Model Registry & Versioning Strategies
Managing the lifecycle. Aliasing, Tagging, and Rollback strategies using Vertex AI Model Registry.
Chaos Control
You have model.h5. Then model_v2.h5. Then model_final.h5.
Stop doing this.
1. Vertex AI Model Registry
The Registry is the Single Source of Truth.
- Version 1: Created Jan 1. Accuracy 80%.
- Version 2: Created Jan 15. Accuracy 85%.
Aliases
You don't deploy "Version 2". You deploy "Production".
- You assign the alias
defaultorprodtoVersion 2. - Your deployment script says:
deploy(model_name='churn_model', alias='prod'). - Benefit: If V2 breaks, you just move the
prodalias back to V1. The deployment script doesn't need to change hardcoded IDs.
2. Evaluation inside Registry
Before you alias a model as "Prod", you must Evaluate it. Vertex AI Model Registry allows you to attach Evaluation Metrics to a version. You can compare V1 vs V2 side-by-side in the UI (Confusion Matrix, ROC Curve).
Knowledge Check
?Knowledge Check
You have a CI/CD pipeline that automatically deploys the model tagged as 'Staging' to a QA environment. A data scientist trains a new model version. How should they trigger the deployment?