OpenAI deploys Pando Git backend across its monorepo

OpenAI has deployed Pando, a new Git-compatible repository backend, to power all development in its monorepo, rebuilding its source-control infrastructure to support the concurrent engineering work generated by Codex.

The company says the switch cut median push latency by nearly 80%, improved merge reliability and gave the system capacity to deliver a petabyte of Git data per day with a handful of CPU cores. Those are OpenAI-reported production measurements and were not independently audited in the sources reviewed.

OpenAI said its previous monorepo infrastructure had struggled as Codex enabled engineers and agents to work on more changes in parallel. The resulting increase in branches, pushes, merges, fetches and clones caused slow operations and failed merges, with reliability at times falling below 90%, according to the company.

Pando changes how work is coordinated. Rather than enforce a single repository-wide order for every update, it serializes changes to each individual Git reference, such as a branch pointer. Pushes to unrelated branches can therefore proceed independently. A client reading several branches is not guaranteed to see all of them at the same point in time, but a client fetching one branch receives the history required for that revision.

For each push, Pando validates the incoming Git packfile while streaming it to blob storage. It then records the new branch version, the packfile and its dependencies in a horizontally scalable metadata store before atomically updating the branch pointer. OpenAI says the synchronous path requires a constant number of metadata reads and writes, regardless of push size.

The read path avoids generating a custom packfile for every client. Pando instead selects stored packs that cover the difference between the client’s existing history and requested branch version, then returns their locations through Git’s protocol-v2 packfile-URI feature. The Git project’s documentation says that mechanism lets a server substitute URIs for part of a packfile response, which can shift delivery to caches or a content delivery network and reduce server bandwidth and CPU work.

OpenAI calls its regional pack-delivery layer Express. The company reports cache-hit rates above 99% and says clone and fetch tests on standard cloud virtual machines saturated 10 Gbps network links. It also says Pando has served more than 57 terabytes of Git packs in one hour, with more than 99% of those bytes coming from local disk cache.

Reusing pushed packs creates a tradeoff. When a client pushes a feature branch, Pando normally advertises only a designated trunk branch rather than every reference in the repository. That keeps coordination bounded and makes the resulting pack reusable, but can cause clients to resend objects. OpenAI says the median push pack in its monorepo remained below 20 KiB and the 99th percentile below 3 MiB.

Background workers compact the small packs created by pushes into progressively larger covers. OpenAI says its monorepo currently uses levels through L4, improving compression while allowing fetch planning to skip across branch versions in logarithmic time.

Pando was built in Rust and moved from kickoff to production in 10 weeks, according to OpenAI. It began serving production traffic in April 2026. The company first ran it as a read- and write-through proxy that kept the previous system synchronized, then migrated Git clients and hundreds of integrations in phases without downtime.

OpenAI has also published its Git fork and a patch tracker for client changes and fixes prompted by the project, including work on concurrent packfile-URI downloads and tracing. Pando itself has not been publicly released; OpenAI said only that it is exploring ways to share the technology.