7. Glossary

A/B Testing — Comparing two model versions in production by splitting live traffic between them. See §5.1.3.

Amazon Bedrock — Managed service for accessing and customizing foundation models from multiple providers. See §3.1.3. High exam relevance.

Amazon Comprehend — NLP service for sentiment analysis, entity recognition, key phrases and PII detection. Its topic modeling feature is no longer available to new customers; AWS points to Amazon Bedrock models for topic detection, and the SageMaker AI NTM and LDA algorithms also discover topics. See §3.1.4.

Amazon Comprehend Medical — Specialized NLP service for extracting medical entities from clinical text. See §3.1.4.

Amazon EMR — Managed Hadoop/Spark framework for large-scale data processing. See §2.2.4.

Amazon Kinesis — Managed streaming data platform for real-time data ingestion and processing. See §2.1.3. High exam relevance.

Amazon Macie — Service that uses ML to discover and protect sensitive data (PII/PHI) in S3. See §5.3.4.

Amazon Rekognition — Pre-built computer vision service for image and video analysis. See §3.1.4.

Amazon SageMaker — Comprehensive ML platform covering data prep through deployment and monitoring. See §1.3. Highest exam relevance — appears in 60-80% of questions.

Amazon Textract — Service for extracting text, tables, and forms from scanned documents. See §3.1.4.

Asynchronous Endpoint — SageMaker endpoint type for large payloads (up to 1 GB) and long processing times (up to 1 hour), with auto-scale-to-zero capability. See §4.1.1. High exam relevance.

Auto Scaling — Automatically adjusting compute resources based on demand metrics. For SageMaker endpoints, commonly based on InvocationsPerInstance. See §4.2.1.

Automatic Model Tuning (AMT) — SageMaker's hyperparameter optimization service using Bayesian optimization, random search, or grid search. See §3.2.2.

AWS CloudFormation — Infrastructure as code service using declarative JSON/YAML templates. See §4.2.2.

AWS CloudTrail — Service that logs all API calls in an AWS account for audit and compliance. See §5.2.1, §5.3.4. High exam relevance for security questions.

AWS CodeBuild — Managed build service for compiling code, running tests, and producing artifacts. See §4.3.1.

AWS CodeDeploy — Service for automating application deployments with strategies like blue/green and canary. See §4.3.1.

AWS CodePipeline — CI/CD orchestration service for automating release pipelines. See §4.3.1.

AWS Config — Service that monitors and evaluates resource configurations against compliance rules. See §5.3.4.

AWS Cost Explorer — Tool for visualizing and analyzing AWS spending. See §5.2.2.

AWS Glue — Serverless ETL service for data integration, built on Apache Spark. See §2.2.4. High exam relevance.

AWS Glue DataBrew — Visual data preparation tool with 250+ built-in transformations. See §2.2.4.

AWS Glue Data Quality — Service for defining and monitoring data quality rules. See §2.3.2.

AWS IAM — Identity and Access Management service for controlling access to AWS resources. See §5.3.1. High exam relevance for security questions.

AWS KMS — Key Management Service for creating and controlling encryption keys. See §5.3.3. High exam relevance.

AWS Lambda — Serverless compute service for running code without managing servers. See §4.1.1.

Batch Transform — SageMaker feature for large-scale offline inference without maintaining a persistent endpoint. See §4.1.1.

Bayesian Optimization — Hyperparameter search strategy that uses past results to intelligently choose next parameter combinations. Default strategy for SageMaker AMT. See §3.2.2.

Bias Drift — Change in model fairness metrics over time in production, detected by recomputing post-training bias metrics per segment on a schedule (SageMaker Model Monitor with Clarify for existing customers). See §5.1.1, §5.1.2.

Blue/Green Deployment — Deployment strategy maintaining two environments; traffic switches from old (blue) to new (green) with instant rollback capability. See §4.3.3. High exam relevance.

Bring Your Own Container (BYOC) — Custom Docker container for SageMaker training or inference when pre-built containers don't support your framework. See §4.1.3.

Canary Deployment — Deployment strategy that routes a small percentage of traffic to the new version, monitoring for errors before increasing. See §4.3.3.

Class Imbalance (CI) — Pre-training bias metric measuring uneven representation between FACET values (for example male vs female), not between the target label’s classes. See §2.3.1.

CloudWatch — Monitoring service for AWS resources, providing metrics, logs, and alarms. See §5.2.1. High exam relevance.

Concept Drift — Change in the relationship between input features and target variable over time. See §5.1.1.

Confusion Matrix — Table comparing predicted vs. actual classifications, enabling calculation of precision, recall, and F1 score. See §3.3.1.

Customer-Managed Key (CMK) — KMS key created and managed by the customer, providing full control over key policies. See §5.3.3.

Data Capture — SageMaker endpoint feature that stores a sample of inference requests and responses in S3, the input for drift and model-quality monitoring. See §5.1.2.

Data Drift — Change in the statistical distribution of input features compared to training data. See §5.1.1. High exam relevance.

Data Wrangler — SageMaker visual tool for data exploration, transformation, and feature engineering with minimal code. See §2.2.4, §1.3.2. High exam relevance.

Difference in Proportions of Labels (DPL) — Pre-training bias metric measuring label imbalance between demographic groups. See §2.3.1.

Distributed Training — Splitting training across multiple GPUs or instances using data parallelism or model parallelism. See §3.2.4.

