Scenario practice questions

A retail company wants to build a product recommendation system using BigQuery ML for their e-commerce platform. The data includes customer purchase history, product metadata, and clickstream logs. The ML engineer needs to minimize manual feature engineering and leverage pre-built solutions. Which approach should the engineer take?

Trap 1: Use a pre-built recommendation model from Vertex AI Model Garden…

This is not a BigQuery ML solution and requires additional infrastructure.

Trap 2: Write a custom TensorFlow model using the Vertex AI Training…

This requires significant custom coding, not low-code.

Trap 3: Export the data to CSV and use AutoML Tables to train a…

AutoML Tables is not integrated with BigQuery ML; exporting data adds complexity.

A
Use a pre-built recommendation model from Vertex AI Model Garden and deploy it to an endpoint.
Why wrong: This is not a BigQuery ML solution and requires additional infrastructure.
B
Write a custom TensorFlow model using the Vertex AI Training service and deploy it via Vertex AI Prediction.
Why wrong: This requires significant custom coding, not low-code.
C
Export the data to CSV and use AutoML Tables to train a recommendation model.
Why wrong: AutoML Tables is not integrated with BigQuery ML; exporting data adds complexity.
D
Use BigQuery ML's matrix factorization model (CREATE MODEL with model_type='matrix_factorization') to train directly on historical interaction data.
BigQuery ML provides low-code matrix factorization for recommendations.

Question 2hardmulti select

A team is training a custom TensorFlow model on Vertex AI using a pre-built container. They need to use a TPU pod slice (v3-32). What THREE actions are required to set up the training job correctly?

Trap 1: Set the training worker pool to use only one worker

TPU pod slices often require multiple workers for best performance, but the question asks for required actions. One worker can use a TPU pod, but it's not a requirement.

Trap 2: Use a custom container with TensorFlow 2.12 and TPU support

A custom container is not necessary because Vertex AI's pre-built TensorFlow containers already support TPU training. Using a pre-built container simplifies setup and is the recommended approach.

Trap 3: Set the machine type to a high-memory VM with NVIDIA A100 GPUs

TPU training does not use GPUs; the accelerator type should be TPU.

A
Configure TF_CONFIG for distributed training
When using TPU pods with multi-worker, TF_CONFIG must be set to coordinate workers.
B
Set the training worker pool to use only one worker
Why wrong: TPU pod slices often require multiple workers for best performance, but the question asks for required actions. One worker can use a TPU pod, but it's not a requirement.
C
Specify the accelerator type as TPU_V3 and topology as '2x2x4'
For a v3-32 pod slice, topology '2x2x4' is correct (32 TPU chips).
D
Use a custom container with TensorFlow 2.12 and TPU support
Why wrong: A custom container is not necessary because Vertex AI's pre-built TensorFlow containers already support TPU training. Using a pre-built container simplifies setup and is the recommended approach.
E
Set the machine type to a high-memory VM with NVIDIA A100 GPUs
Why wrong: TPU training does not use GPUs; the accelerator type should be TPU.

Question 3mediummultiple choice

You have deployed a regression model that predicts house prices. Over the past month, the model's predictions have been consistently too high. You suspect data drift in the input features. Which monitoring metric should you prioritize to confirm this?

Trap 1: Monitor prediction drift (prediction distribution)

Prediction drift indicates output change but not which feature drifted.

Trap 2: Monitor feature attribution drift using SHAP values

Feature attribution drift measures impact on predictions, not input distribution.

Trap 3: Monitor residual distribution drift

Residuals measure error, not input drift.

A
Monitor prediction drift (prediction distribution)
Why wrong: Prediction drift indicates output change but not which feature drifted.
B
Monitor feature distribution drift using a divergence metric like Jensen-Shannon divergence
Feature drift measures input distribution change.
C
Monitor feature attribution drift using SHAP values
Why wrong: Feature attribution drift measures impact on predictions, not input distribution.
D
Monitor residual distribution drift
Why wrong: Residuals measure error, not input drift.

Question 4mediummultiple choice

A team wants to perform hyperparameter tuning on a Vertex AI custom training job with 100 trials. They require an algorithm that efficiently explores the search space by learning from previous trials. Which algorithm should they select in the study configuration?

Trap 1: RANDOM_SEARCH

Random search is simple but does not learn from previous trials, less efficient than Bayesian optimization.

Trap 2: HYPERBAND

Hyperband is not directly supported by Vertex AI Vizier; it uses Bayesian optimization, grid, or random search.

Trap 3: GRID_SEARCH

Grid search is exhaustive and inefficient for 100 trials; it works best for small discrete spaces.

A
RANDOM_SEARCH
Why wrong: Random search is simple but does not learn from previous trials, less efficient than Bayesian optimization.
B
HYPERBAND
Why wrong: Hyperband is not directly supported by Vertex AI Vizier; it uses Bayesian optimization, grid, or random search.
C
ALGORITHM_UNSPECIFIED (defaults to Bayesian optimization)
Bayesian optimization is the default and most efficient algorithm for hyperparameter tuning.
D
GRID_SEARCH
Why wrong: Grid search is exhaustive and inefficient for 100 trials; it works best for small discrete spaces.

