OpenAI develops Pando, a new Git backend, to handle increased monorepo commits

Addressing Git Bottlenecks at OpenAI

OpenAI announced the development of Pando, a new Git backend designed to manage the growing volume of commits to its monorepo. The company observed a substantial increase in code contributions from both human developers and AI agents, with quarterly commits projected to rise by over 2,400% by the end of the year compared to Q1 2024. This surge in activity led to performance issues with their previous Git infrastructure.

Pando's Performance Improvements

Pando has significantly improved the efficiency of OpenAI's development workflow. The new backend reduced push latency by nearly five times and currently serves approximately one petabyte of Git data per day. These improvements were necessary as the reliability of their previous Git backend had fallen below 90%, causing slow fetches and failed merges.

Maintaining Git Compatibility

Despite the performance challenges, OpenAI chose to remain within the Git ecosystem. The team determined that the core Git client and protocol were not the primary issues, but rather the conventional architecture of Git servers. Pando was developed to address these server-side limitations while preserving compatibility with existing Git tools and workflows, allowing developers to continue building on established practices.

How Pando Operates

Pando deviates from traditional Git server requirements in two key areas. It does not insist on keeping every branch in a consistent state simultaneously, and it avoids building a custom packfile for every client fetching code. These architectural changes contribute to its improved performance and scalability in handling large-scale monorepos.