPU resources, model loading, and request concurrency to handle summaries requested by the support assistant.B. Host the Llama model on GKE with GPU workers, and configure the inference service, workload scaling, and cluster autoscaling to handle assistant summary requests.C. Use the Llama model's managed API in Model Garden with pay-as-you-go billing, and send summary requests directly from the support assistant.D. Deploy the Llama model from Model Garden to an Agent Platform endpoint, and configure the machine type and GPU resources to handle summary requests.

Want to check your readiness?

Take the free 20-question diagnostic

A readiness score is a study signal, not an exam-outcome guarantee.