Google Professional Machine Learning Engineer Question 304
Single answerGoogle Cloud PlatformYou have trained two machine learning models: a PyTorch model for image classification and an XGBoost model for customer churn prediction. You need to deploy both models on Google Cloud and ensure they can be served simultaneously with minimal operational complexity while scaling independently. Which approach should you take?
- A
Deploy both models using Vertex AI Prediction with custom containers.
- B
Export the PyTorch model to TensorFlow SavedModel format and deploy both models using Vertex AI Prediction with pre-built containers.
- C
Deploy the PyTorch model using Vertex AI and the XGBoost model using Cloud Run.
- D
Package both models together into a single custom container and deploy it on Google Kubernetes Engine (GKE).
Show answer and explanation
Correct answer: A
Explanation
Vertex AI Prediction with custom containers is the optimal solution for deploying models built with different frameworks. It simplifies operations by providing a unified deployment platform, supports independent scaling of models, and reduces the need for manual management. Other approaches either introduce unnecessary complexity or lack scalability and manageability.
- A. Correct.
This is the correct approach. Vertex AI Prediction supports custom containers, allowing you to deploy models built with different frameworks like PyTorch and XGBoost while ensuring scalability and operational simplicity.
- B. Incorrect.
Exporting the PyTorch model to TensorFlow SavedModel format introduces unnecessary complexity and is not required, as Vertex AI Prediction supports custom containers for serving models built with different frameworks.
- C. Incorrect.
Deploying the models on different services (Vertex AI and Cloud Run) adds operational complexity and makes scaling them independently harder to manage compared to using a unified solution like Vertex AI Prediction with custom containers.
- D. Incorrect.
Packaging both models into a single custom container and deploying them on GKE increases operational complexity, as you would need to manage scaling and routing requests manually, making this option suboptimal.