Architecting for high availability on Amazon S3
summary
Amazon S3 claims multi‑decade, multi‑region availability at scale — over 15 years, hundreds of trillions of objects, across 20+ regions and 60+ availability zones with storage classes replicating across >3 AZs and four nines availability. Availability is measured as request error rate per-customer over fine‑grained intervals, backed by continuous canaries, an elastic repair fleet that auto‑scales re‑replication, and cell-based deployments with rollback‑first releases, SDK tuning (JVM DNS TTL ~60s) and least‑privilege operator tooling; these practices reduce blast radius and speed recovery.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:53Measuring availability
- 7:38How things fail
- 14:22Operational best practices
- 14:52Cellbased architecture
- 16:00Good tooling
- 16:45Mitigation recovery
- 17:35Organizational culture
- 18:46Applications
- 19:15Case Study 1
- 20:38Case Study 2
- 21:53Case Study 3
- 23:11Summary
- 23:55Questions
related talks
44mMulti‑region architectures on AWS
49mPatterns for resilient, highly available cloud architectures
30mResilient Architectures on Amazon EKS: High Availability Patterns and Strategies
44mAvailability, Latency and Cost: Withstanding Outages
01h 03mPatterns for Resilient Architecture
25mExplore fault isolation boundaries on AWS
53mResiliency and Availability Design Patterns for the Cloud
29mMulti Region Cloud Architecture for Data Residency Compliance