Amazon SageMaker

Learning Objectives

• Understand what SageMaker is and how it fits into AWS's AI product portfolio alongside Bedrock
• Identify SageMaker's key components: Studio, JumpStart, Pipelines, and HyperPod
• Evaluate when to use SageMaker versus Bedrock for ML workloads

What Is Amazon SageMaker?

Amazon SageMaker is AWS's fully managed platform for the entire machine learning lifecycle — from data preparation and labeling through model training, tuning, deployment, and monitoring. While Amazon Bedrock provides model-as-a-service (call an API, get a response), SageMaker is the platform for teams that need to build, train, and operate their own ML systems.

SageMaker has been available since 2017 and is one of the most widely used enterprise ML platforms globally. It provides the infrastructure, tools, and managed services that let ML teams focus on model development rather than infrastructure management.

Key Components

SageMaker Studio

A web-based IDE for ML development:

• Jupyter notebooks with managed compute (no infrastructure setup)
• Visual experiment tracking and model comparison
• Integrated debugging and profiling tools
• Collaboration features for ML teams

SageMaker JumpStart

A model hub and solution catalog:

• Hundreds of pre-trained models (Llama, Mistral, Stable Diffusion, Hugging Face models)
• One-click deployment for foundation models
• Pre-built ML solutions for common use cases (fraud detection, demand forecasting, image classification)
• Fine-tuning workflows for customizing models on your data

SageMaker Pipelines

MLOps automation:

• Define ML workflows as code (data processing, training, evaluation, deployment)
• Automated model retraining on new data
• Model registry for versioning and approval workflows
• Integration with CI/CD for production ML deployments

SageMaker HyperPod

Distributed training infrastructure:

• Managed GPU/Trainium clusters for training large models
• Automatic node health monitoring and replacement
• Optimized for multi-node training of foundation models
• Reduces the operational burden of managing training clusters

SageMaker Canvas

No-code ML for business analysts:

• Visual interface for building ML models without writing code
• Point-and-click data import, model training, and prediction generation
• Supports tabular data (forecasting, classification, regression)
• Connects to data in S3, Redshift, and other AWS data stores

Pricing

SageMaker uses pay-as-you-go pricing across multiple dimensions:

SageMaker vs. Alternatives

Strengths

• End-to-end ML platform — covers data prep, training, deployment, monitoring, and MLOps in one service
• JumpStart model hub — hundreds of pre-trained models deployable with one click
• HyperPod for large-scale training — managed clusters for training foundation models on GPU/Trainium
• Canvas for no-code ML — accessible to business analysts without ML expertise
• Deep AWS integration — native connections to S3, Redshift, Glue, Lambda, and other AWS services
• Mature and battle-tested — available since 2017; used by thousands of enterprises in production

Limitations & Considerations

• AWS lock-in — deeply integrated with AWS; migrating SageMaker workloads to another cloud is significant effort
• Complexity — the breadth of features means a steep learning curve; many teams only use a fraction of capabilities
• Cost management — pay-per-use across many dimensions (notebooks, training, inference, storage) can be difficult to predict
• Not for simple LLM use cases — if you just need to call an LLM API, use Bedrock instead; SageMaker is for custom ML development
• Overhead for small teams — the platform is designed for enterprise ML teams; solo developers may find it heavyweight

Key Takeaways

• Amazon SageMaker is AWS's fully managed ML platform — covering the entire lifecycle from data preparation through model training, deployment, and monitoring
• Distinct from Bedrock: SageMaker is for building and training custom ML systems; Bedrock is for consuming foundation models via API
• JumpStart provides one-click access to hundreds of pre-trained models; HyperPod manages distributed training infrastructure; Canvas offers no-code ML for business users
• Most compelling for enterprise ML teams on AWS; solo developers and simple LLM use cases are better served by Bedrock or Hugging Face