Design and implement ML solutions
This domain tests whether you can translate a business problem into an Azure Machine Learning solution that meets latency, cost, security, and operability constraints. Design decisions include workspace and environment topology, compute choice, data and model assets, identity, and serving target.
Core rules
- Use separate workspaces or at least separate environments for dev, test, and prod; promote code, environments, data assets, and registered models, not manual portal state.
- Choose managed online endpoints for interactive low-latency scoring; batch endpoints for scheduled bulk scoring; Kubernetes online endpoints when you must control cluster networking or existing AKS.
- Prefer managed identity and Azure Key Vault over stored secrets. Put private endpoints and managed virtual networks in the design when data cannot traverse public networks.
- Pin environments, version data assets, register models, and keep training code in Git so a deployment can be reproduced.
Worked example
A fraud model must score card transactions in under 200 ms and use PCI data. The design uses a private-endpoint workspace, managed online endpoint with autoscale, model from the registry, managed identity to Key Vault, and Application Insights for latency and error telemetry. Batch was rejected because it cannot meet interactive scoring. A single shared prod/dev workspace was rejected because it weakens quota isolation and access control.
Traps
- Selecting batch for an interactive workload or online for nightly bulk scoring.
- Hard-coding keys in YAML or scoring code.
- Designing only the model and omitting quota, region, identity, and rollback.