Authentication and authorization for online endpoints

APPLIES TO: Azure CLI ml extension v2 (current) Python SDK azure-ai-ml v2 (current)

This article explains the concepts of identity and permission in the context of Azure Machine Learning online endpoints. The article discusses Microsoft Entra IDs that support role-based access control and permissions. A Microsoft Entra ID is called either a user identity or an endpoint identity, depending on its purpose.

• A user identity is a Microsoft Entra ID that can create an endpoint and its deployments, or interact with endpoints or workspaces. A user identity issues requests to endpoints, deployments, or workspaces. The user identity needs proper permissions to perform control plane and data plane operations on the endpoints or workspaces.

• An endpoint identity is a Microsoft Entra ID that runs the user container in deployments. The user container uses the endpoint identity for the deployment. The endpoint identity also needs proper permissions for the user container to interact with resources as needed. For example, the endpoint identity needs the proper permissions to pull images from Azure Container Registry or to interact with other Azure services.

The user identity and endpoint identity have separate permission requirements. For more information on managing identities and permissions, see How to authenticate clients for online endpoints.

Permissions and scope for authorization

Azure role-based access control (RBAC) allows you to define and assign roles with a set of allowed and/or denied actions on specific scopes. You can customize these roles and scopes according to your business needs. The following examples serve as a starting point that you can extend as necessary.

For user identity:

• To manage control plane and data plane operations, you can use the built-in role AzureML Data Scientist that includes the permission action Microsoft.MachineLearningServices/workspaces/onlineEndpoints/*/actions.
• To control the operations for a specific endpoint, use the scope /subscriptions/<subscriptionId>/resourcegroups/<resourceGroupName>/providers/Microsoft.MachineLearningServices/workspaces/<workspaceName>/onlineEndpoints/<endpointName>.
• To control the operations for all endpoints in a workspace, use the scope /subscriptions/<subscriptionId>/resourcegroups/<resourceGroupName>/providers/Microsoft.MachineLearningServices/workspaces/<workspaceName>.

For endpoint identity, to allow the user container to read blobs, the built-in role Storage Blob Data Reader includes the permission data action Microsoft.Storage/storageAccounts/blobServices/containers/blobs/read.

For more information on managing authorization to Azure Machine Learning workspaces, see Manage access to Azure Machine Learning. For more information on role definition, scope, and role assignment, see Azure RBAC. To understand the scope for assigned roles, see Understand scope for Azure RBAC.

Permissions needed for user identity

When you sign in to your Azure tenant with your Microsoft account, for example by using az login, you complete the authn user authentication step that determines your identity as a user. To create an online endpoint under an Azure Machine Learning workspace, your identity needs the proper permission, also called authorization or authz. User identities need appropriate permissions to perform both control plane and data plane operations.

Control plane operations

Control plane operations control and change the online endpoints. These operations include create, read, update, and delete (CRUD) operations on online endpoints and online deployments. For online endpoints and deployments, requests to perform control plane operations go to the Azure Machine Learning workspace.

Authentication for control plane operations

For control plane operations, you use a Microsoft Entra token to authenticate a client to the workspace. Depending on your use case, you can choose from several authentication workflows to get this token. The user identity also needs to have the proper Azure RBAC role assigned to access resources.

Authorization for control plane operations

For control plane operations, your user identity needs to have the proper Azure RBAC role assigned to access your resources. Specifically, for CRUD operations on online endpoints and deployments, the user identity needs to have roles assigned for the following actions:

* You can fetch your Microsoft Entra aad_token directly from Microsoft Entra ID once you sign in, so that operation doesn't require Azure RBAC permission on the workspace.

Permissions for user identity to enforce access to default secret stores

If you use the secret injection feature, and you set the flag to enforce access to the default secret stores while creating endpoints, your user identity must have permission to read secrets from workspace connections.

An endpoint identity can be either a system-assigned identity (SAI) or a user-assigned identity (UAI). When the endpoint is created with a SAI and the flag to enforce access to the default secret stores is set, a user identity must have permissions to read secrets from workspace connections when creating an endpoint and deployments. This restriction ensures that only a user identity with the permission to read secrets can grant the endpoint identity the permission to read secrets.

If a user identity that doesn't have permission to read secrets from workspace connections tries to create an endpoint or a deployment with a SAI, and the endpoint's flag is set to enforce access to the default secret stores, the endpoint, or deployment creation is rejected.

If the endpoint is created with a UAI, or the endpoint uses a SAI but the flag to enforce access to the default secret stores isn't set, the user identity doesn't need to be able to read secrets from workspace connections to create an endpoint or deployment. In this case, the endpoint identity isn't automatically granted the permission to read secrets, but can be manually granted this permission by assigning the proper role.

Regardless of whether the role assignment is automatic or manual, the secret retrieval, and injection is triggered if you mapped the environment variables with secret references in the endpoint or deployment definition. The secret injection feature uses the endpoint identity to do the secret retrieval and injection. For more information on secret injection, see Secret injection in online endpoints.

Data plane operations

Data plane operations don't change the online endpoints, but use data that interacts with the endpoints. An example of a data plane operation is to send a scoring request to an online endpoint and get a response from it. For online endpoints and deployments, requests to perform data plane operations go to the endpoint's scoring URI.

Authentication for data plane operations

For data plane operations, you can choose from the following ways to authenticate a client to send requests to an endpoint's scoring URI:

• Key
• Azure Machine Learning aml_token
• Microsoft Entra aad_token

For more information on how to authenticate clients for data plane operations, see How to authenticate clients for online endpoints.

Authorization for data plane operations

For data plane operations, your user identity needs proper Azure RBAC roles to allow access to your resources only if the endpoint is set to use Microsoft Entra aad_token. For data plane operations on online endpoints and deployments, the user identity needs to have a role assigned with the following actions:

Permissions needed for endpoint identity

An online deployment runs your user container with the endpoint identity, that is, the managed identity associated with the endpoint. The endpoint identity is a Microsoft Entra ID that supports Azure RBAC. Therefore, you can assign Azure roles to the endpoint identity to control permissions that are required to perform operations. This endpoint identity can be either a SAI or a UAI. You can decide whether to use an SAI or a UAI when you create the endpoint.

• For a SAI, the identity is created automatically when you create the endpoint, and roles with fundamental permissions, such as the Container Registry pull permission AcrPull and the Storage Blob Data Reader, are automatically assigned.
• For a UAI, you need to create the identity first, and then associate it with the endpoint when you create the endpoint. You're also responsible for assigning proper roles to the UAI as needed.

Automatic role assignment for endpoint identity

If the endpoint identity is a SAI, the following roles are assigned to the endpoint identity for convenience.

• If the endpoint identity is a SAI, and the enforce flag isn't set or the user identity doesn't have permission to read secrets, there's no automatic role assignment for the Azure Machine Learning Workspace Connection Secrets Reader role. For more information, see How to deploy online endpoint with secret injection.
• If the endpoint identity is a UAI, there's no automatic role assignment for the Azure Machine Learning Workspace Connection Secrets Reader role. In this case, you need to manually assign roles to the endpoint identity as needed.

For more information on the Azure Machine Learning Workspace Connection Secrets Reader role, see Assign permissions to the identity.

Related content

• Set up authentication
• How to authenticate to an online endpoint
• How to deploy an online endpoint