Skip to main content

Managed AML Infrastructure

Overview​

AI Platform's Traditional AI/ML journey runs on a managed Azure ML infrastructure. It enables fast and secure development and deployments of structured machine learning (ML) projects.

  • Each instance of a Traditional AI/ML project on AI Platform is an Azure ML deployment that typically consists of four resource groups (dev, test, prod, and shrd), each with an Azure ML Workspace.
  • Each workspace meets compliance and access requirements, ensuring the secure deployment of production-ready ML models across various environments.
  • Additionally, each instance has an ML registry located in the shrd resource group, allowing for easy sharing of assets, such as models or data, between the separate workspaces. Alternative configurations with an added QA layer or a basic exploratory option with just a dev layer is also readily available.
note

The managed Azure ML infrastructure is based on Team EurekaML's Archimedes Platform. Some components within the platform may still carry names or Ids based on the Archimedes Platform.

Diagram​

Click on image to zoom in.

High-level Traditional AI/ML architecture

Managed Azure ML architecture diagram

Description​

  • The main parts of the Traditional AI/ML architecture are the machine learning workspaces and the machine learning registry. There are three workspaces located in separate resource groups for development, testing and production, respectively.
  • The workspaces are connected via a machine learning registry, located in the shared resource group. The machine learning registry allows for sharing of assets between the three workspaces. This allows for an MLOps cycle where assets developed in the development workspace can be shared to the registry and tested in the test workspace, before being used in the production workspace.
  • For convenience, two datalakes are set up as datastores by default, namely the inbox and prediction datalakes. These are intended for data ingested to and output from the workspaces, respectively. The datalakes (and all other resources) found in the production resource group also exists in the development and test resource groups.
  • For resource monitoring, a log analytics workspace and application insights are available in all workspaces. A Grafana workspace may be deployed to the shared resource group if ordered by the user.
  • The platform infrastructure is configured using Azure Bicep, and deployed by GitHub Actions using a service principal for authentication. Users are advised to use a GitHub repository for interacting with their instance.
  • AKS clusters are typically used as compute targets.
tip

For help with files used within the infrastructure, go to the Template Repository page..