Azure Machine Learning is Microsoft's managed ML platform for training, tracking, deploying, and monitoring models. It covers the full classical-ML lifecycle — data preparation, AutoML, hyperparameter sweeps, distributed training, managed endpoints, and MLOps pipelines — and underpins Azure AI Foundry's generative-AI tooling.
from azure.ai.ml import MLClient, command, Input
from azure.ai.ml.constants import AssetTypes
from azure.identity import DefaultAzureCredential
ml = MLClient(DefaultAzureCredential(),
subscription_id="<sub>", resource_group_name="<rg>",
workspace_name="my-aml-ws")
job = command(
code="./src", # folder with train.py
command="python train.py --data ${{inputs.data}} --epochs 10",
environment="azureml://registries/azureml/environments/sklearn-1.5/versions/latest",
compute="cpu-cluster",
inputs={"data": Input(type=AssetTypes.URI_FOLDER, path="azureml:titanic:1")},
experiment_name="titanic-gbm",
display_name="gbm-sweep-base",
)
returned = ml.jobs.create_or_update(job)
print(returned.studio_url)
from azure.ai.ml.sweep import Choice, Uniform
sweep = job.sweep(
compute="cpu-cluster",
sampling_algorithm="bayesian",
primary_metric="val_auc",
goal="Maximize",
max_total_trials=20,
max_concurrent_trials=4,
)
sweep.search_space = {
"learning_rate": Uniform(min_value=0.01, max_value=0.3),
"max_depth": Choice(values=[3, 5, 7, 9]),
}
ml.jobs.create_or_update(sweep)
from azure.ai.ml import automl, Input
from azure.ai.ml.constants import AssetTypes
automl_job = automl.classification(
compute="cpu-cluster",
experiment_name="churn-automl",
training_data=Input(type=AssetTypes.MLTABLE, path="azureml:churn-train:1"),
target_column_name="churned",
primary_metric="AUC_weighted",
n_cross_validations=5,
enable_model_explainability=True,
)
automl_job.set_limits(timeout_minutes=60, max_trials=30, max_concurrent_trials=4)
ml.jobs.create_or_update(automl_job)
from azure.ai.ml.entities import ManagedOnlineEndpoint, ManagedOnlineDeployment, Model, Environment
endpoint = ManagedOnlineEndpoint(name="churn-endpoint", auth_mode="key")
ml.online_endpoints.begin_create_or_update(endpoint).result()
deployment = ManagedOnlineDeployment(
name="blue",
endpoint_name="churn-endpoint",
model=Model(path="azureml:churn-gbm:3"),
environment=Environment(image="mcr.microsoft.com/azureml/inference-base:latest"),
instance_type="Standard_DS3_v2",
instance_count=1,
)
ml.online_deployments.begin_create_or_update(deployment).result()
ml.online_endpoints.begin_update(ManagedOnlineEndpoint(name="churn-endpoint", traffic={"blue": 100})).result()
from azure.ai.ml import dsl, Input, Output
prep = command(code="./prep", command="python prep.py --in ${{inputs.raw}} --out ${{outputs.clean}}",
environment="...", inputs={"raw": Input(type="uri_folder")},
outputs={"clean": Output(type="uri_folder")})
train = command(code="./train", command="python train.py --data ${{inputs.data}}",
environment="...", inputs={"data": Input(type="uri_folder")},
outputs={"model": Output(type="uri_folder")})
@dsl.pipeline(compute="cpu-cluster")
def churn_pipeline(raw_data):
p = prep(raw=raw_data)
t = train(data=p.outputs.clean)
return {"model": t.outputs.model}
pipeline_job = churn_pipeline(raw_data=Input(type="uri_folder", path="azureml:raw-churn:1"))
ml.jobs.create_or_update(pipeline_job)
Everything in Azure ML hangs off a workspace. Data assets point to data that stays in its storage service; jobs mount or download it at run time. An environment pins the container image and packages, and compute supplies the machines: compute instances for notebooks, auto-scaling clusters that scale to zero when idle, attached Kubernetes, or serverless compute where Azure provisions the nodes per job. A job combines code, environment, compute and inputs. Metrics and artifacts are logged through MLflow, which is the native tracking API in SDK v2, so standard mlflow.log_metric calls work unchanged. Trained models are registered with lineage back to the job that produced them and served through managed endpoints.
Access is controlled with Azure role-based access control at the workspace level, and jobs authenticate to data with either the submitting user's identity or a managed identity attached to the compute, which avoids storing storage keys in code.
# job.yml
$schema: https://azuremlschemas.azureedge.net/latest/commandJob.schema.json
code: ./src
command: python train.py --data ${{inputs.data}} --epochs 10
inputs:
data:
type: uri_folder
path: azureml:titanic:1
environment: azureml://registries/azureml/environments/sklearn-1.5/versions/latest
compute: azureml:cpu-cluster
experiment_name: titanic-gbmaz extension add -n ml
az ml job create -f job.yml -g my-rg -w my-aml-ws --stream
az ml model create --name churn-gbm --path azureml://jobs/<job-name>/outputs/artifacts/paths/model/ --type mlflow_model -g my-rg -w my-aml-ws
az ml online-endpoint invoke --name churn-endpoint --request-file sample.json -g my-rg -w my-aml-wsThe YAML and the Python SDK are two views of the same resource definitions, so teams often keep YAML files in Git and submit them from CI. The ${{inputs.data}} syntax is resolved by Azure ML on the compute target, not by the shell.
A managed online endpoint is a stable URL with key or Microsoft Entra token authentication; behind it sit one or more deployments, each with its own model version, environment, instance type and count. Traffic can be split between deployments for blue/green or canary releases, and mirrored traffic lets a new deployment receive a copy of live requests without affecting responses. Batch endpoints run scoring jobs over files on a compute cluster. Pipelines chain components (reusable, versioned steps) so data preparation, training and evaluation run as one reproducible graph, and schedules or Azure DevOps and GitHub Actions trigger them.
| Option | Best for | Trade-off |
|---|---|---|
| Azure Machine Learning | custom training, AutoML, governed MLOps on Azure | platform concepts to learn; several linked resources |
| Azure AI Foundry | generative AI apps, model catalog, agents, evaluation | less focused on classical tabular training |
| Azure OpenAI Service | calling OpenAI models directly | no custom training beyond fine-tuning |
| Databricks on Azure | Spark-heavy data engineering plus ML | separate workspace and billing model |
| AWS SageMaker / Vertex AI | equivalent platforms on other clouds | cloud lock-in either way |
versions/latest is convenient but not reproducible; pin a version for production.