I85PREP
⌘ K
Cloudera Data Operator study materials
Cloudera Data Operator study resources

Cloudera Data Operator: Enterprise AI on Cloudera Machine Learning

Cloudera Data Operator · Work through your exam, one topic at a time.

All guide sections

Enterprise AI on Cloudera Machine Learning

The Enterprise AI domain tests how Cloudera Machine Learning (CML) supports the model lifecycle on CDP: experimentation, reproducible environments, deployment, and governance. Questions often ask where a step happens-project, session, job, model, or deployment-and how AI artifacts inherit lakehouse security.

Core concepts

  • CML is a Kubernetes-based data science service in CDP. It provides projects (Git-backed), sessions (interactive compute engines), jobs (scheduled or triggered runs), experiments (tracking parameters and metrics), models (registered artifacts), and deployments (REST endpoints).
  • Engines are containerized runtime templates (Python, R, Scala, Spark) with resource profiles for CPU/GPU and memory. Sessions are ephemeral; jobs are for repeatable pipelines; deployments run model replicas behind an endpoint.
  • AMPs (Applied ML Prototypes) are end-to-end use-case templates with code, data, and instructions. They accelerate starts but are not automatically production-hardened.
  • Governance: CML integrates with SDX. Ranger controls data access; Atlas captures lineage from data to model; the model registry versions artifacts; deployments run under a defined identity and can be monitored for drift in supported versions.
  • Reproducibility: projects track code via Git; engines pin dependencies; experiments log parameters and metrics; model builds capture artifacts and environment.

Worked example

A data scientist must train on a Hive table containing PII and deploy a REST endpoint without exposing raw PII. Use a CML project; create a Spark session; read the Hive table under Ranger-authorized access; mask or tokenize PII in the feature pipeline; log the experiment; register the model; deploy it as a model endpoint. Atlas records lineage from Hive table to model; Ranger enforces column masking for other users. The endpoint does not bypass governance because it runs under a defined identity.

Traps

  • Thinking CML sessions persist indefinitely; they are ephemeral and resource-bounded.
  • Confusing model registration with model deployment; registering does not automatically create a serving endpoint.
  • Assuming AMPs are production-ready; they need review, data substitution, and testing.
  • Ignoring that Git-backed projects are the source of truth; manual edits outside Git break reproducibility.
  • Believing CML bypasses Ranger because it is a data science tool; all data access is governed.