ResearchOpenAI

OpenAI builds Pando to scale Git for Codex

Original title: Scaling Git at OpenAI

Plain language

What it is: OpenAI created a new system called Pando to run its internal Git repositories. It helps the company handle the massive amount of code development enabled by Codex.
How it works: The system allows different teams and automated agents to save their work to different branches at the same time without waiting for each other. It also reuses pre-packaged files to let users download code updates much faster.
Where it is available: This system is currently used internally to power all software development in OpenAI's main code repository.
Example: OpenAI uses Pando to let its engineers and automated agents push code changes to different branches simultaneously, which cut their average wait time for saving code by nearly five times.

What it is: OpenAI developed Pando, a Git-compatible backend designed to scale source control for high-concurrency workflows. It replaces traditional Git backends for OpenAI's monorepo.
How it works: Pando uses a causal consistency model where updates to unrelated branches are coordinated independently and stored as immutable blobs in a blob store. It scales reads by reusing packfiles across fetches instead of dynamically generating unique packfiles for each client.
Where it is available: Pando is deployed internally and powers all development within OpenAI's monorepo.
Example: The system is used to handle concurrent pushes and fetches from developers, research jobs, and continuous integration systems, processing nearly a petabyte of Git data daily.

The headline and explanation on this page are based on OpenAI's post. The full details are in the original.

Scaling Git at OpenAI