AWS re:Invent 2025 - Amazon S3 performance: Architecture, design, and optimization (STG335)

Optimizing Amazon S3 Performance: Architecture, Design, and Strategies

S3 Architecture and Scale

• S3 is composed of three main components:
  • Front-end services that handle requests and orchestrate processing
  • Index that maps object metadata to storage locations
  • Storage subsystem that manages disks and data placement
• S3 currently stores over 500 trillion objects, equating to hundreds of exabytes of storage
• S3 serves over 200 million requests per second, with customers running over 1 million data lakes on AWS

Driving High Throughput on S3

• Leverage S3's scale by parallelizing connections across multiple IP addresses
• Use multipart uploads to split large objects into smaller parts and upload in parallel
  • Improves throughput and recovery time if individual connections fail
  • Allows starting uploads before having all data in memory
• Use range gets to download large objects in parallel across multiple connections
  • Reduces download time from 5 seconds to 1 second in the example

Architectural Patterns for High Performance

• Use the AWS Common Runtime (CRT) client library to automatically handle connection management, parallelization, retries, and more
  • Provides a "target throughput" configuration to automatically scale connections
  • Available in SDKs like Java, C++, Python, and the AWS CLI
• Leverage S3's prefix-based partitioning to scale throughput
  • Each prefix can handle up to 3,500 puts or 5,500 gets per second
  • Strategically structure prefixes to avoid "sharp edges" when data patterns change over time

S3 Express One Zone

• S3 Express One Zone is a high-performance storage class offering:
  • Single-digit millisecond access times
  • Up to 2 million requests per second per directory bucket
  • Ability to append data to existing objects
  • Constant-time object renaming
• Key use cases:
  • Machine learning training
  • Interactive querying and analytics
  • Log and media streaming
  • Model loading for inference pipelines
• Architectural considerations:
  • Single availability zone storage
  • Directory buckets with pre-scaled throughput
  • Session-based authentication for low-latency access

Key Takeaways

• Leverage S3's scale and parallelism to achieve high throughput and low latency
• Use the AWS CRT client library to simplify performance optimization
• Structure prefix hierarchies strategically to avoid throughput bottlenecks
• Consider S3 Express One Zone for latency-sensitive, bursty workloads
• Choose storage classes and architectural patterns based on specific performance requirements