Dropout — Regularization technique that randomly disables neurons during training to prevent overfitting. See §3.2.3.

Early Stopping — Training technique that halts training when validation loss stops improving, preventing overfitting and saving compute. See §3.2.2.

Epoch — One complete pass through the entire training dataset. See §3.2.1.

Evidently AI — Open-source library used in AWS's SageMaker AI monitoring solutions to detect data drift (PSI, KS) and compute model quality metrics against a baseline. See §5.1.2.

F1 Score — Harmonic mean of precision and recall, balancing both metrics. See §3.3.1.

Feature Engineering — Process of creating, transforming, and selecting input variables to improve model performance. See §2.2.2.

Feature Store — SageMaker service for storing, sharing, and reusing features across teams and models. See §1.3.2, §2.2.4.

Foundation Model — Large pre-trained model that can be fine-tuned for specific tasks (e.g., via Bedrock or JumpStart). See §3.1.3.

Ground Truth — SageMaker service for human data-labeling workflows (private or vendor workforces, plus automated labeling). Closed to new customers since 30 July 2026; existing customers can continue to use it. See §2.3.3.

Hyperparameter — Model configuration value set before training (e.g., learning rate, number of trees). See §3.2.2.

Inference Recommender — SageMaker tool that load-tests models across instance types to find optimal cost-performance. See §5.2.2.

JumpStart — SageMaker hub of pre-trained models, solution templates, and example notebooks. See §3.1.3.

L1/L2 Regularization — Techniques that add penalty terms to the loss function to prevent overfitting. L1 promotes sparsity; L2 penalizes large weights. See §3.2.3.

Managed Spot Training — SageMaker feature for using Spot Instances for training with automatic checkpointing and resumption. See §5.2.2.

MLflow on SageMaker AI — Managed MLflow (including serverless MLflow Apps) for logging parameters, metrics and artifacts, comparing training runs, and registering models. See §3.3.3, §3.4.5.

Model Monitor — SageMaker service for scheduled monitoring of deployed models (data quality, model quality, bias drift and feature attribution drift). No longer open to new customers; AWS recommends its open-source SageMaker AI monitoring solutions with CloudWatch for new teams. See §5.1.2. High exam relevance for the monitoring concepts.

Model Registry — SageMaker service for versioning, cataloging, and managing approval workflows for models. See §3.2.5.

Network Isolation — SageMaker configuration that prevents containers from making any outbound network calls. See §5.3.2.

One-Hot Encoding — Encoding technique that converts categorical values into binary vectors. See §2.2.3.

Overfitting — Model performs well on training data but poorly on unseen data due to memorizing noise. See §3.2.3.

Parquet — Columnar data format optimized for analytics and ML workloads with efficient compression. See §2.1.1.

Pipe Mode — SageMaker training mode that streams data directly from S3 rather than downloading to local disk. See §3.2.4.

Production Variant — A model version deployed behind a SageMaker endpoint that receives a configurable percentage of traffic. See §5.1.3.

Regularization — Techniques to prevent model overfitting by constraining model complexity. See §3.2.3.

RMSE (Root Mean Square Error) — Regression metric measuring the square root of average squared prediction errors. See §3.3.1.

ROC-AUC — Classification metric measuring the area under the receiver operating characteristic curve, evaluating model discrimination ability. See §3.3.1.

SageMaker Clarify — Service for computing bias metrics on data and models and explaining predictions with SHAP values. No longer open to new customers; AWS recommends computing the same metrics with pandas and scikit-learn and the SHAP library, logged to a SageMaker AI MLflow App. See §2.3.1, §3.3.2. High exam relevance for the bias and SHAP concepts.

SageMaker Debugger — Service that captured tensor data during training and applied built-in rules and a profiler to detect convergence and utilization problems. No longer open to new customers; AWS recommends MLflow for metric logging and run comparison, TensorBoard in SageMaker AI for gradient and weight analysis, and CloudWatch metrics and alarms for resource profiling and automated alerts. See §3.3.3.

SageMaker Neo — Service for compiling and optimizing models for deployment on edge devices. See §4.1.4.

SageMaker Pipelines — ML workflow orchestration service for building, automating, and managing end-to-end ML pipelines. See §4.3.2. High exam relevance.

Script Mode — SageMaker training approach where you provide your own training script with a supported framework (TensorFlow, PyTorch). See §3.2.1.

Serverless Endpoint — SageMaker endpoint type that auto-scales to zero when idle, suitable for intermittent traffic. See §4.1.1.

Shadow Variant — A model version that receives production traffic copies for evaluation but doesn't serve predictions to users. See §5.1.3.

SHAP Values (Shapley Values) — Method for explaining individual predictions by quantifying each feature's contribution. See §3.3.2.

SMOTE — Synthetic Minority Over-sampling Technique for addressing class imbalance by generating synthetic training examples. See §2.3.1.

Spot Instances — EC2 instances available at up to 90% discount but subject to 2-minute interruption notice. See §5.2.2.

TensorBoard in SageMaker AI — Managed TensorBoard application for inspecting training internals (gradient and weight histograms, scalars) of PyTorch, TensorFlow and Hugging Face Transformers jobs; AWS's recommended replacement for Debugger's tensor analysis. See §3.3.3.

Underfitting — Model is too simple to capture underlying data patterns, performing poorly on both training and test data. See §3.2.3.

VPC Mode — SageMaker configuration where training jobs and endpoints run inside a customer's VPC for network isolation. See §5.3.2. High exam relevance.