PMLE Practice Test 36 — 15 Questions

Question 1

An ML engineer runs this command to upload a model. The model artifact in Cloud Storage is a directory containing model.pkl and a custom preprocessing script. What will happen when he later deploys this model to an endpoint and sends a prediction request?

Accepted Answer

The prediction will fail because the custom preprocessing script is not a standard scikit-learn serialized object.. Option C is correct because the pre-built container for scikit-learn expects a single serialized model file (e.g., model.pkl) as the artifact. A directory containing a custom preprocessing script is not a standard scikit-learn serialized object, so the container cannot load or execute it, causing the prediction to fail.

Answer

The prediction will succeed because the pre-built container automatically detects and uses the custom preprocessing script.

Answer

The prediction will succeed only if he also specifies a custom prediction routine.

Answer

The prediction will fail because the artifact URI must point to a single file not a directory.

Question 2

A machine learning engineer is scaling a prototype natural language processing model that uses a transformer encoder. The prototype was trained on a small corpus on a single GPU. For production, they need to train on a much larger corpus using TPUs on Vertex AI. They convert the TensorFlow code to work with TPUStrategy. The training starts but after a few steps, the loss becomes NaN and training diverges. The learning rate scheduler uses a warm-up and then linear decay. The initial learning rate is 1e-4. The batch size per TPU core is 32, with 8 cores total (batch size 256). What is the most likely cause?

Accepted Answer

The learning rate is too high for the batch size.. When scaling from a single GPU to 8 TPU cores, the global batch size increases from 32 to 256. The learning rate of 1e-4, which was appropriate for batch size 32, becomes too high for the larger batch size. This violates the linear scaling rule (learning rate should be scaled proportionally to batch size), causing gradient updates to overshoot minima and leading to NaN loss and divergence.

Answer

The batch size is too small for TPU.

Answer

The learning rate schedule should be cosine instead of linear.

Answer

The warm-up steps are insufficient.

Question 3

Which THREE factors should be considered when choosing a compute option for serving a deep learning model in production on Google Cloud? (Choose three.)

Accepted Answer

Integration with Vertex AI for model monitoring. A is correct because Vertex AI provides integrated model monitoring capabilities, including feature drift detection, prediction skew analysis, and outlier detection, which are essential for maintaining model performance in production. Without this integration, you would need to build custom monitoring pipelines, increasing operational complexity.

Answer

The programming language used for training

Answer

The color of the team's logo

Question 4

A large e-commerce company deploys a recommendation model on Vertex AI with autoscaling enabled. During Black Friday, traffic spikes rapidly. The autoscaler adds new instances, but new instances take several minutes to become ready (cold start). As a result, many requests time out. What should they do to mitigate this issue?

Accepted Answer

Set a higher minimum number of instances to handle the expected peak.. Option D is correct because setting a higher minimum number of instances ensures that a baseline capacity is always running and ready to serve traffic. This pre-warms instances, eliminating the cold-start latency during rapid traffic spikes, such as Black Friday, because new instances do not need to initialize from scratch.

Answer

Use a larger machine type to reduce the number of instances needed.

Answer

Configure the autoscaler to use CPU utilization metric instead of request count.

Answer

Increase the health check grace period for new instances.

Question 5

An organization runs a batch prediction job on Vertex AI for a large dataset (10 TB). The job is configured to use a cluster of 100 n1-standard-16 machines. Midway through, the job fails with 'Out of memory' errors. What is the most effective mitigation strategy?

Accepted Answer

Use a machine type with more memory per instance.. The 'Out of memory' error indicates that individual worker nodes are running out of RAM when processing their assigned data shards. Using a machine type with more memory per instance (e.g., n1-highmem-16) directly addresses the root cause by providing each node with sufficient memory to hold the model and its intermediate computations, without changing the data distribution or parallelism strategy.

Answer

Split the input data into smaller chunks and run multiple jobs.

Answer

Enable model parallelism within the prediction script.

Answer

Increase the number of machines to distribute data more.

Question 6

A team has deployed a model with autoscaling configured as shown. They notice that during off-peak hours, the endpoint consistently runs 3 instances instead of scaling down to 1. What is the most likely cause?

Accepted Answer

There is a sustained request rate that prevents scaling down.. The autoscaling configuration is likely based on a target metric (e.g., requests per second or CPU utilization). During off-peak hours, if there is a sustained but low request rate that still exceeds the scale-down threshold, the model will not reduce instances below the number needed to handle that load. The endpoint runs 3 instances because the sustained request rate prevents the scaling-down logic from triggering, even though the traffic is lower than peak.

Answer

The `enableAccessLogging` flag increases resource usage.

Answer

The `minReplicaCount` is set too high.

Answer

The model is too large to fit on a single instance.

Question 7

A team has a prototype image classification model trained on a small dataset using TensorFlow Keras on a single GPU. They need to train on a larger dataset (1 million images) using a distributed strategy on Vertex AI with 8 GPUs. They implement a MirroredStrategy for data parallelism. During the first few epochs, the training speed does not improve significantly compared to a single GPU, and GPU utilization is low. The data is stored as JPEG files in Cloud Storage, and the input pipeline uses tf.data with map to decode images. What is the most likely cause?

