Model Registration & Versioning

From experiment to production artifact

Registering an MLflow model

After a successful training run, register the model in the Azure ML model registry:

from azure.ai.ml.entities import Model
from azure.ai.ml.constants import AssetTypes

# Register from an MLflow run
model = Model(
    path="azureml://jobs/{run_id}/outputs/artifacts/churn-model",
    name="churn-predictor",
    version="3",
    type=AssetTypes.MLFLOW_MODEL,
    description="Churn prediction model — LightGBM, F1=0.961",
    tags={
        "algorithm": "lightgbm",
        "dataset_version": "v2",
        "f1_score": "0.961"
    }
)
registered_model = ml_client.models.create_or_update(model)

What’s happening:

• Line 6: Points to the model artifact from a specific training run — full lineage
• Line 9: MLFLOW_MODEL type — includes the model file, conda environment, and MLmodel metadata
• Lines 11-14: Tags make the model searchable and auditable

Or register directly from MLflow:

import mlflow

# During training, log and register in one step
with mlflow.start_run():
    # ... train model ...
    mlflow.sklearn.log_model(model, "model",
        registered_model_name="churn-predictor")

💡 Exam tip: MLflow model format

The MLflow model format is a directory containing:

• MLmodel — metadata file listing flavors (sklearn, pytorch, etc.)
• model.pkl (or equivalent) — the serialised model
• conda.yaml — environment dependencies
• requirements.txt — pip dependencies

This standard format means the model can be deployed to any MLflow-compatible platform. The exam favours MLflow model format over custom serialisation.

Feature retrieval specification

A feature retrieval spec bundles feature definitions with the model, ensuring that inference uses the correct features in the correct format:

from azure.ai.ml.entities import Model

model = Model(
    path="./model_output",
    name="churn-predictor",
    version="4",
    type=AssetTypes.MLFLOW_MODEL,
    properties={
        "feature_retrieval_spec": "feature_retrieval_spec.yaml"
    }
)

The spec file defines which features the model expects and where to retrieve them:

# feature_retrieval_spec.yaml
feature_store_uri: azureml://featurestores/churn-features
features:
  - name: customer_tenure_months
    source: customer_features:v2
  - name: monthly_charges
    source: billing_features:v1
  - name: support_tickets_30d
    source: support_features:v3

Scenario: Kai packages the churn model

Kai’s churn model needs 12 features to make predictions. Without a feature retrieval spec, the deployment team has to guess which features to send. With the spec bundled into the model:

1. Training time: model is trained on features from the feature store
2. Registration: feature spec is packaged with the model
3. Deployment time: the endpoint reads the spec and automatically retrieves the right features
4. Result: zero ambiguity about what the model needs

Priya (CTO): “No more deployment bugs because someone sent the wrong features.”

Model lifecycle management

Models go through stages as they mature:

Stage	Meaning	Action
None (default)	Just registered, not validated	Run evaluation before promoting
Staging	Being tested, not serving production traffic	Deploy to staging endpoint, run tests
Production	Actively serving predictions	Deploy to production endpoint
Archived	Retired, kept for audit/compliance	Remove from endpoints, retain in registry

# Update model stage
from azure.ai.ml.entities import Model

model = ml_client.models.get(name="churn-predictor", version="3")

# Archive an old version
ml_client.models.archive(name="churn-predictor", version="1")

# Restore if needed (for audit or rollback)
ml_client.models.restore(name="churn-predictor", version="1")

Scenario: Dr. Fatima's model governance at Meridian

Meridian Financial has strict model governance requirements. Dr. Fatima implements a lifecycle policy:

1. Registration — every model must include: algorithm, dataset version, metrics, and feature spec
2. Staging gate — model must pass responsible AI evaluation (Module 11) before staging
3. Production gate — model must pass A/B test against current production model
4. Archiving — retired models kept for 7 years (regulatory requirement)
5. Version limit — max 10 active versions per model; older ones auto-archived

James Chen (CISO): “Every model decision is auditable. That’s what regulators want to see.”

Key terms flashcards

Question

What is the MLflow model format?

Click or press Enter to reveal answer

Answer

A standardised directory containing: MLmodel (metadata), the serialised model file, conda.yaml (environment), and requirements.txt (pip deps). Deployable to any MLflow-compatible platform.

Click to flip back

Question

What is a feature retrieval specification?

Click or press Enter to reveal answer

Answer

A YAML file bundled with a model that defines which features the model needs and where to retrieve them. Ensures inference uses the correct features in the correct format.

Click to flip back

Question

What are the model lifecycle stages?

Click or press Enter to reveal answer

Answer

None (just registered) → Staging (testing) → Production (serving) → Archived (retired, kept for audit). Models can be restored from archive for rollback.

Click to flip back

Knowledge check

Knowledge Check

Kai's deployment team keeps sending the wrong features to the churn model endpoint, causing prediction errors. What should Kai bundle with the registered model?

AA feature retrieval specification that defines which features to retrieve and from whereBA README file listing the expected featuresCA copy of the training datasetDA test script that validates the input format

Check Answer

Knowledge Check

Dr. Fatima needs to retire an old model version but must keep it available for regulatory audits for 7 years. What should she do?

ADelete the model from the registry and keep a backup in blob storageBArchive the model version in the registryCMove the model to a separate compliance workspaceDExport the model to a Docker image and store it

Check Answer

🎬 Video coming soon

---

Next up: Model Approval & Responsible AI Gates — ensuring models are fair, explainable, and safe before deployment.