CAP Model Deployment & Lifecycle Management Flashcards

What is the primary purpose of a model registry in an MLOps pipeline?

Answer: To centrally track model versions, metadata, and deployment status

A model registry provides a single source of truth for model versions, lineage, and stage transitions (staging, production, archived).

Which metric directly measures concept drift in a deployed classification model?

Answer: Accuracy or AUC degradation over time on live labels

Tracking predictive performance metrics (accuracy, AUC) on incoming labeled data over time directly reveals concept drift.

Shadow mode deployment means:

Answer: Running a new model in parallel without using its outputs for decisions

In shadow mode, the new model receives real inputs and generates predictions, but only the existing model's outputs drive actual decisions.

Which trigger should initiate a model retraining cycle?

Answer: Statistically significant performance degradation detected by monitoring

Retraining should be triggered by evidence of performance degradation, not arbitrary schedules, to ensure resources are spent when needed.

Feature stores in production analytics primarily serve to:

Answer: Provide consistent, reusable feature definitions across training and serving

Feature stores centralize computed feature logic so training and serving pipelines use identical transformations, preventing training-serving skew.

Which practice reduces training-serving skew in deployed models?

Answer: Sharing a single feature transformation pipeline for both training and serving

Sharing the same feature transformation pipeline ensures training and serving see identically processed features, eliminating skew.