Google Professional Machine Learning Engineer Question 323
Select 3Google Cloud PlatformYou are designing an online model serving system on Google Cloud for a recommendation engine that handles highly variable traffic, including sudden spikes during promotional events. Which combination of strategies should you prioritize to ensure both scalability and cost-efficiency?
- A
Use autoscaling for the model serving infrastructure to handle variable traffic loads.
- B
Deploy the model on a single high-performance VM to maximize resource utilization.
- C
Enable request batching to improve throughput and reduce computational overhead.
- D
Leverage a managed service like Vertex AI Prediction to handle scalability and infrastructure management.
- E
Pre-provision a fixed number of nodes in a Kubernetes cluster to handle peak traffic.
Show answer and explanation
Correct answers: A, C, D
Explanation
To ensure scalability and cost-efficiency for online model serving with highly variable traffic, it is critical to use solutions that dynamically manage resources and improve efficiency. Autoscaling ensures the infrastructure adapts to traffic fluctuations. Request batching helps optimize performance during high traffic. Vertex AI Prediction provides a managed, scalable service for model serving, reducing operational complexity and improving reliability. Relying on fixed resources or single machines is inefficient and does not scale effectively for sudden spikes.
- A. Correct.
Autoscaling dynamically adjusts resources based on traffic demand, ensuring the system can handle both spikes and low traffic periods efficiently.
- B. Incorrect.
Using a single high-performance VM is not scalable for sudden traffic spikes and might lead to underutilization during low traffic periods.
- C. Correct.
Request batching groups multiple requests into a single inference batch, improving throughput and reducing computational overhead, especially during high-traffic periods.
- D. Correct.
Vertex AI Prediction is a managed service designed for scalable model serving, reducing operational overhead and ensuring reliable traffic handling.
- E. Incorrect.
Pre-provisioning fixed nodes does not adapt dynamically to changes in traffic, leading to either over-provisioning (wasting resources) or under-provisioning (failing during spikes).