Google Cloud Professional Machine Learning Engineer — Free Practice Questions
Free Google Professional ML Engineer practice questions, every answer explained. Covers Gemini Enterprise Agent Platform (formerly Vertex AI) and MLOps.
96 questions · every answer explained · free to practise · Google Cloud ML hub
Collaborating within and across teams to manage data and models
Scaling prototypes into ML models
Serving and scaling models
Automating and orchestrating ML pipelines
Monitoring AI solutions
Sample questions with answers
8 of the 96 questions in this set, with the correct answer marked and every option explained.
1. You are an ML engineer at a media company. You need to build an ML model to analyze video content frame-by-frame, identify objects, and alert users if there is inappropriate content. Which Google Cloud products should you use to build this project?
✓Pub/Sub, Cloud Function, Video Intelligence API, Cloud Logging
C is correct as Video Intelligence API can find inappropriate components and other components satisfy the requirements of real-time processing and notification
✗Pub/Sub, Cloud Function, AutoML Video Intelligence, Cloud Logging
AutoML Video Intelligence is for custom labels you train yourself. If the pre-built API already recognises what you need, training a custom model is unnecessary work.
✗Pub/Sub, Cloud Function, Cloud Vision API
The Vision API analyses still images. Frame-by-frame analysis of video is what Video Intelligence is built for.
Cloud IoT is for device telemetry, not video ingestion, and Vision API is again the wrong service for video.
2. What is a difference between example (training) data and test data?
✗Test data have labels attached and examples do not.
Backwards, and both are labelled in supervised learning. Training labels teach the model; test labels are held back to score it.
✗Examples are used to validate training models, and test data is used to train the learning model.
The two roles are swapped. Training on your test set destroys the only unbiased estimate of performance you have.
✓Example (training) data uses its labels to train the machine learning model, while test data uses its labels to validate the model’s accuracy.
Examples have a correct label attached to them, which an ML model uses to learn how to associate features with the label. While test data also has labels, it is used to validate accuracy of the model, but not to train it.
✗Examples are used for feature engineering, and test data is used for hyperparameter tuning.
Both halves are off. Feature engineering is a preprocessing step applied across the data, and hyperparameter tuning should use a separate validation set — reusing the test set for tuning contaminates it.
3. You work for a manufacturing company that owns a high-value machine which has several machine settings and multiple sensors. A history of the machine’s hourly sensor readings and known failure event data are stored in BigQuery. You need to predict if the machine will fail within the next 3 days in order to schedule maintenance before the machine fails. Which data preparation and model training steps should you take?
✗Data preparation: Daily max value feature engineering with DataPrep; Model training: AutoML classification with BQML
A single daily maximum discards the trend information that predicts gradual machine degradation. A rolling average captures it.
✗Data preparation: Daily min value feature engineering with DataPrep; Model training: Logistic regression with BQML and AUTO_CLASS_WEIGHTS set to True
Same problem as the daily maximum — one extreme value per day loses the shape of the signal.
✓Data preparation: Rolling average feature engineering with DataPrep; Model training: Logistic regression with BQML and AUTO_CLASS_WEIGHTS set to True
D is correct because it uses the rolling average of the sensor data and balances the weights using the BQML auto class weight balance parameter.
✗Data preparation: Rolling average feature engineering with DataPrep; Model training: Logistic regression with BQML and AUTO_CLASS_WEIGHTS set to False
The feature engineering is right, but failure data is heavily imbalanced. Disabling automatic class weights lets the model predict 'no failure' every time and still look accurate.
4. You work on a team where the process for deploying a model into production starts with data scientists training different versions of models in a Kubeflow pipeline. The workflow then stores the new model artifact into the corresponding Cloud Storage bucket. You need to build the next steps of the pipeline after the submitted model is ready to be tested and deployed in production on Agent Platform. How should you configure the architecture before deploying the model to production?
✗Create a new Agent Platform model version - > Deploy model in test environment -> Validate model
Creating the production model version first means an unvalidated model already exists in the registry, which defeats the point of a gate.
