AWS Machine Learning Specialty (AWS) key terms

Durable object storage service that holds raw datasets, training inputs, and model artifacts; exam scenarios often use it as the source and sink for SageMaker jobs and pipelines.

Tiering option such as Standard, Intelligent-Tiering, or Glacier that trades retrieval latency and cost; choosing one for training data affects both pipeline speed and monthly spend.

Serverless data integration service that catalogs datasets and runs Apache Spark extract-transform-load jobs, commonly used to prepare features before model training in SageMaker.

Central metadata repository storing table definitions, schemas, and partition information; services such as Athena, Redshift Spectrum, and SageMaker read from it to locate data.

Automated process that scans data stores, infers schemas and partitions, and writes table definitions into the Data Catalog so downstream query and training tools can use them.

Columnar file format with compression and schema embedding that reduces scan volume and cost; converting CSV training data to Parquet is a frequent optimization in exam scenarios.

Serverless query service that runs SQL directly against files in S3 using table definitions from the Data Catalog, useful for ad hoc exploration without loading a warehouse.

Real-time streaming service that ingests ordered records at high throughput; feature pipelines and online inference systems read from it to process events within seconds.

Managed delivery service that buffers streaming records and writes them to destinations such as S3, Redshift, or OpenSearch, often landing raw events for later batch training.

Service that centralizes fine-grained permission management over data lake tables in S3 and the Data Catalog, letting teams share training data without copying it.

Managed labeling workflow that combines human annotators, automated labeling, and active learning to produce high-quality labeled datasets for supervised training.

Managed run that provisions compute, pulls training data, executes a container's training script, and writes model artifacts to S3, with metrics and logs emitted to CloudWatch.

Python SDK object that declares a training job's container image, instance type and count, hyperparameters, and input channels before calling fit to launch it.

Hyperparameter optimization feature that runs many training jobs using Bayesian or random search strategies to find configurations optimizing a chosen objective metric.

Configuration value set before training, such as learning rate, tree depth, or batch size, that governs how an algorithm learns rather than being learned from data.

Failure mode where a model memorizes training noise and generalizes poorly; detected by a widening gap between training and validation metrics and addressed with regularization or more data.

Technique that penalizes model complexity, through L1 or L2 weight penalties, dropout, or early stopping, to reduce variance and improve performance on unseen data.

Training control that halts when validation performance stops improving for a set number of rounds, saving compute and limiting overfitting without changing the model family.

Resampling procedure that rotates held-out folds so every observation serves as validation once, producing a more reliable estimate of generalization than a single split.

Partitioning data into a fitting set, a tuning set for model selection, and an untouched test set for final unbiased evaluation; leakage between them inflates reported accuracy.

Contamination where information unavailable at prediction time, such as target-derived features or future rows, enters training and produces optimistic offline metrics that fail in production.

Situation where one label vastly outnumbers others, making accuracy misleading; remedies include resampling, class weights, and choosing precision-recall metrics instead of raw accuracy.

Table of predicted versus actual classes showing true positives, false positives, true negatives, and false negatives, the basis for deriving precision, recall, and related metrics.

Precision measures how many predicted positives are correct; recall measures how many actual positives were found. Their tradeoff depends on the business cost of false positives versus misses.

Harmonic mean of precision and recall, giving a single balanced number that is useful when classes are imbalanced and both false positives and false negatives matter.

Area under the receiver operating characteristic curve, measuring how well a classifier ranks positives above negatives across all thresholds; 0.5 is random and 1.0 is perfect.

Creating, transforming, or selecting input variables, through encoding, scaling, binning, or aggregation, so that patterns become easier for the algorithm to learn.

Repository with online and offline stores that keeps feature definitions consistent between training and real-time inference, reducing skew from separately computed pipelines.

Divergence between how features are computed during training and during inference, which silently degrades live predictions; shared feature pipelines and stored features prevent it.

Hosted HTTPS inference service backed by one or more model instances; supports real-time prediction with autoscaling and can run multiple models behind a single URL.

Offline inference mode that loads a model, processes a large S3 dataset in batches, and writes predictions back to S3 without keeping a persistent endpoint running.

Policy-driven adjustment of endpoint instance count based on CloudWatch metrics such as invocations per instance, balancing latency targets against cost during traffic swings.

Monitoring service collecting metrics, logs, and alarms; used to watch training loss, endpoint latency, and data drift signals and to trigger automated responses.

Capability that compares live endpoint traffic against a baseline and raises alarms for data drift, model quality decay, bias drift, and feature attribution drift.

Workflow orchestration service that chains processing, training, evaluation, and registration steps as a directed acyclic graph with lineage tracking and repeatable execution.

Catalog of trained model versions with metadata and approval status, enabling controlled promotion from development to production and supporting automated deployment workflows.

Using discounted spare EC2 capacity for training jobs with checkpointing so work resumes after interruption, cutting cost substantially for fault-tolerant workloads.

Feature that keeps provisioned instances alive between jobs so subsequent training or processing starts quickly, trading idle cost for reduced startup latency.

AWS identity assumed by a SageMaker job or endpoint that grants exactly the S3, ECR, and CloudWatch permissions needed, following least-privilege access principles.