Overview
- OpenAI disclosed Monday that Pando, a custom Git backend it put into production in April 2026, cut median push latency by about 80% and now serves nearly one petabyte of Git data per day.
- Pando preserves Git protocol compatibility while changing server-side coordination by handling each branch independently and advertising only the trunk branch so servers can reuse and cache packfiles.
- The system uses packfile-uri links and blob storage to serve prebuilt packfiles from an Express delivery layer with regional caches, achieving over 99% cache-hit rates and peaks above 57 TB of pack data served in an hour.
- OpenAI built Pando in roughly 10 weeks in Rust with a single lead engineer and substantial implementation help from Codex agents, and it has contributed many client- and protocol fixes back to the upstream Git project while keeping Pando proprietary.
- The move highlights a broader industry shift: Git at very large scale is becoming a bottleneck, prompting firms and startups to redesign version-control infrastructure and indicating potential follow-on changes in client behavior, tooling, and upstream Git protocols.