Google Professional Machine Learning Engineer Question 306
Single answerGoogle Cloud PlatformYou are tasked with deploying a machine learning model for real-time predictions on Google Cloud. The model was trained using PyTorch, but your team has other models in production that were trained using XGBoost. You need to select a serving infrastructure that accommodates both frameworks efficiently while minimizing operational complexity. Which solution should you choose?
- A
Use AI Platform Prediction and export both models to TensorFlow SavedModel format before deployment.
- B
Deploy both models to Vertex AI Prediction, which natively supports multiple frameworks including PyTorch and XGBoost.
- C
Deploy the PyTorch model to a custom Compute Engine instance and the XGBoost model to a prebuilt AI Platform Prediction container.
- D
Use Vertex AI Workbench to serve both models directly, as it supports model serving out-of-the-box.
Show answer and explanation
Correct answer: B
Explanation
Vertex AI Prediction is the recommended solution for serving models trained with different frameworks like PyTorch and XGBoost on Google Cloud. It provides native support for multiple frameworks, removes the need for exporting models to a common format, and simplifies operational overhead by using a unified serving platform.
- A. Incorrect.
AI Platform Prediction requires models to be exported to TensorFlow SavedModel format, which is unnecessary for frameworks like PyTorch and XGBoost since Vertex AI Prediction supports them natively.
- B. Correct.
Vertex AI Prediction natively supports multiple frameworks, including PyTorch and XGBoost, making it the best choice for minimizing operational complexity while deploying models trained in different frameworks.
- C. Incorrect.
While deploying models to separate infrastructure is possible, it increases operational complexity and is not an efficient solution when Vertex AI Prediction supports both frameworks natively.
- D. Incorrect.
Vertex AI Workbench is primarily a development environment for building and experimenting with models, not a serving infrastructure for production use.