Databricks Machine Learning Professional Question 213
Single answerYou are building a machine learning model deployment pipeline that requires real-time predictions for a high-traffic e-commerce application. The solution must handle dynamic scaling, ensure high availability, and integrate seamlessly with other cloud services. Which of the following approaches is the most suitable for deploying this model in production?
- A
Deploy the model as a RESTful API service within a cloud-managed container orchestration platform like Kubernetes.
- B
Serve the model by embedding it directly into the e-commerce application's backend monolithic codebase.
- C
Deploy the model as a batch processing job that writes predictions to a database for later retrieval.
- D
Host the model on a local on-premises server and expose it as a RESTful API.
Show answer and explanation
Correct answer: A
Explanation
Deploying a machine learning model as a RESTful API service within a cloud-managed container orchestration platform (e.g., Kubernetes) is the best practice for production-grade real-time deployments. This approach allows for dynamic scaling, ensures high availability, and integrates easily with other cloud-native solutions. It decouples the model serving layer from the application logic, making it more flexible and maintainable.
- A. Correct.
Deploying the model as a RESTful API service within a cloud-managed container orchestration platform like Kubernetes is the best option for production-grade real-time deployments. It ensures scalability, fault tolerance, and seamless integration with other cloud services.
- B. Incorrect.
Embedding the model directly into the e-commerce application's backend monolithic codebase is not ideal for production-grade real-time deployments as it tightly couples the model with the application, making updates and scaling challenging.
- C. Incorrect.
Deploying the model as a batch processing job does not meet the requirement for real-time predictions, as batch jobs are designed for offline processing and delayed results.
- D. Incorrect.
Hosting the model on a local on-premises server may limit scalability, availability, and integration with cloud services, making it unsuitable for high-traffic real-time applications.