Model Versioning & Rollback

Tier 1 DEPLOY

What This Requires

Implement model versioning with semantic versioning (major.minor.patch), maintain version history, and test rollback procedures. Require ability to roll back to previous version within 15 minutes.

Why It Matters

Model updates can introduce regressions or performance degradation. Without versioning and rollback, teams are stuck with broken deployments. Quick rollback minimizes customer impact.

How To Implement

Versioning Scheme

Use semantic versioning: MAJOR (breaking changes), MINOR (new features), PATCH (bug fixes). Tag models in registry (AWS ECR, Azure ACR, MLflow).

Version Metadata

Store metadata with each version: training date, dataset version, performance metrics, deployment date, rollback history. Use Model Card or YAML manifest.

Rollback Procedure

Document rollback steps: update deployment config to previous version, restart services, validate health. Automate via script or CI/CD pipeline. Test rollback quarterly.

Immutable Artifacts

Store model artifacts immutably (no overwrite). Retain N previous versions (e.g., last 5). Archive older versions to cold storage.

Evidence & Audit

  • Model versioning policy with semantic versioning scheme
  • Model registry showing version history and metadata
  • Rollback procedure documentation with automation script
  • Rollback test records (quarterly) with success criteria
  • Artifact retention policy and archive logs

Related Controls