Workspaces, Datastores, Compute and Access

What it is

An Azure Machine Learning workspace is the top-level resource for machine learning work. It holds jobs, models, endpoints and assets, and it is also an MLflow tracking server. A workspace depends on a storage account, a key vault and an Application Insights instance. A container registry is optional and is created when it is first needed to build environment images.

Datastores are named references to storage, so jobs never contain raw connection strings. Compute targets run the work. Identities and roles decide who and what may touch all of it.

How it works

Create a workspace

# workspace.yml
$schema: https://azuremlschemas.azureedge.net/latest/workspace.schema.json
name: mlw-prod
location: eastus
display_name: Production ML workspace
description: Training and endpoints for the churn model

az extension add --name ml
az ml workspace create --file workspace.yml --resource-group rg-ml

In Python SDK v2, connect with MLClient(DefaultAzureCredential(), subscription_id, resource_group, workspace_name) and create a workspace with ml_client.workspaces.begin_create(Workspace(name="mlw-prod", location="eastus")).

Datastores

Every workspace gets default datastores, including workspaceblobstore (blob) and workspacefilestore (files). A datastore authenticates either with credentials (account key, SAS token, service principal) or with identity, where you save no secret and the user's or workspace's identity must hold a data role on the storage.

$schema: https://azuremlschemas.azureedge.net/latest/azureBlob.schema.json
name: raw_data
type: azure_blob
account_name: stmlrawdata
container_name: raw

az ml datastore create --file datastore.yml --workspace-name mlw-prod --resource-group rg-ml

Because no credentials appear in the file above, the datastore is identity based. The identity needs Storage Blob Data Reader (read) or Storage Blob Data Contributor (write) on the container.

Compute targets

$schema: https://azuremlschemas.azureedge.net/latest/amlCompute.schema.json
name: cpu-cluster
type: amlcompute
size: STANDARD_DS3_V2
min_instances: 0
max_instances: 4
idle_time_before_scale_down: 120
tier: low_priority

tier: low_priority uses cheaper preemptible VMs that can be taken away, so it suits fault tolerant training that writes checkpoints. A job can request serverless compute by omitting compute and setting resources with instance_type and instance_count.

Identity and access

Access is controlled with Azure RBAC at the workspace.

Identities come in three kinds: the user (Microsoft Entra ID), the workspace managed identity (used to reach storage, Key Vault and the registry), and compute managed identities (system or user assigned) that jobs use to read data. Prefer managed identities over keys, and give each one only the data role it needs.

A worked example

A team wants nightly training that reads a private data lake and costs nothing when idle. They create an ADLS Gen2 datastore with no credentials, a cluster with min_instances: 0 and a user assigned managed identity, and grant that identity Storage Blob Data Reader on the lake. Analysts get AzureML Data Scientist so they can submit jobs but cannot delete the cluster.

How the exam asks about this

• "Analysts must submit jobs but not create or delete compute." Assign AzureML Data Scientist.
• "Training must not incur cost when no jobs run." A compute cluster with minimum instances 0.
• "Remove stored keys from the datastore." Use identity based access and grant a storage data role.