AWS SageMaker

Provides fully managed Jupyter-based notebook instances hosted on AWS infrastructure with integrated Amazon Q Developer assistant for code generation, data exploration, and ML pipeline creation. Notebooks are pre-configured with common ML libraries and direct S3/Redshift access, eliminating local environment setup. The built-in AI agent generates SQL queries, discovers data sources, and scaffolds training code through natural language prompts.

Orchestrates distributed training jobs across multiple compute instances using a managed training job abstraction that handles data distribution, checkpoint management, and fault recovery. Automatic Model Tuning (AMT) layer runs Bayesian optimization over hyperparameter search spaces, launching parallel training jobs and selecting best-performing configurations based on user-defined metrics. Training jobs pull data from S3, log metrics to CloudWatch, and persist models back to S3 automatically.

Deploys multiple trained models to a single inference endpoint, enabling efficient resource utilization and simplified model management. Models are loaded into shared container instances and invoked by specifying the target model name in the request. Supports independent scaling per model and A/B testing across models. Reduces infrastructure costs by consolidating multiple low-traffic models onto shared instances.

Continuously monitors deployed model endpoints for data drift (input distribution changes), prediction drift (output distribution changes), and feature attribution drift. Compares production data against training data baselines and alerts when drift exceeds configured thresholds. Integrates with CloudWatch for alerting and provides dashboards for drift visualization. Supports custom metrics and drift detection algorithms.

Enables asynchronous model inference for long-running predictions by accepting requests from S3 input locations and writing predictions to S3 output locations. Clients submit inference requests with S3 URIs and receive output location URIs without waiting for completion. Useful for batch-like inference with unpredictable latency or large payloads. Automatically scales inference capacity based on queue depth.

Provides managed compute clusters optimized for large-scale model training and development, handling infrastructure provisioning, networking, and fault recovery. Clusters support distributed training frameworks (PyTorch, TensorFlow) and enable researchers to focus on model development without managing infrastructure. Includes automatic node provisioning, inter-node networking optimization, and checkpoint management.

Converts trained model artifacts into production-ready inference endpoints through a declarative deployment abstraction that handles container orchestration, auto-scaling configuration, and traffic routing. Users specify model artifact location, instance type, and initial capacity; SageMaker provisions infrastructure, exposes REST/gRPC endpoints, and manages rolling updates. Endpoints automatically scale based on request volume (auto-scaling specifics undocumented) and support A/B testing via traffic splitting.

Processes large datasets through trained models without maintaining persistent endpoints by submitting batch inference jobs that read input data from S3, invoke the model on mini-batches, and write predictions back to S3. Jobs automatically partition data across multiple instances for parallel processing and handle fault recovery. Useful for offline scoring, feature generation, or periodic model evaluation on large datasets.

Defines machine learning workflows as directed acyclic graphs (DAGs) where nodes represent training jobs, batch transforms, model evaluations, or conditional logic, and edges define data dependencies. Pipelines are defined declaratively (YAML or Python SDK), stored in version control, and executed on a managed orchestration engine that handles job scheduling, data passing between steps, and conditional branching. Integrates with SageMaker training, tuning, and deployment steps natively.

Generative AI assistant integrated into SageMaker notebooks and development environments that generates ML training code, SQL queries, and data pipeline definitions from natural language prompts. The assistant has built-in knowledge of AWS data sources (S3 buckets, Redshift schemas, DataZone catalogs) and SageMaker APIs, enabling context-aware code generation without manual schema specification. Supports data discovery queries like 'find tables with customer demographics' and generates corresponding SQL or data loading code.

Centralized registry built on Amazon DataZone that enables teams to register, catalog, and discover ML models, datasets, and data pipelines with metadata, lineage, and access controls. Assets are tagged with business context (owner, use case, quality metrics), searchable by natural language queries, and governed through approval workflows. Integrates with SageMaker training and deployment to track model lineage back to source datasets and training configurations.

Managed feature repository that stores pre-computed features in online (low-latency) and offline (batch) storage, enabling ML teams to define features once and reuse across training and inference. Features are organized into feature groups with schemas, versioning, and lineage tracking. Training jobs and inference endpoints can fetch features by entity ID without writing custom data loading code. Supports feature transformations and point-in-time joins for training data consistency.

Catalog of pre-trained foundation models and industry-specific ML solution templates that can be deployed with minimal configuration. Models include computer vision, NLP, and time-series models from AWS and third-party providers. Solutions are packaged with training notebooks, deployment code, and example datasets. Users can fine-tune pre-trained models on custom data or deploy them directly to endpoints.

Evaluates trained models against user-defined metrics (accuracy, precision, recall, F1, custom metrics) and compares performance across model versions or hyperparameter configurations. Evaluation can be triggered automatically after training or run on-demand against holdout test sets. Results are visualized in dashboards and can be used to gate model promotion (e.g., deploy only if accuracy improves by >1%).