Accepted Answer

The data loading from Cloud Storage is a bottleneck.. Option B is correct because reading and decoding JPEG images from Cloud Storage can be I/O-bound, causing low GPU utilization. Option A is wrong because large batch size per GPU could cause memory issues but not low utilization. Option C is wrong because MirroredStrategy is typically configured correctly. Option D is wrong because even if the model is small, distributed training should still improve throughput if the pipeline is not bottlenecked.

Answer

The batch size per GPU is too large.

Answer

The MirroredStrategy is not properly configured.

Answer

The model is too small for distributed training.

Question 8

Your team manages multiple ML models in Vertex AI Model Registry. Each model has several versions deployed to different endpoints for testing and production. You need to implement a process where a model version can be promoted from a staging environment to production only after it has passed automated validation tests and been approved by a designated reviewer. The team uses CI/CD pipelines (Cloud Build) for training and deployment. Currently, model versions are deployed to endpoints using Vertex AI Endpoints with a single traffic split configuration. You want to track promotion requests and enforce approval gates. What should you do?

Accepted Answer

Use Vertex AI Model Registry version aliases ('staging', 'production') and configure Cloud Build to trigger a Cloud Run service that handles approval logic, then update the alias upon approval.. Option D is correct because Vertex AI Model Registry version aliases (e.g., 'staging', 'production') are designed to track model version lifecycle stages. By integrating Cloud Build to trigger a Cloud Run service that enforces approval logic before updating the alias, you create a clear promotion gate. This approach natively supports tracking promotion requests and enforcing approval without custom databases or manual scripts, aligning with CI/CD best practices.

Answer

Deploy each model version to a separate endpoint, and use a custom database to track which endpoint is 'production'. Then use migration scripts to switch traffic.

Answer

Store the model version metadata in a BigQuery table and use a scheduled query to automatically update the endpoint deployment based on validation results.

Answer

Use Vertex AI Model Registry labels to mark versions as 'staging' or 'production', and create a Cloud Function that checks the label before deploying to the endpoint.

Question 9

A financial services company uses Vertex AI to build credit risk models. They have a team of 10 data scientists and 3 ML engineers. They use multiple notebooks in Vertex AI Workbench, storing data in Cloud Storage and BigQuery. The team reports that training jobs sometimes fail with 'Permission denied' errors when reading from certain Cloud Storage buckets. The error occurs intermittently and only for some users. The team uses custom service accounts for each user's notebook instance, but the permissions seem inconsistent. The IT security team has enforced that all service accounts must have least privilege. What is the most effective course of action to resolve the permission issues while maintaining security?

Accepted Answer

Implement resource-level IAM policies on the specific Cloud Storage buckets used, and audit the existing service account permissions.. Option C is correct because implementing resource-level IAM policies on the specific Cloud Storage buckets used will provide the necessary permissions without overgranting. Auditing current permissions helps identify gaps. Option A grants too broad access. Option B reduces security by using a single powerful account. Option D changes the data architecture unnecessarily.

Answer

Create a single service account with broad permissions for all notebook instances and have users impersonate it.

Answer

Grant all data scientists the 'Storage Admin' role on the project to ensure they can access any bucket.

Answer

Move all training data to BigQuery to avoid Cloud Storage permission issues.

Question 10

A financial services company uses a custom container to serve a fraud detection model on Vertex AI Endpoints. The model requires a feature store lookup for each prediction. Recently, the feature store (Cloud Bigtable) experienced a brief outage, causing some predictions to fail. After the outage resolved, the endpoint's CPU utilization dropped significantly, and prediction latency improved. However, the model's false positive rate increased sharply. The ML engineer suspects the model is using stale features because the feature store outage caused missing lookups. Cloud Monitoring for the endpoint shows no errors after the outage, but the number of feature store read requests per prediction decreased by 30%. Which metric should the engineer use to confirm the hypothesis of stale features?

Accepted Answer

Use Vertex AI Model Monitoring to compare the prediction distribution before and after the outage; significant drift indicates stale features.. Option B is correct because Vertex AI Model Monitoring can detect prediction distribution drift, which directly indicates that the model is receiving different input features than expected. A significant drift after the outage, combined with the 30% drop in feature store read requests, confirms that stale or default features were substituted for missing lookups, causing the false positive rate to spike.

Answer

Monitor the prediction request latency to see if it remains low.

Answer

Verify the feature store's read throughput and latency metrics to ensure it is healthy.

Answer

Check the error rate for the endpoint; if no errors, then features were retrieved correctly.

Question 11

A retail company has deployed a machine learning model using Vertex AI Endpoints to predict inventory demand. The model was trained on data from the past two years and has been in production for six months. The team has enabled Vertex AI Model Monitoring to track prediction drift with an alert threshold of 0.2. Last week, they received an alert that the prediction drift score reached 0.35, exceeding the threshold. The engineer checks the monitoring dashboard and sees that the distribution of predictions has shifted noticeably compared to the training data. The engineer also notices that the model's accuracy metrics, computed from weekly ground truth data, have remained within acceptable range. What should the engineer do first?

