Google Professional Machine Learning Engineer Question 325
Single answerGoogle Cloud PlatformYou are building a real-time fraud detection system using Google Cloud services. The system uses a machine learning model that requires features such as user transaction history, geolocation, and device details. You need to ensure these features are stored, updated, and served with low latency for both training and online prediction. Which approach would best meet these requirements?
- A
Store the features in BigQuery and query them directly during model inference.
- B
Use Vertex AI Feature Store to store and serve the features for both training and online prediction.
- C
Store the features in Cloud SQL and retrieve them using SQL queries during predictions.
- D
Manually manage feature storage in a Cloud Storage bucket and access it during training and prediction.
Show answer and explanation
Correct answer: B
Explanation
Vertex AI Feature Store is the ideal solution for storing and serving machine learning features in both training and prediction workflows. It is designed to handle low-latency lookups for online prediction, as well as feature versioning and reuse for training. This makes it the best choice for real-time fraud detection systems where accurate and timely predictions are critical.
- A. Incorrect.
BigQuery is optimized for analytical queries and is not ideal for low-latency retrieval during online prediction. It may result in high latency for serving features in real-time scenarios.
- B. Correct.
Vertex AI Feature Store is specifically designed to store, manage, and serve machine learning features with low latency. It supports both online and offline use cases, making it the best fit for real-time fraud detection.
- C. Incorrect.
Cloud SQL is a relational database service and is not optimized for managing machine learning features or ensuring low-latency access during real-time predictions.
- D. Incorrect.
Manually managing features in Cloud Storage is error-prone and does not provide the automated scaling, low-latency serving, or feature versioning capabilities required for a production-grade machine learning system.