SageMaker

Provides fully managed Jupyter notebook instances that automatically scale compute resources without requiring infrastructure provisioning. Notebooks are hosted on AWS infrastructure with built-in IAM authentication, S3 integration, and pre-installed ML libraries (scikit-learn, TensorFlow, PyTorch). Users can start notebooks immediately without managing EC2 instances or container orchestration, with automatic shutdown policies to control costs.

Manages end-to-end distributed training execution across multiple compute instances (CPU and GPU) using a declarative job submission model. SageMaker Training handles resource provisioning, distributed training framework setup (TensorFlow, PyTorch, MXNet), data distribution across nodes, and automatic cleanup. Users define training scripts, specify instance types/counts, and SageMaker orchestrates the entire lifecycle including spot instance management for cost optimization.

Centralized repository for storing, versioning, and retrieving ML features (engineered data) for training and inference. The Feature Store manages feature definitions, handles feature versioning, and provides both batch and real-time feature retrieval APIs. Features are computed once and reused across multiple models, reducing redundant computation and ensuring consistency between training and inference feature sets.

Provides an AI-powered assistant integrated into SageMaker notebooks and the AWS console that helps users discover data, build training models, generate SQL queries, and create data pipeline jobs through natural language prompts. Q generates Python code, training configurations, and pipeline definitions based on user intent, reducing boilerplate and accelerating ML workflow setup. The assistant is trained on AWS documentation and SageMaker best practices.

Centralized discovery and governance platform (built on Amazon DataZone) for finding datasets, models, and ML artifacts across the organization. The Catalog enables data lineage tracking, access control, and metadata management for all ML assets. Users can search for datasets by business domain, view data quality metrics, and request access through approval workflows integrated with IAM.

Runs batch prediction jobs on large datasets without requiring real-time endpoints. Batch transform jobs read data from S3, invoke the model on each record, and write predictions back to S3. Supports data transformation before/after inference and automatic parallelization across multiple instances. Ideal for offline prediction scenarios (nightly scoring, bulk recommendations).

Enables deploying SageMaker models across multiple AWS accounts and regions for disaster recovery, compliance, and low-latency serving. Models are registered in a central account and deployed to endpoints in regional or cross-account environments. Supports model replication and automatic failover between regions.

Automatically tunes model hyperparameters by launching multiple training jobs with different parameter combinations and selecting optimal configurations using Bayesian optimization. SageMaker Hyperparameter Tuning evaluates objective metrics (accuracy, loss, F1) across training jobs, applies early stopping to terminate unpromising runs, and returns ranked hyperparameter sets. The service manages all training job provisioning, metric collection, and optimization algorithm execution.

Centralized repository for storing, versioning, and governing trained models with metadata tracking (training parameters, metrics, data lineage) and approval workflows. The Model Registry integrates with SageMaker training jobs to automatically register models, supports semantic versioning, and enables role-based access control (RBAC) for model promotion across environments (dev → staging → production). Models are stored with full provenance including training job ID, dataset version, and hyperparameters.

Deploys trained models as REST API endpoints with automatic scaling based on request volume and latency metrics. SageMaker Endpoints manage containerization, load balancing, and health checks across multiple instances. The service supports multi-model endpoints (hosting multiple model versions on shared infrastructure) and A/B testing by routing traffic between endpoint variants. Endpoints integrate with CloudWatch for monitoring and support custom inference code via Docker containers.

Defines and executes multi-step ML workflows as directed acyclic graphs (DAGs) using SageMaker Pipelines, which orchestrates data preprocessing, training, evaluation, and model registration steps. Pipelines support conditional execution (branching based on metrics), parameter injection, and integration with SageMaker training/processing jobs. Workflows are defined in Python using a declarative API and can be triggered manually or on schedules via EventBridge.

Processes large datasets through trained models asynchronously using Batch Transform, which reads input data from S3, distributes inference across multiple instances, and writes predictions back to S3. The service handles data partitioning, parallel processing, and automatic cleanup of compute resources. Batch Transform supports CSV, JSON, and Parquet formats and can process datasets ranging from gigabytes to terabytes without requiring real-time API endpoints.

Provides a curated catalog of pre-trained models (LLMs, vision models, NLP models) that can be deployed directly or fine-tuned on custom data. JumpStart models are packaged with inference containers and training scripts, enabling one-click deployment or transfer learning without model architecture knowledge. The service supports fine-tuning on custom datasets with automatic hyperparameter selection and handles model versioning and updates.

Provides a visual, spreadsheet-like interface for non-technical users to build ML models without writing code. Canvas handles data upload, automatic feature engineering, model training, and prediction generation through a drag-and-drop UI. The service automatically selects algorithms, tunes hyperparameters, and generates predictions that can be exported to CSV or integrated with business applications via APIs.

Manages large-scale data labeling workflows using a combination of human annotators, automated labeling, and active learning to reduce labeling costs. Ground Truth routes data to human workers (via Amazon Mechanical Turk or private workforces), applies consensus voting for quality control, and uses active learning to identify high-value samples for annotation. The service integrates with SageMaker training to automatically generate labeled datasets for model training.