Artificial Intelligence
OpenAI details Pando Git backend
OpenAI said its Pando system was built to handle more commits from developers and agents while keeping Git compatibility, reducing push latency and changing how its servers store and serve repository data.
- OpenAI said Pando cut push latency by nearly five times and currently serves close to a petabyte of Git data each day.
- The company said quarterly commits to its monorepo will be more than 2,400% higher by the end of this year than in the first quarter of 2024.
- OpenAI said its older Git backend had fallen below 90% reliability, with slow fetches and failed merges.
- The company said Pando went into production in April and was built in Rust in 10 weeks.
- OpenAI said it has sent some of its Git changes upstream, but Pando itself is not open sourced.
OpenAI said on Monday that it had built Pando, a new Git backend for its monorepo, to cope with the volume of commits made by its developers and agents. The company said the system reduced push latency by nearly five times and already handled close to a petabyte of Git data a day.
The company said its quarterly commit volume would end the year more than 2,400% above the first quarter of 2024. It said the old backend had become less reliable, with slow fetches and merge failures, but it chose to keep using Git rather than move away from it.
OpenAI said Pando changed two assumptions common in conventional Git servers: it did not force every branch to stay in one consistent state, and it did not build a separate packfile for each client fetch. Instead, it validated incoming data as it streamed it into blob storage, stored metadata about branch tips and dependencies, and reused stored packs for later downloads.
The company said a caching layer called Express served those packs from regional caches with a hit rate above 99%, while background workers merged many small packs into larger ones over time. OpenAI also said developers still used standard Git commands, although it now ships its own Git build to developer machines and the ChatGPT desktop app.
Named in this story
People
- Thomas Dohmke
- was identified as the former GitHub chief behind Entire
- Aaron Friel
- co-authored OpenAI’s announcement about the backend
- Taylor Blau
- co-authored OpenAI’s announcement about the backend
- Ted Nyman
- co-authored OpenAI’s announcement about the backend
Companies
- OpenAI
- built the Pando backend and described how it changed its Git infrastructure
- GitHub
- a comparator whose infrastructure work was cited as another response to scale
- Entire
- a Git-based startup cited as another example in the market
- Cursor
- the company associated with the Origin project
- Anthropic
- was cited as taking a different approach with its own architecture
Products and systems
- Pando
- the new Git backend OpenAI built for its monorepo
- Git
- the version control system OpenAI kept compatibility with
- Origin
- a project from Cursor mentioned as another Git-based effort
- Codex
- the agent system used by OpenAI engineers to implement parts of the backend
- Express
- the caching service that served stored packs from regional caches
- ChatGPT desktop app
- now ships OpenAI’s Git build to developer machines
- packfile-uri
- the Git protocol feature Pando relied on for separate pack downloads
- Scion
- Anthropic’s unreleased backend mentioned in comparison
How the source tells it
Mostly a neutral technical explainer that relayed OpenAI’s own performance claims and design choices without strong emotive framing.
No emotive framing was found in the original.