Google Professional Machine Learning Engineer Question 318
Single answerGoogle Cloud PlatformYou are deploying two versions of a machine learning model, Model A (current version) and Model B (new version), in Google Cloud. You want to conduct an A/B test to compare their performance using live traffic while ensuring a seamless user experience. What is the best approach to implement this in Google Cloud?
- A
Use Vertex AI Prediction to deploy both models and configure traffic splitting to route a percentage of the traffic to Model B.
- B
Deploy both models on Compute Engine and use a custom load balancer to split traffic between them.
- C
Use BigQuery ML to train both models and compare their performance using historical data instead of live traffic.
- D
Deploy both models in Dataflow and implement a custom pipeline to distribute traffic between the two.
Show answer and explanation
Correct answer: A
Explanation
A/B testing in Google Cloud is best implemented using Vertex AI Prediction. It allows you to deploy multiple versions of a model and seamlessly split traffic between them using traffic splitting configurations. This approach ensures minimal manual effort and leverages managed services for scalability and reliability. Other options, such as Compute Engine or Dataflow, would require unnecessary custom development and are not optimized for this use case. BigQuery ML, while useful for training and evaluating models, is not applicable for live traffic testing.
- A. Correct.
This is the correct option. Vertex AI Prediction allows you to deploy multiple versions of a model and configure traffic splitting, which is ideal for A/B testing with live traffic.
- B. Incorrect.
This option is incorrect because while you can deploy on Compute Engine and use a custom load balancer, it requires significant manual effort and does not leverage Google Cloud's managed services for A/B testing.
- C. Incorrect.
This option is incorrect because BigQuery ML is used for training and evaluating models on historical data, not for live traffic A/B testing.
- D. Incorrect.
This option is incorrect because Dataflow is primarily used for data processing pipelines, not for deploying models or splitting traffic for live testing.