✗Validate model -> Deploy model in test environment -> Create a new Agent Platform model version
You cannot meaningfully validate a model before it is deployed somewhere to be exercised.
✓Deploy model in test environment -> Validate model -> Create a new Agent Platform model version
the model can be validated after it is deployed to the test environment, and the release version is established before the model is deployed in production
✗Create a new Agent Platform model version -> Validate model -> Deploy model in test environment
Deploying to test after the production version already exists inverts the promotion path.
5. Production grade machine learning challenges are addressed by implementing an important concept:
✗Directed Acyclic Graphs (DAGs)
The structure a pipeline is expressed as — a graph of steps with no cycles. It describes the shape rather than the concept being applied.
✗Orchestrators
The tools that execute pipelines, such as Kubeflow Pipelines or Airflow. Implementation machinery rather than the underlying idea.
✓Machine learning pipelines
ML pipelines provide support for automating, monitoring and maintaining a model as you continue to train it over its lifetime.
✗Tensorflow Extended (TFX)
One specific framework for building ML pipelines. A concrete product, whereas the question asks for the general concept it implements.
6. More often than not, ML systems can fail the users it serves. In this context, what is representational harm?
✗Giving skewed outputs more frequently for certain groups of users
Describes quality-of-service harm, where a system simply works less well for some groups. Real, but distinct from reinforcing a stereotype.
✓The amplification or negative reflection of certain groups stereotypes.
Representative harms occur when systems reinforce the subordination of some groups along the lines of identity. The amplification or negative reflection of certain groups stereotypes. This is a prototypical way an ML system may fail the users it serves.
✗Making predictions and decisions that preclude certain groups from accessing resources or opportunities.
Allocative harm — the withholding of jobs, credit or services. This concerns material outcomes rather than how groups are represented.
✗Inferring prejudicial links between certain demographic traits and user behaviors.
Close, and often a mechanism that produces representational harm. But it describes the inference the model makes, whereas representational harm is the effect on how a group is portrayed.
7. Your team has proprietary support-ticket data in BigQuery and wants to adapt a Gemini model to your company's product vocabulary. The data must not leave BigQuery, and the team's skills are SQL-first with no ML engineering capacity. What is the most appropriate approach?
✓Fine-tune the Gemini model directly from BigQuery
The exam guide lists 'fine-tuning Gemini models using BigQuery' under low-code AI solutions precisely for this case: a SQL-first team adapting a foundation model without moving the data out.
✗Export the tickets to Cloud Storage and run a custom training job on Agent Platform
This moves the data out of BigQuery, which the requirement forbids, and custom training demands exactly the ML engineering capacity the team does not have.
✗Use prompt engineering alone with few-shot examples pulled from BigQuery at request time
Few-shot prompting is a reasonable first thing to try, but it does not adapt the model to a large proprietary vocabulary and it pays the token cost of the examples on every single call.
✗Train an AutoML text model on the tickets and use it instead of Gemini
AutoML would build a separate task-specific classifier. The requirement is to adapt Gemini's language to your domain, not to replace it with a narrower model.
8. You are standing up notebook environments for a team of twelve data scientists who handle regulated customer data. You need per-user isolation, VPC-internal networking, and the ability to shut down idle instances to control cost. Which environment best fits?
✓Agent Platform Workbench instances with idle shutdown and VPC Service Controls
Workbench gives each scientist a managed instance you can place inside your VPC perimeter and configure to stop when idle — the combination of isolation, network control and cost control the scenario asks for.
✗A single shared Workbench instance that the whole team logs into
A shared instance destroys per-user isolation and makes it impossible to attribute access to regulated data to an individual, which is usually the point of the audit requirement.
✗Colab Enterprise notebooks with default settings
Colab Enterprise is a genuine option for collaborative notebooks, but taking it with default settings skips the VPC and isolation configuration the regulated-data requirement turns on.
✗Local Jupyter installs on each analyst's laptop, pulling data through the BigQuery API
This pulls regulated data onto unmanaged endpoints, which is the outcome VPC Service Controls exists to prevent.
88 more questions in the app
Practise the full 96-question set with a timer, scoring and progress tracking.
