AI-300 Practice Question 21: Environment Asset Reproducibility
Microsoft Machine Learning Operations Engineer Associate (AI-300) · Test 3 · question 1 of 10
Data scientists on a team start every training script with a block of pip install commands. The same script run two weeks apart produced noticeably different evaluation metrics, and nobody can now determine which package versions the earlier run used. What should the team adopt so that a job's software dependencies are reproducible and auditable later?
Show answer and explanation
Answer:
An environment encapsulates the Python packages and software settings a training or scoring script runs with, and "the Machine Learning workspace manages and versions these environments, enabling reproducible, auditable, and portable machine learning workflows across various compute targets". Microsoft lists exactly these uses for an Environment object: reuse the same environment for training at scale, deploy the model with that same environment, and revisit the environment in which an existing model was trained (option D). The environment, compute target, and script together are the job's full specification.
- A is wrong because the dependency list stays inside the script, where it is neither a versioned asset nor attached to the job as something you can query later. It also leaves transitive dependencies unpinned, so the resolved package set can still shift between runs. - B is wrong because it records the symptom after the fact. You learn what the run happened to install, but there is no reusable artifact to rerun the job with or to deploy the model in. - C is wrong because the node image doesn't determine the Python packages; the pip install lines do, and they resolve afresh on every run. Standardizing the compute changes nothing about the dependency drift or the audit trail.
Reference: About Azure Machine Learning environments