OpenAI Details Pando, Its In-House Git Backend: Median Push Latency Down Nearly 5x, Nearly 1 PB Served Daily, Built With Codex in 10 Weeks

Background and Industry Pain Points

Coding agents have made it normal for one engineer to run several branches at once, and every push can set off several CI jobs fetching the same commit, so source control has become a new bottleneck. In a November 2 engineering post, OpenAI says that from late last year the systems behind its monorepo could not keep up with the concurrency Codex created: fetches and pushes slowed, merges failed, and reliability at times fell below 90%. Instead of leaving Git, OpenAI rebuilt the server side as Pando, a new backend compatible with the Git protocol and client, so engineers and agents use Git exactly as before.

Architecture Highlights and Underlying Mechanism

Pando drops two things most Git servers treat as requirements. First, repository-wide consistency: it coordinates each branch independently, so only pushes to the same branch conflict, at the cost that a reader of several branches may see them at different points in time. Second, a unique pack per client: pushed packfiles are validated while streamed into blob storage, a metadata store records ref versions and dependencies, and later fetches select existing packs and hand clients packfile-uri links to a regional cache called Express, with hit rates above 99%. Background compaction, similar to an LSM tree, merges small packs into larger covers (up to L4 today) so fetch planning runs in logarithmic time.

Benchmarks and Measured Results

All figures are OpenAI's own: median push latency down nearly 5x (about 80%); nearly 1 PB of Git data per day on a handful of CPU cores; a peak of more than 57 TB in one hour with over 99% of bytes from local disk cache; and clones and fetches that saturate 10 Gbps links on standard cloud VMs. The comparison with three-phase commit and write-ahead-log designs is a teaching simulation that OpenAI says is not a production benchmark. Independent reproduction: not published.

Developer Playbook and Getting Started

Pando is internal to OpenAI for now; the company says it is exploring ways to share the technology with users, and there is no product to try yet. What you can use today is OpenAI's public Git fork and the packfile-uri improvements already upstream. For teams running many coding agents, the useful lessons are per-branch rather than per-repository consistency, advertising only trunk on pushes so packs can be reused, and migrating behind a read- and write-through proxy without downtime. The build process is notable too: one engineer designed the core, teams of Codex agents implemented it, and agents wrote Lean and TLA+ specs to formally check the system.