OpenAI Builds Custom Git Backend Pando to Cut Push Latency by Nearly 80%

OpenAI revealed its in-house Git backend system, Pando. As Codex-assisted engineering generated a surge in branches, pushes, and continuous integration (CI) traffic within its monorepo, the prior architecture faced latency and reliability bottlenecks in late 2025. Pando preserves standard Git compatibility while transitioning branch coordination to causal consistency, enabling unrelated branches to update in parallel. Additionally, Pando retains received packfiles, merges them in the background, and distributes them via packfile-uri caching, significantly cutting real-time server computation overhead. OpenAI stated that median push latency dropped by nearly 80%, with over 57 TB transferred in a single hour and more than 99% served from local cache. Pando currently supports internal monorepo operations and has not been announced for external release.

Key takeaways

• OpenAI built Pando to address performance and reliability bottlenecks caused by Codex-driven high-concurrency workflows.
• The architecture adopts causal consistency for parallel branch updates, trading away global point-in-time multi-branch snapshots on the client side.
• Pando reuses received packfiles with caching and packfile-uri distribution, avoiding repetitive server-side packaging overhead.
• Handling production traffic since April of that year, Pando reduced median push latency by nearly 80%; the system itself has not been announced for external availability.