Why A is the answer

In Git

Git gives training code and pipelines history, review and a single source of truth.

In detail

• Git gives history, review and a single source of truth.
• Compute should hold no unique work.

See it in different ways

mindmap
  root((Reproducible run))
    Code
      Git
    Data
      Versioned data assets
    Environment
      Registered environment
    Compute
      Cluster that scales to zero
    Results
      MLflow

flowchart LR
  S0["Write code"] --> S1["Commit"] --> S2["Review"] --> S3["Pipeline runs it"]

Why the other options are wrong

B. Only in a notebook on one compute instance

One compute instance is a single point of loss.

C. In the description of a registered model

A description is not code storage.

D. In the workspace, as notebooks that each person edits

Notebooks edited by each person have no reviewed history.

Related questions

• Question 9What is the main benefit of managing Azure Machine Learning workspaces with Bicep?
• Question 12Training needs GPUs for a few hours a week. Which compute fits?
• Question 13Which asset keeps a training job's dependencies consistent across runs?
• Question 16Datasets used for training must be reproducible. What should you register?

Learn more

• Our lesson: MLOps infrastructure
• MLflow in Azure Machine Learning