Design and Implement an MLOps Infrastructure
MLOps infrastructure questions test whether you can build a repeatable, governed path from a data scientist's notebook to a production deployment. Know the resource hierarchy: a workspace holds compute, environments, datastores, and job definitions; a registry shares models, components, and environments across workspaces; datastores point at storage while data assets are versioned references to actual data. Distinguish compute instance (single-user dev VM with idle shutdown) from compute cluster (autoscaling multi-node for training jobs) and serverless compute (no quota to manage).
CI/CD is typically a repo (Azure DevOps or GitHub) plus infrastructure-as-code (Bicep or Terraform) plus CLI/SDK pipelines. A defensible split: CI validates code, builds and versions an environment image, runs unit tests and a training smoke job; CD promotes the same artifacts through dev → test → prod workflows. Reproducibility requires that you version three things together - code, environment, and data reference - and keep lineage from model version back to the run that produced it.
Traps. Single-workspace-everything vs. isolated dev/prod workspaces (the exam favors isolation for prod governance). Confusing a datastore with a registered data asset. Choosing a compute instance for a scheduled retraining job. Forgetting private endpoints and managed virtual networks when the scenario mentions data exfiltration or a VNet. Assuming a managed identity works before you assign the RBAC role on the target resource.
Rule of thumb: if the question asks where does this thing live and who can touch it, it is an infrastructure question - answer with scope (workspace vs. registry vs. subscription), identity, and network boundary before naming a service.