• Sources: primary
  • Summary: OpenAI built Pando because Codex-driven parallel work pushed monorepo reliability at times below 90 percent. Pando streams each pushed packfile to blob storage, records ref versions in a horizontally scalable metadata store, serializes updates only per ref, and reuses stored packs across fetches instead of building one per client. OpenAI says it cut median push latency by nearly 5x, serves nearly a petabyte of Git data a day, keeps the Git protocol and client unchanged, and will upstream client changes to Git.
  • Why it matters: It is a first-party design for forges under agent-scale write load, and it trades repository-wide ref consistency for parallel pushes, a tradeoff any tool reading several branches at once must tolerate.