Question 5easymultiple choice

An organization wants to use Cloud Composer (Airflow) to orchestrate a machine learning workflow that includes running a Vertex AI Pipeline, followed by a BigQuery job, and then a Dataflow pipeline. What is the primary advantage of using Cloud Composer for this orchestration?

Trap 1: It automatically caches the outputs of each step to avoid…

Caching is a feature of Vertex AI Pipelines, not Cloud Composer.

Trap 2: It integrates natively with the Vertex AI Model Registry for model…

Cloud Composer can interact with Model Registry via API, but it is not a native integration feature.

Trap 3: It provides a serverless execution environment for ML pipelines.

Cloud Composer is not serverless; it runs on GKE. Vertex AI Pipelines itself is serverless for ML pipelines.

A
It allows orchestrating heterogeneous workflows across multiple GCP services with dependencies and retries.
Cloud Composer excels at orchestrating complex DAGs that span multiple services like Vertex AI, BigQuery, and Dataflow.
B
It automatically caches the outputs of each step to avoid recomputation.
Why wrong: Caching is a feature of Vertex AI Pipelines, not Cloud Composer.
C
It integrates natively with the Vertex AI Model Registry for model versioning.
Why wrong: Cloud Composer can interact with Model Registry via API, but it is not a native integration feature.
D
It provides a serverless execution environment for ML pipelines.
Why wrong: Cloud Composer is not serverless; it runs on GKE. Vertex AI Pipelines itself is serverless for ML pipelines.

Question 6hardmultiple choice

You need to deploy a TensorFlow model to edge devices for real-time inference with minimal latency. The model is currently trained on Vertex AI. Which approach should you use?

Trap 1: Export the model as a SavedModel and deploy it to Vertex AI Edge…

Edge Manager requires the model in a format compatible with the target device; SavedModel may not be optimized for edge.

Trap 2: Deploy the model to Vertex AI endpoint and use Cloud IoT Core to…

This introduces network latency, not suitable for real-time on-device inference.

Trap 3: Use Vertex AI Model Optimization to quantize the model to INT8 and…

Vertex AI Model Optimization can quantize, but deployment to edge is best done via Edge Manager.

A
Convert the model to TensorFlow Lite using the TF Lite converter, then deploy to edge devices via Vertex AI Edge Manager.
TFLite is optimized for on-device inference; Edge Manager can deploy it to edge devices.
B
Export the model as a SavedModel and deploy it to Vertex AI Edge Manager using the Edge Manager API.
Why wrong: Edge Manager requires the model in a format compatible with the target device; SavedModel may not be optimized for edge.
C
Deploy the model to Vertex AI endpoint and use Cloud IoT Core to stream data to the cloud for inference.
Why wrong: This introduces network latency, not suitable for real-time on-device inference.
D
Use Vertex AI Model Optimization to quantize the model to INT8 and then deploy as a web service on a Raspberry Pi.
Why wrong: Vertex AI Model Optimization can quantize, but deployment to edge is best done via Edge Manager.

Question 7hardmultiple choice

A company has a prototype ML model that achieves 85% accuracy on historical data. In production, accuracy drops to 70% after two weeks due to data drift. They need an automated retraining pipeline with minimal manual oversight. Which solution is most cost-effective?

Trap 1: Use Cloud Functions to trigger a Dataflow job that trains the model…

Dataflow is for data processing, not model training; would need complex custom setup.

Trap 2: Deploy the model on a GPU-equipped Compute Engine VM and run…

Constant GPU cost, manual setup, no drift detection.

Trap 3: Schedule a weekly Cloud Composer DAG that runs a new training job…

Scheduled retraining doesn't adapt to actual drift, may waste resources.

A
Use Cloud Functions to trigger a Dataflow job that trains the model using custom containers
Why wrong: Dataflow is for data processing, not model training; would need complex custom setup.
B
Deploy the model on a GPU-equipped Compute Engine VM and run retraining every time new data arrives
Why wrong: Constant GPU cost, manual setup, no drift detection.
C
Set up Vertex AI Model Monitoring to detect drift, which triggers a Cloud Function that submits a Vertex AI Training job with new data
Monitoring detects drift, automation triggers retraining with new data, cost-effective.
D
Schedule a weekly Cloud Composer DAG that runs a new training job with all available data
Why wrong: Scheduled retraining doesn't adapt to actual drift, may waste resources.

Question 8mediummultiple choice

A developer sees this error when calling the endpoint. What is the most likely cause?

Exhibit

Refer to the exhibit:
{
  "textPayload": "Prediction failed: Model not ready or not deployed.",
  "resource": {
    "type": "ai_platform_endpoint",
    "labels": {
      "endpoint_id": "12345678",
      "model_id": "87654321"
    }
  },
  "severity": "ERROR"
}

Trap 1: The model is still in training

Wrong: A model in training is not deployed to an endpoint.

Trap 2: The endpoint has no deployed model

Wrong: Would give error like 'No model deployed'.

Trap 3: The request payload size exceeds limit

Wrong: Would be a different error message.

A
The model is still in training
Why wrong: Wrong: A model in training is not deployed to an endpoint.
B
The model is deployed but not yet serving
Correct: Model deployment is still initializing.
C
The endpoint has no deployed model
Why wrong: Wrong: Would give error like 'No model deployed'.
D
The request payload size exceeds limit
Why wrong: Wrong: Would be a different error message.