What's changed: Added per-section figures (cert-figure-retrofit). New AI-300 Chapter 2 (Domain 2 "ML model lifecycle and operations": training orchestration = MLflow/AutoML/sweep-early termination/distributed training/training pipelines/model comparison; registration and versioning = MLflow models/feature retrieval spec/responsible AI evaluation/archiving; production deployment and monitoring = online-batch endpoints/progressive rollout-traffic split/safe rollback/data drift/retraining-alert triggers)
2.2Model registration, versioning, and responsible AI evaluation
Understand registering an MLflow model, packaging a feature retrieval specification with the model artifact, evaluating a model using responsible AI principles, and managing the model lifecycle including archiving.
Before production use, register and version trained models in the model registry. This tracks which model came from which experiment and clarifies the target for deployment and rollback.
2.2.1Model registration and feature retrieval spec
Registering as an MLflow model stores inference metadata (signature, dependencies) alongside it, easing deployment. Packaging a feature retrieval specification (defining where to fetch features at inference) with the model artifact prevents training/serving skew (mismatch between training-time and inference-time features). Each registration gets a version for traceability.
2.2.2Responsible AI evaluation and lifecycle
Before deployment, evaluate the model with responsible AI principles (fairness, reliability & safety, explainability). The AML Responsible AI dashboard visualizes error analysis, fairness, and interpretability (feature importance). Operationally, archive old models you no longer need and manage the lifecycle: register → evaluate → deploy → monitor → (retrain/archive).
| Goal | Means | Key point |
|---|---|---|
| Register for easy deploy | MLflow model | Signature/dependencies included |
| Prevent feature skew | Package feature retrieval spec | Training/serving consistency |
| Evaluate fairness/interpretability | Responsible AI dashboard | Error/fairness/feature importance |
| Traceability | Versioning | Trace experiment→model |
| Tidy unused models | Archiving | Lifecycle management |
Watch the mix-ups: (1) model registration (registry) vs deployment to an endpoint are separate steps. (2) packaging the feature retrieval spec counters training/serving skew. (3) Responsible AI evaluation is a quality gate before deployment.
Map requirement → means: "register with inference metadata for easy deploy" = MLflow model; "avoid feature mismatch between training and inference" = package feature retrieval spec; "evaluate fairness and error" = Responsible AI dashboard; "trace which experiment a model came from" = versioning; "tidy old models" = archiving.
2.2.3Section summary
- Register as MLflow models (with metadata) + version for traceability
- Package the feature retrieval spec to prevent training/serving skew
- Evaluate with responsible AI before deployment; archive unused models
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. Which model registration format stores inference signature and dependencies together to ease deployment?
Q2. Which defines where to fetch features at inference and is packaged with the model to prevent training/serving skew?
Q3. Which AML feature evaluates fairness, error, and feature importance before deployment?
Q4. What is attached to model registrations to make it traceable which experiment a model came from?
Q5. Which operation tidies old, unneeded models and manages the lifecycle?
Q6. Which correctly describes registration vs deployment of a model?