Manage Azure ML Resources with Python SDK

Azure Machine Learning SDK v2 for Python is a client library for managing Azure ML resources: workspaces, jobs, models, data, and compute. It authenticates via MLClient with DefaultAzureCredential plus an explicit subscription ID, resource group, and workspace name, or via MLClient.from_config() reading a local config.json.

It manages workspaces (create and list), data assets (register a file or a whole folder as a versioned Data asset pointing at a datastore path), the model registry (register and list versioned models), compute clusters (an AmlCompute cluster with a min/max instance range and an idle-time-before-scale-down setting to control cost), custom environments (a Docker image plus a conda file for reproducible training), datastores (list them or fetch the workspace's default), and jobs - a command job runs a script against a compute target with typed inputs, and a @dsl.pipeline-decorated function chains multiple components into a multi-step pipeline job. Job progress can be streamed directly to the console, and every client entity - workspaces, jobs, models, data, compute, environments, datastores, components - exposes a consistent create_or_update/get/list operation set, with jobs additionally supporting stream and cancel and models supporting archive.

Use it when you need to manage Azure ML resources programmatically - provisioning compute, registering data (as an AssetTypes.URI_FILE or AssetTypes.URI_FOLDER) and models, submitting training jobs, or registering reusable components for pipelines - rather than doing it all through the Azure ML Studio UI. It is scoped to resource management and job orchestration, not to the model training code itself; the actual training script is still yours to write and is simply pointed to by a command or pipeline-component definition.

Input is Azure ML resource definitions (workspace, data, model, compute, environment, or job specs) plus the Azure subscription/resource-group/workspace identifiers. Output is the created or updated Azure ML resource - a running job with a studio URL, a registered model or data version, or a provisioned compute cluster. Install it with:

pip install azure-ai-ml

Its own best practices: version data, models, and environments explicitly; configure idle scale-down on compute clusters to control cost; use custom environments for reproducible training; stream job logs to monitor progress; register models only after a successful training job; use pipelines for multi-step workflows; and tag resources for organization and cost tracking.

It wraps the Azure ML control plane via azure-ai-ml and azure-identity, and its command jobs reference prebuilt or custom container environments such as AzureML-sklearn-1.0-ubuntu20.04-py38-cpu.

ML engineers and data scientists who manage Azure ML workspaces, data, models, compute, and training jobs programmatically from Python rather than through the Studio UI, and who need reproducible, versioned, cost-aware ML infrastructure.