Google Cloud Professional Machine Learning Engineer · 13% practice allocation

Google Cloud Professional Machine Learning Engineer: Monitoring ML solutions Practice Questions

5 free sample questions from the Monitoring ML solutions 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.

A team is monitoring costs for a Vertex AI training workload that uses a large cluster of A100 GPUs. They notice that GPU utilization reported in Cloud Monitoring averages only 35% throughout training runs. Which THREE actions would most effectively improve GPU utilization and reduce the cost per training run? (Select THREE)

A. Enable Vertex AI TensorBoard to visualize training curves and identify convergence issues
B. Migrate from synchronous data-parallel training to a larger single-GPU instance to avoid communication overhead
C. Profile the training job using Cloud Profiler or TensorBoard's trace viewer to identify bottlenecks in the data pipeline or host-to-device transfers✓ Correct
D. Increase the per-device batch size to better saturate GPU memory and compute✓ Correct
E. Switch to preemptible (Spot) VMs for the training cluster to reduce per-hour costs
F. Use a tf.data pipeline with prefetching and parallel data loading to eliminate CPU-bound data bottlenecks✓ Correct

Explanation

Low GPU utilization (35%) typically means GPUs are starved for data or compute is bottlenecked elsewhere. Increasing batch size (D) directly increases the amount of work per GPU step, improving utilization. Optimizing the data pipeline with tf.data prefetching and parallel loading (F) eliminates CPU-bound I/O bottlenecks that leave GPUs idle waiting for data. Profiling with Cloud Profiler or TensorBoard's trace viewer (C) is essential to identify the exact bottleneck (e.g., data loading, host-to-device transfers, kernel launch overhead) before applying further fixes. Option A (TensorBoard for training curves) helps with convergence analysis but does not directly address GPU utilization. Option B (migrating to a single GPU) reduces parallelism and would increase training time, not improve efficiency; it may also not be feasible for large models. Option E (Spot VMs) reduces hourly cost but does not improve GPU utilization percentage and introduces the risk of preemption mid-run, which can waste already-consumed compute.

Your organization maintains a Vertex AI endpoint serving a tabular regression model that predicts customer lifetime value. You've configured prediction logging to a BigQuery table and set up alerting based on prediction statistics. You notice that the 95th percentile of predictions has jumped from $8,500 to $45,000 overnight. The model architecture and weights have not changed. Which of the following are the most likely explanations? (Select TWO)

A. A batch of high-value enterprise customers were onboarded to production, shifting the input feature distribution toward higher-income segments✓ Correct
B. The feature store used for serving predictions is returning NULL values for missing categorical features, which the model interprets as the most recent observed value, causing systematic overestimation
C. The model's prediction logging configuration is sampling only 10% of requests, and by chance, this sample contains disproportionately high-value customers
D. A recent data pipeline update changed how currency conversion is applied to the input features, causing all monetary feature values to be multiplied by a scaling factor✓ Correct
E. The endpoint's online prediction latency has increased from 50ms to 500ms, forcing the model to make inferences on stale feature data
F. The training dataset used for model development contained outliers in the high-value range, and production traffic now resembles those outlier patterns more closely

Explanation

The correct answers are options A and D. Both represent legitimate data or feature distribution changes that would cause prediction values to shift without model weight changes. Option A reflects a natural shift in the input data distribution (covariate shift), which is expected to produce higher predictions if the model encounters customers with genuinely higher lifetime value indicators. Option D indicates a systematic feature transformation issue (feature drift) where the same underlying business data is being represented differently in serving than in training. Option B is incorrect—NULL handling wouldn't cause systematic overestimation; it would more likely cause systematic bias in a single direction and would show up as errors. Option C is incorrect because sampling bias wouldn't cause the 95th percentile to uniformly shift; it might increase variance but the true distribution shift would persist. Option E is incorrect because prediction latency doesn't directly affect prediction values, only the staleness of input features (though this could cause some drift, the sudden overnight jump suggests a more fundamental data change). Option F is plausible but is less likely than A and D because training outliers are typically identified and handled; production data suddenly 'resembling outliers more' would be a consequence of A.

