PDE Practice Test 26 — 15 Questions

Question 1

A company needs to serve predictions for a model that runs an expensive computation on each request. The model is used by a batch job that processes millions of records each night, and also by a real-time API for a few thousand queries per hour. Which prediction strategy minimizes cost and latency for both use cases?

Accepted Answer

Use Vertex AI batch prediction for the nightly job and a separate online endpoint with auto-scaling for the real-time API.. Using batch prediction for the batch job and a separate online endpoint with a smaller machine or auto-scaling for real-time queries optimizes cost and latency. Option D is correct. Option A is wrong because batch prediction alone doesn't serve real-time. Option B is wrong because online endpoint for millions of records is expensive. Option C is wrong because using the same endpoint for both may cause interference.

Answer

Deploy two identical models, one on a Compute Engine VM for batch, one on Vertex AI for online, and synchronize updates.

Answer

Use Vertex AI batch prediction for both workloads.

Answer

Use a single online Vertex AI endpoint with auto-scaling to handle both workloads.

Question 2

A production model deployed on Vertex AI Endpoint is experiencing high latency during traffic spikes. The current configuration uses a single replica. What is the most efficient solution?

Accepted Answer

Enable autoscaling with minReplicaCount=1 and maxReplicaCount=10. Enabling autoscaling with a min replica count ensures always-on capacity and scales up during spikes. Using a larger machine type might help but is less dynamic. Using batch prediction doesn't solve real-time latency. Increasing min replicas without autoscaling leaves resources idle at quiet times.

Answer

Set a higher min replica count (e.g., 3)

Answer

Use a larger machine type (e.g., n1-highmem-8)

Answer

Switch to batch prediction to handle spikes

Question 3

Refer to the exhibit. What is the cause of this error?

Accepted Answer

The machine type flag is only used during model deployment, not endpoint creation. The --machine-type flag is not valid for the endpoints create command; it should be specified when deploying a model to the endpoint using 'gcloud ai endpoints deploy-model'. The user must first create an endpoint without machine type, then deploy a model.

Answer

The endpoint name already exists

Answer

The user must specify a model name

Answer

The region is missing

Question 4

Your company uses Vertex AI Pipelines to automate the ML lifecycle. The pipeline includes training, evaluation, and deployment steps. You want to ensure that if a pipeline run fails due to a transient error (e.g., resource quota shortage), it automatically retries before marking the run as failed. What is the best way to implement this?

Accepted Answer

In the pipeline component code, implement retry logic using exponential backoff for specific exceptions.. Vertex AI Pipelines does not have built-in retry logic for failed steps. You can wrap each step's logic to catch transient errors and retry, or use a retry mechanism in the container itself. Kubeflow Pipelines' retry policy can be specified. Modifying pipeline code is the most direct way.

Answer

Configure Vertex AI Pipelines to automatically restart failed runs.

Answer

Set a high timeout value for the pipeline so that transient errors resolve before timeout.

Answer

Use Cloud Tasks to schedule pipeline runs and retry upon failure.

Question 5

A team uses Vertex AI Pipelines to automate retraining of a model every month. The pipeline includes data preprocessing, training, and deployment steps. After a recent update, the pipeline fails intermittently with a timeout error during the deployment step. What is the most likely cause?

Accepted Answer

The trained model size has increased due to more data, causing the deployment step to time out. Option D is correct because a model with increased size due to training on more data can cause the deployment step to time out. Option A (BigQuery quotas) would affect preprocessing, not deployment. Option B (insufficient service account permissions) would cause persistent errors, not intermittent. Option C (pipeline step order) would cause consistent failure.

Answer

The service account used by the pipeline lacks permissions to deploy the model

Answer

The pipeline is configured to run steps in parallel, leading to resource contention

Answer

BigQuery query quotas are being exceeded during data preprocessing

Question 6

A team is training a large model using a custom container with TensorFlow on Vertex AI Training. They need to use multiple GPUs across several machines. Which strategy should they implement to maximize training throughput?

Accepted Answer

Use Vertex AI Training with a custom job specifying workerPoolSpecs and MultiWorkerMirroredStrategy. Vertex AI supports multi-worker distributed training with the 'distribution_strategy' argument in the custom job config. Using a single VM with multiple GPUs is limited by that machine's capabilities. The 'MirroredStrategy' addresses single-machine multi-GPU, not multi-machine.

Answer

Use Cloud TPU Pods for distributed training

Answer

Use Dataflow for distributed training

Answer

Use a single worker with multiple GPUs and TensorFlow MirroredStrategy

Question 7

You are designing a system to serve predictions from a large language model (LLM) with a latency SLO of 500ms. The model does not fit on a single GPU and requires model parallelism. You are considering using Vertex AI Endpoints with a custom container. What additional setup is required to achieve the latency target?

Accepted Answer

Use a machine type with multiple GPUs and configure the container to use tensor parallelism.. Model parallelism across multiple GPUs on a single machine can be handled by the container using libraries like TensorFlow Distribution Strategies. Sharding across endpoints would incur network latency. Using TPUs is an alternative but not necessarily required. The key is to configure multi-GPU in the machine type.

Answer

Compile the model using TensorFlow XLA to optimize for single GPU execution.

Answer

Deploy the model across multiple endpoints and use a load balancer to send requests to different parts of the model.

Answer

Use Vertex AI Prediction as a service for LLMs, which automatically handles hardware selection.

