OpenAI and Anthropic Take Opposite Routes to Scaling Git
A source report compares OpenAI’s Pando, a Git-compatible hosting backend, with Anthropic’s Scion, which moves commits, rebases, merges, history queries, and diffs to the server. OpenAI says Pando coordinates branches independently and reuses packfiles, reducing median push latency to one-fifth of the prior level while replacing repository-level consistency with causal consistency. Anthropic says Scion eliminates the merge queue’s global lock by moving validation to after merging and before deployment, while reducing CI checkout times and infrastructure use. The performance figures are vendor-reported claims presented in the report.
The comparison illustrates how agent-generated code increases pressure on source-control architecture. Pando preserves Git commands and protocols but requires explicit checks for workflows that depend on a consistent cross-branch snapshot, while Scion replaces major client-side functions and accepts a period in which the main branch may be broken after merges.
摘要:OpenAI 把中位推送延迟压到原来的五分之一,代价是仓库不再有一个统一时间点;Anthropic 干脆换掉了 Git,换来的是取消合并队列的全局锁。同一个症状,两家给出了相反的诊断。
下午四点,你在等 CI。三个分支挂着,push 上去要等十几秒,fetch 卡在 resolving deltas 上转圈。
你骂了一句:Git 真慢。
把这个量级放大到几千倍,就是 OpenAI 去年年底的状态。而他们给出的结论,跟你的第一反应正好相反。
先看两家被压成什么样
两个月前,OpenAI 工程博客发了一篇《Scaling Git at OpenAI》。11 月 2 日,作者是 Aaron Friel、Taylor Blau 和 Ted Nyman 三个人。