AWS ML Engineer Associate key terms

A key path segment used to organize objects in a bucket, commonly treated as a logical folder. Training jobs and pipelines point here to read datasets and write model artifacts.

Organizing dataset objects into prefixes by values such as date or category so query engines and distributed training jobs read only relevant files, reducing scan volume and improving throughput.

A training input channel mode that streams record data directly from S3 to the algorithm instead of downloading the full dataset first, reducing startup time and disk usage.

A training input channel mode that downloads the entire S3 dataset to local storage before training begins. Suitable for small datasets or algorithms needing random access.

A managed data-labeling service that combines human labelers with automated labeling and active learning to produce high-quality annotated datasets for supervised training.

A labeling strategy where the model identifies unlabeled examples it is least confident about, so human effort focuses on the most informative samples and total annotation cost drops.

A visual data preparation tool for importing, profiling, transforming, and exporting tabular and text data, with built-in analyses that detect data quality and leakage issues.

A modeling error where information unavailable at prediction time, such as target-derived columns or post-outcome fields, is used during training, inflating measured accuracy and degrading production performance.

Dividing a dataset into subsets for fitting parameters, tuning hyperparameters, and estimating final generalization performance, keeping the test set untouched until evaluation.

A resampling technique that rotates which data folds serve as validation across several training rounds, giving a more stable performance estimate than a single split.

A condition where one target class greatly outnumbers others, causing models to favor the majority class; addressed with resampling, class weights, or metrics like recall and F1.

Selecting a representative subset of records for training, evaluation, or labeling, used to control cost and class balance while preserving statistical fidelity.

A repository for storing, discovering, and serving curated features consistently for training and real-time inference, with online and offline stores kept in sync.

The low-latency portion of a feature repository that serves the latest feature values per entity for real-time inference requests, typically backed by a fast key-value database.

The storage layer holding historical feature values in files for batch training and analysis, enabling point-in-time joins so training data reflects what was known then.

A managed compute task that runs a training container on specified instance types, reads input channels from S3, and writes model artifacts and logs to S3.

The output files produced by training, such as serialized weights and configuration, packaged as a compressed archive in S3 and later used to create a deployable model.

A hyperparameter optimization feature that runs many training jobs using search strategies such as Bayesian optimization to find the configuration that best optimizes a chosen objective metric.

Systematically searching combinations of settings like learning rate and tree depth that govern training, using search strategies and early stopping to reduce cost.

Choosing the measurement that reflects business success, such as precision, recall, F1, or RMSE, and using it consistently for tuning, thresholding, and comparing candidate models.

A capability that captures tensors and system metrics during training and applies built-in or custom rules to detect problems such as vanishing gradients, overfitting, or stalled progress.

A tracking service that records training runs with their parameters, metrics, and artifacts, letting teams compare trials and reproduce the best configuration.

A catalog for versioning trained models, storing metadata and approval status, and managing promotion across environments so deployment references an approved version.

A governance process where a registered model version is reviewed against quality and compliance criteria and marked approved before a pipeline deploys it.

A managed workflow service for defining ML steps such as processing, training, evaluation, and registration as a directed acyclic graph that can be scheduled and rerun.

A managed task for running data transformation, feature engineering, or model evaluation scripts on dedicated compute, writing outputs to S3 for downstream steps.

Automation that builds, tests, and deploys model and pipeline changes through environments, extending software delivery practices to models and their supporting infrastructure.

A persistently hosted HTTPS inference service backed by one or more instances, providing low-latency predictions with configurable instance count and autoscaling.

An endpoint option that provisions compute on demand and scales to zero when idle, suited to intermittent or low-volume traffic where cold-start latency is acceptable.

An offline inference mode that processes a large S3 dataset with a deployed model and writes predictions to S3, without maintaining a persistent endpoint.

An endpoint mode that queues large or long-running requests, returns results via S3 when complete, and can scale to zero between requests.

Dynamically adjusting the number or size of instances behind an endpoint in response to traffic metrics, balancing latency targets against cost.

Routing a copy of live traffic to a new model version while the current version serves users, letting teams compare behavior before switching production traffic.

Deploying a new model version to a separate environment, validating it, then shifting traffic over so rollback means returning traffic to the unchanged original environment.

Reverting production to a previously known-good model version after a regression or failure, which requires versioned artifacts and a tested reversal path.

A service that captures endpoint or batch inference data and compares it against a baseline to detect drift in data quality, model quality, bias, and feature attribution.

A change in the statistical distribution of input features over time relative to training data, which can degrade predictions even when the underlying relationship is unchanged.

A change over time in the relationship between input features and the target, so the same inputs should now map to different outputs and the model must be retrained.

A reference set of metric values, such as accuracy or error thresholds, captured from validation data and used to judge whether live performance has degraded.

An automated condition, such as drift exceeding a threshold or a metric falling below target, that starts a pipeline run to refresh the model.

A tool that computes bias metrics on data and models and generates feature attribution explanations, helping teams detect unfair outcomes and justify predictions.

An identity assumed by a SageMaker resource to access other AWS services, whose attached policies determine which buckets, repositories, and endpoints the job can use.

A private network connection that lets training and hosting resources reach S3 and other services without traversing the public internet, supporting network isolation requirements.

Key management that encrypts data at rest in S3, EBS volumes, and feature stores, with key policies controlling which roles may decrypt training data and artifacts.