OpenAI has built Pando, a new backend for Git repositories, to handle the increased volume of parallel development enabled by Codex. The system now powers all development in the company’s monorepo and was introduced without interrupting engineering work or changing how staff use Git.
OpenAI says Pando was designed to support simultaneous pushes to different branches while serving repository data quickly to many readers. In production, it has cut median push latency by nearly five times and delivers nearly a petabyte of Git data each day.
The company says growing development activity brought more branches, pushes, merges, fetches and clones from engineering environments, research jobs and CI. Its previous monorepo systems began struggling around the end of last year, with slow operations, failed merges and reliability that at times fell below 90%.
Pando retains compatibility with the Git protocol and client. Its design allows updates to unrelated branches to proceed independently, while reusing stored packfiles across fetches rather than generating a separate packfile for every client. This can mean a client downloads some objects it already has, but it reduces the computation required to serve the same data to many readers.
The architecture stores incoming packfiles in blob storage and records branch versions in a metadata store. On a fetch, Pando selects the stored packs a client needs, while a separate service called Express serves them through regional caches. OpenAI says the approach avoids requiring servers to unpack and repack data into local repositories for each operation.
Pando uses causal consistency rather than providing a single point-in-time view of every branch across the repository. Updates to an individual branch remain serializable, but clients reading several branches may see them at different points in time. OpenAI says this tradeoff lets unrelated branch updates avoid contending on a repository-wide coordination point.
The company says simulations and testing on its monorepo showed Pando handling many thousands of concurrent clients without performance degradation under load. OpenAI also plans to continue contributing improvements to Git, saying that maintaining compatibility allows engineers and agents to adopt the system through existing workflows.
Sources: OpenAI News.