13.1 Organizing & Versioning Models in Model Registry

The exam guide lists organizing and versioning models in Gemini Enterprise Agent Platform Model Registry. Model Registry answers operational questions: which models exist, which version is in production, how each version performed, and how to roll back.

What Model Registry Manages

• Custom-trained models (prebuilt or custom serving containers)
• AutoML models (tabular and image)
• BigQuery ML models, registered without export
• Imported models trained elsewhere, when artifacts meet prebuilt-container requirements or come with a custom container

From a version's detail page you can evaluate, deploy to an endpoint, run batch inference, and view performance metrics. Knowledge Catalog makes models searchable across projects and regions.

Models, Versions, Aliases, and Labels

Alias rules

• The first version of a new model automatically gets the default alias.
• Exactly one version carries default at all times. When a command doesn't name a version, the default version is used.
• An alias points to one version per model. Assigning an existing alias to another version moves it there.
• Aliases must match the format [a-z][a-z0-9-]{0,126}[a-z0-9] and can't be purely numeric, so they're never confused with version IDs.
• Aliases aren't labels. Aliases select a version by role. Labels describe and group.

Creating Versions

model_v8 = aiplatform.Model.upload(
    display_name="fraud-detector",
    parent_model="projects/p/locations/us-central1/models/1234",
    artifact_uri="gs://ml-artifacts/fraud/2026-09-17/",
    serving_container_image_uri=XGB_PREBUILT_IMAGE,
    version_aliases=["challenger"],
    is_default_version=False,
    labels={"data_version": "v13", "team": "risk"},
)

Versions and Evaluations

Attach model evaluations to each version, either from the registry or through the evaluation pipeline component (Chapter 7). You can then compare versions side by side before promotion. Combined with ML Metadata lineage (Chapter 6), each version links back to its dataset, code, and training run.

A Promotion Workflow

1. The training pipeline uploads a new version with the alias challenger. It is not default.
2. The evaluation component compares challenger with champion on the same test set.
3. If it meets the gates, a canary rollout sends a small share of endpoint traffic to the challenger (Section 13.2).
4. After online checks pass, move the champion alias (and default) to the new version.
5. Rollback means moving the alias back and shifting traffic to the previous deployed version, with no retraining.

Because deployments and batch jobs can reference versions by alias, moving an alias is a clean, auditable promotion step.

Copying Models Across Regions and Projects

Copy model duplicates a registered model into another location in the same project or another project. Use it to deploy the exact same trained model in several regions, or to promote from a dev project to a prod project. Retraining isn't a substitute, because training is non-deterministic. If you don't name a version, the default version is copied. BigQuery ML models can't be copied this way.

Organizing at Scale

• One model resource per use case (such as churn-mobile), with versions over time. Don't create a new model resource for every retrain.
• Labels for ownership and context: team, cost center, data version, regulatory scope.
• Separate projects for environments (dev, staging, prod), with copy model or CI/CD promotion between them.
• IAM to control who can upload versions or move production aliases.
• Retire old versions that are neither deployed nor needed for audits, keeping records required by policy.

Governance Questions the Registry Should Answer

Exam Traps

• Deploying by version ID in scripts that should follow the "current production" model. Use aliases so promotion doesn't require code changes.
• Using labels to mark the production model. Labels don't enforce one-version-per-role, but aliases do.
• Retraining in each region to "replicate" a model instead of copying it.
• Forgetting that a registered BigQuery ML model can be deployed to an endpoint directly from the registry without export.