Govern models with MLflow and Amazon SageMaker AI Model Registry sync: Part 2
aws ·
This article extends previous guidance on governing machine learning models by introducing two cross-account topologies for organizations that separate development and production environments. The first topology, a hub-and-spoke pattern, centralizes governance by sharing a single MLflow application across multiple development accounts using AWS Resource Access Manager. In this setup, data scientists register models in their local accounts, which automatically sync to a central registry where governance officers approve them before deployment. The second topology is a hybrid approach designed for regulated environments that require strict isolation between development and production accounts. This method allows developers to register and locally approve models within their own accounts, after which an automated workflow copies the approved artifacts to the central hub for final validation. Both patterns utilize continuous integration and delivery pipelines to deploy approved models to SageMaker endpoints. The post details the specific roles of administrators, data scientists, model owners, and governance officers in executing these workflows. It concludes by providing a comparison of the two topologies to help organizations select the architecture that best fits their compliance and operational needs.
aws