Serving Models PROFESSIONAL-MACHINE-LEARNING-ENGINEER Practice Questions (Page 8)

Questions 36–40

1. 36foundation · easy

   What is a primary benefit of using GKE for model serving compared to fully managed services?

   • AIt requires no infrastructure management
   • BIt automatically scales to zero
   • CIt is the only service that supports online inference
   • DIt provides greater control over the serving infrastructure
   Reveal Solution Discussion Select an answer first
2. 37application · medium

   An ML team needs to roll back to a previous model version quickly if the new version performs poorly in production. They have multiple versions of the same model. What is the best way to ensure this capability?

   • AKeep all model versions in the Vertex AI Model Registry and use traffic splitting to revert to the previous version.
   • BStore only the latest model version in the registry and keep backups in Cloud Storage.
   • CRedeploy the previous model from a backup file each time a rollback is needed.
   • DUse Cloud Run revisions to manage model versions.
   Reveal Solution Discussion Select an answer first
4. 38foundation · easy

   What is the primary purpose of Vertex AI Prediction?

   • ATo train machine learning models
   • BTo visualize model metrics
   • CTo serve model predictions
   • DTo store training data
   Reveal Solution Discussion Select an answer first
5. 39foundation · easy

   How does the Model Registry handle multiple versions of the same model?

   • AIt stores each version as a separate entry with a version number
   • BIt requires a separate registry for each version
   • CIt overwrites the previous version
   • DIt only keeps the latest version
   Reveal Solution Discussion Select an answer first
6. 40application · medium

   A company wants to roll out a new model version to production with minimal risk. They plan to start with 5% of traffic, monitor error rates and latency, and gradually increase to 100% if no issues are detected. What is the best way to implement this?

   • AUse Cloud Run revisions and split traffic between the old and new container images.
   • BDeploy the new version to a staging environment, run load tests, and then promote it to production.
   • CDeploy the new version to a separate endpoint and update the client application to send 5% of requests there.
   • DUse Vertex AI Prediction's traffic splitting feature to gradually shift traffic from the old to the new version.
   Reveal Solution Discussion Select an answer first