Skip to content

Back to the catalog

openai.comarticle

Scaling Git at OpenAI

Pando is OpenAI's new Git backend, built to handle the push and fetch load Codex puts on its monorepo. It stores pushed packfiles in blob storage, tracks per-branch ref versions with causal consistency, and reuses packs across fetches. OpenAI reports nearly 5x lower median push latency and about a petabyte of Git data served per day. It was built in 10 weeks with Codex agents.

TL;DR Describes Pando, OpenAI's Git-compatible backend that scales concurrent pushes and cached pack fetches for agent-driven monorepo traffic.
Open original ↗

Organisations named here

  • OpenAI
  • Git community

Technologies named here

  • GPT-6 Astra
  • Codex
  • Git
  • Lean
  • Rust
  • ChatGPT desktop app
  • ChatGPT
  • Daybreak
  • Daybreak-capable models
  • Express
  • packfile-uri
  • Pando
  • TLA+
  • Turtle

Techniques named here

  • Compare-and-swap
  • Consistent hashing
  • deterministic simulation testing
  • Log-structured merge-tree compaction
  • three-phase commit
  • write-ahead log

Concepts named here

  • causal consistency
  • Explicit haves
  • Packfile reuse
  • Read-your-own-write consistency
  • Ref advertisement
  • Ref versions

Related in ai-infra