Free Professional Machine Learning Engineer exam Practice Questions

Sample Practice Questions

Try these 10 sample questions from the Professional Machine Learning Engineer exam practice bank. Each includes the answer and explanation. Continue with the interactive practice above to review the full question set.

1. 1An analyst must predict a continuous delivery cost from tabular data using BigQuery ML. Which model directly matches this target?

   1. A. Logistic regression
   2. B. K-means clustering
   3. C. Linear regression
   4. D. Principal component analysis

   Answer: C

   Explanation: The documentation states, "Linear regression is for predicting the value of a numerical metric for new data". Delivery cost is a numerical target; logistic regression predicts classes, k-means finds segments, and PCA reduces dimensions.

   Reference: Introduction to BigQuery ML — Supported models > Internally trained models > Linear regression

2. 2Where is historical offline feature data maintained in the current Agent Platform Feature Store architecture?

   1. A. In a required separate Feature Store offline database
   2. B. In the endpoint's prediction request logs
   3. C. In BigQuery tables or views
   4. D. In the model's container image

   Answer: C

   Explanation: The overview states, "the BigQuery tables or views containing the feature data collectively form the offline store." Feature Store uses this BigQuery data rather than requiring a separate offline database or treating logs or model images as the feature store.

   Reference: Feature Store overview — Feature Store overview > Introduction

3. 3A research team requires its own Python model architecture and training loop. It wants full control of both while using managed Google Cloud infrastructure. Which approach fits?

   1. A. Use managed AutoML with its supported architecture search
   2. B. Tune the parameters of a hosted pretrained foundation model
   3. C. Run the team's Python code with managed custom training
   4. D. Train a built-in model type through BigQuery ML SQL

   Answer: C

   Explanation: The overview says that providing your own training code gives "full control and flexibility over your model's architecture and training logic". Managed custom training runs that code. AutoML search, hosted model tuning, and built-in SQL model types do not give this same control over an arbitrary Python architecture and training loop.

   Reference: Training overview — Run custom training code on Agent Platform

4. 4A custom inference container listens only on 127.0.0.1 and cannot receive Agent Platform probes. Which server binding meets the documented requirement?

   1. A. Bind to 0.0.0.0 on the configured container port
   2. B. Bind to 127.0.0.1 on the configured container port
   3. C. Bind to 0.0.0.0 on a different container port
   4. D. Bind to the loopback interface on a different port

   Answer: A

   Explanation: The requirements state, "Your HTTP server must listen for requests on 0.0.0.0". The listener must also use the configured port. Loopback-only binding prevents external probes from reaching the server, and a mismatched port sends probes to a different listener address.

   Reference: Custom container requirements for inference — Container image requirements > Run the HTTP server

5. 5A pipeline's evaluation task must receive the model produced by its training task as a declared model-artifact input. How should that dependency be represented?

   1. A. Place evaluation after training in the displayed task list
   2. B. Bind the training model output to the evaluation model input
   3. C. Add only an execution-order dependency, leaving the model input unbound
   4. D. Give both components the same model-path default without connecting tasks

   Answer: B

   Explanation: The guide describes tasks as "interconnected using input-output dependencies". Binding the model output to the evaluation input supplies both the artifact and its data dependency. Display order, order-only sequencing without the model input, or a shared default path does not express the required artifact transfer.

   Reference: Introduction to Agent Platform Pipelines — Structure of an ML pipeline

6. 6A security team wants to detect malicious instructions that attempt to override an LLM's guardrails. Which Model Armor filter addresses that risk directly?

   1. A. Prompt injection and jailbreak detection
   2. B. Malicious URL detection alone
   3. C. Sensitive Data Protection's credit-card detector alone
   4. D. A topicality rule restricted to a product catalog

   Answer: A

   Explanation: The overview describes prompt injection as crafted commands that "make the AI ignore its usual instructions". Prompt injection and jailbreak detection targets that behavior. URL reputation, a specific PII detector, and topicality alone do not detect instruction override as such.

   Reference: Model Armor overview — Model Armor filters > Prompt injection and jailbreak detection

7. 7A BigQuery ML model must assign each transaction to one of three known risk categories using labeled examples. Which model type fits this requirement?

   1. A. Logistic regression
   2. B. Linear regression
   3. C. K-means clustering
   4. D. Matrix factorization

   Answer: A

   Explanation: The documentation states, "Logistic regression is for the classification of two or more possible values". The three known risk categories form a supervised multiclass target. Linear regression predicts numerical values, k-means finds unlabeled clusters, and matrix factorization learns recommendation factors rather than this classifier.

   Reference: Introduction to BigQuery ML — Supported models > Internally trained models > Logistic regression

8. 8A team serves features directly from a BigQuery source without registering them in Feature Registry. Which requirement must its ID column satisfy?

   1. A. The ID alone must uniquely identify each row
   2. B. Only the ID-and-timestamp pair must be unique
   3. C. IDs need be unique only within each source partition
   4. D. Repeated IDs are allowed if feature values differ

   Answer: A

   Explanation: The overview cautions, "If you choose not to register your features, you must ensure that every row contains a unique value in the ID column." Uniqueness is required for the ID itself across the source. Adding timestamps, partition boundaries, or different feature values does not satisfy that requirement when IDs repeat.

   Reference: Feature Store overview — Feature Registry setup > Caution

9. 9A tabular classification team wants automated data preparation, model selection, and hyperparameter tuning with minimal training-code development. Which approach directly provides that automation?

   1. A. Use Agent Platform AutoML's model-development workflow
   2. B. Implement preprocessing and search in a custom training job
   3. C. Write a notebook loop that manually compares trial models
   4. D. Specify preprocessing and tuning for a self-managed training cluster

   Answer: A

   Explanation: The overview states, "AutoML can automate tasks like data preparation, model selection, hyperparameter tuning, and deployment". AutoML directly provides the requested development automation. The alternatives require the team to implement or manage the preparation and search workflow itself.

   Reference: Training overview — AutoML

10. 10A custom container's entrypoint launches its HTTP server in the background and then exits. The container immediately stops. What should be changed?

    1. A. Start the server in the background and end the startup shell
    2. B. Run the server as the foreground process of the entrypoint
    3. C. Run a one-time server check and then end the entrypoint
    4. D. Add a short startup sleep and then exit the entrypoint

    Answer: B

    Explanation: The requirements state, "ensure that this ENTRYPOINT instruction runs indefinitely." Keeping the server as the foreground process keeps the entrypoint alive. A background launch, one-time check, or finite startup delay followed by exit still terminates the container's main process.

    Reference: Custom container requirements for inference — Container image requirements > ENTRYPOINT