AWS Machine Learning Specialty (AWS) key terms

Durable object storage that holds datasets as objects under key prefixes; it is the usual landing zone for a data lake and the source for Athena, Glue, EMR, and SageMaker training jobs.

A managed metadata repository of databases, tables, schemas, and partition definitions that Athena, Redshift Spectrum, EMR, and Glue jobs consult to locate data without moving it.

A job that scans a data store such as S3 or a JDBC source, infers file formats and column types, and writes or updates table definitions in the Glue Data Catalog.

An open columnar file format with per-column compression and statistics, so analytical queries read only the needed columns and skip irrelevant row groups, cutting scan cost on S3.

A real-time streaming service where records are appended to shards and retained for a configurable period, letting producers and consumers decouple at high throughput.

A serverless query service that runs standard SQL directly against files in S3 using table definitions from the Glue Data Catalog, billed per terabyte scanned.

Storing objects under prefixes keyed by column values such as date or region so query engines skip entire prefixes; this is the main lever for reducing bytes scanned and cost.

A permission layer over S3 data lakes that grants table- and column-level access through the Glue Data Catalog and centrally audits fine-grained access for analytics services.

A visual tool inside SageMaker Studio for importing, profiling, cleaning, and transforming tabular data, with built-in analyses and export to a processing job or pipeline.

Creating or transforming input variables such as ratios, aggregates, lags, or text embeddings so a model can learn a relationship more easily from raw data.

Replacing absent values with a substitute such as the mean, median, mode, or a model-predicted value so algorithms that reject nulls can still train on the record.

A condition where one label greatly outnumbers another, which inflates accuracy and requires resampling, class weights, or threshold tuning to be evaluated properly.

Including a feature that would not be available at prediction time, or that encodes the label, producing excellent offline scores that collapse in production.

A capability that computes bias metrics across demographic groups and produces feature-attribution explanations such as SHAP values for models trained and hosted in SageMaker.

Pre-packaged containers maintained by AWS, including XGBoost, Linear Learner, k-NN, BlazingText, and DeepAR, that train without custom code using standard record formats.

A gradient-boosted decision tree algorithm that builds trees sequentially to correct earlier errors; a strong default for tabular classification and regression and available as a SageMaker built-in.

A capability that launches many training jobs across a hyperparameter range and uses Bayesian optimization to find the configuration that best optimizes a chosen objective metric.

An automated machine learning feature that explores preprocessing and algorithms, generates candidate models, and exposes the underlying notebooks so results remain explainable and editable.

Splitting data into k folds, training k times with a different fold held out each time, and averaging the scores to estimate generalization better than a single split.

Adding a penalty on model complexity, such as L1 or L2 weight terms, or limiting tree depth, to reduce variance and stop the model from fitting noise.

The probability that a classifier ranks a random positive above a random negative; it is threshold-independent, so it summarizes ranking quality for imbalanced binary problems.

Starting from weights trained on a large source dataset and fine-tuning on a smaller target dataset, which reduces training time and labeled data needs for vision and text tasks.

A catalog of pretrained models, solution templates, and example notebooks that can be deployed to an endpoint or fine-tuned with a few clicks inside SageMaker Studio.

A managed run that provisions compute instances, pulls the container image and input data, executes the training script, and saves resulting model artifacts to S3.

A training input mode that streams records directly from S3 into the algorithm instead of downloading every file first, reducing startup time and local disk usage.

Training on spare EC2 capacity at a steep discount, with SageMaker checkpointing to S3 and resuming after interruptions, trading longer runtime for lower cost.

A hosted HTTPS inference service backed by one or more instances, offered in real-time, serverless, and asynchronous variants with autoscaling and optional data capture.

An offline inference job that runs a trained model over a large S3 dataset and writes predictions back to S3, avoiding the cost of a persistent endpoint.

A versioned catalog of model packages holding metadata, approval status, and lineage, used to gate promotion of a model from experimentation into production pipelines.

A workflow service for chaining processing, training, tuning, evaluation, and registration steps into a reproducible directed acyclic graph with parameters and step caching.

A tracking feature that records runs, parameters, metrics, and artifacts so many training attempts can be compared and organized into trials and experiment groups.

A monitoring tool that captures tensors and system metrics during training and raises built-in rules for problems such as vanishing gradients, overfitting, or excessive GPU load.

A feature that compares production endpoint inputs and outputs against a baseline and alerts on data drift, model quality decay, bias drift, and feature attribution drift.

A compilation service that optimizes a trained model for a specific hardware target such as an edge device, producing a smaller, faster artifact for on-device inference.

A managed repository that stores curated features online for low-latency serving and offline for training, keeping both consistent and letting teams reuse feature definitions.

A tool that benchmarks a model on candidate instance types and configurations, then recommends the least expensive option that still meets latency and throughput targets.