You are managing costs for a Vertex AI training job that trains a large language model using 8 A100 GPUs for 72 hours. The job completes successfully but you notice the average GPU utilization is only 40%, and the job took 72 hours when you estimated 48 hours based on similar historical jobs. Logs show no errors or warnings. Which of the following might explain the difference, and what action should you take? (Select TWO)

A. The job may be bottlenecked on I/O operations (data loading from Cloud Storage). Review the data pipeline configuration and consider using Vertex AI TensorBoard to profile data loading.✓ Correct
B. Batch size is too small. Increasing the batch size will improve GPU utilization and reduce training time.
C. The model was deployed to a region with lower-performance A100 GPUs. Use gcloud to check the GPU generation and relocate to a region with A100-80GB GPUs.
D. The job may be bottlenecked on CPU or synchronization overhead (gradient aggregation, communication between GPUs). Review distributed training configuration and consider adjusting the batch size or gradient accumulation steps.✓ Correct

Explanation

Correct answers are A and D. Low GPU utilization (40%) despite successful completion indicates a bottleneck in the training pipeline, most commonly I/O-bound data loading (A) or CPU-side operations like gradient synchronization in distributed training (D). Both are common causes of GPUs sitting idle waiting for data or computation. B is partially related (batch size affects both) but is vague—increasing batch size without understanding the bottleneck might not help; if bottlenecked on data loading, larger batches might make it worse. C is incorrect because the job completed successfully without errors, suggesting the GPU type itself is not the issue; also, A100s in a region don't vary significantly in performance, and the problem is likely not GPU hardware but rather pipeline bottlenecks.

You deployed a sentiment analysis model on Vertex AI that processes customer reviews in English. After two weeks, you notice that prediction latency has increased from 50ms to 200ms, but the model version, hardware (n1-standard-4 machines), and input data size have not changed. CPU utilization is at 45% and memory usage is at 30%. What should you investigate FIRST?

A. The model container needs more CPU. Scale up to n1-standard-8 machines immediately.
B. Check the prediction logging configuration. Excessive logging or inefficient serialization of prediction objects may be consuming latency overhead.✓ Correct
C. The input data quality has degraded, causing the model to require more inference steps. Validate the input schema and check for unexpected null values.
D. Model performance has drifted. Retrain the model to restore latency.

Explanation

The correct answer is B. Since model version, hardware, and input size are unchanged, but latency has increased 4x while resource utilization remains low, the most likely culprit is the prediction pipeline configuration—specifically prediction logging overhead. Logging can add significant latency through serialization, I/O, or network operations if misconfigured. A is incorrect because CPU and memory utilization are low (45% and 30%), indicating the instance has available capacity; scaling up would not be the first troubleshooting step. C is wrong because there's no indication of data quality changes, and degraded data quality would not cause latency to increase while the model logic remains the same. D is incorrect because model drift affects prediction accuracy, not latency, and the model version is unchanged.

Your team runs a Vertex AI custom training job that trains a large recommendation model on 500GB of data using a single A100 GPU. The job runs for approximately 12 hours and costs $150 per training run. Recently, you've noticed that training loss converges after about 6 hours, and additional training time yields minimal performance improvement. You want to reduce both training time and cost while maintaining model quality. What is the recommended approach?

A. Switch to a smaller GPU (V100) to reduce per-hour costs, and accept the longer training time
B. Use Vertex AI Hyperparameter Tuning to automatically identify optimal learning rate and batch size, which will accelerate convergence
C. Reduce the batch size to force more gradient updates, allowing the model to converge faster
D. Implement early stopping with a validation set and adjust the patience parameter to stop training when validation loss plateaus✓ Correct

Explanation

The correct answer is option D. Early stopping is the standard approach to avoid wasting compute resources when a model has converged. By monitoring validation loss and stopping when improvement plateaus, you reduce both training time and cost without sacrificing quality. Option C is incorrect because reducing batch size increases training time (more updates needed) and doesn't address the fundamental issue that the model has already converged. Option A is incorrect because it reduces hardware cost but increases wall-clock time, which doesn't solve the actual problem of wasted compute after convergence. Option B is incorrect because while hyperparameter tuning could improve convergence, it's an orthogonal concern; the scenario already shows convergence is happening—we just need to stop early.

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

Monitoring ML solutions study guide (Professional Machine Learning Engineer) →

Other Professional Machine Learning Engineer domains

← All Google Cloud Professional Machine Learning Engineer study resources