Google Professional Machine Learning Engineer Question 194
Single answerGoogle Cloud PlatformYou are a Machine Learning Engineer building a model training pipeline on Google Cloud. You need to track, compare, and analyze various hyperparameter tuning experiments to identify the best-performing model. Additionally, you want to ensure the solution is fully managed and requires minimal DevOps overhead. Which approach should you use?
- A
Use Vertex AI Experiments to track the experiments and analyze their performance.
- B
Use Kubeflow Pipelines along with a custom-built logging system to track experiment metadata.
- C
Deploy a self-hosted MLFlow server on Google Kubernetes Engine (GKE) to track experiment runs and results.
- D
Rely on Cloud Logging to store experiment details and manually compare results.
Show answer and explanation
Correct answer: A
Explanation
Vertex AI Experiments is the best solution because it is a fully managed, purpose-built service for tracking and analyzing machine learning experiments on Google Cloud. It satisfies the requirements for minimal DevOps overhead and provides seamless integration with other Google Cloud tools, making it an ideal choice for this scenario.
- A. Correct.
Vertex AI Experiments is a fully managed service that allows you to track, compare, and analyze machine learning experiments without requiring additional infrastructure setup. It integrates seamlessly with other Vertex AI tools, making it the most appropriate choice in this scenario.
- B. Incorrect.
Kubeflow Pipelines is a powerful tool for orchestrating ML workflows, but managing a custom-built logging system adds DevOps overhead, which contradicts the requirement for a fully managed solution.
- C. Incorrect.
While MLFlow is a popular tool for experiment tracking, deploying and managing a self-hosted MLFlow server on GKE requires significant DevOps effort, violating the requirement for minimal DevOps overhead.
- D. Incorrect.
Cloud Logging is designed for storing logs but is not purpose-built for tracking and comparing ML experiments. Manually comparing results is inefficient and error-prone.