Model Deployment and Serving (Cloudera AI)
What the domain tests
This domain covers turning a registered model into a governed inference endpoint in Cloudera AI. Expect questions on model versions, deployment configuration, replicas, resource profiles, autoscaling, request and response formats, authentication, rollout and rollback, and the difference between batch scoring and real-time serving.
Core rules
- Deploy a specific model version, not 'latest' by default. A deployment should be traceable to a registered version and its artifacts.
- Match the serving runtime to the training environment. Missing libraries, wrong Python versions, or absent preprocessing artifacts are common failure causes.
- Size replicas and resource profiles for latency and throughput. One replica gives no availability during restarts; too many replicas waste quota. Load-test before setting autoscaling limits.
- Validate input at the endpoint. Reject malformed payloads with clear errors rather than letting the model fail unpredictably.
- Secure the endpoint with authentication and least-privilege access. Do not expose an unauthenticated public URL.
- Use canary or blue-green rollout when changing versions, and keep the previous version available for rollback.
Worked example
A registered churn model version 3 is deployed with two replicas on a CPU profile. The engineer sends a JSON payload with the same fields used in training, checks the response schema, and runs a small load test. After a week, version 4 is deployed to a canary receiving 10 percent of traffic; latency and error rate are compared before full cutover. Reasoning: two replicas avoid a single point of failure; the canary limits blast radius; the previous version remains for rollback.
Traps
- Deploying a model that references local training paths or project files not packaged in the artifact.
- Forgetting to include encoders, scalers, or feature-order metadata.
- Testing only with one happy-path request and no malformed input.
- Setting autoscaling without a maximum, exhausting cluster quota.
- Changing a model version in place with no rollback plan.
- Assuming batch scoring and real-time serving have the same timeout and payload limits.