We built Pando, a new Git backend for our monorepo, to handle more pushes and fetches without changing how our engineers and agents use Git.

Pando cut push latency by ~80% and now serves nearly a petabyte of Git data a day. We can grow Git capacity alongside agent usage.

If engineers push to different branches, Pando can update both in parallel. If we push to the same branch, it still puts those updates in order.

On a fetch, you get the history for each branch tip, but two tips may reflect different points in time.

The Git client tells Pando which commits it has and which it wants. Pando returns links to stored packfiles that cover the gap, including missing branch dependencies.

These packfiles may include some objects a client already has; but by reusing these across many readers, Pando provides CDN-like performance.

We’ve also contributed fixes back to Git for cases like cleaning up after an interrupted packfile download.

openai.com/index/scaling-…