OpenAI engineers develop Pando to handle massive growth in source control demands

OpenAI has engineered Pando, a novel backend for Git repositories, to manage the escalating demands of its rapidly growing monorepo, which was necessitated by the scale enabled by Codex. This new system was implemented without disrupting existing engineering workflows. Pando achieves two primary objectives: supporting horizontally scalable writes for concurrent branch pushes and maximizing read throughput to saturate network bandwidth. In production, Pando reportedly reduced median push latency by nearly fivefold and now handles close to a petabyte of Git data daily. The development addressed performance bottlenecks that surfaced toward the end of 2025, where the existing systems experienced slow fetches, pushes, and reliability dipping below 90%. Pando distinguishes itself from conventional scaling methods like three-phase commit (3PC) and shared write-ahead logs (WAL) by adopting a relaxed causal consistency model for writes and reusing packfiles for reads. This design allows independent updates to unrelated branches to proceed in parallel. By storing incoming packfiles as immutable blobs and recording ref versions in a metadata store, Pando scales both storage and metadata horizontally. This architecture enables clients to reconstruct repository states by replaying ref versions, ensuring serializability for individual ref updates while supporting massive concurrent client activity.