Simon Eskildsen

Eskildsen worked at Shopify from 2013 to 2021 as its infrastructure grew from about a thousand to a million requests per second. He then helped several friends’ companies scale their systems, including Readwise, Causal and Replicate. He cofounded turbopuffer in 2023 after encountering the cost of adding semantic search to a large document collection.[1][2]

Estimate performance before choosing an implementation

Napkin Math collects rounded measurements, benchmarks and worked problems for reasoning about latency, throughput and cost. Instead of memorizing a precise answer for every system, a reader combines estimates for memory, disks, networks and computation. The current tables identify the machines and measurement period, and the repository explains how to rerun the benchmarks. The numbers are references for reasoning, with workload and hardware context.[3]

Use object storage as the durable search foundation

In his turbopuffer design essay, Eskildsen explains keeping data in object storage while serving frequently accessed data through memory and SSD caches. The design responds to the economics of large searchable collections, including vectors used for semantic retrieval. It accepts a tradeoff between cold access and warm queries rather than keeping every document in the most expensive storage tier.[2]