
Hardware latency and speed-of-light performance

Distributed databases with Peter Mattis
Topic 5 of 13
Published Sep 30, 2026 · Clip 24:01–30:05 (6:04)
Peter explains that distributed file system latency is bounded by hardware — S3 and GCS have 20–30ms first-read latencies on hard disks, while NVMe SSDs are orders of magnitude faster at tens of microseconds. He stresses designing to machine scale rather than human-perceptible latency, calling these "speed of light numbers," and shares the fact that sending data via satellites (Starlink-style) beats fiber for global transfer since light is faster in vacuum. He also discusses performance layers: avoiding synchronization, lock-free programming, and respecting the full cache hierarchy from registers through L1/L2/L3 to disk.