Google Cloud Professional Machine Learning Engineer · 14% practice allocation

Google Cloud Professional Machine Learning Engineer: Collaborating within and across teams Practice Questions

5 free sample questions from the Collaborating within and across teams domain, with answers and explanations. The full Professional Machine Learning Engineer bank has 251 questions across 6 domains. Questions are scenario-based and written from the official Google Cloud exam guide.

Your team is implementing a TFX data validation component in a production pipeline to detect data drift between training and serving datasets. The team is debating how to handle schema mismatches. Which two of the following are valid approaches within TFX's responsible AI and data quality framework? (Select TWO.)

A. Disable schema validation in production to avoid false positives from minor data drift
B. Use the TFX SchemaGen component to generate a schema from training data, then use ExampleValidator to compare serving data against it and alert on anomalies✓ Correct
C. Use TFDV (TensorFlow Data Validation) to compute detailed statistics and anomalies, integrating alerts into your monitoring workflow✓ Correct
D. Manually inspect serving data each week and update the schema in a text file when drift is detected
E. Configure the ExampleValidator to silently drop any serving examples that violate the learned schema

Explanation

Options 0 and 3 are correct because they represent systematic, automated approaches to detecting and responding to data quality issues. SchemaGen + ExampleValidator create a reproducible baseline; TFDV provides detailed statistical analysis and is the standard TFX component for drift detection. Option 1 is wrong because manual inspection is not scalable and does not fit the TFX pipeline paradigm. Option 2 is wrong because silently dropping data introduces hidden biases and violates responsible AI principles of transparency. Option 4 is wrong because disabling validation removes safeguards against serving data degradation, a key risk in production systems.

Your data science team has discovered that a classification model for loan approval shows significantly different false positive rates across demographic groups. Before fixing the model, the team wants to understand and document this fairness issue transparently for stakeholders. Which combination of tools and practices best supports this responsible AI goal? (Select THREE.)

A. Integrate SHAP or similar explainability library in your Vertex AI Experiments metadata to track which features differ in importance across demographic groups
B. Create a Colab Enterprise notebook documenting the fairness analysis, check it into Git, and share the link with stakeholders so they can reproduce the findings✓ Correct
C. Rebalance the training dataset by removing examples from overrepresented demographic groups
D. Use Vertex AI Explainable AI (feature importance analysis) to identify which features are driving disparate predictions across demographic groups✓ Correct
E. Use Vertex AI Model Monitoring to set up alerts for fairness metrics by demographic group, enabling proactive detection of future drift✓ Correct
F. Document the fairness issue in a static PDF report and distribute it once without allowing feedback

Explanation

Options 0, 1, and 3 are correct because they form a responsible AI framework: explainability reveals why disparities occur (option 0), collaborative documentation enables transparency and reproducibility (option 1), and monitoring ensures ongoing fairness (option 3). Option 2 is wrong because silently removing data introduces hidden bias and violates the principle of transparency required for responsible AI. Option 4 is wrong because one-time static reporting does not engage stakeholders in an ongoing conversation about fairness. Option 5 is wrong because while SHAP integration is valuable, it was not presented as one of the three best tools in the standard GCP responsible AI stack; the question tests knowledge of Vertex AI's native explainability and monitoring services.

You are mentoring a junior engineer who asks why the team uses Vertex AI Workbench instead of running Jupyter locally. Which reasons are most relevant to collaborative ML development? (Select TWO)

A. Workbench provides built-in Git integration, environment management, and shareable links, enabling reproducible collaboration without each engineer managing local dependencies✓ Correct
B. Workbench automatically ensures that all team members use identical hardware specifications for training
C. Workbench integrates with Vertex AI services (Feature Store, Experiments, training pipelines), reducing the need to manually export/import data and enabling seamless collaboration with production systems✓ Correct
D. Workbench eliminates the need for version control in projects
E. Workbench is required to use Python (local Jupyter also supports Python)

Explanation

Options A and C are correct. Workbench provides Git integration, environment reproducibility, and tight integration with other Vertex AI services—all essential for collaborative, production-ready ML work. Option E is false (local Jupyter supports Python equally). Option B is false (Workbench doesn't enforce hardware specs). Option D is false (version control is always necessary). Workbench's value lies in reproducibility, collaboration, and ecosystem integration—not in replacing version control or hardware control.

