MLOps
Free Databricks Certified Machine Learning Professional practice — 6 questions on MLOps, with explanations. No sign-up. Full 12-question mixed test →
A retail company manages fraud-detection models in Unity Catalog. They want zero-downtime promotion between a currently serving model and a newly trained candidate, avoiding the deprecated Staging/Production stage model entirely, while allowing serving code to always reference the 'current best' model without redeploying. Which approach BEST manages this Champion/Challenger workflow?
Unity Catalog model registry uses aliases (mutable pointers to a specific version) instead of the legacy stage model. set_registered_model_alias lets you repoint 'champion' to a new version instantly, and code referencing models:/<name>@champion automatically picks up the new version with zero code changes and zero downtime.
You configure a Lakehouse Monitoring inference profile on a table with a continuous numeric feature (transaction_amount) and a categorical feature (payment_method). Which statistical tests does Lakehouse Monitoring apply by default to detect distributional drift on these two columns respectively?
Lakehouse Monitoring's drift detection uses the Kolmogorov-Smirnov (KS) test to compare distributions of continuous numeric columns and the chi-square test to compare distributions of categorical columns between the baseline and current windows.
A fraud-detection model's performance degrades unpredictably due to data drift rather than on a fixed schedule. The ML team wants retraining triggered automatically only when drift is detected, to minimize unnecessary compute costs. Which architecture BEST meets this requirement?
Lakehouse Monitoring can compute drift metrics on inference tables and raise alerts when thresholds are breached. Wiring that alert to trigger a Databricks Job run (via the Jobs API or a webhook-based automation) creates an event-driven, drift-triggered retraining pipeline that avoids unnecessary scheduled runs and eliminates manual intervention.
Your ML team needs to deploy the same MLflow pipeline (training job, model registration, and serving endpoint) across dev, staging, and prod workspaces with environment-specific cluster sizes and model alias targets, fully version-controlled and deployable via CI/CD. Which Databricks tool is purpose-built for this?
Databricks Asset Bundles (DABs) are the source-controlled, YAML-based IaC framework for defining jobs, pipelines, and deployments, with 'targets' that let you override cluster configs, parameters, and alias references per environment (dev/staging/prod) and deploy consistently through CI/CD (e.g., `databricks bundle deploy -t prod`).
In a CI pipeline for a Databricks ML project, which test type is MOST appropriate for verifying that a custom PyFunc model's predict() method returns correctly shaped output for a fixed input schema, without invoking Spark or any external data source?
A unit test isolates a single function or method — here, predict() — from external systems (Spark clusters, databases, serving infrastructure) and checks its output shape/type against known input fixtures, making it fast, deterministic, and appropriate for CI without needing live infrastructure.
Before promoting a new model version to the @Champion alias in Unity Catalog, the team requires an automated validation job (checking accuracy, fairness metrics, and inference latency) to pass successfully. Which implementation enforces this gate as part of CI/CD?
Unity Catalog does not have a built-in automated validation gate — validation logic must be implemented explicitly. Encoding the validation checks and conditional alias assignment (set_registered_model_alias only on success) inside a Databricks Job, deployed and orchestrated via a Databricks Asset Bundle CI/CD pipeline, creates a repeatable, auditable, code-driven promotion gate that blocks bad models before they reach @Champion.
Ready for the real thing?
The full course has two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed answer explanations.
Start my full course on Udemy →