Skip to main content

Azure Machine Learning: workspaces, compute and pipelines

Take a demand-forecasting model for a retail chain from an empty resource group to a scheduled weekly pipeline, a managed online endpoint, and a batch endpoint — using the Azure ML SDK v2 and the az ml CLI.

Course Duration: 7h​

What You'll Learn​

  • Lay out an Azure ML workspace and its four attached resources cleanly, with the right identities and roles
  • Pick between a compute instance, an auto-scaling cluster, serverless compute, and low-priority nodes for training
  • Register file, folder and MLTable data assets against Blob Storage and ADLS, and read them by identity
  • Build reproducible environments from a base image, a conda file, or a full Dockerfile
  • Submit command and sweep jobs, log runs to MLflow natively, and read them in Studio
  • Decide when AutoML wins over a hand-built model and when it does not, at a cost you can measure
  • Register model versions with tags, lineage and cross-workspace sharing
  • Deploy a managed online endpoint with traffic splitting, and a batch endpoint on a cluster
  • Compose reusable components into a pipeline, schedule it weekly, or trigger it on new data
  • Monitor model drift, read cores quotas per region, and cost-tag every resource

Prerequisites​

  • MLOps foundations (Course 20)
  • Basic Azure: resource groups, storage accounts, role assignments
  • Python and the command line

Course Modules​

  1. Azure ML workspace and resource organization
  2. Compute targets: instances, clusters, serverless compute
  3. Data assets and datastores
  4. Environments and container images
  5. Training jobs and run tracking
  6. AutoML: relevant uses and limits
  7. Model registry and versioning
  8. Online and batch endpoints
  9. Pipelines and scheduling
  10. Monitoring, quotas and costs

The running example​

A single project runs through the ten modules: a weekly demand-forecasting model for a chain of grocery stores. It predicts the number of units sold per store, per product, for the next seven days, using two years of point-of-sale history. Module 1 stands up the workspace. Module 2 picks the right compute. Module 3 registers the sales data. Module 4 pins the environment. Module 5 trains a gradient-boosted regressor and logs it with MLflow. Module 6 pits AutoML against that baseline. Module 7 registers the winner. Module 8 exposes it as an online endpoint for ad-hoc queries and a batch endpoint for the weekly re-forecast. Module 9 packages every step as a pipeline that reruns each Sunday night. Module 10 puts a cost tag on everything and adds a drift monitor.

Tools you will use​

Azure ML SDK v2 (azure-ai-ml) for programmatic control, az ml for scripted operations, Azure ML Studio for exploration and debugging. MLflow is built into the training runs; Azure Monitor and Cost Management cover the observability side. Every tool is introduced when the running project needs it, never in isolation.

Assessment and certificate​

The course ends with a 40-question exam covering every module. On success, a certificate of completion is issued; its number is verifiable on the platform.

Free courses, by contrast, end with a 5-question quiz and a preview of the certificate, without certification.