Serving Models PROFESSIONAL-MACHINE-LEARNING-ENGINEER Practice Questions (Page 3)

Questions 11–15

1. 11application · medium

   A company has a model that requires a proprietary C++ library for inference. They want to deploy it as an online service with autoscaling. The serving code must also perform feature engineering before calling the model. What is the best approach?

   • ADeploy to Vertex AI Prediction using a prebuilt container and install the C++ library at runtime.
   • BUse Vertex AI Batch Prediction with a custom container.
   • CUse GKE with a custom container, but manage autoscaling manually.
   • DBuild a custom container with the C++ library and feature engineering code, then deploy to Cloud Run.
   Reveal Solution Discussion Select an answer first
2. 12application · medium

   A team has a trained scikit-learn model and wants to deploy it to Vertex AI Prediction for online inference. They want to minimize effort and avoid building a custom container. What should they do?

   • AUse Vertex AI Batch Prediction with the prebuilt scikit-learn container.
   • BUse a prebuilt container for scikit-learn if available, otherwise build a custom container.
   • CDeploy the model to Cloud Run using a prebuilt scikit-learn container.
   • DUse the prebuilt XGBoost container and convert the model to XGBoost format.
   Reveal Solution Discussion Select an answer first
4. 13application · medium

   A startup has a custom PyTorch model with a complex preprocessing step that requires a specific library not available in Vertex AI's prebuilt containers. They want a serverless, autoscaling online endpoint with minimal infrastructure management. What should they do?

   • AUse Vertex AI Batch Prediction with a custom container.
   • BDeploy the model to Vertex AI Prediction using the prebuilt PyTorch container.
   • CDeploy to GKE with a custom container and configure cluster autoscaling.
   • DBuild a custom container with the required library and deploy it to Cloud Run.
   Reveal Solution Discussion Select an answer first
5. 14foundation · easy

   A web application needs to return a model prediction to a user within 200 milliseconds of a request. Which inference approach is most appropriate?

   • ABatch inference
   • BScheduled inference
   • COffline inference
   • DOnline inference
   Reveal Solution Discussion Select an answer first
6. 15application · medium

   A team has a TensorFlow model that expects normalized features. They want to deploy it on Cloud Run, but the incoming requests contain raw, unnormalized data. What is the best way to handle this?

   • AUse Vertex AI Feature Store to normalize the data automatically.
   • BModify the client application to normalize the data before sending it to the model.
   • CTrain the model to accept raw data by adding a normalization layer.
   • DInclude normalization logic in the container's serving code that runs before inference.
   Reveal Solution Discussion Select an answer first