I85PREP
⌘ K
DE MLO study materials
DE MLO study resources

DE MLO: ML Governance and Reproducibility

DE MLO · Work through your exam, one topic at a time.

All guide sections

ML Governance and Reproducibility

What the exam tests

This domain tests whether you can govern models and reproduce results later. Questions cover Unity Catalog permissions and lineage, model ownership, audit logs, immutable versions, Git integration, environment capture, and the evidence needed to explain how a deployed model was produced.

Core rules

  • Unity Catalog governs models, tables, and functions with a three-level namespace. Model versions inherit governance from the registered model and schema, and permissions such as USE CATALOG, USE_SCHEMA, CREATE MODEL, and EXECUTE control access.
  • Lineage connects catalog objects. A model version links to its MLflow run, and the run should record source table versions, Git commit, parameters, metrics, and artifacts. This is the audit trail for reproduction.
  • Model versions are immutable. Governance means creating a new version for changes and moving aliases, not overwriting an existing artifact.
  • Capture the environment: library versions, cluster or runtime configuration, and framework versions. A notebook alone is not a reproducible environment.
  • Use Git folders or logged Git metadata so code changes are traceable. Record random seeds where supported.
  • Protect sensitive data with Unity Catalog privileges, row filters, and column masks. Audit logs show who accessed or changed governed objects.

Worked example

An auditor asks how main.default.fraud_model version 7 was built and why it is live. The team shows alias champion points to version 7. The version links to an MLflow run with Git commit a1b2c3, training table version 42, hyperparameters, and validation metrics. The run used a recorded runtime and library set. Unity Catalog lineage shows the source tables and downstream serving endpoint. Audit logs show the alias change and who approved it. To reproduce, the team checks out the commit, restores the table version, and reruns with the same parameters and seed. Reasoning: the version, run, lineage, and audit records together provide evidence; a notebook path alone does not.

Traps that cost marks

  • Relying on notebook state or an untracked interactive session as the source of truth.
  • Overwriting training data or model artifacts, which destroys reproducibility.
  • Logging metrics without Git commit, data version, or environment details.
  • Granting broad schema or catalog privileges instead of least-privilege access.
  • Ignoring lineage when investigating a production incident.
  • Treating an alias change as invisible; it should be governed and auditable.

Key points

  • Unity Catalog provides permissions, lineage, and auditability for models and data.
  • Model versions are immutable; promote by creating versions and moving aliases.
  • Reproduce with Git commit, data version, environment, parameters, and seeds.
  • Row filters and column masks protect sensitive data without broad access.
  • Audit logs and lineage are evidence for governance reviews and incident response.