Google Professional Machine Learning Engineer
Get started today
Ultimate access to all questions.
You are a machine learning engineer at a company that has developed an ML model to detect the sentiment of users' posts on your company's social media page. The goal is to identify potential outages or bugs through sentiment analysis. For real-time predictions, you are using Dataflow to process data ingested from Pub/Sub. You plan to update the model regularly with new training iterations and keep the latest two versions live after every update. You need to split the incoming traffic between the two versions in an 80:20 ratio, with the newest model getting 80% of the traffic. Additionally, you want the solution to require minimal management and keep the pipeline as simple as possible. What should you do?
You are a machine learning engineer at a company that has developed an ML model to detect the sentiment of users' posts on your company's social media page. The goal is to identify potential outages or bugs through sentiment analysis. For real-time predictions, you are using Dataflow to process data ingested from Pub/Sub. You plan to update the model regularly with new training iterations and keep the latest two versions live after every update. You need to split the incoming traffic between the two versions in an 80:20 ratio, with the newest model getting 80% of the traffic. Additionally, you want the solution to require minimal management and keep the pipeline as simple as possible. What should you do?
Explanation:
The correct answer is A. The recommended approach to achieve the desired outcome is to deploy the ML models to a Vertex AI endpoint and configure the traffic splitting using the traffic-split parameter. Vertex AI natively supports traffic splitting between deployed models, and the traffic-split parameter allows you to specify the desired traffic distribution percentage between different versions. In this scenario, setting the traffic-split parameter to 0=80 will send 80% of the traffic to the newest model, while the PREVIOUS_MODEL_ID=20 configuration will direct the remaining 20% to the previous version. This method ensures a minimal management pipeline, meeting all the requirements specified in the question.