Back to news
researchOpenAI2026-11-03

OpenAI rebuilds Git backend as Pando, cutting median push latency nearly 80 percent

OpenAI disclosed on November 2 that its in-house Git backend Pando now serves all development traffic on its monorepo, cutting median push latency nearly 80 percent and serving close to 1 PB of Git data daily.

OpenAI published the engineering article Scaling Git at OpenAI on November 2, disclosing Pando, a new Git backend it built for its internal monorepo that now serves all development traffic on that repository.

OpenAI said that from around the end of 2025 the existing source control system struggled to keep up with monorepo concurrency, with fetch and push slowing, merges failing and reliability at times below 90 percent. The cause is that one engineer can have multiple agents preparing changes at once, each push triggers multiple CI jobs pulling the same commit, and branches, writes and reads all grow together. Using Q1 2024 as a baseline, commits landed on the monorepo trunk rose every quarter, reaching 1378 percent in Q3 2026.

The key trade-off in Pando is that a single Git operation does not need to see all repository refs at one instant snapshot. It records versions per ref and protects ref pointers with atomic conditional updates, so two agents pushing different branches can proceed in parallel. Packs are saved on write and reused on read, and when a push arrives Pando validates the packfile during streaming receipt and writes it as an immutable object to blob storage. On the read path it returns download locations through the Git packfile-uri mechanism, with a service named Express serving bytes from regional caches achieving over 99 percent cache hit rate, and clone and fetch able to saturate a 10 Gbps network link on standard cloud virtual machines.

Pando is implemented in Rust and took 10 weeks from project start to production, with one engineer designing the core, agent teams helping implement it and a small number of engineers handling deployment, and began carrying production traffic in April 2026. Reported results include median push latency down nearly 80 percent, over 57 TB of Git pack data served in a single hour with more than 99 percent of bytes coming from local disk cache. These are disclosures from OpenAI internal environments with no reproducible public benchmark provided.

OpenAIPandoGit 后端monorepo工程基础设施