Govern models with MLflow and Amazon SageMaker AI Model Registry sync: Part 2
This article extends the single‑account MLflow‑SageMaker AI Model Registry integration from Part 1 to two cross‑account governance topologies—a hub‑and‑spoke pattern that shares a single Managed MLflow app via AWS Resource Access Manager, and a hybrid pattern that isolates development accounts while still allowing model promotion to a central hub. It details the required IAM condition keys, AWS RAM shares, bucket policies, and CloudFormation stack deployment for both a development (spoke) and governance (hub) account, and shows how an approved model moves from the registry to a SageMaker endpoint through CI/CD. Engineers can now enforce strict separation of development and production workloads while maintaining a unified model registry across accounts.
⚡ Key Takeaways
- Hub‑and‑spoke topology uses AWS RAM to share one Managed MLflow app across multiple development accounts.
- Hybrid topology keeps each development account isolated; model owners approve models locally before promotion to the hub.
- Cross‑account setup requires a one‑time administrator configuration of AWS RAM shares, bucket policies, and destination groups via console or CLI.
- Model promotion to a deployed endpoint is demonstrated through a CI/CD pipeline that pulls the approved model from the SageMaker AI Model Registry.
- Prerequisites include two AWS accounts, named AWS CLI profiles (e.g., mlops‑spoke, mlops‑hub), and deployment of the provided CloudFormation template (cfn/sagemaker-studio-mlflow.yaml) in each account.
By automating cross‑account model registration and promotion, teams can enforce governance policies at scale without sacrificing the agility of independent development environments, reducing risk of accidental production writes.
✅ Practical Steps
- Deploy the CloudFormation stack (cfn/sagemaker-studio-mlflow.yaml) in both the development and governance accounts to provision SageMaker Studio, user profiles, execution roles, and a Managed MLflow app with AutoModelRegistrationEnabled.
- Configure AWS RAM shares, bucket policies, and destination groups using the AWS CLI or console as described in the repository’s “Account and profile setup” section.
- Run the CI/CD workflow (referenced in the GitHub repo) to promote an approved model from the SageMaker AI Model Registry to a SageMaker endpoint.
Want the full story? Read the original article.
Read on AWS ML Blog ↗