Skip to main content

Model rollback (Coming soon)

warning

WIP: requires examples

Model rollback refers to the process of reverting to a previous version of a machine learning model when an issue arises due to changes in code, environment, or data. This is a critical capability for ensuring stability and traceability in production systems.

By tightly integrating GitHub versioning with Azure Machine Learning model management, AI Platform ensures that every model is traceable, reproducible, and easily restorable in the event of an issue.

When to perform a model rollback​

You may need to roll back a model if:

  • A bug is introduced in the code or training logic
  • Environment changes (e.g., library updates) cause inconsistencies
  • Training data changes lead to unexpected model behavior
  • A model deployed to production is underperforming or generating errors

Rollback workflow​

Identify the model version to roll back to.​

  • Use Azure Machine Learning workspace to view registered model versions, properties, and tags.
  • Look for a specific commit ID in the model’s tags to tie the Azure ML model version back to the source code version in GitHub.

Locate the corresponding GitHub commit.​

  • Navigate to the matching commit ID in the GitHub repository.
  • This commit should include:
    • The exact training code
    • The correct environment definition (e.g., conda or Docker config)
    • Any associated parameter settings or config files

Restore the model​

  • Move pointer of GitHub repository.

Best practices​

  • Always tag models with git-commit-id and model-name during training
  • Keep training scripts and environments versioned in GitHub
  • Use the Azure ML model registry to manage model versions, tags, and metadata
  • Maintain consistent naming conventions between GitHub and Azure ML assets