1.14 Registering a Model in the Unity Catalog Registry

1.14 Registering a Model in the Unity Catalog Registry

The transition from experimental machine learning to production requires rigorous model lifecycle governance. Organizations must maintain absolute traceability over which dataset version and code commit trained a model, who reviewed and approved it, and which specific model version is currently serving live production traffic.

Models in Unity Catalog integrate the MLflow Model Registry directly into Databricks' centralized governance framework. This architecture unifies model management under the standard 3-level namespace (catalog.schema.model_name), deprecating the legacy, workspace-confined Model Registry.

+---------------------------------------------------------------------------------------------------+
|                             MODELS IN UNITY CATALOG ARCHITECTURE                                 |
+---------------------------------------------------------------------------------------------------+
| `prod_catalog` (Catalog: Environment Isolation)                                                   |
|    |                                                                                              |
|    +-- `fraud_detection` (Schema: Domain Grouping)                                                |
|          |                                                                                        |
|          +-- `payment_risk_classifier` (Registered Model)                                         |
|                |                                                                                  |
|                +-- Version 1 (Trained 2026-01-10) -> Tag: `status: deprecated`                    |
|                +-- Version 2 (Trained 2026-02-15) -> Alias: `@champion` (Serving Live Traffic)   |
|                +-- Version 3 (Trained 2026-03-01) -> Alias: `@challenger` (Shadow Evaluation)     |
+---------------------------------------------------------------------------------------------------+

---

Legacy Workspace Model Registry vs. Models in Unity Catalog

The Databricks ML certification heavily emphasizes the modernization from legacy workspace-scoped registries to Unity Catalog:

---

Registering Models in Unity Catalog

Models can be registered in Unity Catalog using two primary programmatic methods:

Method 1: Inline Registration During Training (log_model)

import mlflow
import mlflow.sklearn
from sklearn.ensemble import RandomForestClassifier

# Configure active registry destination to Unity Catalog
mlflow.set_registry_uri("databricks-uc")

full_model_name = "prod_catalog.fraud_detection.payment_risk_classifier"

with mlflow.start_run(run_name="rf_production_candidate") as run:
    model = RandomForestClassifier(n_estimators=150, max_depth=8)
    model.fit(X_train, y_train)
    
    # Log and register model in a single atomic operation
    mlflow.sklearn.log_model(
        sk_model=model,
        artifact_path="model",
        registered_model_name=full_model_name,
        input_example=X_train.iloc[:5]
    )

Method 2: Registering an Existing Run Artifact (register_model)

from mlflow.tracking import MlflowClient

client = MlflowClient()
run_id = "3a8f1b2c4d5e6f7a8b9c0d1e2f3a4b5c"
model_uri = f"runs:/{run_id}/model"

# Register existing run artifact to Unity Catalog 3-level model name
model_version = mlflow.register_model(
    model_uri=model_uri,
    name="prod_catalog.fraud_detection.payment_risk_classifier"
)
print(f"Successfully created Model Version: {model_version.version}")

---

Prerequisites the Exam Tests

Registration into Unity Catalog fails for predictable reasons, and each one is a plausible distractor:

Privileges: Registering vs. Consuming

Two different roles need two different grants, and mixing them up is a common exam distractor:

-- The training job needs to create versions
GRANT USE CATALOG ON CATALOG prod_catalog TO `sp-training-job`;
GRANT USE SCHEMA ON SCHEMA prod_catalog.fraud_detection TO `sp-training-job`;
GRANT CREATE MODEL ON SCHEMA prod_catalog.fraud_detection TO `sp-training-job`;

-- The batch scoring job only needs to load and score
GRANT EXECUTE ON MODEL prod_catalog.fraud_detection.payment_risk_classifier
TO `sp-batch-inference-runner`;

EXECUTE is the privilege that permits mlflow.pyfunc.load_model, mlflow.pyfunc.spark_udf, and Model Serving to use the model. It does not permit creating new versions or reassigning aliases.

---

Access Control and Permissions in Unity Catalog

Unity Catalog enforces role-based access control (RBAC) on registered models via standard SQL:

-- Grant scoring permission to production service principal
GRANT EXECUTE ON MODEL prod_catalog.fraud_detection.payment_risk_classifier 
TO `sp-batch-inference-runner`;

-- Grant read-only exploration access to junior analysts
GRANT READ_METADATA ON MODEL prod_catalog.fraud_detection.payment_risk_classifier 
TO `junior-analysts`;