Machine Learning Operations (MLOps) Flashcards

Which AWS service can be used to create event-driven model retraining pipelines by reacting to new data uploaded to S3?

Answer: Amazon EventBridge

Amazon EventBridge can detect S3 events (such as new file uploads) and trigger downstream actions like starting a SageMaker Pipeline for retraining automatically.

What does 'training-serving skew' mean in ML operations?

Answer: When the features or data preprocessing used during training differ from those used during inference

Training-serving skew occurs when the feature engineering or data transformations applied at training time differ from those at inference time, causing unexpected model behavior in production.

Which SageMaker Model Monitor schedule type detects changes in data quality of live inference input data?

Answer: Data Quality Monitor

The Data Quality Monitor compares statistical properties of real-time inference inputs against a baseline to detect data drift such as changes in distributions, missing values, or type violations.

In MLOps, what is the purpose of a model approval gate in a deployment pipeline?

Answer: To require a human or automated review before a trained model is promoted to production

A model approval gate (such as the 'Approved' status in SageMaker Model Registry) ensures that trained models undergo validation — automated or manual — before being deployed to production.

What is the benefit of using Amazon SageMaker Pipelines over AWS Step Functions for ML workflows?

Answer: SageMaker Pipelines has native integrations for ML-specific steps like training, processing, and model registration

SageMaker Pipelines provides first-class ML step types (TrainingStep, ProcessingStep, RegisterModel, etc.) with lineage tracking and native SageMaker integration, making it purpose-built for ML workflows.

What is ML lineage tracking and why is it important?

Answer: Recording the complete chain of artifacts, steps, and parameters that produced a specific model

ML lineage tracking records end-to-end provenance — what data, code, hyperparameters, and steps produced each model — enabling reproducibility, debugging, and regulatory compliance.

Which statement best describes the 'continuous delivery' (CD) component specific to MLOps?

Answer: Automatically deploying approved models to a staging or production environment after passing validation gates

In MLOps, CD automates the deployment of models that have passed evaluation and approval gates to staging and eventually production, ensuring consistent and repeatable releases.