MLA-C01 Question 344
Select 3You are a Machine Learning Engineer at a company that uses AWS for deploying ML models. Your team has implemented a CI/CD pipeline to streamline model training, validation, and deployment. Which of the following steps should be included in your CI/CD pipeline to ensure a robust ML workflow?
- A
Automate data preprocessing and feature engineering to ensure consistency across environments.
- B
Use a manual approval step to deploy the model directly to the production environment without testing.
- C
Integrate automated model testing to validate performance metrics before deployment.
- D
Enable versioning for data, models, and code to track changes and ensure reproducibility.
- E
Exclude retraining models in the pipeline to reduce complexity and runtime.
Show answer and explanation
Correct answers: A, C, D
Explanation
A CI/CD pipeline in ML workflows should automate key processes like data preprocessing, testing, and versioning to ensure consistent, robust, and reproducible deployments. Testing and validation are critical to ensure that models meet requirements, while versioning helps manage changes effectively. Manual steps or skipping important phases like retraining can lead to errors, inconsistencies, or degraded model performance in production.
- A. Correct.
Automating data preprocessing and feature engineering is a best practice in ML workflows to ensure consistent results across development, staging, and production environments.
- B. Incorrect.
Skipping testing by using a manual approval step for deployment is not a recommended practice. Robust CI/CD pipelines should validate models automatically to ensure reliability.
- C. Correct.
Automated model testing is a critical step to validate performance metrics and ensure that only models meeting business requirements are deployed.
- D. Correct.
Versioning for data, models, and code is essential in ML workflows to track changes, enable reproducibility, and facilitate debugging.
- E. Incorrect.
Excluding retraining models in the pipeline is not a best practice. Retraining is often necessary to account for data drift, and pipelines can be designed to handle this efficiently.