Episode Summary
Executive Summary: The episode explores the hidden engineering behind Amazon S3 at extraordinary scale: 500T+ objects, hundreds of exabytes, millions of servers, and 200+ microservices. It covers S3’s shift from eventual to strong consistency, the use of formal methods to prove correctness, failure handling and correlated failures, and S3’s evolution into tables and vectors for modern analytics and AI.
Main Topics: S3 at planetary scale: Mylon describes the immense physical and logical scale behind S3, from hard drives and servers to regions and availability zones, and explains why customers experience it as simply 'it just works.' From eventual to strong consistency: The discussion details S3’s original eventual consistency model, why it optimized availability, and the major engineering effort required to introduce strong consistency without charging customers more or materially increasing latency. Formal methods and correctness: S3 relies on automated reasoning and formal methods to prove that consistency, replication, and APIs behave correctly, with proofs integrated into check-ins and critical code paths. Failure tolerance, correlated failures, and crash consistency: The episode explains how S3 engineers design for the reality that servers fail constantly, using replication, quorum, cache coherency, failure allowances, and physical fault-domain separation to preserve service reliability. Economics and architecture of storage tiers: Mylon explains how S3’s pricing strategy and architecture enable Glacier, intelligent tiering, and lower cost of ownership by optimizing every layer of the stack, including hardware and operations. S3’s evolution into tables and vectors: S3 is expanding beyond blobs into structured and semantic data with S3 Tables (Iceberg/Parquet) and S3 Vectors, reflecting customer demand for SQL access and AI-native storage. Team culture and engineering principles: The S3 team balances 'respect what came before' with 'technically fearless,' emphasizing simplicity, ownership, curiosity, and continuous evolution while preserving core traits.
Key Arguments: S3’s scale is so large that reliability must be engineered as a default state, not an exceptional feature. Strong consistency required new distributed data structures and cache coherency protocols, but AWS chose to make it free and broadly available to all customers. Formal methods are essential at S3 scale because informal testing cannot cover the combinatorics of every edge case. Correlated failures are more dangerous than isolated failures, so S3 designs around fault domains and replication across zones and regions. Storage economics are a major product feature: lower prices, Glacier, and intelligent tiering encourage customers to retain and use more data. S3 is no longer just object storage; it is becoming a platform for tabular data and semantic/vector data to support analytics and AI. Customer behavior, not only product vision, drives roadmap decisions; new primitives are created when users adopt S3 in unexpected ways. Simplicity remains the guiding UX and systems principle even as the internals become highly complex.
Data Points: Objects stored in S3: over 500 trillion objects - Current scale of S3’s stored object count Data stored in S3: hundreds of exabytes - Current total storage capacity/usage Transactions served worldwide: hundreds of millions per second - Global request rate described by Mylon Annual requests processed: over a quadrillion per year - Total yearly S3 request volume Hard drives managed: tens of millions - Underlying hardware fleet Servers managed: millions - Underlying infrastructure scale Availability zones: 120 - S3 footprint across cloud fault domains Regions: 38 - Global deployment footprint Object storage launch price: 15 cents per GB per month - Initial S3 pricing in 2006 Typical storage rate at launch: 50–75 cents per GB per month - Competing rates at the time First-day developer signups: 12,000 - Adoption immediately after launch Current storage price mentioned: about 2 to 2.3 cents per GB per month - Current pricing reference in discussion Data growth rate cited from IDC: 27% year over year - Industry-wide data growth benchmark Intelligent tiering discount: up to 40% - Automatic discount when data is not accessed for a month S3 object size limit: 50 terabytes - Largest object size supported today Original object size limit: 5 terabytes - Size limit at S3 launch S3 Tables launch year: 2024 - Introduced in December 2024 S3 Tables feature additions: over 15 new features this year - Rapid post-launch iteration S3 Vectors preview launch: July 2025 - Preview launch timing mentioned in the episode S3 Vectors general availability: last week - GA status at time of conversation S3 regional microservices: over 200 microservices - Subsystems behind one S3 regional endpoint S3 Vectors scale per index: up to 2 billion vectors - Per-index capacity S3 Vectors bucket scale: up to 20 trillion vectors - Maximum bucket scale Warm query latency for S3 Vectors: 100 ms or less - Performance target for warm queries
Pivotal Quotes: "The mission of S3 is to provide the best storage service on the planet." — Mylon: Explaining why S3 keeps lowering prices and expanding capability "Scale is to your advantage." — Mylon: Describing an S3 engineering tenet that shapes system design "Respect what came before." — Host paraphrasing Amazon engineering tenet: Discussing the cultural balance between stability and innovation on the S3 team
Implications: S3 is evolving from object storage into a foundational data platform for analytics and AI. Its engineering model suggests that correctness, fault-domain design, and formal verification will become increasingly important at large-scale infrastructure companies.