Azure Machine Learning Platform
End-to-end ML lifecycle management using Azure ML workspaces, managed endpoints, MLflow integration, and Responsible AI tooling.
Overview
The Azure ML workspace
The Azure Machine Learning workspace is the top-level resource that organizes every ML asset: experiments, datasets, environments, models, compute targets and endpoints. Anything created from the SDK, the CLI or the Studio UI is registered there and visible to the whole team.
- Azure Storage Account — Run artifacts
- Azure Container Registry (ACR) — Environment images
- Azure Key Vault — Secrets
- Application Insights — Endpoint telemetry
Access and network isolation
- Data scientists: submit runs and register models.
- ML engineers: deploy endpoints.
- Administrators: manage compute and quotas.
- Network: workspaces support private endpoints and VNet integration for enterprise isolation.
> [!NOTE] Role-based access control (RBAC) is set at the workspace level, which is what separates these three roles.
Experiments, runs and environments
An experiment is a named group of training runs. A run is one execution of a training script and records its parameters, metrics and artifacts.
> [!TIP] Azure ML uses MLflow as its native tracking backend, so mlflow.log_metric(), mlflow.log_param() and mlflow.log_artifact() work unchanged and experiments stay portable.
Environments define the Python packages, conda dependencies and Docker base image for a training job or endpoint.
- Curated — Pre-built for PyTorch, TensorFlow, HuggingFace and cached in ACR — Faster startup
- Custom — Defined in YAML, built once — Reused across runs
Compute targets
- Compute instance — Interactive development — Single-node VM
- Compute cluster (AmlCompute) — Batch training — Auto-scales; scales to zero when idle
- Attached Kubernetes (AKS) — Serving — Attached cluster
- Azure Databricks — Large-scale data processing — Spark clusters
> [!TIP] Compute clusters scale to zero when idle, so teams pay only for active training time.
Azure ML pipelines
Pipelines chain steps into a reproducible DAG. Each step runs on its own compute target and environment, so one pipeline can mix Spark and GPU work.
Managed endpoints
Online endpoints serve real-time inference with SLA-backed availability. A deployment specifies the model, environment, compute SKU and instance count, and traffic is split across deployments by percentage.
- Online — Real-time, SLA-backed — Chosen SKU and instance count — Response per request
- Batch — Asynchronous over large files — AmlCompute — Predictions written to blob storage
> [!NOTE] Percentage traffic splits give blue/green and canary rollouts natively, with no extra routing layer.
Responsible AI dashboard
The Responsible AI dashboard in Azure ML Studio combines four analyses in one view. It is generated by a pipeline component that runs after training and is attached to the registered model.
Azure Databricks integration
Databricks is the preferred platform for large-scale feature engineering and distributed training.
- Attached compute: Databricks clusters can be used as Azure ML training targets.
- Shared tracking: runs started in Databricks appear in the Azure ML experiment view.
- Shared features: Feature Store tables feed Azure ML training jobs and online endpoints directly.
Interview questions
- Design an end-to-end MLOps pipeline on Azure that covers data ingestion, training, evaluation, registration, and deployment with CI/CD gates.
- How would you monitor an Azure ML online endpoint in production and respond to model drift?
- Explain how you would use Azure ML's Responsible AI Dashboard to detect and mitigate bias in a credit scoring model.
Sign up free to read the full topic
More MLOps & LLMOps topics
Start free on Cariara · Coding problems · System design studio