MLOps v3 documentation

This site explains how to deploy and manage Azure Machine Learning MLOps with the MLOps v3 Accelerator, a PowerShell module that turns one inputs file into a ready-to-run repository with bootstrap, infrastructure (Terraform or Bicep on Azure Verified Modules), CI/CD (GitHub Actions or Azure DevOps) and a training workload.

Use the navigation on the left to explore the documentation. For an introduction to the concepts this accelerator builds on, see the MLOps v2 guidance on Learn.

The MLOps journey

The journey has several steps. It starts with planning and bootstrapping your environment and ends with workloads that train and serve models.

Plan your deployment

Decide the scenario, network mode, environments, data source and tooling before you run anything. Each decision maps to a key in inputs.yaml.

See the planning section and the prerequisites for more information.

Bootstrap your environment

The bootstrap is the first step of the journey. It sets up the foundational components used to deploy and manage the platform: the repository or project, a managed identity with federated credentials, environments, and a remote state store. This step needs the correct access and a subscription to work with.

See the bootstrap section for more information.

The MLOps platform

The platform is the core of the journey. It provides the governance, networking and security components that workloads build on.

Included in the platform are:

• Azure Machine Learning workspace - one per environment (dev, test, prod), with storage, Key Vault, container registry and monitoring.
• Policy - guardrail policy assignments that audit or enforce secure settings. See policy.
• Role-based access control (RBAC) - least-privilege roles for the pipeline identity and the workspace.
• Networking - public, managed virtual network or bring-your-own virtual network, depending on the scenario.
• Budgets and alerts - cost guardrails per environment.

Workloads

Workloads are the models and pipelines that run on the platform. Each workload gets its own folder, Azure ML assets and train and deploy pipelines. Several workloads can share one platform through the workloads input.

See run and the customization options for more information.

Questions or problems? See the FAQ, troubleshooting, advanced options and known issues.

To get started, look at the quick start or the bootstrap your environment process.

What gets created

Each environment then deploys an Azure ML workspace with storage, Key Vault, container registry and monitoring. Private scenarios add networking.