Generative models get the headlines; the predictive models, churn, demand, fraud, pricing, still carry the business, and they need the full software lifecycle: versioned data, reproducible training, gated promotion, monitored serving. Azure Machine Learning is that lifecycle as a service, and MLOps is applying to models the discipline this series applies to everything.

Workspace Architecture and Compute

The workspace is the governance unit, and the pattern matches the estate: workspaces per team or domain, dev and prod separated, deployed by Terraform with managed VNet isolation, private endpoints to the storage and Key Vault they attach, and Entra groups for the roles, the hub and project model from yesterday’s Foundry post shares this substrate. Compute divides cleanly: compute instances are personal dev boxes for notebooks, sized modestly with the idle shutdown enforced by policy because they are the space heaters of ML; compute clusters autoscale from zero for training jobs, spot backed for the experiments and on demand for the deadline runs, the Batch post economics exactly; and serverless compute removes the cluster management for teams that just want jobs to run. Data flows through datastores (pointers to the ADLS containers from September, identity based access, no keys) and versioned data assets, because “which data trained this model” must be an answer, not an investigation.

Pipelines, Environments, Registry

Training is a pipeline of components, data prep, train, evaluate, register, each a versioned piece with typed inputs and outputs, defined in YAML and Python, run in the CLI from your CI system on triggers: code merge, new data version, schedule, or the drift signal below. Environments pin the runtime, curated images plus your conda spec, built and cached, so the reproducibility story covers code, data, and dependencies together, the same triad the IaC posts demanded. The model registry is the promotion gate: training registers candidates with their metrics, lineage, and the evaluation report; promotion to production tags follows a comparison against the champion on the golden dataset plus the human approval your risk profile demands, and the deployment pipeline (tomorrow’s post) consumes only registered, promoted versions. MLflow compatibility means the tracking and packaging conventions transfer, and the registries feature shares blessed models and components across workspaces, the module registry pattern for ML assets.

The Loop That Makes It MLOps

A trained model is inventory; the loop is the product. Monitoring on the serving side captures inputs and predictions (sampled, privacy filtered) and computes drift, input distributions shifting from training data, prediction distributions shifting from history, and where ground truth arrives later, the actual performance metrics as labels land. Thresholds on those signals raise the retraining trigger, which fires the pipeline, which registers a candidate, which passes or fails the gate, and the humans in the loop review the diff, not the ceremony. Close it with the estate’s machinery: job and endpoint diagnostics into Log Analytics, cost per training run and per thousand inferences on the FinOps dashboards, responsible AI reports (explanations, fairness metrics on the sensitive features you declare) attached to registry entries where regulation or ethics demands them. Most organizations’ ML problem is not modeling, it is that models decay silently while everyone builds new ones; the loop is the fix, and it is pipelines, gates, and monitors, the same nouns as the rest of this series.

Cheers
Osama