OpenAI Builds In-House Git Backend Pando: Cuts Push Latency by Nearly 80%, Processes Close to 1PB of Data Daily
OpenAI has developed Pando, a Git-compatible backend purpose-built for its monorepo. It handles branches independently, reuses packfiles across separate fetch requests, cuts median push latency by nearly 80%, and currently processes close to 1PB of Git data every day.
As reported by Michal Sutter of MarkTechPost, OpenAI built a Git-compatible backend named Pando for its in-house monorepo. It does not bypass the Git protocol, and its core goal is to make Git run far more smoothly on ultra-large repositories.
A monorepo stores a large number of different projects inside a single repository. This approach is very low-friction to manage when you only have a small number of projects. Once the project count, branch volume, and commit frequency all scale up, Git's native fetch, push and branch management operations turn into critical infrastructure pain points. OpenAI's operational scale has clearly reached this exact stage.
The key highlights of Pando are as follows:
- Independent branch processing.
- Reuse of already stored packfiles across multiple fetch requests. A packfile is the data file Git uses to package objects, and reusing packfiles eliminates redundant data transfer and duplicate computation work.
- Median push latency drops by nearly 80%.
- It now processes close to 1PB of Git data on a daily basis.
All these metrics come from the MarkTechPost report, which does not go into further details about Pando's implementation. What is confirmed is that Pando remains fully Git-compatible — this is a very important design choice. The biggest risk for large teams switching version control backends is having to completely rebuild their entire toolchain and abandon long-established working habits. Full Git compatibility keeps migration costs fully under control.
The truly notable takeaway from this launch is that once a monorepo grows past a certain size, Git's default native implementation becomes a performance bottleneck, to the point that even OpenAI has to build a dedicated custom backend layer on top of it. Most common projects will never come close to hitting this level of scale. For professionals working on platform engineering, developer tools and code hosting services, this is a clear signal: once repository size crosses a certain threshold, version control itself has to be designed as a full-fledged standalone product.
Original article: OpenAI Introduces Pando: A Git Backend That Cuts Median Push Latency ~80%
Author: Michal Sutter
发布时间: 2026-11-03 13:51