OpenAI builds custom Git backend to handle massive agentic coding volume

Scaling for the agent era

OpenAI recently detailed the creation of Pando, a custom Git backend designed to power its internal monorepo. The company needed a solution to handle the massive influx of commits generated by both human developers and AI agents. As a result of this shift toward agentic coding, OpenAI expects its quarterly commits to increase by over 2,400% by the end of this year.

The move comes as the company faced reliability issues with standard Git infrastructure. The team reported that their existing backend reliability had fallen below 90%, plagued by slow fetches and failed merges. Rather than abandoning the Git ecosystem entirely, they chose to rebuild the server architecture to solve these specific bottlenecks.

The team noted that the core Git protocol was not the primary problem. Instead, they identified that conventional Git servers were not built to handle the scale of a modern monorepo. By building Pando, they aimed to keep the familiar developer experience while fundamentally changing how the data is served behind the scenes.

How Pando handles data

Pando achieves its speed by removing two standard requirements of traditional Git servers. First, it does not require every branch in a repository to be in a consistent state at the same time. Second, it avoids building a custom packfile for every single client that fetches code, which is a common and expensive operation in standard Git.

In a typical setup, pushes to different branches can end up waiting on each other because the server tries to order updates across the entire repository. Pando handles each branch independently, allowing pushes to unrelated branches to proceed without coordination. While this means a client might see different branches at slightly different points in time, the team argues that users only care about the order of updates on a single branch.

When a client pushes data, Pando validates the incoming packfile as it streams into blob storage without unpacking it into a local repository. It then records the new branch tip and metadata in a separate store. This allows Pando to serve download links through a protocol feature called packfile-uri, which lets clients download the specific pieces of data they need from regional caches.

Optimizing for high volume

To manage the sheer volume of data, OpenAI uses a caching service called Express. This service maintains a hit rate above 99% by serving packs from regional caches. On standard cloud VMs, clones and fetches can max out a 10 Gbps network link, but a busy monorepo would normally force clients to download tens of thousands of small packs every day.

To prevent this, Pando uses background workers to merge small packs into larger ones over time. This process is similar to how a log-structured database compacts its data to improve efficiency. By doing this, Pando can figure out which packs a client needs in logarithmic time, significantly speeding up the retrieval process for large repositories.

Pando also avoids the heavy compute and disk I/O required by regular Git servers for repacking and garbage collection. The system never throws a pack away, keeping them at every compaction level. This ensures that the server remains fast even as the repository grows to include millions of branches and massive amounts of historical data.

Impact on the Git project

While Pando is a private backend, OpenAI is contributing many of its technical fixes back to the main Git project. Because Pando relies heavily on the packfile-uri feature, the team had to fix several rough edges in the Git client. These fixes include handling concurrent downloads, cleaning up interrupted transfers, and adding tracing to identify where slow fetches are occurring.

Some of these improvements have already been sent upstream to the official Git project. Others remain in OpenAI's public repository for now, including a new feature called "explicit haves." This feature allows a server to tell a pushing client where a branch points without forcing the client to leave its current history out of the pack.

OpenAI is currently tracking all of these technical adjustments in a public patch tracker. While they are not open-sourcing the Pando backend itself, they are looking into ways to share the underlying technology with their users. This work highlights the growing need for infrastructure that can support the rapid-fire pace of AI-driven software engineering.