Domain 2 of 5 · Chapter 2 of 4

Model Registration and Versioning

Unlock the complete study guide + 1,040 practice questions across 16 full exams.

Bundled into the existing Operationalizing Machine Learning and Generative AI Solutions premium course — no separate purchase.

14-day money-back guarantee — no questions asked.

Included in this chapter:

  • The registered model: name and version
  • Register an MLflow model from a run
  • Evaluate with Responsible AI before you register
  • Package a feature-retrieval spec with the model
  • Manage the model lifecycle: archive and promote
  • Exam pattern recognition

Responsible AI dashboard: evaluation components

ComponentQuestion it answersHow it splits the dataReach for it when
Error analysisWhere does the model fail?Any input feature (error tree + heatmap)Finding high-error cohorts beyond one accuracy number
Model interpretabilityWhich features drove the prediction?Global and per-prediction (local)Explaining the model overall or one decision
Fairness assessmentDoes it harm a protected group?Sensitive features only (via Fairlearn)Comparing error and selection rates across cohorts
Data analysisIs the training data skewed?Feature distributions and cohortsSpotting over- or under-representation first

Decision tree

Explain an individual prediction? Model interpretability global + local importance Compare across sensitive groups? Fairness assessment Fairlearn · sensitive cohorts Find high-error cohorts on any feature? Error analysis error tree + heatmap Data analysis distribution / representation Yes No Yes No Yes No
Cheat sheet

Cheat sheet

  • Register a model directly from a run's output artifacts
  • MLflow model type enables no-code deployment
  • Registering under an existing name increments the version
  • Promote a registered model version across workspaces through a registry
  • Archiving hides a model version but retains it
  • There is no delete/unregister for a registered model version
  • feature_retrieval_spec.yaml is packaged at the model artifact root
  • Feature store serves an offline and an online store
  • Interpretability produces feature-importance for prediction transparency
  • The Responsible AI dashboard bundles several evaluation components
  • The Responsible AI error-analysis component locates high-error cohorts on any feature
  • Fairness evaluates error/selection rates across sensitive cohorts
  • ThresholdOptimizer mitigates unfairness without retraining

Unlock with Premium — includes all practice exams and the complete study guide.

References

  1. Register and work with models - Azure Machine Learning
  2. Deploy MLflow models to real-time endpoints - Azure Machine Learning
  3. Assess AI Systems and Make Data-Driven Decisions with Azure Machine Learning Responsible AI Dashboard - Azure Machine Learning
  4. Assess errors in machine learning models - Azure Machine Learning
  5. Model interpretability - Azure Machine Learning
  6. Machine learning fairness - Azure Machine Learning
  7. Tutorial 2: Experiment and train models by using features - Azure Machine Learning managed feature store - basics
  8. What is managed feature store? - Azure Machine Learning
  9. MLOps machine learning model management - Azure Machine Learning