The Pragmatic Engineer How AWS S3 is built
181 snips
Jan 21, 2026 
Mai-Lan Tomsen Bukovec, VP of Data and Analytics at AWS, leads the charge on Amazon S3, one of the largest distributed systems globally. She delves into S3’s impressive scale, with 500 trillion objects and millions of servers, and the evolution of its simplicity and consistency. Mai-Lan shares insights on optimizing costs, creating robust archival solutions like Glacier, and employing formal methods for correctness. She also discusses how S3 is evolving with new data primitives like tables and vectors, shaping the future of data management.
AI Snips
Chapters
Transcript
Episode notes
S3's Unimaginable Physical Scale
- S3 stores hundreds of exabytes across tens of millions of drives and millions of servers.
- Mai-Lan Tomsen Bukovec says this scale makes everyday failures routine and shapes their engineering decisions.
How S3 Began With Eventual Consistency
- S3 launched in 2006 to store unstructured assets like images and backups and optimized for availability and durability.
- The original design used eventual consistency because human-facing UX could tolerate delayed listings.
From Blobs To Tables And Vectors
- Customers began storing tabular data (Parquet) on S3, spawning Iceberg and S3 Tables.
- Mai-Lan explains S3 evolved to support native table semantics and vectors as first‑class primitives.