Question 8

A financial services company deploys a fraud detection model on Vertex AI using a custom prediction container that runs a PyTorch model. The model requires GPU acceleration. The deployment succeeds but predictions return an error: 'CUDA error: out of memory'. What should the team do to resolve this issue?

Accepted Answer

Increase the GPU machine type to one with more memory (e.g., from NVIDIA T4 to A100). Option A is correct because the GPU memory is insufficient; using a machine with more GPU memory or optimizing the model is the solution. Option B (enabling model monitoring) does not fix memory. Option C (adding more CPUs) does not address GPU memory. Option D (using CPU-only) would defeat the purpose of GPU acceleration.

Answer

Change the container to use a CPU-only image to avoid CUDA errors

Answer

Enable Vertex AI Model Monitoring to automatically scale the endpoint

Answer

Add CPU replicas to distribute the inferencing load

Question 9

Which THREE metrics should be monitored to detect model drift in a production ML system?

Accepted Answer

Prediction distribution (prediction drift).. Prediction drift (distribution of predictions), feature drift (distribution of input features), and model performance metrics (e.g., accuracy) are key indicators. Infrastructure metrics (CPU usage) and training loss are not directly drift indicators.

Answer

Training loss convergence.

Answer

CPU utilization of the serving nodes.

Question 10

A company wants to version its ML models and track lineage from training data to deployed model. Which Google Cloud service should they use?

Accepted Answer

Vertex AI ML Metadata. Option B is correct because Vertex AI ML Metadata manages lineage and artifacts. Option A (Cloud Storage) is for storage only. Option C (Artifact Registry) is for container images, not ML models. Option D (Data Catalog) is for data discovery.

Answer

Cloud Storage with object versioning

Answer

Data Catalog

Answer

Artifact Registry

Question 11

An MLOps team manages a pipeline that retrains an XGBoost classifier weekly using BigQuery data. The pipeline is orchestrated with Cloud Composer and deploys the new model to Vertex AI Endpoint if validation metrics (AUC > 0.9) are met. Over the past month, the deployed model's AUC has dropped from 0.95 to 0.88, despite the training pipeline consistently reporting AUC > 0.9. Which THREE steps should the team take to diagnose and fix this issue?

Accepted Answer

Add a canary deployment step where new model version receives a small percentage of traffic before full rollout.

Answer

Review the training pipeline's hyperparameter tuning configuration to ensure it is not overfitting to stale data.

Answer

Retrain the model using a longer training history to include older data that may still be relevant.

Question 12

Which THREE steps are essential for implementing a continuous training pipeline with Vertex AI?

Accepted Answer

If the new model passes evaluation, deploy it to a production endpoint.. A continuous training pipeline involves automated retraining, evaluation, and deployment when new data or model improvements occur. Manual approval is optional, not essential. One-time manual deployment is not continuous. The three essential steps are: trigger on new data, train, and evaluate/promote.

Answer

Manually approve each new model version before deployment.

Answer

Deploy the original model once and set it to auto-update.

Question 13

A data science team wants to deploy a model that requires a custom container with specific NVIDIA CUDA version. They build the image and push to Artifact Registry. When deploying to Vertex AI, the model fails to load with an error: 'Failed to start container: invalid ELF header'. What is the most likely cause?

Accepted Answer

The container image was built for a different CPU architecture (e.g., ARM64) than the Vertex AI machine (x86_64). Option A is correct because the image was built for the wrong architecture (e.g., building on an ARM Mac for a x86 deployment). Option B (CUDA version mismatch) would cause a different error. Option C (container permissions) would cause a permission denied error. Option D (model file format) would cause loading errors but not container startup failure.

Answer

The model file (saved as .pkl) is corrupted

Answer

The CUDA version in the container is incompatible with the GPU on the machine

Answer

The container does not have the necessary permissions to access the model file in Cloud Storage

Question 14

Your company has deployed a machine learning model on Vertex AI Endpoint to serve real-time predictions for a mobile application. The model was trained using TensorFlow and the prediction requests include raw images that are preprocessed by the client before sending. Recently, the application developers reported that the predictions are becoming less accurate over time. They suspect the issue is related to changes in the client-side preprocessing code. You need to verify this hypothesis and monitor for future regressions. What should you do?

Accepted Answer

Enable Vertex AI Model Monitoring for feature attribution and set up alerting on skew detection.

Answer

Retrain the model using the latest client data to adapt to any changes in preprocessing.

Answer

Roll back to a previous model version that was known to work well and disable automatic retraining.

Answer

Ask the developers to provide the exact preprocessing code and manually compare it with the training pipeline's preprocessing.

Question 15

A company uses Vertex AI Feature Store for serving features to both training and prediction. The team notices that predictions made shortly after training use different feature values, causing a training-serving skew. What is the most effective way to prevent this skew?

Accepted Answer

Configure the Feature Store to use point-in-time lookup using the training timestamp. Option C is correct because using a feature timestamp to serve the exact feature point-in-time that was used during training ensures consistency. Option A (retraining more often) does not prevent skew. Option B (batch prediction) still uses current features. Option D (identical compute resources) does not affect feature values.

Answer

Retrain the model more frequently to adapt to the new feature distributions

Answer

Use batch prediction instead of online prediction to ensure consistent features

Answer

Ensure that the training and prediction environments use identical compute resources