SageMaker Model Registry

Mental model

• Governance layer: versioned models + metadata + approval gates.
• Bridges training pipelines and deployment workflows.

Where it fits

• ML CI/CD: evaluation → register → approve → deploy (canary/blue-green).
• Audit: “what model is in prod and why?”

Knobs that matter

• Model package groups (one per model family)
• Approval status (manual gates)
• Metadata: metrics, datasets, git commit, image digest, eval reports
• Cross-account promotion: dev → staging → prod (common)

Pricing mental model

• Registry is usually not the cost driver; the underlying artifacts (S3) and jobs are.

Terraform template (model package group)

resource "aws_sagemaker_model_package_group" "group" {
  model_package_group_name = var.group_name
  model_package_group_description = "Registry for ${var.group_name}"
}

variable "group_name" { type = string }

---