Scaling Git at OpenAI
OpenAI describes Pando, a Rust Git backend that scales concurrent branch pushes by giving up a repository-wide point-in-time view while keeping updates to each individual ref serializable.
It stores incoming packfiles in blob storage, tracks their dependencies, and reuses them for cached fetches; background compaction bounds the number of packs clients need. The trade-offs include potentially redundant data on pushes and different branches being observed at different times. OpenAI reports nearly 5× lower median push latency in its own deployment and nearly a petabyte of Git data served per day; the article’s interactive comparisons are simulations, not production benchmarks.
The team says it used deterministic simulation and formal methods alongside Git conformance tests. HN readers praised the technical detail but also questioned whether the growth in agent-generated work creates a maintenance liability.