Accepted Answer

Investigate the input feature distributions for the recent serving requests to identify if data drift is the underlying cause of the prediction drift.. The prediction drift alert indicates a shift in prediction distribution, but accuracy is stable. This suggests data drift (change in input features) rather than concept drift. The engineer should first investigate input feature distributions to confirm if data drift is the cause. Retraining (A) is premature without root cause analysis. Increasing the threshold (C) ignores the underlying issue. Rolling back (D) may not help if the previous version also suffers from the same data drift.

Answer

Increase the prediction drift alert threshold to 0.4 to reduce the number of false alerts.

Answer

Retrain the model using the latest three months of data to incorporate recent trends.

Answer

Roll back to an earlier model version that had lower prediction drift.

Question 12

Your team has deployed a PyTorch model using a custom container on Vertex AI Prediction. The model uses dynamic batching to combine incoming requests. You notice that the average latency is 150 ms, but the 99th percentile latency is 2 seconds. Cloud Monitoring shows that the CPU is idle much of the time, and GPU utilization is around 70%. The model is deployed on a single n1-standard-4 with a T4 GPU. You suspect the issue is related to request queuing. Which change would most effectively reduce tail latency?

Accepted Answer

Add a second replica to share the load.. Option C is correct because adding a replica reduces the queue length per replica, thus reducing waiting time for requests. Option A might increase tail latency if timeout is too long. Option B could reduce processing time but not queuing delay. Option D adds complexity and doesn't address root cause.

Answer

Increase the batch timeout to allow larger batches to form, reducing the number of batches.

Answer

Decrease the batch size to reduce processing time per batch.

Answer

Implement a priority queue to handle high-priority requests first.

Question 13

You are using Vertex AI Training to train a model and then automatically deploy the best candidate to a Vertex AI Prediction endpoint via the Vertex AI Model Registry. However, after deployment, you notice that the endpoint returns predictions for the new model, but they are significantly different from the evaluation metrics computed during training. The training scripts used TensorFlow with a serving input function. What is the most likely issue and how would you fix it?

Accepted Answer

The serving input function's preprocessing steps do not match the training preprocessing; you should verify and align them.. Option B is correct because the serving input function's preprocessing must match training preprocessing exactly; any mismatch causes prediction errors. Option A is wrong because the model saved includes evaluation metrics. Option C is possible but less likely given the consistency of difference. Option D is unlikely as numerical precision differences are minimal.

Answer

The endpoint is using a different machine type affecting numerical precision; you should use the same machine type as training.

Answer

The model registry deployed a different version; you should check the alias.

Answer

The model was saved with training-only metrics; you should retrain with evaluation metrics.

Question 14

Your team has deployed a scikit-learn model using a custom container on Vertex AI Prediction. The model receives about 100 requests per second, and the endpoint is configured with a single n1-standard-4 machine. You notice that response times are around 200 ms on average, but occasionally spike to over 10 seconds during traffic bursts. You have set the min replicas to 1 and max replicas to 10. Despite this, spikes still occur. What is the most likely cause and the best course of action?

Accepted Answer

The autoscaling is too slow to react; you should increase the max replicas to 20 and reduce the cooldown period.. Option A is correct because the occasional spikes suggest that autoscaling is too slow; enabling batching reduces the number of inference calls and smooths out bursts. Option B could help but the autoscaling may still be too slow. Option C is not necessarily needed if average latency is acceptable. Option D is unlikely to cause intermittent spikes.

Answer

The model is not optimized for parallel inference; you should enable batching in the custom container.

Answer

The machine type is insufficient for the model size; you should switch to a n1-highmem-8.

Answer

The container has a memory leak; you should restart the container periodically.

Question 15

You are responsible for deploying a real-time recommendation model that uses a large embedding table (5 GB) and a small neural network. The model is served through a custom container on Vertex AI Prediction. The end-to-end latency requirement is under 200 ms. During load testing with 500 QPS, you observe that latency increases linearly with batch size. You are currently using a single replica with an n1-standard-8 machine and one T4 GPU. The embedding table is loaded entirely in GPU memory. However, CPU utilization is at 100% while GPU is at 30%. What is the best approach to meet the latency requirement at scale?

Accepted Answer

Use a machine type with more CPU cores to parallelize embedding lookups.. Option D is correct because CPU is the bottleneck; using a machine type with more CPU cores (e.g., n1-highcpu-16) allows parallel embedding lookups and reduces latency. Option A increases resources but not in the bottleneck area. Option B increases replicas but each would still be CPU-bound. Option C is expensive and may not improve latency if model not T PU-compatible.

Answer

Increase the number of replicas and use a global load balancer to distribute traffic.

Answer

Use a custom container that partitions the embedding table across multiple GPUs within a single replica.

Answer

Switch to a TPU v2-8 pod slice to accelerate embedding lookups.