OpenAI releases Pando, a new Git backend for agent-scale parallel pushes
The interesting shift is architectural: instead of abandoning Git workflows, OpenAI traded repo-wide snapshot consistency for write parallelism and reusable packfiles for read scaling — a pattern other teams running agent-driven CI will face as machine-generated commits multiply. Developers operating monorepos or agent pipelines can study the trade-offs around packfile reuse and ref-level serialization. Upstream contributions like parallel pack downloads and explicit haves may land in standard Git tooling.
OpenAI built Pando, a Git-compatible backend for its monorepo, after Codex-driven concurrent pushes and fetches degraded the old system's reliability below 90%. Pando keeps the Git protocol and clients unchanged while relaxing repo-wide snapshot consistency to per-ref causally consistent updates. In production it cut median push latency by about 80% (roughly 5x faster) and now serves nearly 1 PB of Git data daily. Written in Rust, it went from project start to production in about 10 weeks, with traffic cutover beginning around April 2026 and single-hour output exceeding 57 TB with over 99% local cache hits.
OpenAI 发布 Pando：给 Agent 规模并行推送重做的 Git 后端，中位推送约快 5 倍 Codex 把 OpenAI 单体仓推到「多分支并发写 + 海量并发读」的极限后，他们自研 Pando，兼容 Git 协议与客户端的新后端：分支级因果一致性、跨客户端复用 packfile；生产上中位推送延迟约降到原来的 1/5，日均可吐近 1 PB Git 数据。 openai.com/index/scaling-… Codex 让并行改动、分支、推送、CI 拉取同时放大，旧系统去年底起慢推慢拉、合并失败，可靠性一度低于 90%。他们没换掉 Git 工作流，而是换后端：无关分支可并行更新；读侧尽量用同一份已存 pack 吃满带宽。 对比过两条常见路：三阶段提交靠多副本多数派锁 ref，全仓更新偏串行，磁盘与协调压在推送路径；共享 WAL/清单能并行上传 pack，但无关分支仍抢同一 compare-and-swap，副本跟进建索引跟不上就拖「刚推完就能拉」。Pando 松掉两点假设（全仓某时刻一致、以及按客户端现算专用 pack）改成：单条 ref 可串行化，全仓只保因果一致；pack 进对象存储，元数据记 tip、依赖与包指针。推不同分支互不堵；读多条分支时，不必假设同一快照。 写入默认只广告主干作基，广告不随百万级分支膨胀，pack 更易复用；代价是叠分支可能多传。他们仓里中位 pack <20 KiB，P99 <3 MiB。读侧把 have/want 映射成已存 ref version，靠 packfile-uri 下发可缓存包（Express 命中率 >99%），标准机可吃满 10 Gbps。后台按 LSM 思路把 L0 合到更高等级（至 L4），既压缩又当跳跃索引，规划约 O(log Δ)。 约 10 周用 Codex 从立项到生产（约 4 月起接流量），Rust 重写，先双写代理再分期切流。效果含合并更稳、单小时曾送出 >57 TB（>99% 本地缓存）。改进回馈上游（并行下包、explicit haves 等）。仿真与生产口径需分开看；核心取舍是用因果一致换写并行、用可复用 pack 换读扩展，同时保住 Git 客户端兼容。 OpenAI Developers @OpenAIDevs 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. Your browser does not support the video tag. 🔗 View on Twitter 🔗 View Quoted Tweet 💬 13 🔄 0 ❤️ 12 👀 989 📊 12 ⚡
