ML Ops Study Guide — Machine Learning Professional

ML Ops – Databricks ML Professional Study Guide

This domain accounts for 44% of the exam (tied for the largest) and is what most distinguishes the Professional from the Associate: operationalizing models. It covers the MLflow Model Registry, model lifecycle governance, testing and CI/CD, automated retraining, environment management with Databricks Asset Bundles, and Lakehouse Monitoring for drift. Think in terms of the full production loop, not just training.

The MLflow Model Registry

The central store for managing the model lifecycle:

• Registered models and versions — each training produces a version.
• Aliases (Unity Catalog registry) — named pointers like @champion or @production to a specific version. Deployment references the alias; repointing it promotes or rolls back without changing consumer code.
• Governance — Unity Catalog permissions/grants control who can use or promote a model; lineage links a model to its training run, data version, and code.

Rollback pattern: because serving references an alias, repointing the production alias to the previous known-good version is an instant, safe rollback.

Lifecycle Automation

• Registry webhooks (or event-triggered Jobs) fire on lifecycle events (e.g., a stage transition) to run tests, notify, or deploy.
• Champion / Challenger — the Challenger is a candidate evaluated against the production Champion; promote it only if it passes validation and outperforms.
• Gated promotion — a CI/validation job runs on the candidate and only updates the production alias/stage if checks pass (e.g., meets a minimum metric on a holdout set).