3.3.3. Model Management and Deployment

After training, models must be managed and deployed for use. Think of model management like software version control—you need to track what you've built, deploy it reliably, and monitor it in production.

Model management:

• Model registration: Store trained models in a central registry with metadata (accuracy, training date, dataset used)
• Model versioning: Track different versions as you improve models over time
• Model lineage: Record which data and parameters created each model
• Model comparison: Compare performance across versions to choose the best one

Why registration matters: Without a registry, you end up with models scattered across laptops and storage accounts. When something breaks in production, you can't answer "Which model version is deployed?" or "What data was it trained on?"

Model deployment options:

Model deployment concepts:

• Endpoints: URLs where applications send data for predictions (like API endpoints)
• Containers: Models packaged with dependencies for consistent deployment
• Scaling: Automatically add compute resources when demand increases
• Blue/green deployment: Run two versions simultaneously to safely switch

Model monitoring: Once deployed, models can degrade over time as real-world data changes:

• Data drift: Input data patterns shift from training data
• Concept drift: The relationship between inputs and outputs changes
• Performance monitoring: Track accuracy, latency, and error rates
• Alerts: Notify when metrics exceed thresholds

Example workflow:

1. Train model in Azure ML workspace
2. Register model with version tag (v1.0)
3. Deploy to real-time endpoint
4. Monitor performance metrics
5. Train improved model, register as v1.1
6. Compare v1.0 vs v1.1 performance
7. Deploy v1.1 using blue/green deployment
8. Retire v1.0 after successful transition

⚠️ Exam Tip: Questions about "tracking model versions" or "storing models" point to model registration. Questions about "making predictions available to applications" point to deployment endpoints.