An ML engineer is configuring Vertex AI Feature Store to serve features for a real-time recommendation system. The serving latency requirement is under 10 milliseconds at the 99th percentile. After deployment, the team observes p99 latency of 35 milliseconds. Which THREE configuration changes are MOST likely to reduce online serving latency to meet the target? (Choose THREE.)

A. Compress feature values using gzip before writing them to the online store.
B. Co-locate the model serving endpoint and the Feature Store in the same GCP region.✓ Correct
C. Reduce the number of features fetched per entity by removing unused features from the serving request.✓ Correct
D. Enable feature caching at the application layer (e.g., using Memorystore) for frequently accessed entity IDs.
E. Switch the feature store's storage backend from BigQuery to Cloud Spanner.
F. Increase the number of online serving nodes in the Feature Store configuration.✓ Correct

Explanation

Increasing online serving nodes (Option F) directly increases throughput and reduces queuing latency under load. Reducing the number of features per request (Option C) reduces the data retrieval payload and deserialization time, directly lowering latency. Co-locating the serving endpoint and Feature Store in the same region (Option B) eliminates cross-region network round-trip time, which is a major latency contributor. Option E is incorrect because Vertex AI Feature Store's online store already uses a low-latency storage backend; users cannot swap it for Spanner directly—this is not a configurable option. Option D (application-layer caching) can help but is not a Vertex AI Feature Store configuration change; it is an external architectural decision and may introduce staleness issues for a recommendation system. Option A (compressing values) would actually increase CPU overhead during decompression and is not a recommended practice for reducing online serving latency.

An ML team uses Vertex AI Experiments to track training runs. After several weeks, a junior engineer notices that two runs with identical hyperparameters produce slightly different evaluation metrics. The team uses TensorFlow and trains on Vertex AI Training. Which TWO actions should the team take to improve reproducibility of experiment results? (Choose TWO.)

A. Store the exact training container image digest (not just a tag) in the experiment metadata.
B. Enable deterministic operations in TensorFlow by setting the TF_DETERMINISTIC_OPS environment variable to '1'.✓ Correct
C. Switch from Vertex AI Training to on-premises GPU servers to eliminate cloud variability.
D. Set a fixed random seed in TensorFlow (tf.random.set_seed) and NumPy before training starts.✓ Correct
E. Increase the learning rate to reduce the number of training steps and minimize stochasticity.

Explanation

Setting a fixed random seed (Option D) controls stochasticity in weight initialization, data shuffling, and dropout, which are common sources of run-to-run variance in TensorFlow. Enabling TF_DETERMINISTIC_OPS (Option B) forces TensorFlow to use deterministic CUDA kernels, eliminating non-determinism from GPU parallelism. Together these two actions address the most common sources of irreproducibility. Option C is incorrect because moving to on-premises hardware does not inherently solve non-determinism and introduces new operational burdens. Option E is a training strategy change that does not address stochasticity and could degrade model quality. Option A (storing the container image digest) is a good practice for environment reproducibility but does not explain or fix metric differences when the code and environment are already the same between runs.

Test yourself across every Professional Machine Learning Engineer domain

The free Professional Machine Learning Engineer practice test mixes about 20 questions weighted like the exam. The full 251-question bank adds instant grading, flashcards and a timed mock exam.

Take the free Professional Machine Learning Engineer practice test Practice the full bank

Study this domain

Collaborating within and across teams study guide (Professional Machine Learning Engineer) →

Other Professional Machine Learning Engineer domains

← All Google Cloud Professional Machine Learning Engineer study resources